1 /* 2 * builtin-stat.c 3 * 4 * Builtin stat command: Give a precise performance counters summary 5 * overview about any workload, CPU or specific PID. 6 * 7 * Sample output: 8 9 $ perf stat ./hackbench 10 10 11 Time: 0.118 12 13 Performance counter stats for './hackbench 10': 14 15 1708.761321 task-clock # 11.037 CPUs utilized 16 41,190 context-switches # 0.024 M/sec 17 6,735 CPU-migrations # 0.004 M/sec 18 17,318 page-faults # 0.010 M/sec 19 5,205,202,243 cycles # 3.046 GHz 20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle 21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle 22 2,603,501,247 instructions # 0.50 insns per cycle 23 # 1.48 stalled cycles per insn 24 484,357,498 branches # 283.455 M/sec 25 6,388,934 branch-misses # 1.32% of all branches 26 27 0.154822978 seconds time elapsed 28 29 * 30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 31 * 32 * Improvements and fixes by: 33 * 34 * Arjan van de Ven <arjan@linux.intel.com> 35 * Yanmin Zhang <yanmin.zhang@intel.com> 36 * Wu Fengguang <fengguang.wu@intel.com> 37 * Mike Galbraith <efault@gmx.de> 38 * Paul Mackerras <paulus@samba.org> 39 * Jaswinder Singh Rajput <jaswinder@kernel.org> 40 * 41 * Released under the GPL v2. (and only v2, not any later version) 42 */ 43 44 #include "perf.h" 45 #include "builtin.h" 46 #include "util/util.h" 47 #include "util/parse-options.h" 48 #include "util/parse-events.h" 49 #include "util/event.h" 50 #include "util/evlist.h" 51 #include "util/evsel.h" 52 #include "util/debug.h" 53 #include "util/color.h" 54 #include "util/header.h" 55 #include "util/cpumap.h" 56 #include "util/thread.h" 57 #include "util/thread_map.h" 58 59 #include <sys/prctl.h> 60 #include <math.h> 61 #include <locale.h> 62 63 #define DEFAULT_SEPARATOR " " 64 #define CNTR_NOT_SUPPORTED "<not supported>" 65 #define CNTR_NOT_COUNTED "<not counted>" 66 67 static struct perf_event_attr default_attrs[] = { 68 69 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 70 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 71 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 72 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 73 74 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 75 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 76 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 77 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 78 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 79 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 80 81 }; 82 83 /* 84 * Detailed stats (-d), covering the L1 and last level data caches: 85 */ 86 static struct perf_event_attr detailed_attrs[] = { 87 88 { .type = PERF_TYPE_HW_CACHE, 89 .config = 90 PERF_COUNT_HW_CACHE_L1D << 0 | 91 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 93 94 { .type = PERF_TYPE_HW_CACHE, 95 .config = 96 PERF_COUNT_HW_CACHE_L1D << 0 | 97 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 98 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 99 100 { .type = PERF_TYPE_HW_CACHE, 101 .config = 102 PERF_COUNT_HW_CACHE_LL << 0 | 103 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 105 106 { .type = PERF_TYPE_HW_CACHE, 107 .config = 108 PERF_COUNT_HW_CACHE_LL << 0 | 109 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 110 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 111 }; 112 113 /* 114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 115 */ 116 static struct perf_event_attr very_detailed_attrs[] = { 117 118 { .type = PERF_TYPE_HW_CACHE, 119 .config = 120 PERF_COUNT_HW_CACHE_L1I << 0 | 121 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 123 124 { .type = PERF_TYPE_HW_CACHE, 125 .config = 126 PERF_COUNT_HW_CACHE_L1I << 0 | 127 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 128 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 129 130 { .type = PERF_TYPE_HW_CACHE, 131 .config = 132 PERF_COUNT_HW_CACHE_DTLB << 0 | 133 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 135 136 { .type = PERF_TYPE_HW_CACHE, 137 .config = 138 PERF_COUNT_HW_CACHE_DTLB << 0 | 139 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 140 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 141 142 { .type = PERF_TYPE_HW_CACHE, 143 .config = 144 PERF_COUNT_HW_CACHE_ITLB << 0 | 145 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 147 148 { .type = PERF_TYPE_HW_CACHE, 149 .config = 150 