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 "util/cgroup.h" 45 #include <subcmd/parse-options.h> 46 #include "util/parse-events.h" 47 #include "util/pmus.h" 48 #include "util/pmu.h" 49 #include "util/tool_pmu.h" 50 #include "util/event.h" 51 #include "util/evlist.h" 52 #include "util/evsel.h" 53 #include "util/debug.h" 54 #include "util/color.h" 55 #include "util/stat.h" 56 #include "util/header.h" 57 #include "util/cpumap.h" 58 #include "util/thread_map.h" 59 #include "util/counts.h" 60 #include "util/topdown.h" 61 #include "util/session.h" 62 #include "util/tool.h" 63 #include "util/string2.h" 64 #include "util/metricgroup.h" 65 #include "util/synthetic-events.h" 66 #include "util/target.h" 67 #include "util/time-utils.h" 68 #include "util/top.h" 69 #include "util/affinity.h" 70 #include "util/pfm.h" 71 #include "util/bpf_counter.h" 72 #include "util/iostat.h" 73 #include "util/util.h" 74 #include "util/intel-tpebs.h" 75 #include "asm/bug.h" 76 77 #include <linux/time64.h> 78 #include <linux/zalloc.h> 79 #include <api/fs/fs.h> 80 #include <errno.h> 81 #include <signal.h> 82 #include <stdlib.h> 83 #include <sys/prctl.h> 84 #include <inttypes.h> 85 #include <locale.h> 86 #include <math.h> 87 #include <sys/types.h> 88 #include <sys/stat.h> 89 #include <sys/wait.h> 90 #include <unistd.h> 91 #include <sys/time.h> 92 #include <sys/resource.h> 93 #include <linux/err.h> 94 95 #include <linux/ctype.h> 96 #include <perf/evlist.h> 97 #include <internal/threadmap.h> 98 99 #define DEFAULT_SEPARATOR " " 100 #define FREEZE_ON_SMI_PATH "bus/event_source/devices/cpu/freeze_on_smi" 101 102 static void print_counters(struct timespec *ts, int argc, const char **argv); 103 104 static struct evlist *evsel_list; 105 static struct parse_events_option_args parse_events_option_args = { 106 .evlistp = &evsel_list, 107 }; 108 109 static bool all_counters_use_bpf = true; 110 111 static struct target target; 112 113 static volatile sig_atomic_t child_pid = -1; 114 static int detailed_run = 0; 115 static bool transaction_run; 116 static bool topdown_run = false; 117 static bool smi_cost = false; 118 static bool smi_reset = false; 119 static int big_num_opt = -1; 120 static const char *pre_cmd = NULL; 121 static const char *post_cmd = NULL; 122 static bool sync_run = false; 123 static bool forever = false; 124 static bool force_metric_only = false; 125 static struct timespec ref_time; 126 static bool append_file; 127 static bool interval_count; 128 static const char *output_name; 129 static int output_fd; 130 static char *metrics; 131 132 struct perf_stat { 133 bool record; 134 struct perf_data data; 135 struct perf_session *session; 136 u64 bytes_written; 137 struct perf_tool tool; 138 bool maps_allocated; 139 struct perf_cpu_map *cpus; 140 struct perf_thread_map *threads; 141 enum aggr_mode aggr_mode; 142 u32 aggr_level; 143 }; 144 145 static struct perf_stat perf_stat; 146 #define STAT_RECORD perf_stat.record 147 148 static volatile sig_atomic_t done = 0; 149 150 /* Options set from the command line. */ 151 struct opt_aggr_mode { 152 bool node, socket, die, cluster, cache, core, thread, no_aggr; 153 }; 154 155 /* Turn command line option into most generic aggregation mode setting. */ 156 static enum aggr_mode opt_aggr_mode_to_aggr_mode(struct opt_aggr_mode *opt_mode) 157 { 158 enum aggr_mode mode = AGGR_GLOBAL; 159 160 if (opt_mode->node) 161 mode = AGGR_NODE; 162 if (opt_mode->socket) 163 mode = AGGR_SOCKET; 164 if (opt_mode->die) 165 mode = AGGR_DIE; 166 if (opt_mode->cluster) 167 mode = AGGR_CLUSTER; 168 if (opt_mode->cache) 169 mode = AGGR_CACHE; 170 if (opt_mode->core) 171 mode = AGGR_CORE; 172 if (opt_mode->thread) 173 mode = AGGR_THREAD; 174 if (opt_mode->no_aggr) 175 mode = AGGR_NONE; 176 return mode; 177 } 178 179 static void evlist__check_cpu_maps(struct evlist *evlist) 180 { 181 struct evsel *evsel, *warned_leader = NULL; 182 183 evlist__for_each_entry(evlist, evsel) { 184 struct evsel *leader = evsel__leader(evsel); 185 186 /* Check that leader matches cpus with each member. */ 187 if (leader == evsel) 188 continue; 189 if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus)) 190 continue; 191 192 /* If there's mismatch disable the group and warn user. */ 193 if (warned_leader != leader) { 194 char buf[200]; 195 196 pr_warning("WARNING: grouped events cpus do not match.\n" 197 "Events with CPUs not matching the leader will " 198 "be removed from the group.\n"); 199 evsel__group_desc(leader, buf, sizeof(buf)); 200 pr_warning(" %s\n", buf); 201 warned_leader = leader; 202 } 203 if (verbose > 0) { 204 char buf[200]; 205 206 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf)); 207 pr_warning(" %s: %s\n", leader->name, buf); 208 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf)); 209 pr_warning(" %s: %s\n", evsel->name, buf); 210 } 211 212 evsel__remove_from_group(evsel, leader); 213 } 214 } 215 216 static inline void diff_timespec(struct timespec *r, struct timespec *a, 217 struct timespec *b) 218 { 219 r->tv_sec = a->tv_sec - b->tv_sec; 220 if (a->tv_nsec < b->tv_nsec) { 221 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 222 r->tv_sec--; 223 } else { 224 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 225 } 226 } 227 228 static void perf_stat__reset_stats(void) 229 { 230 evlist__reset_stats(evsel_list); 231 perf_stat__reset_shadow_stats(); 232 } 233 234 static int process_synthesized_event(const struct perf_tool *tool __maybe_unused, 235 union perf_event *event, 236 struct perf_sample *sample __maybe_unused, 237 struct machine *machine __maybe_unused) 238 { 239 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 240 pr_err("failed to write perf data, error: %m\n"); 241 return -1; 242 } 243 244 perf_stat.bytes_written += event->header.size; 245 return 0; 246 } 247 248 static int write_stat_round_event(u64 tm, u64 type) 249 { 250 return perf_event__synthesize_stat_round(NULL, tm, type, 251 process_synthesized_event, 252 NULL); 253 } 254 255 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 256 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 257 258 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 259 260 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread, 261 struct perf_counts_values *count) 262 { 263 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread); 264 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx); 265 266 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 267 process_synthesized_event, NULL); 268 } 269 270 static int read_single_counter(struct evsel *counter, int cpu_map_idx, int thread) 271 { 272 int err = evsel__read_counter(counter, cpu_map_idx, thread); 273 274 /* 275 * Reading user and system time will fail when the process 276 * terminates. Use the wait4 values in that case. 277 */ 278 if (err && cpu_map_idx == 0 && 279 (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME || 280 evsel__tool_event(counter) == TOOL_PMU__EVENT_SYSTEM_TIME)) { 281 u64 val, *start_time; 282 struct perf_counts_values *count = 283 perf_counts(counter->counts, cpu_map_idx, thread); 284 285 start_time = xyarray__entry(counter->start_times, cpu_map_idx, thread); 286 if (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME) 287 val = ru_stats.ru_utime_usec_stat.mean; 288 else 289 val = ru_stats.ru_stime_usec_stat.mean; 290 count->ena = count->run = *start_time + val; 291 count->val = val; 292 return 0; 293 } 294 return err; 295 } 296 297 /* 298 * Read out the results of a single counter: 299 * do not aggregate counts across CPUs in system-wide mode 300 */ 301 static int read_counter_cpu(struct evsel *counter, int cpu_map_idx) 302 { 303 int nthreads = perf_thread_map__nr(evsel_list->core.threads); 304 int thread; 305 306 if (!counter->supported) 307 return -ENOENT; 308 309 for (thread = 0; thread < nthreads; thread++) { 310 struct perf_counts_values *count; 311 312 count = perf_counts(counter->counts, cpu_map_idx, thread); 313 314 /* 315 * The leader's group read loads data into its group members 316 * (via evsel__read_counter()) and sets their count->loaded. 317 */ 318 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) && 319 read_single_counter(counter, cpu_map_idx, thread)) { 320 counter->counts->scaled = -1; 321 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0; 322 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0; 323 return -1; 324 } 325 326 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false); 327 328 if (STAT_RECORD) { 329 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) { 330 pr_err("failed to write stat event\n"); 331 return -1; 332 } 333 } 334 335 if (verbose > 1) { 336 fprintf(stat_config.output, 337 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 338 evsel__name(counter), 339 perf_cpu_map__cpu(evsel__cpus(counter), 340 cpu_map_idx).cpu, 341 count->val, count->ena, count->run); 342 } 343 } 344 345 return 0; 346 } 347 348 static int read_affinity_counters(void) 349 { 350 struct evlist_cpu_iterator evlist_cpu_itr; 351 struct affinity saved_affinity, *affinity; 352 353 if (all_counters_use_bpf) 354 return 0; 355 356 if (!target__has_cpu(&target) || target__has_per_thread(&target)) 357 affinity = NULL; 358 else if (affinity__setup(&saved_affinity) < 0) 359 return -1; 360 else 361 affinity = &saved_affinity; 362 363 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 364 struct evsel *counter = evlist_cpu_itr.evsel; 365 366 if (evsel__is_bpf(counter)) 367 continue; 368 369 if (!counter->err) 370 counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx); 371 } 372 if (affinity) 373 affinity__cleanup(&saved_affinity); 374 375 return 0; 376 } 377 378 static int read_bpf_map_counters(void) 379 { 380 struct evsel *counter; 381 int err; 382 383 evlist__for_each_entry(evsel_list, counter) { 384 if (!evsel__is_bpf(counter)) 385 continue; 386 387 err = bpf_counter__read(counter); 388 if (err) 389 return err; 390 } 391 return 0; 392 } 393 394 static int read_counters(void) 395 { 396 if (!stat_config.stop_read_counter) { 397 if (read_bpf_map_counters() || 398 read_affinity_counters()) 399 return -1; 400 } 401 return 0; 402 } 403 404 static void process_counters(void) 405 { 406 struct evsel *counter; 407 408 evlist__for_each_entry(evsel_list, counter) { 409 if (counter->err) 410 pr_debug("failed to read counter %s\n", counter->name); 411 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) 412 pr_warning("failed to process counter %s\n", counter->name); 413 counter->err = 0; 414 } 415 416 perf_stat_merge_counters(&stat_config, evsel_list); 417 perf_stat_process_percore(&stat_config, evsel_list); 418 } 419 420 static void process_interval(void) 421 { 422 struct timespec ts, rs; 423 424 clock_gettime(CLOCK_MONOTONIC, &ts); 425 diff_timespec(&rs, &ts, &ref_time); 426 427 evlist__reset_aggr_stats(evsel_list); 428 429 if (read_counters() == 0) 430 process_counters(); 431 432 if (STAT_RECORD) { 433 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 434 pr_err("failed to write stat round event\n"); 435 } 436 437 init_stats(&walltime_nsecs_stats); 438 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL); 439 print_counters(&rs, 0, NULL); 440 } 441 442 static bool handle_interval(unsigned int interval, int *times) 443 { 444 if (interval) { 445 process_interval(); 446 if (interval_count && !