1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * builtin-stat.c 4 * 5 * Builtin stat command: Give a precise performance counters summary 6 * overview about any workload, CPU or specific PID. 7 * 8 * Sample output: 9 10 $ perf stat ./hackbench 10 11 12 Time: 0.118 13 14 Performance counter stats for './hackbench 10': 15 16 1708.761321 task-clock # 11.037 CPUs utilized 17 41,190 context-switches # 0.024 M/sec 18 6,735 CPU-migrations # 0.004 M/sec 19 17,318 page-faults # 0.010 M/sec 20 5,205,202,243 cycles # 3.046 GHz 21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle 22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle 23 2,603,501,247 instructions # 0.50 insns per cycle 24 # 1.48 stalled cycles per insn 25 484,357,498 branches # 283.455 M/sec 26 6,388,934 branch-misses # 1.32% of all branches 27 28 0.154822978 seconds time elapsed 29 30 * 31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 32 * 33 * Improvements and fixes by: 34 * 35 * Arjan van de Ven <arjan@linux.intel.com> 36 * Yanmin Zhang <yanmin.zhang@intel.com> 37 * Wu Fengguang <fengguang.wu@intel.com> 38 * Mike Galbraith <efault@gmx.de> 39 * Paul Mackerras <paulus@samba.org> 40 * Jaswinder Singh Rajput <jaswinder@kernel.org> 41 */ 42 43 #include "builtin.h" 44 #include "util/cgroup.h" 45 #include <subcmd/parse-options.h> 46 #include "util/parse-events.h" 47 #include "util/pmus.h" 48 #include "util/pmu.h" 49 #include "util/tool_pmu.h" 50 #include "util/event.h" 51 #include "util/evlist.h" 52 #include "util/evsel.h" 53 #include "util/debug.h" 54 #include "util/color.h" 55 #include "util/stat.h" 56 #include "util/header.h" 57 #include "util/cpumap.h" 58 #include "util/thread_map.h" 59 #include "util/counts.h" 60 #include "util/topdown.h" 61 #include "util/session.h" 62 #include "util/tool.h" 63 #include "util/string2.h" 64 #include "util/metricgroup.h" 65 #include "util/synthetic-events.h" 66 #include "util/target.h" 67 #include "util/time-utils.h" 68 #include "util/top.h" 69 #include "util/affinity.h" 70 #include "util/pfm.h" 71 #include "util/bpf_counter.h" 72 #include "util/iostat.h" 73 #include "util/util.h" 74 #include "util/intel-tpebs.h" 75 #include "asm/bug.h" 76 77 #include <linux/time64.h> 78 #include <linux/zalloc.h> 79 #include <api/fs/fs.h> 80 #include <errno.h> 81 #include <signal.h> 82 #include <stdlib.h> 83 #include <sys/prctl.h> 84 #include <inttypes.h> 85 #include <locale.h> 86 #include <math.h> 87 #include <sys/types.h> 88 #include <sys/stat.h> 89 #include <sys/wait.h> 90 #include <unistd.h> 91 #include <sys/time.h> 92 #include <sys/resource.h> 93 #include <linux/err.h> 94 95 #include <linux/ctype.h> 96 #include <perf/evlist.h> 97 #include <internal/threadmap.h> 98 99 #define DEFAULT_SEPARATOR " " 100 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi" 101 102 static void print_counters(struct timespec *ts, int argc, const char **argv); 103 104 static struct evlist *evsel_list; 105 static struct parse_events_option_args parse_events_option_args = { 106 .evlistp = &evsel_list, 107 }; 108 109 static bool all_counters_use_bpf = true; 110 111 static struct target target = { 112 .uid = UINT_MAX, 113 }; 114 115 static volatile sig_atomic_t child_pid = -1; 116 static int detailed_run = 0; 117 static bool transaction_run; 118 static bool topdown_run = false; 119 static bool smi_cost = false; 120 static bool smi_reset = false; 121 static int big_num_opt = -1; 122 static const char *pre_cmd = NULL; 123 static const char *post_cmd = NULL; 124 static bool sync_run = false; 125 static bool forever = false; 126 static bool force_metric_only = false; 127 static struct timespec ref_time; 128 static bool append_file; 129 static bool interval_count; 130 static const char *output_name; 131 static int output_fd; 132 static char *metrics; 133 134 struct perf_stat { 135 bool record; 136 struct perf_data data; 137 struct perf_session *session; 138 u64 bytes_written; 139 struct perf_tool tool; 140 bool maps_allocated; 141 struct perf_cpu_map *cpus; 142 struct perf_thread_map *threads; 143 enum aggr_mode aggr_mode; 144 u32 aggr_level; 145 }; 146 147 static struct perf_stat perf_stat; 148 #define STAT_RECORD perf_stat.record 149 150 static volatile sig_atomic_t done = 0; 151 152 /* Options set from the command line. */ 153 struct opt_aggr_mode { 154 bool node, socket, die, cluster, cache, core, thread, no_aggr; 155 }; 156 157 /* Turn command line option into most generic aggregation mode setting. */ 158 static enum aggr_mode opt_aggr_mode_to_aggr_mode(struct opt_aggr_mode *opt_mode) 159 { 160 enum aggr_mode mode = AGGR_GLOBAL; 161 162 if (opt_mode->node) 163 mode = AGGR_NODE; 164 if (opt_mode->socket) 165 mode = AGGR_SOCKET; 166 if (opt_mode->die) 167 mode = AGGR_DIE; 168 if (opt_mode->cluster) 169 mode = AGGR_CLUSTER; 170 if (opt_mode->cache) 171 mode = AGGR_CACHE; 172 if (opt_mode->core) 173 mode = AGGR_CORE; 174 if (opt_mode->thread) 175 mode = AGGR_THREAD; 176 if (opt_mode->no_aggr) 177 mode = AGGR_NONE; 178 return mode; 179 } 180 181 static void evlist__check_cpu_maps(struct evlist *evlist) 182 { 183 struct evsel *evsel, *warned_leader = NULL; 184 185 evlist__for_each_entry(evlist, evsel) { 186 struct evsel *leader = evsel__leader(evsel); 187 188 /* Check that leader matches cpus with each member. */ 189 if (leader == evsel) 190 continue; 191 if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus)) 192 continue; 193 194 /* If there's mismatch disable the group and warn user. */ 195 if (warned_leader != leader) { 196 char buf[200]; 197 198 pr_warning("WARNING: grouped events cpus do not match.\n" 199 "Events with CPUs not matching the leader will " 200 "be removed from the group.\n"); 201 evsel__group_desc(leader, buf, sizeof(buf)); 202 pr_warning(" %s\n", buf); 203 warned_leader = leader; 204 } 205 if (verbose > 0) { 206 char buf[200]; 207 208 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf)); 209 pr_warning(" %s: %s\n", leader->name, buf); 210 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf)); 211 pr_warning(" %s: %s\n", evsel->name, buf); 212 } 213 214 evsel__remove_from_group(evsel, leader); 215 } 216 } 217 218 static inline void diff_timespec(struct timespec *r, struct timespec *a, 219 struct timespec *b) 220 { 221 r->tv_sec = a->tv_sec - b->tv_sec; 222 if (a->tv_nsec < b->tv_nsec) { 223 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 224 r->tv_sec--; 225 } else { 226 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 227 } 228 } 229 230 static void perf_stat__reset_stats(void) 231 { 232 evlist__reset_stats(evsel_list); 233 perf_stat__reset_shadow_stats(); 234 } 235 236 static int process_synthesized_event(const struct perf_tool *tool __maybe_unused, 237 union perf_event *event, 238 struct perf_sample *sample __maybe_unused, 239 struct machine *machine __maybe_unused) 240 { 241 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 242 pr_err("failed to write perf data, error: %m\n"); 243 return -1; 244 } 245 246 perf_stat.bytes_written += event->header.size; 247 return 0; 248 } 249 250 static int write_stat_round_event(u64 tm, u64 type) 251 { 252 return perf_event__synthesize_stat_round(NULL, tm, type, 253 process_synthesized_event, 254 NULL); 255 } 256 257 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 258 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 259 260 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 261 262 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread, 263 struct perf_counts_values *count) 264 { 265 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread); 266 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx); 267 268 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 269 process_synthesized_event, NULL); 270 } 271 272 static int read_single_counter(struct evsel *counter, int cpu_map_idx, int thread) 273 { 274 int err = evsel__read_counter(counter, cpu_map_idx, thread); 275 276 /* 277 * Reading user and system time will fail when the process 278 * terminates. Use the wait4 values in that case. 279 */ 280 if (err && cpu_map_idx == 0 && 281 (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME || 282 evsel__tool_event(counter) == TOOL_PMU__EVENT_SYSTEM_TIME)) { 283 u64 val, *start_time; 284 struct perf_counts_values *count = 285 perf_counts(counter->counts, cpu_map_idx, thread); 286 287 start_time = xyarray__entry(counter->start_times, cpu_map_idx, thread); 288 if (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME) 289 val = ru_stats.ru_utime_usec_stat.mean; 290 else 291 val = ru_stats.ru_stime_usec_stat.mean; 292 count->ena = count->run = *start_time + val; 293 count->val = val; 294 return 0; 295 } 296 return err; 297 } 298 299 /* 300 * Read out the results of a single counter: 301 * do not aggregate counts across CPUs in system-wide mode 302 */ 303 static int read_counter_cpu(struct evsel *counter, int cpu_map_idx) 304 { 305 int nthreads = perf_thread_map__nr(evsel_list->core.threads); 306 int thread; 307 308 if (!counter->supported) 309 return -ENOENT; 310 311 for (thread = 0; thread < nthreads; thread++) { 312 struct perf_counts_values *count; 313 314 count = perf_counts(counter->counts, cpu_map_idx, thread); 315 316 /* 317 * The leader's group read loads data into its group members 318 * (via evsel__read_counter()) and sets their count->loaded. 