PERF_COUNT_HW_CACHE_ITLB << 0 | 151 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 152 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 153 154 }; 155 156 /* 157 * Very, very detailed stats (-d -d -d), adding prefetch events: 158 */ 159 static struct perf_event_attr very_very_detailed_attrs[] = { 160 161 { .type = PERF_TYPE_HW_CACHE, 162 .config = 163 PERF_COUNT_HW_CACHE_L1D << 0 | 164 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 166 167 { .type = PERF_TYPE_HW_CACHE, 168 .config = 169 PERF_COUNT_HW_CACHE_L1D << 0 | 170 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 171 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 172 }; 173 174 175 176 struct perf_evlist *evsel_list; 177 178 static bool system_wide = false; 179 static int run_idx = 0; 180 181 static int run_count = 1; 182 static bool no_inherit = false; 183 static bool scale = true; 184 static bool no_aggr = false; 185 static pid_t target_pid = -1; 186 static pid_t target_tid = -1; 187 static pid_t child_pid = -1; 188 static bool null_run = false; 189 static int detailed_run = 0; 190 static bool sync_run = false; 191 static bool big_num = true; 192 static int big_num_opt = -1; 193 static const char *cpu_list; 194 static const char *csv_sep = NULL; 195 static bool csv_output = false; 196 static bool group = false; 197 198 static volatile int done = 0; 199 200 struct stats 201 { 202 double n, mean, M2; 203 }; 204 205 struct perf_stat { 206 struct stats res_stats[3]; 207 }; 208 209 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) 210 { 211 evsel->priv = zalloc(sizeof(struct perf_stat)); 212 return evsel->priv == NULL ? -ENOMEM : 0; 213 } 214 215 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) 216 { 217 free(evsel->priv); 218 evsel->priv = NULL; 219 } 220 221 static void update_stats(struct stats *stats, u64 val) 222 { 223 double delta; 224 225 stats->n++; 226 delta = val - stats->mean; 227 stats->mean += delta / stats->n; 228 stats->M2 += delta*(val - stats->mean); 229 } 230 231 static double avg_stats(struct stats *stats) 232 { 233 return stats->mean; 234 } 235 236 /* 237 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 238 * 239 * (\Sum n_i^2) - ((\Sum n_i)^2)/n 240 * s^2 = ------------------------------- 241 * n - 1 242 * 243 * http://en.wikipedia.org/wiki/Stddev 244 * 245 * The std dev of the mean is related to the std dev by: 246 * 247 * s 248 * s_mean = ------- 249 * sqrt(n) 250 * 251 */ 252 static double stddev_stats(struct stats *stats) 253 { 254 double variance = stats->M2 / (stats->n - 1); 255 double variance_mean = variance / stats->n; 256 257 return sqrt(variance_mean); 258 } 259 260 struct stats runtime_nsecs_stats[MAX_NR_CPUS]; 261 struct stats runtime_cycles_stats[MAX_NR_CPUS]; 262 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; 263 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; 264 struct stats runtime_branches_stats[MAX_NR_CPUS]; 265 struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; 266 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; 267 struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; 268 struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; 269 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; 270 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; 271 struct stats walltime_nsecs_stats; 272 273 static int create_perf_stat_counter(struct perf_evsel *evsel) 274 { 275 struct perf_event_attr *attr = &evsel->attr; 276 277 if (scale) 278 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 279 PERF_FORMAT_TOTAL_TIME_RUNNING; 280 281 attr->inherit = !no_inherit; 282 283 if (system_wide) 284 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group); 285 286 if (target_pid == -1 && target_tid == -1) { 287 attr->disabled = 1; 288 attr->enable_on_exec = 1; 289 } 290 291 return perf_evsel__open_per_thread(evsel, evsel_list->threads, group); 292 } 293 294 /* 295 * Does the counter have nsecs as a unit? 296 */ 297 static inline int nsec_counter(struct perf_evsel *evsel) 298 { 299 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || 300 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 301 return 1; 302 303 return 0; 304 } 305 306 /* 307 * Update various tracking values we maintain to print 308 * more semantic information such as miss/hit ratios, 309 * instruction rates, etc: 310 */ 311 static void update_shadow_stats(struct perf_evsel *counter, u64 *count) 312 { 313 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) 314 update_stats(&runtime_nsecs_stats[0], count[0]); 315 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) 316 update_stats(&runtime_cycles_stats[0], count[0]); 317 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) 318 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); 319 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) 320 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); 321 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) 322 update_stats(&runtime_branches_stats[0], count[0]); 323 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) 324 update_stats(&runtime_cacherefs_stats[0], count[0]); 325 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) 326 update_stats(&runtime_l1_dcache_stats[0], count[0]); 327 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) 328 update_stats(&runtime_l1_icache_stats[0], count[0]); 329 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) 330 update_stats(&runtime_ll_cache_stats[0], count[0]); 331 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) 332 update_stats(&runtime_dtlb_cache_stats[0], count[0]); 333 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) 334 update_stats(&runtime_itlb_cache_stats[0], count[0]); 335 } 336 337 /* 338 * Read out the results of a single counter: 339 * aggregate counts across CPUs in system-wide mode 340 */ 341 static int read_counter_aggr(struct perf_evsel *counter) 342 { 343 struct perf_stat *ps = counter->priv; 344 u64 *count = counter->counts->aggr.values; 345 int i; 346 347 if (__perf_evsel__read(counter, evsel_list->cpus->nr, 348 evsel_list->threads->nr, scale) < 0) 349 return -1; 350 351 for (i = 0; i < 3; i++) 352 update_stats(&ps->res_stats[i], count[i]); 353 354 if (verbose) { 355 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 356 event_name(counter), count[0], count[1], count[2]); 357 } 358 359 /* 360 * Save the full runtime - to allow normalization during printout: 361 */ 362 update_shadow_stats(counter, count); 363 364 return 0; 365 } 366 367 /* 368 * Read out the results of a single counter: 369 * do not aggregate counts across CPUs in system-wide mode 370 */ 371 static int read_counter(struct perf_evsel *counter) 372 { 373 u64 *count; 374 int cpu; 375 376 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { 377 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) 378 return -1; 379 380 count = counter->counts->cpu[cpu].values; 381 382 update_shadow_stats(counter, count); 383 } 384 385 return 0; 386 } 387 388 static int run_perf_stat(int argc __used, const char **argv) 389 { 390 unsigned long long t0, t1; 391 struct perf_evsel *counter; 392 int status = 0; 393 int child_ready_pipe[2], go_pipe[2]; 394 const bool forks = (argc > 0); 395 char buf; 396 397 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 398 perror("failed to create pipes"); 399 exit(1); 400 } 401 402 if (forks) { 403 if ((child_pid = fork()) < 0) 404 perror("failed to fork"); 405 406 if (!child_pid) { 407 close(child_ready_pipe[0]); 408 close(go_pipe[1]); 409 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 410 411 /* 412 * Do a dummy execvp to get the PLT entry resolved, 413 * so we avoid the resolver overhead on the real 414 * execvp call. 415 */ 416 execvp("", (char **)argv); 417 418 /* 419 * Tell the parent we're ready to go 420 */ 421 close(child_ready_pipe[1]); 422 423 /* 424 * Wait until the parent tells us to go. 425 */ 426 if (read(go_pipe[0], &buf, 1) == -1) 427 perror("unable to read pipe"); 428 429 execvp(argv[0], (char **)argv); 430 431 perror(argv[0]); 432 exit(-1); 433 } 434 435 if (target_tid == -1 && target_pid == -1 && !system_wide) 436 evsel_list->threads->map[0] = child_pid; 437 438 /* 439 * Wait for the child to be ready to exec. 440 */ 441 close(child_ready_pipe[1]); 442 close(go_pipe[0]); 443 if (read(child_ready_pipe[0], &buf, 1) == -1) 444 perror("unable to read pipe"); 445 close(child_ready_pipe[0]); 446 } 447 448 list_for_each_entry(counter, &evsel_list->entries, node) { 449 if (create_perf_stat_counter(counter) < 0) { 450 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) { 451 if (verbose) 452 ui__warning("%s event is not supported by the kernel.\n", 453 event_name(counter)); 454 counter->supported = false; 455 continue; 456 } 457 458 if (errno == EPERM || errno == EACCES) { 459 error("You may not have permission to collect %sstats.