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_counter_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 #define MAX_RUN                 100
  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 int                      target_pid                      = -1;
  75 static int                      null_run                        =  0;
  76
  77 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  78
  79 static u64                      runtime_nsecs[MAX_RUN];
  80 static u64                      walltime_nsecs[MAX_RUN];
  81 static u64                      runtime_cycles[MAX_RUN];
  82
  83 static u64                      event_res[MAX_RUN][MAX_COUNTERS][3];
  84 static u64                      event_scaled[MAX_RUN][MAX_COUNTERS];
  85
  86 struct stats
  87 {
  88         double sum;
  89         double sum_sq;
  90 };
  91
  92 static double avg_stats(struct stats *stats)
  93 {
  94         return stats->sum / run_count;
  95 }
  96
  97 /*
  98  * stddev = sqrt(1/N (\Sum n_i^2) - avg(n)^2)
  99  */
 100 static double stddev_stats(struct stats *stats)
 101 {
 102         double avg = stats->sum / run_count;
 103
 104         return sqrt(stats->sum_sq/run_count - avg*avg);
 105 }
 106
 107 struct stats                    event_res_stats[MAX_COUNTERS][3];
 108 struct stats                    event_scaled_stats[MAX_COUNTERS];
 109 struct stats                    runtime_nsecs_stats;
 110 struct stats                    walltime_nsecs_stats;
 111 struct stats                    runtime_cycles_stats;
 112
 113 #define MATCH_EVENT(t, c, counter)                      \
 114         (attrs[counter].type == PERF_TYPE_##t &&        \
 115          attrs[counter].config == PERF_COUNT_##c)
 116
 117 #define ERR_PERF_OPEN \
 118 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
 119
 120 static void create_perf_stat_counter(int counter, int pid)
 121 {
 122         struct perf_counter_attr *attr = attrs + counter;
 123
 124         if (scale)
 125                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 126                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 127
 128         if (system_wide) {
 129                 unsigned int cpu;
 130
 131                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 132                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 133                         if (fd[cpu][counter] < 0 && verbose)
 134                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 135                                         fd[cpu][counter], strerror(errno));
 136                 }
 137         } else {
 138                 attr->inherit        = inherit;
 139                 attr->disabled       = 1;
 140                 attr->enable_on_exec = 1;
 141
 142                 fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
 143                 if (fd[0][counter] < 0 && verbose)
 144                         fprintf(stderr, ERR_PERF_OPEN, counter,
 145                                 fd[0][counter], strerror(errno));
 146         }
 147 }
 148
 149 /*
 150  * Does the counter have nsecs as a unit?
 151  */
 152 static inline int nsec_counter(int counter)
 153 {
 154         if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
 155             MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 156                 return 1;
 157
 158         return 0;
 159 }
 160
 161 /*
 162  * Read out the results of a single counter:
 163  */
 164 static void read_counter(int counter)
 165 {
 166         u64 *count, single_count[3];
 167         unsigned int cpu;
 168         size_t res, nv;
 169         int scaled;
 170
 171         count = event_res[run_idx][counter];
 172
 173         count[0] = count[1] = count[2] = 0;
 174
 175         nv = scale ? 3 : 1;
 176         for (cpu = 0; cpu < nr_cpus; cpu++) {
 177                 if (fd[cpu][counter] < 0)
 178                         continue;
 179
 180                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 181                 assert(res == nv * sizeof(u64));
 182
 183                 close(fd[cpu][counter]);
 184                 fd[cpu][counter] = -1;
 185
 186                 count[0] += single_count[0];
 187                 if (scale) {
 188                         count[1] += single_count[1];
 189                         count[2] += single_count[2];
 190                 }
 191         }
 192
 193         scaled = 0;
 194         if (scale) {
 195                 if (count[2] == 0) {
 196                         event_scaled[run_idx][counter] = -1;
 197                         count[0] = 0;
 198                         return;
 199                 }
 200
 201                 if (count[2] < count[1]) {
 202                         event_scaled[run_idx][counter] = 1;
 203                         count[0] = (unsigned long long)
 204                                 ((double)count[0] * count[1] / count[2] + 0.5);
 205                 }
 206         }
 207         /*
 208          * Save the full runtime - to allow normalization during printout:
 209          */
 210         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 211                 runtime_nsecs[run_idx] = count[0];
 212         if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 213                 runtime_cycles[run_idx] = count[0];
 214 }
 215
 216 static int run_perf_stat(int argc __used, const char **argv)
 217 {
 218         unsigned long long t0, t1;
 219         int status = 0;
 220         int counter;
 221         int pid;
 222         int child_ready_pipe[2], go_pipe[2];
 223         char buf;
 224
 225         if (!system_wide)
 226                 nr_cpus = 1;
 227
 228         if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
 229                 perror("failed to create pipes");
 230                 exit(1);
 231         }
 232
 233         if ((pid = fork()) < 0)
 234                 perror("failed to fork");
 235
 236         if (!pid) {
 237                 close(child_ready_pipe[0]);
 238                 close(go_pipe[1]);
 239                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 240
 241                 /*
 242                  * Do a dummy execvp to get the PLT entry resolved,
 243                  * so we avoid the resolver overhead on the real
 244                  * execvp call.