319 */ 320 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) && 321 read_single_counter(counter, cpu_map_idx, thread)) { 322 counter->counts->scaled = -1; 323 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0; 324 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0; 325 return -1; 326 } 327 328 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false); 329 330 if (STAT_RECORD) { 331 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) { 332 pr_err("failed to write stat event\n"); 333 return -1; 334 } 335 } 336 337 if (verbose > 1) { 338 fprintf(stat_config.output, 339 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 340 evsel__name(counter), 341 perf_cpu_map__cpu(evsel__cpus(counter), 342 cpu_map_idx).cpu, 343 count->val, count->ena, count->run); 344 } 345 } 346 347 return 0; 348 } 349 350 static int read_affinity_counters(void) 351 { 352 struct evlist_cpu_iterator evlist_cpu_itr; 353 struct affinity saved_affinity, *affinity; 354 355 if (all_counters_use_bpf) 356 return 0; 357 358 if (!target__has_cpu(&target) || target__has_per_thread(&target)) 359 affinity = NULL; 360 else if (affinity__setup(&saved_affinity) < 0) 361 return -1; 362 else 363 affinity = &saved_affinity; 364 365 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 366 struct evsel *counter = evlist_cpu_itr.evsel; 367 368 if (evsel__is_bpf(counter)) 369 continue; 370 371 if (!counter->err) 372 counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx); 373 } 374 if (affinity) 375 affinity__cleanup(&saved_affinity); 376 377 return 0; 378 } 379 380 static int read_bpf_map_counters(void) 381 { 382 struct evsel *counter; 383 int err; 384 385 evlist__for_each_entry(evsel_list, counter) { 386 if (!evsel__is_bpf(counter)) 387 continue; 388 389 err = bpf_counter__read(counter); 390 if (err) 391 return err; 392 } 393 return 0; 394 } 395 396 static int read_counters(void) 397 { 398 if (!stat_config.stop_read_counter) { 399 if (read_bpf_map_counters() || 400 read_affinity_counters()) 401 return -1; 402 } 403 return 0; 404 } 405 406 static void process_counters(void) 407 { 408 struct evsel *counter; 409 410 evlist__for_each_entry(evsel_list, counter) { 411 if (counter->err) 412 pr_debug("failed to read counter %s\n", counter->name); 413 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) 414 pr_warning("failed to process counter %s\n", counter->name); 415 counter->err = 0; 416 } 417 418 perf_stat_merge_counters(&stat_config, evsel_list); 419 perf_stat_process_percore(&stat_config, evsel_list); 420 } 421 422 static void process_interval(void) 423 { 424 struct timespec ts, rs; 425 426 clock_gettime(CLOCK_MONOTONIC, &ts); 427 diff_timespec(&rs, &ts, &ref_time); 428 429 evlist__reset_aggr_stats(evsel_list); 430 431 if (read_counters() == 0) 432 process_counters(); 433 434 if (STAT_RECORD) { 435 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 436 pr_err("failed to write stat round event\n"); 437 } 438 439 init_stats(&walltime_nsecs_stats); 440 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL); 441 print_counters(&rs, 0, NULL); 442 } 443 444 static bool handle_interval(unsigned int interval, int *times) 445 { 446 if (interval) { 447 process_interval(); 448 if (interval_count && !(--(*times))) 449 return true; 450 } 451 return false; 452 } 453 454 static int enable_counters(void) 455 { 456 struct evsel *evsel; 457 int err; 458 459 evlist__for_each_entry(evsel_list, evsel) { 460 if (!evsel__is_bpf(evsel)) 461 continue; 462 463 err = bpf_counter__enable(evsel); 464 if (err) 465 return err; 466 } 467 468 if (!target__enable_on_exec(&target)) { 469 if (!all_counters_use_bpf) 470 evlist__enable(evsel_list); 471 } 472 return 0; 473 } 474 475 static void disable_counters(void) 476 { 477 struct evsel *counter; 478 479 /* 480 * If we don't have tracee (attaching to task or cpu), counters may 481 * still be running. To get accurate group ratios, we must stop groups 482 * from counting before reading their constituent counters. 483 */ 484 if (!target__none(&target)) { 485 evlist__for_each_entry(evsel_list, counter) 486 bpf_counter__disable(counter); 487 if (!all_counters_use_bpf) 488 evlist__disable(evsel_list); 489 } 490 } 491 492 static volatile sig_atomic_t workload_exec_errno; 493 494 /* 495 * evlist__prepare_workload will send a SIGUSR1 496 * if the fork fails, since we asked by setting its 497 * want_signal to true. 498 */ 499 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 500 void *ucontext __maybe_unused) 501 { 502 workload_exec_errno = info->si_value.sival_int; 503 } 504 505 static bool evsel__should_store_id(struct evsel *counter) 506 { 507 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 508 } 509 510 static bool is_target_alive(struct target *_target, 511 struct perf_thread_map *threads) 512 { 513 struct stat st; 514 int i; 515 516 if (!target__has_task(_target)) 517 return true; 518 519 for (i = 0; i < threads->nr; i++) { 520 char path[PATH_MAX]; 521 522 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 523 threads->map[i].pid); 524 525 if (!stat(path, &st)) 526 return true; 527 } 528 529 return false; 530 } 531 532 static void process_evlist(struct evlist *evlist, unsigned int interval) 533 { 534 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED; 535 536 if (evlist__ctlfd_process(evlist, &cmd) > 0) { 537 switch (cmd) { 538 case EVLIST_CTL_CMD_ENABLE: 539 fallthrough; 540 case EVLIST_CTL_CMD_DISABLE: 541 if (interval) 542 process_interval(); 543 break; 544 case EVLIST_CTL_CMD_SNAPSHOT: 545 case EVLIST_CTL_CMD_ACK: 546 case EVLIST_CTL_CMD_UNSUPPORTED: 547 case EVLIST_CTL_CMD_EVLIST: 548 case EVLIST_CTL_CMD_STOP: 549 case EVLIST_CTL_CMD_PING: 550 default: 551 break; 552 } 553 } 554 } 555 556 static void compute_tts(struct timespec *time_start, struct timespec *time_stop, 557 int *time_to_sleep) 558 { 559 int tts = *time_to_sleep; 560 struct timespec time_diff; 561 562 diff_timespec(&time_diff, time_stop, time_start); 563 564 tts -= time_diff.tv_sec * MSEC_PER_SEC + 565 time_diff.tv_nsec / NSEC_PER_MSEC; 566 567 if (tts < 0) 568 tts = 0; 569 570 *time_to_sleep = tts; 571 } 572 573 static int dispatch_events(bool forks, int timeout, int interval, int *times) 574 { 575 int child_exited = 0, status = 0; 576 int time_to_sleep, sleep_time; 577 struct timespec time_start, time_stop; 578 579 if (interval) 580 sleep_time = interval; 581 else if (timeout) 582 sleep_time = timeout; 583 else 584 sleep_time = 1000; 585 586 time_to_sleep = sleep_time; 587 588 while (!done) { 589 if (forks) 590 child_exited = waitpid(child_pid, &status, WNOHANG); 591 else 592 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0; 593 594 if (child_exited) 595 break; 596 597 clock_gettime(CLOCK_MONOTONIC, &time_start); 598 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */ 599 if (timeout || handle_interval(interval, times)) 600 break; 601 time_to_sleep = sleep_time; 602 } else { /* fd revent */ 603 process_evlist(evsel_list, interval); 604 clock_gettime(CLOCK_MONOTONIC, &time_stop); 605 compute_tts(&time_start, &time_stop, &time_to_sleep); 606 } 607 } 608 609 return status; 610 } 611 612 enum counter_recovery { 613 COUNTER_SKIP, 614 COUNTER_RETRY, 615 COUNTER_FATAL, 616 }; 617 618 static enum counter_recovery stat_handle_error(struct evsel *counter) 619 { 620 char msg[BUFSIZ]; 621 /* 622 * PPC returns ENXIO for HW counters until 2.6.37 623 * (behavior changed with commit b0a873e). 624 */ 625 if (errno == EINVAL || errno == ENOSYS || 626 errno == ENOENT || errno == ENXIO) { 627 if (verbose > 0) 628 ui__warning("%s event is not supported by the kernel.\n", 629 evsel__name(counter)); 630 counter->supported = false; 631 /* 632 * errored is a sticky flag that means one of the counter's 633 * cpu event had a problem and needs to be reexamined. 634 */ 635 counter->errored = true; 636 637 if ((evsel__leader(counter) != counter) || 638 !(counter->core.leader->nr_members > 1)) 639 return COUNTER_SKIP; 640 } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) { 641 if (verbose > 0) 642 ui__warning("%s\n", msg); 643 return COUNTER_RETRY; 644 } else if (target__has_per_thread(&target) && errno != EOPNOTSUPP && 645 evsel_list->core.threads && 646 evsel_list->core.threads->err_thread != -1) { 647 /* 648 * For global --per-thread case, skip current 649 * error thread. 650 */ 651 if (!thread_map__remove(evsel_list->core.threads, 652 evsel_list->core.threads->err_thread)) { 653 evsel_list->core.threads->err_thread = -1; 654 return COUNTER_RETRY; 655 } 656 } else if (counter->skippable) { 657 if (verbose > 0) 658 ui__warning("skipping event %s that kernel failed to open .\n", 659 evsel__name(counter)); 660 counter->supported = false; 661 counter->errored = true; 662 return COUNTER_SKIP; 663 } 664 665 if (errno == EOPNOTSUPP) { 666 if (verbose > 0) { 667 ui__warning("%s event is not supported by the kernel.\n", 668 evsel__name(counter)); 669 } 670 counter->supported = false; 671 counter->errored = true; 672 673 if ((evsel__leader(counter) != counter) || 674 !(counter->core.leader->nr_members > 1)) 675 return COUNTER_SKIP; 676 } 677 678 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg)); 679 ui__error("%s\n", msg); 680 681 if (child_pid != -1) 682 kill(child_pid, SIGTERM); 683 684 tpebs_delete(); 685 686 return COUNTER_FATAL; 687 } 688 689 static int __run_perf_stat(int argc, const char **argv, int run_idx) 690 { 691 int interval = stat_config.interval; 692 int times = stat_config.times; 693 int timeout = stat_config.timeout; 694 char msg[BUFSIZ]; 695 unsigned long long t0, t1; 696 struct evsel *counter; 697 size_t l; 698 int status = 0; 699 const bool forks = (argc > 0); 700 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; 701 struct evlist_cpu_iterator evlist_cpu_itr; 702 struct affinity saved_affinity, *affinity = NULL; 703 int err; 704 bool second_pass = false; 705 706 if (forks) { 707 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) { 708 perror("failed to prepare workload"); 709 return -1; 710 } 711 child_pid = evsel_list->workload.pid; 712 } 713 714 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) { 715 if (affinity__setup(&saved_affinity) < 0) { 716 err = -1; 717 goto err_out; 718 } 719 affinity = &saved_affinity; 720 } 721 722 evlist__for_each_entry(evsel_list, counter) { 723 counter->reset_group = false; 724 if (bpf_counter__load(counter, &target)) { 725 err = -1; 726 goto err_out; 727 } 728 if (!(evsel__is_bperf(counter))) 729 all_counters_use_bpf = false; 730 } 731 732 evlist__reset_aggr_stats(evsel_list); 733 734 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 735 counter = evlist_cpu_itr.evsel; 736 737 /* 738 * bperf calls evsel__open_per_cpu() in bperf__load(), so 739 * no need to call it again here. 740 */ 741 if (target.use_bpf) 742 break; 743 744 if (counter->reset_group || counter->errored) 745 continue; 746 if (evsel__is_bperf(counter)) 747 continue; 748 try_again: 749 if (create_perf_stat_counter(counter, &stat_config, &target, 750 evlist_cpu_itr.cpu_map_idx) < 0) { 751 752 /* 753 * Weak group failed. We cannot just undo this here 754 * because earlier CPUs might be in group mode, and the kernel 755 * doesn't support mixing group and non group reads. Defer 756 * it to later. 757 * Don't close here because we're in the wrong affinity. 758 */ 759 if ((errno == EINVAL || errno == EBADF) && 760 evsel__leader(counter) != counter && 761 counter->weak_group) { 762 evlist__reset_weak_group(evsel_list, counter, false); 763 assert(counter->reset_group); 764 second_pass = true; 765 continue; 766 } 767 768 switch (stat_handle_error(counter)) { 769 case COUNTER_FATAL: 770 err = -1; 771 goto err_out; 772 case COUNTER_RETRY: 773 goto try_again; 774 case COUNTER_SKIP: 775 continue; 776 default: 777 break; 778 } 779 780 } 781 counter->supported = true; 782 } 783 784 if (second_pass) { 785 /* 786 * Now redo all the weak group after closing them, 787 * and also close errored counters. 