(--(*times))) 447 return true; 448 } 449 return false; 450 } 451 452 static int enable_counters(void) 453 { 454 struct evsel *evsel; 455 int err; 456 457 evlist__for_each_entry(evsel_list, evsel) { 458 if (!evsel__is_bpf(evsel)) 459 continue; 460 461 err = bpf_counter__enable(evsel); 462 if (err) 463 return err; 464 } 465 466 if (!target__enable_on_exec(&target)) { 467 if (!all_counters_use_bpf) 468 evlist__enable(evsel_list); 469 } 470 return 0; 471 } 472 473 static void disable_counters(void) 474 { 475 struct evsel *counter; 476 477 /* 478 * If we don't have tracee (attaching to task or cpu), counters may 479 * still be running. To get accurate group ratios, we must stop groups 480 * from counting before reading their constituent counters. 481 */ 482 if (!target__none(&target)) { 483 evlist__for_each_entry(evsel_list, counter) 484 bpf_counter__disable(counter); 485 if (!all_counters_use_bpf) 486 evlist__disable(evsel_list); 487 } 488 } 489 490 static volatile sig_atomic_t workload_exec_errno; 491 492 /* 493 * evlist__prepare_workload will send a SIGUSR1 494 * if the fork fails, since we asked by setting its 495 * want_signal to true. 496 */ 497 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 498 void *ucontext __maybe_unused) 499 { 500 workload_exec_errno = info->si_value.sival_int; 501 } 502 503 static bool evsel__should_store_id(struct evsel *counter) 504 { 505 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 506 } 507 508 static bool is_target_alive(struct target *_target, 509 struct perf_thread_map *threads) 510 { 511 struct stat st; 512 int i; 513 514 if (!target__has_task(_target)) 515 return true; 516 517 for (i = 0; i < threads->nr; i++) { 518 char path[PATH_MAX]; 519 520 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 521 threads->map[i].pid); 522 523 if (!stat(path, &st)) 524 return true; 525 } 526 527 return false; 528 } 529 530 static void process_evlist(struct evlist *evlist, unsigned int interval) 531 { 532 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED; 533 534 if (evlist__ctlfd_process(evlist, &cmd) > 0) { 535 switch (cmd) { 536 case EVLIST_CTL_CMD_ENABLE: 537 fallthrough; 538 case EVLIST_CTL_CMD_DISABLE: 539 if (interval) 540 process_interval(); 541 break; 542 case EVLIST_CTL_CMD_SNAPSHOT: 543 case EVLIST_CTL_CMD_ACK: 544 case EVLIST_CTL_CMD_UNSUPPORTED: 545 case EVLIST_CTL_CMD_EVLIST: 546 case EVLIST_CTL_CMD_STOP: 547 case EVLIST_CTL_CMD_PING: 548 default: 549 break; 550 } 551 } 552 } 553 554 static void compute_tts(struct timespec *time_start, struct timespec *time_stop, 555 int *time_to_sleep) 556 { 557 int tts = *time_to_sleep; 558 struct timespec time_diff; 559 560 diff_timespec(&time_diff, time_stop, time_start); 561 562 tts -= time_diff.tv_sec * MSEC_PER_SEC + 563 time_diff.tv_nsec / NSEC_PER_MSEC; 564 565 if (tts < 0) 566 tts = 0; 567 568 *time_to_sleep = tts; 569 } 570 571 static int dispatch_events(bool forks, int timeout, int interval, int *times) 572 { 573 int child_exited = 0, status = 0; 574 int time_to_sleep, sleep_time; 575 struct timespec time_start, time_stop; 576 577 if (interval) 578 sleep_time = interval; 579 else if (timeout) 580 sleep_time = timeout; 581 else 582 sleep_time = 1000; 583 584 time_to_sleep = sleep_time; 585 586 while (!done) { 587 if (forks) 588 child_exited = waitpid(child_pid, &status, WNOHANG); 589 else 590 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0; 591 592 if (child_exited) 593 break; 594 595 clock_gettime(CLOCK_MONOTONIC, &time_start); 596 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */ 597 if (timeout || handle_interval(interval, times)) 598 break; 599 time_to_sleep = sleep_time; 600 } else { /* fd revent */ 601 process_evlist(evsel_list, interval); 602 clock_gettime(CLOCK_MONOTONIC, &time_stop); 603 compute_tts(&time_start, &time_stop, &time_to_sleep); 604 } 605 } 606 607 return status; 608 } 609 610 enum counter_recovery { 611 COUNTER_SKIP, 612 COUNTER_RETRY, 613 }; 614 615 static enum counter_recovery stat_handle_error(struct evsel *counter, int err) 616 { 617 char msg[BUFSIZ]; 618 619 assert(!counter->supported); 620 621 /* 622 * PPC returns ENXIO for HW counters until 2.6.37 623 * (behavior changed with commit b0a873e). 624 */ 625 if (err == EINVAL || err == ENOSYS || err == ENOENT || err == ENXIO) { 626 if (verbose > 0) { 627 ui__warning("%s event is not supported by the kernel.\n", 628 evsel__name(counter)); 629 } 630 return COUNTER_SKIP; 631 } 632 if (evsel__fallback(counter, &target, err, msg, sizeof(msg))) { 633 if (verbose > 0) 634 ui__warning("%s\n", msg); 635 counter->supported = true; 636 return COUNTER_RETRY; 637 } 638 if (target__has_per_thread(&target) && err != EOPNOTSUPP && 639 evsel_list->core.threads && evsel_list->core.threads->err_thread != -1) { 640 /* 641 * For global --per-thread case, skip current 642 * error thread. 643 */ 644 if (!thread_map__remove(evsel_list->core.threads, 645 evsel_list->core.threads->err_thread)) { 646 evsel_list->core.threads->err_thread = -1; 647 counter->supported = true; 648 return COUNTER_RETRY; 649 } 650 } 651 if (verbose > 0) { 652 ui__warning(err == EOPNOTSUPP 653 ? "%s event is not supported by the kernel.\n" 654 : "skipping event %s that kernel failed to open.\n", 655 evsel__name(counter)); 656 } 657 return COUNTER_SKIP; 658 } 659 660 static int create_perf_stat_counter(struct evsel *evsel, 661 struct perf_stat_config *config, 662 int cpu_map_idx) 663 { 664 struct perf_event_attr *attr = &evsel->core.attr; 665 struct evsel *leader = evsel__leader(evsel); 666 667 /* Reset supported flag as creating a stat counter is retried. */ 668 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 669 PERF_FORMAT_TOTAL_TIME_RUNNING; 670 671 /* 672 * The event is part of non trivial group, let's enable 673 * the group read (for leader) and ID retrieval for all 674 * members. 675 */ 676 if (leader->core.nr_members > 1) 677 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP; 678 679 attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list); 680 681 /* 682 * Some events get initialized with sample_(period/type) set, 683 * like tracepoints. Clear it up for counting. 684 */ 685 attr->sample_period = 0; 686 687 if (config->identifier) 688 attr->sample_type = PERF_SAMPLE_IDENTIFIER; 689 690 if (config->all_user) { 691 attr->exclude_kernel = 1; 692 attr->exclude_user = 0; 693 } 694 695 if (config->all_kernel) { 696 attr->exclude_kernel = 0; 697 attr->exclude_user = 1; 698 } 699 700 /* 701 * Disabling all counters initially, they will be enabled 702 * either manually by us or by kernel via enable_on_exec 703 * set later. 704 */ 705 if (evsel__is_group_leader(evsel)) { 706 attr->disabled = 1; 707 708 if (target__enable_on_exec(&target)) 709 attr->enable_on_exec = 1; 710 } 711 712 return evsel__open_per_cpu_and_thread(evsel, evsel__cpus(evsel), cpu_map_idx, 713 evsel->core.threads); 714 } 715 716 static int __run_perf_stat(int argc, const char **argv, int run_idx) 717 { 718 int interval = stat_config.interval; 719 int times = stat_config.times; 720 int timeout = stat_config.timeout; 721 char msg[BUFSIZ]; 722 unsigned long long t0, t1; 723 struct evsel *counter; 724 size_t l; 725 int status = 0; 726 const bool forks = (argc > 0); 727 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; 728 struct evlist_cpu_iterator evlist_cpu_itr; 729 struct affinity saved_affinity, *affinity = NULL; 730 int err, open_err = 0; 731 bool second_pass = false, has_supported_counters; 732 733 if (forks) { 734 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) { 735 perror("failed to prepare workload"); 736 return -1; 737 } 738 child_pid = evsel_list->workload.pid; 739 } 740 741 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) { 742 if (affinity__setup(&saved_affinity) < 0) { 743 err = -1; 744 goto err_out; 745 } 746 affinity = &saved_affinity; 747 } 748 749 evlist__for_each_entry(evsel_list, counter) { 750 counter->reset_group = false; 751 if (bpf_counter__load(counter, &target)) { 752 err = -1; 753 goto err_out; 754 } 755 if (!(evsel__is_bperf(counter))) 756 all_counters_use_bpf = false; 757 } 758 759 evlist__reset_aggr_stats(evsel_list); 760 761 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 762 counter = evlist_cpu_itr.evsel; 763 764 /* 765 * bperf calls evsel__open_per_cpu() in bperf__load(), so 766 * no need to call it again here. 767 */ 768 if (target.use_bpf) 769 break; 770 771 if (counter->reset_group || !counter->supported) 772 continue; 773 if (evsel__is_bperf(counter)) 774 continue; 775 776 while (true) { 777 if (create_perf_stat_counter(counter, &stat_config, 778 evlist_cpu_itr.cpu_map_idx) == 0) 779 break; 780 781 open_err = errno; 782 /* 783 * Weak group failed. We cannot just undo this here 784 * because earlier CPUs might be in group mode, and the kernel 785 * doesn't support mixing group and non group reads. Defer 786 * it to later. 787 * Don't close here because we're in the wrong affinity. 788 */ 789 if ((open_err == EINVAL || open_err == EBADF) && 790 evsel__leader(counter) != counter && 791 counter->weak_group) { 792 evlist__reset_weak_group(evsel_list, counter, false); 793 assert(counter->reset_group); 794 counter->supported = true; 795 second_pass = true; 796 break; 797 } 798 799 if (stat_handle_error(counter, open_err) != COUNTER_RETRY) 800 break; 801 } 802 } 803 804 if (second_pass) { 805 /* 806 * Now redo all the weak group after closing them, 807 * and also close errored counters. 808 */ 809 810 /* First close errored or weak retry */ 811 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 812 counter = evlist_cpu_itr.evsel; 813 814 if (!counter->reset_group && counter->supported) 815 continue; 816 817 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx); 818 } 819 /* Now reopen weak */ 820 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 821 counter = evlist_cpu_itr.evsel; 822 823 if (!counter->reset_group) 824 continue; 825 826 while (true) { 827 pr_debug2("reopening weak %s\n", evsel__name(counter)); 828 if (create_perf_stat_counter(counter, &stat_config, 829 evlist_cpu_itr.cpu_map_idx) == 0) 830 break; 831 832 open_err = errno; 833 if (stat_handle_error(counter, open_err) != COUNTER_RETRY) 834 break; 835 } 836 } 837 } 838 affinity__cleanup(affinity); 839 affinity = NULL; 840 841 has_supported_counters = false; 842 evlist__for_each_entry(evsel_list, counter) { 843 if (!counter->supported) { 844 perf_evsel__free_fd(&counter->core); 845 continue; 846 } 847 has_supported_counters = true; 848 849 l = strlen(counter->unit); 850 if (l > stat_config.unit_width) 851 stat_config.unit_width = l; 852 853 if (evsel__should_store_id(counter) && 854 evsel__store_ids(counter, evsel_list)) { 855 err = -1; 856 goto err_out; 857 } 858 } 859 if (!has_supported_counters) { 860 evsel__open_strerror(evlist__first(evsel_list), &target, open_err, 861 msg, sizeof(msg)); 862 ui__error("No supported events found.