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
  46 #include <sys/prctl.h>
  47 #include <math.h>
  48
  49 static struct perf_counter_attr default_attrs[] = {
  50
  51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
  52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
  53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS  },
  54   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
  55
  56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
  57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS    },
  58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
  59   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES    },
  60
  61 };
  62
  63 #define MAX_RUN                 100
  64
  65 static int                      system_wide                     =  0;
  66 static int                      verbose                         =  0;
  67 static int                      nr_cpus                         =  0;
  68 static int                      run_idx                         =  0;
  69
  70 static int                      run_count                       =  1;
  71 static int                      inherit                         =  1;
  72 static int                      scale                           =  1;
  73 static int                      target_pid                      = -1;
  74 static int                      null_run                        =  0;
  75
  76 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  77
  78 static u64                      runtime_nsecs[MAX_RUN];
  79 static u64                      walltime_nsecs[MAX_RUN];
  80 static u64                      runtime_cycles[MAX_RUN];
  81
  82 static u64                      event_res[MAX_RUN][MAX_COUNTERS][3];
  83 static u64                      event_scaled[MAX_RUN][MAX_COUNTERS];
  84
  85 static u64                      event_res_avg[MAX_COUNTERS][3];
  86 static u64                      event_res_noise[MAX_COUNTERS][3];
  87
  88 static u64                      event_scaled_avg[MAX_COUNTERS];
  89
  90 static u64                      runtime_nsecs_avg;
  91 static u64                      runtime_nsecs_noise;
  92
  93 static u64                      walltime_nsecs_avg;
  94 static u64                      walltime_nsecs_noise;
  95
  96 static u64                      runtime_cycles_avg;
  97 static u64                      runtime_cycles_noise;
  98
  99 #define ERR_PERF_OPEN \
 100 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
 101
 102 static void create_perf_stat_counter(int counter, int pid)
 103 {
 104         struct perf_counter_attr *attr = attrs + counter;
 105
 106         if (scale)
 107                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 108                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 109
 110         if (system_wide) {
 111                 int cpu;
 112                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 113                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 114                         if (fd[cpu][counter] < 0 && verbose)
 115                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 116                                         fd[cpu][counter], strerror(errno));
 117                 }
 118         } else {
 119                 attr->inherit        = inherit;
 120                 attr->disabled       = 1;
 121                 attr->enable_on_exec = 1;
 122
 123                 fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
 124                 if (fd[0][counter] < 0 && verbose)
 125                         fprintf(stderr, ERR_PERF_OPEN, counter,
 126                                 fd[0][counter], strerror(errno));
 127         }
 128 }
 129
 130 /*
 131  * Does the counter have nsecs as a unit?
 132  */
 133 static inline int nsec_counter(int counter)
 134 {
 135         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 136                 return 0;
 137
 138         if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
 139                 return 1;
 140
 141         if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 142                 return 1;
 143
 144         return 0;
 145 }
 146
 147 /*
 148  * Read out the results of a single counter:
 149  */
 150 static void read_counter(int counter)
 151 {
 152         u64 *count, single_count[3];
 153         ssize_t res;
 154         int cpu, nv;
 155         int scaled;
 156
 157         count = event_res[run_idx][counter];
 158
 159         count[0] = count[1] = count[2] = 0;
 160
 161         nv = scale ? 3 : 1;
 162         for (cpu = 0; cpu < nr_cpus; cpu++) {
 163                 if (fd[cpu][counter] < 0)
 164                         continue;
 165
