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/event.h" 50 #include "util/evlist.h" 51 #include "util/evsel.h" 52 #include "util/debug.h" 53 #include "util/color.h" 54 #include "util/stat.h" 55 #include "util/header.h" 56 #include "util/cpumap.h" 57 #include "util/thread_map.h" 58 #include "util/counts.h" 59 #include "util/topdown.h" 60 #include "util/session.h" 61 #include "util/tool.h" 62 #include "util/string2.h" 63 #include "util/metricgroup.h" 64 #include "util/synthetic-events.h" 65 #include "util/target.h" 66 #include "util/time-utils.h" 67 #include "util/top.h" 68 #include "util/affinity.h" 69 #include "util/pfm.h" 70 #include "util/bpf_counter.h" 71 #include "util/iostat.h" 72 #include "util/util.h" 73 #include "asm/bug.h" 74 75 #include <linux/time64.h> 76 #include <linux/zalloc.h> 77 #include <api/fs/fs.h> 78 #include <errno.h> 79 #include <signal.h> 80 #include <stdlib.h> 81 #include <sys/prctl.h> 82 #include <inttypes.h> 83 #include <locale.h> 84 #include <math.h> 85 #include <sys/types.h> 86 #include <sys/stat.h> 87 #include <sys/wait.h> 88 #include <unistd.h> 89 #include <sys/time.h> 90 #include <sys/resource.h> 91 #include <linux/err.h> 92 93 #include <linux/ctype.h> 94 #include <perf/evlist.h> 95 #include <internal/threadmap.h> 96 97 #define DEFAULT_SEPARATOR " " 98 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi" 99 100 static void print_counters(struct timespec *ts, int argc, const char **argv); 101 102 static struct evlist *evsel_list; 103 static struct parse_events_option_args parse_events_option_args = { 104 .evlistp = &evsel_list, 105 }; 106 107 static bool all_counters_use_bpf = true; 108 109 static struct target target = { 110 .uid = UINT_MAX, 111 }; 112 113 #define METRIC_ONLY_LEN 20 114 115 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 static struct perf_stat_config stat_config = { 153 .aggr_mode = AGGR_GLOBAL, 154 .aggr_level = MAX_CACHE_LVL + 1, 155 .scale = true, 156 .unit_width = 4, /* strlen("unit") */ 157 .run_count = 1, 158 .metric_only_len = METRIC_ONLY_LEN, 159 .walltime_nsecs_stats = &walltime_nsecs_stats, 160 .ru_stats = &ru_stats, 161 .big_num = true, 162 .ctl_fd = -1, 163 .ctl_fd_ack = -1, 164 .iostat_run = false, 165 }; 166 167 static bool cpus_map_matched(struct evsel *a, struct evsel *b) 168 { 169 if (!a->core.cpus && !b->core.cpus) 170 return true; 171 172 if (!a->core.cpus || !b->core.cpus) 173 return false; 174 175 if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus)) 176 return false; 177 178 for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) { 179 if (perf_cpu_map__cpu(a->core.cpus, i).cpu != 180 perf_cpu_map__cpu(b->core.cpus, i).cpu) 181 return false; 182 } 183 184 return true; 185 } 186 187 static void evlist__check_cpu_maps(struct evlist *evlist) 188 { 189 struct evsel *evsel, *warned_leader = NULL; 190 191 evlist__for_each_entry(evlist, evsel) { 192 struct evsel *leader = evsel__leader(evsel); 193 194 /* Check that leader matches cpus with each member. */ 195 if (leader == evsel) 196 continue; 197 if (cpus_map_matched(leader, evsel)) 198 continue; 199 200 /* If there's mismatch disable the group and warn user. */ 201 if (warned_leader != leader) { 202 char buf[200]; 203 204 pr_warning("WARNING: grouped events cpus do not match.\n" 205 "Events with CPUs not matching the leader will " 206 "be removed from the group.\n"); 207 evsel__group_desc(leader, buf, sizeof(buf)); 208 pr_warning(" %s\n", buf); 209 warned_leader = leader; 210 } 211 if (verbose > 0) { 212 char buf[200]; 213 214 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf)); 215 pr_warning(" %s: %s\n", leader->name, buf); 216 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf)); 217 pr_warning(" %s: %s\n", evsel->name, buf); 218 } 219 220 evsel__remove_from_group(evsel, leader); 221 } 222 } 223 224 static inline void diff_timespec(struct timespec *r, struct timespec *a, 225 struct timespec *b) 226 { 227 r->tv_sec = a->tv_sec - b->tv_sec; 228 if (a->tv_nsec < b->tv_nsec) { 229 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 230 r->tv_sec--; 231 } else { 232 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 233 } 234 } 235 236 static void perf_stat__reset_stats(void) 237 { 238 evlist__reset_stats(evsel_list); 239 perf_stat__reset_shadow_stats(); 240 } 241 242 static int process_synthesized_event(struct perf_tool *tool __maybe_unused, 243 union perf_event *event, 244 struct perf_sample *sample __maybe_unused, 245 struct machine *machine __maybe_unused) 246 { 247 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 248 pr_err("failed to write perf data, error: %m\n"); 249 return -1; 250 } 251 252 perf_stat.bytes_written += event->header.size; 253 return 0; 254 } 255 256 static int write_stat_round_event(u64 tm, u64 type) 257 { 258 return perf_event__synthesize_stat_round(NULL, tm, type, 259 process_synthesized_event, 260 NULL); 261 } 262 263 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 264 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 265 266 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 267 268 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread, 269 struct perf_counts_values *count) 270 { 271 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread); 272 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx); 273 274 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 275 process_synthesized_event, NULL); 276 } 277 278 static int read_single_counter(struct evsel *counter, int cpu_map_idx, 279 int thread, struct timespec *rs) 280 { 281 switch(counter->tool_event) { 282 case PERF_TOOL_DURATION_TIME: { 283 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL; 284 struct perf_counts_values *count = 285 perf_counts(counter->counts, cpu_map_idx, thread); 286 count->ena = count->run = val; 287 count->val = val; 288 return 0; 289 } 290 case PERF_TOOL_USER_TIME: 291 case PERF_TOOL_SYSTEM_TIME: { 292 u64 val; 293 struct perf_counts_values *count = 294 perf_counts(counter->counts, cpu_map_idx, thread); 295 if (counter->tool_event == PERF_TOOL_USER_TIME) 296 val = ru_stats.ru_utime_usec_stat.mean; 297 else 298 val = ru_stats.ru_stime_usec_stat.mean; 299 count->ena = count->run = val; 300 count->val = val; 301 return 0; 302 } 303 default: 304 case PERF_TOOL_NONE: 305 return evsel__read_counter(counter, cpu_map_idx, thread); 306 case PERF_TOOL_MAX: 307 /* This should never be reached */ 308 return 0; 309 } 310 } 311 312 /* 313 * Read out the results of a single counter: 314 * do not aggregate counts across CPUs in system-wide mode 315 */ 316 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx) 317 { 318 int nthreads = perf_thread_map__nr(evsel_list->core.threads); 319 int thread; 320 321 if (!counter->supported) 322 return -ENOENT; 323 324 for (thread = 0; thread < nthreads; thread++) { 325 struct perf_counts_values *count; 326 327 count = perf_counts(counter->counts, cpu_map_idx, thread); 328 329 /* 330 * The leader's group read loads data into its group members 331 * (via evsel__read_counter()) and sets their count->loaded. 332 */ 333 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) && 334 read_single_counter(counter, cpu_map_idx, thread, rs)) { 335 counter->counts->scaled = -1; 336 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0; 337 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0; 338 return -1; 339 } 340 341 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false); 342 343 if (STAT_RECORD) { 344 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) { 345 pr_err("failed to write stat event\n"); 346 return -1; 347 } 348 } 349 350 if (verbose > 1) { 351 fprintf(stat_config.output, 352 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 353 evsel__name(counter), 354 perf_cpu_map__cpu(evsel__cpus(counter), 355 cpu_map_idx).cpu, 356 count->val, count->ena, count->run); 357 } 358 } 359 360 return 0; 361 } 362 363 static int read_affinity_counters(struct timespec *rs) 364 { 365 struct evlist_cpu_iterator evlist_cpu_itr; 366 struct affinity saved_affinity, *affinity; 367 368 if (all_counters_use_bpf) 369 return 0; 370 371 if (!target__has_cpu(&target) || target__has_per_thread(&target)) 372 affinity = NULL; 373 else if (affinity__setup(&saved_affinity) < 0) 374 return -1; 375 else 376 affinity = &saved_affinity; 377 378 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { 379 struct evsel *counter = evlist_cpu_itr.evsel; 380 381 if (evsel__is_bpf(counter)) 382 continue; 383 384 if (!counter->err) { 385 counter->err = read_counter_cpu(counter, rs, 386 evlist_cpu_itr.cpu_map_idx); 387 } 388 } 389 if (affinity) 390 affinity__cleanup(&saved_affinity); 391 392 return 0; 393 } 394 395 static int read_bpf_map_counters(void) 396 { 397 struct evsel *counter; 398 int err; 399 400 evlist__for_each_entry(evsel_list, counter) { 401 if (!evsel__is_bpf(counter)) 402 continue; 403 404 err = bpf_counter__read(counter); 405 if (err) 406 return err; 407 } 408 return 0; 409 } 410 411 static int read_counters(struct timespec *rs) 412 { 413 if (!stat_config.stop_read_counter) { 414 if (read_bpf_map_counters() || 415 read_affinity_counters(rs)) 416 return -1; 417 } 418 return 0; 419 } 420 421 static void process_counters(void) 422 { 423 struct evsel *counter; 424 425 evlist__for_each_entry(evsel_list, counter) { 426 if (counter->err) 427 pr_debug("failed to read counter %s\n", counter->name); 428 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) 429 pr_warning("failed to process counter %s\n", counter->name); 430 counter->err = 0; 431 } 432 433 perf_stat_merge_counters(&stat_config, evsel_list); 434 perf_stat_process_percore(&stat_config, evsel_list); 435 } 436 437 static void process_interval(void) 438 { 439 struct timespec ts, rs; 440 441 clock_gettime(CLOCK_MONOTONIC, &ts); 442 diff_timespec(&rs, &ts, &ref_time); 443 444 evlist__reset_aggr_stats(evsel_list); 445 446 if (read_counters(&rs) == 0) 447 process_counters(); 448 449 if (STAT_RECORD) { 450 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 451 pr_err("failed to write stat round event\n"); 452 } 453 454 init_stats(&walltime_nsecs_stats); 455 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL); 456 print_counters(&rs, 0, NULL); 457 } 458 459 static bool handle_interval(unsigned int interval, int *times) 460 { 461 if (interval) { 462 process_interval(); 463 if (interval_count && !