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/stat.h" 55 #include "util/header.h" 56 #include "util/cpumap.h" 57 #include "util/thread.h" 58 #include "util/thread_map.h" 59 60 #include <stdlib.h> 61 #include <sys/prctl.h> 62 #include <locale.h> 63 64 #define DEFAULT_SEPARATOR " " 65 #define CNTR_NOT_SUPPORTED "<not supported>" 66 #define CNTR_NOT_COUNTED "<not counted>" 67 68 static struct perf_evlist *evsel_list; 69 70 static struct perf_target target = { 71 .uid = UINT_MAX, 72 }; 73 74 static int run_count = 1; 75 static bool no_inherit = false; 76 static bool scale = true; 77 static bool no_aggr = false; 78 static pid_t child_pid = -1; 79 static bool null_run = false; 80 static int detailed_run = 0; 81 static bool big_num = true; 82 static int big_num_opt = -1; 83 static const char *csv_sep = NULL; 84 static bool csv_output = false; 85 static bool group = false; 86 static FILE *output = NULL; 87 static const char *pre_cmd = NULL; 88 static const char *post_cmd = NULL; 89 static bool sync_run = false; 90 91 static volatile int done = 0; 92 93 struct perf_stat { 94 struct stats res_stats[3]; 95 }; 96 97 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) 98 { 99 evsel->priv = zalloc(sizeof(struct perf_stat)); 100 return evsel->priv == NULL ? -ENOMEM : 0; 101 } 102 103 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) 104 { 105 free(evsel->priv); 106 evsel->priv = NULL; 107 } 108 109 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel) 110 { 111 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus; 112 } 113 114 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel) 115 { 116 return perf_evsel__cpus(evsel)->nr; 117 } 118 119 static struct stats runtime_nsecs_stats[MAX_NR_CPUS]; 120 static struct stats runtime_cycles_stats[MAX_NR_CPUS]; 121 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; 122 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; 123 static struct stats runtime_branches_stats[MAX_NR_CPUS]; 124 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; 125 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; 126 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; 127 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; 128 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; 129 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; 130 static struct stats walltime_nsecs_stats; 131 132 static int create_perf_stat_counter(struct perf_evsel *evsel) 133 { 134 struct perf_event_attr *attr = &evsel->attr; 135 bool exclude_guest_missing = false; 136 int ret; 137 138 if (scale) 139 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 140 PERF_FORMAT_TOTAL_TIME_RUNNING; 141 142 attr->inherit = !no_inherit; 143 144 retry: 145 if (exclude_guest_missing) 146 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0; 147 148 if (perf_target__has_cpu(&target)) { 149 ret = perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); 150 if (ret) 151 goto check_ret; 152 return 0; 153 } 154 155 if (!perf_target__has_task(&target) && 156 !perf_evsel__is_group_member(evsel)) { 157 attr->disabled = 1; 158 attr->enable_on_exec = 1; 159 } 160 161 ret = perf_evsel__open_per_thread(evsel, evsel_list->threads); 162 if (!ret) 163 return 0; 164 /* fall through */ 165 check_ret: 166 if (ret && errno == EINVAL) { 167 if (!exclude_guest_missing && 168 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) { 169 pr_debug("Old kernel, cannot exclude " 170 "guest or host samples.\n"); 171 exclude_guest_missing = true; 172 goto retry; 173 } 174 } 175 return ret; 176 } 177 178 /* 179 * Does the counter have nsecs as a unit? 180 */ 181 static inline int nsec_counter(struct perf_evsel *evsel) 182 { 183 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || 184 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 185 return 1; 186 187 return 0; 188 } 189 190 /* 191 * Update various tracking values we maintain to print 192 * more semantic information such as miss/hit ratios, 193 * instruction rates, etc: 194 */ 195 static void update_shadow_stats(struct perf_evsel *counter, u64 *count) 196 { 197 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) 198 update_stats(&runtime_nsecs_stats[0], count[0]); 199 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) 200 update_stats(&runtime_cycles_stats[0], count[0]); 201 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) 202 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); 