788 */ 789 790 /* First close errored or weak retry */ 791 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 792 counter = evlist_cpu_itr.evsel; 793 794 if (!counter->reset_group && !counter->errored) 795 continue; 796 797 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx); 798 } 799 /* Now reopen weak */ 800 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 801 counter = evlist_cpu_itr.evsel; 802 803 if (!counter->reset_group) 804 continue; 805 try_again_reset: 806 pr_debug2("reopening weak %s\n", evsel__name(counter)); 807 if (create_perf_stat_counter(counter, &stat_config, &target, 808 evlist_cpu_itr.cpu_map_idx) < 0) { 809 810 switch (stat_handle_error(counter)) { 811 case COUNTER_FATAL: 812 err = -1; 813 goto err_out; 814 case COUNTER_RETRY: 815 goto try_again_reset; 816 case COUNTER_SKIP: 817 continue; 818 default: 819 break; 820 } 821 } 822 counter->supported = true; 823 } 824 } 825 affinity__cleanup(affinity); 826 affinity = NULL; 827 828 evlist__for_each_entry(evsel_list, counter) { 829 if (!counter->supported) { 830 perf_evsel__free_fd(&counter->core); 831 continue; 832 } 833 834 l = strlen(counter->unit); 835 if (l > stat_config.unit_width) 836 stat_config.unit_width = l; 837 838 if (evsel__should_store_id(counter) && 839 evsel__store_ids(counter, evsel_list)) { 840 err = -1; 841 goto err_out; 842 } 843 } 844 845 if (evlist__apply_filters(evsel_list, &counter, &target)) { 846 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n", 847 counter->filter, evsel__name(counter), errno, 848 str_error_r(errno, msg, sizeof(msg))); 849 return -1; 850 } 851 852 if (STAT_RECORD) { 853 int fd = perf_data__fd(&perf_stat.data); 854 855 if (is_pipe) { 856 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); 857 } else { 858 err = perf_session__write_header(perf_stat.session, evsel_list, 859 fd, false); 860 } 861 862 if (err < 0) 863 goto err_out; 864 865 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list, 866 process_synthesized_event, is_pipe); 867 if (err < 0) 868 goto err_out; 869 870 } 871 872 if (target.initial_delay) { 873 pr_info(EVLIST_DISABLED_MSG); 874 } else { 875 err = enable_counters(); 876 if (err) { 877 err = -1; 878 goto err_out; 879 } 880 } 881 882 /* Exec the command, if any */ 883 if (forks) 884 evlist__start_workload(evsel_list); 885 886 if (target.initial_delay > 0) { 887 usleep(target.initial_delay * USEC_PER_MSEC); 888 err = enable_counters(); 889 if (err) { 890 err = -1; 891 goto err_out; 892 } 893 894 pr_info(EVLIST_ENABLED_MSG); 895 } 896 897 t0 = rdclock(); 898 clock_gettime(CLOCK_MONOTONIC, &ref_time); 899 900 if (forks) { 901 if (interval || timeout || evlist__ctlfd_initialized(evsel_list)) 902 status = dispatch_events(forks, timeout, interval, ×); 903 if (child_pid != -1) { 904 if (timeout) 905 kill(child_pid, SIGTERM); 906 wait4(child_pid, &status, 0, &stat_config.ru_data); 907 } 908 909 if (workload_exec_errno) { 910 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); 911 pr_err("Workload failed: %s\n", emsg); 912 err = -1; 913 goto err_out; 914 } 915 916 if (WIFSIGNALED(status)) 917 psignal(WTERMSIG(status), argv[0]); 918 } else { 919 status = dispatch_events(forks, timeout, interval, ×); 920 } 921 922 disable_counters(); 923 924 t1 = rdclock(); 925 926 if (stat_config.walltime_run_table) 927 stat_config.walltime_run[run_idx] = t1 - t0; 928 929 if (interval && stat_config.summary) { 930 stat_config.interval = 0; 931 stat_config.stop_read_counter = true; 932 init_stats(&walltime_nsecs_stats); 933 update_stats(&walltime_nsecs_stats, t1 - t0); 934 935 evlist__copy_prev_raw_counts(evsel_list); 936 evlist__reset_prev_raw_counts(evsel_list); 937 evlist__reset_aggr_stats(evsel_list); 938 } else { 939 update_stats(&walltime_nsecs_stats, t1 - t0); 940 update_rusage_stats(&ru_stats, &stat_config.ru_data); 941 } 942 943 /* 944 * Closing a group leader splits the group, and as we only disable 945 * group leaders, results in remaining events becoming enabled. To 946 * avoid arbitrary skew, we must read all counters before closing any 947 * group leaders. 948 */ 949 if (read_counters() == 0) 950 process_counters(); 951 952 /* 953 * We need to keep evsel_list alive, because it's processed 954 * later the evsel_list will be closed after. 955 */ 956 if (!STAT_RECORD) 957 evlist__close(evsel_list); 958 959 return WEXITSTATUS(status); 960 961 err_out: 962 if (forks) 963 evlist__cancel_workload(evsel_list); 964 965 affinity__cleanup(affinity); 966 return err; 967 } 968 969 /* 970 * Returns -1 for fatal errors which signifies to not continue 971 * when in repeat mode. 972 * 973 * Returns < -1 error codes when stat record is used. These 974 * result in the stat information being displayed, but writing 975 * to the file fails and is non fatal. 976 */ 977 static int run_perf_stat(int argc, const char **argv, int run_idx) 978 { 979 int ret; 980 981 if (pre_cmd) { 982 ret = system(pre_cmd); 983 if (ret) 984 return ret; 985 } 986 987 if (sync_run) 988 sync(); 989 990 ret = __run_perf_stat(argc, argv, run_idx); 991 if (ret) 992 return ret; 993 994 if (post_cmd) { 995 ret = system(post_cmd); 996 if (ret) 997 return ret; 998 } 999 1000 return ret; 1001 } 1002 1003 static void print_counters(struct timespec *ts, int argc, const char **argv) 1004 { 1005 /* Do not print anything if we record to the pipe. */ 1006 if (STAT_RECORD && perf_stat.data.is_pipe) 1007 return; 1008 if (quiet) 1009 return; 1010 1011 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv); 1012 } 1013 1014 static volatile sig_atomic_t signr = -1; 1015 1016 static void skip_signal(int signo) 1017 { 1018 if ((child_pid == -1) || stat_config.interval) 1019 done = 1; 1020 1021 signr = signo; 1022 /* 1023 * render child_pid harmless 1024 * won't send SIGTERM to a random 1025 * process in case of race condition 1026 * and fast PID recycling 1027 */ 1028 child_pid = -1; 1029 } 1030 1031 static void sig_atexit(void) 1032 { 1033 sigset_t set, oset; 1034 1035 /* 1036 * avoid race condition with SIGCHLD handler 1037 * in skip_signal() which is modifying child_pid 1038 * goal is to avoid send SIGTERM to a random 1039 * process 1040 */ 1041 sigemptyset(&set); 1042 sigaddset(&set, SIGCHLD); 1043 sigprocmask(SIG_BLOCK, &set, &oset); 1044 1045 if (child_pid != -1) 1046 kill(child_pid, SIGTERM); 1047 1048 sigprocmask(SIG_SETMASK, &oset, NULL); 1049 1050 if (signr == -1) 1051 return; 1052 1053 signal(signr, SIG_DFL); 1054 kill(getpid(), signr); 1055 } 1056 1057 static int stat__set_big_num(const struct option *opt __maybe_unused, 1058 const char *s __maybe_unused, int unset) 1059 { 1060 big_num_opt = unset ? 0 : 1; 1061 perf_stat__set_big_num(!unset); 1062 return 0; 1063 } 1064 1065 static int enable_metric_only(const struct option *opt __maybe_unused, 1066 const char *s __maybe_unused, int unset) 1067 { 1068 force_metric_only = true; 1069 stat_config.metric_only = !unset; 1070 return 0; 1071 } 1072 1073 static int append_metric_groups(const struct option *opt __maybe_unused, 1074 const char *str, 1075 int unset __maybe_unused) 1076 { 1077 if (metrics) { 1078 char *tmp; 1079 1080 if (asprintf(&tmp, "%s,%s", metrics, str) < 0) 1081 return -ENOMEM; 1082 free(metrics); 1083 metrics = tmp; 1084 } else { 1085 metrics = strdup(str); 1086 if (!metrics) 1087 return -ENOMEM; 1088 } 1089 return 0; 1090 } 1091 1092 static int parse_control_option(const struct option *opt, 1093 const char *str, 1094 int unset __maybe_unused) 1095 { 1096 struct perf_stat_config *config = opt->value; 1097 1098 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close); 1099 } 1100 1101 static int parse_stat_cgroups(const struct option *opt, 1102 const char *str, int unset) 1103 { 1104 if (stat_config.cgroup_list) { 1105 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 1106 return -1; 1107 } 1108 1109 return parse_cgroups(opt, str, unset); 1110 } 1111 1112 static int parse_cputype(const struct option *opt, 1113 const char *str, 1114 int unset __maybe_unused) 1115 { 1116 const struct perf_pmu *pmu; 1117 struct evlist *evlist = *(struct evlist **)opt->value; 1118 1119 if (!list_empty(&evlist->core.entries)) { 1120 fprintf(stderr, "Must define cputype before events/metrics\n"); 1121 return -1; 1122 } 1123 1124 pmu = perf_pmus__pmu_for_pmu_filter(str); 1125 if (!pmu) { 1126 fprintf(stderr, "--cputype %s is not supported!\n", str); 1127 return -1; 1128 } 1129 parse_events_option_args.pmu_filter = pmu->name; 1130 1131 return 0; 1132 } 1133 1134 static int parse_cache_level(const struct option *opt, 1135 const char *str, 1136 int unset __maybe_unused) 1137 { 1138 int level; 1139 struct opt_aggr_mode *opt_aggr_mode = (struct opt_aggr_mode *)opt->value; 1140 u32 *aggr_level = (u32 *)opt->data; 1141 1142 /* 1143 * If no string is specified, aggregate based on the topology of 1144 * Last Level Cache (LLC). Since the LLC level can change from 1145 * architecture to architecture, set level greater than 1146 * MAX_CACHE_LVL which will be interpreted as LLC. 1147 */ 1148 if (str == NULL) { 1149 level = MAX_CACHE_LVL + 1; 1150 goto out; 1151 } 1152 1153 /* 1154 * The format to specify cache level is LX or lX where X is the 1155 * cache level. 1156 */ 1157 if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) { 1158 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1159 MAX_CACHE_LVL, 1160 MAX_CACHE_LVL); 1161 return -EINVAL; 1162 } 1163 1164 level = atoi(&str[1]); 1165 if (level < 1) { 1166 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1167 MAX_CACHE_LVL, 1168 MAX_CACHE_LVL); 1169 return -EINVAL; 1170 } 1171 1172 if (level > MAX_CACHE_LVL) { 1173 pr_err("perf only supports max cache level of %d.\n" 1174 "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL); 1175 return -EINVAL; 1176 } 1177 out: 1178 opt_aggr_mode->cache = true; 1179 *aggr_level = level; 1180 return 0; 1181 } 1182 1183 /** 1184 * Calculate the cache instance ID from the map in 1185 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1186 * Cache instance ID is the first CPU reported in the shared_cpu_list file. 1187 */ 1188 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map) 1189 { 1190 int id; 1191 struct perf_cpu_map *cpu_map = perf_cpu_map__new(map); 1192 1193 /* 1194 * If the map contains no CPU, consider the current CPU to 1195 * be the first online CPU in the cache domain else use the 1196 * first online CPU of the cache domain as the ID. 1197 */ 1198 id = perf_cpu_map__min(cpu_map).cpu; 1199 if (id == -1) 1200 id = cpu.cpu; 1201 1202 /* Free the perf_cpu_map used to find the cache ID */ 1203 perf_cpu_map__put(cpu_map); 1204 1205 return id; 1206 } 1207 1208 /** 1209 * cpu__get_cache_id - Returns 0 if successful in populating the 1210 * cache level and cache id. Cache level is read from 1211 * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID 1212 * is the first CPU reported by 1213 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1214 */ 1215 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache) 1216 { 1217 int ret = 0; 1218 u32 cache_level = stat_config.aggr_level; 1219 struct cpu_cache_level caches[MAX_CACHE_LVL]; 1220 u32 i = 0, caches_cnt = 0; 1221 1222 cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1223 cache->cache = -1; 1224 1225 ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt); 1226 if (ret) { 1227 /* 1228 * If caches_cnt is not 0, cpu_cache_level data 1229 * was allocated when building the topology. 