(--(*times))) 464 return true; 465 } 466 return false; 467 } 468 469 static int enable_counters(void) 470 { 471 struct evsel *evsel; 472 int err; 473 474 evlist__for_each_entry(evsel_list, evsel) { 475 if (!evsel__is_bpf(evsel)) 476 continue; 477 478 err = bpf_counter__enable(evsel); 479 if (err) 480 return err; 481 } 482 483 if (!target__enable_on_exec(&target)) { 484 if (!all_counters_use_bpf) 485 evlist__enable(evsel_list); 486 } 487 return 0; 488 } 489 490 static void disable_counters(void) 491 { 492 struct evsel *counter; 493 494 /* 495 * If we don't have tracee (attaching to task or cpu), counters may 496 * still be running. To get accurate group ratios, we must stop groups 497 * from counting before reading their constituent counters. 498 */ 499 if (!target__none(&target)) { 500 evlist__for_each_entry(evsel_list, counter) 501 bpf_counter__disable(counter); 502 if (!all_counters_use_bpf) 503 evlist__disable(evsel_list); 504 } 505 } 506 507 static volatile sig_atomic_t workload_exec_errno; 508 509 /* 510 * evlist__prepare_workload will send a SIGUSR1 511 * if the fork fails, since we asked by setting its 512 * want_signal to true. 513 */ 514 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 515 void *ucontext __maybe_unused) 516 { 517 workload_exec_errno = info->si_value.sival_int; 518 } 519 520 static bool evsel__should_store_id(struct evsel *counter) 521 { 522 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 523 } 524 525 static bool is_target_alive(struct target *_target, 526 struct perf_thread_map *threads) 527 { 528 struct stat st; 529 int i; 530 531 if (!target__has_task(_target)) 532 return true; 533 534 for (i = 0; i < threads->nr; i++) { 535 char path[PATH_MAX]; 536 537 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 538 threads->map[i].pid); 539 540 if (!stat(path, &st)) 541 return true; 542 } 543 544 return false; 545 } 546 547 static void process_evlist(struct evlist *evlist, unsigned int interval) 548 { 549 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED; 550 551 if (evlist__ctlfd_process(evlist, &cmd) > 0) { 552 switch (cmd) { 553 case EVLIST_CTL_CMD_ENABLE: 554 fallthrough; 555 case EVLIST_CTL_CMD_DISABLE: 556 if (interval) 557 process_interval(); 558 break; 559 case EVLIST_CTL_CMD_SNAPSHOT: 560 case EVLIST_CTL_CMD_ACK: 561 case EVLIST_CTL_CMD_UNSUPPORTED: 562 case EVLIST_CTL_CMD_EVLIST: 563 case EVLIST_CTL_CMD_STOP: 564 case EVLIST_CTL_CMD_PING: 565 default: 566 break; 567 } 568 } 569 } 570 571 static void compute_tts(struct timespec *time_start, struct timespec *time_stop, 572 int *time_to_sleep) 573 { 574 int tts = *time_to_sleep; 575 struct timespec time_diff; 576 577 diff_timespec(&time_diff, time_stop, time_start); 578 579 tts -= time_diff.tv_sec * MSEC_PER_SEC + 580 time_diff.tv_nsec / NSEC_PER_MSEC; 581 582 if (tts < 0) 583 tts = 0; 584 585 *time_to_sleep = tts; 586 } 587 588 static int dispatch_events(bool forks, int timeout, int interval, int *times) 589 { 590 int child_exited = 0, status = 0; 591 int time_to_sleep, sleep_time; 592 struct timespec time_start, time_stop; 593 594 if (interval) 595 sleep_time = interval; 596 else if (timeout) 597 sleep_time = timeout; 598 else 599 sleep_time = 1000; 600 601 time_to_sleep = sleep_time; 602 603 while (!done) { 604 if (forks) 605 child_exited = waitpid(child_pid, &status, WNOHANG); 606 else 607 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0; 608 609 if (child_exited) 610 break; 611 612 clock_gettime(CLOCK_MONOTONIC, &time_start); 613 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */ 614 if (timeout || handle_interval(interval, times)) 615 break; 616 time_to_sleep = sleep_time; 617 } else { /* fd revent */ 618 process_evlist(evsel_list, interval); 619 clock_gettime(CLOCK_MONOTONIC, &time_stop); 620 compute_tts(&time_start, &time_stop, &time_to_sleep); 621 } 622 } 623 624 return status; 625 } 626 627 enum counter_recovery { 628 COUNTER_SKIP, 629 COUNTER_RETRY, 630 COUNTER_FATAL, 631 }; 632 633 static enum counter_recovery stat_handle_error(struct evsel *counter) 634 { 635 char msg[BUFSIZ]; 636 /* 637 * PPC returns ENXIO for HW counters until 2.6.37 638 * (behavior changed with commit b0a873e). 639 */ 640 if (errno == EINVAL || errno == ENOSYS || 641 errno == ENOENT || errno == EOPNOTSUPP || 642 errno == ENXIO) { 643 if (verbose > 0) 644 ui__warning("%s event is not supported by the kernel.\n", 645 evsel__name(counter)); 646 counter->supported = false; 647 /* 648 * errored is a sticky flag that means one of the counter's 649 * cpu event had a problem and needs to be reexamined. 650 */ 651 counter->errored = true; 652 653 if ((evsel__leader(counter) != counter) || 654 !(counter->core.leader->nr_members > 1)) 655 return COUNTER_SKIP; 656 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) { 657 if (verbose > 0) 658 ui__warning("%s\n", msg); 659 return COUNTER_RETRY; 660 } else if (target__has_per_thread(&target) && 661 evsel_list->core.threads && 662 evsel_list->core.threads->err_thread != -1) { 663 /* 664 * For global --per-thread case, skip current 665 * error thread. 666 */ 667 if (!thread_map__remove(evsel_list->core.threads, 668 evsel_list->core.threads->err_thread)) { 669 evsel_list->core.threads->err_thread = -1; 670 return COUNTER_RETRY; 671 } 672 } else if (counter->skippable) { 673 if (verbose > 0) 674 ui__warning("skipping event %s that kernel failed to open .\n", 675 evsel__name(counter)); 676 counter->supported = false; 677 counter->errored = true; 678 return COUNTER_SKIP; 679 } 680 681 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg)); 682 ui__error("%s\n", msg); 683 684 if (child_pid != -1) 685 kill(child_pid, SIGTERM); 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)) { 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(&(struct timespec) { .tv_nsec = t1-t0 }) == 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 static int run_perf_stat(int argc, const char **argv, int run_idx) 970 { 971 int ret; 972 973 if (pre_cmd) { 974 ret = system(pre_cmd); 975 if (ret) 976 return ret; 977 } 978 979 if (sync_run) 980 sync(); 981 982 ret = __run_perf_stat(argc, argv, run_idx); 983 if (ret) 984 return ret; 985 986 if (post_cmd) { 987 ret = system(post_cmd); 988 if (ret) 989 return ret; 990 } 991 992 return ret; 993 } 994 995 static void print_counters(struct timespec *ts, int argc, const char **argv) 996 { 997 /* Do not print anything if we record to the pipe. */ 998 if (STAT_RECORD && perf_stat.data.is_pipe) 999 return; 1000 if (quiet) 1001 return; 1002 1003 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv); 1004 } 1005 1006 static volatile sig_atomic_t signr = -1; 1007 1008 static void skip_signal(int signo) 1009 { 1010 if ((child_pid == -1) || stat_config.interval) 1011 done = 1; 1012 1013 signr = signo; 1014 /* 1015 * render child_pid harmless 1016 * won't send SIGTERM to a random 1017 * process in case of race condition 1018 * and fast PID recycling 1019 */ 1020 child_pid = -1; 1021 } 1022 1023 static void sig_atexit(void) 1024 { 1025 sigset_t set, oset; 1026 1027 /* 1028 * avoid race condition with SIGCHLD handler 1029 * in skip_signal() which is modifying child_pid 1030 * goal is to avoid send SIGTERM to a random 1031 * process 1032 */ 1033 sigemptyset(&set); 1034 sigaddset(&set, SIGCHLD); 1035 sigprocmask(SIG_BLOCK, &set, &oset); 1036 1037 if (child_pid != -1) 1038 kill(child_pid, SIGTERM); 1039 1040 sigprocmask(SIG_SETMASK, &oset, NULL); 1041 1042 if (signr == -1) 1043 return; 1044 1045 signal(signr, SIG_DFL); 1046 kill(getpid(), signr); 1047 } 1048 1049 void perf_stat__set_big_num(int set) 1050 { 1051 stat_config.big_num = (set != 0); 1052 } 1053 1054 void perf_stat__set_no_csv_summary(int set) 1055 { 1056 stat_config.no_csv_summary = (set != 0); 1057 } 1058 1059 static int stat__set_big_num(const struct option *opt __maybe_unused, 1060 const char *s __maybe_unused, int unset) 1061 { 1062 big_num_opt = unset ? 