\n%s\n", msg); 863 864 if (child_pid != -1) 865 kill(child_pid, SIGTERM); 866 err = -1; 867 goto err_out; 868 } 869 870 if (evlist__apply_filters(evsel_list, &counter, &target)) { 871 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n", 872 counter->filter, evsel__name(counter), errno, 873 str_error_r(errno, msg, sizeof(msg))); 874 return -1; 875 } 876 877 if (STAT_RECORD) { 878 int fd = perf_data__fd(&perf_stat.data); 879 880 if (is_pipe) { 881 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); 882 } else { 883 err = perf_session__write_header(perf_stat.session, evsel_list, 884 fd, false); 885 } 886 887 if (err < 0) 888 goto err_out; 889 890 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list, 891 process_synthesized_event, is_pipe); 892 if (err < 0) 893 goto err_out; 894 895 } 896 897 if (target.initial_delay) { 898 pr_info(EVLIST_DISABLED_MSG); 899 } else { 900 err = enable_counters(); 901 if (err) { 902 err = -1; 903 goto err_out; 904 } 905 } 906 907 /* Exec the command, if any */ 908 if (forks) 909 evlist__start_workload(evsel_list); 910 911 if (target.initial_delay > 0) { 912 usleep(target.initial_delay * USEC_PER_MSEC); 913 err = enable_counters(); 914 if (err) { 915 err = -1; 916 goto err_out; 917 } 918 919 pr_info(EVLIST_ENABLED_MSG); 920 } 921 922 t0 = rdclock(); 923 clock_gettime(CLOCK_MONOTONIC, &ref_time); 924 925 if (forks) { 926 if (interval || timeout || evlist__ctlfd_initialized(evsel_list)) 927 status = dispatch_events(forks, timeout, interval, ×); 928 if (child_pid != -1) { 929 if (timeout) 930 kill(child_pid, SIGTERM); 931 wait4(child_pid, &status, 0, &stat_config.ru_data); 932 } 933 934 if (workload_exec_errno) { 935 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); 936 pr_err("Workload failed: %s\n", emsg); 937 err = -1; 938 goto err_out; 939 } 940 941 if (WIFSIGNALED(status)) 942 psignal(WTERMSIG(status), argv[0]); 943 } else { 944 status = dispatch_events(forks, timeout, interval, ×); 945 } 946 947 disable_counters(); 948 949 t1 = rdclock(); 950 951 if (stat_config.walltime_run_table) 952 stat_config.walltime_run[run_idx] = t1 - t0; 953 954 if (interval && stat_config.summary) { 955 stat_config.interval = 0; 956 stat_config.stop_read_counter = true; 957 init_stats(&walltime_nsecs_stats); 958 update_stats(&walltime_nsecs_stats, t1 - t0); 959 960 evlist__copy_prev_raw_counts(evsel_list); 961 evlist__reset_prev_raw_counts(evsel_list); 962 evlist__reset_aggr_stats(evsel_list); 963 } else { 964 update_stats(&walltime_nsecs_stats, t1 - t0); 965 update_rusage_stats(&ru_stats, &stat_config.ru_data); 966 } 967 968 /* 969 * Closing a group leader splits the group, and as we only disable 970 * group leaders, results in remaining events becoming enabled. To 971 * avoid arbitrary skew, we must read all counters before closing any 972 * group leaders. 973 */ 974 if (read_counters() == 0) 975 process_counters(); 976 977 /* 978 * We need to keep evsel_list alive, because it's processed 979 * later the evsel_list will be closed after. 980 */ 981 if (!STAT_RECORD) 982 evlist__close(evsel_list); 983 984 return WEXITSTATUS(status); 985 986 err_out: 987 if (forks) 988 evlist__cancel_workload(evsel_list); 989 990 affinity__cleanup(affinity); 991 return err; 992 } 993 994 /* 995 * Returns -1 for fatal errors which signifies to not continue 996 * when in repeat mode. 997 * 998 * Returns < -1 error codes when stat record is used. These 999 * result in the stat information being displayed, but writing 1000 * to the file fails and is non fatal. 1001 */ 1002 static int run_perf_stat(int argc, const char **argv, int run_idx) 1003 { 1004 int ret; 1005 1006 if (pre_cmd) { 1007 ret = system(pre_cmd); 1008 if (ret) 1009 return ret; 1010 } 1011 1012 if (sync_run) 1013 sync(); 1014 1015 ret = __run_perf_stat(argc, argv, run_idx); 1016 if (ret) 1017 return ret; 1018 1019 if (post_cmd) { 1020 ret = system(post_cmd); 1021 if (ret) 1022 return ret; 1023 } 1024 1025 return ret; 1026 } 1027 1028 static void print_counters(struct timespec *ts, int argc, const char **argv) 1029 { 1030 /* Do not print anything if we record to the pipe. */ 1031 if (STAT_RECORD && perf_stat.data.is_pipe) 1032 return; 1033 if (quiet) 1034 return; 1035 1036 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv); 1037 } 1038 1039 static volatile sig_atomic_t signr = -1; 1040 1041 static void skip_signal(int signo) 1042 { 1043 if ((child_pid == -1) || stat_config.interval) 1044 done = 1; 1045 1046 signr = signo; 1047 /* 1048 * render child_pid harmless 1049 * won't send SIGTERM to a random 1050 * process in case of race condition 1051 * and fast PID recycling 1052 */ 1053 child_pid = -1; 1054 } 1055 1056 static void sig_atexit(void) 1057 { 1058 sigset_t set, oset; 1059 1060 /* 1061 * avoid race condition with SIGCHLD handler 1062 * in skip_signal() which is modifying child_pid 1063 * goal is to avoid send SIGTERM to a random 1064 * process 1065 */ 1066 sigemptyset(&set); 1067 sigaddset(&set, SIGCHLD); 1068 sigprocmask(SIG_BLOCK, &set, &oset); 1069 1070 if (child_pid != -1) 1071 kill(child_pid, SIGTERM); 1072 1073 sigprocmask(SIG_SETMASK, &oset, NULL); 1074 1075 if (signr == -1) 1076 return; 1077 1078 signal(signr, SIG_DFL); 1079 kill(getpid(), signr); 1080 } 1081 1082 static int stat__set_big_num(const struct option *opt __maybe_unused, 1083 const char *s __maybe_unused, int unset) 1084 { 1085 big_num_opt = unset ? 0 : 1; 1086 perf_stat__set_big_num(!unset); 1087 return 0; 1088 } 1089 1090 static int enable_metric_only(const struct option *opt __maybe_unused, 1091 const char *s __maybe_unused, int unset) 1092 { 1093 force_metric_only = true; 1094 stat_config.metric_only = !unset; 1095 return 0; 1096 } 1097 1098 static int append_metric_groups(const struct option *opt __maybe_unused, 1099 const char *str, 1100 int unset __maybe_unused) 1101 { 1102 if (metrics) { 1103 char *tmp; 1104 1105 if (asprintf(&tmp, "%s,%s", metrics, str) < 0) 1106 return -ENOMEM; 1107 free(metrics); 1108 metrics = tmp; 1109 } else { 1110 metrics = strdup(str); 1111 if (!metrics) 1112 return -ENOMEM; 1113 } 1114 return 0; 1115 } 1116 1117 static int parse_control_option(const struct option *opt, 1118 const char *str, 1119 int unset __maybe_unused) 1120 { 1121 struct perf_stat_config *config = opt->value; 1122 1123 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close); 1124 } 1125 1126 static int parse_stat_cgroups(const struct option *opt, 1127 const char *str, int unset) 1128 { 1129 if (stat_config.cgroup_list) { 1130 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 1131 return -1; 1132 } 1133 1134 return parse_cgroups(opt, str, unset); 1135 } 1136 1137 static int parse_cputype(const struct option *opt, 1138 const char *str, 1139 int unset __maybe_unused) 1140 { 1141 const struct perf_pmu *pmu; 1142 struct evlist *evlist = *(struct evlist **)opt->value; 1143 1144 if (!list_empty(&evlist->core.entries)) { 1145 fprintf(stderr, "Must define cputype before events/metrics\n"); 1146 return -1; 1147 } 1148 1149 pmu = perf_pmus__pmu_for_pmu_filter(str); 1150 if (!pmu) { 1151 fprintf(stderr, "--cputype %s is not supported!\n", str); 1152 return -1; 1153 } 1154 parse_events_option_args.pmu_filter = pmu->name; 1155 1156 return 0; 1157 } 1158 1159 static int parse_cache_level(const struct option *opt, 1160 const char *str, 1161 int unset __maybe_unused) 1162 { 1163 int level; 1164 struct opt_aggr_mode *opt_aggr_mode = (struct opt_aggr_mode *)opt->value; 1165 u32 *aggr_level = (u32 *)opt->data; 1166 1167 /* 1168 * If no string is specified, aggregate based on the topology of 1169 * Last Level Cache (LLC). Since the LLC level can change from 1170 * architecture to architecture, set level greater than 1171 * MAX_CACHE_LVL which will be interpreted as LLC. 1172 */ 1173 if (str == NULL) { 1174 level = MAX_CACHE_LVL + 1; 1175 goto out; 1176 } 1177 1178 /* 1179 * The format to specify cache level is LX or lX where X is the 1180 * cache level. 1181 */ 1182 if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) { 1183 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1184 MAX_CACHE_LVL, 1185 MAX_CACHE_LVL); 1186 return -EINVAL; 1187 } 1188 1189 level = atoi(&str[1]); 1190 if (level < 1) { 1191 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1192 MAX_CACHE_LVL, 1193 MAX_CACHE_LVL); 1194 return -EINVAL; 1195 } 1196 1197 if (level > MAX_CACHE_LVL) { 1198 pr_err("perf only supports max cache level of %d.\n" 1199 "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL); 1200 return -EINVAL; 1201 } 1202 out: 1203 opt_aggr_mode->cache = true; 1204 *aggr_level = level; 1205 return 0; 1206 } 1207 1208 /** 1209 * Calculate the cache instance ID from the map in 1210 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1211 * Cache instance ID is the first CPU reported in the shared_cpu_list file. 1212 */ 1213 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map) 1214 { 1215 int id; 1216 struct perf_cpu_map *cpu_map = perf_cpu_map__new(map); 1217 1218 /* 1219 * If the map contains no CPU, consider the current CPU to 1220 * be the first online CPU in the cache domain else use the 1221 * first online CPU of the cache domain as the ID. 1222 */ 1223 id = perf_cpu_map__min(cpu_map).cpu; 1224 if (id == -1) 1225 id = cpu.cpu; 1226 1227 /* Free the perf_cpu_map used to find the cache ID */ 1228 perf_cpu_map__put(cpu_map); 1229 1230 return id; 1231 } 1232 1233 /** 1234 * cpu__get_cache_id - Returns 0 if successful in populating the 1235 * cache level and cache id. Cache level is read from 1236 * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID 1237 * is the first CPU reported by 1238 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1239 */ 1240 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache) 1241 { 1242 int ret = 0; 1243 u32 cache_level = stat_config.aggr_level; 1244 struct cpu_cache_level caches[MAX_CACHE_LVL]; 1245 u32 i = 0, caches_cnt = 0; 1246 1247 cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1248 cache->cache = -1; 1249 1250 ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt); 1251 if (ret) { 1252 /* 1253 * If caches_cnt is not 0, cpu_cache_level data 1254 * was allocated when building the topology. 