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/debug.h"
  47 #include "util/header.h"
  48 #include "util/cpumap.h"
  49 #include "util/thread.h"
  50
  51 #include <sys/prctl.h>
  52 #include <math.h>
  53 #include <locale.h>
  54
  55 static struct perf_event_attr default_attrs[] = {
  56
  57   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  58   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  59   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  60   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  61
  62   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  63   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  64   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  65   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  66   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES        },
  67   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES            },
  68
  69 };
  70
  71 static bool                     system_wide                     =  false;
  72 static unsigned int             nr_cpus                         =  0;
  73 static int                      run_idx                         =  0;
  74
  75 static int                      run_count                       =  1;
  76 static bool                     no_inherit                      = false;
  77 static bool                     scale                           =  true;
  78 static pid_t                    target_pid                      = -1;
  79 static pid_t                    target_tid                      = -1;
  80 static pid_t                    *all_tids                       =  NULL;
  81 static int                      thread_num                      =  0;
  82 static pid_t                    child_pid                       = -1;
  83 static bool                     null_run                        =  false;
  84 static bool                     big_num                         =  false;
  85
  86
  87 static int                      *fd[MAX_NR_CPUS][MAX_COUNTERS];
  88
  89 static int                      event_scaled[MAX_COUNTERS];
  90
  91 static volatile int done = 0;
  92
  93 struct stats
  94 {
  95         double n, mean, M2;
  96 };
  97
  98 static void update_stats(struct stats *stats, u64 val)
  99 {
 100         double delta;
 101
 102         stats->n++;
 103         delta = val - stats->mean;
 104         stats->mean += delta / stats->n;
 105         stats->M2 += delta*(val - stats->mean);
 106 }
 107
 108 static double avg_stats(struct stats *stats)
 109 {
 110         return stats->mean;
 111 }
 112
 113 /*
 114  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 115  *
 116  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 117  * s^2 = -------------------------------
 118  *                  n - 1
 119  *
 120  * http://en.wikipedia.org/wiki/Stddev
 121  *
 122  * The std dev of the mean is related to the std dev by:
 123  *
 124  *             s
 125  * s_mean = -------
 126  *          sqrt(n)
 127  *
 128  */
 129 static double stddev_stats(struct stats *stats)
 130 {
 131         double variance = stats->M2 / (stats->n - 1);
 132         double variance_mean = variance / stats->n;
 133
 134         return sqrt(variance_mean);
 135 }
 136
 137 struct stats                    event_res_stats[MAX_COUNTERS][3];
 138 struct stats                    runtime_nsecs_stats;
 139 struct stats                    walltime_nsecs_stats;
 140 struct stats                    runtime_cycles_stats;
 141 struct stats                    runtime_branches_stats;
 142
 143 #define MATCH_EVENT(t, c, counter)                      \
 144         (attrs[counter].type == PERF_TYPE_##t &&        \
 145          attrs[counter].config == PERF_COUNT_##c)
 146
 147 #define ERR_PERF_OPEN \
 148 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
 149
 150 static int create_perf_stat_counter(int counter)
 151 {
 152         struct perf_event_attr *attr = attrs + counter;
 153         int thread;
 154         int ncreated = 0;
 155
 156         if (scale)
 157                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 158                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 159
 160         if (system_wide) {
 161                 unsigned int cpu;
 162
 163                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 164                         fd[cpu][counter][0] = sys_perf_event_open(attr,
 165                                         -1, cpumap[cpu], -1, 0);
 166                         if (fd[cpu][counter][0] < 0)
 167                                 pr_debug(ERR_PERF_OPEN, counter,
 168                                          fd[cpu][counter][0], strerror(errno));
 169                         else
 170                                 ++ncreated;
 171                 }
 172         } else {
 173                 attr->inherit = !no_inherit;
 174                 if (target_pid == -1 && target_tid == -1) {
 175                         attr->disabled = 1;
 176                         attr->enable_on_exec = 1;
 177                 }
 178                 for (thread = 0; thread < thread_num; thread++) {
 179                         fd[0][counter][thread] = sys_perf_event_open(attr,
 180                                 all_tids[thread], -1, -1, 0);
 181                         if (fd[0][counter][thread] < 0)
 182                                 pr_debug(ERR_PERF_OPEN, counter,
 183                                          fd[0][counter][thread],
 184                                          strerror(errno));
 185                         else
 186                                 ++ncreated;
 187                 }
 188         }
 189
 190         return ncreated;
 191 }
 192
 193 /*
 194  * Does the counter have nsecs as a unit?