1230 * Free the allocated data before returning. 1231 */ 1232 if (caches_cnt) 1233 goto free_caches; 1234 1235 return ret; 1236 } 1237 1238 if (!caches_cnt) 1239 return -1; 1240 1241 /* 1242 * Save the data for the highest level if no 1243 * level was specified by the user. 1244 */ 1245 if (cache_level > MAX_CACHE_LVL) { 1246 int max_level_index = 0; 1247 1248 for (i = 1; i < caches_cnt; ++i) { 1249 if (caches[i].level > caches[max_level_index].level) 1250 max_level_index = i; 1251 } 1252 1253 cache->cache_lvl = caches[max_level_index].level; 1254 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map); 1255 1256 /* Reset i to 0 to free entire caches[] */ 1257 i = 0; 1258 goto free_caches; 1259 } 1260 1261 for (i = 0; i < caches_cnt; ++i) { 1262 if (caches[i].level == cache_level) { 1263 cache->cache_lvl = cache_level; 1264 cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1265 } 1266 1267 cpu_cache_level__free(&caches[i]); 1268 } 1269 1270 free_caches: 1271 /* 1272 * Free all the allocated cpu_cache_level data. 1273 */ 1274 while (i < caches_cnt) 1275 cpu_cache_level__free(&caches[i++]); 1276 1277 return ret; 1278 } 1279 1280 /** 1281 * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache 1282 * level, die and socket populated with the cache instache ID, cache level, 1283 * die and socket for cpu. The function signature is compatible with 1284 * aggr_cpu_id_get_t. 1285 */ 1286 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data) 1287 { 1288 int ret; 1289 struct aggr_cpu_id id; 1290 struct perf_cache cache; 1291 1292 id = aggr_cpu_id__die(cpu, data); 1293 if (aggr_cpu_id__is_empty(&id)) 1294 return id; 1295 1296 ret = cpu__get_cache_details(cpu, &cache); 1297 if (ret) 1298 return id; 1299 1300 id.cache_lvl = cache.cache_lvl; 1301 id.cache = cache.cache; 1302 return id; 1303 } 1304 1305 static const char *const aggr_mode__string[] = { 1306 [AGGR_CORE] = "core", 1307 [AGGR_CACHE] = "cache", 1308 [AGGR_CLUSTER] = "cluster", 1309 [AGGR_DIE] = "die", 1310 [AGGR_GLOBAL] = "global", 1311 [AGGR_NODE] = "node", 1312 [AGGR_NONE] = "none", 1313 [AGGR_SOCKET] = "socket", 1314 [AGGR_THREAD] = "thread", 1315 [AGGR_UNSET] = "unset", 1316 }; 1317 1318 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 1319 struct perf_cpu cpu) 1320 { 1321 return aggr_cpu_id__socket(cpu, /*data=*/NULL); 1322 } 1323 1324 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 1325 struct perf_cpu cpu) 1326 { 1327 return aggr_cpu_id__die(cpu, /*data=*/NULL); 1328 } 1329 1330 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused, 1331 struct perf_cpu cpu) 1332 { 1333 return aggr_cpu_id__cache(cpu, /*data=*/NULL); 1334 } 1335 1336 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused, 1337 struct perf_cpu cpu) 1338 { 1339 return aggr_cpu_id__cluster(cpu, /*data=*/NULL); 1340 } 1341 1342 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 1343 struct perf_cpu cpu) 1344 { 1345 return aggr_cpu_id__core(cpu, /*data=*/NULL); 1346 } 1347 1348 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused, 1349 struct perf_cpu cpu) 1350 { 1351 return aggr_cpu_id__node(cpu, /*data=*/NULL); 1352 } 1353 1354 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused, 1355 struct perf_cpu cpu) 1356 { 1357 return aggr_cpu_id__global(cpu, /*data=*/NULL); 1358 } 1359 1360 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused, 1361 struct perf_cpu cpu) 1362 { 1363 return aggr_cpu_id__cpu(cpu, /*data=*/NULL); 1364 } 1365 1366 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config, 1367 aggr_get_id_t get_id, struct perf_cpu cpu) 1368 { 1369 struct aggr_cpu_id id; 1370 1371 /* per-process mode - should use global aggr mode */ 1372 if (cpu.cpu == -1) 1373 return get_id(config, cpu); 1374 1375 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu])) 1376 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu); 1377 1378 id = config->cpus_aggr_map->map[cpu.cpu]; 1379 return id; 1380 } 1381 1382 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config, 1383 struct perf_cpu cpu) 1384 { 1385 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu); 1386 } 1387 1388 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config, 1389 struct perf_cpu cpu) 1390 { 1391 return perf_stat__get_aggr(config, perf_stat__get_die, cpu); 1392 } 1393 1394 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config, 1395 struct perf_cpu cpu) 1396 { 1397 return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu); 1398 } 1399 1400 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config, 1401 struct perf_cpu cpu) 1402 { 1403 return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu); 1404 } 1405 1406 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config, 1407 struct perf_cpu cpu) 1408 { 1409 return perf_stat__get_aggr(config, perf_stat__get_core, cpu); 1410 } 1411 1412 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config, 1413 struct perf_cpu cpu) 1414 { 1415 return perf_stat__get_aggr(config, perf_stat__get_node, cpu); 1416 } 1417 1418 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config, 1419 struct perf_cpu cpu) 1420 { 1421 return perf_stat__get_aggr(config, perf_stat__get_global, cpu); 1422 } 1423 1424 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config, 1425 struct perf_cpu cpu) 1426 { 1427 return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu); 1428 } 1429 1430 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode) 1431 { 1432 switch (aggr_mode) { 1433 case AGGR_SOCKET: 1434 return aggr_cpu_id__socket; 1435 case AGGR_DIE: 1436 return aggr_cpu_id__die; 1437 case AGGR_CLUSTER: 1438 return aggr_cpu_id__cluster; 1439 case AGGR_CACHE: 1440 return aggr_cpu_id__cache; 1441 case AGGR_CORE: 1442 return aggr_cpu_id__core; 1443 case AGGR_NODE: 1444 return aggr_cpu_id__node; 1445 case AGGR_NONE: 1446 return aggr_cpu_id__cpu; 1447 case AGGR_GLOBAL: 1448 return aggr_cpu_id__global; 1449 case AGGR_THREAD: 1450 case AGGR_UNSET: 1451 case AGGR_MAX: 1452 default: 1453 return NULL; 1454 } 1455 } 1456 1457 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode) 1458 { 1459 switch (aggr_mode) { 1460 case AGGR_SOCKET: 1461 return perf_stat__get_socket_cached; 1462 case AGGR_DIE: 1463 return perf_stat__get_die_cached; 1464 case AGGR_CLUSTER: 1465 return perf_stat__get_cluster_cached; 1466 case AGGR_CACHE: 1467 return perf_stat__get_cache_id_cached; 1468 case AGGR_CORE: 1469 return perf_stat__get_core_cached; 1470 case AGGR_NODE: 1471 return perf_stat__get_node_cached; 1472 case AGGR_NONE: 1473 return perf_stat__get_cpu_cached; 1474 case AGGR_GLOBAL: 1475 return perf_stat__get_global_cached; 1476 case AGGR_THREAD: 1477 case AGGR_UNSET: 1478 case AGGR_MAX: 1479 default: 1480 return NULL; 1481 } 1482 } 1483 1484 static int perf_stat_init_aggr_mode(void) 1485 { 1486 int nr; 1487 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode); 1488 1489 if (get_id) { 1490 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1491 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1492 get_id, /*data=*/NULL, needs_sort); 1493 if (!stat_config.aggr_map) { 1494 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1495 return -1; 1496 } 1497 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode); 1498 } 1499 1500 if (stat_config.aggr_mode == AGGR_THREAD) { 1501 nr = perf_thread_map__nr(evsel_list->core.threads); 1502 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1503 if (stat_config.aggr_map == NULL) 1504 return -ENOMEM; 1505 1506 for (int s = 0; s < nr; s++) { 1507 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1508 1509 id.thread_idx = s; 1510 stat_config.aggr_map->map[s] = id; 1511 } 1512 return 0; 1513 } 1514 1515 /* 1516 * The evsel_list->cpus is the base we operate on, 1517 * taking the highest cpu number to be the size of 1518 * the aggregation translate cpumap. 1519 */ 1520 if (!perf_cpu_map__is_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus)) 1521 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu; 1522 else 1523 nr = 0; 1524 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1); 1525 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 1526 } 1527 1528 static void cpu_aggr_map__delete(struct cpu_aggr_map *map) 1529 { 1530 free(map); 1531 } 1532 1533 static void perf_stat__exit_aggr_mode(void) 1534 { 1535 cpu_aggr_map__delete(stat_config.aggr_map); 1536 cpu_aggr_map__delete(stat_config.cpus_aggr_map); 1537 stat_config.aggr_map = NULL; 1538 stat_config.cpus_aggr_map = NULL; 1539 } 1540 1541 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data) 1542 { 1543 struct perf_env *env = data; 1544 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1545 1546 if (cpu.cpu != -1) 1547 id.socket = env->cpu[cpu.cpu].socket_id; 1548 1549 return id; 1550 } 1551 1552 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data) 1553 { 1554 struct perf_env *env = data; 1555 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1556 1557 if (cpu.cpu != -1) { 1558 /* 1559 * die_id is relative to socket, so start 1560 * with the socket ID and then add die to 1561 * make a unique ID. 1562 */ 1563 id.socket = env->cpu[cpu.cpu].socket_id; 1564 id.die = env->cpu[cpu.cpu].die_id; 1565 } 1566 1567 return id; 1568 } 1569 1570 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env, 1571 u32 cache_level, struct aggr_cpu_id *id) 1572 { 1573 int i; 1574 int caches_cnt = env->caches_cnt; 1575 struct cpu_cache_level *caches = env->caches; 1576 1577 id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1578 id->cache = -1; 1579 1580 if (!caches_cnt) 1581 return; 1582 1583 for (i = caches_cnt - 1; i > -1; --i) { 1584 struct perf_cpu_map *cpu_map; 1585 int map_contains_cpu; 1586 1587 /* 1588 * If user has not specified a level, find the fist level with 1589 * the cpu in the map. Since building the map is expensive, do 1590 * this only if levels match. 