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  *
  36  * Released under the GPL v2. (and only v2, not any later version)
  37  */
  38
  39 #include "perf.h"
  40 #include "builtin.h"
  41 #include "util/util.h"
  42 #include "util/parse-options.h"
  43 #include "util/parse-events.h"
  44
  45 #include <sys/prctl.h>
  46 #include <math.h>
  47
  48 static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
  49
  50   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
  51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
  52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS  },
  53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
  54
  55   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
  56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS    },
  57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
  58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES    },
  59
  60 };
  61
  62 #define MAX_RUN                 100
  63
  64 static int                      system_wide                     =  0;
  65 static int                      verbose                         =  0;
  66 static 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 int                      target_pid                      = -1;
  73 static int                      null_run                        =  0;
  74
  75 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  76
  77 static u64                      runtime_nsecs[MAX_RUN];
  78 static u64                      walltime_nsecs[MAX_RUN];
  79 static u64                      runtime_cycles[MAX_RUN];
  80
  81 static u64                      event_res[MAX_RUN][MAX_COUNTERS][3];
  82 static u64                      event_scaled[MAX_RUN][MAX_COUNTERS];
  83
  84 static u64                      event_res_avg[MAX_COUNTERS][3];
  85 static u64                      event_res_noise[MAX_COUNTERS][3];
  86
  87 static u64                      event_scaled_avg[MAX_COUNTERS];
  88
  89 static u64                      runtime_nsecs_avg;
  90 static u64                      runtime_nsecs_noise;
  91
  92 static u64                      walltime_nsecs_avg;
  93 static u64                      walltime_nsecs_noise;
  94
  95 static u64                      runtime_cycles_avg;
  96 static u64                      runtime_cycles_noise;
  97
  98 #define ERR_PERF_OPEN \
  99 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
 100
 101 static void create_perf_stat_counter(int counter)
 102 {
 103         struct perf_counter_attr *attr = attrs + counter;
 104
 105         if (scale)
 106                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 107                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 108
 109         if (system_wide) {
 110                 int cpu;
 111                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 112                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 113                         if (fd[cpu][counter] < 0 && verbose)
 114                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 115                                         fd[cpu][counter], strerror(errno));
 116                 }
 117         } else {
 118                 attr->inherit   = inherit;
 119                 attr->disabled  = 1;
 120
 121                 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
 122                 if (fd[0][counter] < 0 && verbose)
 123                         fprintf(stderr, ERR_PERF_OPEN, counter,
 124                                 fd[0][counter], strerror(errno));
 125         }
 126 }
 127
 128 /*
 129  * Does the counter have nsecs as a unit?
 130  */
 131 static inline int nsec_counter(int counter)
 132 {
 133         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 134                 return 0;
 135
 136         if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
 137                 return 1;
 138
 139         if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 140                 return 1;
 141
 142         return 0;
 143 }
 144
 145 /*
 146  * Read out the results of a single counter:
 147  */
 148 static void read_counter(int counter)
 149 {
 150         u64 *count, single_count[3];
 151         ssize_t res;
 152         int cpu, nv;
 153         int scaled;
 154
 155         count = event_res[run_idx][counter];
 156
 157         count[0] = count[1] = count[2] = 0;
 158
 159         nv = scale ? 3 : 1;
 160         for (cpu = 0; cpu < nr_cpus; cpu++) {
 161                 if (fd[cpu][counter] < 0)
 162                         continue;
 163
 164                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 165                 assert(res == nv * sizeof(u64));
 166                 close(fd[cpu][counter]);
 167                 fd[cpu][counter] = -1;
 168
 169                 count[0] += single_count[0];
 170                 if (scale) {
 171                         count[1] += single_count[1];
 172                         count[2] += single_count[2];
 173                 }
 174         }
 175
 176         scaled = 0;
 177         if (scale) {
 178                 if (count[2] == 0) {
 179                         event_scaled[run_idx][counter] = -1;
 180                         count[0] = 0;
 181                         return;
 182                 }
 183
 184                 if (count[2] < count[1]) {
 185                         event_scaled[run_idx][counter] = 1;
 186                         count[0] = (unsigned long long)
 187                                 ((double)count[0] * count[1] / count[2] + 0.5);
 188                 }
 189         }
 190         /*
 191          * Save the full runtime - to allow normalization during printout:
 192          */
