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/list_sort.h> 78 #include <linux/time64.h> 79 #include <linux/zalloc.h> 80 #include <api/fs/fs.h> 81 #include <errno.h> 82 #include <signal.h> 83 #include <stdlib.h> 84 #include <sys/prctl.h> 85 #include <inttypes.h> 86 #include <locale.h> 87 #include <math.h> 88 #include <sys/types.h> 89 #include <sys/stat.h> 90 #include <sys/wait.h> 91 #include <unistd.h> 92 #include <sys/time.h> 93 #include <sys/resource.h> 94 #include <linux/err.h> 95 96 #include <linux/ctype.h> 97 #include <perf/evlist.h> 98 #include <internal/threadmap.h> 99 100 #ifdef HAVE_BPF_SKEL 101 #include "util/bpf_skel/bperf_cgroup.h" 102 #endif 103 104 #define DEFAULT_SEPARATOR " " 105 #define FREEZE_ON_SMI_PATH "bus/event_source/devices/cpu/freeze_on_smi" 106 107 struct rusage_stats { 108 struct stats ru_utime_usec_stat; 109 struct stats ru_stime_usec_stat; 110 }; 111 112 static void print_counters(struct timespec *ts, int argc, const char **argv); 113 114 static struct evlist *evsel_list; 115 static struct parse_events_option_args parse_events_option_args = { 116 .evlistp = &evsel_list, 117 }; 118 119 static bool all_counters_use_bpf = true; 120 121 static struct target target; 122 123 static volatile sig_atomic_t child_pid = -1; 124 static int detailed_run = 0; 125 static bool transaction_run; 126 static bool topdown_run = false; 127 static bool smi_cost = false; 128 static bool smi_reset = false; 129 static int big_num_opt = -1; 130 static const char *pre_cmd = NULL; 131 static const char *post_cmd = NULL; 132 static bool sync_run = false; 133 static bool forever = false; 134 static bool force_metric_only = false; 135 static struct timespec ref_time; 136 static bool append_file; 137 static bool interval_count; 138 static const char *output_name; 139 static int output_fd; 140 static char *metrics; 141 static struct rusage_stats ru_stats; 142 143 struct perf_stat { 144 bool record; 145 struct perf_data data; 146 struct perf_session *session; 147 u64 bytes_written; 148 struct perf_tool tool; 149 bool maps_allocated; 150 struct perf_cpu_map *cpus; 151 struct perf_thread_map *threads; 152 enum aggr_mode aggr_mode; 153 u32 aggr_level; 154 }; 155 156 static struct perf_stat perf_stat; 157 #define STAT_RECORD perf_stat.record 158 159 static volatile sig_atomic_t done = 0; 160 161 /* Options set from the command line. */ 162 struct opt_aggr_mode { 163 bool node, socket, die, cluster, cache, core, thread, no_aggr; 164 }; 165 166 /* Turn command line option into most generic aggregation mode setting. */ 167 static enum aggr_mode opt_aggr_mode_to_aggr_mode(struct opt_aggr_mode *opt_mode) 168 { 169 enum aggr_mode mode = AGGR_GLOBAL; 170 171 if (opt_mode->node) 172 mode = AGGR_NODE; 173 if (opt_mode->socket) 174 mode = AGGR_SOCKET; 175 if (opt_mode->die) 176 mode = AGGR_DIE; 177 if (opt_mode->cluster) 178 mode = AGGR_CLUSTER; 179 if (opt_mode->cache) 180 mode = AGGR_CACHE; 181 if (opt_mode->core) 182 mode = AGGR_CORE; 183 if (opt_mode->thread) 184 mode = AGGR_THREAD; 185 if (opt_mode->no_aggr) 186 mode = AGGR_NONE; 187 return mode; 188 } 189 190 static void evlist__check_cpu_maps(struct evlist *evlist) 191 { 192 struct evsel *evsel, *warned_leader = NULL; 193 194 evlist__for_each_entry(evlist, evsel) { 195 struct evsel *leader = evsel__leader(evsel); 196 197 /* Check that leader matches cpus with each member. */ 198 if (leader == evsel) 199 continue; 200 if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus)) 201 continue; 202 203 /* If there's mismatch disable the group and warn user. */ 204 if (warned_leader != leader) { 205 char buf[200]; 206 207 pr_warning("WARNING: grouped events cpus do not match.\n" 208 "Events with CPUs not matching the leader will " 209 "be removed from the group.\n"); 210 evsel__group_desc(leader, buf, sizeof(buf)); 211 pr_warning(" %s\n", buf); 212 warned_leader = leader; 213 } 214 if (verbose > 0) { 215 char buf[200]; 216 217 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf)); 218 pr_warning(" %s: %s\n", leader->name, buf); 219 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf)); 220 pr_warning(" %s: %s\n", evsel->name, buf); 221 } 222 223 evsel__remove_from_group(evsel, leader); 224 } 225 } 226 227 static inline void diff_timespec(struct timespec *r, struct timespec *a, 228 struct timespec *b) 229 { 230 r->tv_sec = a->tv_sec - b->tv_sec; 231 if (a->tv_nsec < b->tv_nsec) { 232 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 233 r->tv_sec--; 234 } else { 235 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 236 } 237 } 238 239 static void perf_stat__reset_stats(void) 240 { 241 evlist__reset_stats(evsel_list); 242 memset(stat_config.walltime_nsecs_stats, 0, sizeof(*stat_config.walltime_nsecs_stats)); 243 } 244 245 static int process_synthesized_event(const struct perf_tool *tool __maybe_unused, 246 union perf_event *event, 247 struct perf_sample *sample __maybe_unused, 248 struct machine *machine __maybe_unused) 249 { 250 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 251 pr_err("failed to write perf data, error: %m\n"); 252 return -1; 253 } 254 255 perf_stat.bytes_written += event->header.size; 256 return 0; 257 } 258 259 static int write_stat_round_event(u64 tm, u64 type) 260 { 261 return perf_event__synthesize_stat_round(NULL, tm, type, 262 process_synthesized_event, 263 NULL); 264 } 265 266 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 267 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 268 269 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 270 271 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread, 272 struct perf_counts_values *count) 273 { 274 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread); 275 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx); 276 277 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 278 process_synthesized_event, NULL); 279 } 280 281 static int read_single_counter(struct evsel *counter, int cpu_map_idx, int thread) 282 { 283 int err = evsel__read_counter(counter, cpu_map_idx, thread); 284 285 /* 286 * Reading user and system time will fail when the process 287 * terminates. Use the wait4 values in that case. 288 */ 289 if (err && cpu_map_idx == 0 && 290 (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME || 291 evsel__tool_event(counter) == TOOL_PMU__EVENT_SYSTEM_TIME)) { 292 struct perf_counts_values *count = 293 perf_counts(counter->counts, cpu_map_idx, thread); 294 struct perf_counts_values *old_count = NULL; 295 u64 val; 296 297 if (counter->prev_raw_counts) 298 old_count = perf_counts(counter->prev_raw_counts, cpu_map_idx, thread); 299 300 if (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME) 301 val = ru_stats.ru_utime_usec_stat.mean; 302 else 303 val = ru_stats.ru_stime_usec_stat.mean; 304 305 count->val = val; 306 if (old_count) { 307 count->run = old_count->run + 1; 308 count->ena = old_count->ena + 1; 309 } else { 310 count->run++; 311 count->ena++; 312 } 313 return 0; 314 } 315 return err; 316 } 317 318 /* 319 * Read out the results of a single counter: 320 * do not aggregate counts across CPUs in system-wide mode 321 */ 322 static int read_counter_cpu(struct evsel *counter, int cpu_map_idx) 323 { 324 int nthreads = perf_thread_map__nr(evsel_list->core.threads); 325 int thread; 326 327 if (!counter->supported) 328 return -ENOENT; 329 330 for (thread = 0; thread < nthreads; thread++) { 331 struct perf_counts_values *count; 332 333 count = perf_counts(counter->counts, cpu_map_idx, thread); 334 335 /* 336 * The leader's group read loads data into its group members 337 * (via evsel__read_counter()) and sets their count->loaded. 338 */ 339 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) && 340 read_single_counter(counter, cpu_map_idx, thread)) { 341 counter->counts->scaled = -1; 342 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0; 343 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0; 344 return -1; 345 } 346 347 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false); 348 349 if (STAT_RECORD) { 350 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) { 351 pr_err("failed to write stat event\n"); 352 return -1; 353 } 354 } 355 356 if (verbose > 1) { 357 fprintf(stat_config.output, 358 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 359 evsel__name(counter), 360 perf_cpu_map__cpu(evsel__cpus(counter), 361 cpu_map_idx).cpu, 362 count->val, count->ena, count->run); 363 } 364 } 365 366 return 0; 367 } 368 369 static int read_counters_with_affinity(void) 370 { 371 struct evlist_cpu_iterator evlist_cpu_itr; 372 373 if (all_counters_use_bpf) 374 return 0; 375 376 evlist__for_each_cpu(evlist_cpu_itr, evsel_list) { 377 struct evsel *counter = evlist_cpu_itr.evsel; 378 379 if (evsel__is_bpf(counter)) 380 continue; 381 382 if (evsel__is_tool(counter)) 383 continue; 384 385 if (!counter->err) 386 counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx); 387 } 388 389 return 0; 390 } 391 392 static int read_bpf_map_counters(void) 393 { 394 struct evsel *counter; 395 int err; 396 397 evlist__for_each_entry(evsel_list, counter) { 398 if (!evsel__is_bpf(counter)) 399 continue; 400 401 err = bpf_counter__read(counter); 402 if (err) 403 return err; 404 } 405 return 0; 406 } 407 408 static int read_tool_counters(void) 409 { 410 struct evsel *counter; 411 412 evlist__for_each_entry(evsel_list, counter) { 413 int idx; 414 415 if (!evsel__is_tool(counter)) 416 continue; 417 418 perf_cpu_map__for_each_idx(idx, counter->core.cpus) { 419 if (!counter->err) 420 counter->err = read_counter_cpu(counter, idx); 421 } 422 } 423 return 0; 424 } 425 426 static int read_counters(void) 427 { 428 int ret; 429 430 if (stat_config.stop_read_counter) 431 return 0; 432 433 // Read all BPF counters first. 434 ret = read_bpf_map_counters(); 435 if (ret) 436 return ret; 437 438 // Read non-BPF and non-tool counters next. 439 ret = read_counters_with_affinity(); 440 if (ret) 441 return ret; 442 443 // Read the tool counters last. This way the duration_time counter 444 // should always be greater than any other counter's enabled time. 445 return read_tool_counters(); 446 } 447 448 static void process_counters(void) 449 { 450 struct evsel *counter; 451 452 evlist__for_each_entry(evsel_list, counter) { 453 if (counter->err) 454 pr_debug("failed to read counter %s\n", counter->name); 455 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) 456 pr_warning("failed to process counter %s\n", counter->name); 457 counter->err = 0; 458 } 459 460 perf_stat_merge_counters(&stat_config, evsel_list); 461 perf_stat_process_percore(&stat_config, evsel_list); 462 } 463 464 static void process_interval(void) 465 { 466 struct timespec ts, rs; 467 468 clock_gettime(CLOCK_MONOTONIC, &ts); 469 diff_timespec(&rs, &ts, &ref_time); 470 471 evlist__reset_aggr_stats(evsel_list); 472 473 if (read_counters() == 0) 474 process_counters(); 475 476 if (STAT_RECORD) { 477 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 478 pr_err("failed to write stat round event\n"); 479 } 480 481 init_stats(stat_config.walltime_nsecs_stats); 482 update_stats(stat_config.walltime_nsecs_stats, stat_config.interval * 1000000ULL); 483 print_counters(&rs, 0, NULL); 484 } 485 486 static bool handle_interval(unsigned int interval, int *times) 487 { 488 if (interval) { 489 process_interval(); 490 if (interval_count && !