1591 */ 1592 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level) 1593 continue; 1594 1595 cpu_map = perf_cpu_map__new(caches[i].map); 1596 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu); 1597 perf_cpu_map__put(cpu_map); 1598 1599 if (map_contains_cpu != -1) { 1600 id->cache_lvl = caches[i].level; 1601 id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1602 return; 1603 } 1604 } 1605 } 1606 1607 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu, 1608 void *data) 1609 { 1610 struct perf_env *env = data; 1611 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1612 1613 if (cpu.cpu != -1) { 1614 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level; 1615 1616 id.socket = env->cpu[cpu.cpu].socket_id; 1617 id.die = env->cpu[cpu.cpu].die_id; 1618 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id); 1619 } 1620 1621 return id; 1622 } 1623 1624 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu, 1625 void *data) 1626 { 1627 struct perf_env *env = data; 1628 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1629 1630 if (cpu.cpu != -1) { 1631 id.socket = env->cpu[cpu.cpu].socket_id; 1632 id.die = env->cpu[cpu.cpu].die_id; 1633 id.cluster = env->cpu[cpu.cpu].cluster_id; 1634 } 1635 1636 return id; 1637 } 1638 1639 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data) 1640 { 1641 struct perf_env *env = data; 1642 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1643 1644 if (cpu.cpu != -1) { 1645 /* 1646 * core_id is relative to socket, die and cluster, we need a 1647 * global id. So we set socket, die id, cluster id and core id. 1648 */ 1649 id.socket = env->cpu[cpu.cpu].socket_id; 1650 id.die = env->cpu[cpu.cpu].die_id; 1651 id.cluster = env->cpu[cpu.cpu].cluster_id; 1652 id.core = env->cpu[cpu.cpu].core_id; 1653 } 1654 1655 return id; 1656 } 1657 1658 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data) 1659 { 1660 struct perf_env *env = data; 1661 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1662 1663 if (cpu.cpu != -1) { 1664 /* 1665 * core_id is relative to socket and die, 1666 * we need a global id. So we set 1667 * socket, die id and core id 1668 */ 1669 id.socket = env->cpu[cpu.cpu].socket_id; 1670 id.die = env->cpu[cpu.cpu].die_id; 1671 id.core = env->cpu[cpu.cpu].core_id; 1672 id.cpu = cpu; 1673 } 1674 1675 return id; 1676 } 1677 1678 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data) 1679 { 1680 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1681 1682 id.node = perf_env__numa_node(data, cpu); 1683 return id; 1684 } 1685 1686 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused, 1687 void *data __maybe_unused) 1688 { 1689 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1690 1691 /* it always aggregates to the cpu 0 */ 1692 id.cpu = (struct perf_cpu){ .cpu = 0 }; 1693 return id; 1694 } 1695 1696 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1697 struct perf_cpu cpu) 1698 { 1699 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1700 } 1701 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1702 struct perf_cpu cpu) 1703 { 1704 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1705 } 1706 1707 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused, 1708 struct perf_cpu cpu) 1709 { 1710 return perf_env__get_cluster_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1711 } 1712 1713 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused, 1714 struct perf_cpu cpu) 1715 { 1716 return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1717 } 1718 1719 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1720 struct perf_cpu cpu) 1721 { 1722 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1723 } 1724 1725 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused, 1726 struct perf_cpu cpu) 1727 { 1728 return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1729 } 1730 1731 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, 1732 struct perf_cpu cpu) 1733 { 1734 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1735 } 1736 1737 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused, 1738 struct perf_cpu cpu) 1739 { 1740 return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1741 } 1742 1743 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode) 1744 { 1745 switch (aggr_mode) { 1746 case AGGR_SOCKET: 1747 return perf_env__get_socket_aggr_by_cpu; 1748 case AGGR_DIE: 1749 return perf_env__get_die_aggr_by_cpu; 1750 case AGGR_CLUSTER: 1751 return perf_env__get_cluster_aggr_by_cpu; 1752 case AGGR_CACHE: 1753 return perf_env__get_cache_aggr_by_cpu; 1754 case AGGR_CORE: 1755 return perf_env__get_core_aggr_by_cpu; 1756 case AGGR_NODE: 1757 return perf_env__get_node_aggr_by_cpu; 1758 case AGGR_GLOBAL: 1759 return perf_env__get_global_aggr_by_cpu; 1760 case AGGR_NONE: 1761 return perf_env__get_cpu_aggr_by_cpu; 1762 case AGGR_THREAD: 1763 case AGGR_UNSET: 1764 case AGGR_MAX: 1765 default: 1766 return NULL; 1767 } 1768 } 1769 1770 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode) 1771 { 1772 switch (aggr_mode) { 1773 case AGGR_SOCKET: 1774 return perf_stat__get_socket_file; 1775 case AGGR_DIE: 1776 return perf_stat__get_die_file; 1777 case AGGR_CLUSTER: 1778 return perf_stat__get_cluster_file; 1779 case AGGR_CACHE: 1780 return perf_stat__get_cache_file; 1781 case AGGR_CORE: 1782 return perf_stat__get_core_file; 1783 case AGGR_NODE: 1784 return perf_stat__get_node_file; 1785 case AGGR_GLOBAL: 1786 return perf_stat__get_global_file; 1787 case AGGR_NONE: 1788 return perf_stat__get_cpu_file; 1789 case AGGR_THREAD: 1790 case AGGR_UNSET: 1791 case AGGR_MAX: 1792 default: 1793 return NULL; 1794 } 1795 } 1796 1797 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1798 { 1799 struct perf_env *env = &st->session->header.env; 1800 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode); 1801 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1802 1803 if (stat_config.aggr_mode == AGGR_THREAD) { 1804 int nr = perf_thread_map__nr(evsel_list->core.threads); 1805 1806 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1807 if (stat_config.aggr_map == NULL) 1808 return -ENOMEM; 1809 1810 for (int s = 0; s < nr; s++) { 1811 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1812 1813 id.thread_idx = s; 1814 stat_config.aggr_map->map[s] = id; 1815 } 1816 return 0; 1817 } 1818 1819 if (!get_id) 1820 return 0; 1821 1822 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1823 get_id, env, needs_sort); 1824 if (!stat_config.aggr_map) { 1825 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1826 return -1; 1827 } 1828 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode); 1829 return 0; 1830 } 1831 1832 /* 1833 * Add default events, if there were no attributes specified or 1834 * if -d/--detailed, -d -d or -d -d -d is used: 1835 */ 1836 static int add_default_events(void) 1837 { 1838 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 1839 struct parse_events_error err; 1840 struct evlist *evlist = evlist__new(); 1841 struct evsel *evsel; 1842 int ret = 0; 1843 1844 if (!evlist) 1845 return -ENOMEM; 1846 1847 parse_events_error__init(&err); 1848 1849 /* Set attrs if no event is selected and !null_run: */ 1850 if (stat_config.null_run) 1851 goto out; 1852 1853 if (transaction_run) { 1854 /* Handle -T as -M transaction. Once platform specific metrics 1855 * support has been added to the json files, all architectures 1856 * will use this approach. To determine transaction support 1857 * on an architecture test for such a metric name. 1858 */ 1859 if (!metricgroup__has_metric(pmu, "transaction")) { 1860 pr_err("Missing transaction metrics\n"); 1861 ret = -1; 1862 goto out; 1863 } 1864 ret = metricgroup__parse_groups(evlist, pmu, "transaction", 1865 stat_config.metric_no_group, 1866 stat_config.metric_no_merge, 1867 stat_config.metric_no_threshold, 1868 stat_config.user_requested_cpu_list, 1869 stat_config.system_wide, 1870 stat_config.hardware_aware_grouping, 1871 &stat_config.metric_events); 1872 goto out; 1873 } 1874 1875 if (smi_cost) { 1876 int smi; 1877 1878 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 1879 pr_err("freeze_on_smi is not supported.\n"); 1880 ret = -1; 1881 goto out; 1882 } 1883 1884 if (!smi) { 1885 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 1886 pr_err("Failed to set freeze_on_smi.\n"); 1887 ret = -1; 1888 goto out; 1889 } 1890 smi_reset = true; 1891 } 1892 1893 if (!metricgroup__has_metric(pmu, "smi")) { 1894 pr_err("Missing smi metrics\n"); 1895 ret = -1; 1896 goto out; 1897 } 1898 1899 if (!force_metric_only) 1900 stat_config.metric_only = true; 1901 1902 ret = metricgroup__parse_groups(evlist, pmu, "smi", 1903 stat_config.metric_no_group, 1904 stat_config.metric_no_merge, 1905 stat_config.metric_no_threshold, 1906 stat_config.user_requested_cpu_list, 1907 stat_config.system_wide, 1908 stat_config.hardware_aware_grouping, 1909 &stat_config.metric_events); 1910 goto out; 1911 } 1912 1913 if (topdown_run) { 1914 unsigned int max_level = metricgroups__topdown_max_level(); 1915 char str[] = "TopdownL1"; 1916 1917 if (!force_metric_only) 1918 stat_config.metric_only = true; 1919 1920 if (!max_level) { 1921 pr_err("Topdown requested but the topdown metric groups aren't present.\n" 1922 "(See perf list the metric groups have names like TopdownL1)\n"); 1923 ret = -1; 1924 goto out; 1925 } 1926 if (stat_config.topdown_level > max_level) { 1927 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level); 1928 ret = -1; 1929 goto out; 1930 } else if (!stat_config.topdown_level) { 1931 stat_config.topdown_level = 1; 1932 } 1933 if (!stat_config.interval && !stat_config.metric_only) { 1934 fprintf(stat_config.output, 1935 "Topdown accuracy may decrease when measuring long periods.\n" 1936 "Please print the result regularly, e.g. -I1000\n"); 1937 } 1938 str[8] = stat_config.topdown_level + '0'; 1939 if (metricgroup__parse_groups(evlist, 1940 pmu, str, 1941 /*metric_no_group=*/false, 1942 /*metric_no_merge=*/false, 1943 /*metric_no_threshold=*/true, 1944 stat_config.user_requested_cpu_list, 1945 stat_config.system_wide, 1946 stat_config.hardware_aware_grouping, 1947 &stat_config.metric_events) < 0) { 1948 ret = -1; 1949 goto out; 1950 } 1951 } 1952 1953 if (!stat_config.topdown_level) 1954 stat_config.topdown_level = 1; 1955 1956 if (!evlist->core.nr_entries && !evsel_list->core.nr_entries) { 1957 /* No events so add defaults. */ 1958 if (target__has_cpu(&target)) 1959 ret = parse_events(evlist, "cpu-clock", &err); 1960 else 1961 ret = parse_events(evlist, "task-clock", &err); 1962 if (ret) 1963 goto out; 1964 1965 ret = parse_events(evlist, 1966 "context-switches," 1967 "cpu-migrations," 1968 "page-faults," 1969 "instructions," 1970 "cycles," 1971 "stalled-cycles-frontend," 1972 "stalled-cycles-backend," 1973 "branches," 1974 "branch-misses", 1975 &err); 1976 if (ret) 1977 goto out; 1978 1979 /* 1980 * Add TopdownL1 metrics if they exist. To minimize 1981 * multiplexing, don't request threshold computation. 