 166                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 167                 assert(res == nv * sizeof(u64));
 168                 close(fd[cpu][counter]);
 169                 fd[cpu][counter] = -1;
 170
 171                 count[0] += single_count[0];
 172                 if (scale) {
 173                         count[1] += single_count[1];
 174                         count[2] += single_count[2];
 175                 }
 176         }
 177
 178         scaled = 0;
 179         if (scale) {
 180                 if (count[2] == 0) {
 181                         event_scaled[run_idx][counter] = -1;
 182                         count[0] = 0;
 183                         return;
 184                 }
 185
 186                 if (count[2] < count[1]) {
 187                         event_scaled[run_idx][counter] = 1;
 188                         count[0] = (unsigned long long)
 189                                 ((double)count[0] * count[1] / count[2] + 0.5);
 190                 }
 191         }
 192         /*
 193          * Save the full runtime - to allow normalization during printout:
 194          */
 195         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 196                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 197                 runtime_nsecs[run_idx] = count[0];
 198         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 199                 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
 200                 runtime_cycles[run_idx] = count[0];
 201 }
 202
 203 static int run_perf_stat(int argc, const char **argv)
 204 {
 205         unsigned long long t0, t1;
 206         int status = 0;
 207         int counter;
 208         int pid;
 209         int child_ready_pipe[2], go_pipe[2];
 210         char buf;
 211
 212         if (!system_wide)
 213                 nr_cpus = 1;
 214
 215         if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
 216                 perror("failed to create pipes");
 217                 exit(1);
 218         }
 219
 220         if ((pid = fork()) < 0)
 221                 perror("failed to fork");
 222
 223         if (!pid) {
 224                 close(child_ready_pipe[0]);
 225                 close(go_pipe[1]);
 226                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 227
 228                 /*
 229                  * Do a dummy execvp to get the PLT entry resolved,
 230                  * so we avoid the resolver overhead on the real
 231                  * execvp call.
 232                  */
 233                 execvp("", (char **)argv);
 234
 235                 /*
 236                  * Tell the parent we're ready to go
 237                  */
 238                 close(child_ready_pipe[1]);
 239
 240                 /*
 241                  * Wait until the parent tells us to go.
 242                  */
 243                 read(go_pipe[0], &buf, 1);
 244
 245                 execvp(argv[0], (char **)argv);
 246
 247                 perror(argv[0]);
 248                 exit(-1);
 249         }
 250
 251         /*
 252          * Wait for the child to be ready to exec.
 253          */
 254         close(child_ready_pipe[1]);
 255         close(go_pipe[0]);
 256         read(child_ready_pipe[0], &buf, 1);
 257         close(child_ready_pipe[0]);
 258
 259         for (counter = 0; counter < nr_counters; counter++)
 260                 create_perf_stat_counter(counter, pid);
 261
 262         /*
 263          * Enable counters and exec the command:
 264          */
 265         t0 = rdclock();
 266
 267         close(go_pipe[1]);
 268         wait(&status);
 269
 270         t1 = rdclock();
 271
 272         walltime_nsecs[run_idx] = t1 - t0;
 273
 274         for (counter = 0; counter < nr_counters; counter++)
 275                 read_counter(counter);
 276
 277         return WEXITSTATUS(status);
 278 }
 279
 280 static void print_noise(u64 *count, u64 *noise)
 281 {
 282         if (run_count > 1)
 283                 fprintf(stderr, "   ( +- %7.3f%% )",
 284                         (double)noise[0]/(count[0]+1)*100.0);
 285 }
 286
 287 static void nsec_printout(int counter, u64 *count, u64 *noise)
 288 {
 289         double msecs = (double)count[0] / 1000000;
 290
 291         fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
 292
 293         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 294                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
 295
 296                 if (walltime_nsecs_avg)
 297                         fprintf(stderr, " # %10.3f CPUs ",
 298                                 (double)count[0] / (double)walltime_nsecs_avg);
 299         }
 300         print_noise(count, noise);
 301 }
 302
