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 const char *target_pid; 186 static const char *target_tid; 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 bool exclude_guest_missing = false; 287 int ret; 288 289 if (group && evsel != first) 290 group_fd = first->fd; 291 292 if (scale) 293 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 294 PERF_FORMAT_TOTAL_TIME_RUNNING; 295 296 attr->inherit = !no_inherit; 297 298 retry: 299 if (exclude_guest_missing) 300 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0; 301 302 if (system_wide) { 303 ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus, 304 group, group_fd); 305 if (ret) 306 goto check_ret; 307 return 0; 308 } 309 310 if (!target_pid && !target_tid && (!group || evsel == first)) { 311 attr->disabled = 1; 312 attr->enable_on_exec = 1; 313 } 314 315 ret = perf_evsel__open_per_thread(evsel, evsel_list->threads, 316 group, group_fd); 317 if (!ret) 318 return 0; 319 /* fall through */ 320 check_ret: 321 if (ret && errno == EINVAL) { 322 if (!exclude_guest_missing && 323 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) { 324 pr_debug("Old kernel, cannot exclude " 325 "guest or host samples.\n"); 326 exclude_guest_missing = true; 327 goto retry; 328 } 329 } 330 return ret; 331 } 332 333 /* 334 * Does the counter have nsecs as a unit? 335 */ 336 static inline int nsec_counter(struct perf_evsel *evsel) 337 { 338 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || 339 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 340 return 1; 341 342 return 0; 343 } 344 345 /* 346 * Update various tracking values we maintain to print 347 * more semantic information such as miss/hit ratios, 348 * instruction rates, etc: 349 */ 350 static void update_shadow_stats(struct perf_evsel *counter, u64 *count) 351 { 352 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) 353 update_stats(&runtime_nsecs_stats[0], count[0]); 354 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) 355 update_stats(&runtime_cycles_stats[0], count[0]); 356 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) 357 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); 358 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) 359 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); 360 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) 361 update_stats(&runtime_branches_stats[0], count[0]); 362 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) 363 update_stats(&runtime_cacherefs_stats[0], count[0]); 364 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) 365 update_stats(&runtime_l1_dcache_stats[0], count[0]); 366 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) 367 update_stats(&runtime_l1_icache_stats[0], count[0]); 368 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) 369 update_stats(&runtime_ll_cache_stats[0], count[0]); 370 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) 371 update_stats(&runtime_dtlb_cache_stats[0], count[0]); 372 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) 373 update_stats(&runtime_itlb_cache_stats[0], count[0]); 374 } 375 376 /* 377 * Read out the results of a single counter: 378 * aggregate counts across CPUs in system-wide mode 379 */ 380 static int read_counter_aggr(struct perf_evsel *counter) 381 { 382 struct perf_stat *ps = counter->priv; 383 u64 *count = counter->counts->aggr.values; 384 int i; 385 386 if (__perf_evsel__read(counter, evsel_list->cpus->nr, 387 evsel_list->threads->nr, scale) < 0) 388 return -1; 389 390 for (i = 0; i < 3; i++) 391 update_stats(&ps->res_stats[i], count[i]); 392 393 if (verbose) { 394 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 395 event_name(counter), count[0], count[1], count[2]); 396 } 397 398 /* 399 * Save the full runtime - to allow normalization during printout: 400 */ 401 update_shadow_stats(counter, count); 402 403 return 0; 404 } 405 406 /* 407 * Read out the results of a single counter: 408 * do not aggregate counts across CPUs in system-wide mode 409 */ 410 static int read_counter(struct perf_evsel *counter) 411 { 412 u64 *count; 413 int cpu; 414 415 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { 416 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) 417 return -1; 418 419 count = counter->counts->cpu[cpu].values; 420 421 update_shadow_stats(counter, count); 422 } 423 424 return 0; 425 } 426 427 static int run_perf_stat(int argc __used, const char **argv) 428 { 429 unsigned long long t0, t1; 430 struct perf_evsel *counter, *first; 431 int status = 0; 432 int child_ready_pipe[2], go_pipe[2]; 433 const bool forks = (argc > 0); 434 char buf; 435 436 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 437 perror("failed to create pipes"); 438 exit(1); 439 } 440 441 if (forks) { 442 if ((child_pid = fork()) < 0) 443 perror("failed to fork"); 444 445 if (!child_pid) { 446 close(child_ready_pipe[0]); 447 close(go_pipe[1]); 448 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 449 450 /* 451 * Do a dummy execvp to get the PLT entry resolved, 452 * so we avoid the resolver overhead on the real 453 * execvp call. 454 */ 455 execvp("", (char **)argv); 456 457 /* 458 * Tell the parent we're ready to go 459 */ 460 close(child_ready_pipe[1]); 461 462 /* 463 * Wait until the parent tells us to go. 464 */ 465 if (read(go_pipe[0], &buf, 1) == -1) 466 perror("unable to read pipe"); 467 468 execvp(argv[0], (char **)argv); 469 470 perror(argv[0]); 471 exit(-1); 472 } 473 474 if (!target_tid && !target_pid && !system_wide) 475 evsel_list->threads->map[0] = child_pid; 476 477 /* 478 * Wait for the child to be ready to exec. 479 */ 480 close(child_ready_pipe[1]); 481 close(go_pipe[0]); 482 if (read(child_ready_pipe[0], &buf, 1) == -1) 483 perror("unable to read pipe"); 484 close(child_ready_pipe[0]); 485 } 486 487 first = list_entry(evsel_list->entries.next, struct perf_evsel, node); 488 489 list_for_each_entry(counter, &evsel_list->entries, node) { 490 if (create_perf_stat_counter(counter, first) < 0) { 491 /* 492 * PPC returns ENXIO for HW counters until 2.6.37 493 * (behavior changed with commit b0a873e). 494 */ 495 if (errno == EINVAL || errno == ENOSYS || 496 errno == ENOENT || errno == EOPNOTSUPP || 497 errno == ENXIO) { 498 if (verbose) 499 ui__warning("%s event is not supported by the kernel.\n", 500 event_name(counter)); 501 counter->supported = false; 502 continue; 503 } 504 505 if (errno == EPERM || errno == EACCES) { 506 error("You may not have permission to collect %sstats.\n" 507 "\t Consider tweaking" 508 " /proc/sys/kernel/perf_event_paranoid or running as root.", 509 system_wide ? "system-wide " : ""); 510 } else { 511 error("open_counter returned with %d (%s). " 512 "/bin/dmesg may provide additional information.\n", 513 errno, strerror(errno)); 514 } 515 if (child_pid != -1) 516 kill(child_pid, SIGTERM); 517 die("Not all events could be opened.\n"); 518 return -1; 519 } 520 counter->supported = true; 521 } 522 523 if (perf_evlist__set_filters(evsel_list)) { 524 error("failed to set filter with %d (%s)\n", errno, 525 strerror(errno)); 526 return -1; 527 } 528 529 /* 530 * Enable counters and exec the command: 531 */ 532 t0 = rdclock(); 533 534 if (forks) { 535 close(go_pipe[1]); 536 wait(&status); 537 if (WIFSIGNALED(status)) 538 psignal(WTERMSIG(status), argv[0]); 539 } else { 540 while(!done) sleep(1); 541 } 542 543 t1 = rdclock(); 544 545 update_stats(&walltime_nsecs_stats, t1 - t0); 546 547 if (no_aggr) { 548 list_for_each_entry(counter, &evsel_list->entries, node) { 549 read_counter(counter); 550 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1); 551 } 552 } else { 553 list_for_each_entry(counter, &evsel_list->entries, node) { 554 read_counter_aggr(counter); 555 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 556 evsel_list->threads->nr); 557 } 558 } 559 560 return WEXITSTATUS(status); 561 } 562 563 static void print_noise_pct(double total, double avg) 564 { 565 double pct = 0.0; 566 567 if (avg) 568 pct = 100.0*total/avg; 569 570 if (csv_output) 571 fprintf(output, "%s%.2f%%", csv_sep, pct); 572 else if (pct) 573 fprintf(output, " ( +-%6.2f%% )", pct); 574 } 575 576 static void print_noise(struct perf_evsel *evsel, double avg) 577 { 578 struct perf_stat *ps; 579 580 if (run_count == 1) 581 return; 582 583 ps = evsel->priv; 584 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); 585 } 586 587 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) 588 { 589 double msecs = avg / 1e6; 590 char cpustr[16] = { '\0', }; 591 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; 592 593 if (no_aggr) 594 sprintf(cpustr, "CPU%*d%s", 595 csv_output ? 