1982 */ 1983 if (metricgroup__has_metric(pmu, "Default")) { 1984 struct evlist *metric_evlist = evlist__new(); 1985 1986 if (!metric_evlist) { 1987 ret = -ENOMEM; 1988 goto out; 1989 } 1990 if (metricgroup__parse_groups(metric_evlist, pmu, "Default", 1991 /*metric_no_group=*/false, 1992 /*metric_no_merge=*/false, 1993 /*metric_no_threshold=*/true, 1994 stat_config.user_requested_cpu_list, 1995 stat_config.system_wide, 1996 stat_config.hardware_aware_grouping, 1997 &stat_config.metric_events) < 0) { 1998 ret = -1; 1999 goto out; 2000 } 2001 2002 evlist__for_each_entry(metric_evlist, evsel) 2003 evsel->default_metricgroup = true; 2004 2005 evlist__splice_list_tail(evlist, &metric_evlist->core.entries); 2006 evlist__delete(metric_evlist); 2007 } 2008 } 2009 2010 /* Detailed events get appended to the event list: */ 2011 2012 if (!ret && detailed_run >= 1) { 2013 /* 2014 * Detailed stats (-d), covering the L1 and last level data 2015 * caches: 2016 */ 2017 ret = parse_events(evlist, 2018 "L1-dcache-loads," 2019 "L1-dcache-load-misses," 2020 "LLC-loads," 2021 "LLC-load-misses", 2022 &err); 2023 } 2024 if (!ret && detailed_run >= 2) { 2025 /* 2026 * Very detailed stats (-d -d), covering the instruction cache 2027 * and the TLB caches: 2028 */ 2029 ret = parse_events(evlist, 2030 "L1-icache-loads," 2031 "L1-icache-load-misses," 2032 "dTLB-loads," 2033 "dTLB-load-misses," 2034 "iTLB-loads," 2035 "iTLB-load-misses", 2036 &err); 2037 } 2038 if (!ret && detailed_run >= 3) { 2039 /* 2040 * Very, very detailed stats (-d -d -d), adding prefetch events: 2041 */ 2042 ret = parse_events(evlist, 2043 "L1-dcache-prefetches," 2044 "L1-dcache-prefetch-misses", 2045 &err); 2046 } 2047 out: 2048 if (!ret) { 2049 evlist__for_each_entry(evlist, evsel) { 2050 /* 2051 * Make at least one event non-skippable so fatal errors are visible. 2052 * 'cycles' always used to be default and non-skippable, so use that. 2053 */ 2054 if (strcmp("cycles", evsel__name(evsel))) 2055 evsel->skippable = true; 2056 } 2057 } 2058 parse_events_error__exit(&err); 2059 evlist__splice_list_tail(evsel_list, &evlist->core.entries); 2060 evlist__delete(evlist); 2061 return ret; 2062 } 2063 2064 static const char * const stat_record_usage[] = { 2065 "perf stat record [<options>]", 2066 NULL, 2067 }; 2068 2069 static void init_features(struct perf_session *session) 2070 { 2071 int feat; 2072 2073 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 2074 perf_header__set_feat(&session->header, feat); 2075 2076 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 2077 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 2078 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 2079 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 2080 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 2081 } 2082 2083 static int __cmd_record(const struct option stat_options[], struct opt_aggr_mode *opt_mode, 2084 int argc, const char **argv) 2085 { 2086 struct perf_session *session; 2087 struct perf_data *data = &perf_stat.data; 2088 2089 argc = parse_options(argc, argv, stat_options, stat_record_usage, 2090 PARSE_OPT_STOP_AT_NON_OPTION); 2091 stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(opt_mode); 2092 2093 if (output_name) 2094 data->path = output_name; 2095 2096 if (stat_config.run_count != 1 || forever) { 2097 pr_err("Cannot use -r option with perf stat record.\n"); 2098 return -1; 2099 } 2100 2101 session = perf_session__new(data, NULL); 2102 if (IS_ERR(session)) { 2103 pr_err("Perf session creation failed\n"); 2104 return PTR_ERR(session); 2105 } 2106 2107 init_features(session); 2108 2109 session->evlist = evsel_list; 2110 perf_stat.session = session; 2111 perf_stat.record = true; 2112 return argc; 2113 } 2114 2115 static int process_stat_round_event(struct perf_session *session, 2116 union perf_event *event) 2117 { 2118 struct perf_record_stat_round *stat_round = &event->stat_round; 2119 struct timespec tsh, *ts = NULL; 2120 const char **argv = session->header.env.cmdline_argv; 2121 int argc = session->header.env.nr_cmdline; 2122 2123 process_counters(); 2124 2125 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 2126 update_stats(&walltime_nsecs_stats, stat_round->time); 2127 2128 if (stat_config.interval && stat_round->time) { 2129 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 2130 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 2131 ts = &tsh; 2132 } 2133 2134 print_counters(ts, argc, argv); 2135 return 0; 2136 } 2137 2138 static 2139 int process_stat_config_event(struct perf_session *session, 2140 union perf_event *event) 2141 { 2142 const struct perf_tool *tool = session->tool; 2143 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2144 2145 perf_event__read_stat_config(&stat_config, &event->stat_config); 2146 2147 if (perf_cpu_map__is_empty(st->cpus)) { 2148 if (st->aggr_mode != AGGR_UNSET) 2149 pr_warning("warning: processing task data, aggregation mode not set\n"); 2150 } else if (st->aggr_mode != AGGR_UNSET) { 2151 stat_config.aggr_mode = st->aggr_mode; 2152 } 2153 2154 if (perf_stat.data.is_pipe) 2155 perf_stat_init_aggr_mode(); 2156 else 2157 perf_stat_init_aggr_mode_file(st); 2158 2159 if (stat_config.aggr_map) { 2160 int nr_aggr = stat_config.aggr_map->nr; 2161 2162 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) { 2163 pr_err("cannot allocate aggr counts\n"); 2164 return -1; 2165 } 2166 } 2167 return 0; 2168 } 2169 2170 static int set_maps(struct perf_stat *st) 2171 { 2172 if (!st->cpus || !st->threads) 2173 return 0; 2174 2175 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 2176 return -EINVAL; 2177 2178 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 2179 2180 if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true)) 2181 return -ENOMEM; 2182 2183 st->maps_allocated = true; 2184 return 0; 2185 } 2186 2187 static 2188 int process_thread_map_event(struct perf_session *session, 2189 union perf_event *event) 2190 { 2191 const struct perf_tool *tool = session->tool; 2192 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2193 2194 if (st->threads) { 2195 pr_warning("Extra thread map event, ignoring.\n"); 2196 return 0; 2197 } 2198 2199 st->threads = thread_map__new_event(&event->thread_map); 2200 if (!st->threads) 2201 return -ENOMEM; 2202 2203 return set_maps(st); 2204 } 2205 2206 static 2207 int process_cpu_map_event(struct perf_session *session, 2208 union perf_event *event) 2209 { 2210 const struct perf_tool *tool = session->tool; 2211 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2212 struct perf_cpu_map *cpus; 2213 2214 if (st->cpus) { 2215 pr_warning("Extra cpu map event, ignoring.\n"); 2216 return 0; 2217 } 2218 2219 cpus = cpu_map__new_data(&event->cpu_map.data); 2220 if (!cpus) 2221 return -ENOMEM; 2222 2223 st->cpus = cpus; 2224 return set_maps(st); 2225 } 2226 2227 static const char * const stat_report_usage[] = { 2228 "perf stat report [<options>]", 2229 NULL, 2230 }; 2231 2232 static struct perf_stat perf_stat = { 2233 .aggr_mode = AGGR_UNSET, 2234 .aggr_level = 0, 2235 }; 2236 2237 static int __cmd_report(int argc, const char **argv) 2238 { 2239 struct perf_session *session; 2240 const struct option options[] = { 2241 OPT_STRING('i', "input", &input_name, "file", "input file name"), 2242 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 2243 "aggregate counts per processor socket", AGGR_SOCKET), 2244 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 2245 "aggregate counts per processor die", AGGR_DIE), 2246 OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode, 2247 "aggregate counts perf processor cluster", AGGR_CLUSTER), 2248 OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level, 2249 "cache level", 2250 "aggregate count at this cache level (Default: LLC)", 2251 parse_cache_level), 2252 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 2253 "aggregate counts per physical processor core", AGGR_CORE), 2254 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, 2255 "aggregate counts per numa node", AGGR_NODE), 2256 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 2257 "disable CPU count aggregation", AGGR_NONE), 2258 OPT_END() 2259 }; 2260 struct stat st; 2261 int ret; 2262 2263 argc = parse_options(argc, argv, options, stat_report_usage, 0); 2264 2265 if (!input_name || !strlen(input_name)) { 2266 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 2267 input_name = "-"; 2268 else 2269 input_name = "perf.data"; 2270 } 2271 2272 perf_stat.data.path = input_name; 2273 perf_stat.data.mode = PERF_DATA_MODE_READ; 2274 2275 perf_tool__init(&perf_stat.tool, /*ordered_events=*/false); 2276 perf_stat.tool.attr = perf_event__process_attr; 2277 perf_stat.tool.event_update = perf_event__process_event_update; 2278 perf_stat.tool.thread_map = process_thread_map_event; 2279 perf_stat.tool.cpu_map = process_cpu_map_event; 2280 perf_stat.tool.stat_config = process_stat_config_event; 2281 perf_stat.tool.stat = perf_event__process_stat_event; 2282 perf_stat.tool.stat_round = process_stat_round_event; 2283 2284 session = perf_session__new(&perf_stat.data, &perf_stat.tool); 2285 if (IS_ERR(session)) 2286 return PTR_ERR(session); 2287 2288 perf_stat.session = session; 2289 stat_config.output = stderr; 2290 evlist__delete(evsel_list); 2291 evsel_list = session->evlist; 2292 2293 ret = perf_session__process_events(session); 2294 if (ret) 2295 return ret; 2296 2297 perf_session__delete(session); 2298 return 0; 2299 } 2300 2301 static void setup_system_wide(int forks) 2302 { 2303 /* 2304 * Make system wide (-a) the default target if 2305 * no target was specified and one of following 2306 * conditions is met: 2307 * 2308 * - there's no workload specified 2309 * - there is workload specified but all requested 2310 * events are system wide events 2311 */ 2312 if (!target__none(&target)) 2313 return; 2314 2315 if (!forks) 2316 target.system_wide = true; 2317 else { 2318 struct evsel *counter; 2319 2320 evlist__for_each_entry(evsel_list, counter) { 2321 if (!counter->core.requires_cpu && 2322 !evsel__name_is(counter, "duration_time")) { 2323 return; 2324 } 2325 } 2326 2327 if (evsel_list->core.nr_entries) 2328 target.system_wide = true; 2329 } 2330 } 2331 2332 int cmd_stat(int argc, const char **argv) 2333 { 2334 struct opt_aggr_mode opt_mode = {}; 2335 struct option stat_options[] = { 2336 OPT_BOOLEAN('T', "transaction", &transaction_run, 2337 "hardware transaction statistics"), 2338 OPT_CALLBACK('e', "event", &parse_events_option_args, "event", 2339 "event selector. use 'perf list' to list available events", 2340 parse_events_option), 2341 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 