 303 static void abs_printout(int counter, u64 *count, u64 *noise)
 304 {
 305         fprintf(stderr, " %14Ld  %-24s", count[0], event_name(counter));
 306
 307         if (runtime_cycles_avg &&
 308                 attrs[counter].type == PERF_TYPE_HARDWARE &&
 309                         attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
 310
 311                 fprintf(stderr, " # %10.3f IPC  ",
 312                         (double)count[0] / (double)runtime_cycles_avg);
 313         } else {
 314                 if (runtime_nsecs_avg) {
 315                         fprintf(stderr, " # %10.3f M/sec",
 316                                 (double)count[0]/runtime_nsecs_avg*1000.0);
 317                 }
 318         }
 319         print_noise(count, noise);
 320 }
 321
 322 /*
 323  * Print out the results of a single counter:
 324  */
 325 static void print_counter(int counter)
 326 {
 327         u64 *count, *noise;
 328         int scaled;
 329
 330         count = event_res_avg[counter];
 331         noise = event_res_noise[counter];
 332         scaled = event_scaled_avg[counter];
 333
 334         if (scaled == -1) {
 335                 fprintf(stderr, " %14s  %-24s\n",
 336                         "<not counted>", event_name(counter));
 337                 return;
 338         }
 339
 340         if (nsec_counter(counter))
 341                 nsec_printout(counter, count, noise);
 342         else
 343                 abs_printout(counter, count, noise);
 344
 345         if (scaled)
 346                 fprintf(stderr, "  (scaled from %.2f%%)",
 347                         (double) count[2] / count[1] * 100);
 348
 349         fprintf(stderr, "\n");
 350 }
 351
 352 /*
 353  * normalize_noise noise values down to stddev:
 354  */
 355 static void normalize_noise(u64 *val)
 356 {
 357         double res;
 358
 359         res = (double)*val / (run_count * sqrt((double)run_count));
 360
 361         *val = (u64)res;
 362 }
 363
 364 static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
 365 {
 366         *avg += *val;
 367
 368         if (verbose > 1)
 369                 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
 370 }
 371 /*
 372  * Calculate the averages and noises:
 373  */
 374 static void calc_avg(void)
 375 {
 376         int i, j;
 377
 378         if (verbose > 1)
 379                 fprintf(stderr, "\n");
 380
 381         for (i = 0; i < run_count; i++) {
 382                 update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
 383                 update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
 384                 update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
 385
 386                 for (j = 0; j < nr_counters; j++) {
 387                         update_avg("counter/0", j,
 388                                 event_res_avg[j]+0, event_res[i][j]+0);
 389                         update_avg("counter/1", j,
 390                                 event_res_avg[j]+1, event_res[i][j]+1);
 391                         update_avg("counter/2", j,
 392                                 event_res_avg[j]+2, event_res[i][j]+2);
 393                         if (event_scaled[i][j] != -1)
 394                                 update_avg("scaled", j,
 395                                         event_scaled_avg + j, event_scaled[i]+j);
 396                         else
 397                                 event_scaled_avg[j] = -1;
 398                 }
 399         }
 400         runtime_nsecs_avg /= run_count;
 401         walltime_nsecs_avg /= run_count;
 402         runtime_cycles_avg /= run_count;
 403
 404         for (j = 0; j < nr_counters; j++) {
 405                 event_res_avg[j][0] /= run_count;
 406                 event_res_avg[j][1] /= run_count;
 407                 event_res_avg[j][2] /= run_count;
 408         }
 409
 410         for (i = 0; i < run_count; i++) {
 411                 runtime_nsecs_noise +=
 412                         abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
 413                 walltime_nsecs_noise +=
 414                         abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
 415                 runtime_cycles_noise +=
 416                         abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
 417
 418                 for (j = 0; j < nr_counters; j++) {
 419                         event_res_noise[j][0] +=
 420                                 abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
 421                         event_res_noise[j][1] +=
 422                                 abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
 423                         event_res_noise[j][2] +=
 424                                 abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