203 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) 204 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); 205 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) 206 update_stats(&runtime_branches_stats[0], count[0]); 207 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) 208 update_stats(&runtime_cacherefs_stats[0], count[0]); 209 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) 210 update_stats(&runtime_l1_dcache_stats[0], count[0]); 211 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) 212 update_stats(&runtime_l1_icache_stats[0], count[0]); 213 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) 214 update_stats(&runtime_ll_cache_stats[0], count[0]); 215 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) 216 update_stats(&runtime_dtlb_cache_stats[0], count[0]); 217 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) 218 update_stats(&runtime_itlb_cache_stats[0], count[0]); 219 } 220 221 /* 222 * Read out the results of a single counter: 223 * aggregate counts across CPUs in system-wide mode 224 */ 225 static int read_counter_aggr(struct perf_evsel *counter) 226 { 227 struct perf_stat *ps = counter->priv; 228 u64 *count = counter->counts->aggr.values; 229 int i; 230 231 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter), 232 evsel_list->threads->nr, scale) < 0) 233 return -1; 234 235 for (i = 0; i < 3; i++) 236 update_stats(&ps->res_stats[i], count[i]); 237 238 if (verbose) { 239 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 240 perf_evsel__name(counter), count[0], count[1], count[2]); 241 } 242 243 /* 244 * Save the full runtime - to allow normalization during printout: 245 */ 246 update_shadow_stats(counter, count); 247 248 return 0; 249 } 250 251 /* 252 * Read out the results of a single counter: 253 * do not aggregate counts across CPUs in system-wide mode 254 */ 255 static int read_counter(struct perf_evsel *counter) 256 { 257 u64 *count; 258 int cpu; 259 260 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { 261 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) 262 return -1; 263 264 count = counter->counts->cpu[cpu].values; 265 266 update_shadow_stats(counter, count); 267 } 268 269 return 0; 270 } 271 272 static int __run_perf_stat(int argc __maybe_unused, const char **argv) 273 { 274 unsigned long long t0, t1; 275 struct perf_evsel *counter; 276 int status = 0; 277 int child_ready_pipe[2], go_pipe[2]; 278 const bool forks = (argc > 0); 279 char buf; 280 281 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 282 perror("failed to create pipes"); 283 return -1; 284 } 285 286 if (forks) { 287 if ((child_pid = fork()) < 0) 288 perror("failed to fork"); 289 290 if (!child_pid) { 291 close(child_ready_pipe[0]); 292 close(go_pipe[1]); 293 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 294 295 /* 296 * Do a dummy execvp to get the PLT entry resolved, 297 * so we avoid the resolver overhead on the real 298 * execvp call. 299 */ 300 execvp("", (char **)argv); 301 302 /* 303 * Tell the parent we're ready to go 304 */ 305 close(child_ready_pipe[1]); 306 307 /* 308 * Wait until the parent tells us to go. 309 */ 310 if (read(go_pipe[0], &buf, 1) == -1) 311 perror("unable to read pipe"); 312 313 execvp(argv[0], (char **)argv); 314 315 perror(argv[0]); 316 exit(-1); 317 } 318 319 if (perf_target__none(&target)) 320 evsel_list->threads->map[0] = child_pid; 321 322 /* 323 * Wait for the child to be ready to exec. 324 */ 325 close(child_ready_pipe[1]); 326 close(go_pipe[0]); 327 if (read(child_ready_pipe[0], &buf, 1) == -1) 328 perror("unable to read pipe"); 329 close(child_ready_pipe[0]); 330 } 331 332 if (group) 333 perf_evlist__set_leader(evsel_list); 334 335 list_for_each_entry(counter, &evsel_list->entries, node) { 336 if (create_perf_stat_counter(counter) < 0) { 337 /* 338 * PPC returns ENXIO for HW counters until 2.6.37 339 * (behavior changed with commit b0a873e). 340 */ 341 if (errno == EINVAL || errno == ENOSYS || 342 errno == ENOENT || errno == EOPNOTSUPP || 343 errno == ENXIO) { 344 if (verbose) 345 ui__warning("%s event is not supported by the kernel.\n", 346 perf_evsel__name(counter)); 347 counter->supported = false; 348 continue; 349 } 350 351 if (errno == EPERM || errno == EACCES) { 352 error("You may not have permission to collect %sstats.