SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 static int                      system_wide                     =  0;
  63 static int                      inherit                         =  1;
  64 static int                      verbose                         =  0;
  65
  66 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  67
  68 static int                      target_pid                      = -1;
  69 static int                      nr_cpus                         =  0;
  70 static unsigned int             page_size;
  71
  72 static int                      scale                           =  1;
  73
  74 static const unsigned int default_count[] = {
  75         1000000,
  76         1000000,
  77           10000,
  78           10000,
  79         1000000,
  80           10000,
  81 };
  82
  83 #define MAX_RUN 100
  84
  85 static int                      run_count               =  1;
  86 static int                      run_idx                 =  0;
  87
  88 static __u64                    event_res[MAX_RUN][MAX_COUNTERS][3];
  89 static __u64                    event_scaled[MAX_RUN][MAX_COUNTERS];
  90
  91 //static __u64                  event_hist[MAX_RUN][MAX_COUNTERS][3];
  92
  93
  94 static __u64                    runtime_nsecs[MAX_RUN];
  95 static __u64                    walltime_nsecs[MAX_RUN];
  96 static __u64                    runtime_cycles[MAX_RUN];
  97
  98 static __u64                    event_res_avg[MAX_COUNTERS][3];
  99 static __u64                    event_res_noise[MAX_COUNTERS][3];
 100
 101 static __u64                    event_scaled_avg[MAX_COUNTERS];
 102
 103 static __u64                    runtime_nsecs_avg;
 104 static __u64                    runtime_nsecs_noise;
 105
 106 static __u64                    walltime_nsecs_avg;
 107 static __u64                    walltime_nsecs_noise;
 108
 109 static __u64                    runtime_cycles_avg;
 110 static __u64                    runtime_cycles_noise;
 111
 112 static void create_perf_stat_counter(int counter)
 113 {
 114         struct perf_counter_attr *attr = attrs + counter;
 115
 116         if (scale)
 117                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 118                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 119
 120         if (system_wide) {
 121                 int cpu;
 122                 for (cpu = 0; cpu < nr_cpus; cpu ++) {
 123                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 124                         if (fd[cpu][counter] < 0 && verbose) {
 125                                 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno));
 126                         }
 127                 }
 128         } else {
 129                 attr->inherit   = inherit;
 130                 attr->disabled  = 1;
 131
 132                 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
 133                 if (fd[0][counter] < 0 && verbose) {
 134                         printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno));
 135                 }
 136         }
 137 }
 138
 139 /*
 140  * Does the counter have nsecs as a unit?
 141  */
 142 static inline int nsec_counter(int counter)
 143 {
 144         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 145                 return 0;
 146
 147         if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
 148                 return 1;
 149
 150         if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 151                 return 1;
 152
 153         return 0;
 154 }
 155
 156 /*
 157  * Read out the results of a single counter:
 158  */
 159 static void read_counter(int counter)
 160 {
 161         __u64 *count, single_count[3];
 162         ssize_t res;
 163         int cpu, nv;
 164         int scaled;
 165
 166         count = event_res[run_idx][counter];
 167
 168         count[0] = count[1] = count[2] = 0;
 169
 170         nv = scale ? 3 : 1;
 171         for (cpu = 0; cpu < nr_cpus; cpu ++) {
 172                 if (fd[cpu][counter] < 0)
 173                         continue;
 174
 175                 res = read(fd[cpu][counter], single_count, nv * sizeof(__u64));
 176                 assert(res == nv * sizeof(__u64));
 177                 close(fd[cpu][counter]);
 178                 fd[cpu][counter] = -1;
 179
 180                 count[0] += single_count[0];
 181                 if (scale) {
 182                         count[1] += single_count[1];
 183                         count[2] += single_count[2];
 184                 }
 185         }
 186
 187         scaled = 0;
 188         if (scale) {
 189                 if (count[2] == 0) {
 190                         event_scaled[run_idx][counter] = -1;
 191                         count[0] = 0;
 192                         return;
 193                 }
 194
 195                 if (count[2] < count[1]) {
 196                         event_scaled[run_idx][counter] = 1;
 197                         count[0] = (unsigned long long)
 198                                 ((double)count[0] * count[1] / count[2] + 0.5);
 199                 }
 200         }
 201         /*
 202          * Save the full runtime - to allow normalization during printout:
 203          */
