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  *   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 #include "util/event.h"
  46 #include "util/debug.h"
  47
  48 #include <sys/prctl.h>
  49 #include <math.h>
  50
  51 static struct perf_event_attr default_attrs[] = {
  52
  53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
  54   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
  55   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS  },
  56   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
  57
  58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
  59   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS    },
  60   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
  61   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES    },
  62
  63 };
  64
  65 static int                      system_wide                     =  0;
  66 static unsigned 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 pid_t                    target_pid                      = -1;
  73 static pid_t                    child_pid                       = -1;
  74 static int                      null_run                        =  0;
  75
  76 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  77
  78 static int                      event_scaled[MAX_COUNTERS];
  79
  80 struct stats
  81 {
  82         double n, mean, M2;
  83 };
  84
  85 static void update_stats(struct stats *stats, u64 val)
  86 {
  87         double delta;
  88
  89         stats->n++;
  90         delta = val - stats->mean;
  91         stats->mean += delta / stats->n;
  92         stats->M2 += delta*(val - stats->mean);
  93 }
  94
  95 static double avg_stats(struct stats *stats)
  96 {
  97         return stats->mean;
  98 }
  99
 100 /*
 101  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 102  *
 103  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 104  * s^2 = -------------------------------
 105  *                  n - 1
 106  *
 107  * http://en.wikipedia.org/wiki/Stddev
 108  *
 109  * The std dev of the mean is related to the std dev by:
 110  *
 111  *             s
 112  * s_mean = -------
 113  *          sqrt(n)
 114  *
 115  */
 116 static double stddev_stats(struct stats *stats)
 117 {
 118         double variance = stats->M2 / (stats->n - 1);
 119         double variance_mean = variance / stats->n;
 120
 121         return sqrt(variance_mean);
 122 }
 123
 124 struct stats                    event_res_stats[MAX_COUNTERS][3];
 125 struct stats                    runtime_nsecs_stats;
 126 struct stats                    walltime_nsecs_stats;
 127 struct stats                    runtime_cycles_stats;
 128
 129 #define MATCH_EVENT(t, c, counter)                      \
 130         (attrs[counter].type == PERF_TYPE_##t &&        \
 131          attrs[counter].config == PERF_COUNT_##c)
 132
 133 #define ERR_PERF_OPEN \
 134 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
 135
 136 static void create_perf_stat_counter(int counter, int pid)
 137 {
 138         struct perf_event_attr *attr = attrs + counter;
 139
 140         if (scale)
 141                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 142                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 143
 144         if (system_wide) {
 145                 unsigned int cpu;
 146
 147                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 148                         fd[cpu][counter] = sys_perf_event_open(attr, -1, cpu, -1, 0);
 149                         if (fd[cpu][counter] < 0 && verbose)
 150                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 151                                         fd[cpu][counter], strerror(errno));
 152                 }
 153         } else {
 154                 attr->inherit        = inherit;
 155                 attr->disabled       = 1;
 156                 attr->enable_on_exec = 1;
 157
 158                 fd[0][counter] = sys_perf_event_open(attr, pid, -1, -1, 0);
 159                 if (fd[0][counter] < 0 && verbose)
 160                         fprintf(stderr, ERR_PERF_OPEN, counter,
 161                                 fd[0][counter], strerror(errno));
 162         }
 163 }
 164
 165 /*
 166  * Does the counter have nsecs as a unit?