\n" 460 "\t Consider tweaking" 461 " /proc/sys/kernel/perf_event_paranoid or running as root.", 462 system_wide ? "system-wide " : ""); 463 } else { 464 error("open_counter returned with %d (%s). " 465 "/bin/dmesg may provide additional information.\n", 466 errno, strerror(errno)); 467 } 468 if (child_pid != -1) 469 kill(child_pid, SIGTERM); 470 die("Not all events could be opened.\n"); 471 return -1; 472 } 473 counter->supported = true; 474 } 475 476 if (perf_evlist__set_filters(evsel_list)) { 477 error("failed to set filter with %d (%s)\n", errno, 478 strerror(errno)); 479 return -1; 480 } 481 482 /* 483 * Enable counters and exec the command: 484 */ 485 t0 = rdclock(); 486 487 if (forks) { 488 close(go_pipe[1]); 489 wait(&status); 490 } else { 491 while(!done) sleep(1); 492 } 493 494 t1 = rdclock(); 495 496 update_stats(&walltime_nsecs_stats, t1 - t0); 497 498 if (no_aggr) { 499 list_for_each_entry(counter, &evsel_list->entries, node) { 500 read_counter(counter); 501 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1); 502 } 503 } else { 504 list_for_each_entry(counter, &evsel_list->entries, node) { 505 read_counter_aggr(counter); 506 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 507 evsel_list->threads->nr); 508 } 509 } 510 511 return WEXITSTATUS(status); 512 } 513 514 static void print_noise_pct(double total, double avg) 515 { 516 double pct = 0.0; 517 518 if (avg) 519 pct = 100.0*total/avg; 520 521 if (csv_output) 522 fprintf(stderr, "%s%.2f%%", csv_sep, pct); 523 else 524 fprintf(stderr, " ( +-%6.2f%% )", pct); 525 } 526 527 static void print_noise(struct perf_evsel *evsel, double avg) 528 { 529 struct perf_stat *ps; 530 531 if (run_count == 1) 532 return; 533 534 ps = evsel->priv; 535 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); 536 } 537 538 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) 539 { 540 double msecs = avg / 1e6; 541 char cpustr[16] = { '\0', }; 542 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; 543 544 if (no_aggr) 545 sprintf(cpustr, "CPU%*d%s", 546 csv_output ? 0 : -4, 547 evsel_list->cpus->map[cpu], csv_sep); 548 549 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel)); 550 551 if (evsel->cgrp) 552 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name); 553 554 if (csv_output) 555 return; 556 557 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 558 fprintf(stderr, " # %8.3f CPUs utilized ", avg / avg_stats(&walltime_nsecs_stats)); 559 } 560 561 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg) 562 { 563 double total, ratio = 0.0; 564 const char *color; 565 566 total = avg_stats(&runtime_cycles_stats[cpu]); 567 568 if (total) 569 ratio = avg / total * 100.0; 570 571 color = PERF_COLOR_NORMAL; 572 if (ratio > 50.0) 573 color = PERF_COLOR_RED; 574 else if (ratio > 30.0) 575 color = PERF_COLOR_MAGENTA; 576 else if (ratio > 10.0) 577 color = PERF_COLOR_YELLOW; 578 579 fprintf(stderr, " # "); 580 color_fprintf(stderr, color, "%6.2f%%", ratio); 581 fprintf(stderr, " frontend cycles idle "); 582 } 583 584 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg) 585 { 586 double total, ratio = 0.0; 587 const char *color; 588 589 total = avg_stats(&runtime_cycles_stats[cpu]); 590 591 if (total) 592 ratio = avg / total * 100.0; 593 594 color = PERF_COLOR_NORMAL; 595 if (ratio > 75.0) 596 color = PERF_COLOR_RED; 597 else if (ratio > 50.0) 598 color = PERF_COLOR_MAGENTA; 599 else if (ratio > 20.0) 600 color = PERF_COLOR_YELLOW; 601 602 fprintf(stderr, " # "); 603 color_fprintf(stderr, color, "%6.2f%%", ratio); 604 fprintf(stderr, " backend cycles idle "); 605 } 606 607 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg) 608 { 609 double total, ratio = 0.0; 610 const char *color; 611 612 total = avg_stats(&runtime_branches_stats[cpu]); 613 614 if (total) 615 ratio = avg / total * 100.0; 616 617 color = PERF_COLOR_NORMAL; 618 if (ratio > 20.0) 619 color = PERF_COLOR_RED; 620 else if (ratio > 10.0) 621 color = PERF_COLOR_MAGENTA; 622 else if (ratio > 5.0) 623 color = PERF_COLOR_YELLOW; 624 625 fprintf(stderr, " # "); 626 color_fprintf(stderr, color, "%6.2f%%", ratio); 627 fprintf(stderr, " of all branches "); 628 } 629 630 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 631 { 632 double total, ratio = 0.0; 633 const char *color; 634 635 total = avg_stats(&runtime_l1_dcache_stats[cpu]); 636 637 if (total) 638 ratio = avg / total * 100.0; 639 640 color = PERF_COLOR_NORMAL; 641 if (ratio > 20.0) 642 color = PERF_COLOR_RED; 643 else if (ratio > 10.0) 644 color = PERF_COLOR_MAGENTA; 645 else if (ratio > 5.0) 646 color = PERF_COLOR_YELLOW; 647 648 fprintf(stderr, " # "); 649 color_fprintf(stderr, color, "%6.2f%%", ratio); 650 fprintf(stderr, " of all L1-dcache hits "); 651 } 652 653 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 654 { 655 double total, ratio = 0.0; 656 const char *color; 657 658 total = avg_stats(&runtime_l1_icache_stats[cpu]); 659 660 if (total) 661 ratio = avg / total * 100.0; 662 663 color = PERF_COLOR_NORMAL; 664 if (ratio > 20.0) 665 color = PERF_COLOR_RED; 666 else if (ratio > 10.0) 667 color = PERF_COLOR_MAGENTA; 668 else if (ratio > 5.0) 669 color = PERF_COLOR_YELLOW; 670 671 fprintf(stderr, " # "); 672 color_fprintf(stderr, color, "%6.2f%%", ratio); 673 fprintf(stderr, " of all L1-icache hits "); 674 } 675 676 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 677 { 678 double total, ratio = 0.0; 679 const char *color; 680 681 total = avg_stats(&runtime_dtlb_cache_stats[cpu]); 682 683 if (total) 684 ratio = avg / total * 100.0; 685 686 color = PERF_COLOR_NORMAL; 687 if (ratio > 20.0) 688 color = PERF_COLOR_RED; 689 else if (ratio > 10.0) 690 color = PERF_COLOR_MAGENTA; 691 else if (ratio > 5.0) 692 color = PERF_COLOR_YELLOW; 693 694 fprintf(stderr, " # "); 695 color_fprintf(stderr, color, "%6.2f%%", ratio); 696 fprintf(stderr, " of all dTLB cache hits "); 697 } 698 699 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 700 { 701 double total, ratio = 0.0; 702 const char *color; 703 704 total = avg_stats(&runtime_itlb_cache_stats[cpu]); 705 706 if (total) 707 ratio = avg / total * 100.0; 708 709 color = PERF_COLOR_NORMAL; 710 if (ratio > 20.0) 711 color = PERF_COLOR_RED; 712 else if (ratio > 10.0) 713 color = PERF_COLOR_MAGENTA; 714 else if (ratio > 5.0) 715 color = PERF_COLOR_YELLOW; 716 717 fprintf(stderr, " # "); 718 color_fprintf(stderr, color, "%6.2f%%", ratio); 719 fprintf(stderr, " of all iTLB cache hits "); 720 } 721 722 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) 723 { 724 double total, ratio = 0.0; 725 const char *color; 726 727 total = avg_stats(&runtime_ll_cache_stats[cpu]); 728 729 if (total) 730 ratio = avg / total * 100.0; 731 732 color = PERF_COLOR_NORMAL; 733 if (ratio > 20.0) 734 color = PERF_COLOR_RED; 735 else if (ratio > 10.0) 736 color = PERF_COLOR_MAGENTA; 737 else if (ratio > 5.0) 738 color = PERF_COLOR_YELLOW; 739 740 fprintf(stderr, " # "); 741 color_fprintf(stderr, color, "%6.2f%%", ratio); 742 fprintf(stderr, " of all LL-cache hits "); 743 } 744 745 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) 746 { 747 double total, ratio = 0.0; 748 char cpustr[16] = { '\0', }; 749 const char *fmt; 750 751 if (csv_output) 752 fmt = "%s%.0f%s%s"; 753 else if (big_num) 754 fmt = "%s%'18.0f%s%-25s"; 755 else 756 fmt = "%s%18.0f%s%-25s"; 757 758 if (no_aggr) 759 sprintf(cpustr, "CPU%*d%s", 760 csv_output ? 0 : -4, 761 evsel_list->cpus->map[cpu], csv_sep); 762 else 763 cpu = 0; 764 765 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel)); 766 767 if (evsel->cgrp) 768 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name); 769 770 if (csv_output) 771 return; 772 773 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { 774 total = avg_stats(&runtime_cycles_stats[cpu]); 775 776 if (total) 777 ratio = avg / total; 778 779 fprintf(stderr, " # %5.2f insns per cycle ", ratio); 780 781 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); 782 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); 783 784 if (total && avg) { 785 ratio = total / avg; 786 fprintf(stderr, "\n # %5.2f stalled cycles per insn", ratio); 787 } 788 789 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && 790 runtime_branches_stats[cpu].n != 0) { 791 print_branch_misses(cpu, evsel, avg); 792 } else if ( 793 evsel->attr.type == PERF_TYPE_HW_CACHE && 794 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | 795 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 796 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 797 runtime_l1_dcache_stats[cpu].n != 0) { 798 print_l1_dcache_misses(cpu, evsel, avg); 799 } else if ( 800 evsel->attr.type == PERF_TYPE_HW_CACHE && 801 