 195  */
 196 static inline int nsec_counter(int counter)
 197 {
 198         if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
 199             MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 200                 return 1;
 201
 202         return 0;
 203 }
 204
 205 /*
 206  * Read out the results of a single counter:
 207  */
 208 static void read_counter(int counter)
 209 {
 210         u64 count[3], single_count[3];
 211         unsigned int cpu;
 212         size_t res, nv;
 213         int scaled;
 214         int i, thread;
 215
 216         count[0] = count[1] = count[2] = 0;
 217
 218         nv = scale ? 3 : 1;
 219         for (cpu = 0; cpu < nr_cpus; cpu++) {
 220                 for (thread = 0; thread < thread_num; thread++) {
 221                         if (fd[cpu][counter][thread] < 0)
 222                                 continue;
 223
 224                         res = read(fd[cpu][counter][thread],
 225                                         single_count, nv * sizeof(u64));
 226                         assert(res == nv * sizeof(u64));
 227
 228                         close(fd[cpu][counter][thread]);
 229                         fd[cpu][counter][thread] = -1;
 230
 231                         count[0] += single_count[0];
 232                         if (scale) {
 233                                 count[1] += single_count[1];
 234                                 count[2] += single_count[2];
 235                         }
 236                 }
 237         }
 238
 239         scaled = 0;
 240         if (scale) {
 241                 if (count[2] == 0) {
 242                         event_scaled[counter] = -1;
 243                         count[0] = 0;
 244                         return;
 245                 }
 246
 247                 if (count[2] < count[1]) {
 248                         event_scaled[counter] = 1;
 249                         count[0] = (unsigned long long)
 250                                 ((double)count[0] * count[1] / count[2] + 0.5);
 251                 }
 252         }
 253
 254         for (i = 0; i < 3; i++)
 255                 update_stats(&event_res_stats[counter][i], count[i]);
 256
 257         if (verbose) {
 258                 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
 259                                 count[0], count[1], count[2]);
 260         }
 261
 262         /*
 263          * Save the full runtime - to allow normalization during printout:
 264          */
 265         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 266                 update_stats(&runtime_nsecs_stats, count[0]);
 267         if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 268                 update_stats(&runtime_cycles_stats, count[0]);
 269         if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
 270                 update_stats(&runtime_branches_stats, count[0]);
 271 }
 272
 273 static int run_perf_stat(int argc __used, const char **argv)
 274 {
 275         unsigned long long t0, t1;
 276         int status = 0;
 277         int counter, ncreated = 0;
 278         int child_ready_pipe[2], go_pipe[2];
 279         const bool forks = (argc > 0);
 280         char buf;
 281
 282         if (!system_wide)
 283                 nr_cpus = 1;
 284
 285         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 286                 perror("failed to create pipes");
 287                 exit(1);
 288         }
 289
 290         if (forks) {
 291                 if ((child_pid = fork()) < 0)
 292                         perror("failed to fork");
 293
 294                 if (!child_pid) {
 295                         close(child_ready_pipe[0]);
 296                         close(go_pipe[1]);
 297                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 298
 299                         /*
 300                          * Do a dummy execvp to get the PLT entry resolved,
 301                          * so we avoid the resolver overhead on the real
 302                          * execvp call.
 303                          */
 304                         execvp("", (char **)argv);
 305
 306                         /*
 307                          * Tell the parent we're ready to go
 308                          */
 309                         close(child_ready_pipe[1]);
 310
 311                         /*
 312                          * Wait until the parent tells us to go.
 313                          */
 314                         if (read(go_pipe[0], &buf, 1) == -1)
 315                                 perror("unable to read pipe");
 316
 317                         execvp(argv[0], (char **)argv);
 318
 319                         perror(argv[0]);
 320                         exit(-1);
 321                 }
 322
 323                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 324                         all_tids[0] = child_pid;
 325
 326                 /*
 327                  * Wait for the child to be ready to exec.
