1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * builtin-stat.c 4 * 5 * Builtin stat command: Give a precise performance counters summary 6 * overview about any workload, CPU or specific PID. 7 * 8 * Sample output: 9 10 $ perf stat ./hackbench 10 11 12 Time: 0.118 13 14 Performance counter stats for './hackbench 10': 15 16 1708.761321 task-clock # 11.037 CPUs utilized 17 41,190 context-switches # 0.024 M/sec 18 6,735 CPU-migrations # 0.004 M/sec 19 17,318 page-faults # 0.010 M/sec 20 5,205,202,243 cycles # 3.046 GHz 21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle 22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle 23 2,603,501,247 instructions # 0.50 insns per cycle 24 # 1.48 stalled cycles per insn 25 484,357,498 branches # 283.455 M/sec 26 6,388,934 branch-misses # 1.32% of all branches 27 28 0.154822978 seconds time elapsed 29 30 * 31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 32 * 33 * Improvements and fixes by: 34 * 35 * Arjan van de Ven <arjan@linux.intel.com> 36 * Yanmin Zhang <yanmin.zhang@intel.com> 37 * Wu Fengguang <fengguang.wu@intel.com> 38 * Mike Galbraith <efault@gmx.de> 39 * Paul Mackerras <paulus@samba.org> 40 * Jaswinder Singh Rajput <jaswinder@kernel.org> 41 */ 42 43 #include "builtin.h" 44 #include "perf.h" 45 #include "util/cgroup.h" 46 #include <subcmd/parse-options.h> 47 #include "util/parse-events.h" 48 #include "util/pmu.h" 49 #include "util/event.h" 50 #include "util/evlist.h" 51 #include "util/evsel.h" 52 #include "util/debug.h" 53 #include "util/color.h" 54 #include "util/stat.h" 55 #include "util/header.h" 56 #include "util/cpumap.h" 57 #include "util/thread_map.h" 58 #include "util/counts.h" 59 #include "util/group.h" 60 #include "util/session.h" 61 #include "util/tool.h" 62 #include "util/string2.h" 63 #include "util/metricgroup.h" 64 #include "util/synthetic-events.h" 65 #include "util/target.h" 66 #include "util/time-utils.h" 67 #include "util/top.h" 68 #include "util/affinity.h" 69 #include "asm/bug.h" 70 71 #include <linux/time64.h> 72 #include <linux/zalloc.h> 73 #include <api/fs/fs.h> 74 #include <errno.h> 75 #include <signal.h> 76 #include <stdlib.h> 77 #include <sys/prctl.h> 78 #include <inttypes.h> 79 #include <locale.h> 80 #include <math.h> 81 #include <sys/types.h> 82 #include <sys/stat.h> 83 #include <sys/wait.h> 84 #include <unistd.h> 85 #include <sys/time.h> 86 #include <sys/resource.h> 87 #include <linux/err.h> 88 89 #include <linux/ctype.h> 90 #include <perf/evlist.h> 91 92 #define DEFAULT_SEPARATOR " " 93 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi" 94 95 static void print_counters(struct timespec *ts, int argc, const char **argv); 96 97 /* Default events used for perf stat -T */ 98 static const char *transaction_attrs = { 99 "task-clock," 100 "{" 101 "instructions," 102 "cycles," 103 "cpu/cycles-t/," 104 "cpu/tx-start/," 105 "cpu/el-start/," 106 "cpu/cycles-ct/" 107 "}" 108 }; 109 110 /* More limited version when the CPU does not have all events. */ 111 static const char * transaction_limited_attrs = { 112 "task-clock," 113 "{" 114 "instructions," 115 "cycles," 116 "cpu/cycles-t/," 117 "cpu/tx-start/" 118 "}" 119 }; 120 121 static const char * topdown_attrs[] = { 122 "topdown-total-slots", 123 "topdown-slots-retired", 124 "topdown-recovery-bubbles", 125 "topdown-fetch-bubbles", 126 "topdown-slots-issued", 127 NULL, 128 }; 129 130 static const char *smi_cost_attrs = { 131 "{" 132 "msr/aperf/," 133 "msr/smi/," 134 "cycles" 135 "}" 136 }; 137 138 static struct evlist *evsel_list; 139 140 static struct target target = { 141 .uid = UINT_MAX, 142 }; 143 144 #define METRIC_ONLY_LEN 20 145 146 static volatile pid_t child_pid = -1; 147 static int detailed_run = 0; 148 static bool transaction_run; 149 static bool topdown_run = false; 150 static bool smi_cost = false; 151 static bool smi_reset = false; 152 static int big_num_opt = -1; 153 static bool group = false; 154 static const char *pre_cmd = NULL; 155 static const char *post_cmd = NULL; 156 static bool sync_run = false; 157 static bool forever = false; 158 static bool force_metric_only = false; 159 static struct timespec ref_time; 160 static bool append_file; 161 static bool interval_count; 162 static const char *output_name; 163 static int output_fd; 164 165 struct perf_stat { 166 bool record; 167 struct perf_data data; 168 struct perf_session *session; 169 u64 bytes_written; 170 struct perf_tool tool; 171 bool maps_allocated; 172 struct perf_cpu_map *cpus; 173 struct perf_thread_map *threads; 174 enum aggr_mode aggr_mode; 175 }; 176 177 static struct perf_stat perf_stat; 178 #define STAT_RECORD perf_stat.record 179 180 static volatile int done = 0; 181 182 static struct perf_stat_config stat_config = { 183 .aggr_mode = AGGR_GLOBAL, 184 .scale = true, 185 .unit_width = 4, /* strlen("unit") */ 186 .run_count = 1, 187 .metric_only_len = METRIC_ONLY_LEN, 188 .walltime_nsecs_stats = &walltime_nsecs_stats, 189 .big_num = true, 190 }; 191 192 static inline void diff_timespec(struct timespec *r, struct timespec *a, 193 struct timespec *b) 194 { 195 r->tv_sec = a->tv_sec - b->tv_sec; 196 if (a->tv_nsec < b->tv_nsec) { 197 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 198 r->tv_sec--; 199 } else { 200 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 201 } 202 } 203 204 static void perf_stat__reset_stats(void) 205 { 206 int i; 207 208 perf_evlist__reset_stats(evsel_list); 209 perf_stat__reset_shadow_stats(); 210 211 for (i = 0; i < stat_config.stats_num; i++) 212 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]); 213 } 214 215 static int process_synthesized_event(struct perf_tool *tool __maybe_unused, 216 union perf_event *event, 217 struct perf_sample *sample __maybe_unused, 218 struct machine *machine __maybe_unused) 219 { 220 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 221 pr_err("failed to write perf data, error: %m\n"); 222 return -1; 223 } 224 225 perf_stat.bytes_written += event->header.size; 226 return 0; 227 } 228 229 static int write_stat_round_event(u64 tm, u64 type) 230 { 231 return perf_event__synthesize_stat_round(NULL, tm, type, 232 process_synthesized_event, 233 NULL); 234 } 235 236 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 237 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 238 239 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 240 241 static int 242 perf_evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread, 243 struct perf_counts_values *count) 244 { 245 struct perf_sample_id *sid = SID(counter, cpu, thread); 246 247 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 248 process_synthesized_event, NULL); 249 } 250 251 static int read_single_counter(struct evsel *counter, int cpu, 252 int thread, struct timespec *rs) 253 { 254 if (counter->tool_event == PERF_TOOL_DURATION_TIME) { 255 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL; 256 struct perf_counts_values *count = 257 perf_counts(counter->counts, cpu, thread); 258 count->ena = count->run = val; 259 count->val = val; 260 return 0; 261 } 262 return perf_evsel__read_counter(counter, cpu, thread); 263 } 264 265 /* 266 * Read out the results of a single counter: 267 * do not aggregate counts across CPUs in system-wide mode 268 */ 269 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu) 270 { 271 int nthreads = perf_thread_map__nr(evsel_list->core.threads); 272 int thread; 273 274 if (!counter->supported) 275 return -ENOENT; 276 277 if (counter->core.system_wide) 278 nthreads = 1; 279 280 for (thread = 0; thread < nthreads; thread++) { 281 struct perf_counts_values *count; 282 283 count = perf_counts(counter->counts, cpu, thread); 284 285 /* 286 * The leader's group read loads data into its group members 287 * (via perf_evsel__read_counter()) and sets their count->loaded. 