(--(*times))) 491 return true; 492 } 493 return false; 494 } 495 496 static int enable_counters(void) 497 { 498 struct evsel *evsel; 499 int err; 500 501 evlist__for_each_entry(evsel_list, evsel) { 502 if (!evsel__is_bpf(evsel)) 503 continue; 504 505 err = bpf_counter__enable(evsel); 506 if (err) 507 return err; 508 } 509 510 if (!target__enable_on_exec(&target)) { 511 if (!all_counters_use_bpf) 512 evlist__enable(evsel_list); 513 } 514 return 0; 515 } 516 517 static void disable_counters(void) 518 { 519 struct evsel *counter; 520 521 /* 522 * If we don't have tracee (attaching to task or cpu), counters may 523 * still be running. To get accurate group ratios, we must stop groups 524 * from counting before reading their constituent counters. 525 */ 526 if (!target__none(&target)) { 527 evlist__for_each_entry(evsel_list, counter) 528 bpf_counter__disable(counter); 529 if (!all_counters_use_bpf) 530 evlist__disable(evsel_list); 531 } 532 } 533 534 static volatile sig_atomic_t workload_exec_errno; 535 536 /* 537 * evlist__prepare_workload will send a SIGUSR1 538 * if the fork fails, since we asked by setting its 539 * want_signal to true. 540 */ 541 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 542 void *ucontext __maybe_unused) 543 { 544 workload_exec_errno = info->si_value.sival_int; 545 } 546 547 static bool evsel__should_store_id(struct evsel *counter) 548 { 549 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 550 } 551 552 static bool is_target_alive(struct target *_target, 553 struct perf_thread_map *threads) 554 { 555 struct stat st; 556 int i; 557 558 if (!target__has_task(_target)) 559 return true; 560 561 for (i = 0; i < threads->nr; i++) { 562 char path[PATH_MAX]; 563 564 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 565 threads->map[i].pid); 566 567 if (!stat(path, &st)) 568 return true; 569 } 570 571 return false; 572 } 573 574 static void process_evlist(struct evlist *evlist, unsigned int interval) 575 { 576 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED; 577 578 if (evlist__ctlfd_process(evlist, &cmd) > 0) { 579 switch (cmd) { 580 case EVLIST_CTL_CMD_ENABLE: 581 fallthrough; 582 case EVLIST_CTL_CMD_DISABLE: 583 if (interval) 584 process_interval(); 585 break; 586 case EVLIST_CTL_CMD_SNAPSHOT: 587 case EVLIST_CTL_CMD_ACK: 588 case EVLIST_CTL_CMD_UNSUPPORTED: 589 case EVLIST_CTL_CMD_EVLIST: 590 case EVLIST_CTL_CMD_STOP: 591 case EVLIST_CTL_CMD_PING: 592 default: 593 break; 594 } 595 } 596 } 597 598 static void compute_tts(struct timespec *time_start, struct timespec *time_stop, 599 int *time_to_sleep) 600 { 601 int tts = *time_to_sleep; 602 struct timespec time_diff; 603 604 diff_timespec(&time_diff, time_stop, time_start); 605 606 tts -= time_diff.tv_sec * MSEC_PER_SEC + 607 time_diff.tv_nsec / NSEC_PER_MSEC; 608 609 if (tts < 0) 610 tts = 0; 611 612 *time_to_sleep = tts; 613 } 614 615 static int dispatch_events(bool forks, int timeout, int interval, int *times) 616 { 617 int child_exited = 0, status = 0; 618 int time_to_sleep, sleep_time; 619 struct timespec time_start, time_stop; 620 621 if (interval) 622 sleep_time = interval; 623 else if (timeout) 624 sleep_time = timeout; 625 else 626 sleep_time = 1000; 627 628 time_to_sleep = sleep_time; 629 630 while (!done) { 631 if (forks) 632 child_exited = waitpid(child_pid, &status, WNOHANG); 633 else 634 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0; 635 636 if (child_exited) 637 break; 638 639 clock_gettime(CLOCK_MONOTONIC, &time_start); 640 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */ 641 if (timeout || handle_interval(interval, times)) 642 break; 643 time_to_sleep = sleep_time; 644 } else { /* fd revent */ 645 process_evlist(evsel_list, interval); 646 clock_gettime(CLOCK_MONOTONIC, &time_stop); 647 compute_tts(&time_start, &time_stop, &time_to_sleep); 648 } 649 } 650 651 return status; 652 } 653 654 enum counter_recovery { 655 COUNTER_SKIP, 656 COUNTER_RETRY, 657 }; 658 659 static enum counter_recovery stat_handle_error(struct evsel *counter, int err) 660 { 661 char msg[BUFSIZ]; 662 663 assert(!counter->supported); 664 665 /* 666 * PPC returns ENXIO for HW counters until 2.6.37 667 * (behavior changed with commit b0a873e). 668 */ 669 if (err == EINVAL || err == ENOSYS || err == ENOENT || err == ENXIO) { 670 if (verbose > 0) { 671 evsel__open_strerror(counter, &target, err, msg, sizeof(msg)); 672 ui__warning("%s event is not supported by the kernel.\n%s\n", 673 evsel__name(counter), msg); 674 } 675 return COUNTER_SKIP; 676 } 677 if (evsel__fallback(counter, &target, err, msg, sizeof(msg))) { 678 if (verbose > 0) 679 ui__warning("%s\n", msg); 680 counter->supported = true; 681 return COUNTER_RETRY; 682 } 683 if (target__has_per_thread(&target) && err != EOPNOTSUPP && 684 evsel_list->core.threads && evsel_list->core.threads->err_thread != -1) { 685 /* 686 * For global --per-thread case, skip current 687 * error thread. 688 */ 689 if (!thread_map__remove(evsel_list->core.threads, 690 evsel_list->core.threads->err_thread)) { 691 evsel_list->core.threads->err_thread = -1; 692 counter->supported = true; 693 return COUNTER_RETRY; 694 } 695 } 696 if (verbose > 0) { 697 evsel__open_strerror(counter, &target, err, msg, sizeof(msg)); 698 ui__warning(err == EOPNOTSUPP 699 ? "%s event is not supported by the kernel.\n%s\n" 700 : "skipping event %s that kernel failed to open.\n%s\n", 701 evsel__name(counter), msg); 702 } 703 return COUNTER_SKIP; 704 } 705 706 static int create_perf_stat_counter(struct evsel *evsel, 707 struct perf_stat_config *config, 708 int cpu_map_idx) 709 { 710 struct perf_event_attr *attr = &evsel->core.attr; 711 struct evsel *leader = evsel__leader(evsel); 712 713 /* Reset supported flag as creating a stat counter is retried. */ 714 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 715 PERF_FORMAT_TOTAL_TIME_RUNNING; 716 717 /* 718 * The event is part of non trivial group, let's enable 719 * the group read (for leader) and ID retrieval for all 720 * members. 721 */ 722 if (leader->core.nr_members > 1) 723 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP; 724 725 attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list); 726 727 /* 728 * Some events get initialized with sample_(period/type) set, 729 * like tracepoints. Clear it up for counting. 730 */ 731 attr->sample_period = 0; 732 733 if (config->identifier) 734 attr->sample_type = PERF_SAMPLE_IDENTIFIER; 735 736 if (config->all_user) { 737 attr->exclude_kernel = 1; 738 attr->exclude_user = 0; 739 } 740 741 if (config->all_kernel) { 742 attr->exclude_kernel = 0; 743 attr->exclude_user = 1; 744 } 745 746 /* 747 * Disabling all counters initially, they will be enabled 748 * either manually by us or by kernel via enable_on_exec 749 * set later. 750 */ 751 if (evsel__is_group_leader(evsel)) { 752 attr->disabled = 1; 753 754 if (target__enable_on_exec(&target)) 755 attr->enable_on_exec = 1; 756 } 757 758 return evsel__open_per_cpu_and_thread(evsel, evsel__cpus(evsel), cpu_map_idx, 759 evsel->core.threads); 760 } 761 762 static void update_rusage_stats(const struct rusage *rusage) 763 { 764 const u64 us_to_ns = 1000; 765 const u64 s_to_ns = 1000000000; 766 767 update_stats(&ru_stats.ru_utime_usec_stat, 768 (rusage->ru_utime.tv_usec * us_to_ns + rusage->ru_utime.tv_sec * s_to_ns)); 769 update_stats(&ru_stats.ru_stime_usec_stat, 770 (rusage->ru_stime.tv_usec * us_to_ns + rusage->ru_stime.tv_sec * s_to_ns)); 771 } 772 773 static int __run_perf_stat(int argc, const char **argv, int run_idx) 774 { 775 int interval = stat_config.interval; 776 int times = stat_config.times; 777 int timeout = stat_config.timeout; 778 char msg[BUFSIZ]; 779 unsigned long long t0, t1; 780 struct evsel *counter; 781 size_t l; 782 int status = 0; 783 const bool forks = (argc > 0); 784 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; 785 struct evlist_cpu_iterator evlist_cpu_itr; 786 int err, open_err = 0; 787 bool second_pass = false, has_supported_counters; 788 789 if (forks) { 790 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) { 791 perror("failed to prepare workload"); 792 return -1; 793 } 794 child_pid = evsel_list->workload.pid; 795 } 796 797 evlist__for_each_entry(evsel_list, counter) { 798 counter->reset_group = false; 799 if (bpf_counter__load(counter, &target)) { 800 err = -1; 801 goto err_out; 802 } 803 if (!(evsel__is_bperf(counter))) 804 all_counters_use_bpf = false; 805 } 806 807 evlist__reset_aggr_stats(evsel_list); 808 809 /* 810 * bperf calls evsel__open_per_cpu() in bperf__load(), so 811 * no need to call it again here. 812 */ 813 if (!target.use_bpf) { 814 evlist__for_each_cpu(evlist_cpu_itr, evsel_list) { 815 counter = evlist_cpu_itr.evsel; 816 817 if (counter->reset_group || !counter->supported) 818 continue; 819 if (evsel__is_bperf(counter)) 820 continue; 821 822 while (true) { 823 if (create_perf_stat_counter(counter, &stat_config, 824 evlist_cpu_itr.cpu_map_idx) == 0) 825 break; 826 827 open_err = errno; 828 /* 829 * Weak group failed. We cannot just undo this 830 * here because earlier CPUs might be in group 831 * mode, and the kernel doesn't support mixing 832 * group and non group reads. Defer it to later. 833 * Don't close here because we're in the wrong 834 * affinity. 835 */ 836 if ((open_err == EINVAL || open_err == EBADF) && 837 evsel__leader(counter) != counter && 838 counter->weak_group) { 839 evlist__reset_weak_group(evsel_list, counter, false); 840 assert(counter->reset_group); 841 counter->supported = true; 842 second_pass = true; 843 break; 844 } 845 846 if (stat_handle_error(counter, open_err) != COUNTER_RETRY) 847 break; 848 } 849 } 850 } 851 if (second_pass) { 852 /* 853 * Now redo all the weak group after closing them, 854 * and also close errored counters. 855 */ 856 857 /* First close errored or weak retry */ 858 evlist__for_each_cpu(evlist_cpu_itr, evsel_list) { 859 counter = evlist_cpu_itr.evsel; 860 861 if (!counter->reset_group && counter->supported) 862 continue; 863 864 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx); 865 } 866 /* Now reopen weak */ 867 evlist__for_each_cpu(evlist_cpu_itr, evsel_list) { 868 counter = evlist_cpu_itr.evsel; 869 870 if (!counter->reset_group) 871 continue; 872 873 while (true) { 874 pr_debug2("reopening weak %s\n", evsel__name(counter)); 875 if (create_perf_stat_counter(counter, &stat_config, 876 evlist_cpu_itr.cpu_map_idx) == 0) { 877 evlist_cpu_iterator__exit(&evlist_cpu_itr); 878 break; 879 } 880 open_err = errno; 881 if (stat_handle_error(counter, open_err) != COUNTER_RETRY) { 882 evlist_cpu_iterator__exit(&evlist_cpu_itr); 883 break; 884 } 885 } 886 } 887 } 888 889 has_supported_counters = false; 890 evlist__for_each_entry(evsel_list, counter) { 891 if (!counter->supported) { 892 perf_evsel__free_fd(&counter->core); 893 continue; 894 } 895 has_supported_counters = true; 896 897 l = strlen(counter->unit); 898 if (l > stat_config.unit_width) 899 stat_config.unit_width = l; 900 901 if (evsel__should_store_id(counter) && 902 evsel__store_ids(counter, evsel_list)) { 903 err = -1; 904 goto err_out; 905 } 906 } 907 if (!has_supported_counters && !stat_config.null_run) { 908 if (open_err) { 909 evsel__open_strerror(evlist__first(evsel_list), &target, open_err, 910 msg, sizeof(msg)); 911 } 912 ui__error("No supported events found.