 204         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 205                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 206                 runtime_nsecs[run_idx] = count[0];
 207         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 208                 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
 209                 runtime_cycles[run_idx] = count[0];
 210 }
 211
 212 static int run_perf_stat(int argc, const char **argv)
 213 {
 214         unsigned long long t0, t1;
 215         int status = 0;
 216         int counter;
 217         int pid;
 218
 219         if (!system_wide)
 220                 nr_cpus = 1;
 221
 222         for (counter = 0; counter < nr_counters; counter++)
 223                 create_perf_stat_counter(counter);
 224
 225         /*
 226          * Enable counters and exec the command:
 227          */
 228         t0 = rdclock();
 229         prctl(PR_TASK_PERF_COUNTERS_ENABLE);
 230
 231         if ((pid = fork()) < 0)
 232                 perror("failed to fork");
 233
 234         if (!pid) {
 235                 if (execvp(argv[0], (char **)argv)) {
 236                         perror(argv[0]);
 237                         exit(-1);
 238                 }
 239         }
 240
 241         wait(&status);
 242
 243         prctl(PR_TASK_PERF_COUNTERS_DISABLE);
 244         t1 = rdclock();
 245
 246         walltime_nsecs[run_idx] = t1 - t0;
 247
 248         for (counter = 0; counter < nr_counters; counter++)
 249                 read_counter(counter);
 250
 251         return WEXITSTATUS(status);
 252 }
 253
 254 static void print_noise(__u64 *count, __u64 *noise)
 255 {
 256         if (run_count > 1)
 257                 fprintf(stderr, "   ( +- %7.3f%% )",
 258                         (double)noise[0]/(count[0]+1)*100.0);
 259 }
 260
 261 static void nsec_printout(int counter, __u64 *count, __u64 *noise)
 262 {
 263         double msecs = (double)count[0] / 1000000;
 264
 265         fprintf(stderr, " %14.6f  %-20s", msecs, event_name(counter));
 266
 267         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 268                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
 269
 270                 if (walltime_nsecs_avg)
 271                         fprintf(stderr, " # %10.3f CPUs ",
 272                                 (double)count[0] / (double)walltime_nsecs_avg);
 273         }
 274         print_noise(count, noise);
 275 }
 276
 277 static void abs_printout(int counter, __u64 *count, __u64 *noise)
 278 {
 279         fprintf(stderr, " %14Ld  %-20s", count[0], event_name(counter));
 280
 281         if (runtime_cycles_avg &&
 282                 attrs[counter].type == PERF_TYPE_HARDWARE &&
 283                         attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
 284
 285                 fprintf(stderr, " # %10.3f IPC  ",
 286                         (double)count[0] / (double)runtime_cycles_avg);
 287         } else {
 288                 if (runtime_nsecs_avg) {
 289                         fprintf(stderr, " # %10.3f M/sec",
 290                                 (double)count[0]/runtime_nsecs_avg*1000.0);
 291                 }
 292         }
 293         print_noise(count, noise);
 294 }
 295
 296 /*
 297  * Print out the results of a single counter:
 298  */
 299 static void print_counter(int counter)
 300 {
 301         __u64 *count, *noise;
 302         int scaled;
 303
 304         count = event_res_avg[counter];
 305         noise = event_res_noise[counter];
 306         scaled = event_scaled_avg[counter];
 307
 308         if (scaled == -1) {
 309                 fprintf(stderr, " %14s  %-20s\n",
 310                         "<not counted>", event_name(counter));
 311                 return;
 312         }
 313
 314         if (nsec_counter(counter))
 315                 nsec_printout(counter, count, noise);
 316         else
 317                 abs_printout(counter, count, noise);
 318
 319         if (scaled)
 320                 fprintf(stderr, "  (scaled from %.2f%%)",
 321                         (double) count[2] / count[1] * 100);
 322
 323         fprintf(stderr, "\n");
 324 }
 325
 326 /*
 327  * normalize_noise noise values down to stddev:
 328  */
 329 static void normalize_noise(__u64 *val)
 330 {
 331         double res;
 332
 333         res = (double)*val / (run_count * sqrt((double)run_count));
 334
 335         *val = (__u64)res;
 336 }
 337
 338 static void update_avg(const char *name, int idx, __u64 *avg, __u64 *val)
 339 {
 340         *avg += *val;
 341
 342         if (verbose > 1)
 343                 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
 344 }
 345 /*
 346  * Calculate the averages and noises:
 347  */
 348 static void calc_avg(void)
 349 {
 350         int i, j;
 351
 352         if (verbose > 1)
 353                 fprintf(stderr, "\n");
 354
 355         for (i = 0; i < run_count; i++) {
 356                 update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
 357                 update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
 358                 update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
 359
 360                 for (j = 0; j < nr_counters; j++) {
 361                         update_avg("counter/0", j,
 362                                 event_res_avg[j]+0, event_res[i][j]+0);