288 */ 289 if (!perf_counts__is_loaded(counter->counts, cpu, thread) && 290 read_single_counter(counter, cpu, thread, rs)) { 291 counter->counts->scaled = -1; 292 perf_counts(counter->counts, cpu, thread)->ena = 0; 293 perf_counts(counter->counts, cpu, thread)->run = 0; 294 return -1; 295 } 296 297 perf_counts__set_loaded(counter->counts, cpu, thread, false); 298 299 if (STAT_RECORD) { 300 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) { 301 pr_err("failed to write stat event\n"); 302 return -1; 303 } 304 } 305 306 if (verbose > 1) { 307 fprintf(stat_config.output, 308 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 309 perf_evsel__name(counter), 310 cpu, 311 count->val, count->ena, count->run); 312 } 313 } 314 315 return 0; 316 } 317 318 static void read_counters(struct timespec *rs) 319 { 320 struct evsel *counter; 321 struct affinity affinity; 322 int i, ncpus, cpu; 323 324 if (affinity__setup(&affinity) < 0) 325 return; 326 327 ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus); 328 if (!target__has_cpu(&target) || target__has_per_thread(&target)) 329 ncpus = 1; 330 evlist__for_each_cpu(evsel_list, i, cpu) { 331 if (i >= ncpus) 332 break; 333 affinity__set(&affinity, cpu); 334 335 evlist__for_each_entry(evsel_list, counter) { 336 if (evsel__cpu_iter_skip(counter, cpu)) 337 continue; 338 if (!counter->err) { 339 counter->err = read_counter_cpu(counter, rs, 340 counter->cpu_iter - 1); 341 } 342 } 343 } 344 affinity__cleanup(&affinity); 345 346 evlist__for_each_entry(evsel_list, counter) { 347 if (counter->err) 348 pr_debug("failed to read counter %s\n", counter->name); 349 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) 350 pr_warning("failed to process counter %s\n", counter->name); 351 counter->err = 0; 352 } 353 } 354 355 static void process_interval(void) 356 { 357 struct timespec ts, rs; 358 359 clock_gettime(CLOCK_MONOTONIC, &ts); 360 diff_timespec(&rs, &ts, &ref_time); 361 362 read_counters(&rs); 363 364 if (STAT_RECORD) { 365 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 366 pr_err("failed to write stat round event\n"); 367 } 368 369 init_stats(&walltime_nsecs_stats); 370 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000); 371 print_counters(&rs, 0, NULL); 372 } 373 374 static void enable_counters(void) 375 { 376 if (stat_config.initial_delay) 377 usleep(stat_config.initial_delay * USEC_PER_MSEC); 378 379 /* 380 * We need to enable counters only if: 381 * - we don't have tracee (attaching to task or cpu) 382 * - we have initial delay configured 383 */ 384 if (!target__none(&target) || stat_config.initial_delay) 385 evlist__enable(evsel_list); 386 } 387 388 static void disable_counters(void) 389 { 390 /* 391 * If we don't have tracee (attaching to task or cpu), counters may 392 * still be running. To get accurate group ratios, we must stop groups 393 * from counting before reading their constituent counters. 394 */ 395 if (!target__none(&target)) 396 evlist__disable(evsel_list); 397 } 398 399 static volatile int workload_exec_errno; 400 401 /* 402 * perf_evlist__prepare_workload will send a SIGUSR1 403 * if the fork fails, since we asked by setting its 404 * want_signal to true. 405 */ 406 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 407 void *ucontext __maybe_unused) 408 { 409 workload_exec_errno = info->si_value.sival_int; 410 } 411 412 static bool perf_evsel__should_store_id(struct evsel *counter) 413 { 414 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 415 } 416 417 static bool is_target_alive(struct target *_target, 418 struct perf_thread_map *threads) 419 { 420 struct stat st; 421 int i; 422 423 if (!target__has_task(_target)) 424 return true; 425 426 for (i = 0; i < threads->nr; i++) { 427 char path[PATH_MAX]; 428 429 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 430 threads->map[i].pid); 431 432 if (!stat(path, &st)) 433 return true; 434 } 435 436 return false; 437 } 438 439 enum counter_recovery { 440 COUNTER_SKIP, 441 COUNTER_RETRY, 442 COUNTER_FATAL, 443 }; 444 445 static enum counter_recovery stat_handle_error(struct evsel *counter) 446 { 447 char msg[BUFSIZ]; 448 /* 449 * PPC returns ENXIO for HW counters until 2.6.37 450 * (behavior changed with commit b0a873e). 451 */ 452 if (errno == EINVAL || errno == ENOSYS || 453 errno == ENOENT || errno == EOPNOTSUPP || 454 errno == ENXIO) { 455 if (verbose > 0) 456 ui__warning("%s event is not supported by the kernel.\n", 457 perf_evsel__name(counter)); 458 counter->supported = false; 459 /* 460 * errored is a sticky flag that means one of the counter's 461 * cpu event had a problem and needs to be reexamined. 462 */ 463 counter->errored = true; 464 465 if ((counter->leader != counter) || 466 !(counter->leader->core.nr_members > 1)) 467 return COUNTER_SKIP; 468 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { 469 if (verbose > 0) 470 ui__warning("%s\n", msg); 471 return COUNTER_RETRY; 472 } else if (target__has_per_thread(&target) && 473 evsel_list->core.threads && 474 evsel_list->core.threads->err_thread != -1) { 475 /* 476 * For global --per-thread case, skip current 477 * error thread. 478 */ 479 if (!thread_map__remove(evsel_list->core.threads, 480 evsel_list->core.threads->err_thread)) { 481 evsel_list->core.threads->err_thread = -1; 482 return COUNTER_RETRY; 483 } 484 } 485 486 perf_evsel__open_strerror(counter, &target, 487 errno, msg, sizeof(msg)); 488 ui__error("%s\n", msg); 489 490 if (child_pid != -1) 491 kill(child_pid, SIGTERM); 492 return COUNTER_FATAL; 493 } 494 495 static int __run_perf_stat(int argc, const char **argv, int run_idx) 496 { 497 int interval = stat_config.interval; 498 int times = stat_config.times; 499 int timeout = stat_config.timeout; 500 char msg[BUFSIZ]; 501 unsigned long long t0, t1; 502 struct evsel *counter; 503 struct timespec ts; 504 size_t l; 505 int status = 0; 506 const bool forks = (argc > 0); 507 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; 508 struct affinity affinity; 509 int i, cpu; 510 bool second_pass = false; 511 512 if (interval) { 513 ts.tv_sec = interval / USEC_PER_MSEC; 514 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC; 515 } else if (timeout) { 516 ts.tv_sec = timeout / USEC_PER_MSEC; 517 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC; 518 } else { 519 ts.tv_sec = 1; 520 ts.tv_nsec = 0; 521 } 522 523 if (forks) { 524 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, 525 workload_exec_failed_signal) < 0) { 526 perror("failed to prepare workload"); 527 return -1; 528 } 529 child_pid = evsel_list->workload.pid; 530 } 531 532 if (group) 533 perf_evlist__set_leader(evsel_list); 534 535 if (affinity__setup(&affinity) < 0) 536 return -1; 537 538 evlist__for_each_cpu (evsel_list, i, cpu) { 539 affinity__set(&affinity, cpu); 540 541 evlist__for_each_entry(evsel_list, counter) { 542 if (evsel__cpu_iter_skip(counter, cpu)) 543 continue; 544 if (counter->reset_group || counter->errored) 545 continue; 546 try_again: 547 if (create_perf_stat_counter(counter, &stat_config, &target, 548 counter->cpu_iter - 1) < 0) { 549 550 /* 551 * Weak group failed. We cannot just undo this here 552 * because earlier CPUs might be in group mode, and the kernel 553 * doesn't support mixing group and non group reads. Defer 554 * it to later. 555 * Don't close here because we're in the wrong affinity. 556 */ 557 if ((errno == EINVAL || errno == EBADF) && 558 counter->leader != counter && 559 counter->weak_group) { 560 perf_evlist__reset_weak_group(evsel_list, counter, false); 561 assert(counter->reset_group); 562 second_pass = true; 563 continue; 564 } 565 566 switch (stat_handle_error(counter)) { 567 case COUNTER_FATAL: 568 return -1; 569 case COUNTER_RETRY: 570 goto try_again; 571 case COUNTER_SKIP: 572 continue; 573 default: 574 break; 575 } 576 577 } 578 counter->supported = true; 579 } 580 } 581 582 if (second_pass) { 583 /* 584 * Now redo all the weak group after closing them, 585 * and also close errored counters. 