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
  48 #include <sys/prctl.h>
  49 #include <math.h>
  50
  51 static struct perf_event_attr default_attrs[] = {
  52
  53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
  54   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
  55   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS  },
  56   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
  57
  58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
  59   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS    },
  60   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
  61   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES    },
  62   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS},
  63   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES   },
  64
  65 };
  66
  67 static int                      system_wide                     =  0;
  68 static unsigned int             nr_cpus                         =  0;
  69 static int                      run_idx                         =  0;
  70
  71 static int                      run_count                       =  1;
  72 static int                      inherit                         =  1;
  73 static int                      scale                           =  1;
  74 static pid_t                    target_pid                      = -1;
  75 static pid_t                    child_pid                       = -1;
  76 static int                      null_run                        =  0;
  77
  78 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  79
  80 static int                      event_scaled[MAX_COUNTERS];
  81
  82 struct stats
  83 {
  84         double n, mean, M2;
  85 };
  86
  87 static void update_stats(struct stats *stats, u64 val)
  88 {
  89         double delta;
  90
  91         stats->n++;
  92         delta = val - stats->mean;
  93         stats->mean += delta / stats->n;
  94         stats->M2 += delta*(val - stats->mean);
  95 }
  96
  97 static double avg_stats(struct stats *stats)
  98 {
  99         return stats->mean;
 100 }
 101
 102 /*
 103  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 104  *
 105  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 106  * s^2 = -------------------------------
 107  *                  n - 1
 108  *
 109  * http://en.wikipedia.org/wiki/Stddev
 110  *
 111  * The std dev of the mean is related to the std dev by:
 112  *
 113  *             s
 114  * s_mean = -------
 115  *          sqrt(n)
 116  *
 117  */
 118 static double stddev_stats(struct stats *stats)
 119 {
 120         double variance = stats->M2 / (stats->n - 1);
 121         double variance_mean = variance / stats->n;
 122
 123         return sqrt(variance_mean);
 124 }
 125
 126 struct stats                    event_res_stats[MAX_COUNTERS][3];
 127 struct stats                    runtime_nsecs_stats;
 128 struct stats                    walltime_nsecs_stats;
 129 struct stats                    runtime_cycles_stats;
 130 struct stats                    runtime_branches_stats;
 131
 132 #define MATCH_EVENT(t, c, counter)                      \
 133         (attrs[counter].type == PERF_TYPE_##t &&        \
 134          attrs[counter].config == PERF_COUNT_##c)
 135
 136 #define ERR_PERF_OPEN \
 137 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
 138
 139 static void create_perf_stat_counter(int counter, int pid)
 140 {
 141         struct perf_event_attr *attr = attrs + counter;
 142
 143         if (scale)
 144                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 145                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 146
 147         if (system_wide) {
 148                 unsigned int cpu;
 149
 150                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 151                         fd[cpu][counter] = sys_perf_event_open(attr, -1, cpu, -1, 0);
 152                         if (fd[cpu][counter] < 0 && verbose)
 153                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 154                                         fd[cpu][counter], strerror(errno));
 155                 }
 156         } else {
 157                 attr->inherit        = inherit;
 158                 attr->disabled       = 1;
 159                 attr->enable_on_exec = 1;
 160
 161                 fd[0][counter] = sys_perf_event_open(attr, pid, -1, -1, 0);
 162                 if (fd[0][counter] < 0 && verbose)
 163                         fprintf(stderr, ERR_PERF_OPEN, counter,
 164                                 fd[0][counter], strerror(errno));
 165         }
 166 }
 167
 168 /*
 169  * Does the counter have nsecs as a unit?