 245                  */
 246                 execvp("", (char **)argv);
 247
 248                 /*
 249                  * Tell the parent we're ready to go
 250                  */
 251                 close(child_ready_pipe[1]);
 252
 253                 /*
 254                  * Wait until the parent tells us to go.
 255                  */
 256                 if (read(go_pipe[0], &buf, 1) == -1)
 257                         perror("unable to read pipe");
 258
 259                 execvp(argv[0], (char **)argv);
 260
 261                 perror(argv[0]);
 262                 exit(-1);
 263         }
 264
 265         /*
 266          * Wait for the child to be ready to exec.
 267          */
 268         close(child_ready_pipe[1]);
 269         close(go_pipe[0]);
 270         if (read(child_ready_pipe[0], &buf, 1) == -1)
 271                 perror("unable to read pipe");
 272         close(child_ready_pipe[0]);
 273
 274         for (counter = 0; counter < nr_counters; counter++)
 275                 create_perf_stat_counter(counter, pid);
 276
 277         /*
 278          * Enable counters and exec the command:
 279          */
 280         t0 = rdclock();
 281
 282         close(go_pipe[1]);
 283         wait(&status);
 284
 285         t1 = rdclock();
 286
 287         walltime_nsecs[run_idx] = t1 - t0;
 288
 289         for (counter = 0; counter < nr_counters; counter++)
 290                 read_counter(counter);
 291
 292         return WEXITSTATUS(status);
 293 }
 294
 295 static void print_noise(double avg, double stddev)
 296 {
 297         if (run_count > 1)
 298                 fprintf(stderr, "   ( +- %7.3f%% )", 100*stddev / avg);
 299 }
 300
 301 static void nsec_printout(int counter, double avg, double stddev)
 302 {
 303         double msecs = avg / 1e6;
 304
 305         fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
 306
 307         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
 308                 fprintf(stderr, " # %10.3f CPUs ",
 309                                 avg / avg_stats(&walltime_nsecs_stats));
 310         }
 311         print_noise(avg, stddev);
 312 }
 313
 314 static void abs_printout(int counter, double avg, double stddev)
 315 {
 316         fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));
 317
 318         if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
 319                 fprintf(stderr, " # %10.3f IPC  ",
 320                                 avg / avg_stats(&runtime_cycles_stats));
 321         } else {
 322                 fprintf(stderr, " # %10.3f M/sec",
 323                                 1000.0 * avg / avg_stats(&runtime_nsecs_stats));
 324         }
 325         print_noise(avg, stddev);
 326 }
 327
 328 /*
 329  * Print out the results of a single counter:
 330  */
 331 static void print_counter(int counter)
 332 {
 333         double avg, stddev;
 334         int scaled;
 335
 336         avg    = avg_stats(&event_res_stats[counter][0]);
 337         stddev = stddev_stats(&event_res_stats[counter][0]);
 338         scaled = avg_stats(&event_scaled_stats[counter]);
 339
 340         if (scaled == -1) {
 341                 fprintf(stderr, " %14s  %-24s\n",
 342                         "<not counted>", event_name(counter));
 343                 return;
 344         }
 345
 346         if (nsec_counter(counter))
 347                 nsec_printout(counter, avg, stddev);
 348         else
 349                 abs_printout(counter, avg, stddev);
 350
 351         if (scaled) {
 352                 double avg_enabled, avg_running;
 353
 354                 avg_enabled = avg_stats(&event_res_stats[counter][1]);
 355                 avg_running = avg_stats(&event_res_stats[counter][2]);
 356
 357                 fprintf(stderr, "  (scaled from %.2f%%)",
 358                                 100 * avg_running / avg_enabled);
 359         }
 360
 361         fprintf(stderr, "\n");
 362 }
 363
 364 static void update_stats(const char *name, int idx, struct stats *stats, u64 *val)
 365 {
 366         double sq = *val;
 367
 368         stats->sum += *val;
 369         stats->sum_sq += sq * sq;
 370
 371         if (verbose > 1)
 372                 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
 373 }
 374
 375 /*
 376  * Calculate the averages and noises:
 377  */
 378 static void calc_avg(void)
 379 {
 380         int i, j;
 381
 382         if (verbose > 1)
 383                 fprintf(stderr, "\n");
 384
 385         for (i = 0; i < run_count; i++) {
 386                 update_stats("runtime", 0, &runtime_nsecs_stats, runtime_nsecs + i);
 387                 update_stats("walltime", 0, &walltime_nsecs_stats, walltime_nsecs + i);
 388                 update_stats("runtime_cycles", 0, &runtime_cycles_stats, runtime_cycles + i);
 389
 390                 for (j = 0; j < nr_counters; j++) {
 391                         update_stats("counter/0", j,