\n%s\n", msg); 913 914 if (child_pid != -1) 915 kill(child_pid, SIGTERM); 916 err = -1; 917 goto err_out; 918 } 919 920 if (evlist__apply_filters(evsel_list, &counter, &target)) { 921 pr_err("failed to set filter \"%s\" on event %s: %m\n", 922 counter->filter, evsel__name(counter)); 923 return -1; 924 } 925 926 if (STAT_RECORD) { 927 int fd = perf_data__fd(&perf_stat.data); 928 929 if (is_pipe) { 930 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); 931 } else { 932 err = perf_session__write_header(perf_stat.session, evsel_list, 933 fd, false); 934 } 935 936 if (err < 0) 937 goto err_out; 938 939 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list, 940 process_synthesized_event, is_pipe); 941 if (err < 0) 942 goto err_out; 943 944 } 945 946 if (target.initial_delay) { 947 pr_info(EVLIST_DISABLED_MSG); 948 } else { 949 err = enable_counters(); 950 if (err) { 951 err = -1; 952 goto err_out; 953 } 954 } 955 956 /* Exec the command, if any */ 957 if (forks) 958 evlist__start_workload(evsel_list); 959 960 if (target.initial_delay > 0) { 961 usleep(target.initial_delay * USEC_PER_MSEC); 962 err = enable_counters(); 963 if (err) { 964 err = -1; 965 goto err_out; 966 } 967 968 pr_info(EVLIST_ENABLED_MSG); 969 } 970 971 t0 = rdclock(); 972 clock_gettime(CLOCK_MONOTONIC, &ref_time); 973 974 if (forks) { 975 if (interval || timeout || evlist__ctlfd_initialized(evsel_list)) 976 status = dispatch_events(forks, timeout, interval, ×); 977 if (child_pid != -1) { 978 if (timeout) 979 kill(child_pid, SIGTERM); 980 wait4(child_pid, &status, 0, &stat_config.ru_data); 981 } 982 983 if (workload_exec_errno) { 984 errno = workload_exec_errno; 985 pr_err("Workload failed: %m\n"); 986 err = -1; 987 goto err_out; 988 } 989 990 if (WIFSIGNALED(status)) { 991 /* 992 * We want to indicate failure to stop a repeat run, 993 * hence negative. We want the value to be the exit code 994 * of perf, which for termination by a signal is 128 995 * plus the signal number. 996 */ 997 err = 0 - (128 + WTERMSIG(status)); 998 psignal(WTERMSIG(status), argv[0]); 999 } else { 1000 err = WEXITSTATUS(status); 1001 } 1002 } else { 1003 err = dispatch_events(forks, timeout, interval, ×); 1004 } 1005 1006 disable_counters(); 1007 1008 t1 = rdclock(); 1009 1010 if (stat_config.walltime_run_table) 1011 stat_config.walltime_run[run_idx] = t1 - t0; 1012 1013 if (interval && stat_config.summary) { 1014 stat_config.interval = 0; 1015 stat_config.stop_read_counter = true; 1016 init_stats(stat_config.walltime_nsecs_stats); 1017 update_stats(stat_config.walltime_nsecs_stats, t1 - t0); 1018 1019 evlist__copy_prev_raw_counts(evsel_list); 1020 evlist__reset_prev_raw_counts(evsel_list); 1021 evlist__reset_aggr_stats(evsel_list); 1022 } else { 1023 update_stats(stat_config.walltime_nsecs_stats, t1 - t0); 1024 update_rusage_stats(&stat_config.ru_data); 1025 } 1026 1027 /* 1028 * Closing a group leader splits the group, and as we only disable 1029 * group leaders, results in remaining events becoming enabled. To 1030 * avoid arbitrary skew, we must read all counters before closing any 1031 * group leaders. 1032 */ 1033 if (read_counters() == 0) 1034 process_counters(); 1035 1036 /* 1037 * We need to keep evsel_list alive, because it's processed 1038 * later the evsel_list will be closed after. 1039 */ 1040 if (!STAT_RECORD) 1041 evlist__close(evsel_list); 1042 1043 return err; 1044 1045 err_out: 1046 if (forks) 1047 evlist__cancel_workload(evsel_list); 1048 1049 return err; 1050 } 1051 1052 /* 1053 * Returns -1 for fatal errors which signifies to not continue 1054 * when in repeat mode. 1055 * 1056 * Returns < -1 error codes when stat record is used. These 1057 * result in the stat information being displayed, but writing 1058 * to the file fails and is non fatal. 1059 */ 1060 static int run_perf_stat(int argc, const char **argv, int run_idx) 1061 { 1062 int ret; 1063 1064 if (pre_cmd) { 1065 ret = system(pre_cmd); 1066 if (ret) 1067 return ret; 1068 } 1069 1070 if (sync_run) 1071 sync(); 1072 1073 ret = __run_perf_stat(argc, argv, run_idx); 1074 if (ret) 1075 return ret; 1076 1077 if (post_cmd) { 1078 ret = system(post_cmd); 1079 if (ret) 1080 return ret; 1081 } 1082 1083 return ret; 1084 } 1085 1086 static void print_counters(struct timespec *ts, int argc, const char **argv) 1087 { 1088 /* Do not print anything if we record to the pipe. */ 1089 if (STAT_RECORD && perf_stat.data.is_pipe) 1090 return; 1091 if (quiet) 1092 return; 1093 1094 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv); 1095 } 1096 1097 static volatile sig_atomic_t signr = -1; 1098 1099 static void skip_signal(int signo) 1100 { 1101 if ((child_pid == -1) || stat_config.interval) 1102 done = 1; 1103 1104 signr = signo; 1105 /* 1106 * render child_pid harmless 1107 * won't send SIGTERM to a random 1108 * process in case of race condition 1109 * and fast PID recycling 1110 */ 1111 child_pid = -1; 1112 } 1113 1114 static void sig_atexit(void) 1115 { 1116 sigset_t set, oset; 1117 1118 /* 1119 * avoid race condition with SIGCHLD handler 1120 * in skip_signal() which is modifying child_pid 1121 * goal is to avoid send SIGTERM to a random 1122 * process 1123 */ 1124 sigemptyset(&set); 1125 sigaddset(&set, SIGCHLD); 1126 sigprocmask(SIG_BLOCK, &set, &oset); 1127 1128 if (child_pid != -1) 1129 kill(child_pid, SIGTERM); 1130 1131 sigprocmask(SIG_SETMASK, &oset, NULL); 1132 1133 if (signr == -1) 1134 return; 1135 1136 signal(signr, SIG_DFL); 1137 kill(getpid(), signr); 1138 } 1139 1140 static int stat__set_big_num(const struct option *opt __maybe_unused, 1141 const char *s __maybe_unused, int unset) 1142 { 1143 big_num_opt = unset ? 0 : 1; 1144 perf_stat__set_big_num(!unset); 1145 return 0; 1146 } 1147 1148 static int enable_metric_only(const struct option *opt __maybe_unused, 1149 const char *s __maybe_unused, int unset) 1150 { 1151 force_metric_only = true; 1152 stat_config.metric_only = !unset; 1153 return 0; 1154 } 1155 1156 static int append_metric_groups(const struct option *opt __maybe_unused, 1157 const char *str, 1158 int unset __maybe_unused) 1159 { 1160 if (metrics) { 1161 char *tmp; 1162 1163 if (asprintf(&tmp, "%s,%s", metrics, str) < 0) 1164 return -ENOMEM; 1165 free(metrics); 1166 metrics = tmp; 1167 } else { 1168 metrics = strdup(str); 1169 if (!metrics) 1170 return -ENOMEM; 1171 } 1172 return 0; 1173 } 1174 1175 static int parse_control_option(const struct option *opt, 1176 const char *str, 1177 int unset __maybe_unused) 1178 { 1179 struct perf_stat_config *config = opt->value; 1180 1181 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close); 1182 } 1183 1184 static int parse_stat_cgroups(const struct option *opt, 1185 const char *str, int unset) 1186 { 1187 if (stat_config.cgroup_list) { 1188 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 1189 return -1; 1190 } 1191 1192 return parse_cgroups(opt, str, unset); 1193 } 1194 1195 static int parse_cputype(const struct option *opt, 1196 const char *str, 1197 int unset __maybe_unused) 1198 { 1199 const struct perf_pmu *pmu; 1200 struct evlist *evlist = *(struct evlist **)opt->value; 1201 1202 if (!list_empty(&evlist->core.entries)) { 1203 fprintf(stderr, "Must define cputype before events/metrics\n"); 1204 return -1; 1205 } 1206 1207 pmu = perf_pmus__pmu_for_pmu_filter(str); 1208 if (!pmu) { 1209 fprintf(stderr, "--cputype %s is not supported!\n", str); 1210 return -1; 1211 } 1212 parse_events_option_args.pmu_filter = pmu->name; 1213 1214 return 0; 1215 } 1216 1217 static int parse_cache_level(const struct option *opt, 1218 const char *str, 1219 int unset __maybe_unused) 1220 { 1221 int level; 1222 struct opt_aggr_mode *opt_aggr_mode = (struct opt_aggr_mode *)opt->value; 1223 u32 *aggr_level = (u32 *)opt->data; 1224 1225 /* 1226 * If no string is specified, aggregate based on the topology of 1227 * Last Level Cache (LLC). Since the LLC level can change from 1228 * architecture to architecture, set level greater than 1229 * MAX_CACHE_LVL which will be interpreted as LLC. 1230 */ 1231 if (str == NULL) { 1232 level = MAX_CACHE_LVL + 1; 1233 goto out; 1234 } 1235 1236 /* 1237 * The format to specify cache level is LX or lX where X is the 1238 * cache level. 1239 */ 1240 if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) { 1241 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1242 MAX_CACHE_LVL, 1243 MAX_CACHE_LVL); 1244 return -EINVAL; 1245 } 1246 1247 level = atoi(&str[1]); 1248 if (level < 1) { 1249 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1250 MAX_CACHE_LVL, 1251 MAX_CACHE_LVL); 1252 return -EINVAL; 1253 } 1254 1255 if (level > MAX_CACHE_LVL) { 1256 pr_err("perf only supports max cache level of %d.\n" 1257 "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL); 1258 return -EINVAL; 1259 } 1260 out: 1261 opt_aggr_mode->cache = true; 1262 *aggr_level = level; 1263 return 0; 1264 } 1265 1266 /** 1267 * Calculate the cache instance ID from the map in 1268 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1269 * Cache instance ID is the first CPU reported in the shared_cpu_list file. 1270 */ 1271 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map) 1272 { 1273 int id; 1274 struct perf_cpu_map *cpu_map = perf_cpu_map__new(map); 1275 1276 /* 1277 * If the map contains no CPU, consider the current CPU to 1278 * be the first online CPU in the cache domain else use the 1279 * first online CPU of the cache domain as the ID. 1280 */ 1281 id = perf_cpu_map__min(cpu_map).cpu; 1282 if (id == -1) 1283 id = cpu.cpu; 1284 1285 /* Free the perf_cpu_map used to find the cache ID */ 1286 perf_cpu_map__put(cpu_map); 1287 1288 return id; 1289 } 1290 1291 /** 1292 * cpu__get_cache_id - Returns 0 if successful in populating the 1293 * cache level and cache id. Cache level is read from 1294 * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID 1295 * is the first CPU reported by 1296 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1297 */ 1298 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache) 1299 { 1300 int ret = 0; 1301 u32 cache_level = stat_config.aggr_level; 1302 struct cpu_cache_level caches[MAX_CACHE_LVL]; 1303 u32 i = 0, caches_cnt = 0; 1304 1305 cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1306 cache->cache = -1; 1307 1308 ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt); 1309 if (ret) { 1310 /* 1311 * If caches_cnt is not 0, cpu_cache_level data 1312 * was allocated when building the topology. 1313 * Free the allocated data before returning. 