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/evsel.h"
  47 #include "util/debug.h"
  48 #include "util/header.h"
  49 #include "util/cpumap.h"
  50 #include "util/thread.h"
  51
  52 #include <sys/prctl.h>
  53 #include <math.h>
  54 #include <locale.h>
  55
  56 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  57
  58 #define DEFAULT_SEPARATOR       " "
  59
  60 static struct perf_event_attr default_attrs[] = {
  61
  62   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  63   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  64   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  65   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  66
  67   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  68   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  69   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  70   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  71   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES        },
  72   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES            },
  73
  74 };
  75
  76 static bool                     system_wide                     =  false;
  77 static int                      nr_cpus                         =  0;
  78 static int                      run_idx                         =  0;
  79
  80 static int                      run_count                       =  1;
  81 static bool                     no_inherit                      = false;
  82 static bool                     scale                           =  true;
  83 static bool                     no_aggr                         = false;
  84 static pid_t                    target_pid                      = -1;
  85 static pid_t                    target_tid                      = -1;
  86 static pid_t                    *all_tids                       =  NULL;
  87 static int                      thread_num                      =  0;
  88 static pid_t                    child_pid                       = -1;
  89 static bool                     null_run                        =  false;
  90 static bool                     big_num                         =  true;
  91 static int                      big_num_opt                     =  -1;
  92 static const char               *cpu_list;
  93 static const char               *csv_sep                        = NULL;
  94 static bool                     csv_output                      = false;
  95
  96 struct cpu_counts {
  97         u64 val;
  98         u64 ena;
  99         u64 run;
 100 };
 101
 102 static volatile int done = 0;
 103
 104 struct stats
 105 {
 106         double n, mean, M2;
 107 };
 108
 109 struct perf_stat {
 110         struct stats      res_stats[3];
 111         int               scaled;
 112         struct cpu_counts cpu_counts[];
 113 };
 114
 115 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel, int ncpus)
 116 {
 117         size_t priv_size = (sizeof(struct perf_stat) +
 118                             (ncpus * sizeof(struct cpu_counts)));
 119         evsel->priv = zalloc(priv_size);
 120         return evsel->priv == NULL ? -ENOMEM : 0;
 121 }
 122
 123 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 124 {
 125         free(evsel->priv);
 126         evsel->priv = NULL;
 127 }
 128
 129 static void update_stats(struct stats *stats, u64 val)
 130 {
 131         double delta;
 132
 133         stats->n++;
 134         delta = val - stats->mean;
 135         stats->mean += delta / stats->n;
 136         stats->M2 += delta*(val - stats->mean);
 137 }
 138
 139 static double avg_stats(struct stats *stats)
 140 {
 141         return stats->mean;
 142 }
 143
 144 /*
 145  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 146  *
 147  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 148  * s^2 = -------------------------------
 149  *                  n - 1
 150  *
 151  * http://en.wikipedia.org/wiki/Stddev
 152  *
 153  * The std dev of the mean is related to the std dev by:
 154  *
 155  *             s
 156  * s_mean = -------
 157  *          sqrt(n)
 158  *
 159  */
 160 static double stddev_stats(struct stats *stats)
 161 {
 162         double variance = stats->M2 / (stats->n - 1);
 163         double variance_mean = variance / stats->n;
 164
 165         return sqrt(variance_mean);
 166 }
 167
 168 struct stats                    runtime_nsecs_stats[MAX_NR_CPUS];
 169 struct stats                    runtime_cycles_stats[MAX_NR_CPUS];
 170 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
 171 struct stats                    walltime_nsecs_stats;
 172
 173 #define MATCH_EVENT(t, c, evsel)                        \
 174         (evsel->attr.type == PERF_TYPE_##t &&   \
 175          evsel->attr.config == PERF_COUNT_##c)
 176
 177 #define ERR_PERF_OPEN \
 178 "counter %d, sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information."