 170  */
 171 static inline int nsec_counter(int counter)
 172 {
 173         if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
 174             MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 175                 return 1;
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * Read out the results of a single counter:
 182  */
 183 static void read_counter(int counter)
 184 {
 185         u64 count[3], single_count[3];
 186         unsigned int cpu;
 187         size_t res, nv;
 188         int scaled;
 189         int i;
 190
 191         count[0] = count[1] = count[2] = 0;
 192
 193         nv = scale ? 3 : 1;
 194         for (cpu = 0; cpu < nr_cpus; cpu++) {
 195                 if (fd[cpu][counter] < 0)
 196                         continue;
 197
 198                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 199                 assert(res == nv * sizeof(u64));
 200
 201                 close(fd[cpu][counter]);
 202                 fd[cpu][counter] = -1;
 203
 204                 count[0] += single_count[0];
 205                 if (scale) {
 206                         count[1] += single_count[1];
 207                         count[2] += single_count[2];
 208                 }
 209         }
 210
 211         scaled = 0;
 212         if (scale) {
 213                 if (count[2] == 0) {
 214                         event_scaled[counter] = -1;
 215                         count[0] = 0;
 216                         return;
 217                 }
 218
 219                 if (count[2] < count[1]) {
 220                         event_scaled[counter] = 1;
 221                         count[0] = (unsigned long long)
 222                                 ((double)count[0] * count[1] / count[2] + 0.5);
 223                 }
 224         }
 225
 226         for (i = 0; i < 3; i++)
 227                 update_stats(&event_res_stats[counter][i], count[i]);
 228
 229         if (verbose) {
 230                 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
 231                                 count[0], count[1], count[2]);
 232         }
 233
 234         /*
 235          * Save the full runtime - to allow normalization during printout:
 236          */
 237         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 238                 update_stats(&runtime_nsecs_stats, count[0]);
 239         if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 240                 update_stats(&runtime_cycles_stats, count[0]);
 241         if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
 242                 update_stats(&runtime_branches_stats, count[0]);
 243 }
 244
 245 static int run_perf_stat(int argc __used, const char **argv)
 246 {
 247         unsigned long long t0, t1;
 248         int status = 0;
 249         int counter;
 250         int pid;
 251         int child_ready_pipe[2], go_pipe[2];
 252         char buf;
 253
 254         if (!system_wide)
 255                 nr_cpus = 1;
 256
 257         if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
 258                 perror("failed to create pipes");
 259                 exit(1);
 260         }
 261
 262         if ((pid = fork()) < 0)
 263                 perror("failed to fork");
 264
 265         if (!pid) {
 266                 close(child_ready_pipe[0]);
 267                 close(go_pipe[1]);
 268                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 269
 270                 /*
 271                  * Do a dummy execvp to get the PLT entry resolved,
 272                  * so we avoid the resolver overhead on the real
 273                  * execvp call.
 274                  */
 275                 execvp("", (char **)argv);
 276
 277                 /*
 278                  * Tell the parent we're ready to go
 279                  */
 280                 close(child_ready_pipe[1]);
 281
 282                 /*
 283                  * Wait until the parent tells us to go.
 284                  */
 285                 if (read(go_pipe[0], &buf, 1) == -1)
 286                         perror("unable to read pipe");
 287
 288                 execvp(argv[0], (char **)argv);
 289
 290                 perror(argv[0]);
 291                 exit(-1);
 292         }
 293
 294         child_pid = pid;
 295
 296         /*
 297          * Wait for the child to be ready to exec.
 298          */
 299         close(child_ready_pipe[1]);
 300         close(go_pipe[0]);
 301         if (read(child_ready_pipe[0], &buf, 1) == -1)
 302                 perror("unable to read pipe");
 303         close(child_ready_pipe[0]);
 304
 305         for (counter = 0; counter < nr_counters; counter++)
 306                 create_perf_stat_counter(counter, pid);
 307
 308         /*
 309          * Enable counters and exec the command:
 310          */
 311         t0 = rdclock();
 312
 313         close(go_pipe[1]);
 314         wait(&status);
 315
 316         t1 = rdclock();
 317
 318         update_stats(&walltime_nsecs_stats, t1 - t0);
 319
 320         for (counter = 0; counter < nr_counters; counter++)
 321                 read_counter(counter);
 322
 323         return WEXITSTATUS(status);
 324 }
 325
 326 static void print_noise(int counter, double avg)
 327 {
 328         if (run_count == 1)
 329                 return;
 330
 331         fprintf(stderr, "   ( +- %7.3f%% )",
 332                         100 * stddev_stats(&event_res_stats[counter][0]) / avg);
 333 }
 334
 335 static void nsec_printout(int counter, double avg)
 336 {
 337         double msecs = avg / 1e6;
 338
 339         fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
 340
 341         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
 342                 fprintf(stderr, " # %10.3f CPUs ",
 343                                 avg / avg_stats(&walltime_nsecs_stats));
 344         }
 345 }
 346
 347 static void abs_printout(int counter, double avg)
 348 {
 349         double total, ratio = 0.0;
 350
 351         fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));
 352
 353         if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
 354                 total = avg_stats(&runtime_cycles_stats);
 355
 356                 if (total)
