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 11 Time: 0.118 12 13 Performance counter stats for './hackbench 10': 14 15 1708.761321 task-clock # 11.037 CPUs utilized 16 41,190 context-switches # 0.024 M/sec 17 6,735 CPU-migrations # 0.004 M/sec 18 17,318 page-faults # 0.010 M/sec 19 5,205,202,243 cycles # 3.046 GHz 20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle 21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle 22 2,603,501,247 instructions # 0.50 insns per cycle 23 # 1.48 stalled cycles per insn 24 484,357,498 branches # 283.455 M/sec 25 6,388,934 branch-misses # 1.32% of all branches 26 27 0.154822978 seconds time elapsed 28 29 * 30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 31 * 32 * Improvements and fixes by: 33 * 34 * Arjan van de Ven <arjan@linux.intel.com> 35 * Yanmin Zhang <yanmin.zhang@intel.com> 36 * Wu Fengguang <fengguang.wu@intel.com> 37 * Mike Galbraith <efault@gmx.de> 38 * Paul Mackerras <paulus@samba.org> 39 * Jaswinder Singh Rajput <jaswinder@kernel.org> 40 * 41 * Released under the GPL v2. (and only v2, not any later version) 42 */ 43 44 #include "perf.h" 45 #include "builtin.h" 46 #include "util/util.h" 47 #include "util/parse-options.h" 48 #include "util/parse-events.h" 49 #include "util/event.h" 50 #include "util/evlist.h" 51 #include "util/evsel.h" 52 #include "util/debug.h" 53 #include "util/color.h" 54 #include "util/header.h" 55 #include "util/cpumap.h" 56 #include "util/thread.h" 57 #include "util/thread_map.h" 58 59 #include <sys/prctl.h> 60 #include <math.h> 61 #include <locale.h> 62 63 #define DEFAULT_SEPARATOR " " 64 #define CNTR_NOT_SUPPORTED "<not supported>" 65 #define CNTR_NOT_COUNTED "<not counted>" 66 67 static struct perf_event_attr default_attrs[] = { 68 69 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 70 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 71 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 72 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 73 74 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 75 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 76 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 77 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 78 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 79 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 80 81 }; 82 83 /* 84 * Detailed stats (-d), covering the L1 and last level data caches: 85 */ 86 static struct perf_event_attr detailed_attrs[] = { 87 88 { .type = PERF_TYPE_HW_CACHE, 89 .config = 90 PERF_COUNT_HW_CACHE_L1D << 0 | 91 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 93 94 { .type = PERF_TYPE_HW_CACHE, 95 .config = 96 PERF_COUNT_HW_CACHE_L1D << 0 | 97 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 98 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 99 100 { .type = PERF_TYPE_HW_CACHE, 101 .config = 102 PERF_COUNT_HW_CACHE_LL << 0 | 103 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 105 106 { .type = PERF_TYPE_HW_CACHE, 107 .config = 108 PERF_COUNT_HW_CACHE_LL << 0 | 109 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 110 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 111 }; 112 113 /* 114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 115 */ 116 static struct perf_event_attr very_detailed_attrs[] = { 117 118 { .type = PERF_TYPE_HW_CACHE, 119 .config = 120 PERF_COUNT_HW_CACHE_L1I << 0 | 121 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 123 124 { .type = PERF_TYPE_HW_CACHE, 125 .config = 126 PERF_COUNT_HW_CACHE_L1I << 0 | 127 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 128 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 129 130 { .type = PERF_TYPE_HW_CACHE, 131 .config = 132 PERF_COUNT_HW_CACHE_DTLB << 0 | 133 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 135 136 { .type = PERF_TYPE_HW_CACHE, 137 .config = 138 PERF_COUNT_HW_CACHE_DTLB << 0 | 139 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 140 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 141 142 { .type = PERF_TYPE_HW_CACHE, 143 .config = 144 PERF_COUNT_HW_CACHE_ITLB << 0 | 145 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 147 148 { .type = PERF_TYPE_HW_CACHE, 149 .config = 150 PERF_COUNT_HW_CACHE_ITLB << 0 | 151 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 152 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 153 154 }; 155 156 /* 157 * Very, very detailed stats (-d -d -d), adding prefetch events: 158 */ 159 static struct perf_event_attr very_very_detailed_attrs[] = { 160 161 { .type = PERF_TYPE_HW_CACHE, 162 .config = 163 PERF_COUNT_HW_CACHE_L1D << 0 | 164 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 166 167 { .type = PERF_TYPE_HW_CACHE, 168 .config = 169 PERF_COUNT_HW_CACHE_L1D << 0 | 170 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 171 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 172 }; 173 174 175 176 struct perf_evlist *evsel_list; 177 178 static bool system_wide = false; 179 static int run_idx = 0; 180 181 static int run_count = 1; 182 static bool no_inherit = false; 183 static bool scale = true; 184 static bool no_aggr = false; 185 static pid_t target_pid = -1; 186 static pid_t target_tid = -1; 187 static pid_t child_pid = -1; 188 static bool null_run = false; 189 static int detailed_run = 0; 190 static bool sync_run = false; 191 static bool big_num = true; 192 static int big_num_opt = -1; 193 static const char *cpu_list; 194 static const char *csv_sep = NULL; 195 static bool csv_output = false; 196 static bool group = false; 197 static const char *output_name = NULL; 198 static FILE *output = NULL; 199 static int output_fd; 200 201 static volatile int done = 0; 202 203 struct stats 204 { 205 double n, mean, M2; 206 }; 207 208 struct perf_stat { 209 struct stats res_stats[3]; 210 }; 211 212 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) 213 { 214 evsel->priv = zalloc(sizeof(struct perf_stat)); 215 return evsel->priv == NULL ? -ENOMEM : 0; 216 } 217 218 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) 219 { 220 free(evsel->priv); 221 evsel->priv = NULL; 222 } 223 224 static void update_stats(struct stats *stats, u64 val) 225 { 226 double delta; 227 228 stats->n++; 229 delta = val - stats->mean; 230 stats->mean += delta / stats->n; 231 stats->M2 += delta*(val - stats->mean); 232 } 233 234 static double avg_stats(struct stats *stats) 235 { 236 return stats->mean; 237 } 238 239 /* 240 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 241 * 242 * (\Sum n_i^2) - ((\Sum n_i)^2)/n 243 * s^2 = ------------------------------- 244 * n - 1 245 * 246 * http://en.wikipedia.org/wiki/Stddev 247 * 248 * The std dev of the mean is related to the std dev by: 249 * 250 * s 251 * s_mean = ------- 252 * sqrt(n) 253 * 254 */ 255 static double stddev_stats(struct stats *stats) 256 { 257 double variance, variance_mean; 258 259 if (!stats->n) 260 return 0.0; 261 262 variance = stats->M2 / (stats->n - 1); 263 variance_mean = variance / stats->n; 264 265 return sqrt(variance_mean); 266 } 267 268 struct stats runtime_nsecs_stats[MAX_NR_CPUS]; 269 struct stats runtime_cycles_stats[MAX_NR_CPUS]; 270 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; 271 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; 272 struct stats runtime_branches_stats[MAX_NR_CPUS]; 273 struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; 274 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; 275 struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; 276 struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; 277 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; 278 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; 279 struct stats walltime_nsecs_stats; 280 281 static int create_perf_stat_counter(struct perf_evsel *evsel, 282 struct perf_evsel *first) 283 { 284 struct perf_event_attr *attr = &evsel->attr; 285 struct xyarray *group_fd = NULL; 286 287 if (group && evsel != first) 288 group_fd = first->fd; 289 290 if (scale) 291 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 292 PERF_FORMAT_TOTAL_TIME_RUNNING; 293 294 attr->inherit = !no_inherit; 295 296 if (system_wide) 297 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, 298 group, group_fd); 299 if (target_pid == -1 && target_tid == -1) { 300 attr->disabled = 1; 301 attr->enable_on_exec = 1; 302 } 303 304 return perf_evsel__open_per_thread(evsel, evsel_list->threads, 305 group, group_fd); 306 } 307 308 /* 309 * Does the counter have nsecs as a unit? 310 */ 311 static inline int nsec_counter(struct perf_evsel *evsel) 312 { 313 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || 314 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 315 return 1; 316 317 return 0; 318 } 319 320 /* 321 * Update various tracking values we maintain to print 322 * more semantic information such as miss/hit ratios, 323 * instruction rates, etc: 324 */ 325 static void update_shadow_stats(struct perf_evsel *counter, u64 *count) 326 { 327 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) 328 update_stats(&runtime_nsecs_stats[0], count[0]); 329 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) 330 