1255 * Free the allocated data before returning. 1256 */ 1257 if (caches_cnt) 1258 goto free_caches; 1259 1260 return ret; 1261 } 1262 1263 if (!caches_cnt) 1264 return -1; 1265 1266 /* 1267 * Save the data for the highest level if no 1268 * level was specified by the user. 1269 */ 1270 if (cache_level > MAX_CACHE_LVL) { 1271 int max_level_index = 0; 1272 1273 for (i = 1; i < caches_cnt; ++i) { 1274 if (caches[i].level > caches[max_level_index].level) 1275 max_level_index = i; 1276 } 1277 1278 cache->cache_lvl = caches[max_level_index].level; 1279 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map); 1280 1281 /* Reset i to 0 to free entire caches[] */ 1282 i = 0; 1283 goto free_caches; 1284 } 1285 1286 for (i = 0; i < caches_cnt; ++i) { 1287 if (caches[i].level == cache_level) { 1288 cache->cache_lvl = cache_level; 1289 cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1290 } 1291 1292 cpu_cache_level__free(&caches[i]); 1293 } 1294 1295 free_caches: 1296 /* 1297 * Free all the allocated cpu_cache_level data. 1298 */ 1299 while (i < caches_cnt) 1300 cpu_cache_level__free(&caches[i++]); 1301 1302 return ret; 1303 } 1304 1305 /** 1306 * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache 1307 * level, die and socket populated with the cache instache ID, cache level, 1308 * die and socket for cpu. The function signature is compatible with 1309 * aggr_cpu_id_get_t. 1310 */ 1311 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data) 1312 { 1313 int ret; 1314 struct aggr_cpu_id id; 1315 struct perf_cache cache; 1316 1317 id = aggr_cpu_id__die(cpu, data); 1318 if (aggr_cpu_id__is_empty(&id)) 1319 return id; 1320 1321 ret = cpu__get_cache_details(cpu, &cache); 1322 if (ret) 1323 return id; 1324 1325 id.cache_lvl = cache.cache_lvl; 1326 id.cache = cache.cache; 1327 return id; 1328 } 1329 1330 static const char *const aggr_mode__string[] = { 1331 [AGGR_CORE] = "core", 1332 [AGGR_CACHE] = "cache", 1333 [AGGR_CLUSTER] = "cluster", 1334 [AGGR_DIE] = "die", 1335 [AGGR_GLOBAL] = "global", 1336 [AGGR_NODE] = "node", 1337 [AGGR_NONE] = "none", 1338 [AGGR_SOCKET] = "socket", 1339 [AGGR_THREAD] = "thread", 1340 [AGGR_UNSET] = "unset", 1341 }; 1342 1343 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 1344 struct perf_cpu cpu) 1345 { 1346 return aggr_cpu_id__socket(cpu, /*data=*/NULL); 1347 } 1348 1349 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 1350 struct perf_cpu cpu) 1351 { 1352 return aggr_cpu_id__die(cpu, /*data=*/NULL); 1353 } 1354 1355 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused, 1356 struct perf_cpu cpu) 1357 { 1358 return aggr_cpu_id__cache(cpu, /*data=*/NULL); 1359 } 1360 1361 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused, 1362 struct perf_cpu cpu) 1363 { 1364 return aggr_cpu_id__cluster(cpu, /*data=*/NULL); 1365 } 1366 1367 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 1368 struct perf_cpu cpu) 1369 { 1370 return aggr_cpu_id__core(cpu, /*data=*/NULL); 1371 } 1372 1373 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused, 1374 struct perf_cpu cpu) 1375 { 1376 return aggr_cpu_id__node(cpu, /*data=*/NULL); 1377 } 1378 1379 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused, 1380 struct perf_cpu cpu) 1381 { 1382 return aggr_cpu_id__global(cpu, /*data=*/NULL); 1383 } 1384 1385 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused, 1386 struct perf_cpu cpu) 1387 { 1388 return aggr_cpu_id__cpu(cpu, /*data=*/NULL); 1389 } 1390 1391 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config, 1392 aggr_get_id_t get_id, struct perf_cpu cpu) 1393 { 1394 struct aggr_cpu_id id; 1395 1396 /* per-process mode - should use global aggr mode */ 1397 if (cpu.cpu == -1 || cpu.cpu >= config->cpus_aggr_map->nr) 1398 return get_id(config, cpu); 1399 1400 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu])) 1401 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu); 1402 1403 id = config->cpus_aggr_map->map[cpu.cpu]; 1404 return id; 1405 } 1406 1407 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config, 1408 struct perf_cpu cpu) 1409 { 1410 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu); 1411 } 1412 1413 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config, 1414 struct perf_cpu cpu) 1415 { 1416 return perf_stat__get_aggr(config, perf_stat__get_die, cpu); 1417 } 1418 1419 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config, 1420 struct perf_cpu cpu) 1421 { 1422 return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu); 1423 } 1424 1425 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config, 1426 struct perf_cpu cpu) 1427 { 1428 return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu); 1429 } 1430 1431 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config, 1432 struct perf_cpu cpu) 1433 { 1434 return perf_stat__get_aggr(config, perf_stat__get_core, cpu); 1435 } 1436 1437 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config, 1438 struct perf_cpu cpu) 1439 { 1440 return perf_stat__get_aggr(config, perf_stat__get_node, cpu); 1441 } 1442 1443 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config, 1444 struct perf_cpu cpu) 1445 { 1446 return perf_stat__get_aggr(config, perf_stat__get_global, cpu); 1447 } 1448 1449 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config, 1450 struct perf_cpu cpu) 1451 { 1452 return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu); 1453 } 1454 1455 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode) 1456 { 1457 switch (aggr_mode) { 1458 case AGGR_SOCKET: 1459 return aggr_cpu_id__socket; 1460 case AGGR_DIE: 1461 return aggr_cpu_id__die; 1462 case AGGR_CLUSTER: 1463 return aggr_cpu_id__cluster; 1464 case AGGR_CACHE: 1465 return aggr_cpu_id__cache; 1466 case AGGR_CORE: 1467 return aggr_cpu_id__core; 1468 case AGGR_NODE: 1469 return aggr_cpu_id__node; 1470 case AGGR_NONE: 1471 return aggr_cpu_id__cpu; 1472 case AGGR_GLOBAL: 1473 return aggr_cpu_id__global; 1474 case AGGR_THREAD: 1475 case AGGR_UNSET: 1476 case AGGR_MAX: 1477 default: 1478 return NULL; 1479 } 1480 } 1481 1482 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode) 1483 { 1484 switch (aggr_mode) { 1485 case AGGR_SOCKET: 1486 return perf_stat__get_socket_cached; 1487 case AGGR_DIE: 1488 return perf_stat__get_die_cached; 1489 case AGGR_CLUSTER: 1490 return perf_stat__get_cluster_cached; 1491 case AGGR_CACHE: 1492 return perf_stat__get_cache_id_cached; 1493 case AGGR_CORE: 1494 return perf_stat__get_core_cached; 1495 case AGGR_NODE: 1496 return perf_stat__get_node_cached; 1497 case AGGR_NONE: 1498 return perf_stat__get_cpu_cached; 1499 case AGGR_GLOBAL: 1500 return perf_stat__get_global_cached; 1501 case AGGR_THREAD: 1502 case AGGR_UNSET: 1503 case AGGR_MAX: 1504 default: 1505 return NULL; 1506 } 1507 } 1508 1509 static int perf_stat_init_aggr_mode(void) 1510 { 1511 int nr; 1512 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode); 1513 1514 if (get_id) { 1515 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1516 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1517 get_id, /*data=*/NULL, needs_sort); 1518 if (!stat_config.aggr_map) { 1519 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1520 return -1; 1521 } 1522 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode); 1523 } 1524 1525 if (stat_config.aggr_mode == AGGR_THREAD) { 1526 nr = perf_thread_map__nr(evsel_list->core.threads); 1527 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1528 if (stat_config.aggr_map == NULL) 1529 return -ENOMEM; 1530 1531 for (int s = 0; s < nr; s++) { 1532 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1533 1534 id.thread_idx = s; 1535 stat_config.aggr_map->map[s] = id; 1536 } 1537 return 0; 1538 } 1539 1540 /* 1541 * The evsel_list->cpus is the base we operate on, 1542 * taking the highest cpu number to be the size of 1543 * the aggregation translate cpumap. 1544 */ 1545 nr = perf_cpu_map__max(evsel_list->core.all_cpus).cpu + 1; 1546 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr); 1547 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 1548 } 1549 1550 static void cpu_aggr_map__delete(struct cpu_aggr_map *map) 1551 { 1552 free(map); 1553 } 1554 1555 static void perf_stat__exit_aggr_mode(void) 1556 { 1557 cpu_aggr_map__delete(stat_config.aggr_map); 1558 cpu_aggr_map__delete(stat_config.cpus_aggr_map); 1559 stat_config.aggr_map = NULL; 1560 stat_config.cpus_aggr_map = NULL; 1561 } 1562 1563 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data) 1564 { 1565 struct perf_env *env = data; 1566 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1567 1568 if (cpu.cpu != -1) 1569 id.socket = env->cpu[cpu.cpu].socket_id; 1570 1571 return id; 1572 } 1573 1574 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data) 1575 { 1576 struct perf_env *env = data; 1577 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1578 1579 if (cpu.cpu != -1) { 1580 /* 1581 * die_id is relative to socket, so start 1582 * with the socket ID and then add die to 1583 * make a unique ID. 1584 */ 1585 id.socket = env->cpu[cpu.cpu].socket_id; 1586 id.die = env->cpu[cpu.cpu].die_id; 1587 } 1588 1589 return id; 1590 } 1591 1592 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env, 1593 u32 cache_level, struct aggr_cpu_id *id) 1594 { 1595 int i; 1596 int caches_cnt = env->caches_cnt; 1597 struct cpu_cache_level *caches = env->caches; 1598 1599 id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1600 id->cache = -1; 1601 1602 if (!caches_cnt) 1603 return; 1604 1605 for (i = caches_cnt - 1; i > -1; --i) { 1606 struct perf_cpu_map *cpu_map; 1607 int map_contains_cpu; 1608 1609 /* 1610 * If user has not specified a level, find the fist level with 1611 * the cpu in the map. Since building the map is expensive, do 1612 * this only if levels match. 