0 : 1; 1063 perf_stat__set_big_num(!unset); 1064 return 0; 1065 } 1066 1067 static int enable_metric_only(const struct option *opt __maybe_unused, 1068 const char *s __maybe_unused, int unset) 1069 { 1070 force_metric_only = true; 1071 stat_config.metric_only = !unset; 1072 return 0; 1073 } 1074 1075 static int append_metric_groups(const struct option *opt __maybe_unused, 1076 const char *str, 1077 int unset __maybe_unused) 1078 { 1079 if (metrics) { 1080 char *tmp; 1081 1082 if (asprintf(&tmp, "%s,%s", metrics, str) < 0) 1083 return -ENOMEM; 1084 free(metrics); 1085 metrics = tmp; 1086 } else { 1087 metrics = strdup(str); 1088 if (!metrics) 1089 return -ENOMEM; 1090 } 1091 return 0; 1092 } 1093 1094 static int parse_control_option(const struct option *opt, 1095 const char *str, 1096 int unset __maybe_unused) 1097 { 1098 struct perf_stat_config *config = opt->value; 1099 1100 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close); 1101 } 1102 1103 static int parse_stat_cgroups(const struct option *opt, 1104 const char *str, int unset) 1105 { 1106 if (stat_config.cgroup_list) { 1107 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 1108 return -1; 1109 } 1110 1111 return parse_cgroups(opt, str, unset); 1112 } 1113 1114 static int parse_cputype(const struct option *opt, 1115 const char *str, 1116 int unset __maybe_unused) 1117 { 1118 const struct perf_pmu *pmu; 1119 struct evlist *evlist = *(struct evlist **)opt->value; 1120 1121 if (!list_empty(&evlist->core.entries)) { 1122 fprintf(stderr, "Must define cputype before events/metrics\n"); 1123 return -1; 1124 } 1125 1126 pmu = perf_pmus__pmu_for_pmu_filter(str); 1127 if (!pmu) { 1128 fprintf(stderr, "--cputype %s is not supported!\n", str); 1129 return -1; 1130 } 1131 parse_events_option_args.pmu_filter = pmu->name; 1132 1133 return 0; 1134 } 1135 1136 static int parse_cache_level(const struct option *opt, 1137 const char *str, 1138 int unset __maybe_unused) 1139 { 1140 int level; 1141 u32 *aggr_mode = (u32 *)opt->value; 1142 u32 *aggr_level = (u32 *)opt->data; 1143 1144 /* 1145 * If no string is specified, aggregate based on the topology of 1146 * Last Level Cache (LLC). Since the LLC level can change from 1147 * architecture to architecture, set level greater than 1148 * MAX_CACHE_LVL which will be interpreted as LLC. 1149 */ 1150 if (str == NULL) { 1151 level = MAX_CACHE_LVL + 1; 1152 goto out; 1153 } 1154 1155 /* 1156 * The format to specify cache level is LX or lX where X is the 1157 * cache level. 1158 */ 1159 if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) { 1160 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1161 MAX_CACHE_LVL, 1162 MAX_CACHE_LVL); 1163 return -EINVAL; 1164 } 1165 1166 level = atoi(&str[1]); 1167 if (level < 1) { 1168 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n", 1169 MAX_CACHE_LVL, 1170 MAX_CACHE_LVL); 1171 return -EINVAL; 1172 } 1173 1174 if (level > MAX_CACHE_LVL) { 1175 pr_err("perf only supports max cache level of %d.\n" 1176 "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL); 1177 return -EINVAL; 1178 } 1179 out: 1180 *aggr_mode = AGGR_CACHE; 1181 *aggr_level = level; 1182 return 0; 1183 } 1184 1185 static struct option stat_options[] = { 1186 OPT_BOOLEAN('T', "transaction", &transaction_run, 1187 "hardware transaction statistics"), 1188 OPT_CALLBACK('e', "event", &parse_events_option_args, "event", 1189 "event selector. use 'perf list' to list available events", 1190 parse_events_option), 1191 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 1192 "event filter", parse_filter), 1193 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 1194 "child tasks do not inherit counters"), 1195 OPT_STRING('p', "pid", &target.pid, "pid", 1196 "stat events on existing process id"), 1197 OPT_STRING('t', "tid", &target.tid, "tid", 1198 "stat events on existing thread id"), 1199 #ifdef HAVE_BPF_SKEL 1200 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id", 1201 "stat events on existing bpf program id"), 1202 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf, 1203 "use bpf program to count events"), 1204 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path", 1205 "path to perf_event_attr map"), 1206 #endif 1207 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 1208 "system-wide collection from all CPUs"), 1209 OPT_BOOLEAN(0, "scale", &stat_config.scale, 1210 "Use --no-scale to disable counter scaling for multiplexing"), 1211 OPT_INCR('v', "verbose", &verbose, 1212 "be more verbose (show counter open errors, etc)"), 1213 OPT_INTEGER('r', "repeat", &stat_config.run_count, 1214 "repeat command and print average + stddev (max: 100, forever: 0)"), 1215 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 1216 "display details about each run (only with -r option)"), 1217 OPT_BOOLEAN('n', "null", &stat_config.null_run, 1218 "null run - dont start any counters"), 1219 OPT_INCR('d', "detailed", &detailed_run, 1220 "detailed run - start a lot of events"), 1221 OPT_BOOLEAN('S', "sync", &sync_run, 1222 "call sync() before starting a run"), 1223 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1224 "print large numbers with thousands\' separators", 1225 stat__set_big_num), 1226 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 1227 "list of cpus to monitor in system-wide"), 1228 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, 1229 "disable CPU count aggregation", AGGR_NONE), 1230 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"), 1231 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge, 1232 "Merge identical named hybrid events"), 1233 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 1234 "print counts with custom separator"), 1235 OPT_BOOLEAN('j', "json-output", &stat_config.json_output, 1236 "print counts in JSON format"), 1237 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 1238 "monitor event in cgroup name only", parse_stat_cgroups), 1239 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name", 1240 "expand events for each cgroup"), 1241 OPT_STRING('o', "output", &output_name, "file", "output file name"), 1242 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 1243 OPT_INTEGER(0, "log-fd", &output_fd, 1244 "log output to fd, instead of stderr"), 1245 OPT_STRING(0, "pre", &pre_cmd, "command", 1246 "command to run prior to the measured command"), 1247 OPT_STRING(0, "post", &post_cmd, "command", 1248 "command to run after to the measured command"), 1249 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 1250 "print counts at regular interval in ms " 1251 "(overhead is possible for values <= 100ms)"), 1252 OPT_INTEGER(0, "interval-count", &stat_config.times, 1253 "print counts for fixed number of times"), 1254 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 1255 "clear screen in between new interval"), 1256 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 1257 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 1258 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, 1259 "aggregate counts per processor socket", AGGR_SOCKET), 1260 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode, 1261 "aggregate counts per processor die", AGGR_DIE), 1262 OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level, 1263 "cache level", "aggregate count at this cache level (Default: LLC)", 1264 parse_cache_level), 1265 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, 1266 "aggregate counts per physical processor core", AGGR_CORE), 1267 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, 1268 "aggregate counts per thread", AGGR_THREAD), 1269 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode, 1270 "aggregate counts per numa node", AGGR_NODE), 1271 OPT_INTEGER('D', "delay", &target.initial_delay, 1272 "ms to wait before starting measurement after program start (-1: start with events disabled)"), 1273 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 1274 "Only print computed metrics. No raw values", enable_metric_only), 1275 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group, 1276 "don't group metric events, impacts multiplexing"), 1277 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge, 1278 "don't try to share events between metrics in a group"), 1279 OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold, 1280 "don't try to share events between metrics in a group "), 1281 OPT_BOOLEAN(0, "topdown", &topdown_run, 1282 "measure top-down statistics"), 1283 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level, 1284 "Set the metrics level for the top-down statistics (0: max level)"), 1285 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 1286 "measure SMI cost"), 1287 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 1288 "monitor specified metrics or metric groups (separated by ,)", 1289 append_metric_groups), 1290 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, 1291 "Configure all used events to run in kernel space.", 1292 PARSE_OPT_EXCLUSIVE), 1293 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, 1294 "Configure all used events to run in user space.", 1295 PARSE_OPT_EXCLUSIVE), 1296 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread, 1297 "Use with 'percore' event qualifier to show the event " 1298 "counts of one hardware thread by sum up total hardware " 1299 "threads of same physical core"), 1300 OPT_BOOLEAN(0, "summary", &stat_config.summary, 1301 "print summary for interval mode"), 1302 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary, 1303 "don't print 'summary' for CSV summary output"), 1304 OPT_BOOLEAN(0, "quiet", &quiet, 1305 "don't print any output, messages or warnings (useful with record)"), 1306 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type", 1307 "Only enable events on applying cpu with this type " 1308 "for hybrid platform (e.g. core or atom)", 1309 parse_cputype), 1310 #ifdef HAVE_LIBPFM 1311 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event", 1312 "libpfm4 event selector. use 'perf list' to list available events", 1313 parse_libpfm_events_option), 1314 #endif 1315 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]", 1316 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n" 1317 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n" 1318 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.", 1319 parse_control_option), 1320 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default", 1321 "measure I/O performance