 167  */
 168 static inline int nsec_counter(int counter)
 169 {
 170         if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
 171             MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 172                 return 1;
 173
 174         return 0;
 175 }
 176
 177 /*
 178  * Read out the results of a single counter:
 179  */
 180 static void read_counter(int counter)
 181 {
 182         u64 count[3], single_count[3];
 183         unsigned int cpu;
 184         size_t res, nv;
 185         int scaled;
 186         int i;
 187
 188         count[0] = count[1] = count[2] = 0;
 189
 190         nv = scale ? 3 : 1;
 191         for (cpu = 0; cpu < nr_cpus; cpu++) {
 192                 if (fd[cpu][counter] < 0)
 193                         continue;
 194
 195                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 196                 assert(res == nv * sizeof(u64));
 197
 198                 close(fd[cpu][counter]);
 199                 fd[cpu][counter] = -1;
 200
 201                 count[0] += single_count[0];
 202                 if (scale) {
 203                         count[1] += single_count[1];
 204                         count[2] += single_count[2];
 205                 }
 206         }
 207
 208         scaled = 0;
 209         if (scale) {
 210                 if (count[2] == 0) {
 211                         event_scaled[counter] = -1;
 212                         count[0] = 0;
 213                         return;
 214                 }
 215
 216                 if (count[2] < count[1]) {
 217                         event_scaled[counter] = 1;
 218                         count[0] = (unsigned long long)
 219                                 ((double)count[0] * count[1] / count[2] + 0.5);
 220                 }
 221         }
 222
 223         for (i = 0; i < 3; i++)
 224                 update_stats(&event_res_stats[counter][i], count[i]);
 225
 226         if (verbose) {
 227                 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
 228                                 count[0], count[1], count[2]);
 229         }
 230
 231         /*
 232          * Save the full runtime - to allow normalization during printout:
 233          */
 234         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 235                 update_stats(&runtime_nsecs_stats, count[0]);
 236         if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 237                 update_stats(&runtime_cycles_stats, count[0]);
 238 }
 239
 240 static int run_perf_stat(int argc __used, const char **argv)
 241 {
 242         unsigned long long t0, t1;
 243         int status = 0;
 244         int counter;
 245         int pid;
 246         int child_ready_pipe[2], go_pipe[2];
 247         char buf;
 248
 249         if (!system_wide)
 250                 nr_cpus = 1;
 251
 252         if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
 253                 perror("failed to create pipes");
 254                 exit(1);
 255         }
 256
 257         if ((pid = fork()) < 0)
 258                 perror("failed to fork");
 259
 260         if (!pid) {
 261                 close(child_ready_pipe[0]);
 262                 close(go_pipe[1]);
 263                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 264
 265                 /*
 266                  * Do a dummy execvp to get the PLT entry resolved,
 267                  * so we avoid the resolver overhead on the real
 268                  * execvp call.
 269                  */
 270                 execvp("", (char **)argv);
 271
 272                 /*
 273                  * Tell the parent we're ready to go
 274                  */
 275                 close(child_ready_pipe[1]);
 276
 277                 /*
 278                  * Wait until the parent tells us to go.
 279                  */
 280                 if (read(go_pipe[0], &buf, 1) == -1)
 281                         perror("unable to read pipe");
 282
 283                 execvp(argv[0], (char **)argv);
 284
 285                 perror(argv[0]);
 286                 exit(-1);
 287         }
 288
 289         child_pid = pid;
 290
 291         /*
 292          * Wait for the child to be ready to exec.
 293          */
 294         close(child_ready_pipe[1]);
 295         close(go_pipe[0]);
 296         if (read(child_ready_pipe[0], &buf, 1) == -1)
 297                 perror("unable to read pipe");
 298         close(child_ready_pipe[0]);
 299
 300         for (counter = 0; counter < nr_counters; counter++)
 301                 create_perf_stat_counter(counter, pid);
 302
 303         /*
 304          * Enable counters and exec the command:
 305          */
 306         t0 = rdclock();
 307
 308         close(go_pipe[1]);
 309         wait(&status);
 310
 311         t1 = rdclock();
 312
 313         update_stats(&walltime_nsecs_stats, t1 - t0);
 314
 315         for (counter = 0; counter < nr_counters; counter++)
 316                 read_counter(counter);
 317
 318         return WEXITSTATUS(status);
 319 }
 320
 321 static void print_noise(int counter, double avg)
 322 {
 323         if (run_count == 1)
 324                 return;
 325
 326         fprintf(stderr, "   ( +- %7.3f%% )",
 327                         100 * stddev_stats(&event_res_stats[counter][0]) / avg);
 328 }
 329
 330 static void nsec_printout(int counter, double avg)
 331 {
 332         double msecs = avg / 1e6;
 333
 334         fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
 335
 336         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
 337                 fprintf(stderr, " # %10.3f CPUs ",
 338                                 avg / avg_stats(&walltime_nsecs_stats));
 339         }
 340 }
 341
 342 static void abs_printout(int counter, double avg)
 343 {
 344         double total, ratio = 0.0;
 345
 346         fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));
 347
 348         if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