586 */ 587 588 evlist__for_each_cpu(evsel_list, i, cpu) { 589 affinity__set(&affinity, cpu); 590 /* First close errored or weak retry */ 591 evlist__for_each_entry(evsel_list, counter) { 592 if (!counter->reset_group && !counter->errored) 593 continue; 594 if (evsel__cpu_iter_skip_no_inc(counter, cpu)) 595 continue; 596 perf_evsel__close_cpu(&counter->core, counter->cpu_iter); 597 } 598 /* Now reopen weak */ 599 evlist__for_each_entry(evsel_list, counter) { 600 if (!counter->reset_group && !counter->errored) 601 continue; 602 if (evsel__cpu_iter_skip(counter, cpu)) 603 continue; 604 if (!counter->reset_group) 605 continue; 606 try_again_reset: 607 pr_debug2("reopening weak %s\n", perf_evsel__name(counter)); 608 if (create_perf_stat_counter(counter, &stat_config, &target, 609 counter->cpu_iter - 1) < 0) { 610 611 switch (stat_handle_error(counter)) { 612 case COUNTER_FATAL: 613 return -1; 614 case COUNTER_RETRY: 615 goto try_again_reset; 616 case COUNTER_SKIP: 617 continue; 618 default: 619 break; 620 } 621 } 622 counter->supported = true; 623 } 624 } 625 } 626 affinity__cleanup(&affinity); 627 628 evlist__for_each_entry(evsel_list, counter) { 629 if (!counter->supported) { 630 perf_evsel__free_fd(&counter->core); 631 continue; 632 } 633 634 l = strlen(counter->unit); 635 if (l > stat_config.unit_width) 636 stat_config.unit_width = l; 637 638 if (perf_evsel__should_store_id(counter) && 639 perf_evsel__store_ids(counter, evsel_list)) 640 return -1; 641 } 642 643 if (perf_evlist__apply_filters(evsel_list, &counter)) { 644 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n", 645 counter->filter, perf_evsel__name(counter), errno, 646 str_error_r(errno, msg, sizeof(msg))); 647 return -1; 648 } 649 650 if (STAT_RECORD) { 651 int err, fd = perf_data__fd(&perf_stat.data); 652 653 if (is_pipe) { 654 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); 655 } else { 656 err = perf_session__write_header(perf_stat.session, evsel_list, 657 fd, false); 658 } 659 660 if (err < 0) 661 return err; 662 663 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list, 664 process_synthesized_event, is_pipe); 665 if (err < 0) 666 return err; 667 } 668 669 /* 670 * Enable counters and exec the command: 671 */ 672 t0 = rdclock(); 673 clock_gettime(CLOCK_MONOTONIC, &ref_time); 674 675 if (forks) { 676 perf_evlist__start_workload(evsel_list); 677 enable_counters(); 678 679 if (interval || timeout) { 680 while (!waitpid(child_pid, &status, WNOHANG)) { 681 nanosleep(&ts, NULL); 682 if (timeout) 683 break; 684 process_interval(); 685 if (interval_count && !(--times)) 686 break; 687 } 688 } 689 if (child_pid != -1) 690 wait4(child_pid, &status, 0, &stat_config.ru_data); 691 692 if (workload_exec_errno) { 693 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); 694 pr_err("Workload failed: %s\n", emsg); 695 return -1; 696 } 697 698 if (WIFSIGNALED(status)) 699 psignal(WTERMSIG(status), argv[0]); 700 } else { 701 enable_counters(); 702 while (!done) { 703 nanosleep(&ts, NULL); 704 if (!is_target_alive(&target, evsel_list->core.threads)) 705 break; 706 if (timeout) 707 break; 708 if (interval) { 709 process_interval(); 710 if (interval_count && !(--times)) 711 break; 712 } 713 } 714 } 715 716 disable_counters(); 717 718 t1 = rdclock(); 719 720 if (stat_config.walltime_run_table) 721 stat_config.walltime_run[run_idx] = t1 - t0; 722 723 update_stats(&walltime_nsecs_stats, t1 - t0); 724 725 /* 726 * Closing a group leader splits the group, and as we only disable 727 * group leaders, results in remaining events becoming enabled. To 728 * avoid arbitrary skew, we must read all counters before closing any 729 * group leaders. 730 */ 731 read_counters(&(struct timespec) { .tv_nsec = t1-t0 }); 732 733 /* 734 * We need to keep evsel_list alive, because it's processed 735 * later the evsel_list will be closed after. 736 */ 737 if (!STAT_RECORD) 738 evlist__close(evsel_list); 739 740 return WEXITSTATUS(status); 741 } 742 743 static int run_perf_stat(int argc, const char **argv, int run_idx) 744 { 745 int ret; 746 747 if (pre_cmd) { 748 ret = system(pre_cmd); 749 if (ret) 750 return ret; 751 } 752 753 if (sync_run) 754 sync(); 755 756 ret = __run_perf_stat(argc, argv, run_idx); 757 if (ret) 758 return ret; 759 760 if (post_cmd) { 761 ret = system(post_cmd); 762 if (ret) 763 return ret; 764 } 765 766 return ret; 767 } 768 769 static void print_counters(struct timespec *ts, int argc, const char **argv) 770 { 771 /* Do not print anything if we record to the pipe. */ 772 if (STAT_RECORD && perf_stat.data.is_pipe) 773 return; 774 775 perf_evlist__print_counters(evsel_list, &stat_config, &target, 776 ts, argc, argv); 777 } 778 779 static volatile int signr = -1; 780 781 static void skip_signal(int signo) 782 { 783 if ((child_pid == -1) || stat_config.interval) 784 done = 1; 785 786 signr = signo; 787 /* 788 * render child_pid harmless 789 * won't send SIGTERM to a random 790 * process in case of race condition 791 * and fast PID recycling 792 */ 793 child_pid = -1; 794 } 795 796 static void sig_atexit(void) 797 { 798 sigset_t set, oset; 799 800 /* 801 * avoid race condition with SIGCHLD handler 802 * in skip_signal() which is modifying child_pid 803 * goal is to avoid send SIGTERM to a random 804 * process 805 */ 806 sigemptyset(&set); 807 sigaddset(&set, SIGCHLD); 808 sigprocmask(SIG_BLOCK, &set, &oset); 809 810 if (child_pid != -1) 811 kill(child_pid, SIGTERM); 812 813 sigprocmask(SIG_SETMASK, &oset, NULL); 814 815 if (signr == -1) 816 return; 817 818 signal(signr, SIG_DFL); 819 kill(getpid(), signr); 820 } 821 822 static int stat__set_big_num(const struct option *opt __maybe_unused, 823 const char *s __maybe_unused, int unset) 824 { 825 big_num_opt = unset ? 0 : 1; 826 return 0; 827 } 828 829 static int enable_metric_only(const struct option *opt __maybe_unused, 830 const char *s __maybe_unused, int unset) 831 { 832 force_metric_only = true; 833 stat_config.metric_only = !unset; 834 return 0; 835 } 836 837 static int parse_metric_groups(const struct option *opt, 838 const char *str, 839 int unset __maybe_unused) 840 { 841 return metricgroup__parse_groups(opt, str, &stat_config.metric_events); 842 } 843 844 static struct option stat_options[] = { 845 OPT_BOOLEAN('T', "transaction", &transaction_run, 846 "hardware transaction statistics"), 847 OPT_CALLBACK('e', "event", &evsel_list, "event", 848 "event selector. use 'perf list' to list available events", 849 parse_events_option), 850 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 851 "event filter", parse_filter), 852 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 853 "child tasks do not inherit counters"), 854 OPT_STRING('p', "pid", &target.pid, "pid", 855 "stat events on existing process id"), 856 OPT_STRING('t', "tid", &target.tid, "tid", 857 "stat events on existing thread id"), 858 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 859 "system-wide collection from all CPUs"), 860 OPT_BOOLEAN('g', "group", &group, 861 "put the counters into a counter group"), 862 OPT_BOOLEAN(0, "scale", &stat_config.scale, 863 "Use --no-scale to disable counter scaling for multiplexing"), 864 OPT_INCR('v', "verbose", &verbose, 865 "be more verbose (show counter open errors, etc)"), 866 OPT_INTEGER('r', "repeat", &stat_config.run_count, 867 "repeat command and print average + stddev (max: 100, forever: 0)"), 868 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 869 "display details about each run (only with -r option)"), 870 OPT_BOOLEAN('n', "null", &stat_config.null_run, 871 "null run - dont start any counters"), 872 OPT_INCR('d', "detailed", &detailed_run, 873 "detailed run - start a lot of events"), 874 OPT_BOOLEAN('S', "sync", &sync_run, 875 "call sync() before starting a run"), 876 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 877 "print large numbers with thousands\' separators", 878 stat__set_big_num), 879 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 880 "list of cpus to monitor in system-wide"), 