metrics provided by arch/platform", 1322 iostat_parse), 1323 OPT_END() 1324 }; 1325 1326 /** 1327 * Calculate the cache instance ID from the map in 1328 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1329 * Cache instance ID is the first CPU reported in the shared_cpu_list file. 1330 */ 1331 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map) 1332 { 1333 int id; 1334 struct perf_cpu_map *cpu_map = perf_cpu_map__new(map); 1335 1336 /* 1337 * If the map contains no CPU, consider the current CPU to 1338 * be the first online CPU in the cache domain else use the 1339 * first online CPU of the cache domain as the ID. 1340 */ 1341 if (perf_cpu_map__empty(cpu_map)) 1342 id = cpu.cpu; 1343 else 1344 id = perf_cpu_map__cpu(cpu_map, 0).cpu; 1345 1346 /* Free the perf_cpu_map used to find the cache ID */ 1347 perf_cpu_map__put(cpu_map); 1348 1349 return id; 1350 } 1351 1352 /** 1353 * cpu__get_cache_id - Returns 0 if successful in populating the 1354 * cache level and cache id. Cache level is read from 1355 * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID 1356 * is the first CPU reported by 1357 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list 1358 */ 1359 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache) 1360 { 1361 int ret = 0; 1362 u32 cache_level = stat_config.aggr_level; 1363 struct cpu_cache_level caches[MAX_CACHE_LVL]; 1364 u32 i = 0, caches_cnt = 0; 1365 1366 cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1367 cache->cache = -1; 1368 1369 ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt); 1370 if (ret) { 1371 /* 1372 * If caches_cnt is not 0, cpu_cache_level data 1373 * was allocated when building the topology. 1374 * Free the allocated data before returning. 1375 */ 1376 if (caches_cnt) 1377 goto free_caches; 1378 1379 return ret; 1380 } 1381 1382 if (!caches_cnt) 1383 return -1; 1384 1385 /* 1386 * Save the data for the highest level if no 1387 * level was specified by the user. 1388 */ 1389 if (cache_level > MAX_CACHE_LVL) { 1390 int max_level_index = 0; 1391 1392 for (i = 1; i < caches_cnt; ++i) { 1393 if (caches[i].level > caches[max_level_index].level) 1394 max_level_index = i; 1395 } 1396 1397 cache->cache_lvl = caches[max_level_index].level; 1398 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map); 1399 1400 /* Reset i to 0 to free entire caches[] */ 1401 i = 0; 1402 goto free_caches; 1403 } 1404 1405 for (i = 0; i < caches_cnt; ++i) { 1406 if (caches[i].level == cache_level) { 1407 cache->cache_lvl = cache_level; 1408 cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1409 } 1410 1411 cpu_cache_level__free(&caches[i]); 1412 } 1413 1414 free_caches: 1415 /* 1416 * Free all the allocated cpu_cache_level data. 1417 */ 1418 while (i < caches_cnt) 1419 cpu_cache_level__free(&caches[i++]); 1420 1421 return ret; 1422 } 1423 1424 /** 1425 * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache 1426 * level, die and socket populated with the cache instache ID, cache level, 1427 * die and socket for cpu. The function signature is compatible with 1428 * aggr_cpu_id_get_t. 1429 */ 1430 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data) 1431 { 1432 int ret; 1433 struct aggr_cpu_id id; 1434 struct perf_cache cache; 1435 1436 id = aggr_cpu_id__die(cpu, data); 1437 if (aggr_cpu_id__is_empty(&id)) 1438 return id; 1439 1440 ret = cpu__get_cache_details(cpu, &cache); 1441 if (ret) 1442 return id; 1443 1444 id.cache_lvl = cache.cache_lvl; 1445 id.cache = cache.cache; 1446 return id; 1447 } 1448 1449 static const char *const aggr_mode__string[] = { 1450 [AGGR_CORE] = "core", 1451 [AGGR_CACHE] = "cache", 1452 [AGGR_DIE] = "die", 1453 [AGGR_GLOBAL] = "global", 1454 [AGGR_NODE] = "node", 1455 [AGGR_NONE] = "none", 1456 [AGGR_SOCKET] = "socket", 1457 [AGGR_THREAD] = "thread", 1458 [AGGR_UNSET] = "unset", 1459 }; 1460 1461 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 1462 struct perf_cpu cpu) 1463 { 1464 return aggr_cpu_id__socket(cpu, /*data=*/NULL); 1465 } 1466 1467 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 1468 struct perf_cpu cpu) 1469 { 1470 return aggr_cpu_id__die(cpu, /*data=*/NULL); 1471 } 1472 1473 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused, 1474 struct perf_cpu cpu) 1475 { 1476 return aggr_cpu_id__cache(cpu, /*data=*/NULL); 1477 } 1478 1479 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 1480 struct perf_cpu cpu) 1481 { 1482 return aggr_cpu_id__core(cpu, /*data=*/NULL); 1483 } 1484 1485 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused, 1486 struct perf_cpu cpu) 1487 { 1488 return aggr_cpu_id__node(cpu, /*data=*/NULL); 1489 } 1490 1491 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused, 1492 struct perf_cpu cpu) 1493 { 1494 return aggr_cpu_id__global(cpu, /*data=*/NULL); 1495 } 1496 1497 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused, 1498 struct perf_cpu cpu) 1499 { 1500 return aggr_cpu_id__cpu(cpu, /*data=*/NULL); 1501 } 1502 1503 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config, 1504 aggr_get_id_t get_id, struct perf_cpu cpu) 1505 { 1506 struct aggr_cpu_id id; 1507 1508 /* per-process mode - should use global aggr mode */ 1509 if (cpu.cpu == -1) 1510 return get_id(config, cpu); 1511 1512 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu])) 1513 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu); 1514 1515 id = config->cpus_aggr_map->map[cpu.cpu]; 1516 return id; 1517 } 1518 1519 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config, 1520 struct perf_cpu cpu) 1521 { 1522 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu); 1523 } 1524 1525 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config, 1526 struct perf_cpu cpu) 1527 { 1528 return perf_stat__get_aggr(config, perf_stat__get_die, cpu); 1529 } 1530 1531 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config, 1532 struct perf_cpu cpu) 1533 { 1534 return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu); 1535 } 1536 1537 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config, 1538 struct perf_cpu cpu) 1539 { 1540 return perf_stat__get_aggr(config, perf_stat__get_core, cpu); 1541 } 1542 1543 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config, 1544 struct perf_cpu cpu) 1545 { 1546 return perf_stat__get_aggr(config, perf_stat__get_node, cpu); 1547 } 1548 1549 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config, 1550 struct perf_cpu cpu) 1551 { 1552 return perf_stat__get_aggr(config, perf_stat__get_global, cpu); 1553 } 1554 1555 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config, 1556 struct perf_cpu cpu) 1557 { 1558 return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu); 1559 } 1560 1561 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode) 1562 { 1563 switch (aggr_mode) { 1564 case AGGR_SOCKET: 1565 return aggr_cpu_id__socket; 1566 case AGGR_DIE: 1567 return aggr_cpu_id__die; 1568 case AGGR_CACHE: 1569 return aggr_cpu_id__cache; 1570 case AGGR_CORE: 1571 return aggr_cpu_id__core; 1572 case AGGR_NODE: 1573 return aggr_cpu_id__node; 1574 case AGGR_NONE: 1575 return aggr_cpu_id__cpu; 1576 case AGGR_GLOBAL: 1577 return aggr_cpu_id__global; 1578 case AGGR_THREAD: 1579 case AGGR_UNSET: 1580 case AGGR_MAX: 1581 default: 1582 return NULL; 1583 } 1584 } 1585 1586 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode) 1587 { 1588 switch (aggr_mode) { 1589 case AGGR_SOCKET: 1590 return perf_stat__get_socket_cached; 1591 case AGGR_DIE: 1592 return perf_stat__get_die_cached; 1593 case AGGR_CACHE: 1594 return perf_stat__get_cache_id_cached; 1595 case AGGR_CORE: 1596 return perf_stat__get_core_cached; 1597 case AGGR_NODE: 1598 return perf_stat__get_node_cached; 1599 case AGGR_NONE: 1600 return perf_stat__get_cpu_cached; 1601 case AGGR_GLOBAL: 1602 return perf_stat__get_global_cached; 1603 case AGGR_THREAD: 1604 case AGGR_UNSET: 1605 case AGGR_MAX: 1606 default: 1607 return NULL; 1608 } 1609 } 1610 1611 static int perf_stat_init_aggr_mode(void) 1612 { 1613 int nr; 1614 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode); 1615 1616 if (get_id) { 1617 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1618 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1619 get_id, /*data=*/NULL, needs_sort); 1620 if (!stat_config.aggr_map) { 1621 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1622 return -1; 1623 } 1624 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode); 1625 } 1626 1627 if (stat_config.aggr_mode == AGGR_THREAD) { 1628 nr = perf_thread_map__nr(evsel_list->core.threads); 1629 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1630 if (stat_config.aggr_map == NULL) 1631 return -ENOMEM; 1632 1633 for (int s = 0; s < nr; s++) { 1634 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1635 1636 id.thread_idx = s; 1637 stat_config.aggr_map->map[s] = id; 1638 } 1639 return 0; 1640 } 1641 1642 /* 1643 * The evsel_list->cpus is the base we operate on, 1644 * taking the highest cpu number to be the size of 1645 * the aggregation translate cpumap. 