update_stats(&runtime_cycles_stats[0], count[0]); 331 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) 332 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); 333 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) 334 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); 335 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) 336 update_stats(&runtime_branches_stats[0], count[0]); 337 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) 338 update_stats(&runtime_cacherefs_stats[0], count[0]); 339 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) 340 update_stats(&runtime_l1_dcache_stats[0], count[0]); 341 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) 342 update_stats(&runtime_l1_icache_stats[0], count[0]); 343 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) 344 update_stats(&runtime_ll_cache_stats[0], count[0]); 345 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) 346 update_stats(&runtime_dtlb_cache_stats[0], count[0]); 347 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) 348 update_stats(&runtime_itlb_cache_stats[0], count[0]); 349 } 350 351 /* 352 * Read out the results of a single counter: 353 * aggregate counts across CPUs in system-wide mode 354 */ 355 static int read_counter_aggr(struct perf_evsel *counter) 356 { 357 struct perf_stat *ps = counter->priv; 358 u64 *count = counter->counts->aggr.values; 359 int i; 360 361 if (__perf_evsel__read(counter, evsel_list->cpus->nr, 362 evsel_list->threads->nr, scale) < 0) 363 return -1; 364 365 for (i = 0; i < 3; i++) 366 update_stats(&ps->res_stats[i], count[i]); 367 368 if (verbose) { 369 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 370 event_name(counter), count[0], count[1], count[2]); 371 } 372 373 /* 374 * Save the full runtime - to allow normalization during printout: 375 */ 376 update_shadow_stats(counter, count); 377 378 return 0; 379 } 380 381 /* 382 * Read out the results of a single counter: 383 * do not aggregate counts across CPUs in system-wide mode 384 */ 385 static int read_counter(struct perf_evsel *counter) 386 { 387 u64 *count; 388 int cpu; 389 390 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { 391 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) 392 return -1; 393 394 count = counter->counts->cpu[cpu].values; 395 396 update_shadow_stats(counter, count); 397 } 398 399 return 0; 400 } 401 402 static int run_perf_stat(int argc __used, const char **argv) 403 { 404 unsigned long long t0, t1; 405 struct perf_evsel *counter, *first; 406 int status = 0; 407 int child_ready_pipe[2], go_pipe[2]; 408 const bool forks = (argc > 0); 409 char buf; 410 411 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 412 perror("failed to create pipes"); 413 exit(1); 414 } 415 416 if (forks) { 417 if ((child_pid = fork()) < 0) 418 perror("failed to fork"); 419 420 if (!child_pid) { 421 close(child_ready_pipe[0]); 422 close(go_pipe[1]); 423 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 424 425 /* 426 * Do a dummy execvp to get the PLT entry resolved, 427 * so we avoid the resolver overhead on the real 428 * execvp call. 429 */ 430 execvp("", (char **)argv); 431 432 /* 433 * Tell the parent we're ready to go 434 */ 435 close(child_ready_pipe[1]); 436 437 /* 438 * Wait until the parent tells us to go. 439 */ 440 if (read(go_pipe[0], &buf, 1) == -1) 441 perror("unable to read pipe"); 442 443 execvp(argv[0], (char **)argv); 444 445 perror(argv[0]); 446 exit(-1); 447 } 448 449 if (target_tid == -1 && target_pid == -1 && !system_wide) 450 evsel_list->threads->map[0] = child_pid; 451 452 /* 453 * Wait for the child to be ready to exec. 454 */ 455 close(child_ready_pipe[1]); 456 close(go_pipe[0]); 457 if (read(child_ready_pipe[0], &buf, 1) == -1) 458 perror("unable to read pipe"); 459 close(child_ready_pipe[0]); 460 } 461 462 first = list_entry(evsel_list->entries.next, struct perf_evsel, node); 463 464 list_for_each_entry(counter, &evsel_list->entries, node) { 465 if (create_perf_stat_counter(counter, first) < 0) { 466 if (errno == EINVAL || errno == ENOSYS || 467 errno == ENOENT || errno == EOPNOTSUPP) { 468 if (verbose) 469 ui__warning("%s event is not supported by the kernel.\n", 470 event_name(counter)); 471 counter->supported = false; 472 continue; 473 } 474 475 if (errno == EPERM || errno == EACCES) { 476 error("You may not have permission to collect %sstats.\n" 477 "\t Consider tweaking" 478 " /proc/sys/kernel/perf_event_paranoid or running as root.", 479 system_wide ? "system-wide " : ""); 480 } else { 481 error("open_counter returned with %d (%s). " 482 "/bin/dmesg may provide additional information.