1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * builtin-stat.c 4 * 5 * Builtin stat command: Give a precise performance counters summary 6 * overview about any workload, CPU or specific PID. 7 * 8 * Sample output: 9 10 $ perf stat ./hackbench 10 11 12 Time: 0.118 13 14 Performance counter stats for './hackbench 10': 15 16 1708.761321 task-clock # 11.037 CPUs utilized 17 41,190 context-switches # 0.024 M/sec 18 6,735 CPU-migrations # 0.004 M/sec 19 17,318 page-faults # 0.010 M/sec 20 5,205,202,243 cycles # 3.046 GHz 21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle 22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle 23 2,603,501,247 instructions # 0.50 insns per cycle 24 # 1.48 stalled cycles per insn 25 484,357,498 branches # 283.455 M/sec 26 6,388,934 branch-misses # 1.32% of all branches 27 28 0.154822978 seconds time elapsed 29 30 * 31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 32 * 33 * Improvements and fixes by: 34 * 35 * Arjan van de Ven <arjan@linux.intel.com> 36 * Yanmin Zhang <yanmin.zhang@intel.com> 37 * Wu Fengguang <fengguang.wu@intel.com> 38 * Mike Galbraith <efault@gmx.de> 39 * Paul Mackerras <paulus@samba.org> 40 * Jaswinder Singh Rajput <jaswinder@kernel.org> 41 */ 42 43 #include "builtin.h" 44 #include "perf.h" 45 #include "util/cgroup.h" 46 #include <subcmd/parse-options.h> 47 #include "util/parse-events.h" 48 #include "util/pmu.h" 49 #include "util/event.h" 50 #include "util/evlist.h" 51 #include "util/evlist-hybrid.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/pmu-hybrid.h" 74 #include "asm/bug.h" 75 76 #include <linux/time64.h> 77 #include <linux/zalloc.h> 78 #include <api/fs/fs.h> 79 #include <errno.h> 80 #include <signal.h> 81 #include <stdlib.h> 82 #include <sys/prctl.h> 83 #include <inttypes.h> 84 #include <locale.h> 85 #include <math.h> 86 #include <sys/types.h> 87 #include <sys/stat.h> 88 #include <sys/wait.h> 89 #include <unistd.h> 90 #include <sys/time.h> 91 #include <sys/resource.h> 92 #include <linux/err.h> 93 94 #include <linux/ctype.h> 95 #include <perf/evlist.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 /* Default events used for perf stat -T */ 103 static const char *transaction_attrs = { 104 "task-clock," 105 "{" 106 "instructions," 107 "cycles," 108 "cpu/cycles-t/," 109 "cpu/tx-start/," 110 "cpu/el-start/," 111 "cpu/cycles-ct/" 112 "}" 113 }; 114 115 /* More limited version when the CPU does not have all events. */ 116 static const char * transaction_limited_attrs = { 117 "task-clock," 118 "{" 119 "instructions," 120 "cycles," 121 "cpu/cycles-t/," 122 "cpu/tx-start/" 123 "}" 124 }; 125 126 static const char * topdown_attrs[] = { 127 "topdown-total-slots", 128 "topdown-slots-retired", 129 "topdown-recovery-bubbles", 130 "topdown-fetch-bubbles", 131 "topdown-slots-issued", 132 NULL, 133 }; 134 135 static const char *topdown_metric_attrs[] = { 136 "slots", 137 "topdown-retiring", 138 "topdown-bad-spec", 139 "topdown-fe-bound", 140 "topdown-be-bound", 141 NULL, 142 }; 143 144 static const char *topdown_metric_L2_attrs[] = { 145 "slots", 146 "topdown-retiring", 147 "topdown-bad-spec", 148 "topdown-fe-bound", 149 "topdown-be-bound", 150 "topdown-heavy-ops", 151 "topdown-br-mispredict", 152 "topdown-fetch-lat", 153 "topdown-mem-bound", 154 NULL, 155 }; 156 157 static const char *smi_cost_attrs = { 158 "{" 159 "msr/aperf/," 160 "msr/smi/," 161 "cycles" 162 "}" 163 }; 164 165 static struct evlist *evsel_list; 166 static bool all_counters_use_bpf = true; 167 168 static struct target target = { 169 .uid = UINT_MAX, 170 }; 171 172 #define METRIC_ONLY_LEN 20 173 174 static volatile pid_t child_pid = -1; 175 static int detailed_run = 0; 176 static bool transaction_run; 177 static bool topdown_run = false; 178 static bool smi_cost = false; 179 static bool smi_reset = false; 180 static int big_num_opt = -1; 181 static bool group = false; 182 static const char *pre_cmd = NULL; 183 static const char *post_cmd = NULL; 184 static bool sync_run = false; 185 static bool forever = false; 186 static bool force_metric_only = false; 187 static struct timespec ref_time; 188 static bool append_file; 189 static bool interval_count; 190 static const char *output_name; 191 static int output_fd; 192 193 struct perf_stat { 194 bool record; 195 struct perf_data data; 196 struct perf_session *session; 197 u64 bytes_written; 198 struct perf_tool tool; 199 bool maps_allocated; 200 struct perf_cpu_map *cpus; 201 struct perf_thread_map *threads; 202 enum aggr_mode aggr_mode; 203 }; 204 205 static struct perf_stat perf_stat; 206 #define STAT_RECORD perf_stat.record 207 208 static volatile int done = 0; 209 210 static struct perf_stat_config stat_config = { 211 .aggr_mode = AGGR_GLOBAL, 212 .scale = true, 213 .unit_width = 4, /* strlen("unit") */ 214 .run_count = 1, 215 .metric_only_len = METRIC_ONLY_LEN, 216 .walltime_nsecs_stats = &walltime_nsecs_stats, 217 .big_num = true, 218 .ctl_fd = -1, 219 .ctl_fd_ack = -1, 220 .iostat_run = false, 221 }; 222 223 static bool cpus_map_matched(struct evsel *a, struct evsel *b) 224 { 225 if (!a->core.cpus && !b->core.cpus) 226 return true; 227 228 if (!a->core.cpus || !b->core.cpus) 229 return false; 230 231 if (a->core.cpus->nr != b->core.cpus->nr) 232 return false; 233 234 for (int i = 0; i < a->core.cpus->nr; i++) { 235 if (a->core.cpus->map[i] != b->core.cpus->map[i]) 236 return false; 237 } 238 239 return true; 240 } 241 242 static void evlist__check_cpu_maps(struct evlist *evlist) 243 { 244 struct evsel *evsel, *pos, *leader; 245 char buf[1024]; 246 247 if (evlist__has_hybrid(evlist)) 248 evlist__warn_hybrid_group(evlist); 249 250 evlist__for_each_entry(evlist, evsel) { 251 leader = evsel->leader; 252 253 /* Check that leader matches cpus with each member. */ 254 if (leader == evsel) 255 continue; 256 if (cpus_map_matched(leader, evsel)) 257 continue; 258 259 /* If there's mismatch disable the group and warn user. */ 260 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n"); 261 evsel__group_desc(leader, buf, sizeof(buf)); 262 pr_warning(" %s\n", buf); 263 264 if (verbose) { 265 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf)); 266 pr_warning(" %s: %s\n", leader->name, buf); 267 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf)); 268 