 392                                 event_res_stats[j]+0, event_res[i][j]+0);
 393                         update_stats("counter/1", j,
 394                                 event_res_stats[j]+1, event_res[i][j]+1);
 395                         update_stats("counter/2", j,
 396                                 event_res_stats[j]+2, event_res[i][j]+2);
 397                         if (event_scaled[i][j] != (u64)-1)
 398                                 update_stats("scaled", j,
 399                                         event_scaled_stats + j, event_scaled[i]+j);
 400                 }
 401         }
 402 }
 403
 404 static void print_stat(int argc, const char **argv)
 405 {
 406         int i, counter;
 407
 408         calc_avg();
 409
 410         fflush(stdout);
 411
 412         fprintf(stderr, "\n");
 413         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 414
 415         for (i = 1; i < argc; i++)
 416                 fprintf(stderr, " %s", argv[i]);
 417
 418         fprintf(stderr, "\'");
 419         if (run_count > 1)
 420                 fprintf(stderr, " (%d runs)", run_count);
 421         fprintf(stderr, ":\n\n");
 422
 423         for (counter = 0; counter < nr_counters; counter++)
 424                 print_counter(counter);
 425
 426         fprintf(stderr, "\n");
 427         fprintf(stderr, " %14.9f  seconds time elapsed",
 428                         avg_stats(&walltime_nsecs_stats)/1e9);
 429         if (run_count > 1) {
 430                 fprintf(stderr, "   ( +- %7.3f%% )",
 431                                 100*stddev_stats(&walltime_nsecs_stats) /
 432                                 avg_stats(&walltime_nsecs_stats));
 433         }
 434         fprintf(stderr, "\n\n");
 435 }
 436
 437 static volatile int signr = -1;
 438
 439 static void skip_signal(int signo)
 440 {
 441         signr = signo;
 442 }
 443
 444 static void sig_atexit(void)
 445 {
 446         if (signr == -1)
 447                 return;
 448
 449         signal(signr, SIG_DFL);
 450         kill(getpid(), signr);
 451 }
 452
 453 static const char * const stat_usage[] = {
 454         "perf stat [<options>] <command>",
 455         NULL
 456 };
 457
 458 static const struct option options[] = {
 459         OPT_CALLBACK('e', "event", NULL, "event",
 460                      "event selector. use 'perf list' to list available events",
 461                      parse_events),
 462         OPT_BOOLEAN('i', "inherit", &inherit,
 463                     "child tasks inherit counters"),
 464         OPT_INTEGER('p', "pid", &target_pid,
 465                     "stat events on existing pid"),
 466         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 467                     "system-wide collection from all CPUs"),
 468         OPT_BOOLEAN('c', "scale", &scale,
 469                     "scale/normalize counters"),
 470         OPT_BOOLEAN('v', "verbose", &verbose,
 471                     "be more verbose (show counter open errors, etc)"),
 472         OPT_INTEGER('r', "repeat", &run_count,
 473                     "repeat command and print average + stddev (max: 100)"),
 474         OPT_BOOLEAN('n', "null", &null_run,
 475                     "null run - dont start any counters"),
 476         OPT_END()
 477 };
 478
 479 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 480 {
 481         int status;
 482
 483         argc = parse_options(argc, argv, options, stat_usage,
 484                 PARSE_OPT_STOP_AT_NON_OPTION);
 485         if (!argc)
 486                 usage_with_options(stat_usage, options);
 487         if (run_count <= 0 || run_count > MAX_RUN)
 488                 usage_with_options(stat_usage, options);
 489
 490         /* Set attrs and nr_counters if no event is selected and !null_run */
 491         if (!null_run && !nr_counters) {
 492                 memcpy(attrs, default_attrs, sizeof(default_attrs));
 493                 nr_counters = ARRAY_SIZE(default_attrs);
 494         }
 495
 496         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 497         assert(nr_cpus <= MAX_NR_CPUS);
 498         assert((int)nr_cpus >= 0);
 499
 500         /*
 501          * We dont want to block the signals - that would cause
 502          * child tasks to inherit that and Ctrl-C would not work.
 503          * What we want is for Ctrl-C to work in the exec()-ed
 504          * task, but being ignored by perf stat itself:
 505          */
 506         atexit(sig_atexit);
 507         signal(SIGINT,  skip_signal);
 508         signal(SIGALRM, skip_signal);
 509         signal(SIGABRT, skip_signal);
 510
 511         status = 0;
 512         for (run_idx = 0; run_idx < run_count; run_idx++) {
 513                 if (run_count != 1 && verbose)
 514                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 515                 status = run_perf_stat(argc, argv);
 516         }
 517
 518         print_stat(argc, argv);
 519
 520         return status;
 521 }