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | 802 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 803 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 804 runtime_l1_icache_stats[cpu].n != 0) { 805 print_l1_icache_misses(cpu, evsel, avg); 806 } else if ( 807 evsel->attr.type == PERF_TYPE_HW_CACHE && 808 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | 809 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 810 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 811 runtime_dtlb_cache_stats[cpu].n != 0) { 812 print_dtlb_cache_misses(cpu, evsel, avg); 813 } else if ( 814 evsel->attr.type == PERF_TYPE_HW_CACHE && 815 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | 816 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 817 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 818 runtime_itlb_cache_stats[cpu].n != 0) { 819 print_itlb_cache_misses(cpu, evsel, avg); 820 } else if ( 821 evsel->attr.type == PERF_TYPE_HW_CACHE && 822 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | 823 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 824 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && 825 runtime_ll_cache_stats[cpu].n != 0) { 826 print_ll_cache_misses(cpu, evsel, avg); 827 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && 828 runtime_cacherefs_stats[cpu].n != 0) { 829 total = avg_stats(&runtime_cacherefs_stats[cpu]); 830 831 if (total) 832 ratio = avg * 100 / total; 833 834 fprintf(stderr, " # %8.3f %% of all cache refs ", ratio); 835 836 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { 837 print_stalled_cycles_frontend(cpu, evsel, avg); 838 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { 839 print_stalled_cycles_backend(cpu, evsel, avg); 840 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { 841 total = avg_stats(&runtime_nsecs_stats[cpu]); 842 843 if (total) 844 ratio = 1.0 * avg / total; 845 846 fprintf(stderr, " # %8.3f GHz ", ratio); 847 } else if (runtime_nsecs_stats[cpu].n != 0) { 848 total = avg_stats(&runtime_nsecs_stats[cpu]); 849 850 if (total) 851 ratio = 1000.0 * avg / total; 852 853 fprintf(stderr, " # %8.3f M/sec ", ratio); 854 } else { 855 fprintf(stderr, " "); 856 } 857 } 858 859 /* 860 * Print out the results of a single counter: 861 * aggregated counts in system-wide mode 862 */ 863 static void print_counter_aggr(struct perf_evsel *counter) 864 { 865 struct perf_stat *ps = counter->priv; 866 double avg = avg_stats(&ps->res_stats[0]); 867 int scaled = counter->counts->scaled; 868 869 if (scaled == -1) { 870 fprintf(stderr, "%*s%s%*s", 871 csv_output ? 0 : 18, 872 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 873 csv_sep, 874 csv_output ? 0 : -24, 875 event_name(counter)); 876 877 if (counter->cgrp) 878 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name); 879 880 fputc('\n', stderr); 881 return; 882 } 883 884 if (nsec_counter(counter)) 885 nsec_printout(-1, counter, avg); 886 else 887 abs_printout(-1, counter, avg); 888 889 print_noise(counter, avg); 890 891 if (csv_output) { 892 fputc('\n', stderr); 893 return; 894 } 895 896 if (scaled) { 897 double avg_enabled, avg_running; 898 899 avg_enabled = avg_stats(&ps->res_stats[1]); 900 avg_running = avg_stats(&ps->res_stats[2]); 901 902 fprintf(stderr, " [%5.2f%%]", 100 * avg_running / avg_enabled); 903 } 904 fprintf(stderr, "\n"); 905 } 906 907 /* 908 * Print out the results of a single counter: 909 * does not use aggregated count in system-wide 910 */ 911 static void print_counter(struct perf_evsel *counter) 912 { 913 u64 ena, run, val; 914 int cpu; 915 916 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { 917 val = counter->counts->cpu[cpu].val; 918 ena = counter->counts->cpu[cpu].ena; 919 run = counter->counts->cpu[cpu].run; 920 if (run == 0 || ena == 0) { 921 fprintf(stderr, "CPU%*d%s%*s%s%*s", 922 csv_output ? 0 : -4, 923 evsel_list->cpus->map[cpu], csv_sep, 924 csv_output ? 0 : 18, 925 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 926 csv_sep, 927 csv_output ? 