\n", 483 errno, strerror(errno)); 484 } 485 if (child_pid != -1) 486 kill(child_pid, SIGTERM); 487 die("Not all events could be opened.\n"); 488 return -1; 489 } 490 counter->supported = true; 491 } 492 493 if (perf_evlist__set_filters(evsel_list)) { 494 error("failed to set filter with %d (%s)\n", errno, 495 strerror(errno)); 496 return -1; 497 } 498 499 /* 500 * Enable counters and exec the command: 501 */ 502 t0 = rdclock(); 503 504 if (forks) { 505 close(go_pipe[1]); 506 wait(&status); 507 if (WIFSIGNALED(status)) 508 psignal(WTERMSIG(status), argv[0]); 509 } else { 510 while(!done) sleep(1); 511 } 512 513 t1 = rdclock(); 514 515 update_stats(&walltime_nsecs_stats, t1 - t0); 516 517 if (no_aggr) { 518 list_for_each_entry(counter, &evsel_list->entries, node) { 519 read_counter(counter); 520 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1); 521 } 522 } else { 523 list_for_each_entry(counter, &evsel_list->entries, node) { 524 read_counter_aggr(counter); 525 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 526 evsel_list->threads->nr); 527 } 528 } 529 530 return WEXITSTATUS(status); 531 } 532 533 static void print_noise_pct(double total, double avg) 534 { 535 double pct = 0.0; 536 537 if (avg) 538 pct = 100.0*total/avg; 539 540 if (csv_output) 541 fprintf(output, "%s%.2f%%", csv_sep, pct); 542 else if (pct) 543 fprintf(output, " ( +-%6.2f%% )", pct); 544 } 545 546 static void print_noise(struct perf_evsel *evsel, double avg) 547 { 548 struct perf_stat *ps; 549 550 if (run_count == 1) 551 return; 552 553 ps = evsel->priv; 554 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); 555 } 556 557 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) 558 { 559 double msecs = avg / 1e6; 560 char cpustr[16] = { '\0', }; 561 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; 562 563 if (no_aggr) 564 sprintf(cpustr, "CPU%*d%s", 565 csv_output ? 0 : -4, 566 evsel_list->cpus->map[cpu], csv_sep); 567 568 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel)); 569 570 if (evsel->cgrp) 571 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); 572 573 if (csv_output) 574 return; 575 576 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 577 fprintf(output, " # %8.3f CPUs utilized ", 578 avg / avg_stats(&walltime_nsecs_stats)); 579 } 580 581 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg) 582 { 583 double total, ratio = 0.0; 584 const char *color; 585 586 total = avg_stats(&runtime_cycles_stats[cpu]); 587 588 if (total) 589 ratio = avg / total * 100.0; 590 591 color = PERF_COLOR_NORMAL; 592 if (ratio > 50.0) 593 color = PERF_COLOR_RED; 594 else if (ratio > 30.0) 595 color = PERF_COLOR_MAGENTA; 596 else if (ratio > 10.0) 597 color = PERF_COLOR_YELLOW; 598 599 fprintf(output, " # "); 600 color_fprintf(output, color, "%6.2f%%", ratio); 601 fprintf(output, " frontend cycles idle "); 602 } 603 604 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg) 605 { 606 double total, ratio = 0.0; 607 const char *color; 608 609 total = avg_stats(&runtime_cycles_stats[cpu]); 610 611 if (total) 612 ratio = avg / total * 100.0; 613 614 color = PERF_COLOR_NORMAL; 615 if (ratio > 75.0) 616 color = PERF_COLOR_RED; 617 else if (ratio > 50.0) 618 color = PERF_COLOR_MAGENTA; 619 else if (ratio > 20.0) 620 color = PERF_COLOR_YELLOW; 621 622 fprintf(output, " # "); 623 color_fprintf(output, color, "%6.2f%%", ratio); 624 fprintf(output, " backend cycles idle "); 625 } 626 627 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg) 628 { 629 double total, ratio = 0.0; 630 const char *color; 631 632 total = avg_stats(&runtime_branches_stats[cpu]); 633 634 if (total) 635 ratio = avg / total * 100.0; 636 637 color = PERF_COLOR_NORMAL; 638 if (ratio > 20.0) 639 color = PERF_COLOR_RED; 640 else if (ratio > 10.0) 641 color = PERF_COLOR_MAGENTA; 642 else if (ratio > 5.0) 643 color = PERF_COLOR_YELLOW; 644 645 fprintf(output, " # "); 646 color_fprintf(output, color, "%6.2f%%", ratio); 647 fprintf(output, " of all branches "); 648 } 649 650 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 651 { 652 double total, ratio = 0.0; 653 const char *color; 654 655 total = avg_stats(&runtime_l1_dcache_stats[cpu]); 656 657 if (total) 658 ratio = avg / total * 100.0; 659 660 color = PERF_COLOR_NORMAL; 661 if (ratio > 20.0) 662 color = PERF_COLOR_RED; 663 else if (ratio > 10.0) 664 color = PERF_COLOR_MAGENTA; 665 else if (ratio > 5.0) 666 color = PERF_COLOR_YELLOW; 667 668 fprintf(output, " # "); 669 color_fprintf(output, color, "%6.2f%%", ratio); 