pr_warning(" %s: %s\n", evsel->name, buf); 269 } 270 271 for_each_group_evsel(pos, leader) { 272 pos->leader = pos; 273 pos->core.nr_members = 0; 274 } 275 evsel->leader->core.nr_members = 0; 276 } 277 } 278 279 static inline void diff_timespec(struct timespec *r, struct timespec *a, 280 struct timespec *b) 281 { 282 r->tv_sec = a->tv_sec - b->tv_sec; 283 if (a->tv_nsec < b->tv_nsec) { 284 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 285 r->tv_sec--; 286 } else { 287 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 288 } 289 } 290 291 static void perf_stat__reset_stats(void) 292 { 293 int i; 294 295 evlist__reset_stats(evsel_list); 296 perf_stat__reset_shadow_stats(); 297 298 for (i = 0; i < stat_config.stats_num; i++) 299 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]); 300 } 301 302 static int process_synthesized_event(struct perf_tool *tool __maybe_unused, 303 union perf_event *event, 304 struct perf_sample *sample __maybe_unused, 305 struct machine *machine __maybe_unused) 306 { 307 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 308 pr_err("failed to write perf data, error: %m\n"); 309 return -1; 310 } 311 312 perf_stat.bytes_written += event->header.size; 313 return 0; 314 } 315 316 static int write_stat_round_event(u64 tm, u64 type) 317 { 318 return perf_event__synthesize_stat_round(NULL, tm, type, 319 process_synthesized_event, 320 NULL); 321 } 322 323 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 324 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 325 326 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 327 328 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread, 329 struct perf_counts_values *count) 330 { 331 struct perf_sample_id *sid = SID(counter, cpu, thread); 332 333 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 334 process_synthesized_event, NULL); 335 } 336 337 static int read_single_counter(struct evsel *counter, int cpu, 338 int thread, struct timespec *rs) 339 { 340 if (counter->tool_event == PERF_TOOL_DURATION_TIME) { 341 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL; 342 struct perf_counts_values *count = 343 perf_counts(counter->counts, cpu, thread); 344 count->ena = count->run = val; 345 count->val = val; 346 return 0; 347 } 348 return evsel__read_counter(counter, cpu, thread); 349 } 350 351 /* 352 * Read out the results of a single counter: 353 * do not aggregate counts across CPUs in system-wide mode 354 */ 355 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu) 356 { 357 int nthreads = perf_thread_map__nr(evsel_list->core.threads); 358 int thread; 359 360 if (!counter->supported) 361 return -ENOENT; 362 363 if (counter->core.system_wide) 364 nthreads = 1; 365 366 for (thread = 0; thread < nthreads; thread++) { 367 struct perf_counts_values *count; 368 369 count = perf_counts(counter->counts, cpu, thread); 370 371 /* 372 * The leader's group read loads data into its group members 373 * (via evsel__read_counter()) and sets their count->loaded. 374 */ 375 if (!perf_counts__is_loaded(counter->counts, cpu, thread) && 376 read_single_counter(counter, cpu, thread, rs)) { 377 counter->counts->scaled = -1; 378 perf_counts(counter->counts, cpu, thread)->ena = 0; 379 perf_counts(counter->counts, cpu, thread)->run = 0; 380 return -1; 381 } 382 383 perf_counts__set_loaded(counter->counts, cpu, thread, false); 384 385 if (STAT_RECORD) { 386 if (evsel__write_stat_event(counter, cpu, thread, count)) { 387 pr_err("failed to write stat event\n"); 388 return -1; 389 } 390 } 391 392 if (verbose > 1) { 393 fprintf(stat_config.output, 394 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 395 evsel__name(counter), 396 cpu, 397 count->val, count->ena, count->run); 398 } 399 } 400 401 return 0; 402 } 403 404 static int read_affinity_counters(struct timespec *rs) 405 { 406 struct evsel *counter; 407 struct affinity affinity; 408 int i, ncpus, cpu; 409 410 if (all_counters_use_bpf) 411 return 0; 412 413 if (affinity__setup(&affinity) < 0) 414 return -1; 415 416 ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus); 417 if (!target__has_cpu(&target) || target__has_per_thread(&target)) 418 ncpus = 1; 419 evlist__for_each_cpu(evsel_list, i, cpu) { 420 if (i >= ncpus) 421 break; 422 affinity__set(&affinity, cpu); 423 424 evlist__for_each_entry(evsel_list, counter) { 425 if (evsel__cpu_iter_skip(counter, cpu)) 426 continue; 427 if (evsel__is_bpf(counter)) 428 continue; 429 if (!counter->err) { 430 counter->err = read_counter_cpu(counter, rs, 431 counter->cpu_iter - 1); 432 } 433 } 434 } 435 affinity__cleanup(&affinity); 436 return 0; 437 } 438 439 static int read_bpf_map_counters(void) 440 { 441 struct evsel *counter; 442 int err; 443 444 evlist__for_each_entry(evsel_list, counter) { 445 if (!evsel__is_bpf(counter)) 446 continue; 447 448 err = bpf_counter__read(counter); 449 if (err) 450 return err; 451 } 452 return 0; 453 } 454 455 static void read_counters(struct timespec *rs) 456 { 457 struct evsel *counter; 458 459 if (!stat_config.stop_read_counter) { 460 if (read_bpf_map_counters() || 461 read_affinity_counters(rs)) 462 return; 463 } 464 465 evlist__for_each_entry(evsel_list, counter) { 466 if (counter->err) 467 pr_debug("failed to read counter %s\n", counter->name); 468 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) 469 pr_warning("failed to process counter %s\n", counter->name); 470 counter->err = 0; 471 } 472 } 473 474 static int runtime_stat_new(struct perf_stat_config *config, int nthreads) 475 { 476 int i; 477 478 config->stats = calloc(nthreads, sizeof(struct runtime_stat)); 479 if (!config->stats) 480 return -1; 481 482 config->stats_num = nthreads; 483 484 for (i = 0; i < nthreads; i++) 485 runtime_stat__init(&config->stats[i]); 486 487 return 0; 488 } 489 490 static void runtime_stat_delete(struct perf_stat_config *config) 491 { 492 int i; 493 494 if (!config->stats) 495 return; 496 497 for (i = 0; i < config->stats_num; i++) 498 runtime_stat__exit(&config->stats[i]); 499 500 zfree(&config->stats); 501 } 502 503 static void runtime_stat_reset(struct perf_stat_config *config) 504 { 505 int i; 506 507 if (!config->stats) 508 return; 509 510 for (i = 0; i < config->stats_num; i++) 511 perf_stat__reset_shadow_per_stat(&config->stats[i]); 512 } 513 514 static void process_interval(void) 515 { 516 struct timespec ts, rs; 517 518 clock_gettime(CLOCK_MONOTONIC, &ts); 519 diff_timespec(&rs, &ts, &ref_time); 520 521 perf_stat__reset_shadow_per_stat(&rt_stat); 522 runtime_stat_reset(&stat_config); 523 read_counters(&rs); 524 525 if (STAT_RECORD) { 526 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 527 pr_err("failed to write stat round event\n"); 528 } 529 530 init_stats(&walltime_nsecs_stats); 531 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL); 532 print_counters(&rs, 0, NULL); 533 } 534 535 static bool handle_interval(unsigned int interval, int *times) 536 { 537 if (interval) { 538 process_interval(); 539 if (interval_count && !