1314 */ 1315 if (caches_cnt) 1316 goto free_caches; 1317 1318 return ret; 1319 } 1320 1321 if (!caches_cnt) 1322 return -1; 1323 1324 /* 1325 * Save the data for the highest level if no 1326 * level was specified by the user. 1327 */ 1328 if (cache_level > MAX_CACHE_LVL) { 1329 int max_level_index = 0; 1330 1331 for (i = 1; i < caches_cnt; ++i) { 1332 if (caches[i].level > caches[max_level_index].level) 1333 max_level_index = i; 1334 } 1335 1336 cache->cache_lvl = caches[max_level_index].level; 1337 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map); 1338 1339 /* Reset i to 0 to free entire caches[] */ 1340 i = 0; 1341 goto free_caches; 1342 } 1343 1344 for (i = 0; i < caches_cnt; ++i) { 1345 if (caches[i].level == cache_level) { 1346 cache->cache_lvl = cache_level; 1347 cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1348 } 1349 1350 cpu_cache_level__free(&caches[i]); 1351 } 1352 1353 free_caches: 1354 /* 1355 * Free all the allocated cpu_cache_level data. 1356 */ 1357 while (i < caches_cnt) 1358 cpu_cache_level__free(&caches[i++]); 1359 1360 return ret; 1361 } 1362 1363 /** 1364 * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache 1365 * level, die and socket populated with the cache instache ID, cache level, 1366 * die and socket for cpu. The function signature is compatible with 1367 * aggr_cpu_id_get_t. 1368 */ 1369 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data) 1370 { 1371 int ret; 1372 struct aggr_cpu_id id; 1373 struct perf_cache cache; 1374 1375 id = aggr_cpu_id__die(cpu, data); 1376 if (aggr_cpu_id__is_empty(&id)) 1377 return id; 1378 1379 ret = cpu__get_cache_details(cpu, &cache); 1380 if (ret) 1381 return id; 1382 1383 id.cache_lvl = cache.cache_lvl; 1384 id.cache = cache.cache; 1385 return id; 1386 } 1387 1388 static const char *const aggr_mode__string[] = { 1389 [AGGR_CORE] = "core", 1390 [AGGR_CACHE] = "cache", 1391 [AGGR_CLUSTER] = "cluster", 1392 [AGGR_DIE] = "die", 1393 [AGGR_GLOBAL] = "global", 1394 [AGGR_NODE] = "node", 1395 [AGGR_NONE] = "none", 1396 [AGGR_SOCKET] = "socket", 1397 [AGGR_THREAD] = "thread", 1398 [AGGR_UNSET] = "unset", 1399 }; 1400 1401 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 1402 struct perf_cpu cpu) 1403 { 1404 return aggr_cpu_id__socket(cpu, /*data=*/NULL); 1405 } 1406 1407 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 1408 struct perf_cpu cpu) 1409 { 1410 return aggr_cpu_id__die(cpu, /*data=*/NULL); 1411 } 1412 1413 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused, 1414 struct perf_cpu cpu) 1415 { 1416 return aggr_cpu_id__cache(cpu, /*data=*/NULL); 1417 } 1418 1419 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused, 1420 struct perf_cpu cpu) 1421 { 1422 return aggr_cpu_id__cluster(cpu, /*data=*/NULL); 1423 } 1424 1425 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 1426 struct perf_cpu cpu) 1427 { 1428 return aggr_cpu_id__core(cpu, /*data=*/NULL); 1429 } 1430 1431 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused, 1432 struct perf_cpu cpu) 1433 { 1434 return aggr_cpu_id__node(cpu, /*data=*/NULL); 1435 } 1436 1437 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused, 1438 struct perf_cpu cpu) 1439 { 1440 return aggr_cpu_id__global(cpu, /*data=*/NULL); 1441 } 1442 1443 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused, 1444 struct perf_cpu cpu) 1445 { 1446 return aggr_cpu_id__cpu(cpu, /*data=*/NULL); 1447 } 1448 1449 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config, 1450 aggr_get_id_t get_id, struct perf_cpu cpu) 1451 { 1452 struct aggr_cpu_id id; 1453 1454 /* per-process mode - should use global aggr mode */ 1455 if (cpu.cpu == -1 || cpu.cpu >= config->cpus_aggr_map->nr) 1456 return get_id(config, cpu); 1457 1458 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu])) 1459 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu); 1460 1461 id = config->cpus_aggr_map->map[cpu.cpu]; 1462 return id; 1463 } 1464 1465 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config, 1466 struct perf_cpu cpu) 1467 { 1468 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu); 1469 } 1470 1471 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config, 1472 struct perf_cpu cpu) 1473 { 1474 return perf_stat__get_aggr(config, perf_stat__get_die, cpu); 1475 } 1476 1477 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config, 1478 struct perf_cpu cpu) 1479 { 1480 return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu); 1481 } 1482 1483 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config, 1484 struct perf_cpu cpu) 1485 { 1486 return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu); 1487 } 1488 1489 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config, 1490 struct perf_cpu cpu) 1491 { 1492 return perf_stat__get_aggr(config, perf_stat__get_core, cpu); 1493 } 1494 1495 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config, 1496 struct perf_cpu cpu) 1497 { 1498 return perf_stat__get_aggr(config, perf_stat__get_node, cpu); 1499 } 1500 1501 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config, 1502 struct perf_cpu cpu) 1503 { 1504 return perf_stat__get_aggr(config, perf_stat__get_global, cpu); 1505 } 1506 1507 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config, 1508 struct perf_cpu cpu) 1509 { 1510 return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu); 1511 } 1512 1513 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode) 1514 { 1515 switch (aggr_mode) { 1516 case AGGR_SOCKET: 1517 return aggr_cpu_id__socket; 1518 case AGGR_DIE: 1519 return aggr_cpu_id__die; 1520 case AGGR_CLUSTER: 1521 return aggr_cpu_id__cluster; 1522 case AGGR_CACHE: 1523 return aggr_cpu_id__cache; 1524 case AGGR_CORE: 1525 return aggr_cpu_id__core; 1526 case AGGR_NODE: 1527 return aggr_cpu_id__node; 1528 case AGGR_NONE: 1529 return aggr_cpu_id__cpu; 1530 case AGGR_GLOBAL: 1531 return aggr_cpu_id__global; 1532 case AGGR_THREAD: 1533 case AGGR_UNSET: 1534 case AGGR_MAX: 1535 default: 1536 return NULL; 1537 } 1538 } 1539 1540 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode) 1541 { 1542 switch (aggr_mode) { 1543 case AGGR_SOCKET: 1544 return perf_stat__get_socket_cached; 1545 case AGGR_DIE: 1546 return perf_stat__get_die_cached; 1547 case AGGR_CLUSTER: 1548 return perf_stat__get_cluster_cached; 1549 case AGGR_CACHE: 1550 return perf_stat__get_cache_id_cached; 1551 case AGGR_CORE: 1552 return perf_stat__get_core_cached; 1553 case AGGR_NODE: 1554 return perf_stat__get_node_cached; 1555 case AGGR_NONE: 1556 return perf_stat__get_cpu_cached; 1557 case AGGR_GLOBAL: 1558 return perf_stat__get_global_cached; 1559 case AGGR_THREAD: 1560 case AGGR_UNSET: 1561 case AGGR_MAX: 1562 default: 1563 return NULL; 1564 } 1565 } 1566 1567 static int perf_stat_init_aggr_mode(void) 1568 { 1569 int nr; 1570 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode); 1571 1572 if (get_id) { 1573 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1574 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1575 get_id, /*data=*/NULL, needs_sort); 1576 if (!stat_config.aggr_map) { 1577 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1578 return -1; 1579 } 1580 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode); 1581 } 1582 1583 if (stat_config.aggr_mode == AGGR_THREAD) { 1584 nr = perf_thread_map__nr(evsel_list->core.threads); 1585 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1586 if (stat_config.aggr_map == NULL) 1587 return -ENOMEM; 1588 1589 for (int s = 0; s < nr; s++) { 1590 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1591 1592 id.thread_idx = s; 1593 stat_config.aggr_map->map[s] = id; 1594 } 1595 return 0; 1596 } 1597 1598 /* 1599 * The evsel_list->cpus is the base we operate on, 1600 * taking the highest cpu number to be the size of 1601 * the aggregation translate cpumap. 1602 */ 1603 nr = perf_cpu_map__max(evsel_list->core.all_cpus).cpu + 1; 1604 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr); 1605 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 1606 } 1607 1608 static void cpu_aggr_map__delete(struct cpu_aggr_map *map) 1609 { 1610 free(map); 1611 } 1612 1613 static void perf_stat__exit_aggr_mode(void) 1614 { 1615 cpu_aggr_map__delete(stat_config.aggr_map); 1616 cpu_aggr_map__delete(stat_config.cpus_aggr_map); 1617 stat_config.aggr_map = NULL; 1618 stat_config.cpus_aggr_map = NULL; 1619 } 1620 1621 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data) 1622 { 1623 struct perf_env *env = data; 1624 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1625 1626 if (cpu.cpu != -1) 1627 id.socket = env->cpu[cpu.cpu].socket_id; 1628 1629 return id; 1630 } 1631 1632 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data) 1633 { 1634 struct perf_env *env = data; 1635 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1636 1637 if (cpu.cpu != -1) { 1638 /* 1639 * die_id is relative to socket, so start 1640 * with the socket ID and then add die to 1641 * make a unique ID. 1642 */ 1643 id.socket = env->cpu[cpu.cpu].socket_id; 1644 id.die = env->cpu[cpu.cpu].die_id; 1645 } 1646 1647 return id; 1648 } 1649 1650 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env, 1651 u32 cache_level, struct aggr_cpu_id *id) 1652 { 1653 int i; 1654 int caches_cnt = env->caches_cnt; 1655 struct cpu_cache_level *caches = env->caches; 1656 1657 id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1658 id->cache = -1; 1659 1660 if (!caches_cnt) 1661 return; 1662 1663 for (i = caches_cnt - 1; i > -1; --i) { 1664 struct perf_cpu_map *cpu_map; 1665 int map_contains_cpu; 1666 1667 /* 1668 * If user has not specified a level, find the fist level with 1669 * the cpu in the map. Since building the map is expensive, do 1670 * this only if levels match. 