670 fprintf(output, " of all L1-dcache hits "); 671 } 672 673 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 674 { 675 double total, ratio = 0.0; 676 const char *color; 677 678 total = avg_stats(&runtime_l1_icache_stats[cpu]); 679 680 if (total) 681 ratio = avg / total * 100.0; 682 683 color = PERF_COLOR_NORMAL; 684 if (ratio > 20.0) 685 color = PERF_COLOR_RED; 686 else if (ratio > 10.0) 687 color = PERF_COLOR_MAGENTA; 688 else if (ratio > 5.0) 689 color = PERF_COLOR_YELLOW; 690 691 fprintf(output, " # "); 692 color_fprintf(output, color, "%6.2f%%", ratio); 693 fprintf(output, " of all L1-icache hits "); 694 } 695 696 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 697 { 698 double total, ratio = 0.0; 699 const char *color; 700 701 total = avg_stats(&runtime_dtlb_cache_stats[cpu]); 702 703 if (total) 704 ratio = avg / total * 100.0; 705 706 color = PERF_COLOR_NORMAL; 707 if (ratio > 20.0) 708 color = PERF_COLOR_RED; 709 else if (ratio > 10.0) 710 color = PERF_COLOR_MAGENTA; 711 else if (ratio > 5.0) 712 color = PERF_COLOR_YELLOW; 713 714 fprintf(output, " # "); 715 color_fprintf(output, color, "%6.2f%%", ratio); 716 fprintf(output, " of all dTLB cache hits "); 717 } 718 719 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 720 { 721 double total, ratio = 0.0; 722 const char *color; 723 724 total = avg_stats(&runtime_itlb_cache_stats[cpu]); 725 726 if (total) 727 ratio = avg / total * 100.0; 728 729 color = PERF_COLOR_NORMAL; 730 if (ratio > 20.0) 731 color = PERF_COLOR_RED; 732 else if (ratio > 10.0) 733 color = PERF_COLOR_MAGENTA; 734 else if (ratio > 5.0) 735 color = PERF_COLOR_YELLOW; 736 737 fprintf(output, " # "); 738 color_fprintf(output, color, "%6.2f%%", ratio); 739 fprintf(output, " of all iTLB cache hits "); 740 } 741 742 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 743 { 744 double total, ratio = 0.0; 745 const char *color; 746 747 total = avg_stats(&runtime_ll_cache_stats[cpu]); 748 749 if (total) 750 ratio = avg / total * 100.0; 751 752 color = PERF_COLOR_NORMAL; 753 if (ratio > 20.0) 754 color = PERF_COLOR_RED; 755 else if (ratio > 10.0) 756 color = PERF_COLOR_MAGENTA; 757 else if (ratio > 5.0) 758 color = PERF_COLOR_YELLOW; 759 760 fprintf(output, " # "); 761 color_fprintf(output, color, "%6.2f%%", ratio); 762 fprintf(output, " of all LL-cache hits "); 763 } 764 765 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) 766 { 767 double total, ratio = 0.0; 768 char cpustr[16] = { '\0', }; 769 const char *fmt; 770 771 if (csv_output) 772 fmt = "%s%.0f%s%s"; 773 else if (big_num) 774 fmt = "%s%'18.0f%s%-25s"; 775 else 776 fmt = "%s%18.0f%s%-25s"; 777 778 if (no_aggr) 779 sprintf(cpustr, "CPU%*d%s", 780 csv_output ? 0 : -4, 781 evsel_list->cpus->map[cpu], csv_sep); 782 else 783 cpu = 0; 784 785 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel)); 786 787 if (evsel->cgrp) 788 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); 789 790 if (csv_output) 791 return; 792 793 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { 794 total = avg_stats(&runtime_cycles_stats[cpu]); 795 796 if (total) 797 ratio = avg / total; 798 799 fprintf(output, " # %5.2f insns per cycle ", ratio); 800 801 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); 802 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); 803 804 if (total && avg) { 805 ratio = total / avg; 806 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio); 807 } 808 809 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && 810 runtime_branches_stats[cpu].n != 0) { 811 print_branch_misses(cpu, evsel, avg); 812 } else if ( 813 evsel->attr.type == PERF_TYPE_HW_CACHE && 814 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | 815 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 816 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 817 runtime_l1_dcache_stats[cpu].n != 0) { 818 print_l1_dcache_misses(cpu, evsel, avg); 819 } else if ( 820 evsel->attr.type == PERF_TYPE_HW_CACHE && 821 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | 822 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 823 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 824 runtime_l1_icache_stats[cpu].n != 0) { 825 print_l1_icache_misses(cpu, evsel, avg); 826 } else if ( 827 evsel->attr.type == PERF_TYPE_HW_CACHE && 828 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | 829 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 830 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 831 runtime_dtlb_cache_stats[cpu].n != 0) { 832 print_dtlb_cache_misses(cpu, evsel, avg); 833 } else if ( 834 evsel->attr.type == PERF_TYPE_HW_CACHE && 835 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | 836 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 837 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 838 runtime_itlb_cache_stats[cpu].n != 0) { 839 print_itlb_cache_misses(cpu, evsel, avg); 840 } else if ( 841 evsel->attr.type == PERF_TYPE_HW_CACHE && 842 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | 843 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 844 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 845 runtime_ll_cache_stats[cpu].n != 0) { 846 print_ll_cache_misses(cpu, evsel, avg); 847 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && 848 runtime_cacherefs_stats[cpu].n != 0) { 849 total = avg_stats(&runtime_cacherefs_stats[cpu]); 850 851 if (total) 852 ratio = avg * 100 / total; 853 854 fprintf(output, " # %8.3f %% of all cache refs ", ratio); 855 856 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { 857 print_stalled_cycles_frontend(cpu, evsel, avg); 858 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { 859 print_stalled_cycles_backend(cpu, evsel, avg); 860 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { 861 total = avg_stats(&runtime_nsecs_stats[cpu]); 862 863 if (total) 864 ratio = 1.0 * avg / total; 865 866 fprintf(output, " # %8.3f GHz ", ratio); 867 } else if (runtime_nsecs_stats[cpu].n != 0) { 868 total = avg_stats(&runtime_nsecs_stats[cpu]); 869 870 if (total) 871 ratio = 1000.0 * avg / total; 872 873 fprintf(output, " # %8.3f M/sec ", ratio); 874 } else { 875 fprintf(output, " "); 876 } 877 } 878 879 /* 880 * Print out the results of a single counter: 881 * aggregated counts in system-wide mode 882 */ 883 static void print_counter_aggr(struct perf_evsel *counter) 884 { 885 struct perf_stat *ps = counter->priv; 886 double avg = avg_stats(&ps->res_stats[0]); 887 int scaled = counter->counts->scaled; 888 889 if (scaled == -1) { 890 fprintf(output, "%*s%s%*s", 891 csv_output ? 0 : 18, 892 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 893 csv_sep, 894 csv_output ? 0 : -24, 895 event_name(counter)); 896 897 if (counter->cgrp) 898 fprintf(output, "%s%s", csv_sep, counter->cgrp->name); 899 900 fputc('\n', output); 901 return; 902 } 903 904 if (nsec_counter(counter)) 905 nsec_printout(-1, counter, avg); 906 else 907 abs_printout(-1, counter, avg); 908 909 print_noise(counter, avg); 910 911 if (csv_output) { 912 fputc('\n', output); 913 return; 914 } 915 916 if (scaled) { 917 double avg_enabled, avg_running; 918 919 avg_enabled = avg_stats(&ps->res_stats[1]); 920 avg_running = avg_stats(&ps->res_stats[2]); 921 922 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); 923 } 924 fprintf(output, "\n"); 925 } 926 927 /* 928 * Print out the results of a single counter: 929 * does not use aggregated count in system-wide 930 */ 931 static void print_counter(struct perf_evsel *counter) 932 { 933 u64 ena, run, val; 934 int cpu; 935 936 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { 937 val = counter->counts->cpu[cpu].val; 938 ena = counter->counts->cpu[cpu].ena; 939 run = counter->counts->cpu[cpu].run; 940 if (run == 0 || ena == 0) { 941 fprintf(output, "CPU%*d%s%*s%s%*s", 942 csv_output ? 0 : -4, 943 evsel_list->cpus->map[cpu], csv_sep, 944 csv_output ? 0 : 18, 945 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 946 csv_sep, 947 csv_output ? 0 : -24, 948 event_name(counter)); 949 950 if (counter->cgrp) 951 fprintf(output, "%s%s", 952 csv_sep, counter->cgrp->name); 953 954 fputc('\n', output); 955 continue; 956 } 957 958 if (nsec_counter(counter)) 959 nsec_printout(cpu, counter, val); 960 else 961 abs_printout(cpu, counter, val); 962 963 if (!csv_output) { 964 print_noise(counter, 1.0); 965 966 if (run != ena) 967 fprintf(output, " (%.2f%%)", 968 100.0 * run / ena); 969 } 970 fputc('\n', output); 971 } 972 } 973 974 static void print_stat(int argc, const char **argv) 975 { 976 struct perf_evsel *counter; 977 int i; 978 979 fflush(stdout); 980 981 if (!csv_output) { 982 fprintf(output, "\n"); 983 fprintf(output, " Performance counter stats for "); 984 if(target_pid == -1 && target_tid == -1) { 985 fprintf(output, "\'%s", argv[0]); 986 for (i = 1; i < argc; i++) 987 fprintf(output, " %s", argv[i]); 988 } else if (target_pid != -1) 989 fprintf(output, "process id \'%d", target_pid); 990 else 991 fprintf(output, "thread id \'%d", target_tid); 992 993 fprintf(output, "\'"); 994 if (run_count > 1) 995 fprintf(output, " (%d runs)", run_count); 996 fprintf(output, ":\n\n"); 997 } 998 999 