0 : -4, 596 evsel_list->cpus->map[cpu], csv_sep); 597 598 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel)); 599 600 if (evsel->cgrp) 601 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); 602 603 if (csv_output) 604 return; 605 606 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 607 fprintf(output, " # %8.3f CPUs utilized ", 608 avg / avg_stats(&walltime_nsecs_stats)); 609 else 610 fprintf(output, " "); 611 } 612 613 /* used for get_ratio_color() */ 614 enum grc_type { 615 GRC_STALLED_CYCLES_FE, 616 GRC_STALLED_CYCLES_BE, 617 GRC_CACHE_MISSES, 618 GRC_MAX_NR 619 }; 620 621 static const char *get_ratio_color(enum grc_type type, double ratio) 622 { 623 static const double grc_table[GRC_MAX_NR][3] = { 624 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, 625 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, 626 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, 627 }; 628 const char *color = PERF_COLOR_NORMAL; 629 630 if (ratio > grc_table[type][0]) 631 color = PERF_COLOR_RED; 632 else if (ratio > grc_table[type][1]) 633 color = PERF_COLOR_MAGENTA; 634 else if (ratio > grc_table[type][2]) 635 color = PERF_COLOR_YELLOW; 636 637 return color; 638 } 639 640 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg) 641 { 642 double total, ratio = 0.0; 643 const char *color; 644 645 total = avg_stats(&runtime_cycles_stats[cpu]); 646 647 if (total) 648 ratio = avg / total * 100.0; 649 650 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); 651 652 fprintf(output, " # "); 653 color_fprintf(output, color, "%6.2f%%", ratio); 654 fprintf(output, " frontend cycles idle "); 655 } 656 657 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg) 658 { 659 double total, ratio = 0.0; 660 const char *color; 661 662 total = avg_stats(&runtime_cycles_stats[cpu]); 663 664 if (total) 665 ratio = avg / total * 100.0; 666 667 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); 668 669 fprintf(output, " # "); 670 color_fprintf(output, color, "%6.2f%%", ratio); 671 fprintf(output, " backend cycles idle "); 672 } 673 674 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg) 675 { 676 double total, ratio = 0.0; 677 const char *color; 678 679 total = avg_stats(&runtime_branches_stats[cpu]); 680 681 if (total) 682 ratio = avg / total * 100.0; 683 684 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 685 686 fprintf(output, " # "); 687 color_fprintf(output, color, "%6.2f%%", ratio); 688 fprintf(output, " of all branches "); 689 } 690 691 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 692 { 693 double total, ratio = 0.0; 694 const char *color; 695 696 total = avg_stats(&runtime_l1_dcache_stats[cpu]); 697 698 if (total) 699 ratio = avg / total * 100.0; 700 701 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 702 703 fprintf(output, " # "); 704 color_fprintf(output, color, "%6.2f%%", ratio); 705 fprintf(output, " of all L1-dcache hits "); 706 } 707 708 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 709 { 710 double total, ratio = 0.0; 711 const char *color; 712 713 total = avg_stats(&runtime_l1_icache_stats[cpu]); 714 715 if (total) 716 ratio = avg / total * 100.0; 717 718 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 719 720 fprintf(output, " # "); 721 color_fprintf(output, color, "%6.2f%%", ratio); 722 fprintf(output, " of all L1-icache hits "); 723 } 724 725 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 726 { 727 double total, ratio = 0.0; 728 const char *color; 729 730 total = avg_stats(&runtime_dtlb_cache_stats[cpu]); 731 732 if (total) 733 ratio = avg / total * 100.0; 734 735 