1613 */ 1614 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level) 1615 continue; 1616 1617 cpu_map = perf_cpu_map__new(caches[i].map); 1618 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu); 1619 perf_cpu_map__put(cpu_map); 1620 1621 if (map_contains_cpu != -1) { 1622 id->cache_lvl = caches[i].level; 1623 id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1624 return; 1625 } 1626 } 1627 } 1628 1629 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu, 1630 void *data) 1631 { 1632 struct perf_env *env = data; 1633 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1634 1635 if (cpu.cpu != -1) { 1636 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level; 1637 1638 id.socket = env->cpu[cpu.cpu].socket_id; 1639 id.die = env->cpu[cpu.cpu].die_id; 1640 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id); 1641 } 1642 1643 return id; 1644 } 1645 1646 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu, 1647 void *data) 1648 { 1649 struct perf_env *env = data; 1650 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1651 1652 if (cpu.cpu != -1) { 1653 id.socket = env->cpu[cpu.cpu].socket_id; 1654 id.die = env->cpu[cpu.cpu].die_id; 1655 id.cluster = env->cpu[cpu.cpu].cluster_id; 1656 } 1657 1658 return id; 1659 } 1660 1661 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data) 1662 { 1663 struct perf_env *env = data; 1664 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1665 1666 if (cpu.cpu != -1) { 1667 /* 1668 * core_id is relative to socket, die and cluster, we need a 1669 * global id. So we set socket, die id, cluster id and core id. 1670 */ 1671 id.socket = env->cpu[cpu.cpu].socket_id; 1672 id.die = env->cpu[cpu.cpu].die_id; 1673 id.cluster = env->cpu[cpu.cpu].cluster_id; 1674 id.core = env->cpu[cpu.cpu].core_id; 1675 } 1676 1677 return id; 1678 } 1679 1680 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data) 1681 { 1682 struct perf_env *env = data; 1683 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1684 1685 if (cpu.cpu != -1) { 1686 /* 1687 * core_id is relative to socket and die, 1688 * we need a global id. So we set 1689 * socket, die id and core id 1690 */ 1691 id.socket = env->cpu[cpu.cpu].socket_id; 1692 id.die = env->cpu[cpu.cpu].die_id; 1693 id.core = env->cpu[cpu.cpu].core_id; 1694 id.cpu = cpu; 1695 } 1696 1697 return id; 1698 } 1699 1700 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data) 1701 { 1702 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1703 1704 id.node = perf_env__numa_node(data, cpu); 1705 return id; 1706 } 1707 1708 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused, 1709 void *data __maybe_unused) 1710 { 1711 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1712 1713 /* it always aggregates to the cpu 0 */ 1714 id.cpu = (struct perf_cpu){ .cpu = 0 }; 1715 return id; 1716 } 1717 1718 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1719 struct perf_cpu cpu) 1720 { 1721 return perf_env__get_socket_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1722 } 1723 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1724 struct perf_cpu cpu) 1725 { 1726 return perf_env__get_die_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1727 } 1728 1729 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused, 1730 struct perf_cpu cpu) 1731 { 1732 return perf_env__get_cluster_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1733 } 1734 1735 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused, 1736 struct perf_cpu cpu) 1737 { 1738 return perf_env__get_cache_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1739 } 1740 1741 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1742 struct perf_cpu cpu) 1743 { 1744 return perf_env__get_core_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1745 } 1746 1747 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused, 1748 struct perf_cpu cpu) 1749 { 1750 return perf_env__get_cpu_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1751 } 1752 1753 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, 1754 struct perf_cpu cpu) 1755 { 1756 return perf_env__get_node_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1757 } 1758 1759 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused, 1760 struct perf_cpu cpu) 1761 { 1762 return perf_env__get_global_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1763 } 1764 1765 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode) 1766 { 1767 switch (aggr_mode) { 1768 case AGGR_SOCKET: 1769 return perf_env__get_socket_aggr_by_cpu; 1770 case AGGR_DIE: 1771 return perf_env__get_die_aggr_by_cpu; 1772 case AGGR_CLUSTER: 1773 return perf_env__get_cluster_aggr_by_cpu; 1774 case AGGR_CACHE: 1775 return perf_env__get_cache_aggr_by_cpu; 1776 case AGGR_CORE: 1777 return perf_env__get_core_aggr_by_cpu; 1778 case AGGR_NODE: 1779 return perf_env__get_node_aggr_by_cpu; 1780 case AGGR_GLOBAL: 1781 return perf_env__get_global_aggr_by_cpu; 1782 case AGGR_NONE: 1783 return perf_env__get_cpu_aggr_by_cpu; 1784 case AGGR_THREAD: 1785 case AGGR_UNSET: 1786 case AGGR_MAX: 1787 default: 1788 return NULL; 1789 } 1790 } 1791 1792 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode) 1793 { 1794 switch (aggr_mode) { 1795 case AGGR_SOCKET: 1796 return perf_stat__get_socket_file; 1797 case AGGR_DIE: 1798 return perf_stat__get_die_file; 1799 case AGGR_CLUSTER: 1800 return perf_stat__get_cluster_file; 1801 case AGGR_CACHE: 1802 return perf_stat__get_cache_file; 1803 case AGGR_CORE: 1804 return perf_stat__get_core_file; 1805 case AGGR_NODE: 1806 return perf_stat__get_node_file; 1807 case AGGR_GLOBAL: 1808 return perf_stat__get_global_file; 1809 case AGGR_NONE: 1810 return perf_stat__get_cpu_file; 1811 case AGGR_THREAD: 1812 case AGGR_UNSET: 1813 case AGGR_MAX: 1814 default: 1815 return NULL; 1816 } 1817 } 1818 1819 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1820 { 1821 struct perf_env *env = perf_session__env(st->session); 1822 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode); 1823 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1824 1825 if (stat_config.aggr_mode == AGGR_THREAD) { 1826 int nr = perf_thread_map__nr(evsel_list->core.threads); 1827 1828 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1829 if (stat_config.aggr_map == NULL) 1830 return -ENOMEM; 1831 1832 for (int s = 0; s < nr; s++) { 1833 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1834 1835 id.thread_idx = s; 1836 stat_config.aggr_map->map[s] = id; 1837 } 1838 return 0; 1839 } 1840 1841 if (!get_id) 1842 return 0; 1843 1844 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1845 get_id, env, needs_sort); 1846 if (!stat_config.aggr_map) { 1847 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1848 return -1; 1849 } 1850 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode); 1851 return 0; 1852 } 1853 1854 /* 1855 * Add default events, if there were no attributes specified or 1856 * if -d/--detailed, -d -d or -d -d -d is used: 1857 */ 1858 static int add_default_events(void) 1859 { 1860 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 1861 struct parse_events_error err; 1862 struct evlist *evlist = evlist__new(); 1863 struct evsel *evsel; 1864 int ret = 0; 1865 1866 if (!evlist) 1867 return -ENOMEM; 1868 1869 parse_events_error__init(&err); 1870 1871 /* Set attrs if no event is selected and !null_run: */ 1872 if (stat_config.null_run) 1873 goto out; 1874 1875 if (transaction_run) { 1876 /* Handle -T as -M transaction. Once platform specific metrics 1877 * support has been added to the json files, all architectures 1878 * will use this approach. To determine transaction support 1879 * on an architecture test for such a metric name. 1880 */ 1881 if (!metricgroup__has_metric_or_groups(pmu, "transaction")) { 1882 pr_err("Missing transaction metrics\n"); 1883 ret = -1; 1884 goto out; 1885 } 1886 ret = metricgroup__parse_groups(evlist, pmu, "transaction", 1887 stat_config.metric_no_group, 1888 stat_config.metric_no_merge, 1889 stat_config.metric_no_threshold, 1890 stat_config.user_requested_cpu_list, 1891 stat_config.system_wide, 1892 stat_config.hardware_aware_grouping); 1893 goto out; 1894 } 1895 1896 if (smi_cost) { 1897 int smi; 1898 1899 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 1900 pr_err("freeze_on_smi is not supported.\n"); 1901 ret = -1; 1902 goto out; 1903 } 1904 1905 if (!smi) { 1906 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 1907 pr_err("Failed to set freeze_on_smi.\n"); 1908 ret = -1; 1909 goto out; 1910 } 1911 smi_reset = true; 1912 } 1913 1914 if (!metricgroup__has_metric_or_groups(pmu, "smi")) { 1915 pr_err("Missing smi metrics\n"); 1916 ret = -1; 1917 goto out; 1918 } 1919 1920 if (!force_metric_only) 1921 stat_config.metric_only = true; 1922 1923 ret = metricgroup__parse_groups(evlist, pmu, "smi", 1924 stat_config.metric_no_group, 1925 stat_config.metric_no_merge, 1926 stat_config.metric_no_threshold, 1927 stat_config.user_requested_cpu_list, 1928 stat_config.system_wide, 1929 stat_config.hardware_aware_grouping); 1930 goto out; 1931 } 1932 1933 if (topdown_run) { 1934 unsigned int max_level = metricgroups__topdown_max_level(); 1935 char str[] = "TopdownL1"; 1936 1937 if (!force_metric_only) 1938 stat_config.metric_only = true; 1939 1940 if (!max_level) { 1941 pr_err("Topdown requested but the topdown metric groups aren't present.\n" 1942 "(See perf list the metric groups have names like TopdownL1)\n"); 1943 ret = -1; 1944 goto out; 1945 } 1946 if (stat_config.topdown_level > max_level) { 1947 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level); 1948 ret = -1; 1949 goto out; 1950 } else if (!stat_config.topdown_level) { 1951 stat_config.topdown_level = 1; 1952 } 1953 if (!stat_config.interval && !stat_config.metric_only) { 1954 fprintf(stat_config.output, 1955 "Topdown accuracy may decrease when measuring long periods.\n" 1956 "Please print the result regularly, e.g. -I1000\n"); 1957 } 1958 str[8] = stat_config.topdown_level + '0'; 1959 if (metricgroup__parse_groups(evlist, 1960 pmu, str, 1961 /*metric_no_group=*/false, 1962 /*metric_no_merge=*/false, 1963 /*metric_no_threshold=*/true, 1964 stat_config.user_requested_cpu_list, 1965 stat_config.system_wide, 1966 stat_config.hardware_aware_grouping) < 0) { 1967 ret = -1; 1968 goto out; 1969 } 1970 } 1971 1972 if (!stat_config.topdown_level) 1973 stat_config.topdown_level = 1; 1974 1975 if (!evlist->core.nr_entries && !evsel_list->core.nr_entries) { 1976 /* No events so add defaults. */ 1977 if (target__has_cpu(&target)) 1978 ret = parse_events(evlist, "cpu-clock", &err); 1979 else 1980 ret = parse_events(evlist, "task-clock", &err); 1981 if (ret) 1982 goto out; 1983 1984 ret = parse_events(evlist, 1985 "context-switches," 1986 "cpu-migrations," 1987 "page-faults," 1988 "instructions," 1989 "cycles," 1990 "stalled-cycles-frontend," 1991 "stalled-cycles-backend," 1992 "branches," 1993 "branch-misses", 1994 &err); 1995 if (ret) 1996 goto out; 1997 1998 /* 1999 * Add TopdownL1 metrics if they exist. To minimize 2000 * multiplexing, don't request threshold computation. 