1671 */ 1672 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level) 1673 continue; 1674 1675 cpu_map = perf_cpu_map__new(caches[i].map); 1676 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu); 1677 perf_cpu_map__put(cpu_map); 1678 1679 if (map_contains_cpu != -1) { 1680 id->cache_lvl = caches[i].level; 1681 id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1682 return; 1683 } 1684 } 1685 } 1686 1687 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu, 1688 void *data) 1689 { 1690 struct perf_env *env = data; 1691 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1692 1693 if (cpu.cpu != -1) { 1694 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level; 1695 1696 id.socket = env->cpu[cpu.cpu].socket_id; 1697 id.die = env->cpu[cpu.cpu].die_id; 1698 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id); 1699 } 1700 1701 return id; 1702 } 1703 1704 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu, 1705 void *data) 1706 { 1707 struct perf_env *env = data; 1708 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1709 1710 if (cpu.cpu != -1) { 1711 id.socket = env->cpu[cpu.cpu].socket_id; 1712 id.die = env->cpu[cpu.cpu].die_id; 1713 id.cluster = env->cpu[cpu.cpu].cluster_id; 1714 } 1715 1716 return id; 1717 } 1718 1719 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data) 1720 { 1721 struct perf_env *env = data; 1722 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1723 1724 if (cpu.cpu != -1) { 1725 /* 1726 * core_id is relative to socket, die and cluster, we need a 1727 * global id. So we set socket, die id, cluster id and core id. 1728 */ 1729 id.socket = env->cpu[cpu.cpu].socket_id; 1730 id.die = env->cpu[cpu.cpu].die_id; 1731 id.cluster = env->cpu[cpu.cpu].cluster_id; 1732 id.core = env->cpu[cpu.cpu].core_id; 1733 } 1734 1735 return id; 1736 } 1737 1738 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data) 1739 { 1740 struct perf_env *env = data; 1741 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1742 1743 if (cpu.cpu != -1) { 1744 /* 1745 * core_id is relative to socket and die, 1746 * we need a global id. So we set 1747 * socket, die id and core id 1748 */ 1749 id.socket = env->cpu[cpu.cpu].socket_id; 1750 id.die = env->cpu[cpu.cpu].die_id; 1751 id.core = env->cpu[cpu.cpu].core_id; 1752 id.cpu = cpu; 1753 } 1754 1755 return id; 1756 } 1757 1758 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data) 1759 { 1760 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1761 1762 id.node = perf_env__numa_node(data, cpu); 1763 return id; 1764 } 1765 1766 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused, 1767 void *data __maybe_unused) 1768 { 1769 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1770 1771 /* it always aggregates to the cpu 0 */ 1772 id.cpu = (struct perf_cpu){ .cpu = 0 }; 1773 return id; 1774 } 1775 1776 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1777 struct perf_cpu cpu) 1778 { 1779 return perf_env__get_socket_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1780 } 1781 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1782 struct perf_cpu cpu) 1783 { 1784 return perf_env__get_die_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1785 } 1786 1787 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused, 1788 struct perf_cpu cpu) 1789 { 1790 return perf_env__get_cluster_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1791 } 1792 1793 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused, 1794 struct perf_cpu cpu) 1795 { 1796 return perf_env__get_cache_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1797 } 1798 1799 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1800 struct perf_cpu cpu) 1801 { 1802 return perf_env__get_core_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1803 } 1804 1805 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused, 1806 struct perf_cpu cpu) 1807 { 1808 return perf_env__get_cpu_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1809 } 1810 1811 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, 1812 struct perf_cpu cpu) 1813 { 1814 return perf_env__get_node_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1815 } 1816 1817 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused, 1818 struct perf_cpu cpu) 1819 { 1820 return perf_env__get_global_aggr_by_cpu(cpu, perf_session__env(perf_stat.session)); 1821 } 1822 1823 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode) 1824 { 1825 switch (aggr_mode) { 1826 case AGGR_SOCKET: 1827 return perf_env__get_socket_aggr_by_cpu; 1828 case AGGR_DIE: 1829 return perf_env__get_die_aggr_by_cpu; 1830 case AGGR_CLUSTER: 1831 return perf_env__get_cluster_aggr_by_cpu; 1832 case AGGR_CACHE: 1833 return perf_env__get_cache_aggr_by_cpu; 1834 case AGGR_CORE: 1835 return perf_env__get_core_aggr_by_cpu; 1836 case AGGR_NODE: 1837 return perf_env__get_node_aggr_by_cpu; 1838 case AGGR_GLOBAL: 1839 return perf_env__get_global_aggr_by_cpu; 1840 case AGGR_NONE: 1841 return perf_env__get_cpu_aggr_by_cpu; 1842 case AGGR_THREAD: 1843 case AGGR_UNSET: 1844 case AGGR_MAX: 1845 default: 1846 return NULL; 1847 } 1848 } 1849 1850 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode) 1851 { 1852 switch (aggr_mode) { 1853 case AGGR_SOCKET: 1854 return perf_stat__get_socket_file; 1855 case AGGR_DIE: 1856 return perf_stat__get_die_file; 1857 case AGGR_CLUSTER: 1858 return perf_stat__get_cluster_file; 1859 case AGGR_CACHE: 1860 return perf_stat__get_cache_file; 1861 case AGGR_CORE: 1862 return perf_stat__get_core_file; 1863 case AGGR_NODE: 1864 return perf_stat__get_node_file; 1865 case AGGR_GLOBAL: 1866 return perf_stat__get_global_file; 1867 case AGGR_NONE: 1868 return perf_stat__get_cpu_file; 1869 case AGGR_THREAD: 1870 case AGGR_UNSET: 1871 case AGGR_MAX: 1872 default: 1873 return NULL; 1874 } 1875 } 1876 1877 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1878 { 1879 struct perf_env *env = perf_session__env(st->session); 1880 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode); 1881 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1882 1883 if (stat_config.aggr_mode == AGGR_THREAD) { 1884 int nr = perf_thread_map__nr(evsel_list->core.threads); 1885 1886 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1887 if (stat_config.aggr_map == NULL) 1888 return -ENOMEM; 1889 1890 for (int s = 0; s < nr; s++) { 1891 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1892 1893 id.thread_idx = s; 1894 stat_config.aggr_map->map[s] = id; 1895 } 1896 return 0; 1897 } 1898 1899 if (!get_id) 1900 return 0; 1901 1902 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1903 get_id, env, needs_sort); 1904 if (!stat_config.aggr_map) { 1905 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1906 return -1; 1907 } 1908 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode); 1909 return 0; 1910 } 1911 1912 static int default_evlist_evsel_cmp(void *priv __maybe_unused, 1913 const struct list_head *l, 1914 const struct list_head *r) 1915 { 1916 const struct perf_evsel *lhs_core = container_of(l, struct perf_evsel, node); 1917 const struct evsel *lhs = container_of(lhs_core, struct evsel, core); 1918 const struct perf_evsel *rhs_core = container_of(r, struct perf_evsel, node); 1919 const struct evsel *rhs = container_of(rhs_core, struct evsel, core); 1920 1921 if (evsel__leader(lhs) == evsel__leader(rhs)) { 1922 /* Within the same group, respect the original order. */ 1923 return lhs_core->idx - rhs_core->idx; 1924 } 1925 1926 /* Sort default metrics evsels first, and default show events before those. */ 1927 if (lhs->default_metricgroup != rhs->default_metricgroup) 1928 return lhs->default_metricgroup ? -1 : 1; 1929 1930 if (lhs->default_show_events != rhs->default_show_events) 1931 return lhs->default_show_events ? -1 : 1; 1932 1933 /* Sort by PMU type (prefers legacy types first). */ 1934 if (lhs->pmu != rhs->pmu) 1935 return lhs->pmu->type - rhs->pmu->type; 1936 1937 /* Sort by name. */ 1938 return strcmp(evsel__name((struct evsel *)lhs), evsel__name((struct evsel *)rhs)); 1939 } 1940 1941 /* 1942 * Add default events, if there were no attributes specified or 1943 * if -d/--detailed, -d -d or -d -d -d is used: 1944 */ 1945 static int add_default_events(void) 1946 { 1947 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 1948 struct parse_events_error err; 1949 struct evlist *evlist = evlist__new(); 1950 struct evsel *evsel; 1951 int ret = 0; 1952 1953 if (!evlist) 1954 return -ENOMEM; 1955 1956 parse_events_error__init(&err); 1957 1958 /* Set attrs if no event is selected and !null_run: */ 1959 if (stat_config.null_run) 1960 goto out; 1961 1962 if (transaction_run) { 1963 /* Handle -T as -M transaction. Once platform specific metrics 1964 * support has been added to the json files, all architectures 1965 * will use this approach. To determine transaction support 1966 * on an architecture test for such a metric name. 1967 */ 1968 if (!metricgroup__has_metric_or_groups(pmu, "transaction")) { 1969 pr_err("Missing transaction metrics\n"); 1970 ret = -1; 1971 goto out; 1972 } 1973 ret = metricgroup__parse_groups(evlist, pmu, "transaction", 1974 stat_config.metric_no_group, 1975 stat_config.metric_no_merge, 1976 stat_config.metric_no_threshold, 1977 stat_config.user_requested_cpu_list, 1978 stat_config.system_wide, 1979 stat_config.hardware_aware_grouping); 1980 goto out; 1981 } 1982 1983 if (smi_cost) { 1984 int smi; 1985 1986 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 1987 pr_err("freeze_on_smi is not supported.\n"); 1988 ret = -1; 1989 goto out; 1990 } 1991 1992 if (!smi) { 1993 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 1994 pr_err("Failed to set freeze_on_smi.\n"); 1995 ret = -1; 1996 goto out; 1997 } 1998 smi_reset = true; 1999 } 2000 2001 if (!metricgroup__has_metric_or_groups(pmu, "smi")) { 2002 pr_err("Missing smi metrics\n"); 2003 ret = -1; 2004 goto out; 2005 } 2006 2007 if (!force_metric_only) 2008 stat_config.metric_only = true; 2009 2010 ret = metricgroup__parse_groups(evlist, pmu, "smi", 2011 stat_config.metric_no_group, 2012 stat_config.metric_no_merge, 2013 stat_config.metric_no_threshold, 2014 stat_config.user_requested_cpu_list, 2015 stat_config.system_wide, 2016 stat_config.hardware_aware_grouping); 2017 goto out; 2018 } 2019 2020 if (topdown_run) { 2021 unsigned int max_level = metricgroups__topdown_max_level(); 2022 char str[] = "TopdownL1"; 2023 2024 if (!force_metric_only) 2025 stat_config.metric_only = true; 2026 2027 if (!max_level) { 2028 pr_err("Topdown requested but the topdown metric groups aren't present.\n" 2029 "(See perf list the metric groups have names like TopdownL1)\n"); 2030 ret = -1; 2031 goto out; 2032 } 2033 if (stat_config.topdown_level > max_level) { 2034 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level); 2035 ret = -1; 2036 goto out; 2037 } else if (!stat_config.topdown_level) { 2038 stat_config.topdown_level = 1; 2039 } 2040 if (!stat_config.interval && !stat_config.metric_only) { 2041 fprintf(stat_config.output, 2042 "Topdown accuracy may decrease when measuring long periods.\n" 2043 "Please print the result regularly, e.g. -I1000\n"); 2044 } 2045 str[8] = stat_config.topdown_level + '0'; 2046 if (metricgroup__parse_groups(evlist, 2047 pmu, str, 2048 /*metric_no_group=*/false, 2049 /*metric_no_merge=*/false, 2050 /*metric_no_threshold=*/true, 2051 stat_config.user_requested_cpu_list, 2052 stat_config.system_wide, 2053 stat_config.hardware_aware_grouping) < 0) { 2054 ret = -1; 2055 goto out; 2056 } 2057 } 2058 2059 if (!stat_config.topdown_level) 2060 stat_config.topdown_level = 1; 2061 2062 if (!evlist->core.nr_entries && !evsel_list->core.nr_entries) { 2063 /* 2064 * Add Default metrics. To minimize multiplexing, don't request 2065 * threshold computation, but it will be computed if the events 2066 * are present. 