(--(*times))) 540 return true; 541 } 542 return false; 543 } 544 545 static int enable_counters(void) 546 { 547 struct evsel *evsel; 548 int err; 549 550 evlist__for_each_entry(evsel_list, evsel) { 551 if (!evsel__is_bpf(evsel)) 552 continue; 553 554 err = bpf_counter__enable(evsel); 555 if (err) 556 return err; 557 } 558 559 if (stat_config.initial_delay < 0) { 560 pr_info(EVLIST_DISABLED_MSG); 561 return 0; 562 } 563 564 if (stat_config.initial_delay > 0) { 565 pr_info(EVLIST_DISABLED_MSG); 566 usleep(stat_config.initial_delay * USEC_PER_MSEC); 567 } 568 569 /* 570 * We need to enable counters only if: 571 * - we don't have tracee (attaching to task or cpu) 572 * - we have initial delay configured 573 */ 574 if (!target__none(&target) || stat_config.initial_delay) { 575 if (!all_counters_use_bpf) 576 evlist__enable(evsel_list); 577 if (stat_config.initial_delay > 0) 578 pr_info(EVLIST_ENABLED_MSG); 579 } 580 return 0; 581 } 582 583 static void disable_counters(void) 584 { 585 struct evsel *counter; 586 587 /* 588 * If we don't have tracee (attaching to task or cpu), counters may 589 * still be running. To get accurate group ratios, we must stop groups 590 * from counting before reading their constituent counters. 591 */ 592 if (!target__none(&target)) { 593 evlist__for_each_entry(evsel_list, counter) 594 bpf_counter__disable(counter); 595 if (!all_counters_use_bpf) 596 evlist__disable(evsel_list); 597 } 598 } 599 600 static volatile int workload_exec_errno; 601 602 /* 603 * evlist__prepare_workload will send a SIGUSR1 604 * if the fork fails, since we asked by setting its 605 * want_signal to true. 606 */ 607 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 608 void *ucontext __maybe_unused) 609 { 610 workload_exec_errno = info->si_value.sival_int; 611 } 612 613 static bool evsel__should_store_id(struct evsel *counter) 614 { 615 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 616 } 617 618 static bool is_target_alive(struct target *_target, 619 struct perf_thread_map *threads) 620 { 621 struct stat st; 622 int i; 623 624 if (!target__has_task(_target)) 625 return true; 626 627 for (i = 0; i < threads->nr; i++) { 628 char path[PATH_MAX]; 629 630 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 631 threads->map[i].pid); 632 633 if (!stat(path, &st)) 634 return true; 635 } 636 637 return false; 638 } 639 640 static void process_evlist(struct evlist *evlist, unsigned int interval) 641 { 642 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED; 643 644 if (evlist__ctlfd_process(evlist, &cmd) > 0) { 645 switch (cmd) { 646 case EVLIST_CTL_CMD_ENABLE: 647 if (interval) 648 process_interval(); 649 break; 650 case EVLIST_CTL_CMD_DISABLE: 651 if (interval) 652 process_interval(); 653 break; 654 case EVLIST_CTL_CMD_SNAPSHOT: 655 case EVLIST_CTL_CMD_ACK: 656 case EVLIST_CTL_CMD_UNSUPPORTED: 657 case EVLIST_CTL_CMD_EVLIST: 658 case EVLIST_CTL_CMD_STOP: 659 case EVLIST_CTL_CMD_PING: 660 default: 661 break; 662 } 663 } 664 } 665 666 static void compute_tts(struct timespec *time_start, struct timespec *time_stop, 667 int *time_to_sleep) 668 { 669 int tts = *time_to_sleep; 670 struct timespec time_diff; 671 672 diff_timespec(&time_diff, time_stop, time_start); 673 674 tts -= time_diff.tv_sec * MSEC_PER_SEC + 675 time_diff.tv_nsec / NSEC_PER_MSEC; 676 677 if (tts < 0) 678 tts = 0; 679 680 *time_to_sleep = tts; 681 } 682 683 static int dispatch_events(bool forks, int timeout, int interval, int *times) 684 { 685 int child_exited = 0, status = 0; 686 int time_to_sleep, sleep_time; 687 struct timespec time_start, time_stop; 688 689 if (interval) 690 sleep_time = interval; 691 else if (timeout) 692 sleep_time = timeout; 693 else 694 sleep_time = 1000; 695 696 time_to_sleep = sleep_time; 697 698 while (!done) { 699 if (forks) 700 child_exited = waitpid(child_pid, &status, WNOHANG); 701 else 702 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0; 703 704 if (child_exited) 705 break; 706 707 clock_gettime(CLOCK_MONOTONIC, &time_start); 708 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */ 709 if (timeout || handle_interval(interval, times)) 710 break; 711 time_to_sleep = sleep_time; 712 } else { /* fd revent */ 713 process_evlist(evsel_list, interval); 714 clock_gettime(CLOCK_MONOTONIC, &time_stop); 715 compute_tts(&time_start, &time_stop, &time_to_sleep); 716 } 717 } 718 719 return status; 720 } 721 722 enum counter_recovery { 723 COUNTER_SKIP, 724 COUNTER_RETRY, 725 COUNTER_FATAL, 726 }; 727 728 static enum counter_recovery stat_handle_error(struct evsel *counter) 729 { 730 char msg[BUFSIZ]; 731 /* 732 * PPC returns ENXIO for HW counters until 2.6.37 733 * (behavior changed with commit b0a873e). 734 */ 735 if (errno == EINVAL || errno == ENOSYS || 736 errno == ENOENT || errno == EOPNOTSUPP || 737 errno == ENXIO) { 738 if (verbose > 0) 739 ui__warning("%s event is not supported by the kernel.\n", 740 evsel__name(counter)); 741 counter->supported = false; 742 /* 743 * errored is a sticky flag that means one of the counter's 744 * cpu event had a problem and needs to be reexamined. 745 */ 746 counter->errored = true; 747 748 if ((counter->leader != counter) || 749 !(counter->leader->core.nr_members > 1)) 750 return COUNTER_SKIP; 751 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) { 752 if (verbose > 0) 753 ui__warning("%s\n", msg); 754 return COUNTER_RETRY; 755 } else if (target__has_per_thread(&target) && 756 evsel_list->core.threads && 757 evsel_list->core.threads->err_thread != -1) { 758 /* 759 * For global --per-thread case, skip current 760 * error thread. 