1646 */ 1647 if (!perf_cpu_map__empty(evsel_list->core.user_requested_cpus)) 1648 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu; 1649 else 1650 nr = 0; 1651 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1); 1652 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 1653 } 1654 1655 static void cpu_aggr_map__delete(struct cpu_aggr_map *map) 1656 { 1657 if (map) { 1658 WARN_ONCE(refcount_read(&map->refcnt) != 0, 1659 "cpu_aggr_map refcnt unbalanced\n"); 1660 free(map); 1661 } 1662 } 1663 1664 static void cpu_aggr_map__put(struct cpu_aggr_map *map) 1665 { 1666 if (map && refcount_dec_and_test(&map->refcnt)) 1667 cpu_aggr_map__delete(map); 1668 } 1669 1670 static void perf_stat__exit_aggr_mode(void) 1671 { 1672 cpu_aggr_map__put(stat_config.aggr_map); 1673 cpu_aggr_map__put(stat_config.cpus_aggr_map); 1674 stat_config.aggr_map = NULL; 1675 stat_config.cpus_aggr_map = NULL; 1676 } 1677 1678 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data) 1679 { 1680 struct perf_env *env = data; 1681 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1682 1683 if (cpu.cpu != -1) 1684 id.socket = env->cpu[cpu.cpu].socket_id; 1685 1686 return id; 1687 } 1688 1689 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data) 1690 { 1691 struct perf_env *env = data; 1692 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1693 1694 if (cpu.cpu != -1) { 1695 /* 1696 * die_id is relative to socket, so start 1697 * with the socket ID and then add die to 1698 * make a unique ID. 1699 */ 1700 id.socket = env->cpu[cpu.cpu].socket_id; 1701 id.die = env->cpu[cpu.cpu].die_id; 1702 } 1703 1704 return id; 1705 } 1706 1707 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env, 1708 u32 cache_level, struct aggr_cpu_id *id) 1709 { 1710 int i; 1711 int caches_cnt = env->caches_cnt; 1712 struct cpu_cache_level *caches = env->caches; 1713 1714 id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level; 1715 id->cache = -1; 1716 1717 if (!caches_cnt) 1718 return; 1719 1720 for (i = caches_cnt - 1; i > -1; --i) { 1721 struct perf_cpu_map *cpu_map; 1722 int map_contains_cpu; 1723 1724 /* 1725 * If user has not specified a level, find the fist level with 1726 * the cpu in the map. Since building the map is expensive, do 1727 * this only if levels match. 1728 */ 1729 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level) 1730 continue; 1731 1732 cpu_map = perf_cpu_map__new(caches[i].map); 1733 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu); 1734 perf_cpu_map__put(cpu_map); 1735 1736 if (map_contains_cpu != -1) { 1737 id->cache_lvl = caches[i].level; 1738 id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map); 1739 return; 1740 } 1741 } 1742 } 1743 1744 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu, 1745 void *data) 1746 { 1747 struct perf_env *env = data; 1748 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1749 1750 if (cpu.cpu != -1) { 1751 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level; 1752 1753 id.socket = env->cpu[cpu.cpu].socket_id; 1754 id.die = env->cpu[cpu.cpu].die_id; 1755 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id); 1756 } 1757 1758 return id; 1759 } 1760 1761 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data) 1762 { 1763 struct perf_env *env = data; 1764 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1765 1766 if (cpu.cpu != -1) { 1767 /* 1768 * core_id is relative to socket and die, 1769 * we need a global id. So we set 1770 * socket, die id and core id 1771 */ 1772 id.socket = env->cpu[cpu.cpu].socket_id; 1773 id.die = env->cpu[cpu.cpu].die_id; 1774 id.core = env->cpu[cpu.cpu].core_id; 1775 } 1776 1777 return id; 1778 } 1779 1780 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data) 1781 { 1782 struct perf_env *env = data; 1783 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1784 1785 if (cpu.cpu != -1) { 1786 /* 1787 * core_id is relative to socket and die, 1788 * we need a global id. So we set 1789 * socket, die id and core id 1790 */ 1791 id.socket = env->cpu[cpu.cpu].socket_id; 1792 id.die = env->cpu[cpu.cpu].die_id; 1793 id.core = env->cpu[cpu.cpu].core_id; 1794 id.cpu = cpu; 1795 } 1796 1797 return id; 1798 } 1799 1800 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data) 1801 { 1802 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1803 1804 id.node = perf_env__numa_node(data, cpu); 1805 return id; 1806 } 1807 1808 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused, 1809 void *data __maybe_unused) 1810 { 1811 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1812 1813 /* it always aggregates to the cpu 0 */ 1814 id.cpu = (struct perf_cpu){ .cpu = 0 }; 1815 return id; 1816 } 1817 1818 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1819 struct perf_cpu cpu) 1820 { 1821 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1822 } 1823 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1824 struct perf_cpu cpu) 1825 { 1826 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1827 } 1828 1829 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused, 1830 struct perf_cpu cpu) 1831 { 1832 return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1833 } 1834 1835 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1836 struct perf_cpu cpu) 1837 { 1838 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1839 } 1840 1841 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused, 1842 struct perf_cpu cpu) 1843 { 1844 return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1845 } 1846 1847 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, 1848 struct perf_cpu cpu) 1849 { 1850 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1851 } 1852 1853 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused, 1854 struct perf_cpu cpu) 1855 { 1856 return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env); 1857 } 1858 1859 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode) 1860 { 1861 switch (aggr_mode) { 1862 case AGGR_SOCKET: 1863 return perf_env__get_socket_aggr_by_cpu; 1864 case AGGR_DIE: 1865 return perf_env__get_die_aggr_by_cpu; 1866 case AGGR_CACHE: 1867 return perf_env__get_cache_aggr_by_cpu; 1868 case AGGR_CORE: 1869 return perf_env__get_core_aggr_by_cpu; 1870 case AGGR_NODE: 1871 return perf_env__get_node_aggr_by_cpu; 1872 case AGGR_GLOBAL: 1873 return perf_env__get_global_aggr_by_cpu; 1874 case AGGR_NONE: 1875 return perf_env__get_cpu_aggr_by_cpu; 1876 case AGGR_THREAD: 1877 case AGGR_UNSET: 1878 case AGGR_MAX: 1879 default: 1880 return NULL; 1881 } 1882 } 1883 1884 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode) 1885 { 1886 switch (aggr_mode) { 1887 case AGGR_SOCKET: 1888 return perf_stat__get_socket_file; 1889 case AGGR_DIE: 1890 return perf_stat__get_die_file; 1891 case AGGR_CACHE: 1892 return perf_stat__get_cache_file; 1893 case AGGR_CORE: 1894 return perf_stat__get_core_file; 1895 case AGGR_NODE: 1896 return perf_stat__get_node_file; 1897 case AGGR_GLOBAL: 1898 return perf_stat__get_global_file; 1899 case AGGR_NONE: 1900 return perf_stat__get_cpu_file; 1901 case AGGR_THREAD: 1902 case AGGR_UNSET: 1903 case AGGR_MAX: 1904 default: 1905 return NULL; 1906 } 1907 } 1908 1909 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1910 { 1911 struct perf_env *env = &st->session->header.env; 1912 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode); 1913 bool needs_sort = stat_config.aggr_mode != AGGR_NONE; 1914 1915 if (stat_config.aggr_mode == AGGR_THREAD) { 1916 int nr = perf_thread_map__nr(evsel_list->core.threads); 1917 1918 stat_config.aggr_map = cpu_aggr_map__empty_new(nr); 1919 if (stat_config.aggr_map == NULL) 1920 return -ENOMEM; 1921 1922 for (int s = 0; s < nr; s++) { 1923 struct aggr_cpu_id id = aggr_cpu_id__empty(); 1924 1925 id.thread_idx = s; 1926 stat_config.aggr_map->map[s] = id; 1927 } 1928 return 0; 1929 } 1930 1931 if (!get_id) 1932 return 0; 1933 1934 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, 1935 get_id, env, needs_sort); 1936 if (!stat_config.aggr_map) { 1937 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]); 1938 return -1; 1939 } 1940 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode); 1941 return 0; 1942 } 1943 1944 /* 1945 * Add default attributes, if there were no attributes specified or 1946 * if -d/--detailed, -d -d or -d -d -d is used: 1947 */ 1948 static int add_default_attributes(void) 1949 { 1950 struct perf_event_attr default_attrs0[] = { 1951 1952 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 1953 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 1954 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 1955 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 1956 1957 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 1958 }; 1959 struct perf_event_attr frontend_attrs[] = { 1960 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 1961 }; 1962 struct perf_event_attr backend_attrs[] = { 1963 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 1964 }; 1965 struct perf_event_attr default_attrs1[] = { 1966 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 1967 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 1968 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 1969 1970 }; 1971 1972 /* 1973 * Detailed stats (-d), covering the L1 and last level data caches: 1974 */ 1975 struct perf_event_attr detailed_attrs[] = { 1976 1977 { .type = PERF_TYPE_HW_CACHE, 1978 .config = 1979 PERF_COUNT_HW_CACHE_L1D << 0 | 1980 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1981 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1982 1983 { .type = PERF_TYPE_HW_CACHE, 1984 .config = 1985 PERF_COUNT_HW_CACHE_L1D << 0 | 1986 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1987 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1988 1989 { .type = PERF_TYPE_HW_CACHE, 1990 .config = 1991 PERF_COUNT_HW_CACHE_LL << 0 | 1992 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1993 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1994 1995 { .type = PERF_TYPE_HW_CACHE, 1996 .config = 1997 PERF_COUNT_HW_CACHE_LL << 0 | 1998 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1999 