2067 */ 2068 const char *default_metricgroup_names[] = { 2069 "Default", "Default2", "Default3", "Default4", 2070 }; 2071 2072 for (size_t i = 0; i < ARRAY_SIZE(default_metricgroup_names); i++) { 2073 struct evlist *metric_evlist; 2074 2075 if (!metricgroup__has_metric_or_groups(pmu, default_metricgroup_names[i])) 2076 continue; 2077 2078 if ((int)i > detailed_run) 2079 break; 2080 2081 metric_evlist = evlist__new(); 2082 if (!metric_evlist) { 2083 ret = -ENOMEM; 2084 break; 2085 } 2086 if (metricgroup__parse_groups(metric_evlist, pmu, default_metricgroup_names[i], 2087 /*metric_no_group=*/false, 2088 /*metric_no_merge=*/false, 2089 /*metric_no_threshold=*/true, 2090 stat_config.user_requested_cpu_list, 2091 stat_config.system_wide, 2092 stat_config.hardware_aware_grouping) < 0) { 2093 evlist__delete(metric_evlist); 2094 ret = -1; 2095 break; 2096 } 2097 2098 evlist__for_each_entry(metric_evlist, evsel) 2099 evsel->default_metricgroup = true; 2100 2101 evlist__splice_list_tail(evlist, &metric_evlist->core.entries); 2102 metricgroup__copy_metric_events(evlist, /*cgrp=*/NULL, 2103 &evlist->metric_events, 2104 &metric_evlist->metric_events); 2105 evlist__delete(metric_evlist); 2106 } 2107 list_sort(/*priv=*/NULL, &evlist->core.entries, default_evlist_evsel_cmp); 2108 2109 } 2110 out: 2111 if (!ret) { 2112 evlist__for_each_entry(evlist, evsel) { 2113 /* 2114 * Make at least one event non-skippable so fatal errors are visible. 2115 * 'cycles' always used to be default and non-skippable, so use that. 2116 */ 2117 if (!evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) 2118 evsel->skippable = true; 2119 } 2120 } 2121 parse_events_error__exit(&err); 2122 evlist__splice_list_tail(evsel_list, &evlist->core.entries); 2123 metricgroup__copy_metric_events(evsel_list, /*cgrp=*/NULL, 2124 &evsel_list->metric_events, 2125 &evlist->metric_events); 2126 evlist__delete(evlist); 2127 return ret; 2128 } 2129 2130 static const char * const stat_record_usage[] = { 2131 "perf stat record [<options>]", 2132 NULL, 2133 }; 2134 2135 static void init_features(struct perf_session *session) 2136 { 2137 int feat; 2138 2139 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 2140 perf_header__set_feat(&session->header, feat); 2141 2142 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 2143 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 2144 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 2145 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 2146 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 2147 } 2148 2149 static int __cmd_record(const struct option stat_options[], struct opt_aggr_mode *opt_mode, 2150 int argc, const char **argv) 2151 { 2152 struct perf_session *session; 2153 struct perf_data *data = &perf_stat.data; 2154 2155 argc = parse_options(argc, argv, stat_options, stat_record_usage, 2156 PARSE_OPT_STOP_AT_NON_OPTION); 2157 stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(opt_mode); 2158 2159 if (output_name) 2160 data->path = output_name; 2161 2162 if (stat_config.run_count != 1 || forever) { 2163 pr_err("Cannot use -r option with perf stat record.\n"); 2164 return -1; 2165 } 2166 2167 session = perf_session__new(data, NULL); 2168 if (IS_ERR(session)) { 2169 pr_err("Perf session creation failed\n"); 2170 return PTR_ERR(session); 2171 } 2172 2173 init_features(session); 2174 2175 session->evlist = evsel_list; 2176 perf_stat.session = session; 2177 perf_stat.record = true; 2178 return argc; 2179 } 2180 2181 static int process_stat_round_event(const struct perf_tool *tool __maybe_unused, 2182 struct perf_session *session, 2183 union perf_event *event) 2184 { 2185 struct perf_record_stat_round *stat_round = &event->stat_round; 2186 struct timespec tsh, *ts = NULL; 2187 struct perf_env *env = perf_session__env(session); 2188 const char **argv = env->cmdline_argv; 2189 int argc = env->nr_cmdline; 2190 2191 process_counters(); 2192 2193 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 2194 update_stats(stat_config.walltime_nsecs_stats, stat_round->time); 2195 2196 if (stat_config.interval && stat_round->time) { 2197 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 2198 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 2199 ts = &tsh; 2200 } 2201 2202 print_counters(ts, argc, argv); 2203 return 0; 2204 } 2205 2206 static 2207 int process_stat_config_event(const struct perf_tool *tool, 2208 struct perf_session *session, 2209 union perf_event *event) 2210 { 2211 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2212 2213 perf_event__read_stat_config(&stat_config, &event->stat_config); 2214 2215 if (perf_cpu_map__is_empty(st->cpus)) { 2216 if (st->aggr_mode != AGGR_UNSET) 2217 pr_warning("warning: processing task data, aggregation mode not set\n"); 2218 } else if (st->aggr_mode != AGGR_UNSET) { 2219 stat_config.aggr_mode = st->aggr_mode; 2220 } 2221 2222 if (perf_stat.data.is_pipe) 2223 perf_stat_init_aggr_mode(); 2224 else 2225 perf_stat_init_aggr_mode_file(st); 2226 2227 if (stat_config.aggr_map) { 2228 int nr_aggr = stat_config.aggr_map->nr; 2229 2230 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) { 2231 pr_err("cannot allocate aggr counts\n"); 2232 return -1; 2233 } 2234 } 2235 return 0; 2236 } 2237 2238 static int set_maps(struct perf_stat *st) 2239 { 2240 if (!st->cpus || !st->threads) 2241 return 0; 2242 2243 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 2244 return -EINVAL; 2245 2246 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 2247 2248 if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true)) 2249 return -ENOMEM; 2250 2251 st->maps_allocated = true; 2252 return 0; 2253 } 2254 2255 static 2256 int process_thread_map_event(const struct perf_tool *tool, 2257 struct perf_session *session __maybe_unused, 2258 union perf_event *event) 2259 { 2260 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2261 2262 if (st->threads) { 2263 pr_warning("Extra thread map event, ignoring.\n"); 2264 return 0; 2265 } 2266 2267 st->threads = thread_map__new_event(&event->thread_map); 2268 if (!st->threads) 2269 return -ENOMEM; 2270 2271 return set_maps(st); 2272 } 2273 2274 static 2275 int process_cpu_map_event(const struct perf_tool *tool, 2276 struct perf_session *session __maybe_unused, 2277 union perf_event *event) 2278 { 2279 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2280 struct perf_cpu_map *cpus; 2281 2282 if (st->cpus) { 2283 pr_warning("Extra cpu map event, ignoring.\n"); 2284 return 0; 2285 } 2286 2287 cpus = cpu_map__new_data(&event->cpu_map.data); 2288 if (!cpus) 2289 return -ENOMEM; 2290 2291 st->cpus = cpus; 2292 return set_maps(st); 2293 } 2294 2295 static const char * const stat_report_usage[] = { 2296 "perf stat report [<options>]", 2297 NULL, 2298 }; 2299 2300 static struct perf_stat perf_stat = { 2301 .aggr_mode = AGGR_UNSET, 2302 .aggr_level = 0, 2303 }; 2304 2305 static int __cmd_report(int argc, const char **argv) 2306 { 2307 struct perf_session *session; 2308 const struct option options[] = { 2309 OPT_STRING('i', "input", &input_name, "file", "input file name"), 2310 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 2311 "aggregate counts per processor socket", AGGR_SOCKET), 2312 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 2313 "aggregate counts per processor die", AGGR_DIE), 2314 OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode, 2315 "aggregate counts perf processor cluster", AGGR_CLUSTER), 2316 OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level, 2317 "cache level", 2318 "aggregate count at this cache level (Default: LLC)", 2319 parse_cache_level), 2320 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 2321 "aggregate counts per physical processor core", AGGR_CORE), 2322 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, 2323 "aggregate counts per numa node", AGGR_NODE), 2324 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 2325 "disable CPU count aggregation", AGGR_NONE), 2326 OPT_END() 2327 }; 2328 struct stat st; 2329 int ret; 2330 2331 argc = parse_options(argc, argv, options, stat_report_usage, 0); 2332 2333 if (!input_name || !strlen(input_name)) { 2334 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 2335 input_name = "-"; 2336 else 2337 input_name = "perf.data"; 2338 } 2339 2340 perf_stat.data.path = input_name; 2341 perf_stat.data.mode = PERF_DATA_MODE_READ; 2342 2343 perf_tool__init(&perf_stat.tool, /*ordered_events=*/false); 2344 perf_stat.tool.attr = perf_event__process_attr; 2345 perf_stat.tool.event_update = perf_event__process_event_update; 2346 perf_stat.tool.thread_map = process_thread_map_event; 2347 perf_stat.tool.cpu_map = process_cpu_map_event; 2348 perf_stat.tool.stat_config = process_stat_config_event; 2349 perf_stat.tool.stat = perf_event__process_stat_event; 2350 perf_stat.tool.stat_round = process_stat_round_event; 2351 2352 session = perf_session__new(&perf_stat.data, &perf_stat.tool); 2353 if (IS_ERR(session)) 2354 return PTR_ERR(session); 2355 2356 perf_stat.session = session; 2357 stat_config.output = stderr; 2358 evlist__delete(evsel_list); 2359 evsel_list = session->evlist; 2360 2361 ret = perf_session__process_events(session); 2362 if (ret) 2363 return ret; 2364 2365 perf_session__delete(session); 2366 return 0; 2367 } 2368 2369 static void setup_system_wide(int forks) 2370 { 2371 /* 2372 * Make system wide (-a) the default target if 2373 * no target was specified and one of following 2374 * conditions is met: 2375 * 2376 * - there's no workload specified 2377 * - there is workload specified but all requested 2378 * events are system wide events 2379 */ 2380 if (!target__none(&target)) 2381 return; 2382 2383 if (!forks) 2384 target.system_wide = true; 2385 else { 2386 struct evsel *counter; 2387 2388 evlist__for_each_entry(evsel_list, counter) { 2389 if (!counter->core.requires_cpu && 2390 !evsel__name_is(counter, "duration_time")) { 2391 return; 2392 } 2393 } 2394 2395 if (evsel_list->core.nr_entries) 2396 target.system_wide = true; 2397 } 2398 } 2399 2400 #ifdef HAVE_ARCH_X86_64_SUPPORT 2401 static int parse_tpebs_mode(const struct option *opt, const char *str, 2402 int unset __maybe_unused) 2403 { 2404 enum tpebs_mode *mode = opt->value; 2405 2406 if (!strcasecmp("mean", str)) { 2407 *mode = TPEBS_MODE__MEAN; 2408 return 0; 2409 } 2410 if (!strcasecmp("min", str)) { 2411 *mode = TPEBS_MODE__MIN; 2412 return 0; 2413 } 2414 if (!strcasecmp("max", str)) { 2415 *mode = TPEBS_MODE__MAX; 2416 return 0; 2417 } 2418 if (!strcasecmp("last", str)) { 2419 *mode = TPEBS_MODE__LAST; 2420 return 0; 2421 } 2422 return -1; 2423 } 2424 #endif // HAVE_ARCH_X86_64_SUPPORT 2425 2426 int cmd_stat(int argc, const char **argv) 2427 { 2428 struct opt_aggr_mode opt_mode = {}; 2429 bool affinity = true, affinity_set = false; 2430 struct option stat_options[] = { 2431 OPT_BOOLEAN('T', "transaction", &transaction_run, 2432 "hardware transaction statistics"), 2433 OPT_CALLBACK('e', "event", &parse_events_option_args, "event", 2434 "event selector. use 'perf list' to list available events", 2435 parse_events_option), 2436 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 2437 "event filter", parse_filter), 2438 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 2439 "child tasks do not inherit counters"), 2440 OPT_STRING('p', "pid", &target.pid, "pid", 2441 "stat events on existing process id"), 2442 OPT_STRING('t', "tid", &target.tid, "tid", 2443 "stat events on existing thread id"), 2444 #ifdef HAVE_BPF_SKEL 2445 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id", 2446 "stat events on existing bpf program id"), 2447 