761 */ 762 if (!thread_map__remove(evsel_list->core.threads, 763 evsel_list->core.threads->err_thread)) { 764 evsel_list->core.threads->err_thread = -1; 765 return COUNTER_RETRY; 766 } 767 } 768 769 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg)); 770 ui__error("%s\n", msg); 771 772 if (child_pid != -1) 773 kill(child_pid, SIGTERM); 774 return COUNTER_FATAL; 775 } 776 777 static int __run_perf_stat(int argc, const char **argv, int run_idx) 778 { 779 int interval = stat_config.interval; 780 int times = stat_config.times; 781 int timeout = stat_config.timeout; 782 char msg[BUFSIZ]; 783 unsigned long long t0, t1; 784 struct evsel *counter; 785 size_t l; 786 int status = 0; 787 const bool forks = (argc > 0); 788 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; 789 struct affinity affinity; 790 int i, cpu, err; 791 bool second_pass = false; 792 793 if (forks) { 794 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) { 795 perror("failed to prepare workload"); 796 return -1; 797 } 798 child_pid = evsel_list->workload.pid; 799 } 800 801 if (group) 802 evlist__set_leader(evsel_list); 803 804 if (affinity__setup(&affinity) < 0) 805 return -1; 806 807 evlist__for_each_entry(evsel_list, counter) { 808 if (bpf_counter__load(counter, &target)) 809 return -1; 810 if (!evsel__is_bpf(counter)) 811 all_counters_use_bpf = false; 812 } 813 814 evlist__for_each_cpu (evsel_list, i, cpu) { 815 /* 816 * bperf calls evsel__open_per_cpu() in bperf__load(), so 817 * no need to call it again here. 818 */ 819 if (target.use_bpf) 820 break; 821 affinity__set(&affinity, cpu); 822 823 evlist__for_each_entry(evsel_list, counter) { 824 if (evsel__cpu_iter_skip(counter, cpu)) 825 continue; 826 if (counter->reset_group || counter->errored) 827 continue; 828 if (evsel__is_bpf(counter)) 829 continue; 830 try_again: 831 if (create_perf_stat_counter(counter, &stat_config, &target, 832 counter->cpu_iter - 1) < 0) { 833 834 /* 835 * Weak group failed. We cannot just undo this here 836 * because earlier CPUs might be in group mode, and the kernel 837 * doesn't support mixing group and non group reads. Defer 838 * it to later. 839 * Don't close here because we're in the wrong affinity. 840 */ 841 if ((errno == EINVAL || errno == EBADF) && 842 counter->leader != counter && 843 counter->weak_group) { 844 evlist__reset_weak_group(evsel_list, counter, false); 845 assert(counter->reset_group); 846 second_pass = true; 847 continue; 848 } 849 850 switch (stat_handle_error(counter)) { 851 case COUNTER_FATAL: 852 return -1; 853 case COUNTER_RETRY: 854 goto try_again; 855 case COUNTER_SKIP: 856 continue; 857 default: 858 break; 859 } 860 861 } 862 counter->supported = true; 863 } 864 } 865 866 if (second_pass) { 867 /* 868 * Now redo all the weak group after closing them, 869 * and also close errored counters. 870 */ 871 872 evlist__for_each_cpu(evsel_list, i, cpu) { 873 affinity__set(&affinity, cpu); 874 /* First close errored or weak retry */ 875 evlist__for_each_entry(evsel_list, counter) { 876 if (!counter->reset_group && !counter->errored) 877 continue; 878 if (evsel__cpu_iter_skip_no_inc(counter, cpu)) 879 continue; 880 perf_evsel__close_cpu(&counter->core, counter->cpu_iter); 881 } 882 /* Now reopen weak */ 883 evlist__for_each_entry(evsel_list, counter) { 884 if (!counter->reset_group && !counter->errored) 885 continue; 886 if (evsel__cpu_iter_skip(counter, cpu)) 887 continue; 888 if (!counter->reset_group) 889 continue; 890 try_again_reset: 891 pr_debug2("reopening weak %s\n", evsel__name(counter)); 892 if (create_perf_stat_counter(counter, &stat_config, &target, 893 counter->cpu_iter - 1) < 0) { 894 895 switch (stat_handle_error(counter)) { 896 case COUNTER_FATAL: 897 return -1; 898 case COUNTER_RETRY: 899 goto try_again_reset; 900 case COUNTER_SKIP: 901 continue; 902 default: 903 break; 904 } 905 } 906 counter->supported = true; 907 } 908 } 909 } 910 affinity__cleanup(&affinity); 911 912 evlist__for_each_entry(evsel_list, counter) { 913 if (!counter->supported) { 914 perf_evsel__free_fd(&counter->core); 915 continue; 916 } 917 918 l = strlen(counter->unit); 919 if (l > stat_config.unit_width) 920 stat_config.unit_width = l; 921 922 if (evsel__should_store_id(counter) && 923 evsel__store_ids(counter, evsel_list)) 924 return -1; 925 } 926 927 if (evlist__apply_filters(evsel_list, &counter)) { 928 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n", 929 counter->filter, evsel__name(counter), errno, 930 str_error_r(errno, msg, sizeof(msg))); 931 return -1; 932 } 933 934 if (STAT_RECORD) { 935 int fd = perf_data__fd(&perf_stat.data); 936 937 if (is_pipe) { 938 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); 939 } else { 940 err = perf_session__write_header(perf_stat.session, evsel_list, 941 fd, false); 942 } 943 944 if (err < 0) 945 return err; 946 947 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list, 948 process_synthesized_event, is_pipe); 949 if (err < 0) 950 return err; 951 } 952 953 /* 954 * Enable counters and exec the command: 955 */ 956 if (forks) { 957 evlist__start_workload(evsel_list); 958 err = enable_counters(); 959 if (err) 960 return -1; 961 962 t0 = rdclock(); 963 clock_gettime(CLOCK_MONOTONIC, &ref_time); 964 965 if (interval || timeout || evlist__ctlfd_initialized(evsel_list)) 966 status = dispatch_events(forks, timeout, interval, ×); 967 if (child_pid != -1) { 968 if (timeout) 969 kill(child_pid, SIGTERM); 970 wait4(child_pid, &status, 0, &stat_config.ru_data); 971 } 972 973 if (workload_exec_errno) { 974 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); 975 pr_err("Workload failed: %s\n", emsg); 976 return -1; 977 } 978 979 if (WIFSIGNALED(status)) 980 psignal(WTERMSIG(status), argv[0]); 981 } else { 982 err = enable_counters(); 983 if (err) 984 return -1; 985 986 t0 = rdclock(); 987 clock_gettime(CLOCK_MONOTONIC, &ref_time); 988 989 status = dispatch_events(forks, timeout, interval, ×); 990 } 991 992 disable_counters(); 993 994 t1 = rdclock(); 995 996 if (stat_config.walltime_run_table) 997 stat_config.walltime_run[run_idx] = t1 - t0; 998 999 if (interval && stat_config.summary) { 1000 stat_config.interval = 0; 1001 stat_config.stop_read_counter = true; 1002 init_stats(&walltime_nsecs_stats); 1003 update_stats(&walltime_nsecs_stats, t1 - t0); 1004 1005 if (stat_config.aggr_mode == AGGR_GLOBAL) 1006 evlist__save_aggr_prev_raw_counts(evsel_list); 1007 1008 evlist__copy_prev_raw_counts(evsel_list); 1009 evlist__reset_prev_raw_counts(evsel_list); 1010 runtime_stat_reset(&stat_config); 1011 perf_stat__reset_shadow_per_stat(&rt_stat); 1012 } else 1013 update_stats(&walltime_nsecs_stats, t1 - t0); 1014 1015 /* 1016 * Closing a group leader splits the group, and as we only disable 1017 * group leaders, results in remaining events becoming enabled. To 1018 * avoid arbitrary skew, we must read all counters before closing any 1019 * group leaders. 