881 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, 882 "disable CPU count aggregation", AGGR_NONE), 883 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"), 884 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 885 "print counts with custom separator"), 886 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 887 "monitor event in cgroup name only", parse_cgroups), 888 OPT_STRING('o', "output", &output_name, "file", "output file name"), 889 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 890 OPT_INTEGER(0, "log-fd", &output_fd, 891 "log output to fd, instead of stderr"), 892 OPT_STRING(0, "pre", &pre_cmd, "command", 893 "command to run prior to the measured command"), 894 OPT_STRING(0, "post", &post_cmd, "command", 895 "command to run after to the measured command"), 896 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 897 "print counts at regular interval in ms " 898 "(overhead is possible for values <= 100ms)"), 899 OPT_INTEGER(0, "interval-count", &stat_config.times, 900 "print counts for fixed number of times"), 901 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 902 "clear screen in between new interval"), 903 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 904 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 905 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, 906 "aggregate counts per processor socket", AGGR_SOCKET), 907 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode, 908 "aggregate counts per processor die", AGGR_DIE), 909 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, 910 "aggregate counts per physical processor core", AGGR_CORE), 911 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, 912 "aggregate counts per thread", AGGR_THREAD), 913 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode, 914 "aggregate counts per numa node", AGGR_NODE), 915 OPT_UINTEGER('D', "delay", &stat_config.initial_delay, 916 "ms to wait before starting measurement after program start"), 917 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 918 "Only print computed metrics. No raw values", enable_metric_only), 919 OPT_BOOLEAN(0, "topdown", &topdown_run, 920 "measure topdown level 1 statistics"), 921 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 922 "measure SMI cost"), 923 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 924 "monitor specified metrics or metric groups (separated by ,)", 925 parse_metric_groups), 926 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, 927 "Configure all used events to run in kernel space.", 928 PARSE_OPT_EXCLUSIVE), 929 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, 930 "Configure all used events to run in user space.", 931 PARSE_OPT_EXCLUSIVE), 932 OPT_END() 933 }; 934 935 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 936 struct perf_cpu_map *map, int cpu) 937 { 938 return cpu_map__get_socket(map, cpu, NULL); 939 } 940 941 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 942 struct perf_cpu_map *map, int cpu) 943 { 944 return cpu_map__get_die(map, cpu, NULL); 945 } 946 947 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 948 struct perf_cpu_map *map, int cpu) 949 { 950 return cpu_map__get_core(map, cpu, NULL); 951 } 952 953 static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused, 954 struct perf_cpu_map *map, int cpu) 955 { 956 return cpu_map__get_node(map, cpu, NULL); 957 } 958 959 static int perf_stat__get_aggr(struct perf_stat_config *config, 960 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx) 961 { 962 int cpu; 963 964 if (idx >= map->nr) 965 return -1; 966 967 cpu = map->map[idx]; 968 969 if (config->cpus_aggr_map->map[cpu] == -1) 970 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx); 971 972 return config->cpus_aggr_map->map[cpu]; 973 } 974 975 static int perf_stat__get_socket_cached(struct perf_stat_config *config, 976 struct perf_cpu_map *map, int idx) 977 { 978 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx); 979 } 980 981 static int perf_stat__get_die_cached(struct perf_stat_config *config, 982 struct perf_cpu_map *map, int idx) 983 { 984 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx); 985 } 986 987 static int perf_stat__get_core_cached(struct perf_stat_config *config, 988 struct perf_cpu_map *map, int idx) 989 { 990 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx); 991 } 992 993 static int perf_stat__get_node_cached(struct perf_stat_config *config, 994 struct perf_cpu_map *map, int idx) 995 { 996 return perf_stat__get_aggr(config, perf_stat__get_node, map, idx); 997 } 998 999 static bool term_percore_set(void) 1000 { 1001 struct evsel *counter; 1002 1003 evlist__for_each_entry(evsel_list, counter) { 1004 if (counter->percore) 1005 return true; 1006 } 1007 1008 return false; 1009 } 1010 1011 static int perf_stat_init_aggr_mode(void) 1012 { 1013 int nr; 1014 1015 switch (stat_config.aggr_mode) { 1016 case AGGR_SOCKET: 1017 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1018 perror("cannot build socket map"); 1019 return -1; 1020 } 1021 stat_config.aggr_get_id = perf_stat__get_socket_cached; 1022 break; 1023 case AGGR_DIE: 1024 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1025 perror("cannot build die map"); 1026 return -1; 1027 } 1028 stat_config.aggr_get_id = perf_stat__get_die_cached; 1029 break; 1030 case AGGR_CORE: 1031 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1032 perror("cannot build core map"); 1033 return -1; 1034 } 1035 stat_config.aggr_get_id = perf_stat__get_core_cached; 1036 break; 1037 case AGGR_NODE: 1038 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1039 perror("cannot build core map"); 1040 return -1; 1041 } 1042 stat_config.aggr_get_id = perf_stat__get_node_cached; 1043 break; 1044 case AGGR_NONE: 1045 if (term_percore_set()) { 1046 if (cpu_map__build_core_map(evsel_list->core.cpus, 1047 &stat_config.aggr_map)) { 1048 perror("cannot build core map"); 1049 return -1; 1050 } 1051 stat_config.aggr_get_id = perf_stat__get_core_cached; 1052 } 1053 break; 1054 case AGGR_GLOBAL: 1055 case AGGR_THREAD: 1056 case AGGR_UNSET: 1057 default: 1058 break; 1059 } 1060 1061 /* 1062 * The evsel_list->cpus is the base we operate on, 1063 * taking the highest cpu number to be the size of 1064 * the aggregation translate cpumap. 1065 */ 1066 nr = perf_cpu_map__max(evsel_list->core.cpus); 1067 stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1); 1068 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 1069 } 1070 1071 static void perf_stat__exit_aggr_mode(void) 1072 { 1073 perf_cpu_map__put(stat_config.aggr_map); 1074 perf_cpu_map__put(stat_config.cpus_aggr_map); 1075 stat_config.aggr_map = NULL; 1076 stat_config.cpus_aggr_map = NULL; 1077 } 1078 1079 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx) 1080 { 1081 int cpu; 1082 1083 if (idx > map->nr) 1084 return -1; 1085 1086 cpu = map->map[idx]; 1087 1088 if (cpu >= env->nr_cpus_avail) 1089 return -1; 1090 1091 return cpu; 1092 } 1093 1094 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data) 1095 { 1096 struct perf_env *env = data; 1097 int cpu = perf_env__get_cpu(env, map, idx); 1098 1099 return cpu == -1 ? -1 : env->cpu[cpu].socket_id; 1100 } 1101 1102 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data) 1103 { 1104 struct perf_env *env = data; 1105 int die_id = -1, cpu = perf_env__get_cpu(env, map, idx); 1106 1107 if (cpu != -1) { 1108 /* 1109 * Encode socket in bit range 15:8 1110 * die_id is relative to socket, 1111 * we need a global id. So we combine 1112 * socket + die id 1113 */ 1114 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n")) 1115 return -1; 1116 1117 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n")) 1118 return -1; 1119 1120 die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff); 1121 } 1122 1123 return die_id; 1124 } 1125 1126 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data) 1127 { 1128 struct perf_env *env = data; 