2001 */ 2002 if (metricgroup__has_metric_or_groups(pmu, "Default")) { 2003 struct evlist *metric_evlist = evlist__new(); 2004 2005 if (!metric_evlist) { 2006 ret = -ENOMEM; 2007 goto out; 2008 } 2009 if (metricgroup__parse_groups(metric_evlist, pmu, "Default", 2010 /*metric_no_group=*/false, 2011 /*metric_no_merge=*/false, 2012 /*metric_no_threshold=*/true, 2013 stat_config.user_requested_cpu_list, 2014 stat_config.system_wide, 2015 stat_config.hardware_aware_grouping) < 0) { 2016 ret = -1; 2017 goto out; 2018 } 2019 2020 evlist__for_each_entry(metric_evlist, evsel) 2021 evsel->default_metricgroup = true; 2022 2023 evlist__splice_list_tail(evlist, &metric_evlist->core.entries); 2024 metricgroup__copy_metric_events(evlist, /*cgrp=*/NULL, 2025 &evlist->metric_events, 2026 &metric_evlist->metric_events); 2027 evlist__delete(metric_evlist); 2028 } 2029 } 2030 2031 /* Detailed events get appended to the event list: */ 2032 2033 if (!ret && detailed_run >= 1) { 2034 /* 2035 * Detailed stats (-d), covering the L1 and last level data 2036 * caches: 2037 */ 2038 ret = parse_events(evlist, 2039 "L1-dcache-loads," 2040 "L1-dcache-load-misses," 2041 "LLC-loads," 2042 "LLC-load-misses", 2043 &err); 2044 } 2045 if (!ret && detailed_run >= 2) { 2046 /* 2047 * Very detailed stats (-d -d), covering the instruction cache 2048 * and the TLB caches: 2049 */ 2050 ret = parse_events(evlist, 2051 "L1-icache-loads," 2052 "L1-icache-load-misses," 2053 "dTLB-loads," 2054 "dTLB-load-misses," 2055 "iTLB-loads," 2056 "iTLB-load-misses", 2057 &err); 2058 } 2059 if (!ret && detailed_run >= 3) { 2060 /* 2061 * Very, very detailed stats (-d -d -d), adding prefetch events: 2062 */ 2063 ret = parse_events(evlist, 2064 "L1-dcache-prefetches," 2065 "L1-dcache-prefetch-misses", 2066 &err); 2067 } 2068 out: 2069 if (!ret) { 2070 evlist__for_each_entry(evlist, evsel) { 2071 /* 2072 * Make at least one event non-skippable so fatal errors are visible. 2073 * 'cycles' always used to be default and non-skippable, so use that. 2074 */ 2075 if (strcmp("cycles", evsel__name(evsel))) 2076 evsel->skippable = true; 2077 } 2078 } 2079 parse_events_error__exit(&err); 2080 evlist__splice_list_tail(evsel_list, &evlist->core.entries); 2081 metricgroup__copy_metric_events(evsel_list, /*cgrp=*/NULL, 2082 &evsel_list->metric_events, 2083 &evlist->metric_events); 2084 evlist__delete(evlist); 2085 return ret; 2086 } 2087 2088 static const char * const stat_record_usage[] = { 2089 "perf stat record [<options>]", 2090 NULL, 2091 }; 2092 2093 static void init_features(struct perf_session *session) 2094 { 2095 int feat; 2096 2097 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 2098 perf_header__set_feat(&session->header, feat); 2099 2100 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 2101 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 2102 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 2103 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 2104 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 2105 } 2106 2107 static int __cmd_record(const struct option stat_options[], struct opt_aggr_mode *opt_mode, 2108 int argc, const char **argv) 2109 { 2110 struct perf_session *session; 2111 struct perf_data *data = &perf_stat.data; 2112 2113 argc = parse_options(argc, argv, stat_options, stat_record_usage, 2114 PARSE_OPT_STOP_AT_NON_OPTION); 2115 stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(opt_mode); 2116 2117 if (output_name) 2118 data->path = output_name; 2119 2120 if (stat_config.run_count != 1 || forever) { 2121 pr_err("Cannot use -r option with perf stat record.\n"); 2122 return -1; 2123 } 2124 2125 session = perf_session__new(data, NULL); 2126 if (IS_ERR(session)) { 2127 pr_err("Perf session creation failed\n"); 2128 return PTR_ERR(session); 2129 } 2130 2131 init_features(session); 2132 2133 session->evlist = evsel_list; 2134 perf_stat.session = session; 2135 perf_stat.record = true; 2136 return argc; 2137 } 2138 2139 static int process_stat_round_event(struct perf_session *session, 2140 union perf_event *event) 2141 { 2142 struct perf_record_stat_round *stat_round = &event->stat_round; 2143 struct timespec tsh, *ts = NULL; 2144 struct perf_env *env = perf_session__env(session); 2145 const char **argv = env->cmdline_argv; 2146 int argc = env->nr_cmdline; 2147 2148 process_counters(); 2149 2150 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 2151 update_stats(&walltime_nsecs_stats, stat_round->time); 2152 2153 if (stat_config.interval && stat_round->time) { 2154 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 2155 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 2156 ts = &tsh; 2157 } 2158 2159 print_counters(ts, argc, argv); 2160 return 0; 2161 } 2162 2163 static 2164 int process_stat_config_event(struct perf_session *session, 2165 union perf_event *event) 2166 { 2167 const struct perf_tool *tool = session->tool; 2168 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2169 2170 perf_event__read_stat_config(&stat_config, &event->stat_config); 2171 2172 if (perf_cpu_map__is_empty(st->cpus)) { 2173 if (st->aggr_mode != AGGR_UNSET) 2174 pr_warning("warning: processing task data, aggregation mode not set\n"); 2175 } else if (st->aggr_mode != AGGR_UNSET) { 2176 stat_config.aggr_mode = st->aggr_mode; 2177 } 2178 2179 if (perf_stat.data.is_pipe) 2180 perf_stat_init_aggr_mode(); 2181 else 2182 perf_stat_init_aggr_mode_file(st); 2183 2184 if (stat_config.aggr_map) { 2185 int nr_aggr = stat_config.aggr_map->nr; 2186 2187 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) { 2188 pr_err("cannot allocate aggr counts\n"); 2189 return -1; 2190 } 2191 } 2192 return 0; 2193 } 2194 2195 static int set_maps(struct perf_stat *st) 2196 { 2197 if (!st->cpus || !st->threads) 2198 return 0; 2199 2200 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 2201 return -EINVAL; 2202 2203 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 2204 2205 if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true)) 2206 return -ENOMEM; 2207 2208 st->maps_allocated = true; 2209 return 0; 2210 } 2211 2212 static 2213 int process_thread_map_event(struct perf_session *session, 2214 union perf_event *event) 2215 { 2216 const struct perf_tool *tool = session->tool; 2217 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2218 2219 if (st->threads) { 2220 pr_warning("Extra thread map event, ignoring.\n"); 2221 return 0; 2222 } 2223 2224 st->threads = thread_map__new_event(&event->thread_map); 2225 if (!st->threads) 2226 return -ENOMEM; 2227 2228 return set_maps(st); 2229 } 2230 2231 static 2232 int process_cpu_map_event(struct perf_session *session, 2233 union perf_event *event) 2234 { 2235 const struct perf_tool *tool = session->tool; 2236 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2237 struct perf_cpu_map *cpus; 2238 2239 if (st->cpus) { 2240 pr_warning("Extra cpu map event, ignoring.\n"); 2241 return 0; 2242 } 2243 2244 cpus = cpu_map__new_data(&event->cpu_map.data); 2245 if (!cpus) 2246 return -ENOMEM; 2247 2248 st->cpus = cpus; 2249 return set_maps(st); 2250 } 2251 2252 static const char * const stat_report_usage[] = { 2253 "perf stat report [<options>]", 2254 NULL, 2255 }; 2256 2257 static struct perf_stat perf_stat = { 2258 .aggr_mode = AGGR_UNSET, 2259 .aggr_level = 0, 2260 }; 2261 2262 static int __cmd_report(int argc, const char **argv) 2263 { 2264 struct perf_session *session; 2265 const struct option options[] = { 2266 OPT_STRING('i', "input", &input_name, "file", "input file name"), 2267 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 2268 "aggregate counts per processor socket", AGGR_SOCKET), 2269 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 2270 "aggregate counts per processor die", AGGR_DIE), 2271 OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode, 2272 "aggregate counts perf processor cluster", AGGR_CLUSTER), 2273 OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level, 2274 "cache level", 2275 "aggregate count at this cache level (Default: LLC)", 2276 parse_cache_level), 2277 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 2278 "aggregate counts per physical processor core", AGGR_CORE), 2279 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, 2280 "aggregate counts per numa node", AGGR_NODE), 2281 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 2282 "disable CPU count aggregation", AGGR_NONE), 2283 OPT_END() 2284 }; 2285 struct stat st; 2286 int ret; 2287 2288 argc = parse_options(argc, argv, options, stat_report_usage, 0); 2289 2290 if (!input_name || !strlen(input_name)) { 2291 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 2292 input_name = "-"; 2293 else 2294 input_name = "perf.data"; 2295 } 2296 2297 perf_stat.data.path = input_name; 2298 perf_stat.data.mode = PERF_DATA_MODE_READ; 2299 2300 perf_tool__init(&perf_stat.tool, /*ordered_events=*/false); 2301 perf_stat.tool.attr = perf_event__process_attr; 2302 perf_stat.tool.event_update = perf_event__process_event_update; 2303 perf_stat.tool.thread_map = process_thread_map_event; 2304 perf_stat.tool.cpu_map = process_cpu_map_event; 2305 perf_stat.tool.stat_config = process_stat_config_event; 2306 perf_stat.tool.stat = perf_event__process_stat_event; 2307 perf_stat.tool.stat_round = process_stat_round_event; 2308 2309 session = perf_session__new(&perf_stat.data, &perf_stat.tool); 2310 if (IS_ERR(session)) 2311 return PTR_ERR(session); 2312 2313 perf_stat.session = session; 2314 stat_config.output = stderr; 2315 evlist__delete(evsel_list); 2316 evsel_list = session->evlist; 2317 2318 ret = perf_session__process_events(session); 2319 if (ret) 2320 return ret; 2321 2322 perf_session__delete(session); 2323 return 0; 2324 } 2325 2326 static void setup_system_wide(int forks) 2327 { 2328 /* 2329 * Make system wide (-a) the default target if 2330 * no target was specified and one of following 2331 * conditions is met: 2332 * 2333 * - there's no workload specified 2334 * - there is workload specified but all requested 2335 * events are system wide events 2336 */ 2337 if (!target__none(&target)) 2338 return; 2339 2340 if (!forks) 2341 target.system_wide = true; 2342 else { 2343 struct evsel *counter; 2344 2345 evlist__for_each_entry(evsel_list, counter) { 2346 if (!counter->core.requires_cpu && 2347 !evsel__name_is(counter, "duration_time")) { 2348 return; 2349 } 2350 } 2351 2352 if (evsel_list->core.nr_entries) 2353 target.system_wide = true; 2354 } 2355 } 2356 2357 #ifdef HAVE_ARCH_X86_64_SUPPORT 2358 static int parse_tpebs_mode(const struct option *opt, const char *str, 2359 int unset __maybe_unused) 2360 { 2361 enum tpebs_mode *mode = opt->value; 2362 2363 if (!strcasecmp("mean", str)) { 2364 *mode = TPEBS_MODE__MEAN; 2365 return 0; 2366 } 2367 if (!strcasecmp("min", str)) { 2368 *mode = TPEBS_MODE__MIN; 2369 return 0; 2370 } 2371 if (!strcasecmp("max", str)) { 2372 *mode = TPEBS_MODE__MAX; 2373 return 0; 2374 } 2375 if (!strcasecmp("last", str)) { 2376 *mode = TPEBS_MODE__LAST; 2377 return 0; 2378 } 2379 return -1; 2380 } 2381 #endif // HAVE_ARCH_X86_64_SUPPORT 2382 2383 int cmd_stat(int argc, const char **argv) 2384 { 2385 struct opt_aggr_mode opt_mode = {}; 2386 struct option stat_options[] = { 2387 OPT_BOOLEAN('T', "transaction", &transaction_run, 2388 "hardware transaction statistics"), 2389 OPT_CALLBACK('e', "event", &parse_events_option_args, "event", 