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 #include "util/header.h" 48 #include "util/cpumap.h" 49 #include "util/thread.h" 50 51 #include <sys/prctl.h> 52 #include <math.h> 53 #include <locale.h> 54 55 static struct perf_event_attr default_attrs[] = { 56 57 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 58 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 59 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 60 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 61 62 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 63 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 64 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 65 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 66 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES }, 67 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES }, 68 69 }; 70 71 static bool system_wide = false; 72 static unsigned int nr_cpus = 0; 73 static int run_idx = 0; 74 75 static int run_count = 1; 76 static bool no_inherit = false; 77 static bool scale = true; 78 static pid_t target_pid = -1; 79 static pid_t target_tid = -1; 80 static pid_t *all_tids = NULL; 81 static int thread_num = 0; 82 static pid_t child_pid = -1; 83 static bool null_run = false; 84 static bool big_num = false; 85 86 87 static int *fd[MAX_NR_CPUS][MAX_COUNTERS]; 88 89 static int event_scaled[MAX_COUNTERS]; 90 91 static volatile int done = 0; 92 93 struct stats 94 { 95 double n, mean, M2; 96 }; 97 98 static void update_stats(struct stats *stats, u64 val) 99 { 100 double delta; 101 102 stats->n++; 103 delta = val - stats->mean; 104 stats->mean += delta / stats->n; 105 stats->M2 += delta*(val - stats->mean); 106 } 107 108 static double avg_stats(struct stats *stats) 109 { 110 return stats->mean; 111 } 112 113 /* 114 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 115 * 116 * (\Sum n_i^2) - ((\Sum n_i)^2)/n 117 * s^2 = ------------------------------- 118 * n - 1 119 * 120 * http://en.wikipedia.org/wiki/Stddev 121 * 122 * The std dev of the mean is related to the std dev by: 123 * 124 * s 125 * s_mean = ------- 126 * sqrt(n) 127 * 128 */ 129 static double stddev_stats(struct stats *stats) 130 { 131 double variance = stats->M2 / (stats->n - 1); 132 double variance_mean = variance / stats->n; 133 134 return sqrt(variance_mean); 135 } 136 137 struct stats event_res_stats[MAX_COUNTERS][3]; 138 struct stats runtime_nsecs_stats; 139 struct stats walltime_nsecs_stats; 140 struct stats runtime_cycles_stats; 141 struct stats runtime_branches_stats; 142 143 #define MATCH_EVENT(t, c, counter) \ 144 (attrs[counter].type == PERF_TYPE_##t && \ 145 attrs[counter].config == PERF_COUNT_##c) 146 147 #define ERR_PERF_OPEN \ 148 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n" 149 150 static int create_perf_stat_counter(int counter) 151 { 152 struct perf_event_attr *attr = attrs + counter; 153 int thread; 154 int ncreated = 0; 155 156 if (scale) 157 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 158 PERF_FORMAT_TOTAL_TIME_RUNNING; 159 160 if (system_wide) { 161 unsigned int cpu; 162 163 for (cpu = 0; cpu < nr_cpus; cpu++) { 164 fd[cpu][counter][0] = sys_perf_event_open(attr, 165 -1, cpumap[cpu], -1, 0); 166 if (fd[cpu][counter][0] < 0) 167 pr_debug(ERR_PERF_OPEN, counter, 168 fd[cpu][counter][0], strerror(errno)); 169 else 170 ++ncreated; 171 } 172 } else { 173 attr->inherit = !no_inherit; 174 if (target_pid == -1 && target_tid == -1) { 175 attr->disabled = 1; 176 attr->enable_on_exec = 1; 177 } 178 for (thread = 0; thread < thread_num; thread++) { 179 fd[0][counter][thread] = sys_perf_event_open(attr, 180 all_tids[thread], -1, -1, 0); 181 if (fd[0][counter][thread] < 0) 182 