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 736 737 fprintf(output, " # "); 738 color_fprintf(output, color, "%6.2f%%", ratio); 739 fprintf(output, " of all dTLB cache hits "); 740 } 741 742 static void print_itlb_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_itlb_cache_stats[cpu]); 748 749 if (total) 750 ratio = avg / total * 100.0; 751 752 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 753 754 fprintf(output, " # "); 755 color_fprintf(output, color, "%6.2f%%", ratio); 756 fprintf(output, " of all iTLB cache hits "); 757 } 758 759 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 760 { 761 double total, ratio = 0.0; 762 const char *color; 763 764 total = avg_stats(&runtime_ll_cache_stats[cpu]); 765 766 if (total) 767 ratio = avg / total * 100.0; 768 769 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 770 771 fprintf(output, " # "); 772 color_fprintf(output, color, "%6.2f%%", ratio); 773 fprintf(output, " of all LL-cache hits "); 774 } 775 776 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) 777 { 778 double total, ratio = 0.0; 779 char cpustr[16] = { '\0', }; 780 const char *fmt; 781 782 if (csv_output) 783 fmt = "%s%.0f%s%s"; 784 else if (big_num) 785 fmt = "%s%'18.0f%s%-25s"; 786 else 787 fmt = "%s%18.0f%s%-25s"; 788 789 if (no_aggr) 790 sprintf(cpustr, "CPU%*d%s", 791 csv_output ? 0 : -4, 792 evsel_list->cpus->map[cpu], csv_sep); 793 else 794 cpu = 0; 795 796 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel)); 797 798 if (evsel->cgrp) 799 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); 800 801 if (csv_output) 802 return; 803 804 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { 805 total = avg_stats(&runtime_cycles_stats[cpu]); 806 807 if (total) 808 ratio = avg / total; 809 810 fprintf(output, " # %5.2f insns per cycle ", ratio); 811 812 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); 813 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); 814 815 if (total && avg) { 816 ratio = total / avg; 817 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio); 818 } 819 820 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && 821 runtime_branches_stats[cpu].n != 0) { 822 print_branch_misses(cpu, evsel, avg); 823 } else if ( 824 evsel->attr.type == PERF_TYPE_HW_CACHE && 825 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | 826 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 827 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 828 runtime_l1_dcache_stats[cpu].n != 0) { 829 print_l1_dcache_misses(cpu, evsel, avg); 830 } else if ( 831 evsel->attr.type == PERF_TYPE_HW_CACHE && 832 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | 833 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 834 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 835 runtime_l1_icache_stats[cpu].n != 0) { 836 print_l1_icache_misses(cpu, evsel, avg); 837 } else if ( 838 evsel->attr.type == PERF_TYPE_HW_CACHE && 839 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | 840 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 841 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 842 runtime_dtlb_cache_stats[cpu].n != 0) { 843 print_dtlb_cache_misses(cpu, evsel, avg); 844 } else if ( 845 evsel->attr.type == PERF_TYPE_HW_CACHE && 846 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | 847 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 848 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 849 runtime_itlb_cache_stats[cpu].n != 0) { 850 print_itlb_cache_misses(cpu, evsel, avg); 851 } else if ( 852 evsel->attr.type == PERF_TYPE_HW_CACHE && 853 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | 854 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 855 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 856 runtime_ll_cache_stats[cpu].n != 0) { 857 print_ll_cache_misses(cpu, evsel, avg); 858 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && 859 runtime_cacherefs_stats[cpu].n != 0) { 860 total = avg_stats(&runtime_cacherefs_stats[cpu]); 861 862 if (total) 863 ratio = avg * 100 / total; 864 865 fprintf(output, " # %8.3f %% of all cache refs ", ratio); 866 867 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { 868 print_stalled_cycles_frontend(cpu, evsel, avg); 869 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { 870 print_stalled_cycles_backend(cpu, evsel, avg); 871 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { 872 total = avg_stats(&runtime_nsecs_stats[cpu]); 873 874 if (total) 875 ratio = 1.0 * avg / total; 876 877 fprintf(output, " # %8.3f GHz ", ratio); 878 } else if (runtime_nsecs_stats[cpu].n != 0) { 879 char unit = 'M'; 880 881 total = avg_stats(&runtime_nsecs_stats[cpu]); 882 883 if (total) 884 ratio = 1000.0 * avg / total; 885 if (ratio < 0.001) { 886 ratio *= 1000; 887 unit = 'K'; 888 } 889 890 fprintf(output, " # %8.3f %c/sec ", ratio, unit); 891 } else { 892 fprintf(output, " "); 893 } 894 } 895 896 /* 897 * Print out the results of a single counter: 898 * aggregated counts in system-wide mode 899 */ 900 static void print_counter_aggr(struct perf_evsel *counter) 901 { 902 struct perf_stat *ps = counter->priv; 903 double avg = avg_stats(&ps->res_stats[0]); 904 int scaled = counter->counts->scaled; 905 906 if (scaled == -1) { 907 fprintf(output, "%*s%s%*s", 908 csv_output ? 0 : 18, 909 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 910 csv_sep, 911 csv_output ? 0 : -24, 912 event_name(counter)); 913 914 if (counter->cgrp) 915 fprintf(output, "%s%s", csv_sep, counter->cgrp->name); 916 917 fputc('\n', output); 918 return; 919 } 920 921 if (nsec_counter(counter)) 922 nsec_printout(-1, counter, avg); 923 else 924 abs_printout(-1, counter, avg); 925 926 print_noise(counter, avg); 927 928 if (csv_output) { 929 fputc('\n', output); 930 return; 931 } 932 933 if (scaled) { 934 double avg_enabled, avg_running; 935 936 avg_enabled = avg_stats(&ps->res_stats[1]); 937 avg_running = avg_stats(&ps->res_stats[2]); 938 939 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); 940 } 941 fprintf(output, "\n"); 942 } 943 944 /* 945 * Print out the results of a single counter: 946 * does not use aggregated count in system-wide 947 */ 948 static void print_counter(struct perf_evsel *counter) 949 { 950 u64 ena, run, val; 951 int cpu; 952 953 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { 954 val = counter->counts->cpu[cpu].val; 955 ena = counter->counts->cpu[cpu].ena; 956 run = counter->counts->cpu[cpu].run; 957 if (run == 0 || ena == 0) { 958 fprintf(output, "CPU%*d%s%*s%s%*s", 959 csv_output ? 0 : -4, 960 evsel_list->cpus->map[cpu], csv_sep, 961 csv_output ? 0 : 18, 962 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 963 csv_sep, 964 csv_output ? 0 : -24, 965 event_name(counter)); 966 967 if (counter->cgrp) 968 fprintf(output, "%s%s", 969 csv_sep, counter->cgrp->name); 970 971 fputc('\n', output); 972 continue; 973 } 974 975 if (nsec_counter(counter)) 976 nsec_printout(cpu, counter, val); 977 else 978 abs_printout(cpu, counter, val); 979 980 if (!csv_output) { 981 print_noise(counter, 1.0); 982 983 if (run != ena) 984 fprintf(output, " (%.2f%%)", 985 100.0 * run / ena); 986 } 987 fputc('\n', output); 988 } 989 } 990 991 static void print_stat(int argc, const char **argv) 992 { 993 struct perf_evsel *counter; 994 int i; 995 996 fflush(stdout); 997 998 if (!csv_output) { 999 fprintf(output, "\n"); 1000 fprintf(output, " Performance counter stats for "); 1001 if (!target_pid && !target_tid) { 1002 fprintf(output, "\'%s", argv[0]); 1003 for (i = 1; i < argc; i++) 1004 fprintf(output, " %s", argv[i]); 1005 } else if (target_pid) 1006 fprintf(output, "process id \'%s", target_pid); 1007 else 1008 fprintf(output, "thread id \'%s", target_tid); 1009 1010 fprintf(output, "\'"); 1011 if (run_count > 1) 1012 fprintf(output, " (%d runs)", run_count); 1013 fprintf(output, ":\n\n"); 1014 } 1015 1016 if (no_aggr) { 1017 list_for_each_entry(counter, &evsel_list->entries, node) 1018 print_counter(counter); 1019 } else { 1020 list_for_each_entry(counter, &evsel_list->entries, node) 1021 print_counter_aggr(counter); 1022 } 1023 1024 if (!csv_output) { 1025 if (!null_run) 1026 fprintf(output, "\n"); 