1020 */ 1021 read_counters(&(struct timespec) { .tv_nsec = t1-t0 }); 1022 1023 /* 1024 * We need to keep evsel_list alive, because it's processed 1025 * later the evsel_list will be closed after. 1026 */ 1027 if (!STAT_RECORD) 1028 evlist__close(evsel_list); 1029 1030 return WEXITSTATUS(status); 1031 } 1032 1033 static int run_perf_stat(int argc, const char **argv, int run_idx) 1034 { 1035 int ret; 1036 1037 if (pre_cmd) { 1038 ret = system(pre_cmd); 1039 if (ret) 1040 return ret; 1041 } 1042 1043 if (sync_run) 1044 sync(); 1045 1046 ret = __run_perf_stat(argc, argv, run_idx); 1047 if (ret) 1048 return ret; 1049 1050 if (post_cmd) { 1051 ret = system(post_cmd); 1052 if (ret) 1053 return ret; 1054 } 1055 1056 return ret; 1057 } 1058 1059 static void print_counters(struct timespec *ts, int argc, const char **argv) 1060 { 1061 /* Do not print anything if we record to the pipe. */ 1062 if (STAT_RECORD && perf_stat.data.is_pipe) 1063 return; 1064 if (stat_config.quiet) 1065 return; 1066 1067 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv); 1068 } 1069 1070 static volatile int signr = -1; 1071 1072 static void skip_signal(int signo) 1073 { 1074 if ((child_pid == -1) || stat_config.interval) 1075 done = 1; 1076 1077 signr = signo; 1078 /* 1079 * render child_pid harmless 1080 * won't send SIGTERM to a random 1081 * process in case of race condition 1082 * and fast PID recycling 1083 */ 1084 child_pid = -1; 1085 } 1086 1087 static void sig_atexit(void) 1088 { 1089 sigset_t set, oset; 1090 1091 /* 1092 * avoid race condition with SIGCHLD handler 1093 * in skip_signal() which is modifying child_pid 1094 * goal is to avoid send SIGTERM to a random 1095 * process 1096 */ 1097 sigemptyset(&set); 1098 sigaddset(&set, SIGCHLD); 1099 sigprocmask(SIG_BLOCK, &set, &oset); 1100 1101 if (child_pid != -1) 1102 kill(child_pid, SIGTERM); 1103 1104 sigprocmask(SIG_SETMASK, &oset, NULL); 1105 1106 if (signr == -1) 1107 return; 1108 1109 signal(signr, SIG_DFL); 1110 kill(getpid(), signr); 1111 } 1112 1113 void perf_stat__set_big_num(int set) 1114 { 1115 stat_config.big_num = (set != 0); 1116 } 1117 1118 void perf_stat__set_no_csv_summary(int set) 1119 { 1120 stat_config.no_csv_summary = (set != 0); 1121 } 1122 1123 static int stat__set_big_num(const struct option *opt __maybe_unused, 1124 const char *s __maybe_unused, int unset) 1125 { 1126 big_num_opt = unset ? 0 : 1; 1127 perf_stat__set_big_num(!unset); 1128 return 0; 1129 } 1130 1131 static int enable_metric_only(const struct option *opt __maybe_unused, 1132 const char *s __maybe_unused, int unset) 1133 { 1134 force_metric_only = true; 1135 stat_config.metric_only = !unset; 1136 return 0; 1137 } 1138 1139 static int parse_metric_groups(const struct option *opt, 1140 const char *str, 1141 int unset __maybe_unused) 1142 { 1143 return metricgroup__parse_groups(opt, str, 1144 stat_config.metric_no_group, 1145 stat_config.metric_no_merge, 1146 &stat_config.metric_events); 1147 } 1148 1149 static int parse_control_option(const struct option *opt, 1150 const char *str, 1151 int unset __maybe_unused) 1152 { 1153 struct perf_stat_config *config = opt->value; 1154 1155 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close); 1156 } 1157 1158 static int parse_stat_cgroups(const struct option *opt, 1159 const char *str, int unset) 1160 { 1161 if (stat_config.cgroup_list) { 1162 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 1163 return -1; 1164 } 1165 1166 return parse_cgroups(opt, str, unset); 1167 } 1168 1169 static struct option stat_options[] = { 1170 OPT_BOOLEAN('T', "transaction", &transaction_run, 1171 "hardware transaction statistics"), 1172 OPT_CALLBACK('e', "event", &evsel_list, "event", 1173 "event selector. use 'perf list' to list available events", 1174 parse_events_option), 1175 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 1176 "event filter", parse_filter), 1177 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 1178 "child tasks do not inherit counters"), 1179 OPT_STRING('p', "pid", &target.pid, "pid", 1180 "stat events on existing process id"), 1181 OPT_STRING('t', "tid", &target.tid, "tid", 1182 "stat events on existing thread id"), 1183 #ifdef HAVE_BPF_SKEL 1184 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id", 1185 "stat events on existing bpf program id"), 1186 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf, 1187 "use bpf program to count events"), 1188 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path", 1189 "path to perf_event_attr map"), 1190 #endif 1191 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 1192 "system-wide collection from all CPUs"), 1193 OPT_BOOLEAN('g', "group", &group, 1194 "put the counters into a counter group"), 1195 OPT_BOOLEAN(0, "scale", &stat_config.scale, 1196 "Use --no-scale to disable counter scaling for multiplexing"), 1197 OPT_INCR('v', "verbose", &verbose, 1198 "be more verbose (show counter open errors, etc)"), 1199 OPT_INTEGER('r', "repeat", &stat_config.run_count, 1200 "repeat command and print average + stddev (max: 100, forever: 0)"), 1201 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 1202 "display details about each run (only with -r option)"), 1203 OPT_BOOLEAN('n', "null", &stat_config.null_run, 1204 "null run - dont start any counters"), 1205 OPT_INCR('d', "detailed", &detailed_run, 1206 "detailed run - start a lot of events"), 1207 OPT_BOOLEAN('S', "sync", &sync_run, 1208 "call sync() before starting a run"), 1209 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1210 "print large numbers with thousands\' separators", 1211 stat__set_big_num), 1212 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 