2390 "event selector. use 'perf list' to list available events", 2391 parse_events_option), 2392 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 2393 "event filter", parse_filter), 2394 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 2395 "child tasks do not inherit counters"), 2396 OPT_STRING('p', "pid", &target.pid, "pid", 2397 "stat events on existing process id"), 2398 OPT_STRING('t', "tid", &target.tid, "tid", 2399 "stat events on existing thread id"), 2400 #ifdef HAVE_BPF_SKEL 2401 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id", 2402 "stat events on existing bpf program id"), 2403 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf, 2404 "use bpf program to count events"), 2405 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path", 2406 "path to perf_event_attr map"), 2407 #endif 2408 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 2409 "system-wide collection from all CPUs"), 2410 OPT_BOOLEAN(0, "scale", &stat_config.scale, 2411 "Use --no-scale to disable counter scaling for multiplexing"), 2412 OPT_INCR('v', "verbose", &verbose, 2413 "be more verbose (show counter open errors, etc)"), 2414 OPT_INTEGER('r', "repeat", &stat_config.run_count, 2415 "repeat command and print average + stddev (max: 100, forever: 0)"), 2416 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 2417 "display details about each run (only with -r option)"), 2418 OPT_BOOLEAN('n', "null", &stat_config.null_run, 2419 "null run - dont start any counters"), 2420 OPT_INCR('d', "detailed", &detailed_run, 2421 "detailed run - start a lot of events"), 2422 OPT_BOOLEAN('S', "sync", &sync_run, 2423 "call sync() before starting a run"), 2424 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 2425 "print large numbers with thousands\' separators", 2426 stat__set_big_num), 2427 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 2428 "list of cpus to monitor in system-wide"), 2429 OPT_BOOLEAN('A', "no-aggr", &opt_mode.no_aggr, 2430 "disable aggregation across CPUs or PMUs"), 2431 OPT_BOOLEAN(0, "no-merge", &opt_mode.no_aggr, 2432 "disable aggregation the same as -A or -no-aggr"), 2433 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge, 2434 "Merge identical named hybrid events"), 2435 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 2436 "print counts with custom separator"), 2437 OPT_BOOLEAN('j', "json-output", &stat_config.json_output, 2438 "print counts in JSON format"), 2439 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 2440 "monitor event in cgroup name only", parse_stat_cgroups), 2441 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name", 2442 "expand events for each cgroup"), 2443 OPT_STRING('o', "output", &output_name, "file", "output file name"), 2444 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 2445 OPT_INTEGER(0, "log-fd", &output_fd, 2446 "log output to fd, instead of stderr"), 2447 OPT_STRING(0, "pre", &pre_cmd, "command", 2448 "command to run prior to the measured command"), 2449 OPT_STRING(0, "post", &post_cmd, "command", 2450 "command to run after to the measured command"), 2451 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 2452 "print counts at regular interval in ms " 2453 "(overhead is possible for values <= 100ms)"), 2454 OPT_INTEGER(0, "interval-count", &stat_config.times, 2455 "print counts for fixed number of times"), 2456 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 2457 "clear screen in between new interval"), 2458 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 2459 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 2460 OPT_BOOLEAN(0, "per-socket", &opt_mode.socket, 2461 "aggregate counts per processor socket"), 2462 OPT_BOOLEAN(0, "per-die", &opt_mode.die, "aggregate counts per processor die"), 2463 OPT_BOOLEAN(0, "per-cluster", &opt_mode.cluster, 2464 "aggregate counts per processor cluster"), 2465 OPT_CALLBACK_OPTARG(0, "per-cache", &opt_mode, &stat_config.aggr_level, 2466 "cache level", "aggregate count at this cache level (Default: LLC)", 2467 parse_cache_level), 2468 OPT_BOOLEAN(0, "per-core", &opt_mode.core, 2469 "aggregate counts per physical processor core"), 2470 OPT_BOOLEAN(0, "per-thread", &opt_mode.thread, "aggregate counts per thread"), 2471 OPT_BOOLEAN(0, "per-node", &opt_mode.node, "aggregate counts per numa node"), 2472 OPT_INTEGER('D', "delay", &target.initial_delay, 2473 "ms to wait before starting measurement after program start (-1: start with events disabled)"), 2474 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 2475 "Only print computed metrics. No raw values", enable_metric_only), 2476 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group, 2477 "don't group metric events, impacts multiplexing"), 2478 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge, 2479 "don't try to share events between metrics in a group"), 2480 OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold, 2481 "disable adding events for the metric threshold calculation"), 2482 OPT_BOOLEAN(0, "topdown", &topdown_run, 2483 "measure top-down statistics"), 2484 #ifdef HAVE_ARCH_X86_64_SUPPORT 2485 OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording, 2486 "enable recording for tpebs when retire_latency required"), 2487 OPT_CALLBACK(0, "tpebs-mode", &tpebs_mode, "tpebs-mode", 2488 "Mode of TPEBS recording: mean, min or max", 2489 parse_tpebs_mode), 2490 #endif 2491 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level, 2492 "Set the metrics level for the top-down statistics (0: max level)"), 2493 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 2494 "measure SMI cost"), 2495 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 2496 "monitor specified metrics or metric groups (separated by ,)", 2497 append_metric_groups), 2498 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, 2499 "Configure all used events to run in kernel space.", 2500 PARSE_OPT_EXCLUSIVE), 2501 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, 2502 "Configure all used events to run in user space.", 2503 PARSE_OPT_EXCLUSIVE), 2504 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread, 2505 "Use with 'percore' event qualifier to show the event " 2506 "counts of one hardware thread by sum up total hardware " 2507 "threads of same physical core"), 2508 OPT_BOOLEAN(0, "summary", &stat_config.summary, 2509 "print summary for interval mode"), 2510 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary, 2511 "don't print 'summary' for CSV summary output"), 2512 OPT_BOOLEAN(0, "quiet", &quiet, 2513 "don't print any output, messages or warnings (useful with record)"), 2514 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type", 2515 "Only enable events on applying cpu with this type " 2516 "for hybrid platform (e.g. core or atom)", 2517 parse_cputype), 2518 #ifdef HAVE_LIBPFM 2519 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event", 2520 "libpfm4 event selector. use 'perf list' to list available events", 2521 parse_libpfm_events_option), 2522 #endif 2523 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]", 2524 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n" 2525 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n" 2526 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.", 2527 parse_control_option), 2528 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default", 2529 "measure I/O performance metrics provided by arch/platform", 2530 iostat_parse), 2531 OPT_END() 2532 }; 2533 const char * const stat_usage[] = { 2534 "perf stat [<options>] [<command>]", 2535 NULL 2536 }; 2537 int status = -EINVAL, run_idx, err; 2538 const char *mode; 2539 FILE *output = stderr; 2540 unsigned int interval, timeout; 2541 const char * const stat_subcommands[] = { "record", "report" }; 2542 char errbuf[BUFSIZ]; 2543 2544 setlocale(LC_ALL, ""); 2545 2546 evsel_list = evlist__new(); 2547 if (evsel_list == NULL) 2548 return -ENOMEM; 2549 2550 parse_events__shrink_config_terms(); 2551 2552 /* String-parsing callback-based options would segfault when negated */ 2553 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 2554 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 2555 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 2556 2557 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 2558 (const char **) stat_usage, 2559 PARSE_OPT_STOP_AT_NON_OPTION); 2560 2561 stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(&opt_mode); 2562 2563 if (stat_config.csv_sep) { 2564 stat_config.csv_output = true; 2565 if (!strcmp(stat_config.csv_sep, "\\t")) 2566 stat_config.csv_sep = "\t"; 2567 } else 2568 stat_config.csv_sep = DEFAULT_SEPARATOR; 2569 2570 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2571 argc = __cmd_record(stat_options, &opt_mode, argc, argv); 2572 if (argc < 0) 2573 return -1; 2574 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0])) 2575 return __cmd_report(argc, argv); 2576 2577 interval = stat_config.interval; 2578 timeout = stat_config.timeout; 2579 2580 /* 2581 * For record command the -o is already taken care of. 2582 */ 2583 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 2584 output = NULL; 2585 2586 if (output_name && output_fd) { 2587 fprintf(stderr, "cannot use both --output and --log-fd\n"); 2588 parse_options_usage(stat_usage, stat_options, "o", 1); 2589 parse_options_usage(NULL, stat_options, "log-fd", 0); 2590 goto out; 2591 } 2592 2593 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 2594 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 2595 goto out; 2596 } 2597 2598 if (stat_config.metric_only && stat_config.run_count > 1) { 2599 fprintf(stderr, "--metric-only is not supported with -r\n"); 2600 goto out; 2601 } 2602 2603 if (stat_config.csv_output || (stat_config.metric_only && stat_config.json_output)) { 2604 /* 2605 * Current CSV and metric-only JSON output doesn't display the 2606 * metric threshold so don't compute it. 2607 */ 2608 stat_config.metric_no_threshold = true; 2609 } 2610 2611 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 2612 fprintf(stderr, "--table is only supported with -r\n"); 2613 parse_options_usage(stat_usage, stat_options, "r", 1); 2614 parse_options_usage(NULL, stat_options, "table", 0); 2615 goto out; 2616 } 2617 2618 if (output_fd < 0) { 2619 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 2620 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 2621 goto out; 2622 } 2623 2624 if (!output && !quiet) { 2625 struct timespec tm; 2626 mode = append_file ? "a" : "w"; 2627 2628 output = fopen(output_name, mode); 2629 if (!output) { 2630 perror("failed to create output file"); 2631 return -1; 2632 } 2633 if (!stat_config.json_output) { 2634 clock_gettime(CLOCK_REALTIME, &tm); 2635 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 2636 } 2637 } else if (output_fd > 0) { 2638 mode = append_file ? "a" : "w"; 2639 output = fdopen(output_fd, mode); 2640 if (!output) { 2641 perror("Failed opening logfd"); 2642 return -errno; 2643 } 2644 } 2645 2646 if (stat_config.interval_clear && !isatty(fileno(output))) { 2647 fprintf(stderr, "--interval-clear does not work with output\n"); 2648 parse_options_usage(stat_usage, stat_options, "o", 1); 2649 parse_options_usage(NULL, stat_options, "log-fd", 0); 2650 parse_options_usage(NULL, stat_options, "interval-clear", 0); 2651 return -1; 2652 } 2653 2654 stat_config.output = output; 2655 2656 /* 2657 * let the spreadsheet do the pretty-printing 2658 */ 2659 if (stat_config.csv_output) { 2660 /* User explicitly passed -B? */ 2661 if (big_num_opt == 1) { 2662 fprintf(stderr, "-B option not supported with -x\n"); 2663 parse_options_usage(stat_usage, stat_options, "B", 1); 2664 parse_options_usage(NULL, stat_options, "x", 1); 2665 goto out; 2666 } else /* Nope, so disable big number formatting */ 2667 stat_config.big_num = false; 2668 } else if (big_num_opt == 0) /* User passed --no-big-num */ 2669 stat_config.big_num = false; 2670 2671 target.inherit = !stat_config.no_inherit; 2672 err = target__validate(&target); 2673 if (err) { 2674 target__strerror(&target, err, errbuf, BUFSIZ); 2675 pr_warning("%s\n", errbuf); 2676 } 2677 2678 setup_system_wide(argc); 2679 2680 /* 2681 * Display user/system times only for single 2682 * run and when there's specified tracee. 