 328                  */
 329                 close(child_ready_pipe[1]);
 330                 close(go_pipe[0]);
 331                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 332                         perror("unable to read pipe");
 333                 close(child_ready_pipe[0]);
 334         }
 335
 336         for (counter = 0; counter < nr_counters; counter++)
 337                 ncreated += create_perf_stat_counter(counter);
 338
 339         if (ncreated == 0) {
 340                 pr_err("No permission to collect %sstats.\n"
 341                        "Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n",
 342                        system_wide ? "system-wide " : "");
 343                 if (child_pid != -1)
 344                         kill(child_pid, SIGTERM);
 345                 return -1;
 346         }
 347
 348         /*
 349          * Enable counters and exec the command:
 350          */
 351         t0 = rdclock();
 352
 353         if (forks) {
 354                 close(go_pipe[1]);
 355                 wait(&status);
 356         } else {
 357                 while(!done) sleep(1);
 358         }
 359
 360         t1 = rdclock();
 361
 362         update_stats(&walltime_nsecs_stats, t1 - t0);
 363
 364         for (counter = 0; counter < nr_counters; counter++)
 365                 read_counter(counter);
 366
 367         return WEXITSTATUS(status);
 368 }
 369
 370 static void print_noise(int counter, double avg)
 371 {
 372         if (run_count == 1)
 373                 return;
 374
 375         fprintf(stderr, "   ( +- %7.3f%% )",
 376                         100 * stddev_stats(&event_res_stats[counter][0]) / avg);
 377 }
 378
 379 static void nsec_printout(int counter, double avg)
 380 {
 381         double msecs = avg / 1e6;
 382
 383         fprintf(stderr, " %18.6f  %-24s", msecs, event_name(counter));
 384
 385         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
 386                 fprintf(stderr, " # %10.3f CPUs ",
 387                                 avg / avg_stats(&walltime_nsecs_stats));
 388         }
 389 }
 390
 391 static void abs_printout(int counter, double avg)
 392 {
 393         double total, ratio = 0.0;
 394
 395         if (big_num)
 396                 fprintf(stderr, " %'18.0f  %-24s", avg, event_name(counter));
 397         else
 398                 fprintf(stderr, " %18.0f  %-24s", avg, event_name(counter));
 399
 400         if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
 401                 total = avg_stats(&runtime_cycles_stats);
 402
 403                 if (total)
 404                         ratio = avg / total;
 405
 406                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 407         } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
 408                         runtime_branches_stats.n != 0) {
 409                 total = avg_stats(&runtime_branches_stats);
 410
 411                 if (total)
 412                         ratio = avg * 100 / total;
 413
 414                 fprintf(stderr, " # %10.3f %%    ", ratio);
 415
 416         } else if (runtime_nsecs_stats.n != 0) {
 417                 total = avg_stats(&runtime_nsecs_stats);
 418
 419                 if (total)
 420                         ratio = 1000.0 * avg / total;
 421
 422                 fprintf(stderr, " # %10.3f M/sec", ratio);
 423         }
 424 }
 425
 426 /*
 427  * Print out the results of a single counter:
 428  */
 429 static void print_counter(int counter)
 430 {
 431         double avg = avg_stats(&event_res_stats[counter][0]);
 432         int scaled = event_scaled[counter];
 433
 434         if (scaled == -1) {
 435                 fprintf(stderr, " %18s  %-24s\n",
 436                         "<not counted>", event_name(counter));
 437                 return;
 438         }
 439
 440         if (nsec_counter(counter))
 441                 nsec_printout(counter, avg);
 442         else
 443                 abs_printout(counter, avg);
 444
 445         print_noise(counter, avg);
 446
 447         if (scaled) {
 448                 double avg_enabled, avg_running;
 449
 450                 avg_enabled = avg_stats(&event_res_stats[counter][1]);
 451                 avg_running = avg_stats(&event_res_stats[counter][2]);
 452
 453                 fprintf(stderr, "  (scaled from %.2f%%)",
 454                                 100 * avg_running / avg_enabled);
 455         }
 456
 457         fprintf(stderr, "\n");
 458 }
 459
 460 static void print_stat(int argc, const char **argv)
 461 {
 462         int i, counter;
 463
 464         fflush(stdout);
 465
 466         fprintf(stderr, "\n");
 467         fprintf(stderr, " Performance counter stats for ");
 468         if(target_pid == -1 && target_tid == -1) {
 469                 fprintf(stderr, "\'%s", argv[0]);
 470                 for (i = 1; i < argc; i++)
 471                         fprintf(stderr, " %s", argv[i]);
 472         } else if (target_pid != -1)
 473                 fprintf(stderr, "process id \'%d", target_pid);
 474         else
 475                 fprintf(stderr, "thread id \'%d", target_tid);
 476
 477         fprintf(stderr, "\'");
 478         if (run_count > 1)
 479                 fprintf(stderr, " (%d runs)", run_count);
 480         fprintf(stderr, ":\n\n");
 481