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf, 2448 "use bpf program to count events"), 2449 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path", 2450 "path to perf_event_attr map"), 2451 #endif 2452 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 2453 "system-wide collection from all CPUs"), 2454 OPT_BOOLEAN(0, "scale", &stat_config.scale, 2455 "Use --no-scale to disable counter scaling for multiplexing"), 2456 OPT_INCR('v', "verbose", &verbose, 2457 "be more verbose (show counter open errors, etc)"), 2458 OPT_INTEGER('r', "repeat", &stat_config.run_count, 2459 "repeat command and print average + stddev (max: 100, forever: 0)"), 2460 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 2461 "display details about each run (only with -r option)"), 2462 OPT_BOOLEAN('n', "null", &stat_config.null_run, 2463 "null run - dont start any counters"), 2464 OPT_INCR('d', "detailed", &detailed_run, 2465 "detailed run - start a lot of events"), 2466 OPT_BOOLEAN('S', "sync", &sync_run, 2467 "call sync() before starting a run"), 2468 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 2469 "print large numbers with thousands\' separators", 2470 stat__set_big_num), 2471 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 2472 "list of cpus to monitor in system-wide"), 2473 OPT_BOOLEAN('A', "no-aggr", &opt_mode.no_aggr, 2474 "disable aggregation across CPUs or PMUs"), 2475 OPT_BOOLEAN(0, "no-merge", &opt_mode.no_aggr, 2476 "disable aggregation the same as -A or -no-aggr"), 2477 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge, 2478 "Merge identical named hybrid events"), 2479 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 2480 "print counts with custom separator"), 2481 OPT_BOOLEAN('j', "json-output", &stat_config.json_output, 2482 "print counts in JSON format"), 2483 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 2484 "monitor event in cgroup name only", parse_stat_cgroups), 2485 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name", 2486 "expand events for each cgroup"), 2487 OPT_STRING('o', "output", &output_name, "file", "output file name"), 2488 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 2489 OPT_INTEGER(0, "log-fd", &output_fd, 2490 "log output to fd, instead of stderr"), 2491 OPT_STRING(0, "pre", &pre_cmd, "command", 2492 "command to run prior to the measured command"), 2493 OPT_STRING(0, "post", &post_cmd, "command", 2494 "command to run after to the measured command"), 2495 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 2496 "print counts at regular interval in ms " 2497 "(overhead is possible for values <= 100ms)"), 2498 OPT_INTEGER(0, "interval-count", &stat_config.times, 2499 "print counts for fixed number of times"), 2500 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 2501 "clear screen in between new interval"), 2502 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 2503 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 2504 OPT_BOOLEAN(0, "per-socket", &opt_mode.socket, 2505 "aggregate counts per processor socket"), 2506 OPT_BOOLEAN(0, "per-die", &opt_mode.die, "aggregate counts per processor die"), 2507 OPT_BOOLEAN(0, "per-cluster", &opt_mode.cluster, 2508 "aggregate counts per processor cluster"), 2509 OPT_CALLBACK_OPTARG(0, "per-cache", &opt_mode, &stat_config.aggr_level, 2510 "cache level", "aggregate count at this cache level (Default: LLC)", 2511 parse_cache_level), 2512 OPT_BOOLEAN(0, "per-core", &opt_mode.core, 2513 "aggregate counts per physical processor core"), 2514 OPT_BOOLEAN(0, "per-thread", &opt_mode.thread, "aggregate counts per thread"), 2515 OPT_BOOLEAN(0, "per-node", &opt_mode.node, "aggregate counts per numa node"), 2516 OPT_INTEGER('D', "delay", &target.initial_delay, 2517 "ms to wait before starting measurement after program start (-1: start with events disabled)"), 2518 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 2519 "Only print computed metrics. No raw values", enable_metric_only), 2520 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group, 2521 "don't group metric events, impacts multiplexing"), 2522 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge, 2523 "don't try to share events between metrics in a group"), 2524 OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold, 2525 "disable adding events for the metric threshold calculation"), 2526 OPT_BOOLEAN(0, "topdown", &topdown_run, 2527 "measure top-down statistics"), 2528 #ifdef HAVE_ARCH_X86_64_SUPPORT 2529 OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording, 2530 "enable recording for tpebs when retire_latency required"), 2531 OPT_CALLBACK(0, "tpebs-mode", &tpebs_mode, "tpebs-mode", 2532 "Mode of TPEBS recording: mean, min or max", 2533 parse_tpebs_mode), 2534 #endif 2535 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level, 2536 "Set the metrics level for the top-down statistics (0: max level)"), 2537 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 2538 "measure SMI cost"), 2539 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 2540 "monitor specified metrics or metric groups (separated by ,)", 2541 append_metric_groups), 2542 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, 2543 "Configure all used events to run in kernel space.", 2544 PARSE_OPT_EXCLUSIVE), 2545 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, 2546 "Configure all used events to run in user space.", 2547 PARSE_OPT_EXCLUSIVE), 2548 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread, 2549 "Use with 'percore' event qualifier to show the event " 2550 "counts of one hardware thread by sum up total hardware " 2551 "threads of same physical core"), 2552 OPT_BOOLEAN(0, "summary", &stat_config.summary, 2553 "print summary for interval mode"), 2554 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary, 2555 "don't print 'summary' for CSV summary output"), 2556 OPT_BOOLEAN(0, "quiet", &quiet, 2557 "don't print any output, messages or warnings (useful with record)"), 2558 OPT_BOOLEAN_SET(0, "affinity", &affinity, &affinity_set, 2559 "enable (default) or disable affinity optimizations to reduce IPIs"), 2560 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type", 2561 "Only enable events on applying cpu with this type " 2562 "for hybrid platform (e.g. core or atom)", 2563 parse_cputype), 2564 #ifdef HAVE_LIBPFM 2565 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event", 2566 "libpfm4 event selector. use 'perf list' to list available events", 2567 parse_libpfm_events_option), 2568 #endif 2569 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]", 2570 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n" 2571 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n" 2572 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.", 2573 parse_control_option), 2574 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default", 2575 "measure I/O performance metrics provided by arch/platform", 2576 iostat_parse), 2577 OPT_END() 2578 }; 2579 const char * const stat_usage[] = { 2580 "perf stat [<options>] [<command>]", 2581 NULL 2582 }; 2583 int status = -EINVAL, run_idx, err; 2584 const char *mode; 2585 FILE *output = stderr; 2586 unsigned int interval, timeout; 2587 const char * const stat_subcommands[] = { "record", "report" }; 2588 char errbuf[BUFSIZ]; 2589 struct evsel *counter; 2590 2591 setlocale(LC_ALL, ""); 2592 2593 evsel_list = evlist__new(); 2594 if (evsel_list == NULL) 2595 return -ENOMEM; 2596 2597 parse_events__shrink_config_terms(); 2598 2599 /* String-parsing callback-based options would segfault when negated */ 2600 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 2601 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 2602 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 2603 2604 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 2605 (const char **) stat_usage, 2606 PARSE_OPT_STOP_AT_NON_OPTION); 2607 2608 stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(&opt_mode); 2609 2610 if (stat_config.csv_sep) { 2611 stat_config.csv_output = true; 2612 if (!strcmp(stat_config.csv_sep, "\\t")) 2613 stat_config.csv_sep = "\t"; 2614 } else 2615 stat_config.csv_sep = DEFAULT_SEPARATOR; 2616 2617 if (affinity_set) 2618 evsel_list->no_affinity = !affinity; 2619 2620 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2621 argc = __cmd_record(stat_options, &opt_mode, argc, argv); 2622 if (argc < 0) 2623 return -1; 2624 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0])) 2625 return __cmd_report(argc, argv); 2626 2627 interval = stat_config.interval; 2628 timeout = stat_config.timeout; 2629 2630 /* 2631 * For record command the -o is already taken care of. 2632 */ 2633 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 2634 output = NULL; 2635 2636 if (output_name && output_fd) { 2637 fprintf(stderr, "cannot use both --output and --log-fd\n"); 2638 parse_options_usage(stat_usage, stat_options, "o", 1); 2639 parse_options_usage(NULL, stat_options, "log-fd", 0); 2640 goto out; 2641 } 2642 2643 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 2644 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 2645 goto out; 2646 } 2647 2648 if (stat_config.metric_only && stat_config.run_count > 1) { 2649 fprintf(stderr, "--metric-only is not supported with -r\n"); 2650 goto out; 2651 } 2652 2653 if (stat_config.csv_output || (stat_config.metric_only && stat_config.json_output)) { 2654 /* 2655 * Current CSV and metric-only JSON output doesn't display the 2656 * metric threshold so don't compute it. 2657 */ 2658 stat_config.metric_no_threshold = true; 2659 } 2660 2661 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 2662 fprintf(stderr, "--table is only supported with -r\n"); 2663 parse_options_usage(stat_usage, stat_options, "r", 1); 2664 parse_options_usage(NULL, stat_options, "table", 0); 2665 goto out; 2666 } 2667 2668 if (output_fd < 0) { 2669 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 2670 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 2671 goto out; 2672 } 2673 2674 if (!output && !quiet) { 2675 struct timespec tm; 2676 mode = append_file ? "a" : "w"; 2677 2678 output = fopen(output_name, mode); 2679 if (!output) { 2680 perror("failed to create output file"); 2681 return -1; 2682 } 2683 if (!stat_config.json_output) { 2684 clock_gettime(CLOCK_REALTIME, &tm); 2685 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 2686 } 2687 } else if (output_fd > 0) { 2688 mode = append_file ? "a" : "w"; 2689 output = fdopen(output_fd, mode); 2690 if (!output) { 2691 perror("Failed opening logfd"); 2692 return -errno; 2693 } 2694 } 2695 2696 if (stat_config.interval_clear && !isatty(fileno(output))) { 2697 fprintf(stderr, "--interval-clear does not work with output\n"); 2698 parse_options_usage(stat_usage, stat_options, "o", 1); 2699 parse_options_usage(NULL, stat_options, "log-fd", 0); 2700 parse_options_usage(NULL, stat_options, "interval-clear", 0); 2701 return -1; 2702 } 2703 2704 stat_config.output = output; 2705 2706 /* 2707 * let the spreadsheet do the pretty-printing 2708 */ 2709 if (stat_config.csv_output) { 2710 /* User explicitly passed -B? */ 2711 if (big_num_opt == 1) { 2712 fprintf(stderr, "-B option not supported with -x\n"); 2713 parse_options_usage(stat_usage, stat_options, "B", 1); 2714 parse_options_usage(NULL, stat_options, "x", 1); 2715 goto out; 2716 } else /* Nope, so disable big number formatting */ 2717 stat_config.big_num = false; 2718 } else if (big_num_opt == 0) /* User passed --no-big-num */ 2719 stat_config.big_num = false; 2720 2721 target.inherit = !stat_config.no_inherit; 2722 err = target__validate(&target); 2723 if (err) { 2724 target__strerror(&target, err, errbuf, BUFSIZ); 2725 pr_warning("%s\n", errbuf); 2726 } 2727 2728 setup_system_wide(argc); 2729 2730 /* 2731 * Display user/system times only for single 2732 * run and when there's specified tracee. 