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