 482         for (counter = 0; counter < nr_counters; counter++)
 483                 print_counter(counter);
 484
 485         fprintf(stderr, "\n");
 486         fprintf(stderr, " %18.9f  seconds time elapsed",
 487                         avg_stats(&walltime_nsecs_stats)/1e9);
 488         if (run_count > 1) {
 489                 fprintf(stderr, "   ( +- %7.3f%% )",
 490                                 100*stddev_stats(&walltime_nsecs_stats) /
 491                                 avg_stats(&walltime_nsecs_stats));
 492         }
 493         fprintf(stderr, "\n\n");
 494 }
 495
 496 static volatile int signr = -1;
 497
 498 static void skip_signal(int signo)
 499 {
 500         if(child_pid == -1)
 501                 done = 1;
 502
 503         signr = signo;
 504 }
 505
 506 static void sig_atexit(void)
 507 {
 508         if (child_pid != -1)
 509                 kill(child_pid, SIGTERM);
 510
 511         if (signr == -1)
 512                 return;
 513
 514         signal(signr, SIG_DFL);
 515         kill(getpid(), signr);
 516 }
 517
 518 static const char * const stat_usage[] = {
 519         "perf stat [<options>] [<command>]",
 520         NULL
 521 };
 522
 523 static const struct option options[] = {
 524         OPT_CALLBACK('e', "event", NULL, "event",
 525                      "event selector. use 'perf list' to list available events",
 526                      parse_events),
 527         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 528                     "child tasks do not inherit counters"),
 529         OPT_INTEGER('p', "pid", &target_pid,
 530                     "stat events on existing process id"),
 531         OPT_INTEGER('t', "tid", &target_tid,
 532                     "stat events on existing thread id"),
 533         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 534                     "system-wide collection from all CPUs"),
 535         OPT_BOOLEAN('c', "scale", &scale,
 536                     "scale/normalize counters"),
 537         OPT_INCR('v', "verbose", &verbose,
 538                     "be more verbose (show counter open errors, etc)"),
 539         OPT_INTEGER('r', "repeat", &run_count,
 540                     "repeat command and print average + stddev (max: 100)"),
 541         OPT_BOOLEAN('n', "null", &null_run,
 542                     "null run - dont start any counters"),
 543         OPT_BOOLEAN('B', "big-num", &big_num,
 544                     "print large numbers with thousands\' separators"),
 545         OPT_END()
 546 };
 547
 548 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 549 {
 550         int status;
 551         int i,j;
 552
 553         setlocale(LC_ALL, "");
 554
 555         argc = parse_options(argc, argv, options, stat_usage,
 556                 PARSE_OPT_STOP_AT_NON_OPTION);
 557         if (!argc && target_pid == -1 && target_tid == -1)
 558                 usage_with_options(stat_usage, options);
 559         if (run_count <= 0)
 560                 usage_with_options(stat_usage, options);
 561
 562         /* Set attrs and nr_counters if no event is selected and !null_run */
 563         if (!null_run && !nr_counters) {
 564                 memcpy(attrs, default_attrs, sizeof(default_attrs));
 565                 nr_counters = ARRAY_SIZE(default_attrs);
 566         }
 567
 568         if (system_wide)
 569                 nr_cpus = read_cpu_map();
 570         else
 571                 nr_cpus = 1;
 572
 573         if (target_pid != -1) {
 574                 target_tid = target_pid;
 575                 thread_num = find_all_tid(target_pid, &all_tids);
 576                 if (thread_num <= 0) {
 577                         fprintf(stderr, "Can't find all threads of pid %d\n",
 578                                         target_pid);
 579                         usage_with_options(stat_usage, options);
 580                 }
 581         } else {
 582                 all_tids=malloc(sizeof(pid_t));
 583                 if (!all_tids)
 584                         return -ENOMEM;
 585
 586                 all_tids[0] = target_tid;
 587                 thread_num = 1;
 588         }
 589
 590         for (i = 0; i < MAX_NR_CPUS; i++) {
 591                 for (j = 0; j < MAX_COUNTERS; j++) {
 592                         fd[i][j] = malloc(sizeof(int)*thread_num);
 593                         if (!fd[i][j])
 594                                 return -ENOMEM;
 595                 }
 596         }
 597
 598         /*
 599          * We dont want to block the signals - that would cause
 600          * child tasks to inherit that and Ctrl-C would not work.
 601          * What we want is for Ctrl-C to work in the exec()-ed
 602          * task, but being ignored by perf stat itself:
 603          */
 604         atexit(sig_atexit);
 605         signal(SIGINT,  skip_signal);
 606         signal(SIGALRM, skip_signal);
 607         signal(SIGABRT, skip_signal);
 608
 609         status = 0;
 610         for (run_idx = 0; run_idx < run_count; run_idx++) {
 611                 if (run_count != 1 && verbose)
 612                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 613                 status = run_perf_stat(argc, argv);
 614         }
 615
 616         if (status != -1)
 617                 print_stat(argc, argv);
 618
 619         return status;
 620 }