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 2000 }; 2001 2002 /* 2003 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 2004 */ 2005 struct perf_event_attr very_detailed_attrs[] = { 2006 2007 { .type = PERF_TYPE_HW_CACHE, 2008 .config = 2009 PERF_COUNT_HW_CACHE_L1I << 0 | 2010 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 2011 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 2012 2013 { .type = PERF_TYPE_HW_CACHE, 2014 .config = 2015 PERF_COUNT_HW_CACHE_L1I << 0 | 2016 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 2017 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 2018 2019 { .type = PERF_TYPE_HW_CACHE, 2020 .config = 2021 PERF_COUNT_HW_CACHE_DTLB << 0 | 2022 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 2023 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 2024 2025 { .type = PERF_TYPE_HW_CACHE, 2026 .config = 2027 PERF_COUNT_HW_CACHE_DTLB << 0 | 2028 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 2029 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 2030 2031 { .type = PERF_TYPE_HW_CACHE, 2032 .config = 2033 PERF_COUNT_HW_CACHE_ITLB << 0 | 2034 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 2035 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 2036 2037 { .type = PERF_TYPE_HW_CACHE, 2038 .config = 2039 PERF_COUNT_HW_CACHE_ITLB << 0 | 2040 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 2041 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 2042 2043 }; 2044 2045 /* 2046 * Very, very detailed stats (-d -d -d), adding prefetch events: 2047 */ 2048 struct perf_event_attr very_very_detailed_attrs[] = { 2049 2050 { .type = PERF_TYPE_HW_CACHE, 2051 .config = 2052 PERF_COUNT_HW_CACHE_L1D << 0 | 2053 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 2054 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 2055 2056 { .type = PERF_TYPE_HW_CACHE, 2057 .config = 2058 PERF_COUNT_HW_CACHE_L1D << 0 | 2059 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 2060 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 2061 }; 2062 2063 struct perf_event_attr default_null_attrs[] = {}; 2064 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 2065 2066 /* Set attrs if no event is selected and !null_run: */ 2067 if (stat_config.null_run) 2068 return 0; 2069 2070 if (transaction_run) { 2071 /* Handle -T as -M transaction. Once platform specific metrics 2072 * support has been added to the json files, all architectures 2073 * will use this approach. To determine transaction support 2074 * on an architecture test for such a metric name. 2075 */ 2076 if (!metricgroup__has_metric(pmu, "transaction")) { 2077 pr_err("Missing transaction metrics\n"); 2078 return -1; 2079 } 2080 return metricgroup__parse_groups(evsel_list, pmu, "transaction", 2081 stat_config.metric_no_group, 2082 stat_config.metric_no_merge, 2083 stat_config.metric_no_threshold, 2084 stat_config.user_requested_cpu_list, 2085 stat_config.system_wide, 2086 &stat_config.metric_events); 2087 } 2088 2089 if (smi_cost) { 2090 int smi; 2091 2092 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 2093 pr_err("freeze_on_smi is not supported.\n"); 2094 return -1; 2095 } 2096 2097 if (!smi) { 2098 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 2099 fprintf(stderr, "Failed to set freeze_on_smi.\n"); 2100 return -1; 2101 } 2102 smi_reset = true; 2103 } 2104 2105 if (!metricgroup__has_metric(pmu, "smi")) { 2106 pr_err("Missing smi metrics\n"); 2107 return -1; 2108 } 2109 2110 if (!force_metric_only) 2111 stat_config.metric_only = true; 2112 2113 return metricgroup__parse_groups(evsel_list, pmu, "smi", 2114 stat_config.metric_no_group, 2115 stat_config.metric_no_merge, 2116 stat_config.metric_no_threshold, 2117 stat_config.user_requested_cpu_list, 2118 stat_config.system_wide, 2119 &stat_config.metric_events); 2120 } 2121 2122 if (topdown_run) { 2123 unsigned int max_level = metricgroups__topdown_max_level(); 2124 char str[] = "TopdownL1"; 2125 2126 if (!force_metric_only) 2127 stat_config.metric_only = true; 2128 2129 if (!max_level) { 2130 pr_err("Topdown requested but the topdown metric groups aren't present.\n" 2131 "(See perf list the metric groups have names like TopdownL1)\n"); 2132 return -1; 2133 } 2134 if (stat_config.topdown_level > max_level) { 2135 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level); 2136 return -1; 2137 } else if (!stat_config.topdown_level) 2138 stat_config.topdown_level = 1; 2139 2140 if (!stat_config.interval && !stat_config.metric_only) { 2141 fprintf(stat_config.output, 2142 "Topdown accuracy may decrease when measuring long periods.\n" 2143 "Please print the result regularly, e.g. -I1000\n"); 2144 } 2145 str[8] = stat_config.topdown_level + '0'; 2146 if (metricgroup__parse_groups(evsel_list, 2147 pmu, str, 2148 /*metric_no_group=*/false, 2149 /*metric_no_merge=*/false, 2150 /*metric_no_threshold=*/true, 2151 stat_config.user_requested_cpu_list, 2152 stat_config.system_wide, 2153 &stat_config.metric_events) < 0) 2154 return -1; 2155 } 2156 2157 if (!stat_config.topdown_level) 2158 stat_config.topdown_level = 1; 2159 2160 if (!evsel_list->core.nr_entries) { 2161 /* No events so add defaults. */ 2162 if (target__has_cpu(&target)) 2163 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; 2164 2165 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0) 2166 return -1; 2167 if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) { 2168 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0) 2169 return -1; 2170 } 2171 if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) { 2172 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0) 2173 return -1; 2174 } 2175 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0) 2176 return -1; 2177 /* 2178 * Add TopdownL1 metrics if they exist. To minimize 2179 * multiplexing, don't request threshold computation. 2180 */ 2181 if (metricgroup__has_metric(pmu, "Default")) { 2182 struct evlist *metric_evlist = evlist__new(); 2183 struct evsel *metric_evsel; 2184 2185 if (!metric_evlist) 2186 return -1; 2187 2188 if (metricgroup__parse_groups(metric_evlist, pmu, "Default", 2189 /*metric_no_group=*/false, 2190 /*metric_no_merge=*/false, 2191 /*metric_no_threshold=*/true, 2192 stat_config.user_requested_cpu_list, 2193 stat_config.system_wide, 2194 &stat_config.metric_events) < 0) 2195 return -1; 2196 2197 evlist__for_each_entry(metric_evlist, metric_evsel) { 2198 metric_evsel->skippable = true; 2199 metric_evsel->default_metricgroup = true; 2200 } 2201 evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries); 2202 evlist__delete(metric_evlist); 2203 } 2204 2205 /* Platform specific attrs */ 2206 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0) 2207 return -1; 2208 } 2209 2210 /* Detailed events get appended to the event list: */ 2211 2212 if (detailed_run < 1) 2213 return 0; 2214 2215 /* Append detailed run extra attributes: */ 2216 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) 2217 return -1; 2218 2219 if (detailed_run < 2) 2220 return 0; 2221 2222 /* Append very detailed run extra attributes: */ 2223 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) 2224 return -1; 2225 2226 if (detailed_run < 3) 2227 return 0; 2228 2229 /* Append very, very detailed run extra attributes: */ 2230 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); 2231 } 2232 2233 static const char * const stat_record_usage[] = { 2234 "perf stat record [<options>]", 2235 NULL, 2236 }; 2237 2238 static void init_features(struct perf_session *session) 2239 { 2240 int feat; 2241 2242 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 2243 perf_header__set_feat(&session->header, feat); 2244 2245 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 2246 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 2247 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 2248 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 2249 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 2250 } 2251 2252 static int __cmd_record(int argc, const char **argv) 2253 { 2254 struct perf_session *session; 2255 struct perf_data *data = &perf_stat.data; 2256 2257 argc = parse_options(argc, argv, stat_options, stat_record_usage, 2258 PARSE_OPT_STOP_AT_NON_OPTION); 2259 2260 if (output_name) 2261 data->path = output_name; 2262 2263 if (stat_config.run_count != 1 || forever) { 2264 pr_err("Cannot use -r option with perf stat record.