if (no_aggr) { 1000 list_for_each_entry(counter, &evsel_list->entries, node) 1001 print_counter(counter); 1002 } else { 1003 list_for_each_entry(counter, &evsel_list->entries, node) 1004 print_counter_aggr(counter); 1005 } 1006 1007 if (!csv_output) { 1008 if (!null_run) 1009 fprintf(output, "\n"); 1010 fprintf(output, " %17.9f seconds time elapsed", 1011 avg_stats(&walltime_nsecs_stats)/1e9); 1012 if (run_count > 1) { 1013 fprintf(output, " "); 1014 print_noise_pct(stddev_stats(&walltime_nsecs_stats), 1015 avg_stats(&walltime_nsecs_stats)); 1016 } 1017 fprintf(output, "\n\n"); 1018 } 1019 } 1020 1021 static volatile int signr = -1; 1022 1023 static void skip_signal(int signo) 1024 { 1025 if(child_pid == -1) 1026 done = 1; 1027 1028 signr = signo; 1029 } 1030 1031 static void sig_atexit(void) 1032 { 1033 if (child_pid != -1) 1034 kill(child_pid, SIGTERM); 1035 1036 if (signr == -1) 1037 return; 1038 1039 signal(signr, SIG_DFL); 1040 kill(getpid(), signr); 1041 } 1042 1043 static const char * const stat_usage[] = { 1044 "perf stat [<options>] [<command>]", 1045 NULL 1046 }; 1047 1048 static int stat__set_big_num(const struct option *opt __used, 1049 const char *s __used, int unset) 1050 { 1051 big_num_opt = unset ? 0 : 1; 1052 return 0; 1053 } 1054 1055 static bool append_file; 1056 1057 static const struct option options[] = { 1058 OPT_CALLBACK('e', "event", &evsel_list, "event", 1059 "event selector. use 'perf list' to list available events", 1060 parse_events_option), 1061 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 1062 "event filter", parse_filter), 1063 OPT_BOOLEAN('i', "no-inherit", &no_inherit, 1064 "child tasks do not inherit counters"), 1065 OPT_INTEGER('p', "pid", &target_pid, 1066 "stat events on existing process id"), 1067 OPT_INTEGER('t', "tid", &target_tid, 1068 "stat events on existing thread id"), 1069 OPT_BOOLEAN('a', "all-cpus", &system_wide, 1070 "system-wide collection from all CPUs"), 1071 OPT_BOOLEAN('g', "group", &group, 1072 "put the counters into a counter group"), 1073 OPT_BOOLEAN('c', "scale", &scale, 1074 "scale/normalize counters"), 1075 OPT_INCR('v', "verbose", &verbose, 1076 "be more verbose (show counter open errors, etc)"), 1077 OPT_INTEGER('r', "repeat", &run_count, 1078 "repeat command and print average + stddev (max: 100)"), 1079 OPT_BOOLEAN('n', "null", &null_run, 1080 "null run - dont start any counters"), 1081 OPT_INCR('d', "detailed", &detailed_run, 1082 "detailed run - start a lot of events"), 1083 OPT_BOOLEAN('S', "sync", &sync_run, 1084 "call sync() before starting a run"), 1085 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1086 "print large numbers with thousands\' separators", 1087 stat__set_big_num), 1088 OPT_STRING('C', "cpu", &cpu_list, "cpu", 1089 "list of cpus to monitor in system-wide"), 1090 OPT_BOOLEAN('A', "no-aggr", &no_aggr, 1091 "disable CPU count aggregation"), 1092 OPT_STRING('x', "field-separator", &csv_sep, "separator", 1093 "print counts with custom separator"), 1094 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 1095 "monitor event in cgroup name only", 1096 parse_cgroups), 1097 OPT_STRING('o', "output", &output_name, "file", 1098 "output file name"), 1099 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 1100 OPT_INTEGER(0, "log-fd", &output_fd, 1101 "log output to fd, instead of stderr"), 1102 OPT_END() 1103 }; 1104 1105 /* 1106 * Add default attributes, if there were no attributes specified or 1107 * if -d/--detailed, -d -d or -d -d -d is used: 1108 */ 1109 static int add_default_attributes(void) 1110 { 1111 struct perf_evsel *pos; 1112 size_t attr_nr = 0; 1113 size_t c; 1114 1115 /* Set attrs if no event is selected and !null_run: */ 1116 if (null_run) 1117 return 0; 1118 1119 if (!evsel_list->nr_entries) { 1120 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) { 1121 pos = perf_evsel__new(default_attrs + c, c + attr_nr); 1122 if (pos == NULL) 1123 return -1; 1124 perf_evlist__add(evsel_list, pos); 1125 } 1126 attr_nr += c; 1127 } 1128 1129 /* Detailed events get appended to the event list: */ 1130 1131 if (detailed_run < 1) 1132 return 0; 1133 1134 /* Append detailed run extra attributes: */ 1135 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) { 1136 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr); 1137 if (pos == NULL) 1138 return -1; 1139 perf_evlist__add(evsel_list, pos); 1140 } 1141 attr_nr += c; 1142 1143 if (detailed_run < 2) 1144 return 0; 1145 1146 /* Append very detailed run extra attributes: */ 1147 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) { 