2733 */ 2734 if ((stat_config.run_count == 1) && target__none(&target)) 2735 stat_config.ru_display = true; 2736 2737 if (stat_config.run_count < 0) { 2738 pr_err("Run count must be a positive number\n"); 2739 parse_options_usage(stat_usage, stat_options, "r", 1); 2740 goto out; 2741 } else if (stat_config.run_count == 0) { 2742 forever = true; 2743 stat_config.run_count = 1; 2744 } 2745 2746 if (stat_config.walltime_run_table) { 2747 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 2748 if (!stat_config.walltime_run) { 2749 pr_err("failed to setup -r option"); 2750 goto out; 2751 } 2752 } 2753 2754 if ((stat_config.aggr_mode == AGGR_THREAD) && 2755 !target__has_task(&target)) { 2756 if (!target.system_wide || target.cpu_list) { 2757 fprintf(stderr, "The --per-thread option is only " 2758 "available when monitoring via -p -t -a " 2759 "options or only --per-thread.\n"); 2760 parse_options_usage(NULL, stat_options, "p", 1); 2761 parse_options_usage(NULL, stat_options, "t", 1); 2762 goto out; 2763 } 2764 } 2765 2766 /* 2767 * no_aggr, cgroup are for system-wide only 2768 * --per-thread is aggregated per thread, we dont mix it with cpu mode 2769 */ 2770 if (((stat_config.aggr_mode != AGGR_GLOBAL && 2771 stat_config.aggr_mode != AGGR_THREAD) || 2772 (nr_cgroups || stat_config.cgroup_list)) && 2773 !target__has_cpu(&target)) { 2774 fprintf(stderr, "both cgroup and no-aggregation " 2775 "modes only available in system-wide mode\n"); 2776 2777 parse_options_usage(stat_usage, stat_options, "G", 1); 2778 parse_options_usage(NULL, stat_options, "A", 1); 2779 parse_options_usage(NULL, stat_options, "a", 1); 2780 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2781 goto out; 2782 } 2783 2784 if (stat_config.iostat_run) { 2785 status = iostat_prepare(evsel_list, &stat_config); 2786 if (status) 2787 goto out; 2788 if (iostat_mode == IOSTAT_LIST) { 2789 iostat_list(evsel_list, &stat_config); 2790 goto out; 2791 } else if (verbose > 0) 2792 iostat_list(evsel_list, &stat_config); 2793 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target)) 2794 target.system_wide = true; 2795 } 2796 2797 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 2798 target.per_thread = true; 2799 2800 stat_config.system_wide = target.system_wide; 2801 if (target.cpu_list) { 2802 stat_config.user_requested_cpu_list = strdup(target.cpu_list); 2803 if (!stat_config.user_requested_cpu_list) { 2804 status = -ENOMEM; 2805 goto out; 2806 } 2807 } 2808 2809 /* 2810 * Metric parsing needs to be delayed as metrics may optimize events 2811 * knowing the target is system-wide. 2812 */ 2813 if (metrics) { 2814 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 2815 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics, 2816 stat_config.metric_no_group, 2817 stat_config.metric_no_merge, 2818 stat_config.metric_no_threshold, 2819 stat_config.user_requested_cpu_list, 2820 stat_config.system_wide, 2821 stat_config.hardware_aware_grouping); 2822 2823 zfree(&metrics); 2824 if (ret) { 2825 status = ret; 2826 goto out; 2827 } 2828 } 2829 2830 if (add_default_events()) 2831 goto out; 2832 2833 if (stat_config.cgroup_list) { 2834 if (nr_cgroups > 0) { 2835 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 2836 parse_options_usage(stat_usage, stat_options, "G", 1); 2837 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2838 goto out; 2839 } 2840 2841 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, true) < 0) { 2842 parse_options_usage(stat_usage, stat_options, 2843 "for-each-cgroup", 0); 2844 goto out; 2845 } 2846 } 2847 #ifdef HAVE_BPF_SKEL 2848 if (target.use_bpf && nr_cgroups && 2849 (evsel_list->core.nr_entries / nr_cgroups) > BPERF_CGROUP__MAX_EVENTS) { 2850 pr_warning("Disabling BPF counters due to more events (%d) than the max (%d)\n", 2851 evsel_list->core.nr_entries / nr_cgroups, BPERF_CGROUP__MAX_EVENTS); 2852 target.use_bpf = false; 2853 } 2854 #endif // HAVE_BPF_SKEL 2855 evlist__warn_user_requested_cpus(evsel_list, target.cpu_list); 2856 2857 evlist__for_each_entry(evsel_list, counter) { 2858 /* 2859 * Setup BPF counters to require CPUs as any(-1) isn't 2860 * supported. evlist__create_maps below will propagate this 2861 * information to the evsels. Note, evsel__is_bperf isn't yet 2862 * set up, and this change must happen early, so directly use 2863 * the bpf_counter variable and target information. 2864 */ 2865 if ((counter->bpf_counter || target.use_bpf) && !target__has_cpu(&target)) 2866 counter->core.requires_cpu = true; 2867 } 2868 2869 if (evlist__create_maps(evsel_list, &target) < 0) { 2870 if (target__has_task(&target)) { 2871 pr_err("Problems finding threads of monitor\n"); 2872 parse_options_usage(stat_usage, stat_options, "p", 1); 2873 parse_options_usage(NULL, stat_options, "t", 1); 2874 } else if (target__has_cpu(&target)) { 2875 perror("failed to parse CPUs map"); 2876 parse_options_usage(stat_usage, stat_options, "C", 1); 2877 parse_options_usage(NULL, stat_options, "a", 1); 2878 } 2879 goto out; 2880 } 2881 2882 evlist__check_cpu_maps(evsel_list); 2883 2884 /* 2885 * Initialize thread_map with comm names, 2886 * so we could print it out on output. 2887 */ 2888 if (stat_config.aggr_mode == AGGR_THREAD) { 2889 thread_map__read_comms(evsel_list->core.threads); 2890 } 2891 2892 if (stat_config.aggr_mode == AGGR_NODE) 2893 cpu__setup_cpunode_map(); 2894 2895 if (stat_config.times && interval) 2896 interval_count = true; 2897 else if (stat_config.times && !interval) { 2898 pr_err("interval-count option should be used together with " 2899 "interval-print.\n"); 2900 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 2901 parse_options_usage(stat_usage, stat_options, "I", 1); 2902 goto out; 2903 } 2904 2905 if (timeout && timeout < 100) { 2906 if (timeout < 10) { 2907 pr_err("timeout must be >= 10ms.\n"); 2908 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2909 goto out; 2910 } else 2911 pr_warning("timeout < 100ms. " 2912 "The overhead percentage could be high in some cases. " 2913 "Please proceed with caution.\n"); 2914 } 2915 if (timeout && interval) { 2916 pr_err("timeout option is not supported with interval-print.\n"); 2917 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2918 parse_options_usage(stat_usage, stat_options, "I", 1); 2919 goto out; 2920 } 2921 2922 if (perf_stat_init_aggr_mode()) 2923 goto out; 2924 2925 if (evlist__alloc_stats(&stat_config, evsel_list, interval)) 2926 goto out; 2927 2928 /* 2929 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 2930 * while avoiding that older tools show confusing messages. 2931 * 2932 * However for pipe sessions we need to keep it zero, 2933 * because script's perf_evsel__check_attr is triggered 2934 * by attr->sample_type != 0, and we can't run it on 2935 * stat sessions. 2936 */ 2937 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 2938 2939 /* 2940 * We dont want to block the signals - that would cause 2941 * child tasks to inherit that and Ctrl-C would not work. 2942 * What we want is for Ctrl-C to work in the exec()-ed 2943 * task, but being ignored by perf stat itself: 2944 */ 2945 atexit(sig_atexit); 2946 if (!forever) 2947 signal(SIGINT, skip_signal); 2948 signal(SIGCHLD, skip_signal); 2949 signal(SIGALRM, skip_signal); 2950 signal(SIGABRT, skip_signal); 2951 2952 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack)) 2953 goto out; 2954 2955 /* Enable ignoring missing threads when -p option is defined. */ 2956 evlist__first(evsel_list)->ignore_missing_thread = target.pid; 2957 status = 0; 2958 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 2959 if (stat_config.run_count != 1 && verbose > 0) 2960 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 2961 run_idx + 1); 2962 2963 if (run_idx != 0) 2964 evlist__reset_prev_raw_counts(evsel_list); 2965 2966 status = run_perf_stat(argc, argv, run_idx); 2967 if (status < 0) 2968 break; 2969 2970 if (forever && !interval) { 2971 print_counters(NULL, argc, argv); 2972 perf_stat__reset_stats(); 2973 } 2974 } 2975 2976 if (!forever && status != -1 && (!interval || stat_config.summary)) { 2977 if (stat_config.run_count > 1) 2978 evlist__copy_res_stats(&stat_config, evsel_list); 2979 print_counters(NULL, argc, argv); 2980 } 2981 2982 evlist__finalize_ctlfd(evsel_list); 2983 2984 if (STAT_RECORD) { 2985 /* 2986 * We synthesize the kernel mmap record just so that older tools 2987 * don't emit warnings about not being able to resolve symbols 2988 * due to /proc/sys/kernel/kptr_restrict settings and instead provide 2989 * a saner message about no samples being in the perf.data file. 2990 * 2991 * This also serves to suppress a warning about f_header.data.size == 0 2992 * in header.c at the moment 'perf stat record' gets introduced, which 2993 * is not really needed once we start adding the stat specific PERF_RECORD_ 2994 * records, but the need to suppress the kptr_restrict messages in older 2995 * tools remain -acme 2996 */ 2997 int fd = perf_data__fd(&perf_stat.data); 2998 2999 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 3000 process_synthesized_event, 3001 &perf_stat.session->machines.host); 3002 if (err) { 3003 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 3004 "older tools may produce warnings about this file\n."); 3005 } 3006 3007 if (!interval) { 3008 if (WRITE_STAT_ROUND_EVENT(stat_config.walltime_nsecs_stats->max, FINAL)) 3009 pr_err("failed to write stat round event\n"); 3010 } 3011 3012 if (!perf_stat.data.is_pipe) { 3013 perf_stat.session->header.data_size += perf_stat.bytes_written; 3014 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 3015 } 3016 3017 evlist__close(evsel_list); 3018 perf_session__delete(perf_stat.session); 3019 } 3020 3021 perf_stat__exit_aggr_mode(); 3022 evlist__free_stats(evsel_list); 3023 out: 3024 if (stat_config.iostat_run) 3025 iostat_release(evsel_list); 3026 3027 zfree(&stat_config.walltime_run); 3028 zfree(&stat_config.user_requested_cpu_list); 3029 3030 if (smi_cost && smi_reset) 3031 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 3032 3033 evlist__delete(evsel_list); 3034 3035 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close); 3036 3037 /* Only the low byte of status becomes the exit code. */ 3038 return abs(status); 3039 } 3040