2342 "event filter", parse_filter), 2343 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 2344 "child tasks do not inherit counters"), 2345 OPT_STRING('p', "pid", &target.pid, "pid", 2346 "stat events on existing process id"), 2347 OPT_STRING('t', "tid", &target.tid, "tid", 2348 "stat events on existing thread id"), 2349 #ifdef HAVE_BPF_SKEL 2350 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id", 2351 "stat events on existing bpf program id"), 2352 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf, 2353 "use bpf program to count events"), 2354 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path", 2355 "path to perf_event_attr map"), 2356 #endif 2357 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 2358 "system-wide collection from all CPUs"), 2359 OPT_BOOLEAN(0, "scale", &stat_config.scale, 2360 "Use --no-scale to disable counter scaling for multiplexing"), 2361 OPT_INCR('v', "verbose", &verbose, 2362 "be more verbose (show counter open errors, etc)"), 2363 OPT_INTEGER('r', "repeat", &stat_config.run_count, 2364 "repeat command and print average + stddev (max: 100, forever: 0)"), 2365 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 2366 "display details about each run (only with -r option)"), 2367 OPT_BOOLEAN('n', "null", &stat_config.null_run, 2368 "null run - dont start any counters"), 2369 OPT_INCR('d', "detailed", &detailed_run, 2370 "detailed run - start a lot of events"), 2371 OPT_BOOLEAN('S', "sync", &sync_run, 2372 "call sync() before starting a run"), 2373 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 2374 "print large numbers with thousands\' separators", 2375 stat__set_big_num), 2376 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 2377 "list of cpus to monitor in system-wide"), 2378 OPT_BOOLEAN('A', "no-aggr", &opt_mode.no_aggr, 2379 "disable aggregation across CPUs or PMUs"), 2380 OPT_BOOLEAN(0, "no-merge", &opt_mode.no_aggr, 2381 "disable aggregation the same as -A or -no-aggr"), 2382 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge, 2383 "Merge identical named hybrid events"), 2384 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 2385 "print counts with custom separator"), 2386 OPT_BOOLEAN('j', "json-output", &stat_config.json_output, 2387 "print counts in JSON format"), 2388 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 2389 "monitor event in cgroup name only", parse_stat_cgroups), 2390 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name", 2391 "expand events for each cgroup"), 2392 OPT_STRING('o', "output", &output_name, "file", "output file name"), 2393 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 2394 OPT_INTEGER(0, "log-fd", &output_fd, 2395 "log output to fd, instead of stderr"), 2396 OPT_STRING(0, "pre", &pre_cmd, "command", 2397 "command to run prior to the measured command"), 2398 OPT_STRING(0, "post", &post_cmd, "command", 2399 "command to run after to the measured command"), 2400 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 2401 "print counts at regular interval in ms " 2402 "(overhead is possible for values <= 100ms)"), 2403 OPT_INTEGER(0, "interval-count", &stat_config.times, 2404 "print counts for fixed number of times"), 2405 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 2406 "clear screen in between new interval"), 2407 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 2408 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 2409 OPT_BOOLEAN(0, "per-socket", &opt_mode.socket, 2410 "aggregate counts per processor socket"), 2411 OPT_BOOLEAN(0, "per-die", &opt_mode.die, "aggregate counts per processor die"), 2412 OPT_BOOLEAN(0, "per-cluster", &opt_mode.cluster, 2413 "aggregate counts per processor cluster"), 2414 OPT_CALLBACK_OPTARG(0, "per-cache", &opt_mode, &stat_config.aggr_level, 2415 "cache level", "aggregate count at this cache level (Default: LLC)", 2416 parse_cache_level), 2417 OPT_BOOLEAN(0, "per-core", &opt_mode.core, 2418 "aggregate counts per physical processor core"), 2419 OPT_BOOLEAN(0, "per-thread", &opt_mode.thread, "aggregate counts per thread"), 2420 OPT_BOOLEAN(0, "per-node", &opt_mode.node, "aggregate counts per numa node"), 2421 OPT_INTEGER('D', "delay", &target.initial_delay, 2422 "ms to wait before starting measurement after program start (-1: start with events disabled)"), 2423 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 2424 "Only print computed metrics. No raw values", enable_metric_only), 2425 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group, 2426 "don't group metric events, impacts multiplexing"), 2427 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge, 2428 "don't try to share events between metrics in a group"), 2429 OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold, 2430 "disable adding events for the metric threshold calculation"), 2431 OPT_BOOLEAN(0, "topdown", &topdown_run, 2432 "measure top-down statistics"), 2433 #ifdef HAVE_ARCH_X86_64_SUPPORT 2434 OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording, 2435 "enable recording for tpebs when retire_latency required"), 2436 #endif 2437 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level, 2438 "Set the metrics level for the top-down statistics (0: max level)"), 2439 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 2440 "measure SMI cost"), 2441 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 2442 "monitor specified metrics or metric groups (separated by ,)", 2443 append_metric_groups), 2444 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, 2445 "Configure all used events to run in kernel space.", 2446 PARSE_OPT_EXCLUSIVE), 2447 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, 2448 "Configure all used events to run in user space.", 2449 PARSE_OPT_EXCLUSIVE), 2450 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread, 2451 "Use with 'percore' event qualifier to show the event " 2452 "counts of one hardware thread by sum up total hardware " 2453 "threads of same physical core"), 2454 OPT_BOOLEAN(0, "summary", &stat_config.summary, 2455 "print summary for interval mode"), 2456 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary, 2457 "don't print 'summary' for CSV summary output"), 2458 OPT_BOOLEAN(0, "quiet", &quiet, 2459 "don't print any output, messages or warnings (useful with record)"), 2460 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type", 2461 "Only enable events on applying cpu with this type " 2462 "for hybrid platform (e.g. core or atom)", 2463 parse_cputype), 2464 #ifdef HAVE_LIBPFM 2465 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event", 2466 "libpfm4 event selector. use 'perf list' to list available events", 2467 parse_libpfm_events_option), 2468 #endif 2469 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]", 2470 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n" 2471 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n" 2472 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.", 2473 parse_control_option), 2474 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default", 2475 "measure I/O performance metrics provided by arch/platform", 2476 iostat_parse), 2477 OPT_END() 2478 }; 2479 const char * const stat_usage[] = { 2480 "perf stat [<options>] [<command>]", 2481 NULL 2482 }; 2483 int status = -EINVAL, run_idx, err; 2484 const char *mode; 2485 FILE *output = stderr; 2486 unsigned int interval, timeout; 2487 const char * const stat_subcommands[] = { "record", "report" }; 2488 char errbuf[BUFSIZ]; 2489 2490 setlocale(LC_ALL, ""); 2491 2492 evsel_list = evlist__new(); 2493 if (evsel_list == NULL) 2494 return -ENOMEM; 2495 2496 parse_events__shrink_config_terms(); 2497 2498 /* String-parsing callback-based options would segfault when negated */ 2499 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 2500 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 2501 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 2502 2503 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 2504 (const char **) stat_usage, 2505 PARSE_OPT_STOP_AT_NON_OPTION); 2506 2507 stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(&opt_mode); 2508 2509 if (stat_config.csv_sep) { 2510 stat_config.csv_output = true; 2511 if (!strcmp(stat_config.csv_sep, "\\t")) 2512 stat_config.csv_sep = "\t"; 2513 } else 2514 stat_config.csv_sep = DEFAULT_SEPARATOR; 2515 2516 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2517 argc = __cmd_record(stat_options, &opt_mode, argc, argv); 2518 if (argc < 0) 2519 return -1; 2520 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0])) 2521 return __cmd_report(argc, argv); 2522 2523 interval = stat_config.interval; 2524 timeout = stat_config.timeout; 2525 2526 /* 2527 * For record command the -o is already taken care of. 2528 */ 2529 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 2530 output = NULL; 2531 2532 if (output_name && output_fd) { 2533 fprintf(stderr, "cannot use both --output and --log-fd\n"); 2534 parse_options_usage(stat_usage, stat_options, "o", 1); 2535 parse_options_usage(NULL, stat_options, "log-fd", 0); 2536 goto out; 2537 } 2538 2539 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 2540 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 2541 goto out; 2542 } 2543 2544 if (stat_config.metric_only && stat_config.run_count > 1) { 2545 fprintf(stderr, "--metric-only is not supported with -r\n"); 2546 goto out; 2547 } 2548 2549 if (stat_config.csv_output || (stat_config.metric_only && stat_config.json_output)) { 2550 /* 2551 * Current CSV and metric-only JSON output doesn't display the 2552 * metric threshold so don't compute it. 2553 */ 2554 stat_config.metric_no_threshold = true; 2555 } 2556 2557 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 2558 fprintf(stderr, "--table is only supported with -r\n"); 2559 parse_options_usage(stat_usage, stat_options, "r", 1); 2560 parse_options_usage(NULL, stat_options, "table", 0); 2561 goto out; 2562 } 2563 2564 if (output_fd < 0) { 2565 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 2566 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 2567 goto out; 2568 } 2569 2570 if (!output && !quiet) { 2571 struct timespec tm; 2572 mode = append_file ? "a" : "w"; 2573 2574 output = fopen(output_name, mode); 2575 if (!output) { 2576 perror("failed to create output file"); 2577 return -1; 2578 } 2579 if (!stat_config.json_output) { 2580 clock_gettime(CLOCK_REALTIME, &tm); 2581 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 2582 } 2583 } else if (output_fd > 0) { 2584 mode = append_file ? "a" : "w"; 2585 output = fdopen(output_fd, mode); 2586 if (!output) { 2587 perror("Failed opening logfd"); 2588 return -errno; 2589 } 2590 } 2591 2592 if (stat_config.interval_clear && !isatty(fileno(output))) { 2593 fprintf(stderr, "--interval-clear does not work with output\n"); 2594 parse_options_usage(stat_usage, stat_options, "o", 1); 2595 parse_options_usage(NULL, stat_options, "log-fd", 0); 2596 parse_options_usage(NULL, stat_options, "interval-clear", 0); 2597 return -1; 2598 } 2599 2600 stat_config.output = output; 2601 2602 /* 2603 * let the spreadsheet do the pretty-printing 2604 */ 2605 if (stat_config.csv_output) { 2606 /* User explicitly passed -B? */ 2607 if (big_num_opt == 1) { 2608 fprintf(stderr, "-B option not supported with -x\n"); 2609 parse_options_usage(stat_usage, stat_options, "B", 1); 2610 parse_options_usage(NULL, stat_options, "x", 1); 2611 goto out; 2612 } else /* Nope, so disable big number formatting */ 2613 stat_config.big_num = false; 2614 } else if (big_num_opt == 0) /* User passed --no-big-num */ 2615 stat_config.big_num = false; 2616 2617 target.inherit = !stat_config.no_inherit; 2618 err = target__validate(&target); 2619 if (err) { 2620 target__strerror(&target, err, errbuf, BUFSIZ); 2621 pr_warning("%s\n", errbuf); 2622 } 2623 2624 setup_system_wide(argc); 2625 2626 /* 2627 * Display user/system times only for single 2628 * run and when there's specified tracee. 