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[MAX_COUNTERS] = {
  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)
 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
 122                 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
 123                 if (fd[0][counter] < 0 && verbose)
 124                         fprintf(stderr, ERR_PERF_OPEN, counter,
 125                                 fd[0][counter], strerror(errno));
 126         }
 127 }
 128
 129 /*
 130  * Does the counter have nsecs as a unit?
 131  */
 132 static inline int nsec_counter(int counter)
 133 {
 134         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 135                 return 0;
 136
 137         if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
 138                 return 1;
 139
 140         if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 141                 return 1;
 142
 143         return 0;
 144 }
 145
 146 /*
 147  * Read out the results of a single counter:
 148  */
 149 static void read_counter(int counter)
 150 {
 151         u64 *count, single_count[3];
 152         ssize_t res;
 153         int cpu, nv;
 154         int scaled;
 155
 156         count = event_res[run_idx][counter];
 157
 158         count[0] = count[1] = count[2] = 0;
 159
 160         nv = scale ? 3 : 1;
 161         for (cpu = 0; cpu < nr_cpus; cpu++) {
 162                 if (fd[cpu][counter] < 0)
 163                         continue;
 164
 165                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 166                 assert(res == nv * sizeof(u64));
 167                 close(fd[cpu][counter]);
 168                 fd[cpu][counter] = -1;
 169
 170                 count[0] += single_count[0];
 171                 if (scale) {
 172                         count[1] += single_count[1];
 173                         count[2] += single_count[2];
 174                 }
 175         }
 176
 177         scaled = 0;
 178         if (scale) {
 179                 if (count[2] == 0) {
 180                         event_scaled[run_idx][counter] = -1;
 181                         count[0] = 0;
 182                         return;
 183                 }
 184
 185                 if (count[2] < count[1]) {
 186                         event_scaled[run_idx][counter] = 1;
 187                         count[0] = (unsigned long long)
 188                                 ((double)count[0] * count[1] / count[2] + 0.5);
 189                 }
 190         }
 191         /*
 192          * Save the full runtime - to allow normalization during printout:
 193          */
 194         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 195                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 196                 runtime_nsecs[run_idx] = count[0];
 197         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 198                 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
 199                 runtime_cycles[run_idx] = count[0];
 200 }
 201
 202 static int run_perf_stat(int argc, const char **argv)
 203 {
 204         unsigned long long t0, t1;
 205         int status = 0;
 206         int counter;
 207         int pid;
 208
 209         if (!system_wide)
 210                 nr_cpus = 1;
 211
 212         for (counter = 0; counter < nr_counters; counter++)
 213                 create_perf_stat_counter(counter);
 214
 215         /*
 216          * Enable counters and exec the command:
 217          */
 218         t0 = rdclock();
 219         prctl(PR_TASK_PERF_COUNTERS_ENABLE);
 220
 221         if ((pid = fork()) < 0)
 222                 perror("failed to fork");
 223
 224         if (!pid) {
 225                 if (execvp(argv[0], (char **)argv)) {
 226                         perror(argv[0]);
 227                         exit(-1);
 228                 }
 229         }
 230
 231         wait(&status);
 232
 233         prctl(PR_TASK_PERF_COUNTERS_DISABLE);
 234         t1 = rdclock();
 235
 236         walltime_nsecs[run_idx] = t1 - t0;
 237
 238         for (counter = 0; counter < nr_counters; counter++)
 239                 read_counter(counter);
 240
 241         return WEXITSTATUS(status);
 242 }
 243
 244 static void print_noise(u64 *count, u64 *noise)
 245 {
 246         if (run_count > 1)
 247                 fprintf(stderr, "   ( +- %7.3f%% )",
 248                         (double)noise[0]/(count[0]+1)*100.0);
 249 }
 250
 251 static void nsec_printout(int counter, u64 *count, u64 *noise)
 252 {
 253         double msecs = (double)count[0] / 1000000;
 254
 255         fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
 256
 257         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 258                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
 259
 260                 if (walltime_nsecs_avg)
 261                         fprintf(stderr, " # %10.3f CPUs ",
 262                                 (double)count[0] / (double)walltime_nsecs_avg);
 263         }
 264         print_noise(count, noise);
 265 }
 266
 267 static void abs_printout(int counter, u64 *count, u64 *noise)
 268 {
 269         fprintf(stderr, " %14Ld  %-24s", count[0], event_name(counter));
 270
 271         if (runtime_cycles_avg &&
 272                 attrs[counter].type == PERF_TYPE_HARDWARE &&
 273                         attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
 274
 275                 fprintf(stderr, " # %10.3f IPC  ",
 276                         (double)count[0] / (double)runtime_cycles_avg);
 277         } else {
 278                 if (runtime_nsecs_avg) {
 279                         fprintf(stderr, " # %10.3f M/sec",
 280                                 (double)count[0]/runtime_nsecs_avg*1000.0);
 281                 }
 282         }
 283         print_noise(count, noise);
 284 }
 285
 286 /*
 287  * Print out the results of a single counter:
 288  */
 289 static void print_counter(int counter)
 290 {
 291         u64 *count, *noise;
 292         int scaled;
 293
 294         count = event_res_avg[counter];
 295         noise = event_res_noise[counter];
 296         scaled = event_scaled_avg[counter];
 297
 298         if (scaled == -1) {
 299                 fprintf(stderr, " %14s  %-24s\n",
 300                         "<not counted>", event_name(counter));
 301                 return;
 302         }
 303
 304         if (nsec_counter(counter))
 305                 nsec_printout(counter, count, noise);