\n" 353 "\t Consider tweaking" 354 " /proc/sys/kernel/perf_event_paranoid or running as root.", 355 target.system_wide ? "system-wide " : ""); 356 } else { 357 error("open_counter returned with %d (%s). " 358 "/bin/dmesg may provide additional information.\n", 359 errno, strerror(errno)); 360 } 361 if (child_pid != -1) 362 kill(child_pid, SIGTERM); 363 364 pr_err("Not all events could be opened.\n"); 365 return -1; 366 } 367 counter->supported = true; 368 } 369 370 if (perf_evlist__apply_filters(evsel_list)) { 371 error("failed to set filter with %d (%s)\n", errno, 372 strerror(errno)); 373 return -1; 374 } 375 376 /* 377 * Enable counters and exec the command: 378 */ 379 t0 = rdclock(); 380 381 if (forks) { 382 close(go_pipe[1]); 383 wait(&status); 384 if (WIFSIGNALED(status)) 385 psignal(WTERMSIG(status), argv[0]); 386 } else { 387 while(!done) sleep(1); 388 } 389 390 t1 = rdclock(); 391 392 update_stats(&walltime_nsecs_stats, t1 - t0); 393 394 if (no_aggr) { 395 list_for_each_entry(counter, &evsel_list->entries, node) { 396 read_counter(counter); 397 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1); 398 } 399 } else { 400 list_for_each_entry(counter, &evsel_list->entries, node) { 401 read_counter_aggr(counter); 402 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 403 evsel_list->threads->nr); 404 } 405 } 406 407 return WEXITSTATUS(status); 408 } 409 410 static int run_perf_stat(int argc __maybe_unused, const char **argv) 411 { 412 int ret; 413 414 if (pre_cmd) { 415 ret = system(pre_cmd); 416 if (ret) 417 return ret; 418 } 419 420 if (sync_run) 421 sync(); 422 423 ret = __run_perf_stat(argc, argv); 424 if (ret) 425 return ret; 426 427 if (post_cmd) { 428 ret = system(post_cmd); 429 if (ret) 430 return ret; 431 } 432 433 return ret; 434 } 435 436 static void print_noise_pct(double total, double avg) 437 { 438 double pct = rel_stddev_stats(total, avg); 439 440 if (csv_output) 441 fprintf(output, "%s%.2f%%", csv_sep, pct); 442 else if (pct) 443 fprintf(output, " ( +-%6.2f%% )", pct); 444 } 445 446 static void print_noise(struct perf_evsel *evsel, double avg) 447 { 448 struct perf_stat *ps; 449 450 if (run_count == 1) 451 return; 452 453 ps = evsel->priv; 454 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); 455 } 456 457 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) 458 { 459 double msecs = avg / 1e6; 460 char cpustr[16] = { '\0', }; 461 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; 462 463 if (no_aggr) 464 sprintf(cpustr, "CPU%*d%s", 465 csv_output ? 0 : -4, 466 perf_evsel__cpus(evsel)->map[cpu], csv_sep); 467 468 fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel)); 469 470 if (evsel->cgrp) 471 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); 472 473 if (csv_output) 474 return; 475 476 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 477 fprintf(output, " # %8.3f CPUs utilized ", 478 avg / avg_stats(&walltime_nsecs_stats)); 479 else 480 fprintf(output, " "); 481 } 482 483 /* used for get_ratio_color() */ 484 enum grc_type { 485 GRC_STALLED_CYCLES_FE, 486 GRC_STALLED_CYCLES_BE, 487 GRC_CACHE_MISSES, 488 GRC_MAX_NR 489 }; 490 491 static const char *get_ratio_color(enum grc_type type, double ratio) 492 { 493 static const double grc_table[GRC_MAX_NR][3] = { 494 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, 495 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, 496 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, 497 }; 498 const char *color = PERF_COLOR_NORMAL; 499 500 if (ratio > grc_table[type][0]) 501 color = PERF_COLOR_RED; 502 else if (ratio > grc_table[type][1]) 503 color = PERF_COLOR_MAGENTA; 504 else if (ratio > grc_table[type][2]) 505 color = PERF_COLOR_YELLOW; 506 507 return color; 508 } 509 510 static void print_stalled_cycles_frontend(int cpu, 511 struct perf_evsel *evsel 512 __maybe_unused, double avg) 513 { 514 double total, ratio = 0.0; 515 const char *color; 516 517 total = avg_stats(&runtime_cycles_stats[cpu]); 518 519 if (total) 520 ratio = avg / total * 100.0; 521 522 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); 523 524 fprintf(output, " # "); 525 color_fprintf(output, color, "%6.2f%%", ratio); 526 fprintf(output, " frontend cycles idle "); 527 } 528 529 static void print_stalled_cycles_backend(int cpu, 530 struct perf_evsel *evsel 531 __maybe_unused, double avg) 532 { 533 double total, ratio = 0.0; 534 const char *color; 535 536 total = avg_stats(&runtime_cycles_stats[cpu]); 537 538 if (total) 539 ratio = avg / total * 