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