 193         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 194                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 195                 runtime_nsecs[run_idx] = count[0];
 196         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 197                 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
 198                 runtime_cycles[run_idx] = count[0];
 199 }
 200
 201 static int run_perf_stat(int argc, const char **argv)
 202 {
 203         unsigned long long t0, t1;
 204         int status = 0;
 205         int counter;
 206         int pid;
 207
 208         if (!system_wide)
 209                 nr_cpus = 1;
 210
 211         for (counter = 0; counter < nr_counters; counter++)
 212                 create_perf_stat_counter(counter);
 213
 214         /*
 215          * Enable counters and exec the command:
 216          */
 217         t0 = rdclock();
 218         prctl(PR_TASK_PERF_COUNTERS_ENABLE);
 219
 220         if ((pid = fork()) < 0)
 221                 perror("failed to fork");
 222
 223         if (!pid) {
 224                 if (execvp(argv[0], (char **)argv)) {
 225                         perror(argv[0]);
 226                         exit(-1);
 227                 }
 228         }
 229
 230         wait(&status);
 231
 232         prctl(PR_TASK_PERF_COUNTERS_DISABLE);
 233         t1 = rdclock();
 234
 235         walltime_nsecs[run_idx] = t1 - t0;
 236
 237         for (counter = 0; counter < nr_counters; counter++)
 238                 read_counter(counter);
 239
 240         return WEXITSTATUS(status);
 241 }
 242
 243 static void print_noise(u64 *count, u64 *noise)
 244 {
 245         if (run_count > 1)
 246                 fprintf(stderr, "   ( +- %7.3f%% )",
 247                         (double)noise[0]/(count[0]+1)*100.0);
 248 }
 249
 250 static void nsec_printout(int counter, u64 *count, u64 *noise)
 251 {
 252         double msecs = (double)count[0] / 1000000;
 253
 254         fprintf(stderr, " %14.6f  %-20s", msecs, event_name(counter));
 255
 256         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 257                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
 258
 259                 if (walltime_nsecs_avg)
 260                         fprintf(stderr, " # %10.3f CPUs ",
 261                                 (double)count[0] / (double)walltime_nsecs_avg);
 262         }
 263         print_noise(count, noise);
 264 }
 265
 266 static void abs_printout(int counter, u64 *count, u64 *noise)
 267 {
 268         fprintf(stderr, " %14Ld  %-20s", count[0], event_name(counter));
 269
 270         if (runtime_cycles_avg &&
 271                 attrs[counter].type == PERF_TYPE_HARDWARE &&
 272                         attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
 273
 274                 fprintf(stderr, " # %10.3f IPC  ",
 275                         (double)count[0] / (double)runtime_cycles_avg);
 276         } else {
 277                 if (runtime_nsecs_avg) {
 278                         fprintf(stderr, " # %10.3f M/sec",
 279                                 (double)count[0]/runtime_nsecs_avg*1000.0);
 280                 }
 281         }
 282         print_noise(count, noise);
 283 }
 284
 285 /*
 286  * Print out the results of a single counter:
 287  */
 288 static void print_counter(int counter)
 289 {
 290         u64 *count, *noise;
 291         int scaled;
 292
 293         count = event_res_avg[counter];
 294         noise = event_res_noise[counter];
 295         scaled = event_scaled_avg[counter];
 296
 297         if (scaled == -1) {
 298                 fprintf(stderr, " %14s  %-20s\n",
 299                         "<not counted>", event_name(counter));
 300                 return;
 301         }
 302
 303         if (nsec_counter(counter))
 304                 nsec_printout(counter, count, noise);
 305         else
 306                 abs_printout(counter, count, noise);
 307
 308         if (scaled)
 309                 fprintf(stderr, "  (scaled from %.2f%%)",
 310                         (double) count[2] / count[1] * 100);
 311
 312         fprintf(stderr, "\n");
 313 }
 314
 315 /*
 316  * normalize_noise noise values down to stddev:
 317  */
 318 static void normalize_noise(u64 *val)
 319 {
 320         double res;
 321
 322         res = (double)*val / (run_count * sqrt((double)run_count));
 323
 324         *val = (u64)res;
 325 }
 326
 327 static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
 328 {
 329         *avg += *val;
 330
 331         if (verbose > 1)
 332                 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
 333 }
 334 /*
 335  * Calculate the averages and noises:
 336  */
 337 static void calc_avg(void)
 338 {
 339         int i, j;
 340
 341         if (verbose > 1)
 342                 fprintf(stderr, "\n");
 343
 344         for (i = 0; i < run_count; i++) {
 345                 update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
 346                 update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
 347                 update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
 348
 349                 for (j = 0; j < nr_counters; j++) {
 350                         update_avg("counter/0", j,
 351                                 event_res_avg[j]+0, event_res[i][j]+0);
 352                         update_avg("counter/1", j,
 353                                 event_res_avg[j]+1, event_res[i][j]+1);
 354                         update_avg("counter/2", j,
 355                                 event_res_avg[j]+2, event_res[i][j]+2);
 356                         update_avg("scaled", j,
 357                                 event_scaled_avg + j, event_scaled[i]+j);
 358                 }