100.0; 540 541 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); 542 543 fprintf(output, " # "); 544 color_fprintf(output, color, "%6.2f%%", ratio); 545 fprintf(output, " backend cycles idle "); 546 } 547 548 static void print_branch_misses(int cpu, 549 struct perf_evsel *evsel __maybe_unused, 550 double avg) 551 { 552 double total, ratio = 0.0; 553 const char *color; 554 555 total = avg_stats(&runtime_branches_stats[cpu]); 556 557 if (total) 558 ratio = avg / total * 100.0; 559 560 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 561 562 fprintf(output, " # "); 563 color_fprintf(output, color, "%6.2f%%", ratio); 564 fprintf(output, " of all branches "); 565 } 566 567 static void print_l1_dcache_misses(int cpu, 568 struct perf_evsel *evsel __maybe_unused, 569 double avg) 570 { 571 double total, ratio = 0.0; 572 const char *color; 573 574 total = avg_stats(&runtime_l1_dcache_stats[cpu]); 575 576 if (total) 577 ratio = avg / total * 100.0; 578 579 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 580 581 fprintf(output, " # "); 582 color_fprintf(output, color, "%6.2f%%", ratio); 583 fprintf(output, " of all L1-dcache hits "); 584 } 585 586 static void print_l1_icache_misses(int cpu, 587 struct perf_evsel *evsel __maybe_unused, 588 double avg) 589 { 590 double total, ratio = 0.0; 591 const char *color; 592 593 total = avg_stats(&runtime_l1_icache_stats[cpu]); 594 595 if (total) 596 ratio = avg / total * 100.0; 597 598 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 599 600 fprintf(output, " # "); 601 color_fprintf(output, color, "%6.2f%%", ratio); 602 fprintf(output, " of all L1-icache hits "); 603 } 604 605 static void print_dtlb_cache_misses(int cpu, 606 struct perf_evsel *evsel __maybe_unused, 607 double avg) 608 { 609 double total, ratio = 0.0; 610 const char *color; 611 612 total = avg_stats(&runtime_dtlb_cache_stats[cpu]); 613 614 if (total) 615 ratio = avg / total * 100.0; 616 617 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 618 619 fprintf(output, " # "); 620 color_fprintf(output, color, "%6.2f%%", ratio); 621 fprintf(output, " of all dTLB cache hits "); 622 } 623 624 static void print_itlb_cache_misses(int cpu, 625 struct perf_evsel *evsel __maybe_unused, 626 double avg) 627 { 628 double total, ratio = 0.0; 629 const char *color; 630 631 total = avg_stats(&runtime_itlb_cache_stats[cpu]); 632 633 if (total) 634 ratio = avg / total * 100.0; 635 636 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 637 638 fprintf(output, " # "); 639 color_fprintf(output, color, "%6.2f%%", ratio); 640 fprintf(output, " of all iTLB cache hits "); 641 } 642 643 static void print_ll_cache_misses(int cpu, 644 struct perf_evsel *evsel __maybe_unused, 645 double avg) 646 { 647 double total, ratio = 0.0; 648 const char *color; 649 650 total = avg_stats(&runtime_ll_cache_stats[cpu]); 651 652 if (total) 653 ratio = avg / total * 100.0; 654 655 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 656 657 fprintf(output, " # "); 658 color_fprintf(output, color, "%6.2f%%", ratio); 659 fprintf(output, " of all LL-cache hits "); 660 } 661 662 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) 663 { 664 double total, ratio = 0.0; 665 char cpustr[16] = { '\0', }; 666 const char *fmt; 667 668 if (csv_output) 669 fmt = "%s%.0f%s%s"; 670 else if (big_num) 671 fmt = "%s%'18.0f%s%-25s"; 672 else 673 fmt = "%s%18.0f%s%-25s"; 674 675 if (no_aggr) 676 sprintf(cpustr, "CPU%*d%s", 677 csv_output ? 0 : -4, 678 perf_evsel__cpus(evsel)->map[cpu], csv_sep); 679 else 680 cpu = 0; 681 682 fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel)); 683 684 if (evsel->cgrp) 685 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); 686 687 if (csv_output) 688 return; 689 690 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { 691 total = avg_stats(&runtime_cycles_stats[cpu]); 692 693 if (total) 694 ratio = avg / total; 695 696 fprintf(output, " # %5.2f insns per cycle ", ratio); 697 698 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); 699 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); 700 701 if (total && avg) { 702 ratio = total / avg; 703 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio); 704 } 705 706 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && 707 runtime_branches_stats[cpu].n != 0) { 708 print_branch_misses(cpu, evsel, avg); 709 } else if ( 710 evsel->attr.type == PERF_TYPE_HW_CACHE && 711 