0 : -24, 928 event_name(counter)); 929 930 if (counter->cgrp) 931 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name); 932 933 fputc('\n', stderr); 934 continue; 935 } 936 937 if (nsec_counter(counter)) 938 nsec_printout(cpu, counter, val); 939 else 940 abs_printout(cpu, counter, val); 941 942 if (!csv_output) { 943 print_noise(counter, 1.0); 944 945 if (run != ena) 946 fprintf(stderr, " (%.2f%%)", 100.0 * run / ena); 947 } 948 fputc('\n', stderr); 949 } 950 } 951 952 static void print_stat(int argc, const char **argv) 953 { 954 struct perf_evsel *counter; 955 int i; 956 957 fflush(stdout); 958 959 if (!csv_output) { 960 fprintf(stderr, "\n"); 961 fprintf(stderr, " Performance counter stats for "); 962 if(target_pid == -1 && target_tid == -1) { 963 fprintf(stderr, "\'%s", argv[0]); 964 for (i = 1; i < argc; i++) 965 fprintf(stderr, " %s", argv[i]); 966 } else if (target_pid != -1) 967 fprintf(stderr, "process id \'%d", target_pid); 968 else 969 fprintf(stderr, "thread id \'%d", target_tid); 970 971 fprintf(stderr, "\'"); 972 if (run_count > 1) 973 fprintf(stderr, " (%d runs)", run_count); 974 fprintf(stderr, ":\n\n"); 975 } 976 977 if (no_aggr) { 978 list_for_each_entry(counter, &evsel_list->entries, node) 979 print_counter(counter); 980 } else { 981 list_for_each_entry(counter, &evsel_list->entries, node) 982 print_counter_aggr(counter); 983 } 984 985 if (!csv_output) { 986 if (!null_run) 987 fprintf(stderr, "\n"); 988 fprintf(stderr, " %17.9f seconds time elapsed", 989 avg_stats(&walltime_nsecs_stats)/1e9); 990 if (run_count > 1) { 991 fprintf(stderr, " "); 992 print_noise_pct(stddev_stats(&walltime_nsecs_stats), 993 avg_stats(&walltime_nsecs_stats)); 994 } 995 fprintf(stderr, "\n\n"); 996 } 997 } 998 999 static volatile int signr = -1; 1000 1001 static void skip_signal(int signo) 1002 { 1003 if(child_pid == -1) 1004 done = 1; 1005 1006 signr = signo; 1007 } 1008 1009 static void sig_atexit(void) 1010 { 1011 if (child_pid != -1) 1012 kill(child_pid, SIGTERM); 1013 1014 if (signr == -1) 1015 return; 1016 1017 signal(signr, SIG_DFL); 1018 kill(getpid(), signr); 1019 } 1020 1021 static const char * const stat_usage[] = { 1022 "perf stat [<options>] [<command>]", 1023 NULL 1024 }; 1025 1026 static int stat__set_big_num(const struct option *opt __used, 1027 const char *s __used, int unset) 1028 { 1029 big_num_opt = unset ? 0 : 1; 1030 return 0; 1031 } 1032 1033 static const struct option options[] = { 1034 OPT_CALLBACK('e', "event", &evsel_list, "event", 1035 "event selector. use 'perf list' to list available events", 1036 parse_events_option), 1037 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 1038 "event filter", parse_filter), 1039 OPT_BOOLEAN('i', "no-inherit", &no_inherit, 1040 "child tasks do not inherit counters"), 1041 OPT_INTEGER('p', "pid", &target_pid, 1042 "stat events on existing process id"), 1043 OPT_INTEGER('t', "tid", &target_tid, 1044 "stat events on existing thread id"), 1045 OPT_BOOLEAN('a', "all-cpus", &system_wide, 1046 "system-wide collection from all CPUs"), 1047 OPT_BOOLEAN('g', "group", &group, 1048 "put the counters into a counter group"), 1049 OPT_BOOLEAN('c', "scale", &scale, 1050 "scale/normalize counters"), 1051 OPT_INCR('v', "verbose", &verbose, 1052 "be more verbose (show counter open errors, etc)"), 1053 OPT_INTEGER('r', "repeat", &run_count, 1054 "repeat command and print average + stddev (max: 100)"), 1055 OPT_BOOLEAN('n', "null", &null_run, 1056 "null run - dont start any counters"), 1057 OPT_INCR('d', "detailed", &detailed_run, 1058 "detailed run - start a lot of events"), 1059 OPT_BOOLEAN('S', "sync", &sync_run, 1060 "call sync() before starting a run"), 1061 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1062 "print large numbers with thousands\' separators", 1063 stat__set_big_num), 1064 OPT_STRING('C', "cpu", &cpu_list, "cpu", 1065 "list of cpus to monitor in system-wide"), 1066 OPT_BOOLEAN('A', "no-aggr", &no_aggr, 1067 "disable CPU count aggregation"), 1068 OPT_STRING('x', "field-separator", &csv_sep, "separator", 1069 "print counts with custom separator"), 1070 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 1071 "monitor event in cgroup name only", 1072 parse_cgroups), 1073 OPT_END() 1074 }; 1075 1076 /* 1077 * Add default attributes, if there were no attributes specified or 1078 * if -d/--detailed, -d -d or -d -d -d is used: 1079 */ 1080 static int add_default_attributes(void) 1081 { 1082 struct perf_evsel *pos; 1083 size_t attr_nr = 0; 1084 size_t c; 1085 1086 /* Set attrs if no event is selected and !null_run: */ 1087 if (null_run) 1088 return 0; 1089 1090 if (!evsel_list->nr_entries) { 1091 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) { 1092 pos = perf_evsel__new(default_attrs + c, c + attr_nr); 1093 if (pos == NULL) 1094 return -1; 1095 perf_evlist__add(evsel_list, pos); 1096 } 1097 attr_nr += c; 1098 } 1099 1100 /* Detailed events get appended to the event list: */ 1101 1102 if (detailed_run < 1) 1103 return 0; 1104 1105 /* Append detailed run extra attributes: */ 1106 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) { 1107 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr); 1108 if (pos == NULL) 1109 return -1; 1110 perf_evlist__add(evsel_list, pos); 1111 } 1112 attr_nr += c; 1113 1114 if (detailed_run < 2) 1115 return 0; 1116 1117 /* Append very detailed run extra attributes: */ 1118 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) { 1119 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr); 1120 if (pos == NULL) 1121 return -1; 1122 perf_evlist__add(evsel_list, pos); 1123 } 1124 1125 if (detailed_run < 3) 1126 return 0; 1127 1128 /* Append very, very detailed run extra attributes: */ 1129 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) { 1130 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr); 1131 if (pos == NULL) 1132 return -1; 1133 perf_evlist__add(evsel_list, pos); 1134 } 1135 1136 1137 return 0; 1138 } 1139 1140 int cmd_stat(int argc, const char **argv, const char *prefix __used) 1141 { 1142 struct perf_evsel *pos; 1143 int status = -ENOMEM; 1144 1145 setlocale(LC_ALL, ""); 1146 1147 evsel_list = perf_evlist__new(NULL, NULL); 1148 if (evsel_list == NULL) 1149 return -ENOMEM; 1150 1151 argc = parse_options(argc, argv, options, stat_usage, 1152 PARSE_OPT_STOP_AT_NON_OPTION); 1153 1154 if (csv_sep) 1155 csv_output = true; 1156 else 1157 csv_sep = DEFAULT_SEPARATOR; 1158 1159 /* 1160 * let the spreadsheet do the pretty-printing 1161 */ 1162 if (csv_output) { 1163 /* User explicitely passed -B? */ 1164 if (big_num_opt == 1) { 1165 fprintf(stderr, "-B option not supported with -x\n"); 1166 usage_with_options(stat_usage, options); 1167 } else /* Nope, so disable big number formatting */ 1168 big_num = false; 1169 } else if (big_num_opt == 0) /* User passed --no-big-num */ 1170 big_num = false; 1171 1172 if (!argc && target_pid == -1 && target_tid == -1) 1173 usage_with_options(stat_usage, options); 1174 if (run_count <= 0) 1175 usage_with_options(stat_usage, options); 1176 1177 /* no_aggr, cgroup are for system-wide only */ 1178 if ((no_aggr || nr_cgroups) && !system_wide) { 1179 fprintf(stderr, "both cgroup and no-aggregation " 1180 "modes only available in system-wide mode\n"); 1181 1182 usage_with_options(stat_usage, options); 1183 } 1184 1185 if (add_default_attributes()) 1186 goto out; 1187 1188 if (target_pid != -1) 1189 target_tid = target_pid; 1190 1191 evsel_list->threads = thread_map__new(target_pid, target_tid); 1192 if (evsel_list->threads == NULL) { 1193 pr_err("Problems finding threads of monitor\n"); 1194 usage_with_options(stat_usage, options); 1195 } 1196 1197 if (system_wide) 1198 evsel_list->cpus = cpu_map__new(cpu_list); 1199 else 1200 evsel_list->cpus = cpu_map__dummy_new(); 1201 1202 if (evsel_list->cpus == NULL) { 1203 perror("failed to parse CPUs map"); 1204 usage_with_options(stat_usage, options); 1205 return -1; 1206 } 1207 1208 list_for_each_entry(pos, &evsel_list->entries, node) { 1209 if (perf_evsel__alloc_stat_priv(pos) < 0 || 1210 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 || 1211 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0) 1212 goto out_free_fd; 1213 } 1214 1215 /* 1216 * We dont want to block the signals - that would cause 1217 * child tasks to inherit that and Ctrl-C would not work. 1218 * What we want is for Ctrl-C to work in the exec()-ed 1219 * task, but being ignored by perf stat itself: 1220 */ 1221 atexit(sig_atexit); 1222 signal(SIGINT, skip_signal); 1223 signal(SIGALRM, skip_signal); 1224 signal(SIGABRT, skip_signal); 1225 1226 status = 0; 1227 for (run_idx = 0; run_idx < run_count; run_idx++) { 1228 if (run_count != 1 && verbose) 1229 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); 1230 1231 if (sync_run) 1232 sync(); 1233 1234 status = run_perf_stat(argc, argv); 1235 } 1236 1237 if (status != -1) 1238 print_stat(argc, argv); 1239 out_free_fd: 1240 list_for_each_entry(pos, &evsel_list->entries, node) 1241 perf_evsel__free_stat_priv(pos); 1242 perf_evlist__delete_maps(evsel_list); 1243 out: 1244 perf_evlist__delete(evsel_list); 1245 return status; 1246 } 1247