1213 "list of cpus to monitor in system-wide"), 1214 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, 1215 "disable CPU count aggregation", AGGR_NONE), 1216 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"), 1217 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 1218 "print counts with custom separator"), 1219 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 1220 "monitor event in cgroup name only", parse_stat_cgroups), 1221 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name", 1222 "expand events for each cgroup"), 1223 OPT_STRING('o', "output", &output_name, "file", "output file name"), 1224 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 1225 OPT_INTEGER(0, "log-fd", &output_fd, 1226 "log output to fd, instead of stderr"), 1227 OPT_STRING(0, "pre", &pre_cmd, "command", 1228 "command to run prior to the measured command"), 1229 OPT_STRING(0, "post", &post_cmd, "command", 1230 "command to run after to the measured command"), 1231 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 1232 "print counts at regular interval in ms " 1233 "(overhead is possible for values <= 100ms)"), 1234 OPT_INTEGER(0, "interval-count", &stat_config.times, 1235 "print counts for fixed number of times"), 1236 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 1237 "clear screen in between new interval"), 1238 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 1239 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 1240 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, 1241 "aggregate counts per processor socket", AGGR_SOCKET), 1242 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode, 1243 "aggregate counts per processor die", AGGR_DIE), 1244 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, 1245 "aggregate counts per physical processor core", AGGR_CORE), 1246 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, 1247 "aggregate counts per thread", AGGR_THREAD), 1248 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode, 1249 "aggregate counts per numa node", AGGR_NODE), 1250 OPT_INTEGER('D', "delay", &stat_config.initial_delay, 1251 "ms to wait before starting measurement after program start (-1: start with events disabled)"), 1252 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 1253 "Only print computed metrics. No raw values", enable_metric_only), 1254 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group, 1255 "don't group metric events, impacts multiplexing"), 1256 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge, 1257 "don't try to share events between metrics in a group"), 1258 OPT_BOOLEAN(0, "topdown", &topdown_run, 1259 "measure top-down statistics"), 1260 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level, 1261 "Set the metrics level for the top-down statistics (0: max level)"), 1262 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 1263 "measure SMI cost"), 1264 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 1265 "monitor specified metrics or metric groups (separated by ,)", 1266 parse_metric_groups), 1267 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, 1268 "Configure all used events to run in kernel space.", 1269 PARSE_OPT_EXCLUSIVE), 1270 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, 1271 "Configure all used events to run in user space.", 1272 PARSE_OPT_EXCLUSIVE), 1273 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread, 1274 "Use with 'percore' event qualifier to show the event " 1275 "counts of one hardware thread by sum up total hardware " 1276 "threads of same physical core"), 1277 OPT_BOOLEAN(0, "summary", &stat_config.summary, 1278 "print summary for interval mode"), 1279 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary, 1280 "don't print 'summary' for CSV summary output"), 1281 OPT_BOOLEAN(0, "quiet", &stat_config.quiet, 1282 "don't print output (useful with record)"), 1283 #ifdef HAVE_LIBPFM 1284 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event", 1285 "libpfm4 event selector. use 'perf list' to list available events", 1286 parse_libpfm_events_option), 1287 #endif 1288 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]", 1289 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n" 1290 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n" 1291 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.", 1292 parse_control_option), 1293 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default", 1294 "measure I/O performance metrics provided by arch/platform", 1295 iostat_parse), 1296 OPT_END() 1297 }; 1298 1299 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 1300 struct perf_cpu_map *map, int cpu) 1301 { 1302 return cpu_map__get_socket(map, cpu, NULL); 1303 } 1304 1305 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 1306 struct perf_cpu_map *map, int cpu) 1307 { 1308 return cpu_map__get_die(map, cpu, NULL); 1309 } 1310 1311 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 1312 struct perf_cpu_map *map, int cpu) 1313 { 1314 return cpu_map__get_core(map, cpu, NULL); 1315 } 1316 1317 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused, 1318 struct perf_cpu_map *map, int cpu) 1319 { 1320 return cpu_map__get_node(map, cpu, NULL); 1321 } 1322 1323 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config, 1324 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx) 1325 { 1326 int cpu; 1327 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id(); 1328 1329 if (idx >= map->nr) 1330 return id; 1331 1332 cpu = map->map[idx]; 1333 1334 if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu])) 1335 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx); 1336 1337 id = config->cpus_aggr_map->map[cpu]; 1338 return id; 1339 } 1340 1341 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config, 1342 struct perf_cpu_map *map, int idx) 1343 { 1344 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx); 1345 } 1346 1347 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config, 1348 struct perf_cpu_map *map, int idx) 1349 { 1350 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx); 1351 } 1352 1353 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config, 1354 struct perf_cpu_map *map, int idx) 1355 { 1356 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx); 1357 } 