pr_debug(ERR_PERF_OPEN, counter, 183 fd[0][counter][thread], 184 strerror(errno)); 185 else 186 ++ncreated; 187 } 188 } 189 190 return ncreated; 191 } 192 193 /* 194 * Does the counter have nsecs as a unit? 195 */ 196 static inline int nsec_counter(int counter) 197 { 198 if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) || 199 MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 200 return 1; 201 202 return 0; 203 } 204 205 /* 206 * Read out the results of a single counter: 207 */ 208 static void read_counter(int counter) 209 { 210 u64 count[3], single_count[3]; 211 unsigned int cpu; 212 size_t res, nv; 213 int scaled; 214 int i, thread; 215 216 count[0] = count[1] = count[2] = 0; 217 218 nv = scale ? 3 : 1; 219 for (cpu = 0; cpu < nr_cpus; cpu++) { 220 for (thread = 0; thread < thread_num; thread++) { 221 if (fd[cpu][counter][thread] < 0) 222 continue; 223 224 res = read(fd[cpu][counter][thread], 225 single_count, nv * sizeof(u64)); 226 assert(res == nv * sizeof(u64)); 227 228 close(fd[cpu][counter][thread]); 229 fd[cpu][counter][thread] = -1; 230 231 count[0] += single_count[0]; 232 if (scale) { 233 count[1] += single_count[1]; 234 count[2] += single_count[2]; 235 } 236 } 237 } 238 239 scaled = 0; 240 if (scale) { 241 if (count[2] == 0) { 242 event_scaled[counter] = -1; 243 count[0] = 0; 244 return; 245 } 246 247 if (count[2] < count[1]) { 248 event_scaled[counter] = 1; 249 count[0] = (unsigned long long) 250 ((double)count[0] * count[1] / count[2] + 0.5); 251 } 252 } 253 254 for (i = 0; i < 3; i++) 255 update_stats(&event_res_stats[counter][i], count[i]); 256 257 if (verbose) { 258 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter), 259 count[0], count[1], count[2]); 260 } 261 262 /* 263 * Save the full runtime - to allow normalization during printout: 264 */ 265 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 266 update_stats(&runtime_nsecs_stats, count[0]); 267 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter)) 268 update_stats(&runtime_cycles_stats, count[0]); 269 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter)) 270 update_stats(&runtime_branches_stats, count[0]); 271 } 272 273 static int run_perf_stat(int argc __used, const char **argv) 274 { 275 unsigned long long t0, t1; 276 int status = 0; 277 int counter, ncreated = 0; 278 int child_ready_pipe[2], go_pipe[2]; 279 const bool forks = (argc > 0); 280 char buf; 281 282 if (!system_wide) 283 nr_cpus = 1; 284 285 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 286 perror("failed to create pipes"); 287 exit(1); 288 } 289 290 if (forks) { 291 if ((child_pid = fork()) < 0) 292 perror("failed to fork"); 293 294 if (!child_pid) { 295 close(child_ready_pipe[0]); 296 close(go_pipe[1]); 297 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 298 299 /* 300 * Do a dummy execvp to get the PLT entry resolved, 301 * so we avoid the resolver overhead on the real 302 * execvp call. 303 */ 304 execvp("", (char **)argv); 305 306 /* 307 * Tell the parent we're ready to go 308 */ 309 close(child_ready_pipe[1]); 310 311 /* 312 * Wait until the parent tells us to go. 313 */ 314 if (read(go_pipe[0], &buf, 1) == -1) 315 perror("unable to read pipe"); 316 317 execvp(argv[0], (char **)argv); 318 319 perror(argv[0]); 320 exit(-1); 321 } 322 323 if (target_tid == -1 && target_pid == -1 && !system_wide) 324 all_tids[0] = child_pid; 325 326 /* 327 * Wait for the child to be ready to exec. 328 */ 329 close(child_ready_pipe[1]); 330 close(go_pipe[0]); 331 if (read(child_ready_pipe[0], &buf, 1) == -1) 332 perror("unable to read pipe"); 333 close(child_ready_pipe[0]); 334 } 335 336 for (counter = 0; counter < nr_counters; counter++) 337 ncreated += create_perf_stat_counter(counter); 338 339 if (ncreated == 0) { 340 pr_err("No permission to collect %sstats.