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | 712 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 713 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 714 runtime_l1_dcache_stats[cpu].n != 0) { 715 print_l1_dcache_misses(cpu, evsel, avg); 716 } else if ( 717 evsel->attr.type == PERF_TYPE_HW_CACHE && 718 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | 719 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 720 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 721 runtime_l1_icache_stats[cpu].n != 0) { 722 print_l1_icache_misses(cpu, evsel, avg); 723 } else if ( 724 evsel->attr.type == PERF_TYPE_HW_CACHE && 725 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | 726 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 727 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 728 runtime_dtlb_cache_stats[cpu].n != 0) { 729 print_dtlb_cache_misses(cpu, evsel, avg); 730 } else if ( 731 evsel->attr.type == PERF_TYPE_HW_CACHE && 732 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | 733 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 734 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 735 runtime_itlb_cache_stats[cpu].n != 0) { 736 print_itlb_cache_misses(cpu, evsel, avg); 737 } else if ( 738 evsel->attr.type == PERF_TYPE_HW_CACHE && 739 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | 740 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 741 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 742 runtime_ll_cache_stats[cpu].n != 0) { 743 print_ll_cache_misses(cpu, evsel, avg); 744 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && 745 runtime_cacherefs_stats[cpu].n != 0) { 746 total = avg_stats(&runtime_cacherefs_stats[cpu]); 747 748 if (total) 749 ratio = avg * 100 / total; 750 751 fprintf(output, " # %8.3f %% of all cache refs ", ratio); 752 753 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { 754 print_stalled_cycles_frontend(cpu, evsel, avg); 755 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { 756 print_stalled_cycles_backend(cpu, evsel, avg); 757 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { 758 total = avg_stats(&runtime_nsecs_stats[cpu]); 759 760 if (total) 761 ratio = 1.0 * avg / total; 762 763 fprintf(output, " # %8.3f GHz ", ratio); 764 } else if (runtime_nsecs_stats[cpu].n != 0) { 765 char unit = 'M'; 766 767 total = avg_stats(&runtime_nsecs_stats[cpu]); 768 769 if (total) 770 ratio = 1000.0 * avg / total; 771 if (ratio < 0.001) { 772 ratio *= 1000; 773 unit = 'K'; 774 } 775 776 fprintf(output, " # %8.3f %c/sec ", ratio, unit); 777 } else { 778 fprintf(output, " "); 779 } 780 } 781 782 /* 783 * Print out the results of a single counter: 784 * aggregated counts in system-wide mode 785 */ 786 static void print_counter_aggr(struct perf_evsel *counter) 787 { 788 struct perf_stat *ps = counter->priv; 789 double avg = avg_stats(&ps->res_stats[0]); 790 int scaled = counter->counts->scaled; 791 792 if (scaled == -1) { 793 fprintf(output, "%*s%s%*s", 794 csv_output ? 0 : 18, 795 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 796 csv_sep, 797 csv_output ? 0 : -24, 798 perf_evsel__name(counter)); 799 800 if (counter->cgrp) 801 fprintf(output, "%s%s", csv_sep, counter->cgrp->name); 802 803 fputc('\n', output); 804 return; 805 } 806 807 if (nsec_counter(counter)) 808 nsec_printout(-1, counter, avg); 809 else 810 abs_printout(-1, counter, avg); 811 812 print_noise(counter, avg); 813 814 if (csv_output) { 815 fputc('\n', output); 816 return; 817 } 818 819 if (scaled) { 820 double avg_enabled, avg_running; 821 822 avg_enabled = avg_stats(&ps->res_stats[1]); 823 avg_running = avg_stats(&ps->res_stats[2]); 824 825 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); 826 } 827 fprintf(output, "\n"); 828 } 829 830 /* 831 * Print out the results of a single counter: 832 * does not use aggregated count in system-wide 833 */ 834 static void print_counter(struct perf_evsel *counter) 835 { 836 u64 ena, run, val; 837 int cpu; 838 839 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { 840 val = counter->counts->cpu[cpu].val; 841 ena = counter->counts->cpu[cpu].ena; 842 run = counter->counts->cpu[cpu].run; 843 if (run == 0 || ena == 0) { 844 fprintf(output, "CPU%*d%s%*s%s%*s", 845 csv_output ? 0 : -4, 846 perf_evsel__cpus(counter)->map[cpu], csv_sep, 847 csv_output ? 0 : 18, 848 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 849 csv_sep, 850 csv_output ? 