1358 1359 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config, 1360 struct perf_cpu_map *map, int idx) 1361 { 1362 return perf_stat__get_aggr(config, perf_stat__get_node, map, idx); 1363 } 1364 1365 static bool term_percore_set(void) 1366 { 1367 struct evsel *counter; 1368 1369 evlist__for_each_entry(evsel_list, counter) { 1370 if (counter->percore) 1371 return true; 1372 } 1373 1374 return false; 1375 } 1376 1377 static int perf_stat_init_aggr_mode(void) 1378 { 1379 int nr; 1380 1381 switch (stat_config.aggr_mode) { 1382 case AGGR_SOCKET: 1383 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1384 perror("cannot build socket map"); 1385 return -1; 1386 } 1387 stat_config.aggr_get_id = perf_stat__get_socket_cached; 1388 break; 1389 case AGGR_DIE: 1390 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1391 perror("cannot build die map"); 1392 return -1; 1393 } 1394 stat_config.aggr_get_id = perf_stat__get_die_cached; 1395 break; 1396 case AGGR_CORE: 1397 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1398 perror("cannot build core map"); 1399 return -1; 1400 } 1401 stat_config.aggr_get_id = perf_stat__get_core_cached; 1402 break; 1403 case AGGR_NODE: 1404 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 1405 perror("cannot build core map"); 1406 return -1; 1407 } 1408 stat_config.aggr_get_id = perf_stat__get_node_cached; 1409 break; 1410 case AGGR_NONE: 1411 if (term_percore_set()) { 1412 if (cpu_map__build_core_map(evsel_list->core.cpus, 1413 &stat_config.aggr_map)) { 1414 perror("cannot build core map"); 1415 return -1; 1416 } 1417 stat_config.aggr_get_id = perf_stat__get_core_cached; 1418 } 1419 break; 1420 case AGGR_GLOBAL: 1421 case AGGR_THREAD: 1422 case AGGR_UNSET: 1423 default: 1424 break; 1425 } 1426 1427 /* 1428 * The evsel_list->cpus is the base we operate on, 1429 * taking the highest cpu number to be the size of 1430 * the aggregation translate cpumap. 1431 */ 1432 nr = perf_cpu_map__max(evsel_list->core.cpus); 1433 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1); 1434 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 1435 } 1436 1437 static void cpu_aggr_map__delete(struct cpu_aggr_map *map) 1438 { 1439 if (map) { 1440 WARN_ONCE(refcount_read(&map->refcnt) != 0, 1441 "cpu_aggr_map refcnt unbalanced\n"); 1442 free(map); 1443 } 1444 } 1445 1446 static void cpu_aggr_map__put(struct cpu_aggr_map *map) 1447 { 1448 if (map && refcount_dec_and_test(&map->refcnt)) 1449 cpu_aggr_map__delete(map); 1450 } 1451 1452 static void perf_stat__exit_aggr_mode(void) 1453 { 1454 cpu_aggr_map__put(stat_config.aggr_map); 1455 cpu_aggr_map__put(stat_config.cpus_aggr_map); 1456 stat_config.aggr_map = NULL; 1457 stat_config.cpus_aggr_map = NULL; 1458 } 1459 1460 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx) 1461 { 1462 int cpu; 1463 1464 if (idx > map->nr) 1465 return -1; 1466 1467 cpu = map->map[idx]; 1468 1469 if (cpu >= env->nr_cpus_avail) 1470 return -1; 1471 1472 return cpu; 1473 } 1474 1475 static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data) 1476 { 1477 struct perf_env *env = data; 1478 int cpu = perf_env__get_cpu(env, map, idx); 1479 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id(); 1480 1481 if (cpu != -1) 1482 id.socket = env->cpu[cpu].socket_id; 1483 1484 return id; 1485 } 1486 1487 static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data) 1488 { 1489 struct perf_env *env = data; 1490 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id(); 1491 int cpu = perf_env__get_cpu(env, map, idx); 1492 1493 if (cpu != -1) { 1494 /* 1495 * die_id is relative to socket, so start 1496 * with the socket ID and then add die to 1497 * make a unique ID. 1498 */ 1499 id.socket = env->cpu[cpu].socket_id; 1500 id.die = env->cpu[cpu].die_id; 1501 } 1502 1503 return id; 1504 } 1505 1506 static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data) 1507 { 1508 struct perf_env *env = data; 1509 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id(); 1510 int cpu = perf_env__get_cpu(env, map, idx); 1511 1512 if (cpu != -1) { 1513 /* 1514 * core_id is relative to socket and die, 1515 * we need a global id. So we set 1516 * socket, die id and core id 1517 */ 1518 id.socket = env->cpu[cpu].socket_id; 1519 id.die = env->cpu[cpu].die_id; 1520 id.core = env->cpu[cpu].core_id; 1521 } 1522 1523 return id; 1524 } 1525 1526 static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data) 1527 { 1528 int cpu = perf_env__get_cpu(data, map, idx); 1529 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id(); 1530 1531 id.node = perf_env__numa_node(data, cpu); 1532 return id; 1533 } 1534 1535 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus, 1536 struct cpu_aggr_map **sockp) 1537 { 1538 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); 1539 } 1540 1541 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus, 1542 struct cpu_aggr_map **diep) 1543 { 1544 return cpu_map__build_map(cpus, diep, perf_env__get_die, env); 1545 } 1546 1547 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus, 1548 struct cpu_aggr_map **corep) 1549 { 1550 return cpu_map__build_map(cpus, corep, perf_env__get_core, env); 1551 } 1552 1553 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus, 1554 struct cpu_aggr_map **nodep) 1555 { 1556 return cpu_map__build_map(cpus, nodep, perf_env__get_node, env); 1557 } 1558 1559 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1560 struct perf_cpu_map *map, int idx) 1561 { 1562 return perf_env__get_socket(map, idx, &perf_stat.session->header.env); 1563 } 1564 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1565 struct perf_cpu_map *map, int idx) 1566 { 1567 return perf_env__get_die(map, idx, &perf_stat.session->header.env); 1568 } 1569 1570 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1571 struct perf_cpu_map *map, int idx) 1572 { 1573 return perf_env__get_core(map, idx, &perf_stat.session->header.env); 1574 } 1575 1576 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, 1577 struct perf_cpu_map *map, int idx) 1578 { 1579 return perf_env__get_node(map, idx, &perf_stat.session->header.env); 1580 } 1581 1582 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1583 { 1584 struct perf_env *env = &st->session->header.env; 1585 1586 switch (stat_config.aggr_mode) { 