1148 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr); 1149 if (pos == NULL) 1150 return -1; 1151 perf_evlist__add(evsel_list, pos); 1152 } 1153 1154 if (detailed_run < 3) 1155 return 0; 1156 1157 /* Append very, very detailed run extra attributes: */ 1158 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) { 1159 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr); 1160 if (pos == NULL) 1161 return -1; 1162 perf_evlist__add(evsel_list, pos); 1163 } 1164 1165 1166 return 0; 1167 } 1168 1169 int cmd_stat(int argc, const char **argv, const char *prefix __used) 1170 { 1171 struct perf_evsel *pos; 1172 int status = -ENOMEM; 1173 const char *mode; 1174 1175 setlocale(LC_ALL, ""); 1176 1177 evsel_list = perf_evlist__new(NULL, NULL); 1178 if (evsel_list == NULL) 1179 return -ENOMEM; 1180 1181 argc = parse_options(argc, argv, options, stat_usage, 1182 PARSE_OPT_STOP_AT_NON_OPTION); 1183 1184 output = stderr; 1185 if (output_name && strcmp(output_name, "-")) 1186 output = NULL; 1187 1188 if (output_name && output_fd) { 1189 fprintf(stderr, "cannot use both --output and --log-fd\n"); 1190 usage_with_options(stat_usage, options); 1191 } 1192 if (!output) { 1193 struct timespec tm; 1194 mode = append_file ? "a" : "w"; 1195 1196 output = fopen(output_name, mode); 1197 if (!output) { 1198 perror("failed to create output file"); 1199 exit(-1); 1200 } 1201 clock_gettime(CLOCK_REALTIME, &tm); 1202 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 1203 } else if (output_fd != 2) { 1204 mode = append_file ? "a" : "w"; 1205 output = fdopen(output_fd, mode); 1206 if (!output) { 1207 perror("Failed opening logfd"); 1208 return -errno; 1209 } 1210 } 1211 1212 if (csv_sep) { 1213 csv_output = true; 1214 if (!strcmp(csv_sep, "\\t")) 1215 csv_sep = "\t"; 1216 } else 1217 csv_sep = DEFAULT_SEPARATOR; 1218 1219 /* 1220 * let the spreadsheet do the pretty-printing 1221 */ 1222 if (csv_output) { 1223 /* User explicitly passed -B? */ 1224 if (big_num_opt == 1) { 1225 fprintf(stderr, "-B option not supported with -x\n"); 1226 usage_with_options(stat_usage, options); 1227 } else /* Nope, so disable big number formatting */ 1228 big_num = false; 1229 } else if (big_num_opt == 0) /* User passed --no-big-num */ 1230 big_num = false; 1231 1232 if (!argc && target_pid == -1 && target_tid == -1) 1233 usage_with_options(stat_usage, options); 1234 if (run_count <= 0) 1235 usage_with_options(stat_usage, options); 1236 1237 /* no_aggr, cgroup are for system-wide only */ 1238 if ((no_aggr || nr_cgroups) && !system_wide) { 1239 fprintf(stderr, "both cgroup and no-aggregation " 1240 "modes only available in system-wide mode\n"); 1241 1242 usage_with_options(stat_usage, options); 1243 } 1244 1245 if (add_default_attributes()) 1246 goto out; 1247 1248 if (target_pid != -1) 1249 target_tid = target_pid; 1250 1251 evsel_list->threads = thread_map__new(target_pid, target_tid); 1252 if (evsel_list->threads == NULL) { 1253 pr_err("Problems finding threads of monitor\n"); 1254 usage_with_options(stat_usage, options); 1255 } 1256 1257 if (system_wide) 1258 evsel_list->cpus = cpu_map__new(cpu_list); 1259 else 1260 evsel_list->cpus = cpu_map__dummy_new(); 1261 1262 if (evsel_list->cpus == NULL) { 1263 perror("failed to parse CPUs map"); 1264 usage_with_options(stat_usage, options); 1265 return -1; 1266 } 1267 1268 list_for_each_entry(pos, &evsel_list->entries, node) { 1269 if (perf_evsel__alloc_stat_priv(pos) < 0 || 1270 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 || 1271 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0) 1272 goto out_free_fd; 1273 } 1274 1275 /* 1276 * We dont want to block the signals - that would cause 1277 * child tasks to inherit that and Ctrl-C would not work. 1278 * What we want is for Ctrl-C to work in the exec()-ed 1279 * task, but being ignored by perf stat itself: 1280 */ 1281 atexit(sig_atexit); 1282 signal(SIGINT, skip_signal); 1283 signal(SIGALRM, skip_signal); 1284 signal(SIGABRT, skip_signal); 1285 1286 status = 0; 1287 for (run_idx = 0; run_idx < run_count; run_idx++) { 1288 if (run_count != 1 && verbose) 1289 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 1290 run_idx + 1); 1291 1292 if (sync_run) 1293 sync(); 1294 1295 status = run_perf_stat(argc, argv); 1296 } 1297 1298 if (status != -1) 1299 print_stat(argc, argv); 1300 out_free_fd: 1301 list_for_each_entry(pos, &evsel_list->entries, node) 1302 perf_evsel__free_stat_priv(pos); 1303 perf_evlist__delete_maps(evsel_list); 1304 out: 1305 perf_evlist__delete(evsel_list); 1306 return status; 1307 } 1308