0 : -24, 851 perf_evsel__name(counter)); 852 853 if (counter->cgrp) 854 fprintf(output, "%s%s", 855 csv_sep, counter->cgrp->name); 856 857 fputc('\n', output); 858 continue; 859 } 860 861 if (nsec_counter(counter)) 862 nsec_printout(cpu, counter, val); 863 else 864 abs_printout(cpu, counter, val); 865 866 if (!csv_output) { 867 print_noise(counter, 1.0); 868 869 if (run != ena) 870 fprintf(output, " (%.2f%%)", 871 100.0 * run / ena); 872 } 873 fputc('\n', output); 874 } 875 } 876 877 static void print_stat(int argc, const char **argv) 878 { 879 struct perf_evsel *counter; 880 int i; 881 882 fflush(stdout); 883 884 if (!csv_output) { 885 fprintf(output, "\n"); 886 fprintf(output, " Performance counter stats for "); 887 if (!perf_target__has_task(&target)) { 888 fprintf(output, "\'%s", argv[0]); 889 for (i = 1; i < argc; i++) 890 fprintf(output, " %s", argv[i]); 891 } else if (target.pid) 892 fprintf(output, "process id \'%s", target.pid); 893 else 894 fprintf(output, "thread id \'%s", target.tid); 895 896 fprintf(output, "\'"); 897 if (run_count > 1) 898 fprintf(output, " (%d runs)", run_count); 899 fprintf(output, ":\n\n"); 900 } 901 902 if (no_aggr) { 903 list_for_each_entry(counter, &evsel_list->entries, node) 904 print_counter(counter); 905 } else { 906 list_for_each_entry(counter, &evsel_list->entries, node) 907 print_counter_aggr(counter); 908 } 909 910 if (!csv_output) { 911 if (!null_run) 912 fprintf(output, "\n"); 913 fprintf(output, " %17.9f seconds time elapsed", 914 avg_stats(&walltime_nsecs_stats)/1e9); 915 if (run_count > 1) { 916 fprintf(output, " "); 917 print_noise_pct(stddev_stats(&walltime_nsecs_stats), 918 avg_stats(&walltime_nsecs_stats)); 919 } 920 fprintf(output, "\n\n"); 921 } 922 } 923 924 static volatile int signr = -1; 925 926 static void skip_signal(int signo) 927 { 928 if(child_pid == -1) 929 done = 1; 930 931 signr = signo; 932 } 933 934 static void sig_atexit(void) 935 { 936 if (child_pid != -1) 937 kill(child_pid, SIGTERM); 938 939 if (signr == -1) 940 return; 941 942 signal(signr, SIG_DFL); 943 kill(getpid(), signr); 944 } 945 946 static int stat__set_big_num(const struct option *opt __maybe_unused, 947 const char *s __maybe_unused, int unset) 948 { 949 big_num_opt = unset ? 0 : 1; 950 return 0; 951 } 952 953 /* 954 * Add default attributes, if there were no attributes specified or 955 * if -d/--detailed, -d -d or -d -d -d is used: 956 */ 957 static int add_default_attributes(void) 958 { 959 struct perf_event_attr default_attrs[] = { 960 961 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 962 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 963 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 964 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 965 966 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 967 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 968 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 969 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 970 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 971 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 972 973 }; 974 975 /* 976 * Detailed stats (-d), covering the L1 and last level data caches: 977 */ 978 struct perf_event_attr detailed_attrs[] = { 979 980 { .type = PERF_TYPE_HW_CACHE, 981 .config = 982 PERF_COUNT_HW_CACHE_L1D << 0 | 983 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 984 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 985 986 { .type = PERF_TYPE_HW_CACHE, 987 .config = 988 PERF_COUNT_HW_CACHE_L1D << 0 | 989 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 990 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 991 992 { .type = PERF_TYPE_HW_CACHE, 993 .config = 994 PERF_COUNT_HW_CACHE_LL << 0 | 995 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 996 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 997 998 { .type = PERF_TYPE_HW_CACHE, 999 .config = 1000 PERF_COUNT_HW_CACHE_LL << 0 | 1001 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1002 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1003 }; 1004 1005 /* 1006 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 1007 */ 1008 struct perf_event_attr very_detailed_attrs[] = { 1009 1010 { .type = PERF_TYPE_HW_CACHE, 1011 .config = 1012 PERF_COUNT_HW_CACHE_L1I << 0 | 1013 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1014 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1015 