1129 int core = -1, cpu = perf_env__get_cpu(env, map, idx); 1130 1131 if (cpu != -1) { 1132 /* 1133 * Encode socket in bit range 31:24 1134 * encode die id in bit range 23:16 1135 * core_id is relative to socket and die, 1136 * we need a global id. So we combine 1137 * socket + die id + core id 1138 */ 1139 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n")) 1140 return -1; 1141 1142 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n")) 1143 return -1; 1144 1145 if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n")) 1146 return -1; 1147 1148 core = (env->cpu[cpu].socket_id << 24) | 1149 (env->cpu[cpu].die_id << 16) | 1150 (env->cpu[cpu].core_id & 0xffff); 1151 } 1152 1153 return core; 1154 } 1155 1156 static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data) 1157 { 1158 int cpu = perf_env__get_cpu(data, map, idx); 1159 1160 return perf_env__numa_node(data, cpu); 1161 } 1162 1163 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus, 1164 struct perf_cpu_map **sockp) 1165 { 1166 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); 1167 } 1168 1169 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus, 1170 struct perf_cpu_map **diep) 1171 { 1172 return cpu_map__build_map(cpus, diep, perf_env__get_die, env); 1173 } 1174 1175 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus, 1176 struct perf_cpu_map **corep) 1177 { 1178 return cpu_map__build_map(cpus, corep, perf_env__get_core, env); 1179 } 1180 1181 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus, 1182 struct perf_cpu_map **nodep) 1183 { 1184 return cpu_map__build_map(cpus, nodep, perf_env__get_node, env); 1185 } 1186 1187 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1188 struct perf_cpu_map *map, int idx) 1189 { 1190 return perf_env__get_socket(map, idx, &perf_stat.session->header.env); 1191 } 1192 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1193 struct perf_cpu_map *map, int idx) 1194 { 1195 return perf_env__get_die(map, idx, &perf_stat.session->header.env); 1196 } 1197 1198 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1199 struct perf_cpu_map *map, int idx) 1200 { 1201 return perf_env__get_core(map, idx, &perf_stat.session->header.env); 1202 } 1203 1204 static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, 1205 struct perf_cpu_map *map, int idx) 1206 { 1207 return perf_env__get_node(map, idx, &perf_stat.session->header.env); 1208 } 1209 1210 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1211 { 1212 struct perf_env *env = &st->session->header.env; 1213 1214 switch (stat_config.aggr_mode) { 1215 case AGGR_SOCKET: 1216 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1217 perror("cannot build socket map"); 1218 return -1; 1219 } 1220 stat_config.aggr_get_id = perf_stat__get_socket_file; 1221 break; 1222 case AGGR_DIE: 1223 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1224 perror("cannot build die map"); 1225 return -1; 1226 } 1227 stat_config.aggr_get_id = perf_stat__get_die_file; 1228 break; 1229 case AGGR_CORE: 1230 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1231 perror("cannot build core map"); 1232 return -1; 1233 } 1234 stat_config.aggr_get_id = perf_stat__get_core_file; 1235 break; 1236 case AGGR_NODE: 1237 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1238 perror("cannot build core map"); 1239 return -1; 1240 } 1241 stat_config.aggr_get_id = perf_stat__get_node_file; 1242 break; 1243 case AGGR_NONE: 1244 case AGGR_GLOBAL: 1245 case AGGR_THREAD: 1246 case AGGR_UNSET: 1247 default: 1248 break; 1249 } 1250 1251 return 0; 1252 } 1253 1254 static int topdown_filter_events(const char **attr, char **str, bool use_group) 1255 { 1256 int off = 0; 1257 int i; 1258 int len = 0; 1259 char *s; 1260 1261 for (i = 0; attr[i]; i++) { 1262 if (pmu_have_event("cpu", attr[i])) { 1263 len += strlen(attr[i]) + 1; 1264 attr[i - off] = attr[i]; 1265 } else 1266 off++; 1267 } 1268 attr[i - off] = NULL; 1269 1270 *str = malloc(len + 1 + 2); 1271 if (!*str) 1272 return -1; 1273 s = *str; 1274 if (i - off == 0) { 1275 *s = 0; 1276 return 0; 1277 } 1278 if (use_group) 1279 *s++ = '{'; 1280 for (i = 0; attr[i]; i++) { 1281 strcpy(s, attr[i]); 1282 s += strlen(s); 1283 *s++ = ','; 1284 } 1285 if (use_group) { 1286 s[-1] = '}'; 1287 *s = 0; 1288 } else 1289 s[-1] = 0; 1290 return 0; 1291 } 1292 1293 __weak bool arch_topdown_check_group(bool *warn) 1294 { 1295 *warn = false; 1296 return false; 1297 } 1298 1299 __weak void arch_topdown_group_warn(void) 1300 { 1301 } 1302 1303 /* 1304 * Add default attributes, if there were no attributes specified or 1305 * if -d/--detailed, -d -d or -d -d -d is used: 1306 */ 1307 static int add_default_attributes(void) 1308 { 1309 int err; 1310 struct perf_event_attr default_attrs0[] = { 1311 1312 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 1313 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 1314 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 1315 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 1316 1317 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 1318 }; 1319 struct perf_event_attr frontend_attrs[] = { 1320 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 1321 }; 1322 struct perf_event_attr backend_attrs[] = { 1323 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 1324 }; 1325 struct perf_event_attr default_attrs1[] = { 1326 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 1327 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 1328 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 1329 1330 }; 1331 1332 /* 1333 * Detailed stats (-d), covering the L1 and last level data caches: 1334 */ 1335 struct perf_event_attr detailed_attrs[] = { 1336 1337 { .type = PERF_TYPE_HW_CACHE, 1338 .config = 1339 PERF_COUNT_HW_CACHE_L1D << 0 | 1340 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1341 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1342 1343 { .type = PERF_TYPE_HW_CACHE, 1344 .config = 1345 PERF_COUNT_HW_CACHE_L1D << 0 | 1346 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1347 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1348 1349 { .type = PERF_TYPE_HW_CACHE, 1350 .config = 1351 PERF_COUNT_HW_CACHE_LL << 0 | 1352 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1353 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1354 1355 { .type = PERF_TYPE_HW_CACHE, 1356 .config = 1357 PERF_COUNT_HW_CACHE_LL << 0 | 1358 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1359 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1360 }; 1361 1362 /* 1363 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 1364 */ 1365 struct perf_event_attr very_detailed_attrs[] = { 1366 1367 { .type = PERF_TYPE_HW_CACHE, 1368 .config = 1369 PERF_COUNT_HW_CACHE_L1I << 0 | 1370 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1371 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1372 1373 { .type = PERF_TYPE_HW_CACHE, 1374 .config = 1375 PERF_COUNT_HW_CACHE_L1I << 0 | 1376 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1377 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1378 1379 { .type = PERF_TYPE_HW_CACHE, 1380 .config = 1381 PERF_COUNT_HW_CACHE_DTLB << 0 | 1382 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1383 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1384 1385 { .type = PERF_TYPE_HW_CACHE, 1386 .config = 1387 PERF_COUNT_HW_CACHE_DTLB << 0 | 1388 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1389 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1390 1391 { .type = PERF_TYPE_HW_CACHE, 1392 .config = 1393 PERF_COUNT_HW_CACHE_ITLB << 0 | 1394 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1395 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1396 1397 { .type = PERF_TYPE_HW_CACHE, 1398 .config = 1399 PERF_COUNT_HW_CACHE_ITLB << 0 | 1400 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1401 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1402 1403 }; 1404 1405 /* 