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