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