 179
 180 static int create_perf_stat_counter(struct perf_evsel *evsel, bool *perm_err)
 181 {
 182         struct perf_event_attr *attr = &evsel->attr;
 183         int thread;
 184         int ncreated = 0;
 185
 186         if (scale)
 187                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 188                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 189
 190         if (system_wide) {
 191                 int cpu;
 192
 193                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 194                         FD(evsel, cpu, 0) = sys_perf_event_open(attr,
 195                                         -1, cpumap[cpu], -1, 0);
 196                         if (FD(evsel, cpu, 0) < 0) {
 197                                 if (errno == EPERM || errno == EACCES)
 198                                         *perm_err = true;
 199                                 error(ERR_PERF_OPEN, evsel->idx,
 200                                         FD(evsel, cpu, 0), strerror(errno));
 201                         } else {
 202                                 ++ncreated;
 203                         }
 204                 }
 205         } else {
 206                 attr->inherit = !no_inherit;
 207                 if (target_pid == -1 && target_tid == -1) {
 208                         attr->disabled = 1;
 209                         attr->enable_on_exec = 1;
 210                 }
 211                 for (thread = 0; thread < thread_num; thread++) {
 212                         FD(evsel, 0, thread) = sys_perf_event_open(attr,
 213                                 all_tids[thread], -1, -1, 0);
 214                         if (FD(evsel, 0, thread) < 0) {
 215                                 if (errno == EPERM || errno == EACCES)
 216                                         *perm_err = true;
 217                                 error(ERR_PERF_OPEN, evsel->idx,
 218                                         FD(evsel, 0, thread),
 219                                          strerror(errno));
 220                         } else {
 221                                 ++ncreated;
 222                         }
 223                 }
 224         }
 225
 226         return ncreated;
 227 }
 228
 229 /*
 230  * Does the counter have nsecs as a unit?
 231  */
 232 static inline int nsec_counter(struct perf_evsel *counter)
 233 {
 234         if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
 235             MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 236                 return 1;
 237
 238         return 0;
 239 }
 240
 241 /*
 242  * Read out the results of a single counter:
 243  * aggregate counts across CPUs in system-wide mode
 244  */
 245 static void read_counter_aggr(struct perf_evsel *counter)
 246 {
 247         struct perf_stat *ps = counter->priv;
 248         u64 count[3], single_count[3];
 249         int cpu;
 250         size_t res, nv;
 251         int scaled;
 252         int i, thread;
 253
 254         count[0] = count[1] = count[2] = 0;
 255
 256         nv = scale ? 3 : 1;
 257         for (cpu = 0; cpu < nr_cpus; cpu++) {
 258                 for (thread = 0; thread < thread_num; thread++) {
 259                         if (FD(counter, cpu, thread) < 0)
 260                                 continue;
 261
 262                         res = read(FD(counter, cpu, thread),
 263                                         single_count, nv * sizeof(u64));
 264                         assert(res == nv * sizeof(u64));
 265
 266                         close(FD(counter, cpu, thread));
 267                         FD(counter, cpu, thread) = -1;
 268
 269                         count[0] += single_count[0];
 270                         if (scale) {
 271                                 count[1] += single_count[1];
 272                                 count[2] += single_count[2];
 273                         }
 274                 }
 275         }
 276
 277         scaled = 0;
 278         if (scale) {
 279                 if (count[2] == 0) {
 280                         ps->scaled = -1;
 281                         count[0] = 0;
 282                         return;
 283                 }
 284
 285                 if (count[2] < count[1]) {
 286                         ps->scaled = 1;
 287                         count[0] = (unsigned long long)
 288                                 ((double)count[0] * count[1] / count[2] + 0.5);
 289                 }
 290         }
 291
 292         for (i = 0; i < 3; i++)
 293                 update_stats(&ps->res_stats[i], count[i]);
 294
 295         if (verbose) {
 296                 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
 297                                 count[0], count[1], count[2]);
 298         }
 299
 300         /*
 301          * Save the full runtime - to allow normalization during printout:
 302          */
 303         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 304                 update_stats(&runtime_nsecs_stats[0], count[0]);
 305         if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 306                 update_stats(&runtime_cycles_stats[0], count[0]);
 307         if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
