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