1406 * Very, very detailed stats (-d -d -d), adding prefetch events: 1407 */ 1408 struct perf_event_attr very_very_detailed_attrs[] = { 1409 1410 { .type = PERF_TYPE_HW_CACHE, 1411 .config = 1412 PERF_COUNT_HW_CACHE_L1D << 0 | 1413 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1414 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1415 1416 { .type = PERF_TYPE_HW_CACHE, 1417 .config = 1418 PERF_COUNT_HW_CACHE_L1D << 0 | 1419 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1420 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1421 }; 1422 struct parse_events_error errinfo; 1423 1424 /* Set attrs if no event is selected and !null_run: */ 1425 if (stat_config.null_run) 1426 return 0; 1427 1428 bzero(&errinfo, sizeof(errinfo)); 1429 if (transaction_run) { 1430 /* Handle -T as -M transaction. Once platform specific metrics 1431 * support has been added to the json files, all archictures 1432 * will use this approach. To determine transaction support 1433 * on an architecture test for such a metric name. 1434 */ 1435 if (metricgroup__has_metric("transaction")) { 1436 struct option opt = { .value = &evsel_list }; 1437 1438 return metricgroup__parse_groups(&opt, "transaction", 1439 &stat_config.metric_events); 1440 } 1441 1442 if (pmu_have_event("cpu", "cycles-ct") && 1443 pmu_have_event("cpu", "el-start")) 1444 err = parse_events(evsel_list, transaction_attrs, 1445 &errinfo); 1446 else 1447 err = parse_events(evsel_list, 1448 transaction_limited_attrs, 1449 &errinfo); 1450 if (err) { 1451 fprintf(stderr, "Cannot set up transaction events\n"); 1452 parse_events_print_error(&errinfo, transaction_attrs); 1453 return -1; 1454 } 1455 return 0; 1456 } 1457 1458 if (smi_cost) { 1459 int smi; 1460 1461 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 1462 fprintf(stderr, "freeze_on_smi is not supported.\n"); 1463 return -1; 1464 } 1465 1466 if (!smi) { 1467 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 1468 fprintf(stderr, "Failed to set freeze_on_smi.\n"); 1469 return -1; 1470 } 1471 smi_reset = true; 1472 } 1473 1474 if (pmu_have_event("msr", "aperf") && 1475 pmu_have_event("msr", "smi")) { 1476 if (!force_metric_only) 1477 stat_config.metric_only = true; 1478 err = parse_events(evsel_list, smi_cost_attrs, &errinfo); 1479 } else { 1480 fprintf(stderr, "To measure SMI cost, it needs " 1481 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n"); 1482 parse_events_print_error(&errinfo, smi_cost_attrs); 1483 return -1; 1484 } 1485 if (err) { 1486 parse_events_print_error(&errinfo, smi_cost_attrs); 1487 fprintf(stderr, "Cannot set up SMI cost events\n"); 1488 return -1; 1489 } 1490 return 0; 1491 } 1492 1493 if (topdown_run) { 1494 char *str = NULL; 1495 bool warn = false; 1496 1497 if (stat_config.aggr_mode != AGGR_GLOBAL && 1498 stat_config.aggr_mode != AGGR_CORE) { 1499 pr_err("top down event configuration requires --per-core mode\n"); 1500 return -1; 1501 } 1502 stat_config.aggr_mode = AGGR_CORE; 1503 if (nr_cgroups || !target__has_cpu(&target)) { 1504 pr_err("top down event configuration requires system-wide mode (-a)\n"); 1505 return -1; 1506 } 1507 1508 if (!force_metric_only) 1509 stat_config.metric_only = true; 1510 if (topdown_filter_events(topdown_attrs, &str, 1511 arch_topdown_check_group(&warn)) < 0) { 1512 pr_err("Out of memory\n"); 1513 return -1; 1514 } 1515 if (topdown_attrs[0] && str) { 1516 if (warn) 1517 arch_topdown_group_warn(); 1518 err = parse_events(evsel_list, str, &errinfo); 1519 if (err) { 1520 fprintf(stderr, 1521 "Cannot set up top down events %s: %d\n", 1522 str, err); 1523 parse_events_print_error(&errinfo, str); 1524 free(str); 1525 return -1; 1526 } 1527 } else { 1528 fprintf(stderr, "System does not support topdown\n"); 1529 return -1; 1530 } 1531 free(str); 1532 } 1533 1534 if (!evsel_list->core.nr_entries) { 1535 if (target__has_cpu(&target)) 1536 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; 1537 1538 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0) 1539 return -1; 1540 if (pmu_have_event("cpu", "stalled-cycles-frontend")) { 1541 if (perf_evlist__add_default_attrs(evsel_list, 1542 frontend_attrs) < 0) 1543 return -1; 1544 } 1545 if (pmu_have_event("cpu", "stalled-cycles-backend")) { 1546 if (perf_evlist__add_default_attrs(evsel_list, 1547 backend_attrs) < 0) 1548 return -1; 1549 } 1550 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0) 1551 return -1; 1552 } 1553 1554 /* Detailed events get appended to the event list: */ 1555 1556 if (detailed_run < 1) 1557 return 0; 1558 1559 /* Append detailed run extra attributes: */ 1560 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) 1561 return -1; 1562 1563 if (detailed_run < 2) 1564 return 0; 1565 1566 /* Append very detailed run extra attributes: */ 1567 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) 1568 return -1; 1569 1570 if (detailed_run < 3) 1571 return 0; 1572 1573 /* Append very, very detailed run extra attributes: */ 1574 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); 1575 } 1576 1577 static const char * const stat_record_usage[] = { 1578 "perf stat record [<options>]", 1579 NULL, 1580 }; 1581 1582 static void init_features(struct perf_session *session) 1583 { 1584 int feat; 1585 1586 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 1587 perf_header__set_feat(&session->header, feat); 1588 1589 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 1590 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 1591 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 1592 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 1593 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 1594 } 1595 1596 static int __cmd_record(int argc, const char **argv) 1597 { 1598 struct perf_session *session; 1599 struct perf_data *data = &perf_stat.data; 1600 1601 argc = parse_options(argc, argv, stat_options, stat_record_usage, 1602 PARSE_OPT_STOP_AT_NON_OPTION); 1603 1604 if (output_name) 1605 data->path = output_name; 1606 1607 if (stat_config.run_count != 1 || forever) { 1608 pr_err("Cannot use -r option with perf stat record.\n"); 1609 return -1; 1610 } 1611 1612 session = perf_session__new(data, false, NULL); 1613 if (IS_ERR(session)) { 1614 pr_err("Perf session creation failed\n"); 1615 return PTR_ERR(session); 1616 } 1617 1618 init_features(session); 1619 1620 session->evlist = evsel_list; 1621 perf_stat.session = session; 1622 perf_stat.record = true; 1623 return argc; 1624 } 1625 1626 static int process_stat_round_event(struct perf_session *session, 1627 union perf_event *event) 1628 { 1629 struct perf_record_stat_round *stat_round = &event->stat_round; 1630 struct evsel *counter; 1631 struct timespec tsh, *ts = NULL; 1632 const char **argv = session->header.env.cmdline_argv; 1633 int argc = session->header.env.nr_cmdline; 1634 1635 evlist__for_each_entry(evsel_list, counter) 1636 perf_stat_process_counter(&stat_config, counter); 1637 1638 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 1639 update_stats(&walltime_nsecs_stats, stat_round->time); 1640 1641 if (stat_config.interval && stat_round->time) { 1642 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 1643 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 1644 ts = &tsh; 1645 } 1646 1647 print_counters(ts, argc, argv); 1648 return 0; 1649 } 1650 1651 static 1652 int process_stat_config_event(struct perf_session *session, 1653 union perf_event *event) 1654 { 1655 struct perf_tool *tool = session->tool; 1656 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 1657 1658 perf_event__read_stat_config(&stat_config, &event->stat_config); 1659 1660 if (perf_cpu_map__empty(st->cpus)) { 1661 if (st->aggr_mode != AGGR_UNSET) 1662 pr_warning("warning: processing task data, aggregation mode not set\n"); 1663 return 0; 1664 } 1665 1666 if (st->aggr_mode != AGGR_UNSET) 1667 stat_config.aggr_mode = st->aggr_mode; 1668 1669 if (perf_stat.data.is_pipe) 1670 perf_stat_init_aggr_mode(); 1671 else 1672 perf_stat_init_aggr_mode_file(st); 1673 1674 return 0; 1675 } 1676 1677 static int set_maps(struct perf_stat *st) 1678 { 1679 if (!st->cpus || !st->threads) 1680 return 0; 1681 1682 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 1683 return -EINVAL; 1684 1685 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 1686 1687 if (perf_evlist__alloc_stats(evsel_list, true)) 1688 return -ENOMEM; 1689 1690 st->maps_allocated = true; 1691 return 0; 1692 } 1693 1694 static 1695 int process_thread_map_event(struct perf_session *session, 1696 union perf_event *event) 1697 { 1698 struct perf_tool *tool = session->tool; 1699 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 1700 1701 if (st->threads) { 1702 pr_warning("Extra thread map event, ignoring.