2683 */ 2684 if ((stat_config.run_count == 1) && target__none(&target)) 2685 stat_config.ru_display = true; 2686 2687 if (stat_config.run_count < 0) { 2688 pr_err("Run count must be a positive number\n"); 2689 parse_options_usage(stat_usage, stat_options, "r", 1); 2690 goto out; 2691 } else if (stat_config.run_count == 0) { 2692 forever = true; 2693 stat_config.run_count = 1; 2694 } 2695 2696 if (stat_config.walltime_run_table) { 2697 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 2698 if (!stat_config.walltime_run) { 2699 pr_err("failed to setup -r option"); 2700 goto out; 2701 } 2702 } 2703 2704 if ((stat_config.aggr_mode == AGGR_THREAD) && 2705 !target__has_task(&target)) { 2706 if (!target.system_wide || target.cpu_list) { 2707 fprintf(stderr, "The --per-thread option is only " 2708 "available when monitoring via -p -t -a " 2709 "options or only --per-thread.\n"); 2710 parse_options_usage(NULL, stat_options, "p", 1); 2711 parse_options_usage(NULL, stat_options, "t", 1); 2712 goto out; 2713 } 2714 } 2715 2716 /* 2717 * no_aggr, cgroup are for system-wide only 2718 * --per-thread is aggregated per thread, we dont mix it with cpu mode 2719 */ 2720 if (((stat_config.aggr_mode != AGGR_GLOBAL && 2721 stat_config.aggr_mode != AGGR_THREAD) || 2722 (nr_cgroups || stat_config.cgroup_list)) && 2723 !target__has_cpu(&target)) { 2724 fprintf(stderr, "both cgroup and no-aggregation " 2725 "modes only available in system-wide mode\n"); 2726 2727 parse_options_usage(stat_usage, stat_options, "G", 1); 2728 parse_options_usage(NULL, stat_options, "A", 1); 2729 parse_options_usage(NULL, stat_options, "a", 1); 2730 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2731 goto out; 2732 } 2733 2734 if (stat_config.iostat_run) { 2735 status = iostat_prepare(evsel_list, &stat_config); 2736 if (status) 2737 goto out; 2738 if (iostat_mode == IOSTAT_LIST) { 2739 iostat_list(evsel_list, &stat_config); 2740 goto out; 2741 } else if (verbose > 0) 2742 iostat_list(evsel_list, &stat_config); 2743 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target)) 2744 target.system_wide = true; 2745 } 2746 2747 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 2748 target.per_thread = true; 2749 2750 stat_config.system_wide = target.system_wide; 2751 if (target.cpu_list) { 2752 stat_config.user_requested_cpu_list = strdup(target.cpu_list); 2753 if (!stat_config.user_requested_cpu_list) { 2754 status = -ENOMEM; 2755 goto out; 2756 } 2757 } 2758 2759 /* 2760 * Metric parsing needs to be delayed as metrics may optimize events 2761 * knowing the target is system-wide. 2762 */ 2763 if (metrics) { 2764 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 2765 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics, 2766 stat_config.metric_no_group, 2767 stat_config.metric_no_merge, 2768 stat_config.metric_no_threshold, 2769 stat_config.user_requested_cpu_list, 2770 stat_config.system_wide, 2771 stat_config.hardware_aware_grouping); 2772 2773 zfree(&metrics); 2774 if (ret) { 2775 status = ret; 2776 goto out; 2777 } 2778 } 2779 2780 if (add_default_events()) 2781 goto out; 2782 2783 if (stat_config.cgroup_list) { 2784 if (nr_cgroups > 0) { 2785 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 2786 parse_options_usage(stat_usage, stat_options, "G", 1); 2787 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2788 goto out; 2789 } 2790 2791 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, true) < 0) { 2792 parse_options_usage(stat_usage, stat_options, 2793 "for-each-cgroup", 0); 2794 goto out; 2795 } 2796 } 2797 2798 evlist__warn_user_requested_cpus(evsel_list, target.cpu_list); 2799 2800 if (evlist__create_maps(evsel_list, &target) < 0) { 2801 if (target__has_task(&target)) { 2802 pr_err("Problems finding threads of monitor\n"); 2803 parse_options_usage(stat_usage, stat_options, "p", 1); 2804 parse_options_usage(NULL, stat_options, "t", 1); 2805 } else if (target__has_cpu(&target)) { 2806 perror("failed to parse CPUs map"); 2807 parse_options_usage(stat_usage, stat_options, "C", 1); 2808 parse_options_usage(NULL, stat_options, "a", 1); 2809 } 2810 goto out; 2811 } 2812 2813 evlist__check_cpu_maps(evsel_list); 2814 2815 /* 2816 * Initialize thread_map with comm names, 2817 * so we could print it out on output. 2818 */ 2819 if (stat_config.aggr_mode == AGGR_THREAD) { 2820 thread_map__read_comms(evsel_list->core.threads); 2821 } 2822 2823 if (stat_config.aggr_mode == AGGR_NODE) 2824 cpu__setup_cpunode_map(); 2825 2826 if (stat_config.times && interval) 2827 interval_count = true; 2828 else if (stat_config.times && !interval) { 2829 pr_err("interval-count option should be used together with " 2830 "interval-print.\n"); 2831 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 2832 parse_options_usage(stat_usage, stat_options, "I", 1); 2833 goto out; 2834 } 2835 2836 if (timeout && timeout < 100) { 2837 if (timeout < 10) { 2838 pr_err("timeout must be >= 10ms.\n"); 2839 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2840 goto out; 2841 } else 2842 pr_warning("timeout < 100ms. " 2843 "The overhead percentage could be high in some cases. " 2844 "Please proceed with caution.\n"); 2845 } 2846 if (timeout && interval) { 2847 pr_err("timeout option is not supported with interval-print.\n"); 2848 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2849 parse_options_usage(stat_usage, stat_options, "I", 1); 2850 goto out; 2851 } 2852 2853 if (perf_stat_init_aggr_mode()) 2854 goto out; 2855 2856 if (evlist__alloc_stats(&stat_config, evsel_list, interval)) 2857 goto out; 2858 2859 /* 2860 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 2861 * while avoiding that older tools show confusing messages. 2862 * 2863 * However for pipe sessions we need to keep it zero, 2864 * because script's perf_evsel__check_attr is triggered 2865 * by attr->sample_type != 0, and we can't run it on 2866 * stat sessions. 2867 */ 2868 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 2869 2870 /* 2871 * We dont want to block the signals - that would cause 2872 * child tasks to inherit that and Ctrl-C would not work. 2873 * What we want is for Ctrl-C to work in the exec()-ed 2874 * task, but being ignored by perf stat itself: 2875 */ 2876 atexit(sig_atexit); 2877 if (!forever) 2878 signal(SIGINT, skip_signal); 2879 signal(SIGCHLD, skip_signal); 2880 signal(SIGALRM, skip_signal); 2881 signal(SIGABRT, skip_signal); 2882 2883 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack)) 2884 goto out; 2885 2886 /* Enable ignoring missing threads when -p option is defined. */ 2887 evlist__first(evsel_list)->ignore_missing_thread = target.pid; 2888 status = 0; 2889 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 2890 if (stat_config.run_count != 1 && verbose > 0) 2891 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 2892 run_idx + 1); 2893 2894 if (run_idx != 0) 2895 evlist__reset_prev_raw_counts(evsel_list); 2896 2897 status = run_perf_stat(argc, argv, run_idx); 2898 if (status == -1) 2899 break; 2900 2901 if (forever && !interval) { 2902 print_counters(NULL, argc, argv); 2903 perf_stat__reset_stats(); 2904 } 2905 } 2906 2907 if (!forever && status != -1 && (!interval || stat_config.summary)) { 2908 if (stat_config.run_count > 1) 2909 evlist__copy_res_stats(&stat_config, evsel_list); 2910 print_counters(NULL, argc, argv); 2911 } 2912 2913 evlist__finalize_ctlfd(evsel_list); 2914 2915 if (STAT_RECORD) { 2916 /* 2917 * We synthesize the kernel mmap record just so that older tools 2918 * don't emit warnings about not being able to resolve symbols 2919 * due to /proc/sys/kernel/kptr_restrict settings and instead provide 2920 * a saner message about no samples being in the perf.data file. 2921 * 2922 * This also serves to suppress a warning about f_header.data.size == 0 2923 * in header.c at the moment 'perf stat record' gets introduced, which 2924 * is not really needed once we start adding the stat specific PERF_RECORD_ 2925 * records, but the need to suppress the kptr_restrict messages in older 2926 * tools remain -acme 2927 */ 2928 int fd = perf_data__fd(&perf_stat.data); 2929 2930 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 2931 process_synthesized_event, 2932 &perf_stat.session->machines.host); 2933 if (err) { 2934 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 2935 "older tools may produce warnings about this file\n."); 2936 } 2937 2938 if (!interval) { 2939 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) 2940 pr_err("failed to write stat round event\n"); 2941 } 2942 2943 if (!perf_stat.data.is_pipe) { 2944 perf_stat.session->header.data_size += perf_stat.bytes_written; 2945 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 2946 } 2947 2948 evlist__close(evsel_list); 2949 perf_session__delete(perf_stat.session); 2950 } 2951 2952 perf_stat__exit_aggr_mode(); 2953 evlist__free_stats(evsel_list); 2954 out: 2955 if (stat_config.iostat_run) 2956 iostat_release(evsel_list); 2957 2958 zfree(&stat_config.walltime_run); 2959 zfree(&stat_config.user_requested_cpu_list); 2960 2961 if (smi_cost && smi_reset) 2962 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 2963 2964 evlist__delete(evsel_list); 2965 2966 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close); 2967 2968 return status; 2969 } 2970