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[] = {
  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 unsigned 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 struct stats
  86 {
  87         double sum;
  88         double sum_sq;
  89 };
  90
  91 static double avg_stats(struct stats *stats)
  92 {
  93         return stats->sum / run_count;
  94 }
  95
  96 /*
  97  * stddev = sqrt(1/N (\Sum n_i^2) - avg(n)^2)
  98  */
  99 static double stddev_stats(struct stats *stats)
 100 {
 101         double avg = stats->sum / run_count;
 102
 103         return sqrt(stats->sum_sq/run_count - avg*avg);
 104 }
 105
 106 struct stats                    event_res_stats[MAX_COUNTERS][3];
 107 struct stats                    event_scaled_stats[MAX_COUNTERS];
 108 struct stats                    runtime_nsecs_stats;
 109 struct stats                    walltime_nsecs_stats;
 110 struct stats                    runtime_cycles_stats;
 111
 112 #define MATCH_EVENT(t, c, counter)                      \
 113         (attrs[counter].type == PERF_TYPE_##t &&        \
 114          attrs[counter].config == PERF_COUNT_##c)
 115
 116 #define ERR_PERF_OPEN \
 117 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
 118
 119 static void create_perf_stat_counter(int counter, int pid)
 120 {
 121         struct perf_counter_attr *attr = attrs + counter;
 122
 123         if (scale)
 124                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 125                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 126
 127         if (system_wide) {
 128                 unsigned int cpu;
 129
 130                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 131                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 132                         if (fd[cpu][counter] < 0 && verbose)
 133                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 134                                         fd[cpu][counter], strerror(errno));
 135                 }
 136         } else {
 137                 attr->inherit        = inherit;
 138                 attr->disabled       = 1;
 139                 attr->enable_on_exec = 1;
 140
 141                 fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
 142                 if (fd[0][counter] < 0 && verbose)
 143                         fprintf(stderr, ERR_PERF_OPEN, counter,
 144                                 fd[0][counter], strerror(errno));
 145         }
 146 }
 147
 148 /*
 149  * Does the counter have nsecs as a unit?
 150  */
 151 static inline int nsec_counter(int counter)
 152 {
 153         if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
 154             MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 155                 return 1;
 156
 157         return 0;
 158 }
 159
 160 /*
 161  * Read out the results of a single counter:
 162  */
 163 static void read_counter(int counter)
 164 {
 165         u64 *count, single_count[3];
 166         unsigned int cpu;
 167         size_t res, nv;
 168         int scaled;
 169
 170         count = event_res[run_idx][counter];
 171
 172         count[0] = count[1] = count[2] = 0;
 173
 174         nv = scale ? 3 : 1;
 175         for (cpu = 0; cpu < nr_cpus; cpu++) {
 176                 if (fd[cpu][counter] < 0)
 177                         continue;
 178
 179                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 180                 assert(res == nv * sizeof(u64));
 181
 182                 close(fd[cpu][counter]);
 183                 fd[cpu][counter] = -1;
 184
 185                 count[0] += single_count[0];
 186                 if (scale) {
 187                         count[1] += single_count[1];
 188                         count[2] += single_count[2];
 189                 }
 190         }
 191
 192         scaled = 0;
 193         if (scale) {
 194                 if (count[2] == 0) {
 195                         event_scaled[run_idx][counter] = -1;
 196                         count[0] = 0;
 197                         return;
 198                 }
 199
 200                 if (count[2] < count[1]) {
 201                         event_scaled[run_idx][counter] = 1;
 202                         count[0] = (unsigned long long)
 203                                 ((double)count[0] * count[1] / count[2] + 0.5);
 204                 }
 205         }
 206         /*
 207          * Save the full runtime - to allow normalization during printout:
 208          */
 209         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
 210                 runtime_nsecs[run_idx] = count[0];
 211         if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
 212                 runtime_cycles[run_idx] = count[0];
 213 }
 214
 215 static int run_perf_stat(int argc __used, const char **argv)
 216 {
 217         unsigned long long t0, t1;
 218         int status = 0;
 219         int counter;
 220         int pid;
 221         int child_ready_pipe[2], go_pipe[2];
 222         char buf;
 223
 224         if (!system_wide)
 225                 nr_cpus = 1;
 226
 227         if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
 228                 perror("failed to create pipes");
 229                 exit(1);
 230         }
 231
 232         if ((pid = fork()) < 0)
 233                 perror("failed to fork");
 234
 235         if (!pid) {
 236                 close(child_ready_pipe[0]);
 237                 close(go_pipe[1]);
 238                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 239
 240                 /*
 241                  * Do a dummy execvp to get the PLT entry resolved,
 242                  * so we avoid the resolver overhead on the real
 243                  * execvp call.