\n" 341 "Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n", 342 system_wide ? "system-wide " : ""); 343 if (child_pid != -1) 344 kill(child_pid, SIGTERM); 345 return -1; 346 } 347 348 /* 349 * Enable counters and exec the command: 350 */ 351 t0 = rdclock(); 352 353 if (forks) { 354 close(go_pipe[1]); 355 wait(&status); 356 } else { 357 while(!done) sleep(1); 358 } 359 360 t1 = rdclock(); 361 362 update_stats(&walltime_nsecs_stats, t1 - t0); 363 364 for (counter = 0; counter < nr_counters; counter++) 365 read_counter(counter); 366 367 return WEXITSTATUS(status); 368 } 369 370 static void print_noise(int counter, double avg) 371 { 372 if (run_count == 1) 373 return; 374 375 fprintf(stderr, " ( +- %7.3f%% )", 376 100 * stddev_stats(&event_res_stats[counter][0]) / avg); 377 } 378 379 static void nsec_printout(int counter, double avg) 380 { 381 double msecs = avg / 1e6; 382 383 fprintf(stderr, " %18.6f %-24s", msecs, event_name(counter)); 384 385 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) { 386 fprintf(stderr, " # %10.3f CPUs ", 387 avg / avg_stats(&walltime_nsecs_stats)); 388 } 389 } 390 391 static void abs_printout(int counter, double avg) 392 { 393 double total, ratio = 0.0; 394 395 if (big_num) 396 fprintf(stderr, " %'18.0f %-24s", avg, event_name(counter)); 397 else 398 fprintf(stderr, " %18.0f %-24s", avg, event_name(counter)); 399 400 if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) { 401 total = avg_stats(&runtime_cycles_stats); 402 403 if (total) 404 ratio = avg / total; 405 406 fprintf(stderr, " # %10.3f IPC ", ratio); 407 } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) && 408 runtime_branches_stats.n != 0) { 409 total = avg_stats(&runtime_branches_stats); 410 411 if (total) 412 ratio = avg * 100 / total; 413 414 fprintf(stderr, " # %10.3f %% ", ratio); 415 416 } else if (runtime_nsecs_stats.n != 0) { 417 total = avg_stats(&runtime_nsecs_stats); 418 419 if (total) 420 ratio = 1000.0 * avg / total; 421 422 fprintf(stderr, " # %10.3f M/sec", ratio); 423 } 424 } 425 426 /* 427 * Print out the results of a single counter: 428 */ 429 static void print_counter(int counter) 430 { 431 double avg = avg_stats(&event_res_stats[counter][0]); 432 int scaled = event_scaled[counter]; 433 434 if (scaled == -1) { 435 fprintf(stderr, " %18s %-24s\n", 436 "<not counted>", event_name(counter)); 437 return; 438 } 439 440 if (nsec_counter(counter)) 441 nsec_printout(counter, avg); 442 else 443 abs_printout(counter, avg); 444 445 print_noise(counter, avg); 446 447 if (scaled) { 448 double avg_enabled, avg_running; 449 450 avg_enabled = avg_stats(&event_res_stats[counter][1]); 451 avg_running = avg_stats(&event_res_stats[counter][2]); 452 453 fprintf(stderr, " (scaled from %.2f%%)", 454 100 * avg_running / avg_enabled); 455 } 456 457 fprintf(stderr, "\n"); 458 } 459 460 static void print_stat(int argc, const char **argv) 461 { 462 int i, counter; 463 464 fflush(stdout); 465 466 fprintf(stderr, "\n"); 467 fprintf(stderr, " Performance counter stats for "); 468 if(target_pid == -1 && target_tid == -1) { 469 fprintf(stderr, "\'%s", argv[0]); 470 for (i = 1; i < argc; i++) 471 fprintf(stderr, " %s", argv[i]); 472 } else if (target_pid != -1) 473 fprintf(stderr, "process id \'%d", target_pid); 474 else 475 fprintf(stderr, "thread id \'%d", target_tid); 476 477 fprintf(stderr, "\'"); 478 if (run_count > 1) 479 fprintf(stderr, " (%d runs)", run_count); 480 fprintf(stderr, ":\n\n"); 481 482 for (counter = 0; counter < nr_counters; counter++) 483 print_counter(counter); 484 485 fprintf(stderr, "\n"); 486 fprintf(stderr, " %18.9f seconds