 359         }
 360         runtime_nsecs_avg /= run_count;
 361         walltime_nsecs_avg /= run_count;
 362         runtime_cycles_avg /= run_count;
 363
 364         for (j = 0; j < nr_counters; j++) {
 365                 event_res_avg[j][0] /= run_count;
 366                 event_res_avg[j][1] /= run_count;
 367                 event_res_avg[j][2] /= run_count;
 368         }
 369
 370         for (i = 0; i < run_count; i++) {
 371                 runtime_nsecs_noise +=
 372                         abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
 373                 walltime_nsecs_noise +=
 374                         abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
 375                 runtime_cycles_noise +=
 376                         abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
 377
 378                 for (j = 0; j < nr_counters; j++) {
 379                         event_res_noise[j][0] +=
 380                                 abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
 381                         event_res_noise[j][1] +=
 382                                 abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
 383                         event_res_noise[j][2] +=
 384                                 abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
 385                 }
 386         }
 387
 388         normalize_noise(&runtime_nsecs_noise);
 389         normalize_noise(&walltime_nsecs_noise);
 390         normalize_noise(&runtime_cycles_noise);
 391
 392         for (j = 0; j < nr_counters; j++) {
 393                 normalize_noise(&event_res_noise[j][0]);
 394                 normalize_noise(&event_res_noise[j][1]);
 395                 normalize_noise(&event_res_noise[j][2]);
 396         }
 397 }
 398
 399 static void print_stat(int argc, const char **argv)
 400 {
 401         int i, counter;
 402
 403         calc_avg();
 404
 405         fflush(stdout);
 406
 407         fprintf(stderr, "\n");
 408         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 409
 410         for (i = 1; i < argc; i++)
 411                 fprintf(stderr, " %s", argv[i]);
 412
 413         fprintf(stderr, "\'");
 414         if (run_count > 1)
 415                 fprintf(stderr, " (%d runs)", run_count);
 416         fprintf(stderr, ":\n\n");
 417
 418         for (counter = 0; counter < nr_counters; counter++)
 419                 print_counter(counter);
 420
 421
 422         fprintf(stderr, "\n");
 423         fprintf(stderr, " %14.9f  seconds time elapsed.\n",
 424                         (double)walltime_nsecs_avg/1e9);
 425         fprintf(stderr, "\n");
 426 }
 427
 428 static volatile int signr = -1;
 429
 430 static void skip_signal(int signo)
 431 {
 432         signr = signo;
 433 }
 434
 435 static void sig_atexit(void)
 436 {
 437         if (signr == -1)
 438                 return;
 439
 440         signal(signr, SIG_DFL);
 441         kill(getpid(), signr);
 442 }
 443
 444 static const char * const stat_usage[] = {
 445         "perf stat [<options>] <command>",
 446         NULL
 447 };
 448
 449 static const struct option options[] = {
 450         OPT_CALLBACK('e', "event", NULL, "event",
 451                      "event selector. use 'perf list' to list available events",
 452                      parse_events),
 453         OPT_BOOLEAN('i', "inherit", &inherit,
 454                     "child tasks inherit counters"),
 455         OPT_INTEGER('p', "pid", &target_pid,
 456                     "stat events on existing pid"),
 457         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 458                     "system-wide collection from all CPUs"),
 459         OPT_BOOLEAN('S', "scale", &scale,
 460                     "scale/normalize counters"),
 461         OPT_BOOLEAN('v', "verbose", &verbose,
 462                     "be more verbose (show counter open errors, etc)"),
 463         OPT_INTEGER('r', "repeat", &run_count,
 464                     "repeat command and print average + stddev (max: 100)"),
 465         OPT_BOOLEAN('n', "null", &null_run,
 466                     "null run - dont start any counters"),
 467         OPT_END()
 468 };
 469
 470 int cmd_stat(int argc, const char **argv, const char *prefix)
 471 {
 472         int status;
 473
 474         memcpy(attrs, default_attrs, sizeof(attrs));
 475
 476         argc = parse_options(argc, argv, options, stat_usage, 0);
 477         if (!argc)
 478                 usage_with_options(stat_usage, options);
 479         if (run_count <= 0 || run_count > MAX_RUN)
 480                 usage_with_options(stat_usage, options);
 481
 482         if (!null_run && !nr_counters)
 483                 nr_counters = 8;
 484
 485         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 486         assert(nr_cpus <= MAX_NR_CPUS);
 487         assert(nr_cpus >= 0);
 488
 489         /*
 490          * We dont want to block the signals - that would cause
 491          * child tasks to inherit that and Ctrl-C would not work.
 492          * What we want is for Ctrl-C to work in the exec()-ed
 493          * task, but being ignored by perf stat itself:
 494          */
 495         atexit(sig_atexit);
 496         signal(SIGINT,  skip_signal);
 497         signal(SIGALRM, skip_signal);
 498         signal(SIGABRT, skip_signal);
 499
 500         status = 0;
 501         for (run_idx = 0; run_idx < run_count; run_idx++) {
 502                 if (run_count != 1 && verbose)
 503                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 504                 status = run_perf_stat(argc, argv);
 505         }
 506
 507         print_stat(argc, argv);
 508
 509         return status;
 510 }