\n"); 2265 return -1; 2266 } 2267 2268 session = perf_session__new(data, NULL); 2269 if (IS_ERR(session)) { 2270 pr_err("Perf session creation failed\n"); 2271 return PTR_ERR(session); 2272 } 2273 2274 init_features(session); 2275 2276 session->evlist = evsel_list; 2277 perf_stat.session = session; 2278 perf_stat.record = true; 2279 return argc; 2280 } 2281 2282 static int process_stat_round_event(struct perf_session *session, 2283 union perf_event *event) 2284 { 2285 struct perf_record_stat_round *stat_round = &event->stat_round; 2286 struct timespec tsh, *ts = NULL; 2287 const char **argv = session->header.env.cmdline_argv; 2288 int argc = session->header.env.nr_cmdline; 2289 2290 process_counters(); 2291 2292 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 2293 update_stats(&walltime_nsecs_stats, stat_round->time); 2294 2295 if (stat_config.interval && stat_round->time) { 2296 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 2297 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 2298 ts = &tsh; 2299 } 2300 2301 print_counters(ts, argc, argv); 2302 return 0; 2303 } 2304 2305 static 2306 int process_stat_config_event(struct perf_session *session, 2307 union perf_event *event) 2308 { 2309 struct perf_tool *tool = session->tool; 2310 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2311 2312 perf_event__read_stat_config(&stat_config, &event->stat_config); 2313 2314 if (perf_cpu_map__empty(st->cpus)) { 2315 if (st->aggr_mode != AGGR_UNSET) 2316 pr_warning("warning: processing task data, aggregation mode not set\n"); 2317 } else if (st->aggr_mode != AGGR_UNSET) { 2318 stat_config.aggr_mode = st->aggr_mode; 2319 } 2320 2321 if (perf_stat.data.is_pipe) 2322 perf_stat_init_aggr_mode(); 2323 else 2324 perf_stat_init_aggr_mode_file(st); 2325 2326 if (stat_config.aggr_map) { 2327 int nr_aggr = stat_config.aggr_map->nr; 2328 2329 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) { 2330 pr_err("cannot allocate aggr counts\n"); 2331 return -1; 2332 } 2333 } 2334 return 0; 2335 } 2336 2337 static int set_maps(struct perf_stat *st) 2338 { 2339 if (!st->cpus || !st->threads) 2340 return 0; 2341 2342 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 2343 return -EINVAL; 2344 2345 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 2346 2347 if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true)) 2348 return -ENOMEM; 2349 2350 st->maps_allocated = true; 2351 return 0; 2352 } 2353 2354 static 2355 int process_thread_map_event(struct perf_session *session, 2356 union perf_event *event) 2357 { 2358 struct perf_tool *tool = session->tool; 2359 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2360 2361 if (st->threads) { 2362 pr_warning("Extra thread map event, ignoring.\n"); 2363 return 0; 2364 } 2365 2366 st->threads = thread_map__new_event(&event->thread_map); 2367 if (!st->threads) 2368 return -ENOMEM; 2369 2370 return set_maps(st); 2371 } 2372 2373 static 2374 int process_cpu_map_event(struct perf_session *session, 2375 union perf_event *event) 2376 { 2377 struct perf_tool *tool = session->tool; 2378 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2379 struct perf_cpu_map *cpus; 2380 2381 if (st->cpus) { 2382 pr_warning("Extra cpu map event, ignoring.\n"); 2383 return 0; 2384 } 2385 2386 cpus = cpu_map__new_data(&event->cpu_map.data); 2387 if (!cpus) 2388 return -ENOMEM; 2389 2390 st->cpus = cpus; 2391 return set_maps(st); 2392 } 2393 2394 static const char * const stat_report_usage[] = { 2395 "perf stat report [<options>]", 2396 NULL, 2397 }; 2398 2399 static struct perf_stat perf_stat = { 2400 .tool = { 2401 .attr = perf_event__process_attr, 2402 .event_update = perf_event__process_event_update, 2403 .thread_map = process_thread_map_event, 2404 .cpu_map = process_cpu_map_event, 2405 .stat_config = process_stat_config_event, 2406 .stat = perf_event__process_stat_event, 2407 .stat_round = process_stat_round_event, 2408 }, 2409 .aggr_mode = AGGR_UNSET, 2410 .aggr_level = 0, 2411 }; 2412 2413 static int __cmd_report(int argc, const char **argv) 2414 { 2415 struct perf_session *session; 2416 const struct option options[] = { 2417 OPT_STRING('i', "input", &input_name, "file", "input file name"), 2418 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 2419 "aggregate counts per processor socket", AGGR_SOCKET), 2420 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 2421 "aggregate counts per processor die", AGGR_DIE), 2422 OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level, 2423 "cache level", 2424 "aggregate count at this cache level (Default: LLC)", 2425 parse_cache_level), 2426 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 2427 "aggregate counts per physical processor core", AGGR_CORE), 2428 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, 2429 "aggregate counts per numa node", AGGR_NODE), 2430 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 2431 "disable CPU count aggregation", AGGR_NONE), 2432 OPT_END() 2433 }; 2434 struct stat st; 2435 int ret; 2436 2437 argc = parse_options(argc, argv, options, stat_report_usage, 0); 2438 2439 if (!input_name || !strlen(input_name)) { 2440 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 2441 input_name = "-"; 2442 else 2443 input_name = "perf.data"; 2444 } 2445 2446 perf_stat.data.path = input_name; 2447 perf_stat.data.mode = PERF_DATA_MODE_READ; 2448 2449 session = perf_session__new(&perf_stat.data, &perf_stat.tool); 2450 if (IS_ERR(session)) 2451 return PTR_ERR(session); 2452 2453 perf_stat.session = session; 2454 stat_config.output = stderr; 2455 evlist__delete(evsel_list); 2456 evsel_list = session->evlist; 2457 2458 ret = perf_session__process_events(session); 2459 if (ret) 2460 return ret; 2461 2462 perf_session__delete(session); 2463 return 0; 2464 } 2465 2466 static void setup_system_wide(int forks) 2467 { 2468 /* 2469 * Make system wide (-a) the default target if 2470 * no target was specified and one of following 2471 * conditions is met: 2472 * 2473 * - there's no workload specified 2474 * - there is workload specified but all requested 2475 * events are system wide events 2476 */ 2477 if (!target__none(&target)) 2478 return; 2479 2480 if (!forks) 2481 target.system_wide = true; 2482 else { 2483 struct evsel *counter; 2484 2485 evlist__for_each_entry(evsel_list, counter) { 2486 if (!counter->core.requires_cpu && 2487 !evsel__name_is(counter, "duration_time")) { 2488 return; 2489 } 2490 } 2491 2492 if (evsel_list->core.nr_entries) 2493 target.system_wide = true; 2494 } 2495 } 2496 2497 int cmd_stat(int argc, const char **argv) 2498 { 2499 const char * const stat_usage[] = { 2500 "perf stat [<options>] [<command>]", 2501 NULL 2502 }; 2503 int status = -EINVAL, run_idx, err; 2504 const char *mode; 2505 FILE *output = stderr; 2506 unsigned int interval, timeout; 2507 const char * const stat_subcommands[] = { "record", "report" }; 2508 char errbuf[BUFSIZ]; 2509 2510 setlocale(LC_ALL, ""); 2511 2512 evsel_list = evlist__new(); 2513 if (evsel_list == NULL) 2514 return -ENOMEM; 2515 2516 parse_events__shrink_config_terms(); 2517 2518 /* String-parsing callback-based options would segfault when negated */ 2519 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 2520 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 2521 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 2522 2523 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 2524 (const char **) stat_usage, 2525 PARSE_OPT_STOP_AT_NON_OPTION); 2526 2527 if (stat_config.csv_sep) { 2528 stat_config.csv_output = true; 2529 if (!strcmp(stat_config.csv_sep, "\\t")) 2530 stat_config.csv_sep = "\t"; 2531 } else 2532 stat_config.csv_sep = DEFAULT_SEPARATOR; 2533 2534 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2535 argc = __cmd_record(argc, argv); 2536 if (argc < 0) 2537 return -1; 2538 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0])) 2539 return __cmd_report(argc, argv); 2540 2541 interval = stat_config.interval; 2542 timeout = stat_config.timeout; 2543 2544 /* 2545 * For record command the -o is already taken care of. 2546 */ 2547 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 2548 output = NULL; 2549 2550 if (output_name && output_fd) { 2551 fprintf(stderr, "cannot use both --output and --log-fd\n"); 2552 parse_options_usage(stat_usage, stat_options, "o", 1); 2553 parse_options_usage(NULL, stat_options, "log-fd", 0); 2554 goto out; 2555 } 2556 2557 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 2558 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 2559 goto out; 2560 } 2561 2562 if (stat_config.metric_only && stat_config.run_count > 1) { 2563 fprintf(stderr, "--metric-only is not supported with -r\n"); 2564 goto out; 2565 } 2566 2567 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 2568 fprintf(stderr, "--table is only supported with -r\n"); 2569 parse_options_usage(stat_usage, stat_options, "r", 1); 2570 parse_options_usage(NULL, stat_options, "table", 0); 2571 goto out; 2572 } 2573 2574 if (output_fd < 0) { 2575 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 2576 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 2577 goto out; 2578 } 2579 2580 if (!output && !quiet) { 2581 struct timespec tm; 2582 mode = append_file ? "a" : "w"; 2583 2584 output = fopen(output_name, mode); 2585 if (!output) { 2586 perror("failed to create output file"); 2587 return -1; 2588 } 2589 if (!stat_config.json_output) { 2590 clock_gettime(CLOCK_REALTIME, &tm); 2591 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 2592 } 2593 } else if (output_fd > 0) { 2594 mode = append_file ? "a" : "w"; 2595 output = fdopen(output_fd, mode); 2596 if (!output) { 2597 perror("Failed opening logfd"); 2598 return -errno; 2599 } 2600 } 2601 2602 if (stat_config.interval_clear && !isatty(fileno(output))) { 2603 fprintf(stderr, "--interval-clear does not work with output\n"); 2604 parse_options_usage(stat_usage, stat_options, "o", 1); 2605 parse_options_usage(NULL, stat_options, "log-fd", 0); 2606 parse_options_usage(NULL, stat_options, "interval-clear", 0); 2607 return -1; 2608 } 2609 2610 stat_config.output = output; 2611 2612 /* 2613 * let the spreadsheet do the pretty-printing 2614 */ 2615 if (stat_config.csv_output) { 2616 /* User explicitly passed -B? */ 2617 if (big_num_opt == 1) { 2618 fprintf(stderr, "-B option not supported with -x\n"); 2619 parse_options_usage(stat_usage, stat_options, "B", 1); 2620 parse_options_usage(NULL, stat_options, "x", 1); 2621 goto out; 2622 } else /* Nope, so disable big number formatting */ 2623 stat_config.big_num = false; 2624 } else if (big_num_opt == 0) /* User passed --no-big-num */ 2625 stat_config.big_num = false; 2626 2627 err = target__validate(&target); 2628 if (err) { 2629 target__strerror(&target, err, errbuf, BUFSIZ); 2630 pr_warning("%s\n", errbuf); 2631 } 2632 2633 setup_system_wide(argc); 2634 2635 /* 2636 * Display user/system times only for single 2637 * run and when there's specified tracee. 