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