 425                 }
 426         }
 427
 428         normalize_noise(&runtime_nsecs_noise);
 429         normalize_noise(&walltime_nsecs_noise);
 430         normalize_noise(&runtime_cycles_noise);
 431
 432         for (j = 0; j < nr_counters; j++) {
 433                 normalize_noise(&event_res_noise[j][0]);
 434                 normalize_noise(&event_res_noise[j][1]);
 435                 normalize_noise(&event_res_noise[j][2]);
 436         }
 437 }
 438
 439 static void print_stat(int argc, const char **argv)
 440 {
 441         int i, counter;
 442
 443         calc_avg();
 444
 445         fflush(stdout);
 446
 447         fprintf(stderr, "\n");
 448         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 449
 450         for (i = 1; i < argc; i++)
 451                 fprintf(stderr, " %s", argv[i]);
 452
 453         fprintf(stderr, "\'");
 454         if (run_count > 1)
 455                 fprintf(stderr, " (%d runs)", run_count);
 456         fprintf(stderr, ":\n\n");
 457
 458         for (counter = 0; counter < nr_counters; counter++)
 459                 print_counter(counter);
 460
 461         fprintf(stderr, "\n");
 462         fprintf(stderr, " %14.9f  seconds time elapsed",
 463                         (double)walltime_nsecs_avg/1e9);
 464         if (run_count > 1) {
 465                 fprintf(stderr, "   ( +- %7.3f%% )",
 466                         100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
 467         }
 468         fprintf(stderr, "\n\n");
 469 }
 470
 471 static volatile int signr = -1;
 472
 473 static void skip_signal(int signo)
 474 {
 475         signr = signo;
 476 }
 477
 478 static void sig_atexit(void)
 479 {
 480         if (signr == -1)
 481                 return;
 482
 483         signal(signr, SIG_DFL);
 484         kill(getpid(), signr);
 485 }
 486
 487 static const char * const stat_usage[] = {
 488         "perf stat [<options>] <command>",
 489         NULL
 490 };
 491
 492 static const struct option options[] = {
 493         OPT_CALLBACK('e', "event", NULL, "event",
 494                      "event selector. use 'perf list' to list available events",
 495                      parse_events),
 496         OPT_BOOLEAN('i', "inherit", &inherit,
 497                     "child tasks inherit counters"),
 498         OPT_INTEGER('p', "pid", &target_pid,
 499                     "stat events on existing pid"),
 500         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 501                     "system-wide collection from all CPUs"),
 502         OPT_BOOLEAN('S', "scale", &scale,
 503                     "scale/normalize counters"),
 504         OPT_BOOLEAN('v', "verbose", &verbose,
 505                     "be more verbose (show counter open errors, etc)"),
 506         OPT_INTEGER('r', "repeat", &run_count,
 507                     "repeat command and print average + stddev (max: 100)"),
 508         OPT_BOOLEAN('n', "null", &null_run,
 509                     "null run - dont start any counters"),
 510         OPT_END()
 511 };
 512
 513 int cmd_stat(int argc, const char **argv, const char *prefix)
 514 {
 515         int status;
 516
 517         argc = parse_options(argc, argv, options, stat_usage, 0);
 518         if (!argc)
 519                 usage_with_options(stat_usage, options);
 520         if (run_count <= 0 || run_count > MAX_RUN)
 521                 usage_with_options(stat_usage, options);
 522
 523         /* Set attrs and nr_counters if no event is selected and !null_run */
 524         if (!null_run && !nr_counters) {
 525                 memcpy(attrs, default_attrs, sizeof(default_attrs));
 526                 nr_counters = ARRAY_SIZE(default_attrs);
 527         }
 528
 529         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 530         assert(nr_cpus <= MAX_NR_CPUS);
 531         assert(nr_cpus >= 0);
 532
 533         /*
 534          * We dont want to block the signals - that would cause
 535          * child tasks to inherit that and Ctrl-C would not work.
 536          * What we want is for Ctrl-C to work in the exec()-ed
 537          * task, but being ignored by perf stat itself:
 538          */
 539         atexit(sig_atexit);
 540         signal(SIGINT,  skip_signal);
 541         signal(SIGALRM, skip_signal);
 542         signal(SIGABRT, skip_signal);
 543
 544         status = 0;
 545         for (run_idx = 0; run_idx < run_count; run_idx++) {
 546                 if (run_count != 1 && verbose)
 547                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 548                 status = run_perf_stat(argc, argv);
 549         }
 550
 551         print_stat(argc, argv);
 552
 553         return status;
 554 }