 308                 update_stats(&runtime_branches_stats[0], count[0]);
 309 }
 310
 311 /*
 312  * Read out the results of a single counter:
 313  * do not aggregate counts across CPUs in system-wide mode
 314  */
 315 static void read_counter(struct perf_evsel *counter)
 316 {
 317         struct cpu_counts *cpu_counts = counter->priv;
 318         u64 count[3];
 319         int cpu;
 320         size_t res, nv;
 321
 322         count[0] = count[1] = count[2] = 0;
 323
 324         nv = scale ? 3 : 1;
 325
 326         for (cpu = 0; cpu < nr_cpus; cpu++) {
 327
 328                 if (FD(counter, cpu, 0) < 0)
 329                         continue;
 330
 331                 res = read(FD(counter, cpu, 0), count, nv * sizeof(u64));
 332
 333                 assert(res == nv * sizeof(u64));
 334
 335                 close(FD(counter, cpu, 0));
 336                 FD(counter, cpu, 0) = -1;
 337
 338                 if (scale) {
 339                         if (count[2] == 0) {
 340                                 count[0] = 0;
 341                         } else if (count[2] < count[1]) {
 342                                 count[0] = (unsigned long long)
 343                                 ((double)count[0] * count[1] / count[2] + 0.5);
 344                         }
 345                 }
 346                 cpu_counts[cpu].val = count[0]; /* scaled count */
 347                 cpu_counts[cpu].ena = count[1];
 348                 cpu_counts[cpu].run = count[2];
 349
 350                 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 351                         update_stats(&runtime_nsecs_stats[cpu], count[0]);
 352                 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 353                         update_stats(&runtime_cycles_stats[cpu], count[0]);
 354                 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
 355                         update_stats(&runtime_branches_stats[cpu], count[0]);
 356         }
 357 }
 358
 359 static int run_perf_stat(int argc __used, const char **argv)
 360 {
 361         unsigned long long t0, t1;
 362         struct perf_evsel *counter;
 363         int status = 0;
 364         int ncreated = 0;
 365         int child_ready_pipe[2], go_pipe[2];
 366         bool perm_err = false;
 367         const bool forks = (argc > 0);
 368         char buf;
 369
 370         if (!system_wide)
 371                 nr_cpus = 1;
 372
 373         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 374                 perror("failed to create pipes");
 375                 exit(1);
 376         }
 377
 378         if (forks) {
 379                 if ((child_pid = fork()) < 0)
 380                         perror("failed to fork");
 381
 382                 if (!child_pid) {
 383                         close(child_ready_pipe[0]);
 384                         close(go_pipe[1]);
 385                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 386
 387                         /*
 388                          * Do a dummy execvp to get the PLT entry resolved,
 389                          * so we avoid the resolver overhead on the real
 390                          * execvp call.
 391                          */
 392                         execvp("", (char **)argv);
 393
 394                         /*
 395                          * Tell the parent we're ready to go
 396                          */
 397                         close(child_ready_pipe[1]);
 398
 399                         /*
 400                          * Wait until the parent tells us to go.
 401                          */
 402                         if (read(go_pipe[0], &buf, 1) == -1)
 403                                 perror("unable to read pipe");
 404
 405                         execvp(argv[0], (char **)argv);
 406
 407                         perror(argv[0]);
 408                         exit(-1);
 409                 }
 410
 411                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 412                         all_tids[0] = child_pid;
 413
 414                 /*
 415                  * Wait for the child to be ready to exec.
 416                  */
 417                 close(child_ready_pipe[1]);
 418                 close(go_pipe[0]);
 419                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 420                         perror("unable to read pipe");
 421                 close(child_ready_pipe[0]);
 422         }
 423
 424         list_for_each_entry(counter, &evsel_list, node)
 425                 ncreated += create_perf_stat_counter(counter, &perm_err);
 426
 427         if (ncreated < nr_counters) {
 428                 if (perm_err)
 429                         error("You may not have permission to collect %sstats.\n"
 430                               "\t Consider tweaking"
 431                               " /proc/sys/kernel/perf_event_paranoid or running as root.",
 432                               system_wide ? "system-wide " : "");
 433                 die("Not all events could be opened.\n");
 434                 if (child_pid != -1)
 435                         kill(child_pid, SIGTERM);
 436                 return -1;
 437         }
 438
 439         /*