\n"); 1703 return 0; 1704 } 1705 1706 st->threads = thread_map__new_event(&event->thread_map); 1707 if (!st->threads) 1708 return -ENOMEM; 1709 1710 return set_maps(st); 1711 } 1712 1713 static 1714 int process_cpu_map_event(struct perf_session *session, 1715 union perf_event *event) 1716 { 1717 struct perf_tool *tool = session->tool; 1718 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 1719 struct perf_cpu_map *cpus; 1720 1721 if (st->cpus) { 1722 pr_warning("Extra cpu map event, ignoring.\n"); 1723 return 0; 1724 } 1725 1726 cpus = cpu_map__new_data(&event->cpu_map.data); 1727 if (!cpus) 1728 return -ENOMEM; 1729 1730 st->cpus = cpus; 1731 return set_maps(st); 1732 } 1733 1734 static int runtime_stat_new(struct perf_stat_config *config, int nthreads) 1735 { 1736 int i; 1737 1738 config->stats = calloc(nthreads, sizeof(struct runtime_stat)); 1739 if (!config->stats) 1740 return -1; 1741 1742 config->stats_num = nthreads; 1743 1744 for (i = 0; i < nthreads; i++) 1745 runtime_stat__init(&config->stats[i]); 1746 1747 return 0; 1748 } 1749 1750 static void runtime_stat_delete(struct perf_stat_config *config) 1751 { 1752 int i; 1753 1754 if (!config->stats) 1755 return; 1756 1757 for (i = 0; i < config->stats_num; i++) 1758 runtime_stat__exit(&config->stats[i]); 1759 1760 zfree(&config->stats); 1761 } 1762 1763 static const char * const stat_report_usage[] = { 1764 "perf stat report [<options>]", 1765 NULL, 1766 }; 1767 1768 static struct perf_stat perf_stat = { 1769 .tool = { 1770 .attr = perf_event__process_attr, 1771 .event_update = perf_event__process_event_update, 1772 .thread_map = process_thread_map_event, 1773 .cpu_map = process_cpu_map_event, 1774 .stat_config = process_stat_config_event, 1775 .stat = perf_event__process_stat_event, 1776 .stat_round = process_stat_round_event, 1777 }, 1778 .aggr_mode = AGGR_UNSET, 1779 }; 1780 1781 static int __cmd_report(int argc, const char **argv) 1782 { 1783 struct perf_session *session; 1784 const struct option options[] = { 1785 OPT_STRING('i', "input", &input_name, "file", "input file name"), 1786 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 1787 "aggregate counts per processor socket", AGGR_SOCKET), 1788 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 1789 "aggregate counts per processor die", AGGR_DIE), 1790 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 1791 "aggregate counts per physical processor core", AGGR_CORE), 1792 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, 1793 "aggregate counts per numa node", AGGR_NODE), 1794 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 1795 "disable CPU count aggregation", AGGR_NONE), 1796 OPT_END() 1797 }; 1798 struct stat st; 1799 int ret; 1800 1801 argc = parse_options(argc, argv, options, stat_report_usage, 0); 1802 1803 if (!input_name || !strlen(input_name)) { 1804 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 1805 input_name = "-"; 1806 else 1807 input_name = "perf.data"; 1808 } 1809 1810 perf_stat.data.path = input_name; 1811 perf_stat.data.mode = PERF_DATA_MODE_READ; 1812 1813 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool); 1814 if (IS_ERR(session)) 1815 return PTR_ERR(session); 1816 1817 perf_stat.session = session; 1818 stat_config.output = stderr; 1819 evsel_list = session->evlist; 1820 1821 ret = perf_session__process_events(session); 1822 if (ret) 1823 return ret; 1824 1825 perf_session__delete(session); 1826 return 0; 1827 } 1828 1829 static void setup_system_wide(int forks) 1830 { 1831 /* 1832 * Make system wide (-a) the default target if 1833 * no target was specified and one of following 1834 * conditions is met: 1835 * 1836 * - there's no workload specified 1837 * - there is workload specified but all requested 1838 * events are system wide events 1839 */ 1840 if (!target__none(&target)) 1841 return; 1842 1843 if (!forks) 1844 target.system_wide = true; 1845 else { 1846 struct evsel *counter; 1847 1848 evlist__for_each_entry(evsel_list, counter) { 1849 if (!counter->core.system_wide) 1850 return; 1851 } 1852 1853 if (evsel_list->core.nr_entries) 1854 target.system_wide = true; 1855 } 1856 } 1857 1858 int cmd_stat(int argc, const char **argv) 1859 { 1860 const char * const stat_usage[] = { 1861 "perf stat [<options>] [<command>]", 1862 NULL 1863 }; 1864 int status = -EINVAL, run_idx; 1865 const char *mode; 1866 FILE *output = stderr; 1867 unsigned int interval, timeout; 1868 const char * const stat_subcommands[] = { "record", "report" }; 1869 1870 setlocale(LC_ALL, ""); 1871 1872 evsel_list = evlist__new(); 1873 if (evsel_list == NULL) 1874 return -ENOMEM; 1875 1876 parse_events__shrink_config_terms(); 1877 1878 /* String-parsing callback-based options would segfault when negated */ 1879 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 1880 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 1881 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 1882 1883 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 1884 (const char **) stat_usage, 1885 PARSE_OPT_STOP_AT_NON_OPTION); 1886 perf_stat__collect_metric_expr(evsel_list); 1887 perf_stat__init_shadow_stats(); 1888 1889 if (stat_config.csv_sep) { 1890 stat_config.csv_output = true; 1891 if (!strcmp(stat_config.csv_sep, "\\t")) 1892 stat_config.csv_sep = "\t"; 1893 } else 1894 stat_config.csv_sep = DEFAULT_SEPARATOR; 1895 1896 if (argc && !strncmp(argv[0], "rec", 3)) { 1897 argc = __cmd_record(argc, argv); 1898 if (argc < 0) 1899 return -1; 1900 } else if (argc && !strncmp(argv[0], "rep", 3)) 1901 return __cmd_report(argc, argv); 1902 1903 interval = stat_config.interval; 1904 timeout = stat_config.timeout; 1905 1906 /* 1907 * For record command the -o is already taken care of. 1908 */ 1909 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 1910 output = NULL; 1911 1912 if (output_name && output_fd) { 1913 fprintf(stderr, "cannot use both --output and --log-fd\n"); 1914 parse_options_usage(stat_usage, stat_options, "o", 1); 1915 parse_options_usage(NULL, stat_options, "log-fd", 0); 1916 goto out; 1917 } 1918 1919 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 1920 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 1921 goto out; 1922 } 1923 1924 if (stat_config.metric_only && stat_config.run_count > 1) { 1925 fprintf(stderr, "--metric-only is not supported with -r\n"); 1926 goto out; 1927 } 1928 1929 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 1930 fprintf(stderr, "--table is only supported with -r\n"); 1931 parse_options_usage(stat_usage, stat_options, "r", 1); 1932 parse_options_usage(NULL, stat_options, "table", 0); 1933 goto out; 1934 } 1935 1936 if (output_fd < 0) { 1937 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 1938 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 1939 goto out; 1940 } 1941 1942 if (!output) { 1943 struct timespec tm; 1944 mode = append_file ? "a" : "w"; 1945 1946 output = fopen(output_name, mode); 1947 if (!output) { 1948 perror("failed to create output file"); 1949 return -1; 1950 } 1951 clock_gettime(CLOCK_REALTIME, &tm); 1952 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 1953 } else if (output_fd > 0) { 1954 mode = append_file ? "a" : "w"; 1955 output = fdopen(output_fd, mode); 1956 if (!output) { 1957 perror("Failed opening logfd"); 1958 return -errno; 1959 } 1960 } 1961 1962 stat_config.output = output; 1963 1964 /* 1965 * let the spreadsheet do the pretty-printing 1966 */ 1967 if (stat_config.csv_output) { 1968 /* User explicitly passed -B? */ 1969 if (big_num_opt == 1) { 1970 fprintf(stderr, "-B option not supported with -x\n"); 1971 parse_options_usage(stat_usage, stat_options, "B", 1); 1972 parse_options_usage(NULL, stat_options, "x", 1); 1973 goto out; 1974 } else /* Nope, so disable big number formatting */ 1975 stat_config.big_num = false; 1976 } else if (big_num_opt == 0) /* User passed --no-big-num */ 1977 stat_config.big_num = false; 1978 1979 setup_system_wide(argc); 1980 1981 /* 1982 * Display user/system times only for single 1983 * run and when there's specified tracee. 