1016 { .type = PERF_TYPE_HW_CACHE, 1017 .config = 1018 PERF_COUNT_HW_CACHE_L1I << 0 | 1019 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1020 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1021 1022 { .type = PERF_TYPE_HW_CACHE, 1023 .config = 1024 PERF_COUNT_HW_CACHE_DTLB << 0 | 1025 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1026 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1027 1028 { .type = PERF_TYPE_HW_CACHE, 1029 .config = 1030 PERF_COUNT_HW_CACHE_DTLB << 0 | 1031 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1032 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1033 1034 { .type = PERF_TYPE_HW_CACHE, 1035 .config = 1036 PERF_COUNT_HW_CACHE_ITLB << 0 | 1037 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1038 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1039 1040 { .type = PERF_TYPE_HW_CACHE, 1041 .config = 1042 PERF_COUNT_HW_CACHE_ITLB << 0 | 1043 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1044 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1045 1046 }; 1047 1048 /* 1049 * Very, very detailed stats (-d -d -d), adding prefetch events: 1050 */ 1051 struct perf_event_attr very_very_detailed_attrs[] = { 1052 1053 { .type = PERF_TYPE_HW_CACHE, 1054 .config = 1055 PERF_COUNT_HW_CACHE_L1D << 0 | 1056 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1057 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1058 1059 { .type = PERF_TYPE_HW_CACHE, 1060 .config = 1061 PERF_COUNT_HW_CACHE_L1D << 0 | 1062 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1063 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1064 }; 1065 1066 /* Set attrs if no event is selected and !null_run: */ 1067 if (null_run) 1068 return 0; 1069 1070 if (!evsel_list->nr_entries) { 1071 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) 1072 return -1; 1073 } 1074 1075 /* Detailed events get appended to the event list: */ 1076 1077 if (detailed_run < 1) 1078 return 0; 1079 1080 /* Append detailed run extra attributes: */ 1081 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) 1082 return -1; 1083 1084 if (detailed_run < 2) 1085 return 0; 1086 1087 /* Append very detailed run extra attributes: */ 1088 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) 1089 return -1; 1090 1091 if (detailed_run < 3) 1092 return 0; 1093 1094 /* Append very, very detailed run extra attributes: */ 1095 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); 1096 } 1097 1098 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) 1099 { 1100 bool append_file = false; 1101 int output_fd = 0; 1102 const char *output_name = NULL; 1103 const struct option options[] = { 1104 OPT_CALLBACK('e', "event", &evsel_list, "event", 1105 "event selector. use 'perf list' to list available events", 1106 parse_events_option), 1107 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 1108 "event filter", parse_filter), 1109 OPT_BOOLEAN('i', "no-inherit", &no_inherit, 1110 "child tasks do not inherit counters"), 1111 OPT_STRING('p', "pid", &target.pid, "pid", 1112 "stat events on existing process id"), 1113 OPT_STRING('t', "tid", &target.tid, "tid", 1114 "stat events on existing thread id"), 1115 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 1116 "system-wide collection from all CPUs"), 1117 OPT_BOOLEAN('g', "group", &group, 1118 "put the counters into a counter group"), 1119 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"), 1120 OPT_INCR('v', "verbose", &verbose, 1121 "be more verbose (show counter open errors, etc)"), 1122 OPT_INTEGER('r', "repeat", &run_count, 1123 "repeat command and print average + stddev (max: 100)"), 1124 OPT_BOOLEAN('n', "null", &null_run, 1125 "null run - dont start any counters"), 1126 OPT_INCR('d', "detailed", &detailed_run, 1127 "detailed run - start a lot of events"), 1128 OPT_BOOLEAN('S', "sync", &sync_run, 1129 "call sync() before starting a run"), 1130 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1131 "print large numbers with thousands\' separators", 1132 stat__set_big_num), 1133 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 1134 "list of cpus to monitor in system-wide"), 1135 OPT_BOOLEAN('A', "no-aggr", &no_aggr, "disable CPU count aggregation"), 1136 OPT_STRING('x', "field-separator", &csv_sep, "separator", 1137 "print counts with custom separator"), 1138 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 1139 "monitor event in