 440          * Enable counters and exec the command:
 441          */
 442         t0 = rdclock();
 443
 444         if (forks) {
 445                 close(go_pipe[1]);
 446                 wait(&status);
 447         } else {
 448                 while(!done) sleep(1);
 449         }
 450
 451         t1 = rdclock();
 452
 453         update_stats(&walltime_nsecs_stats, t1 - t0);
 454
 455         if (no_aggr) {
 456                 list_for_each_entry(counter, &evsel_list, node)
 457                         read_counter(counter);
 458         } else {
 459                 list_for_each_entry(counter, &evsel_list, node)
 460                         read_counter_aggr(counter);
 461         }
 462         return WEXITSTATUS(status);
 463 }
 464
 465 static void print_noise(struct perf_evsel *evsel, double avg)
 466 {
 467         struct perf_stat *ps;
 468
 469         if (run_count == 1)
 470                 return;
 471
 472         ps = evsel->priv;
 473         fprintf(stderr, "   ( +- %7.3f%% )",
 474                         100 * stddev_stats(&ps->res_stats[0]) / avg);
 475 }
 476
 477 static void nsec_printout(int cpu, struct perf_evsel *counter, double avg)
 478 {
 479         double msecs = avg / 1e6;
 480         char cpustr[16] = { '\0', };
 481         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 482
 483         if (no_aggr)
 484                 sprintf(cpustr, "CPU%*d%s",
 485                         csv_output ? 0 : -4,
 486                         cpumap[cpu], csv_sep);
 487
 488         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(counter));
 489
 490         if (csv_output)
 491                 return;
 492
 493         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
 494                 fprintf(stderr, " # %10.3f CPUs ",
 495                                 avg / avg_stats(&walltime_nsecs_stats));
 496         }
 497 }
 498
 499 static void abs_printout(int cpu, struct perf_evsel *counter, double avg)
 500 {
 501         double total, ratio = 0.0;
 502         char cpustr[16] = { '\0', };
 503         const char *fmt;
 504
 505         if (csv_output)
 506                 fmt = "%s%.0f%s%s";
 507         else if (big_num)
 508                 fmt = "%s%'18.0f%s%-24s";
 509         else
 510                 fmt = "%s%18.0f%s%-24s";
 511
 512         if (no_aggr)
 513                 sprintf(cpustr, "CPU%*d%s",
 514                         csv_output ? 0 : -4,
 515                         cpumap[cpu], csv_sep);
 516         else
 517                 cpu = 0;
 518
 519         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(counter));
 520
 521         if (csv_output)
 522                 return;
 523
 524         if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
 525                 total = avg_stats(&runtime_cycles_stats[cpu]);
 526
 527                 if (total)
 528                         ratio = avg / total;
 529
 530                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 531         } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
 532                         runtime_branches_stats[cpu].n != 0) {
 533                 total = avg_stats(&runtime_branches_stats[cpu]);
 534
 535                 if (total)
 536                         ratio = avg * 100 / total;
 537
 538                 fprintf(stderr, " # %10.3f %%    ", ratio);
 539
 540         } else if (runtime_nsecs_stats[cpu].n != 0) {
 541                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 542
 543                 if (total)
 544                         ratio = 1000.0 * avg / total;
 545
 546                 fprintf(stderr, " # %10.3f M/sec", ratio);
 547         }
 548 }
 549
 550 /*
 551  * Print out the results of a single counter:
 552  * aggregated counts in system-wide mode
 553  */
 554 static void print_counter_aggr(struct perf_evsel *counter)
 555 {
 556         struct perf_stat *ps = counter->priv;
 557         double avg = avg_stats(&ps->res_stats[0]);
 558         int scaled = ps->scaled;
 559
 560         if (scaled == -1) {
 561                 fprintf(stderr, "%*s%s%-24s\n",
 562                         csv_output ? 0 : 18,
 563                         "<not counted>", csv_sep, event_name(counter));
 564                 return;
 565         }
 566
 567         if (nsec_counter(counter))
 568                 nsec_printout(-1, counter, avg);
 569         else
 570                 abs_printout(-1, counter, avg);
 571
 572         if (csv_output) {
 573                 fputc('\n', stderr);
 574                 return;
 575         }
 576
 577         print_noise(counter, avg);
 578
 579         if (scaled) {
 580                 double avg_enabled, avg_running;
 581
 582                 avg_enabled = avg_stats(&ps->res_stats[1]);
 583                 avg_running = avg_stats(&ps->res_stats[2]);
 584
 585                 fprintf(stderr, "  (scaled from %.2f%%)",
 586                                 100 * avg_running / avg_enabled);
 587         }
 588
 589         fprintf(stderr, "\n");
 590 }
 591
 592 /*
 593  * Print out the results of a single counter:
 594  * does not use aggregated count in system-wide
 595  */
 596 static void print_counter(struct perf_evsel *counter)
 597 {