 357                         ratio = avg / total;
 358
 359                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 360         } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter)) {
 361                 total = avg_stats(&runtime_branches_stats);
 362
 363                 if (total)
 364                         ratio = avg * 100 / total;
 365
 366                 fprintf(stderr, " # %10.3f %%  ", ratio);
 367
 368         } else {
 369                 total = avg_stats(&runtime_nsecs_stats);
 370
 371                 if (total)
 372                         ratio = 1000.0 * avg / total;
 373
 374                 fprintf(stderr, " # %10.3f M/sec", ratio);
 375         }
 376 }
 377
 378 /*
 379  * Print out the results of a single counter:
 380  */
 381 static void print_counter(int counter)
 382 {
 383         double avg = avg_stats(&event_res_stats[counter][0]);
 384         int scaled = event_scaled[counter];
 385
 386         if (scaled == -1) {
 387                 fprintf(stderr, " %14s  %-24s\n",
 388                         "<not counted>", event_name(counter));
 389                 return;
 390         }
 391
 392         if (nsec_counter(counter))
 393                 nsec_printout(counter, avg);
 394         else
 395                 abs_printout(counter, avg);
 396
 397         print_noise(counter, avg);
 398
 399         if (scaled) {
 400                 double avg_enabled, avg_running;
 401
 402                 avg_enabled = avg_stats(&event_res_stats[counter][1]);
 403                 avg_running = avg_stats(&event_res_stats[counter][2]);
 404
 405                 fprintf(stderr, "  (scaled from %.2f%%)",
 406                                 100 * avg_running / avg_enabled);
 407         }
 408
 409         fprintf(stderr, "\n");
 410 }
 411
 412 static void print_stat(int argc, const char **argv)
 413 {
 414         int i, counter;
 415
 416         fflush(stdout);
 417
 418         fprintf(stderr, "\n");
 419         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 420
 421         for (i = 1; i < argc; i++)
 422                 fprintf(stderr, " %s", argv[i]);
 423
 424         fprintf(stderr, "\'");
 425         if (run_count > 1)
 426                 fprintf(stderr, " (%d runs)", run_count);
 427         fprintf(stderr, ":\n\n");
 428
 429         for (counter = 0; counter < nr_counters; counter++)
 430                 print_counter(counter);
 431
 432         fprintf(stderr, "\n");
 433         fprintf(stderr, " %14.9f  seconds time elapsed",
 434                         avg_stats(&walltime_nsecs_stats)/1e9);
 435         if (run_count > 1) {
 436                 fprintf(stderr, "   ( +- %7.3f%% )",
 437                                 100*stddev_stats(&walltime_nsecs_stats) /
 438                                 avg_stats(&walltime_nsecs_stats));
 439         }
 440         fprintf(stderr, "\n\n");
 441 }
 442
 443 static volatile int signr = -1;
 444
 445 static void skip_signal(int signo)
 446 {
 447         signr = signo;
 448 }
 449
 450 static void sig_atexit(void)
 451 {
 452         if (child_pid != -1)
 453                 kill(child_pid, SIGTERM);
 454
 455         if (signr == -1)
 456                 return;
 457
 458         signal(signr, SIG_DFL);
 459         kill(getpid(), signr);
 460 }
 461
 462 static const char * const stat_usage[] = {
 463         "perf stat [<options>] <command>",
 464         NULL
 465 };
 466
 467 static const struct option options[] = {
 468         OPT_CALLBACK('e', "event", NULL, "event",
 469                      "event selector. use 'perf list' to list available events",
 470                      parse_events),
 471         OPT_BOOLEAN('i', "inherit", &inherit,
 472                     "child tasks inherit counters"),
 473         OPT_INTEGER('p', "pid", &target_pid,
 474                     "stat events on existing pid"),
 475         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 476                     "system-wide collection from all CPUs"),
 477         OPT_BOOLEAN('c', "scale", &scale,
 478                     "scale/normalize counters"),
 479         OPT_BOOLEAN('v', "verbose", &verbose,
 480                     "be more verbose (show counter open errors, etc)"),
 481         OPT_INTEGER('r', "repeat", &run_count,
 482                     "repeat command and print average + stddev (max: 100)"),
 483         OPT_BOOLEAN('n', "null", &null_run,
 484                     "null run - dont start any counters"),
 485         OPT_END()
 486 };
 487
 488 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 489 {
 490         int status;
 491
 492         argc = parse_options(argc, argv, options, stat_usage,
 493                 PARSE_OPT_STOP_AT_NON_OPTION);
 494         if (!argc)
 495                 usage_with_options(stat_usage, options);
 496         if (run_count <= 0)
 497                 usage_with_options(stat_usage, options);
 498
 499         /* Set attrs and nr_counters if no event is selected and !null_run */
 500         if (!null_run && !nr_counters) {
 501                 memcpy(attrs, default_attrs, sizeof(default_attrs));
 502                 nr_counters = ARRAY_SIZE(default_attrs);
 503         }
 504
 505         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 506         assert(nr_cpus <= MAX_NR_CPUS);
 507         assert((int)nr_cpus >= 0);
 508
 509         /*
 510          * We dont want to block the signals - that would cause
 511          * child tasks to inherit that and Ctrl-C would not work.
 512          * What we want is for Ctrl-C to work in the exec()-ed
 513          * task, but being ignored by perf stat itself:
 514          */
 515         atexit(sig_atexit);
 516         signal(SIGINT,  skip_signal);
 517         signal(SIGALRM, skip_signal);
 518         signal(SIGABRT, skip_signal);
 519
 520         status = 0;
 521         for (run_idx = 0; run_idx < run_count; run_idx++) {
 522                 if (run_count != 1 && verbose)
 523                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 524                 status = run_perf_stat(argc, argv);
 525         }
 526
 527         print_stat(argc, argv);
 528
 529         return status;
 530 }