2638 */ 2639 if ((stat_config.run_count == 1) && target__none(&target)) 2640 stat_config.ru_display = true; 2641 2642 if (stat_config.run_count < 0) { 2643 pr_err("Run count must be a positive number\n"); 2644 parse_options_usage(stat_usage, stat_options, "r", 1); 2645 goto out; 2646 } else if (stat_config.run_count == 0) { 2647 forever = true; 2648 stat_config.run_count = 1; 2649 } 2650 2651 if (stat_config.walltime_run_table) { 2652 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 2653 if (!stat_config.walltime_run) { 2654 pr_err("failed to setup -r option"); 2655 goto out; 2656 } 2657 } 2658 2659 if ((stat_config.aggr_mode == AGGR_THREAD) && 2660 !target__has_task(&target)) { 2661 if (!target.system_wide || target.cpu_list) { 2662 fprintf(stderr, "The --per-thread option is only " 2663 "available when monitoring via -p -t -a " 2664 "options or only --per-thread.\n"); 2665 parse_options_usage(NULL, stat_options, "p", 1); 2666 parse_options_usage(NULL, stat_options, "t", 1); 2667 goto out; 2668 } 2669 } 2670 2671 /* 2672 * no_aggr, cgroup are for system-wide only 2673 * --per-thread is aggregated per thread, we dont mix it with cpu mode 2674 */ 2675 if (((stat_config.aggr_mode != AGGR_GLOBAL && 2676 stat_config.aggr_mode != AGGR_THREAD) || 2677 (nr_cgroups || stat_config.cgroup_list)) && 2678 !target__has_cpu(&target)) { 2679 fprintf(stderr, "both cgroup and no-aggregation " 2680 "modes only available in system-wide mode\n"); 2681 2682 parse_options_usage(stat_usage, stat_options, "G", 1); 2683 parse_options_usage(NULL, stat_options, "A", 1); 2684 parse_options_usage(NULL, stat_options, "a", 1); 2685 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2686 goto out; 2687 } 2688 2689 if (stat_config.iostat_run) { 2690 status = iostat_prepare(evsel_list, &stat_config); 2691 if (status) 2692 goto out; 2693 if (iostat_mode == IOSTAT_LIST) { 2694 iostat_list(evsel_list, &stat_config); 2695 goto out; 2696 } else if (verbose > 0) 2697 iostat_list(evsel_list, &stat_config); 2698 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target)) 2699 target.system_wide = true; 2700 } 2701 2702 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 2703 target.per_thread = true; 2704 2705 stat_config.system_wide = target.system_wide; 2706 if (target.cpu_list) { 2707 stat_config.user_requested_cpu_list = strdup(target.cpu_list); 2708 if (!stat_config.user_requested_cpu_list) { 2709 status = -ENOMEM; 2710 goto out; 2711 } 2712 } 2713 2714 /* 2715 * Metric parsing needs to be delayed as metrics may optimize events 2716 * knowing the target is system-wide. 2717 */ 2718 if (metrics) { 2719 const char *pmu = parse_events_option_args.pmu_filter ?: "all"; 2720 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics, 2721 stat_config.metric_no_group, 2722 stat_config.metric_no_merge, 2723 stat_config.metric_no_threshold, 2724 stat_config.user_requested_cpu_list, 2725 stat_config.system_wide, 2726 &stat_config.metric_events); 2727 2728 zfree(&metrics); 2729 if (ret) { 2730 status = ret; 2731 goto out; 2732 } 2733 } 2734 2735 if (add_default_attributes()) 2736 goto out; 2737 2738 if (stat_config.cgroup_list) { 2739 if (nr_cgroups > 0) { 2740 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 2741 parse_options_usage(stat_usage, stat_options, "G", 1); 2742 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2743 goto out; 2744 } 2745 2746 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, 2747 &stat_config.metric_events, true) < 0) { 2748 parse_options_usage(stat_usage, stat_options, 2749 "for-each-cgroup", 0); 2750 goto out; 2751 } 2752 } 2753 2754 evlist__warn_user_requested_cpus(evsel_list, target.cpu_list); 2755 2756 if (evlist__create_maps(evsel_list, &target) < 0) { 2757 if (target__has_task(&target)) { 2758 pr_err("Problems finding threads of monitor\n"); 2759 parse_options_usage(stat_usage, stat_options, "p", 1); 2760 parse_options_usage(NULL, stat_options, "t", 1); 2761 } else if (target__has_cpu(&target)) { 2762 perror("failed to parse CPUs map"); 2763 parse_options_usage(stat_usage, stat_options, "C", 1); 2764 parse_options_usage(NULL, stat_options, "a", 1); 2765 } 2766 goto out; 2767 } 2768 2769 evlist__check_cpu_maps(evsel_list); 2770 2771 /* 2772 * Initialize thread_map with comm names, 2773 * so we could print it out on output. 2774 */ 2775 if (stat_config.aggr_mode == AGGR_THREAD) { 2776 thread_map__read_comms(evsel_list->core.threads); 2777 } 2778 2779 if (stat_config.aggr_mode == AGGR_NODE) 2780 cpu__setup_cpunode_map(); 2781 2782 if (stat_config.times && interval) 2783 interval_count = true; 2784 else if (stat_config.times && !interval) { 2785 pr_err("interval-count option should be used together with " 2786 "interval-print.\n"); 2787 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 2788 parse_options_usage(stat_usage, stat_options, "I", 1); 2789 goto out; 2790 } 2791 2792 if (timeout && timeout < 100) { 2793 if (timeout < 10) { 2794 pr_err("timeout must be >= 10ms.\n"); 2795 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2796 goto out; 2797 } else 2798 pr_warning("timeout < 100ms. " 2799 "The overhead percentage could be high in some cases. " 2800 "Please proceed with caution.\n"); 2801 } 2802 if (timeout && interval) { 2803 pr_err("timeout option is not supported with interval-print.\n"); 2804 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2805 parse_options_usage(stat_usage, stat_options, "I", 1); 2806 goto out; 2807 } 2808 2809 if (perf_stat_init_aggr_mode()) 2810 goto out; 2811 2812 if (evlist__alloc_stats(&stat_config, evsel_list, interval)) 2813 goto out; 2814 2815 /* 2816 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 2817 * while avoiding that older tools show confusing messages. 2818 * 2819 * However for pipe sessions we need to keep it zero, 2820 * because script's perf_evsel__check_attr is triggered 2821 * by attr->sample_type != 0, and we can't run it on 2822 * stat sessions. 2823 */ 2824 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 2825 2826 /* 2827 * We dont want to block the signals - that would cause 2828 * child tasks to inherit that and Ctrl-C would not work. 2829 * What we want is for Ctrl-C to work in the exec()-ed 2830 * task, but being ignored by perf stat itself: 2831 */ 2832 atexit(sig_atexit); 2833 if (!forever) 2834 signal(SIGINT, skip_signal); 2835 signal(SIGCHLD, skip_signal); 2836 signal(SIGALRM, skip_signal); 2837 signal(SIGABRT, skip_signal); 2838 2839 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack)) 2840 goto out; 2841 2842 /* Enable ignoring missing threads when -p option is defined. */ 2843 evlist__first(evsel_list)->ignore_missing_thread = target.pid; 2844 status = 0; 2845 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 2846 if (stat_config.run_count != 1 && verbose > 0) 2847 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 2848 run_idx + 1); 2849 2850 if (run_idx != 0) 2851 evlist__reset_prev_raw_counts(evsel_list); 2852 2853 status = run_perf_stat(argc, argv, run_idx); 2854 if (forever && status != -1 && !interval) { 2855 print_counters(NULL, argc, argv); 2856 perf_stat__reset_stats(); 2857 } 2858 } 2859 2860 if (!forever && status != -1 && (!interval || stat_config.summary)) { 2861 if (stat_config.run_count > 1) 2862 evlist__copy_res_stats(&stat_config, evsel_list); 2863 print_counters(NULL, argc, argv); 2864 } 2865 2866 evlist__finalize_ctlfd(evsel_list); 2867 2868 if (STAT_RECORD) { 2869 /* 2870 * We synthesize the kernel mmap record just so that older tools 2871 * don't emit warnings about not being able to resolve symbols 2872 * due to /proc/sys/kernel/kptr_restrict settings and instead provide 2873 * a saner message about no samples being in the perf.data file. 2874 * 2875 * This also serves to suppress a warning about f_header.data.size == 0 2876 * in header.c at the moment 'perf stat record' gets introduced, which 2877 * is not really needed once we start adding the stat specific PERF_RECORD_ 2878 * records, but the need to suppress the kptr_restrict messages in older 2879 * tools remain -acme 2880 */ 2881 int fd = perf_data__fd(&perf_stat.data); 2882 2883 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 2884 process_synthesized_event, 2885 &perf_stat.session->machines.host); 2886 if (err) { 2887 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 2888 "older tools may produce warnings about this file\n."); 2889 } 2890 2891 if (!interval) { 2892 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) 2893 pr_err("failed to write stat round event\n"); 2894 } 2895 2896 if (!perf_stat.data.is_pipe) { 2897 perf_stat.session->header.data_size += perf_stat.bytes_written; 2898 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 2899 } 2900 2901 evlist__close(evsel_list); 2902 perf_session__delete(perf_stat.session); 2903 } 2904 2905 perf_stat__exit_aggr_mode(); 2906 evlist__free_stats(evsel_list); 2907 out: 2908 if (stat_config.iostat_run) 2909 iostat_release(evsel_list); 2910 2911 zfree(&stat_config.walltime_run); 2912 zfree(&stat_config.user_requested_cpu_list); 2913 2914 if (smi_cost && smi_reset) 2915 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 2916 2917 evlist__delete(evsel_list); 2918 2919 metricgroup__rblist_exit(&stat_config.metric_events); 2920 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close); 2921 2922 return status; 2923 } 2924