 349                 total = avg_stats(&runtime_cycles_stats);
 350
 351                 if (total)
 352                         ratio = avg / total;
 353
 354                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 355         } else {
 356                 total = avg_stats(&runtime_nsecs_stats);
 357
 358                 if (total)
 359                         ratio = 1000.0 * avg / total;
 360
 361                 fprintf(stderr, " # %10.3f M/sec", ratio);
 362         }
 363 }
 364
 365 /*
 366  * Print out the results of a single counter:
 367  */
 368 static void print_counter(int counter)
 369 {
 370         double avg = avg_stats(&event_res_stats[counter][0]);
 371         int scaled = event_scaled[counter];
 372
 373         if (scaled == -1) {
 374                 fprintf(stderr, " %14s  %-24s\n",
 375                         "<not counted>", event_name(counter));
 376                 return;
 377         }
 378
 379         if (nsec_counter(counter))
 380                 nsec_printout(counter, avg);
 381         else
 382                 abs_printout(counter, avg);
 383
 384         print_noise(counter, avg);
 385
 386         if (scaled) {
 387                 double avg_enabled, avg_running;
 388
 389                 avg_enabled = avg_stats(&event_res_stats[counter][1]);
 390                 avg_running = avg_stats(&event_res_stats[counter][2]);
 391
 392                 fprintf(stderr, "  (scaled from %.2f%%)",
 393                                 100 * avg_running / avg_enabled);
 394         }
 395
 396         fprintf(stderr, "\n");
 397 }
 398
 399 static void print_stat(int argc, const char **argv)
 400 {
 401         int i, counter;
 402
 403         fflush(stdout);
 404
 405         fprintf(stderr, "\n");
 406         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 407
 408         for (i = 1; i < argc; i++)
 409                 fprintf(stderr, " %s", argv[i]);
 410
 411         fprintf(stderr, "\'");
 412         if (run_count > 1)
 413                 fprintf(stderr, " (%d runs)", run_count);
 414         fprintf(stderr, ":\n\n");
 415
 416         for (counter = 0; counter < nr_counters; counter++)
 417                 print_counter(counter);
 418
 419         fprintf(stderr, "\n");
 420         fprintf(stderr, " %14.9f  seconds time elapsed",
 421                         avg_stats(&walltime_nsecs_stats)/1e9);
 422         if (run_count > 1) {
 423                 fprintf(stderr, "   ( +- %7.3f%% )",
 424                                 100*stddev_stats(&walltime_nsecs_stats) /
 425                                 avg_stats(&walltime_nsecs_stats));
 426         }
 427         fprintf(stderr, "\n\n");
 428 }
 429
 430 static volatile int signr = -1;
 431
 432 static void skip_signal(int signo)
 433 {
 434         signr = signo;
 435 }
 436
 437 static void sig_atexit(void)
 438 {
 439         if (child_pid != -1)
 440                 kill(child_pid, SIGTERM);
 441
 442         if (signr == -1)
 443                 return;
 444
 445         signal(signr, SIG_DFL);
 446         kill(getpid(), signr);
 447 }
 448
 449 static const char * const stat_usage[] = {
 450         "perf stat [<options>] <command>",
 451         NULL
 452 };
 453
 454 static const struct option options[] = {
 455         OPT_CALLBACK('e', "event", NULL, "event",
 456                      "event selector. use 'perf list' to list available events",
 457                      parse_events),
 458         OPT_BOOLEAN('i', "inherit", &inherit,
 459                     "child tasks inherit counters"),
 460         OPT_INTEGER('p', "pid", &target_pid,
 461                     "stat events on existing pid"),
 462         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 463                     "system-wide collection from all CPUs"),
 464         OPT_BOOLEAN('c', "scale", &scale,
 465                     "scale/normalize counters"),
 466         OPT_BOOLEAN('v', "verbose", &verbose,
 467                     "be more verbose (show counter open errors, etc)"),
 468         OPT_INTEGER('r', "repeat", &run_count,
 469                     "repeat command and print average + stddev (max: 100)"),
 470         OPT_BOOLEAN('n', "null", &null_run,
 471                     "null run - dont start any counters"),
 472         OPT_END()
 473 };
 474
 475 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 476 {
 477         int status;
 478
 479         argc = parse_options(argc, argv, options, stat_usage,
 480                 PARSE_OPT_STOP_AT_NON_OPTION);
 481         if (!argc)
 482                 usage_with_options(stat_usage, options);
 483         if (run_count <= 0)
 484                 usage_with_options(stat_usage, options);
 485
 486         /* Set attrs and nr_counters if no event is selected and !null_run */
 487         if (!null_run && !nr_counters) {
 488                 memcpy(attrs, default_attrs, sizeof(default_attrs));
 489                 nr_counters = ARRAY_SIZE(default_attrs);
 490         }
 491
 492         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 493         assert(nr_cpus <= MAX_NR_CPUS);
 494         assert((int)nr_cpus >= 0);
 495
 496         /*
 497          * We dont want to block the signals - that would cause
 498          * child tasks to inherit that and Ctrl-C would not work.
 499          * What we want is for Ctrl-C to work in the exec()-ed
 500          * task, but being ignored by perf stat itself:
 501          */
 502         atexit(sig_atexit);
 503         signal(SIGINT,  skip_signal);
 504         signal(SIGALRM, skip_signal);
 505         signal(SIGABRT, skip_signal);
 506
 507         status = 0;
 508         for (run_idx = 0; run_idx < run_count; run_idx++) {
 509                 if (run_count != 1 && verbose)
 510                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 511                 status = run_perf_stat(argc, argv);
 512         }
 513
 514         print_stat(argc, argv);
 515
 516         return status;
 517 }