 598         struct perf_stat *ps = counter->priv;
 599         u64 ena, run, val;
 600         int cpu;
 601
 602         for (cpu = 0; cpu < nr_cpus; cpu++) {
 603                 val = ps->cpu_counts[cpu].val;
 604                 ena = ps->cpu_counts[cpu].ena;
 605                 run = ps->cpu_counts[cpu].run;
 606                 if (run == 0 || ena == 0) {
 607                         fprintf(stderr, "CPU%*d%s%*s%s%-24s",
 608                                 csv_output ? 0 : -4,
 609                                 cpumap[cpu], csv_sep,
 610                                 csv_output ? 0 : 18,
 611                                 "<not counted>", csv_sep,
 612                                 event_name(counter));
 613
 614                         fprintf(stderr, "\n");
 615                         continue;
 616                 }
 617
 618                 if (nsec_counter(counter))
 619                         nsec_printout(cpu, counter, val);
 620                 else
 621                         abs_printout(cpu, counter, val);
 622
 623                 if (!csv_output) {
 624                         print_noise(counter, 1.0);
 625
 626                         if (run != ena) {
 627                                 fprintf(stderr, "  (scaled from %.2f%%)",
 628                                         100.0 * run / ena);
 629                         }
 630                 }
 631                 fprintf(stderr, "\n");
 632         }
 633 }
 634
 635 static void print_stat(int argc, const char **argv)
 636 {
 637         struct perf_evsel *counter;
 638         int i;
 639
 640         fflush(stdout);
 641
 642         if (!csv_output) {
 643                 fprintf(stderr, "\n");
 644                 fprintf(stderr, " Performance counter stats for ");
 645                 if(target_pid == -1 && target_tid == -1) {
 646                         fprintf(stderr, "\'%s", argv[0]);
 647                         for (i = 1; i < argc; i++)
 648                                 fprintf(stderr, " %s", argv[i]);
 649                 } else if (target_pid != -1)
 650                         fprintf(stderr, "process id \'%d", target_pid);
 651                 else
 652                         fprintf(stderr, "thread id \'%d", target_tid);
 653
 654                 fprintf(stderr, "\'");
 655                 if (run_count > 1)
 656                         fprintf(stderr, " (%d runs)", run_count);
 657                 fprintf(stderr, ":\n\n");
 658         }
 659
 660         if (no_aggr) {
 661                 list_for_each_entry(counter, &evsel_list, node)
 662                         print_counter(counter);
 663         } else {
 664                 list_for_each_entry(counter, &evsel_list, node)
 665                         print_counter_aggr(counter);
 666         }
 667
 668         if (!csv_output) {
 669                 fprintf(stderr, "\n");
 670                 fprintf(stderr, " %18.9f  seconds time elapsed",
 671                                 avg_stats(&walltime_nsecs_stats)/1e9);
 672                 if (run_count > 1) {
 673                         fprintf(stderr, "   ( +- %7.3f%% )",
 674                                 100*stddev_stats(&walltime_nsecs_stats) /
 675                                 avg_stats(&walltime_nsecs_stats));
 676                 }
 677                 fprintf(stderr, "\n\n");
 678         }
 679 }
 680
 681 static volatile int signr = -1;
 682
 683 static void skip_signal(int signo)
 684 {
 685         if(child_pid == -1)
 686                 done = 1;
 687
 688         signr = signo;
 689 }
 690
 691 static void sig_atexit(void)
 692 {
 693         if (child_pid != -1)
 694                 kill(child_pid, SIGTERM);
 695
 696         if (signr == -1)
 697                 return;
 698
 699         signal(signr, SIG_DFL);
 700         kill(getpid(), signr);
 701 }
 702
 703 static const char * const stat_usage[] = {
 704         "perf stat [<options>] [<command>]",
 705         NULL
 706 };
 707
 708 static int stat__set_big_num(const struct option *opt __used,
 709                              const char *s __used, int unset)
 710 {
 711         big_num_opt = unset ? 0 : 1;
 712         return 0;
 713 }
 714
 715 static const struct option options[] = {
 716         OPT_CALLBACK('e', "event", NULL, "event",
 717                      "event selector. use 'perf list' to list available events",
 718                      parse_events),
 719         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 720                     "child tasks do not inherit counters"),
 721         OPT_INTEGER('p', "pid", &target_pid,
 722                     "stat events on existing process id"),
 723         OPT_INTEGER('t', "tid", &target_tid,
 724                     "stat events on existing thread id"),
 725         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 726                     "system-wide collection from all CPUs"),
 727         OPT_BOOLEAN('c', "scale", &scale,
 728                     "scale/normalize counters"),
 729         OPT_INCR('v', "verbose", &verbose,
 730                     "be more verbose (show counter open errors, etc)"),
 731         OPT_INTEGER('r', "repeat", &run_count,