 363                         update_avg("counter/1", j,
 364                                 event_res_avg[j]+1, event_res[i][j]+1);
 365                         update_avg("counter/2", j,
 366                                 event_res_avg[j]+2, event_res[i][j]+2);
 367                         update_avg("scaled", j,
 368                                 event_scaled_avg + j, event_scaled[i]+j);
 369                 }
 370         }
 371         runtime_nsecs_avg /= run_count;
 372         walltime_nsecs_avg /= run_count;
 373         runtime_cycles_avg /= run_count;
 374
 375         for (j = 0; j < nr_counters; j++) {
 376                 event_res_avg[j][0] /= run_count;
 377                 event_res_avg[j][1] /= run_count;
 378                 event_res_avg[j][2] /= run_count;
 379         }
 380
 381         for (i = 0; i < run_count; i++) {
 382                 runtime_nsecs_noise +=
 383                         abs((__s64)(runtime_nsecs[i] - runtime_nsecs_avg));
 384                 walltime_nsecs_noise +=
 385                         abs((__s64)(walltime_nsecs[i] - walltime_nsecs_avg));
 386                 runtime_cycles_noise +=
 387                         abs((__s64)(runtime_cycles[i] - runtime_cycles_avg));
 388
 389                 for (j = 0; j < nr_counters; j++) {
 390                         event_res_noise[j][0] +=
 391                                 abs((__s64)(event_res[i][j][0] - event_res_avg[j][0]));
 392                         event_res_noise[j][1] +=
 393                                 abs((__s64)(event_res[i][j][1] - event_res_avg[j][1]));
 394                         event_res_noise[j][2] +=
 395                                 abs((__s64)(event_res[i][j][2] - event_res_avg[j][2]));
 396                 }
 397         }
 398
 399         normalize_noise(&runtime_nsecs_noise);
 400         normalize_noise(&walltime_nsecs_noise);
 401         normalize_noise(&runtime_cycles_noise);
 402
 403         for (j = 0; j < nr_counters; j++) {
 404                 normalize_noise(&event_res_noise[j][0]);
 405                 normalize_noise(&event_res_noise[j][1]);
 406                 normalize_noise(&event_res_noise[j][2]);
 407         }
 408 }
 409
 410 static void print_stat(int argc, const char **argv)
 411 {
 412         int i, counter;
 413
 414         calc_avg();
 415
 416         fflush(stdout);
 417
 418         fprintf(stderr, "\n");
 419         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 420
 421         for (i = 1; i < argc; i++)
 422                 fprintf(stderr, " %s", argv[i]);
 423
 424         fprintf(stderr, "\'");
 425         if (run_count > 1)
 426                 fprintf(stderr, " (%d runs)", run_count);
 427         fprintf(stderr, ":\n\n");
 428
 429         for (counter = 0; counter < nr_counters; counter++)
 430                 print_counter(counter);
 431
 432
 433         fprintf(stderr, "\n");
 434         fprintf(stderr, " %14.9f  seconds time elapsed.\n",
 435                         (double)walltime_nsecs_avg/1e9);
 436         fprintf(stderr, "\n");
 437 }
 438
 439 static volatile int signr = -1;
 440
 441 static void skip_signal(int signo)
 442 {
 443         signr = signo;
 444 }
 445
 446 static void sig_atexit(void)
 447 {
 448         if (signr == -1)
 449                 return;
 450
 451         signal(signr, SIG_DFL);
 452         kill(getpid(), signr);
 453 }
 454
 455 static const char * const stat_usage[] = {
 456         "perf stat [<options>] <command>",
 457         NULL
 458 };
 459
 460 static const struct option options[] = {
 461         OPT_CALLBACK('e', "event", NULL, "event",
 462                      "event selector. use 'perf list' to list available events",
 463                      parse_events),
 464         OPT_BOOLEAN('i', "inherit", &inherit,
 465                     "child tasks inherit counters"),
 466         OPT_INTEGER('p', "pid", &target_pid,
 467                     "stat events on existing pid"),
 468         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 469                             "system-wide collection from all CPUs"),
 470         OPT_BOOLEAN('S', "scale", &scale,
 471                             "scale/normalize counters"),
 472         OPT_BOOLEAN('v', "verbose", &verbose,
 473                     "be more verbose (show counter open errors, etc)"),
 474         OPT_INTEGER('r', "repeat", &run_count,
 475                     "repeat command and print average + stddev (max: 100)"),
 476         OPT_END()
 477 };
 478
 479 int cmd_stat(int argc, const char **argv, const char *prefix)
 480 {
 481         int status;
 482
 483         page_size = sysconf(_SC_PAGE_SIZE);
 484
 485         memcpy(attrs, default_attrs, sizeof(attrs));
 486
 487         argc = parse_options(argc, argv, options, stat_usage, 0);
 488         if (!argc)
 489                 usage_with_options(stat_usage, options);
 490         if (run_count <= 0 || run_count > MAX_RUN)
 491                 usage_with_options(stat_usage, options);
 492
 493         if (!nr_counters)
 494                 nr_counters = 8;
 495
 496         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 497         assert(nr_cpus <= MAX_NR_CPUS);
 498         assert(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 }