1027 fprintf(output, " %17.9f seconds time elapsed", 1028 avg_stats(&walltime_nsecs_stats)/1e9); 1029 if (run_count > 1) { 1030 fprintf(output, " "); 1031 print_noise_pct(stddev_stats(&walltime_nsecs_stats), 1032 avg_stats(&walltime_nsecs_stats)); 1033 } 1034 fprintf(output, "\n\n"); 1035 } 1036 } 1037 1038 static volatile int signr = -1; 1039 1040 static void skip_signal(int signo) 1041 { 1042 if(child_pid == -1) 1043 done = 1; 1044 1045 signr = signo; 1046 } 1047 1048 static void sig_atexit(void) 1049 { 1050 if (child_pid != -1) 1051 kill(child_pid, SIGTERM); 1052 1053 if (signr == -1) 1054 return; 1055 1056 signal(signr, SIG_DFL); 1057 kill(getpid(), signr); 1058 } 1059 1060 static const char * const stat_usage[] = { 1061 "perf stat [<options>] [<command>]", 1062 NULL 1063 }; 1064 1065 static int stat__set_big_num(const struct option *opt __used, 1066 const char *s __used, int unset) 1067 { 1068 big_num_opt = unset ? 0 : 1; 1069 return 0; 1070 } 1071 1072 static bool append_file; 1073 1074 static const struct option options[] = { 1075 OPT_CALLBACK('e', "event", &evsel_list, "event", 1076 "event selector. use 'perf list' to list available events", 1077 parse_events_option), 1078 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 1079 "event filter", parse_filter), 1080 OPT_BOOLEAN('i', "no-inherit", &no_inherit, 1081 "child tasks do not inherit counters"), 1082 OPT_STRING('p', "pid", &target_pid, "pid", 1083 "stat events on existing process id"), 1084 OPT_STRING('t', "tid", &target_tid, "tid", 1085 "stat events on existing thread id"), 1086 OPT_BOOLEAN('a', "all-cpus", &system_wide, 1087 "system-wide collection from all CPUs"), 1088 OPT_BOOLEAN('g', "group", &group, 1089 "put the counters into a counter group"), 1090 OPT_BOOLEAN('c', "scale", &scale, 1091 "scale/normalize counters"), 1092 OPT_INCR('v', "verbose", &verbose, 1093 "be more verbose (show counter open errors, etc)"), 1094 OPT_INTEGER('r', "repeat", &run_count, 1095 "repeat command and print average + stddev (max: 100)"), 1096 OPT_BOOLEAN('n', "null", &null_run, 1097 "null run - dont start any counters"), 1098 OPT_INCR('d', "detailed", &detailed_run, 1099 "detailed run - start a lot of events"), 1100 OPT_BOOLEAN('S', "sync", &sync_run, 1101 "call sync() before starting a run"), 1102 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1103 "print large numbers with thousands\' separators", 1104 stat__set_big_num), 1105 OPT_STRING('C', "cpu", &cpu_list, "cpu", 1106 "list of cpus to monitor in system-wide"), 1107 OPT_BOOLEAN('A', "no-aggr", &no_aggr, 1108 "disable CPU count aggregation"), 1109 OPT_STRING('x', "field-separator", &csv_sep, "separator", 1110 "print counts with custom separator"), 1111 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 1112 "monitor event in cgroup name only", 1113 parse_cgroups), 1114 OPT_STRING('o', "output", &output_name, "file", 1115 "output file name"), 1116 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 1117 OPT_INTEGER(0, "log-fd", &output_fd, 1118 "log output to fd, instead of stderr"), 1119 OPT_END() 1120 }; 1121 1122 /* 1123 * Add default attributes, if there were no attributes specified or 1124 * if -d/--detailed, -d -d or -d -d -d is used: 1125 */ 1126 static int add_default_attributes(void) 1127 { 1128 /* Set attrs if no event is selected and !null_run: */ 1129 if (null_run) 1130 return 0; 1131 1132 if (!evsel_list->nr_entries) { 1133 if (perf_evlist__add_attrs_array(evsel_list, default_attrs) < 0) 1134 return -1; 1135 } 1136 1137 /* Detailed events get appended to the event list: */ 1138 1139 if (detailed_run < 1) 1140 return 0; 1141 1142 /* Append detailed run extra attributes: */ 1143 if (perf_evlist__add_attrs_array(evsel_list, detailed_attrs) < 0) 1144 return -1; 1145 1146 if (detailed_run < 2) 1147 return 0; 1148 1149 /* Append very detailed run extra attributes: */ 1150 if (perf_evlist__add_attrs_array(evsel_list, very_detailed_attrs) < 0) 1151 return -1; 1152 1153 if (detailed_run < 3) 1154 return 0; 1155 1156 /* Append very, very detailed run extra attributes: */ 1157 return