2629 */ 2630 if ((stat_config.run_count == 1) && target__none(&target)) 2631 stat_config.ru_display = true; 2632 2633 if (stat_config.run_count < 0) { 2634 pr_err("Run count must be a positive number\n"); 2635 parse_options_usage(stat_usage, stat_options, "r", 1); 2636 goto out; 2637 } else if (stat_config.run_count == 0) { 2638 forever = true; 2639 stat_config.run_count = 1; 2640 } 2641 2642 if (stat_config.walltime_run_table) { 2643 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 2644 if (!stat_config.walltime_run) { 2645 pr_err("failed to setup -r option"); 2646 goto out; 2647 } 2648 } 2649 2650 if ((stat_config.aggr_mode == AGGR_THREAD) && 2651 !target__has_task(&target)) { 2652 if (!target.system_wide || target.cpu_list) { 2653 fprintf(stderr, "The --per-thread option is only " 2654 "available when monitoring via -p -t -a " 2655 "options or only --per-thread.\n"); 2656 parse_options_usage(NULL, stat_options, "p", 1); 2657 parse_options_usage(NULL, stat_options, "t", 1); 2658 goto out; 2659 } 2660 } 2661 2662 /* 2663 * no_aggr, cgroup are for system-wide only 2664 * --per-thread is aggregated per thread, we dont mix it with cpu mode 2665 */ 2666 if (((stat_config.aggr_mode != AGGR_GLOBAL && 2667 stat_config.aggr_mode != AGGR_THREAD) || 2668 (nr_cgroups || stat_config.cgroup_list)) && 2669 !target__has_cpu(&target)) { 2670 fprintf(stderr, "both cgroup and no-aggregation " 2671 "modes only available in system-wide mode\n"); 2672 2673 parse_options_usage(stat_usage, stat_options, "G", 1); 2674 parse_options_usage(NULL, stat_options, "A", 1); 2675 parse_options_usage(NULL, stat_options, "a", 1); 2676 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2677 goto out; 2678 } 2679 2680 if (stat_config.iostat_run) { 2681 status = iostat_prepare(evsel_list, &stat_config); 2682 if (status) 2683 goto out; 2684 if (iostat_mode == IOSTAT_LIST) { 2685 iostat_list(evsel_list, &stat_config); 2686 goto out; 2687 } else if (verbose > 0) 2688 iostat_list(evsel_list, &stat_config); 2689 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target)) 2690 target.system_wide = true; 2691 } 2692 2693 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 2694 target.per_thread = true; 2695 2696 stat_config.system_wide = target.system_wide; 2697 if (target.cpu_list) { 2698 stat_config.user_requested_cpu_list = strdup(target.cpu_list); 2699 if (!stat_config.user_requested_cpu_list) { 2700 status = -ENOMEM; 2701 goto out; 2702 } 2703 } 2704 2705 /* 2706 * Metric parsing needs to be delayed as metrics may optimize events 2707 * knowing the target is system-wide. 2708 */ 2709 if (metrics) { 2710 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 2711 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics, 2712 stat_config.metric_no_group, 2713 stat_config.metric_no_merge, 2714 stat_config.metric_no_threshold, 2715 stat_config.user_requested_cpu_list, 2716 stat_config.system_wide, 2717 stat_config.hardware_aware_grouping, 2718 &stat_config.metric_events); 2719 2720 zfree(&metrics); 2721 if (ret) { 2722 status = ret; 2723 goto out; 2724 } 2725 } 2726 2727 if (add_default_events()) 2728 goto out; 2729 2730 if (stat_config.cgroup_list) { 2731 if (nr_cgroups > 0) { 2732 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 2733 parse_options_usage(stat_usage, stat_options, "G", 1); 2734 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2735 goto out; 2736 } 2737 2738 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, 2739 &stat_config.metric_events, true) < 0) { 2740 parse_options_usage(stat_usage, stat_options, 2741 "for-each-cgroup", 0); 2742 goto out; 2743 } 2744 } 2745 2746 evlist__warn_user_requested_cpus(evsel_list, target.cpu_list); 2747 2748 if (evlist__create_maps(evsel_list, &target) < 0) { 2749 if (target__has_task(&target)) { 2750 pr_err("Problems finding threads of monitor\n"); 2751 parse_options_usage(stat_usage, stat_options, "p", 1); 2752 parse_options_usage(NULL, stat_options, "t", 1); 2753 } else if (target__has_cpu(&target)) { 2754 perror("failed to parse CPUs map"); 2755 parse_options_usage(stat_usage, stat_options, "C", 1); 2756 parse_options_usage(NULL, stat_options, "a", 1); 2757 } 2758 goto out; 2759 } 2760 2761 evlist__check_cpu_maps(evsel_list); 2762 2763 /* 2764 * Initialize thread_map with comm names, 2765 * so we could print it out on output. 2766 */ 2767 if (stat_config.aggr_mode == AGGR_THREAD) { 2768 thread_map__read_comms(evsel_list->core.threads); 2769 } 2770 2771 if (stat_config.aggr_mode == AGGR_NODE) 2772 cpu__setup_cpunode_map(); 2773 2774 if (stat_config.times && interval) 2775 interval_count = true; 2776 else if (stat_config.times && !interval) { 2777 pr_err("interval-count option should be used together with " 2778 "interval-print.\n"); 2779 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 2780 parse_options_usage(stat_usage, stat_options, "I", 1); 2781 goto out; 2782 } 2783 2784 if (timeout && timeout < 100) { 2785 if (timeout < 10) { 2786 pr_err("timeout must be >= 10ms.\n"); 2787 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2788 goto out; 2789 } else 2790 pr_warning("timeout < 100ms. " 2791 "The overhead percentage could be high in some cases. " 2792 "Please proceed with caution.\n"); 2793 } 2794 if (timeout && interval) { 2795 pr_err("timeout option is not supported with interval-print.\n"); 2796 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2797 parse_options_usage(stat_usage, stat_options, "I", 1); 2798 goto out; 2799 } 2800 2801 if (perf_stat_init_aggr_mode()) 2802 goto out; 2803 2804 if (evlist__alloc_stats(&stat_config, evsel_list, interval)) 2805 goto out; 2806 2807 /* 2808 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 2809 * while avoiding that older tools show confusing messages. 2810 * 2811 * However for pipe sessions we need to keep it zero, 2812 * because script's perf_evsel__check_attr is triggered 2813 * by attr->sample_type != 0, and we can't run it on 2814 * stat sessions. 2815 */ 2816 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 2817 2818 /* 2819 * We dont want to block the signals - that would cause 2820 * child tasks to inherit that and Ctrl-C would not work. 2821 * What we want is for Ctrl-C to work in the exec()-ed 2822 * task, but being ignored by perf stat itself: 2823 */ 2824 atexit(sig_atexit); 2825 if (!forever) 2826 signal(SIGINT, skip_signal); 2827 signal(SIGCHLD, skip_signal); 2828 signal(SIGALRM, skip_signal); 2829 signal(SIGABRT, skip_signal); 2830 2831 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack)) 2832 goto out; 2833 2834 /* Enable ignoring missing threads when -p option is defined. */ 2835 evlist__first(evsel_list)->ignore_missing_thread = target.pid; 2836 status = 0; 2837 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 2838 if (stat_config.run_count != 1 && verbose > 0) 2839 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 2840 run_idx + 1); 2841 2842 if (run_idx != 0) 2843 evlist__reset_prev_raw_counts(evsel_list); 2844 2845 status = run_perf_stat(argc, argv, run_idx); 2846 if (status == -1) 2847 break; 2848 2849 if (forever && !interval) { 2850 print_counters(NULL, argc, argv); 2851 perf_stat__reset_stats(); 2852 } 2853 } 2854 2855 if (!forever && status != -1 && (!interval || stat_config.summary)) { 2856 if (stat_config.run_count > 1) 2857 evlist__copy_res_stats(&stat_config, evsel_list); 2858 print_counters(NULL, argc, argv); 2859 } 2860 2861 evlist__finalize_ctlfd(evsel_list); 2862 2863 if (STAT_RECORD) { 2864 /* 2865 * We synthesize the kernel mmap record just so that older tools 2866 * don't emit warnings about not being able to resolve symbols 2867 * due to /proc/sys/kernel/kptr_restrict settings and instead provide 2868 * a saner message about no samples being in the perf.data file. 2869 * 2870 * This also serves to suppress a warning about f_header.data.size == 0 2871 * in header.c at the moment 'perf stat record' gets introduced, which 2872 * is not really needed once we start adding the stat specific PERF_RECORD_ 2873 * records, but the need to suppress the kptr_restrict messages in older 2874 * tools remain -acme 2875 */ 2876 int fd = perf_data__fd(&perf_stat.data); 2877 2878 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 2879 process_synthesized_event, 2880 &perf_stat.session->machines.host); 2881 if (err) { 2882 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 2883 "older tools may produce warnings about this file\n."); 2884 } 2885 2886 if (!interval) { 2887 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) 2888 pr_err("failed to write stat round event\n"); 2889 } 2890 2891 if (!perf_stat.data.is_pipe) { 2892 perf_stat.session->header.data_size += perf_stat.bytes_written; 2893 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 2894 } 2895 2896 evlist__close(evsel_list); 2897 perf_session__delete(perf_stat.session); 2898 } 2899 2900 perf_stat__exit_aggr_mode(); 2901 evlist__free_stats(evsel_list); 2902 out: 2903 if (stat_config.iostat_run) 2904 iostat_release(evsel_list); 2905 2906 zfree(&stat_config.walltime_run); 2907 zfree(&stat_config.user_requested_cpu_list); 2908 2909 if (smi_cost && smi_reset) 2910 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 2911 2912 evlist__delete(evsel_list); 2913 2914 metricgroup__rblist_exit(&stat_config.metric_events); 2915 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close); 2916 2917 return status; 2918 } 2919