time elapsed", 487 avg_stats(&walltime_nsecs_stats)/1e9); 488 if (run_count > 1) { 489 fprintf(stderr, " ( +- %7.3f%% )", 490 100*stddev_stats(&walltime_nsecs_stats) / 491 avg_stats(&walltime_nsecs_stats)); 492 } 493 fprintf(stderr, "\n\n"); 494 } 495 496 static volatile int signr = -1; 497 498 static void skip_signal(int signo) 499 { 500 if(child_pid == -1) 501 done = 1; 502 503 signr = signo; 504 } 505 506 static void sig_atexit(void) 507 { 508 if (child_pid != -1) 509 kill(child_pid, SIGTERM); 510 511 if (signr == -1) 512 return; 513 514 signal(signr, SIG_DFL); 515 kill(getpid(), signr); 516 } 517 518 static const char * const stat_usage[] = { 519 "perf stat [<options>] [<command>]", 520 NULL 521 }; 522 523 static const struct option options[] = { 524 OPT_CALLBACK('e', "event", NULL, "event", 525 "event selector. use 'perf list' to list available events", 526 parse_events), 527 OPT_BOOLEAN('i', "no-inherit", &no_inherit, 528 "child tasks do not inherit counters"), 529 OPT_INTEGER('p', "pid", &target_pid, 530 "stat events on existing process id"), 531 OPT_INTEGER('t', "tid", &target_tid, 532 "stat events on existing thread id"), 533 OPT_BOOLEAN('a', "all-cpus", &system_wide, 534 "system-wide collection from all CPUs"), 535 OPT_BOOLEAN('c', "scale", &scale, 536 "scale/normalize counters"), 537 OPT_INCR('v', "verbose", &verbose, 538 "be more verbose (show counter open errors, etc)"), 539 OPT_INTEGER('r', "repeat", &run_count, 540 "repeat command and print average + stddev (max: 100)"), 541 OPT_BOOLEAN('n', "null", &null_run, 542 "null run - dont start any counters"), 543 OPT_BOOLEAN('B', "big-num", &big_num, 544 "print large numbers with thousands\' separators"), 545 OPT_END() 546 }; 547 548 int cmd_stat(int argc, const char **argv, const char *prefix __used) 549 { 550 int status; 551 int i,j; 552 553 setlocale(LC_ALL, ""); 554 555 argc = parse_options(argc, argv, options, stat_usage, 556 PARSE_OPT_STOP_AT_NON_OPTION); 557 if (!argc && target_pid == -1 && target_tid == -1) 558 usage_with_options(stat_usage, options); 559 if (run_count <= 0) 560 usage_with_options(stat_usage, options); 561 562 /* Set attrs and nr_counters if no event is selected and !null_run */ 563 if (!null_run && !nr_counters) { 564 memcpy(attrs, default_attrs, sizeof(default_attrs)); 565 nr_counters = ARRAY_SIZE(default_attrs); 566 } 567 568 if (system_wide) 569 nr_cpus = read_cpu_map(); 570 else 571 nr_cpus = 1; 572 573 if (target_pid != -1) { 574 target_tid = target_pid; 575 thread_num = find_all_tid(target_pid, &all_tids); 576 if (thread_num <= 0) { 577 fprintf(stderr, "Can't find all threads of pid %d\n", 578 target_pid); 579 usage_with_options(stat_usage, options); 580 } 581 } else { 582 all_tids=malloc(sizeof(pid_t)); 583 if (!all_tids) 584 return -ENOMEM; 585 586 all_tids[0] = target_tid; 587 thread_num = 1; 588 } 589 590 for (i = 0; i < MAX_NR_CPUS; i++) { 591 for (j = 0; j < MAX_COUNTERS; j++) { 592 fd[i][j] = malloc(sizeof(int)*thread_num); 593 if (!fd[i][j]) 594 return -ENOMEM; 595 } 596 } 597 598 /* 599 * We dont want to block the signals - that would cause 600 * child tasks to inherit that and Ctrl-C would not work. 601 * What we want is for Ctrl-C to work in the exec()-ed 602 * task, but being ignored by perf stat itself: 603 */ 604 atexit(sig_atexit); 605 signal(SIGINT, skip_signal); 606 signal(SIGALRM, skip_signal); 607 signal(SIGABRT, skip_signal); 608 609 status = 0; 610 for (run_idx = 0; run_idx < run_count; run_idx++) { 611 if (run_count != 1 && verbose) 612 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); 613 status = run_perf_stat(argc, argv); 614 } 615 616 if (status != -1) 617 print_stat(argc, argv); 618 619 return status; 620 } 621