 306         else
 307                 abs_printout(counter, count, noise);
 308
 309         if (scaled)
 310                 fprintf(stderr, "  (%7.2fx scaled)", (double)count[1]/count[2]);
 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                         if (event_scaled[i][j] != -1)
 357                                 update_avg("scaled", j,
 358                                         event_scaled_avg + j, event_scaled[i]+j);
 359                         else
 360                                 event_scaled_avg[j] = -1;
 361                 }
 362         }
 363         runtime_nsecs_avg /= run_count;
 364         walltime_nsecs_avg /= run_count;
 365         runtime_cycles_avg /= run_count;
 366
 367         for (j = 0; j < nr_counters; j++) {
 368                 event_res_avg[j][0] /= run_count;
 369                 event_res_avg[j][1] /= run_count;
 370                 event_res_avg[j][2] /= run_count;
 371         }
 372
 373         for (i = 0; i < run_count; i++) {
 374                 runtime_nsecs_noise +=
 375                         abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
 376                 walltime_nsecs_noise +=
 377                         abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
 378                 runtime_cycles_noise +=
 379                         abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
 380
 381                 for (j = 0; j < nr_counters; j++) {
 382                         event_res_noise[j][0] +=
 383                                 abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
 384                         event_res_noise[j][1] +=
 385                                 abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
 386                         event_res_noise[j][2] +=
 387                                 abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
 388                 }
 389         }
 390
 391         normalize_noise(&runtime_nsecs_noise);
 392         normalize_noise(&walltime_nsecs_noise);
 393         normalize_noise(&runtime_cycles_noise);
 394
 395         for (j = 0; j < nr_counters; j++) {
 396                 normalize_noise(&event_res_noise[j][0]);
 397                 normalize_noise(&event_res_noise[j][1]);
 398                 normalize_noise(&event_res_noise[j][2]);
 399         }
 400 }
 401
 402 static void print_stat(int argc, const char **argv)
 403 {
 404         int i, counter;
 405
 406         calc_avg();
 407
 408         fflush(stdout);
 409
 410         fprintf(stderr, "\n");
 411         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 412
 413         for (i = 1; i < argc; i++)
 414                 fprintf(stderr, " %s", argv[i]);
 415
 416         fprintf(stderr, "\'");
 417         if (run_count > 1)
 418                 fprintf(stderr, " (%d runs)", run_count);
 419         fprintf(stderr, ":\n\n");
 420
 421         for (counter = 0; counter < nr_counters; counter++)
 422                 print_counter(counter);
 423
 424         fprintf(stderr, "\n");
 425         fprintf(stderr, " %14.9f  seconds time elapsed",
 426                         (double)walltime_nsecs_avg/1e9);
 427         if (run_count > 1) {
 428                 fprintf(stderr, "   ( +- %7.3f%% )",
 429                         100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
 430         }
 431         fprintf(stderr, "\n\n");
 432 }
 433
 434 static volatile int signr = -1;
 435
 436 static void skip_signal(int signo)
 437 {
 438         signr = signo;
 439 }
 440
 441 static void sig_atexit(void)
 442 {
 443         if (signr == -1)
 444                 return;
 445
 446         signal(signr, SIG_DFL);
 447         kill(getpid(), signr);
 448 }
 449
 450 static const char * const stat_usage[] = {
 451         "perf stat [<options>] <command>",
 452         NULL
 453 };
 454
 455 static const struct option options[] = {
 456         OPT_CALLBACK('e', "event", NULL, "event",
 457                      "event selector. use 'perf list' to list available events",
 458                      parse_events),
 459         OPT_BOOLEAN('i', "inherit", &inherit,
 460                     "child tasks inherit counters"),
 461         OPT_INTEGER('p', "pid", &target_pid,
 462                     "stat events on existing pid"),
 463         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 464                     "system-wide collection from all CPUs"),
 465         OPT_BOOLEAN('S', "scale", &scale,
 466                     "scale/normalize counters"),
 467         OPT_BOOLEAN('v', "verbose", &verbose,
 468                     "be more verbose (show counter open errors, etc)"),
 469         OPT_INTEGER('r', "repeat", &run_count,
 470                     "repeat command and print average + stddev (max: 100)"),
 471         OPT_BOOLEAN('n', "null", &null_run,
 472                     "null run - dont start any counters"),
 473         OPT_END()
 474 };
 475
 476 int cmd_stat(int argc, const char **argv, const char *prefix)
 477 {
 478         int status;
 479
 480         memcpy(attrs, default_attrs, sizeof(attrs));
 481
 482         argc = parse_options(argc, argv, options, stat_usage, 0);
 483         if (!argc)
 484                 usage_with_options(stat_usage, options);
 485         if (run_count <= 0 || run_count > MAX_RUN)
 486                 usage_with_options(stat_usage, options);
 487
 488         if (!null_run && !nr_counters)
 489                 nr_counters = 8;
 490
 491         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 492         assert(nr_cpus <= MAX_NR_CPUS);
 493         assert(nr_cpus >= 0);
 494
 495         /*
 496          * We dont want to block the signals - that would cause
 497          * child tasks to inherit that and Ctrl-C would not work.
 498          * What we want is for Ctrl-C to work in the exec()-ed
 499          * task, but being ignored by perf stat itself:
 500          */
 501         atexit(sig_atexit);
 502         signal(SIGINT,  skip_signal);
 503         signal(SIGALRM, skip_signal);
 504         signal(SIGABRT, skip_signal);
 505
 506         status = 0;
 507         for (run_idx = 0; run_idx < run_count; run_idx++) {
 508                 if (run_count != 1 && verbose)
 509                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 510                 status = run_perf_stat(argc, argv);
 511         }
 512
 513         print_stat(argc, argv);
 514
 515         return status;
 516 }