 732                     "repeat command and print average + stddev (max: 100)"),
 733         OPT_BOOLEAN('n', "null", &null_run,
 734                     "null run - dont start any counters"),
 735         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 736                            "print large numbers with thousands\' separators",
 737                            stat__set_big_num),
 738         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 739                     "list of cpus to monitor in system-wide"),
 740         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 741                     "disable CPU count aggregation"),
 742         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 743                    "print counts with custom separator"),
 744         OPT_END()
 745 };
 746
 747 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 748 {
 749         struct perf_evsel *pos;
 750         int status = -ENOMEM;
 751
 752         setlocale(LC_ALL, "");
 753
 754         argc = parse_options(argc, argv, options, stat_usage,
 755                 PARSE_OPT_STOP_AT_NON_OPTION);
 756
 757         if (csv_sep)
 758                 csv_output = true;
 759         else
 760                 csv_sep = DEFAULT_SEPARATOR;
 761
 762         /*
 763          * let the spreadsheet do the pretty-printing
 764          */
 765         if (csv_output) {
 766                 /* User explicitely passed -B? */
 767                 if (big_num_opt == 1) {
 768                         fprintf(stderr, "-B option not supported with -x\n");
 769                         usage_with_options(stat_usage, options);
 770                 } else /* Nope, so disable big number formatting */
 771                         big_num = false;
 772         } else if (big_num_opt == 0) /* User passed --no-big-num */
 773                 big_num = false;
 774
 775         if (!argc && target_pid == -1 && target_tid == -1)
 776                 usage_with_options(stat_usage, options);
 777         if (run_count <= 0)
 778                 usage_with_options(stat_usage, options);
 779
 780         /* no_aggr is for system-wide only */
 781         if (no_aggr && !system_wide)
 782                 usage_with_options(stat_usage, options);
 783
 784         /* Set attrs and nr_counters if no event is selected and !null_run */
 785         if (!null_run && !nr_counters) {
 786                 size_t c;
 787
 788                 nr_counters = ARRAY_SIZE(default_attrs);
 789
 790                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 791                         pos = perf_evsel__new(default_attrs[c].type,
 792                                               default_attrs[c].config,
 793                                               nr_counters);
 794                         if (pos == NULL)
 795                                 goto out;
 796                         list_add(&pos->node, &evsel_list);
 797                 }
 798         }
 799
 800         if (system_wide)
 801                 nr_cpus = read_cpu_map(cpu_list);
 802         else
 803                 nr_cpus = 1;
 804
 805         if (nr_cpus < 1)
 806                 usage_with_options(stat_usage, options);
 807
 808         if (target_pid != -1) {
 809                 target_tid = target_pid;
 810                 thread_num = find_all_tid(target_pid, &all_tids);
 811                 if (thread_num <= 0) {
 812                         fprintf(stderr, "Can't find all threads of pid %d\n",
 813                                         target_pid);
 814                         usage_with_options(stat_usage, options);
 815                 }
 816         } else {
 817                 all_tids=malloc(sizeof(pid_t));
 818                 if (!all_tids)
 819                         return -ENOMEM;
 820
 821                 all_tids[0] = target_tid;
 822                 thread_num = 1;
 823         }
 824
 825         list_for_each_entry(pos, &evsel_list, node) {
 826                 if (perf_evsel__alloc_stat_priv(pos, nr_cpus) < 0 ||
 827                     perf_evsel__alloc_fd(pos, nr_cpus, thread_num) < 0)
 828                         goto out_free_fd;
 829         }
 830
 831         /*
 832          * We dont want to block the signals - that would cause
 833          * child tasks to inherit that and Ctrl-C would not work.
 834          * What we want is for Ctrl-C to work in the exec()-ed
 835          * task, but being ignored by perf stat itself:
 836          */
 837         atexit(sig_atexit);
 838         signal(SIGINT,  skip_signal);
 839         signal(SIGALRM, skip_signal);
 840         signal(SIGABRT, skip_signal);
 841
 842         status = 0;
 843         for (run_idx = 0; run_idx < run_count; run_idx++) {
 844                 if (run_count != 1 && verbose)
 845                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 846                 status = run_perf_stat(argc, argv);
 847         }
 848
 849         if (status != -1)
 850                 print_stat(argc, argv);
 851 out_free_fd:
 852         list_for_each_entry(pos, &evsel_list, node) {
 853                 perf_evsel__free_fd(pos);
 854                 perf_evsel__free_stat_priv(pos);
 855         }
 856 out:
 857         return status;
 858 }