1984 */ 1985 if ((stat_config.run_count == 1) && target__none(&target)) 1986 stat_config.ru_display = true; 1987 1988 if (stat_config.run_count < 0) { 1989 pr_err("Run count must be a positive number\n"); 1990 parse_options_usage(stat_usage, stat_options, "r", 1); 1991 goto out; 1992 } else if (stat_config.run_count == 0) { 1993 forever = true; 1994 stat_config.run_count = 1; 1995 } 1996 1997 if (stat_config.walltime_run_table) { 1998 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 1999 if (!stat_config.walltime_run) { 2000 pr_err("failed to setup -r option"); 2001 goto out; 2002 } 2003 } 2004 2005 if ((stat_config.aggr_mode == AGGR_THREAD) && 2006 !target__has_task(&target)) { 2007 if (!target.system_wide || target.cpu_list) { 2008 fprintf(stderr, "The --per-thread option is only " 2009 "available when monitoring via -p -t -a " 2010 "options or only --per-thread.\n"); 2011 parse_options_usage(NULL, stat_options, "p", 1); 2012 parse_options_usage(NULL, stat_options, "t", 1); 2013 goto out; 2014 } 2015 } 2016 2017 /* 2018 * no_aggr, cgroup are for system-wide only 2019 * --per-thread is aggregated per thread, we dont mix it with cpu mode 2020 */ 2021 if (((stat_config.aggr_mode != AGGR_GLOBAL && 2022 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && 2023 !target__has_cpu(&target)) { 2024 fprintf(stderr, "both cgroup and no-aggregation " 2025 "modes only available in system-wide mode\n"); 2026 2027 parse_options_usage(stat_usage, stat_options, "G", 1); 2028 parse_options_usage(NULL, stat_options, "A", 1); 2029 parse_options_usage(NULL, stat_options, "a", 1); 2030 goto out; 2031 } 2032 2033 if (add_default_attributes()) 2034 goto out; 2035 2036 target__validate(&target); 2037 2038 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 2039 target.per_thread = true; 2040 2041 if (perf_evlist__create_maps(evsel_list, &target) < 0) { 2042 if (target__has_task(&target)) { 2043 pr_err("Problems finding threads of monitor\n"); 2044 parse_options_usage(stat_usage, stat_options, "p", 1); 2045 parse_options_usage(NULL, stat_options, "t", 1); 2046 } else if (target__has_cpu(&target)) { 2047 perror("failed to parse CPUs map"); 2048 parse_options_usage(stat_usage, stat_options, "C", 1); 2049 parse_options_usage(NULL, stat_options, "a", 1); 2050 } 2051 goto out; 2052 } 2053 2054 /* 2055 * Initialize thread_map with comm names, 2056 * so we could print it out on output. 2057 */ 2058 if (stat_config.aggr_mode == AGGR_THREAD) { 2059 thread_map__read_comms(evsel_list->core.threads); 2060 if (target.system_wide) { 2061 if (runtime_stat_new(&stat_config, 2062 perf_thread_map__nr(evsel_list->core.threads))) { 2063 goto out; 2064 } 2065 } 2066 } 2067 2068 if (stat_config.aggr_mode == AGGR_NODE) 2069 cpu__setup_cpunode_map(); 2070 2071 if (stat_config.times && interval) 2072 interval_count = true; 2073 else if (stat_config.times && !interval) { 2074 pr_err("interval-count option should be used together with " 2075 "interval-print.\n"); 2076 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 2077 parse_options_usage(stat_usage, stat_options, "I", 1); 2078 goto out; 2079 } 2080 2081 if (timeout && timeout < 100) { 2082 if (timeout < 10) { 2083 pr_err("timeout must be >= 10ms.\n"); 2084 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2085 goto out; 2086 } else 2087 pr_warning("timeout < 100ms. " 2088 "The overhead percentage could be high in some cases. " 2089 "Please proceed with caution.\n"); 2090 } 2091 if (timeout && interval) { 2092 pr_err("timeout option is not supported with interval-print.\n"); 2093 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2094 parse_options_usage(stat_usage, stat_options, "I", 1); 2095 goto out; 2096 } 2097 2098 if (perf_evlist__alloc_stats(evsel_list, interval)) 2099 goto out; 2100 2101 if (perf_stat_init_aggr_mode()) 2102 goto out; 2103 2104 /* 2105 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 2106 * while avoiding that older tools show confusing messages. 2107 * 2108 * However for pipe sessions we need to keep it zero, 2109 * because script's perf_evsel__check_attr is triggered 2110 * by attr->sample_type != 0, and we can't run it on 2111 * stat sessions. 2112 */ 2113 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 2114 2115 /* 2116 * We dont want to block the signals - that would cause 2117 * child tasks to inherit that and Ctrl-C would not work. 2118 * What we want is for Ctrl-C to work in the exec()-ed 2119 * task, but being ignored by perf stat itself: 2120 */ 2121 atexit(sig_atexit); 2122 if (!forever) 2123 signal(SIGINT, skip_signal); 2124 signal(SIGCHLD, skip_signal); 2125 signal(SIGALRM, skip_signal); 2126 signal(SIGABRT, skip_signal); 2127 2128 status = 0; 2129 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 2130 if (stat_config.run_count != 1 && verbose > 0) 2131 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 2132 run_idx + 1); 2133 2134 if (run_idx != 0) 2135 perf_evlist__reset_prev_raw_counts(evsel_list); 2136 2137 status = run_perf_stat(argc, argv, run_idx); 2138 if (forever && status != -1 && !interval) { 2139 print_counters(NULL, argc, argv); 2140 perf_stat__reset_stats(); 2141 } 2142 } 2143 2144 if (!forever && status != -1 && !interval) 2145 print_counters(NULL, argc, argv); 2146 2147 if (STAT_RECORD) { 2148 /* 2149 * We synthesize the kernel mmap record just so that older tools 2150 * don't emit warnings about not being able to resolve symbols 2151 * due to /proc/sys/kernel/kptr_restrict settings and instear provide 2152 * a saner message about no samples being in the perf.data file. 2153 * 2154 * This also serves to suppress a warning about f_header.data.size == 0 2155 * in header.c at the moment 'perf stat record' gets introduced, which 2156 * is not really needed once we start adding the stat specific PERF_RECORD_ 2157 * records, but the need to suppress the kptr_restrict messages in older 2158 * tools remain -acme 2159 */ 2160 int fd = perf_data__fd(&perf_stat.data); 2161 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 2162 process_synthesized_event, 2163 &perf_stat.session->machines.host); 2164 if (err) { 2165 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 2166 "older tools may produce warnings about this file\n."); 2167 } 2168 2169 if (!interval) { 2170 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) 2171 pr_err("failed to write stat round event\n"); 2172 } 2173 2174 if (!perf_stat.data.is_pipe) { 2175 perf_stat.session->header.data_size += perf_stat.bytes_written; 2176 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 2177 } 2178 2179 evlist__close(evsel_list); 2180 perf_session__delete(perf_stat.session); 2181 } 2182 2183 perf_stat__exit_aggr_mode(); 2184 perf_evlist__free_stats(evsel_list); 2185 out: 2186 zfree(&stat_config.walltime_run); 2187 2188 if (smi_cost && smi_reset) 2189 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 2190 2191 evlist__delete(evsel_list); 2192 2193 runtime_stat_delete(&stat_config); 2194 2195 return status; 2196 } 2197