 244                  */
 245                 execvp("", (char **)argv);
 246
 247                 /*
 248                  * Tell the parent we're ready to go
 249                  */
 250                 close(child_ready_pipe[1]);
 251
 252                 /*
 253                  * Wait until the parent tells us to go.
 254                  */
 255                 if (read(go_pipe[0], &buf, 1) == -1)
 256                         perror("unable to read pipe");
 257
 258                 execvp(argv[0], (char **)argv);
 259
 260                 perror(argv[0]);
 261                 exit(-1);
 262         }
 263
 264         /*
 265          * Wait for the child to be ready to exec.
 266          */
 267         close(child_ready_pipe[1]);
 268         close(go_pipe[0]);
 269         if (read(child_ready_pipe[0], &buf, 1) == -1)
 270                 perror("unable to read pipe");
 271         close(child_ready_pipe[0]);
 272
 273         for (counter = 0; counter < nr_counters; counter++)
 274                 create_perf_stat_counter(counter, pid);
 275
 276         /*
 277          * Enable counters and exec the command:
 278          */
 279         t0 = rdclock();
 280
 281         close(go_pipe[1]);
 282         wait(&status);
 283
 284         t1 = rdclock();
 285
 286         walltime_nsecs[run_idx] = t1 - t0;
 287
 288         for (counter = 0; counter < nr_counters; counter++)
 289                 read_counter(counter);
 290
 291         return WEXITSTATUS(status);
 292 }
 293
 294 static void print_noise(double avg, double stddev)
 295 {
 296         if (run_count > 1)
 297                 fprintf(stderr, "   ( +- %7.3f%% )", 100*stddev / avg);
 298 }
 299
 300 static void nsec_printout(int counter, double avg, double stddev)
 301 {
 302         double msecs = avg / 1e6;
 303
 304         fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
 305
 306         if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
 307                 fprintf(stderr, " # %10.3f CPUs ",
 308                                 avg / avg_stats(&walltime_nsecs_stats));
 309         }
 310         print_noise(avg, stddev);
 311 }
 312
 313 static void abs_printout(int counter, double avg, double stddev)
 314 {
 315         fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));
 316
 317         if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
 318                 fprintf(stderr, " # %10.3f IPC  ",
 319                                 avg / avg_stats(&runtime_cycles_stats));
 320         } else {
 321                 fprintf(stderr, " # %10.3f M/sec",
 322                                 1000.0 * avg / avg_stats(&runtime_nsecs_stats));
 323         }
 324         print_noise(avg, stddev);
 325 }
 326
 327 /*
 328  * Print out the results of a single counter:
 329  */
 330 static void print_counter(int counter)
 331 {
 332         double avg, stddev;
 333         int scaled;
 334
 335         avg    = avg_stats(&event_res_stats[counter][0]);
 336         stddev = stddev_stats(&event_res_stats[counter][0]);
 337         scaled = avg_stats(&event_scaled_stats[counter]);
 338
 339         if (scaled == -1) {
 340                 fprintf(stderr, " %14s  %-24s\n",
 341                         "<not counted>", event_name(counter));
 342                 return;
 343         }
 344
 345         if (nsec_counter(counter))
 346                 nsec_printout(counter, avg, stddev);
 347         else
 348                 abs_printout(counter, avg, stddev);
 349
 350         if (scaled) {
 351                 double avg_enabled, avg_running;
 352
 353                 avg_enabled = avg_stats(&event_res_stats[counter][1]);
 354                 avg_running = avg_stats(&event_res_stats[counter][2]);
 355
 356                 fprintf(stderr, "  (scaled from %.2f%%)",
 357                                 100 * avg_running / avg_enabled);
 358         }
 359
 360         fprintf(stderr, "\n");
 361 }
 362
 363 static void update_stats(const char *name, int idx, struct stats *stats, u64 *val)
 364 {
 365         double sq = *val;
 366
 367         stats->sum += *val;
 368         stats->sum_sq += sq * sq;
 369
 370         if (verbose > 1)
 371                 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
 372 }
 373
 374 /*
 375  * Calculate the averages and noises:
 376  */
 377 static void calc_avg(void)
 378 {
 379         int i, j;
 380
 381         if (verbose > 1)
 382                 fprintf(stderr, "\n");
 383
 384         for (i = 0; i < run_count; i++) {
 385                 update_stats("runtime", 0, &runtime_nsecs_stats, runtime_nsecs + i);
 386                 update_stats("walltime", 0, &walltime_nsecs_stats, walltime_nsecs + i);
 387                 update_stats("runtime_cycles", 0, &runtime_cycles_stats, runtime_cycles + i);
 388
 389                 for (j = 0; j < nr_counters; j++) {
 390                         update_stats("counter/0", j,