perf_evlist__add_attrs_array(evsel_list, very_very_detailed_attrs); 1158 } 1159 1160 int cmd_stat(int argc, const char **argv, const char *prefix __used) 1161 { 1162 struct perf_evsel *pos; 1163 int status = -ENOMEM; 1164 const char *mode; 1165 1166 setlocale(LC_ALL, ""); 1167 1168 evsel_list = perf_evlist__new(NULL, NULL); 1169 if (evsel_list == NULL) 1170 return -ENOMEM; 1171 1172 argc = parse_options(argc, argv, options, stat_usage, 1173 PARSE_OPT_STOP_AT_NON_OPTION); 1174 1175 output = stderr; 1176 if (output_name && strcmp(output_name, "-")) 1177 output = NULL; 1178 1179 if (output_name && output_fd) { 1180 fprintf(stderr, "cannot use both --output and --log-fd\n"); 1181 usage_with_options(stat_usage, options); 1182 } 1183 if (!output) { 1184 struct timespec tm; 1185 mode = append_file ? "a" : "w"; 1186 1187 output = fopen(output_name, mode); 1188 if (!output) { 1189 perror("failed to create output file"); 1190 exit(-1); 1191 } 1192 clock_gettime(CLOCK_REALTIME, &tm); 1193 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 1194 } else if (output_fd != 2) { 1195 mode = append_file ? "a" : "w"; 1196 output = fdopen(output_fd, mode); 1197 if (!output) { 1198 perror("Failed opening logfd"); 1199 return -errno; 1200 } 1201 } 1202 1203 if (csv_sep) { 1204 csv_output = true; 1205 if (!strcmp(csv_sep, "\\t")) 1206 csv_sep = "\t"; 1207 } else 1208 csv_sep = DEFAULT_SEPARATOR; 1209 1210 /* 1211 * let the spreadsheet do the pretty-printing 1212 */ 1213 if (csv_output) { 1214 /* User explicitly passed -B? */ 1215 if (big_num_opt == 1) { 1216 fprintf(stderr, "-B option not supported with -x\n"); 1217 usage_with_options(stat_usage, options); 1218 } else /* Nope, so disable big number formatting */ 1219 big_num = false; 1220 } else if (big_num_opt == 0) /* User passed --no-big-num */ 1221 big_num = false; 1222 1223 if (!argc && !target_pid && !target_tid) 1224 usage_with_options(stat_usage, options); 1225 if (run_count <= 0) 1226 usage_with_options(stat_usage, options); 1227 1228 /* no_aggr, cgroup are for system-wide only */ 1229 if ((no_aggr || nr_cgroups) && !system_wide) { 1230 fprintf(stderr, "both cgroup and no-aggregation " 1231 "modes only available in system-wide mode\n"); 1232 1233 usage_with_options(stat_usage, options); 1234 } 1235 1236 if (add_default_attributes()) 1237 goto out; 1238 1239 if (target_pid) 1240 target_tid = target_pid; 1241 1242 evsel_list->threads = thread_map__new_str(target_pid, 1243 target_tid, UINT_MAX); 1244 if (evsel_list->threads == NULL) { 1245 pr_err("Problems finding threads of monitor\n"); 1246 usage_with_options(stat_usage, options); 1247 } 1248 1249 if (system_wide) 1250 evsel_list->cpus = cpu_map__new(cpu_list); 1251 else 1252 evsel_list->cpus = cpu_map__dummy_new(); 1253 1254 if (evsel_list->cpus == NULL) { 1255 perror("failed to parse CPUs map"); 1256 usage_with_options(stat_usage, options); 1257 return -1; 1258 } 1259 1260 list_for_each_entry(pos, &evsel_list->entries, node) { 1261 if (perf_evsel__alloc_stat_priv(pos) < 0 || 1262 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0) 1263 goto out_free_fd; 1264 } 1265 1266 /* 1267 * We dont want to block the signals - that would cause 1268 * child tasks to inherit that and Ctrl-C would not work. 1269 * What we want is for Ctrl-C to work in the exec()-ed 1270 * task, but being ignored by perf stat itself: 1271 */ 1272 atexit(sig_atexit); 1273 signal(SIGINT, skip_signal); 1274 signal(SIGALRM, skip_signal); 1275 signal(SIGABRT, skip_signal); 1276 1277 status = 0; 1278 for (run_idx = 0; run_idx < run_count; run_idx++) { 1279 if (run_count != 1 && verbose) 1280 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 1281 run_idx + 1); 1282 1283 if (sync_run) 1284 sync(); 1285 1286 status = run_perf_stat(argc, argv); 1287 } 1288 1289 if (status != -1) 1290 print_stat(argc, argv); 1291 out_free_fd: 1292 list_for_each_entry(pos, &evsel_list->entries, node) 1293 perf_evsel__free_stat_priv(pos); 1294 perf_evlist__delete_maps(evsel_list); 1295 out: 1296 perf_evlist__delete(evsel_list); 1297 return status; 1298 } 1299