cgroup name only", parse_cgroups), 1140 OPT_STRING('o', "output", &output_name, "file", "output file name"), 1141 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 1142 OPT_INTEGER(0, "log-fd", &output_fd, 1143 "log output to fd, instead of stderr"), 1144 OPT_STRING(0, "pre", &pre_cmd, "command", 1145 "command to run prior to the measured command"), 1146 OPT_STRING(0, "post", &post_cmd, "command", 1147 "command to run after to the measured command"), 1148 OPT_END() 1149 }; 1150 const char * const stat_usage[] = { 1151 "perf stat [<options>] [<command>]", 1152 NULL 1153 }; 1154 struct perf_evsel *pos; 1155 int status = -ENOMEM, run_idx; 1156 const char *mode; 1157 1158 setlocale(LC_ALL, ""); 1159 1160 evsel_list = perf_evlist__new(NULL, NULL); 1161 if (evsel_list == NULL) 1162 return -ENOMEM; 1163 1164 argc = parse_options(argc, argv, options, stat_usage, 1165 PARSE_OPT_STOP_AT_NON_OPTION); 1166 1167 output = stderr; 1168 if (output_name && strcmp(output_name, "-")) 1169 output = NULL; 1170 1171 if (output_name && output_fd) { 1172 fprintf(stderr, "cannot use both --output and --log-fd\n"); 1173 usage_with_options(stat_usage, options); 1174 } 1175 1176 if (output_fd < 0) { 1177 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 1178 usage_with_options(stat_usage, options); 1179 } 1180 1181 if (!output) { 1182 struct timespec tm; 1183 mode = append_file ? "a" : "w"; 1184 1185 output = fopen(output_name, mode); 1186 if (!output) { 1187 perror("failed to create output file"); 1188 return -1; 1189 } 1190 clock_gettime(CLOCK_REALTIME, &tm); 1191 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 1192 } else if (output_fd > 0) { 1193 mode = append_file ? "a" : "w"; 1194 output = fdopen(output_fd, mode); 1195 if (!output) { 1196 perror("Failed opening logfd"); 1197 return -errno; 1198 } 1199 } 1200 1201 if (csv_sep) { 1202 csv_output = true; 1203 if (!strcmp(csv_sep, "\\t")) 1204 csv_sep = "\t"; 1205 } else 1206 csv_sep = DEFAULT_SEPARATOR; 1207 1208 /* 1209 * let the spreadsheet do the pretty-printing 1210 */ 1211 if (csv_output) { 1212 /* User explicitly passed -B? */ 1213 if (big_num_opt == 1) { 1214 fprintf(stderr, "-B option not supported with -x\n"); 1215 usage_with_options(stat_usage, options); 1216 } else /* Nope, so disable big number formatting */ 1217 big_num = false; 1218 } else if (big_num_opt == 0) /* User passed --no-big-num */ 1219 big_num = false; 1220 1221 if (!argc && !perf_target__has_task(&target)) 1222 usage_with_options(stat_usage, options); 1223 if (run_count <= 0) 1224 usage_with_options(stat_usage, options); 1225 1226 /* no_aggr, cgroup are for system-wide only */ 1227 if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) { 1228 fprintf(stderr, "both cgroup and no-aggregation " 1229 "modes only available in system-wide mode\n"); 1230 1231 usage_with_options(stat_usage, options); 1232 } 1233 1234 if (add_default_attributes()) 1235 goto out; 1236 1237 perf_target__validate(&target); 1238 1239 if (perf_evlist__create_maps(evsel_list, &target) < 0) { 1240 if (perf_target__has_task(&target)) 1241 pr_err("Problems finding threads of monitor\n"); 1242 if (perf_target__has_cpu(&target)) 1243 perror("failed to parse CPUs map"); 1244 1245 usage_with_options(stat_usage, options); 1246 return -1; 1247 } 1248 1249 list_for_each_entry(pos, &evsel_list->entries, node) { 1250 if (perf_evsel__alloc_stat_priv(pos) < 0 || 1251 perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0) 1252 goto out_free_fd; 1253 } 1254 1255 /* 1256 * We dont want to block the signals - that would cause 1257 * child tasks to inherit that and Ctrl-C would not work. 1258 * What we want is for Ctrl-C to work in the exec()-ed 1259 * task, but being ignored by perf stat itself: 1260 */ 1261 atexit(sig_atexit); 1262 signal(SIGINT, skip_signal); 1263 signal(SIGALRM, skip_signal); 1264 signal(SIGABRT, skip_signal); 1265 1266 status = 0; 1267 for (run_idx = 0; run_idx < run_count; run_idx++) { 1268 if (run_count != 1 && verbose) 1269 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 1270 run_idx + 1); 1271 1272 status = run_perf_stat(argc, argv); 1273 } 1274 1275 if (status != -1) 1276 print_stat(argc, argv); 1277 out_free_fd: 1278 list_for_each_entry(pos, &evsel_list->entries, node) 1279 perf_evsel__free_stat_priv(pos); 1280 perf_evlist__delete_maps(evsel_list); 1281 out: 1282 perf_evlist__delete(evsel_list); 1283 return status; 1284 } 1285