 391                                 event_res_stats[j]+0, event_res[i][j]+0);
 392                         update_stats("counter/1", j,
 393                                 event_res_stats[j]+1, event_res[i][j]+1);
 394                         update_stats("counter/2", j,
 395                                 event_res_stats[j]+2, event_res[i][j]+2);
 396                         if (event_scaled[i][j] != (u64)-1)
 397                                 update_stats("scaled", j,
 398                                         event_scaled_stats + j, event_scaled[i]+j);
 399                 }
 400         }
 401 }
 402
 403 static void print_stat(int argc, const char **argv)
 404 {
 405         int i, counter;
 406
 407         calc_avg();
 408
 409         fflush(stdout);
 410
 411         fprintf(stderr, "\n");
 412         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 413
 414         for (i = 1; i < argc; i++)
 415                 fprintf(stderr, " %s", argv[i]);
 416
 417         fprintf(stderr, "\'");
 418         if (run_count > 1)
 419                 fprintf(stderr, " (%d runs)", run_count);
 420         fprintf(stderr, ":\n\n");
 421
 422         for (counter = 0; counter < nr_counters; counter++)
 423                 print_counter(counter);
 424
 425         fprintf(stderr, "\n");
 426         fprintf(stderr, " %14.9f  seconds time elapsed",
 427                         avg_stats(&walltime_nsecs_stats)/1e9);
 428         if (run_count > 1) {
 429                 fprintf(stderr, "   ( +- %7.3f%% )",
 430                                 100*stddev_stats(&walltime_nsecs_stats) /
 431                                 avg_stats(&walltime_nsecs_stats));
 432         }
 433         fprintf(stderr, "\n\n");
 434 }
 435
 436 static volatile int signr = -1;
 437
 438 static void skip_signal(int signo)
 439 {
 440         signr = signo;
 441 }
 442
 443 static void sig_atexit(void)
 444 {
 445         if (signr == -1)
 446                 return;
 447
 448         signal(signr, SIG_DFL);
 449         kill(getpid(), signr);
 450 }
 451
 452 static const char * const stat_usage[] = {
 453         "perf stat [<options>] <command>",
 454         NULL
 455 };
 456
 457 static const struct option options[] = {
 458         OPT_CALLBACK('e', "event", NULL, "event",
 459                      "event selector. use 'perf list' to list available events",
 460                      parse_events),
 461         OPT_BOOLEAN('i', "inherit", &inherit,
 462                     "child tasks inherit counters"),
 463         OPT_INTEGER('p', "pid", &target_pid,
 464                     "stat events on existing pid"),
 465         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 466                     "system-wide collection from all CPUs"),
 467         OPT_BOOLEAN('c', "scale", &scale,
 468                     "scale/normalize counters"),
 469         OPT_BOOLEAN('v', "verbose", &verbose,
 470                     "be more verbose (show counter open errors, etc)"),
 471         OPT_INTEGER('r', "repeat", &run_count,
 472                     "repeat command and print average + stddev (max: 100)"),
 473         OPT_BOOLEAN('n', "null", &null_run,
 474                     "null run - dont start any counters"),
 475         OPT_END()
 476 };
 477
 478 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 479 {
 480         int status;
 481
 482         argc = parse_options(argc, argv, options, stat_usage,
 483                 PARSE_OPT_STOP_AT_NON_OPTION);
 484         if (!argc)
 485                 usage_with_options(stat_usage, options);
 486         if (run_count <= 0 || run_count > MAX_RUN)
 487                 usage_with_options(stat_usage, options);
 488
 489         /* Set attrs and nr_counters if no event is selected and !null_run */
 490         if (!null_run && !nr_counters) {
 491                 memcpy(attrs, default_attrs, sizeof(default_attrs));
 492                 nr_counters = ARRAY_SIZE(default_attrs);
 493         }
 494
 495         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 496         assert(nr_cpus <= MAX_NR_CPUS);
 497         assert((int)nr_cpus >= 0);
 498
 499         /*
 500          * We dont want to block the signals - that would cause
 501          * child tasks to inherit that and Ctrl-C would not work.
 502          * What we want is for Ctrl-C to work in the exec()-ed
 503          * task, but being ignored by perf stat itself:
 504          */
 505         atexit(sig_atexit);
 506         signal(SIGINT,  skip_signal);
 507         signal(SIGALRM, skip_signal);
 508         signal(SIGABRT, skip_signal);
 509
 510         status = 0;
 511         for (run_idx = 0; run_idx < run_count; run_idx++) {
 512                 if (run_count != 1 && verbose)
 513                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 514                 status = run_perf_stat(argc, argv);
 515         }
 516
 517         print_stat(argc, argv);
 518
 519         return status;
 520 }