1587 case AGGR_SOCKET: 1588 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1589 perror("cannot build socket map"); 1590 return -1; 1591 } 1592 stat_config.aggr_get_id = perf_stat__get_socket_file; 1593 break; 1594 case AGGR_DIE: 1595 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1596 perror("cannot build die map"); 1597 return -1; 1598 } 1599 stat_config.aggr_get_id = perf_stat__get_die_file; 1600 break; 1601 case AGGR_CORE: 1602 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1603 perror("cannot build core map"); 1604 return -1; 1605 } 1606 stat_config.aggr_get_id = perf_stat__get_core_file; 1607 break; 1608 case AGGR_NODE: 1609 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1610 perror("cannot build core map"); 1611 return -1; 1612 } 1613 stat_config.aggr_get_id = perf_stat__get_node_file; 1614 break; 1615 case AGGR_NONE: 1616 case AGGR_GLOBAL: 1617 case AGGR_THREAD: 1618 case AGGR_UNSET: 1619 default: 1620 break; 1621 } 1622 1623 return 0; 1624 } 1625 1626 /* 1627 * Add default attributes, if there were no attributes specified or 1628 * if -d/--detailed, -d -d or -d -d -d is used: 1629 */ 1630 static int add_default_attributes(void) 1631 { 1632 int err; 1633 struct perf_event_attr default_attrs0[] = { 1634 1635 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 1636 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 1637 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 1638 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 1639 1640 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 1641 }; 1642 struct perf_event_attr frontend_attrs[] = { 1643 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 1644 }; 1645 struct perf_event_attr backend_attrs[] = { 1646 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 1647 }; 1648 struct perf_event_attr default_attrs1[] = { 1649 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 1650 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 1651 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 1652 1653 }; 1654 struct perf_event_attr default_sw_attrs[] = { 1655 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 1656 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 1657 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 1658 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 1659 }; 1660 1661 /* 1662 * Detailed stats (-d), covering the L1 and last level data caches: 1663 */ 1664 struct perf_event_attr detailed_attrs[] = { 1665 1666 { .type = PERF_TYPE_HW_CACHE, 1667 .config = 1668 PERF_COUNT_HW_CACHE_L1D << 0 | 1669 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1670 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1671 1672 { .type = PERF_TYPE_HW_CACHE, 1673 .config = 1674 PERF_COUNT_HW_CACHE_L1D << 0 | 1675 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1676 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1677 1678 { .type = PERF_TYPE_HW_CACHE, 1679 .config = 1680 PERF_COUNT_HW_CACHE_LL << 0 | 1681 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1682 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1683 1684 { .type = PERF_TYPE_HW_CACHE, 1685 .config = 1686 PERF_COUNT_HW_CACHE_LL << 0 | 1687 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1688 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1689 }; 1690 1691 /* 1692 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 1693 */ 1694 struct perf_event_attr very_detailed_attrs[] = { 1695 1696 { .type = PERF_TYPE_HW_CACHE, 1697 .config = 1698 PERF_COUNT_HW_CACHE_L1I << 0 | 1699 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1700 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1701 1702 { .type = PERF_TYPE_HW_CACHE, 1703 .config = 1704 PERF_COUNT_HW_CACHE_L1I << 0 | 1705 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1706 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1707 1708 { .type = PERF_TYPE_HW_CACHE, 1709 .config = 1710 PERF_COUNT_HW_CACHE_DTLB << 0 | 1711 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1712 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1713 1714 { .type = PERF_TYPE_HW_CACHE, 1715 .config = 1716 PERF_COUNT_HW_CACHE_DTLB << 0 | 1717 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1718 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1719 1720 { .type = PERF_TYPE_HW_CACHE, 1721 .config = 1722 PERF_COUNT_HW_CACHE_ITLB << 0 | 1723 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1724 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1725 1726 { .type = PERF_TYPE_HW_CACHE, 1727 .config = 1728 PERF_COUNT_HW_CACHE_ITLB << 0 | 1729 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1730 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1731 1732 }; 1733 1734 /* 1735 * Very, very detailed stats (-d -d -d), adding prefetch events: 1736 */ 1737 struct perf_event_attr very_very_detailed_attrs[] = { 1738 1739 { .type = PERF_TYPE_HW_CACHE, 1740 .config = 1741 PERF_COUNT_HW_CACHE_L1D << 0 | 1742 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1743 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1744 1745 { .type = PERF_TYPE_HW_CACHE, 1746 .config = 1747 PERF_COUNT_HW_CACHE_L1D << 0 | 1748 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1749 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1750 }; 1751 struct parse_events_error errinfo; 1752 1753 /* Set attrs if no event is selected and !null_run: */ 1754 if (stat_config.null_run) 1755 return 0; 1756 1757 bzero(&errinfo, sizeof(errinfo)); 1758 if (transaction_run) { 1759 /* Handle -T as -M transaction. Once platform specific metrics 1760 * support has been added to the json files, all architectures 1761 * will use this approach. To determine transaction support 1762 * on an architecture test for such a metric name. 1763 */ 1764 if (metricgroup__has_metric("transaction")) { 1765 struct option opt = { .value = &evsel_list }; 1766 1767 return metricgroup__parse_groups(&opt, "transaction", 1768 stat_config.metric_no_group, 1769 stat_config.metric_no_merge, 1770 &stat_config.metric_events); 1771 } 1772 1773 if (pmu_have_event("cpu", "cycles-ct") && 1774 pmu_have_event("cpu", "el-start")) 1775 err = parse_events(evsel_list, transaction_attrs, 1776 &errinfo); 1777 else 1778 err = parse_events(evsel_list, 1779 transaction_limited_attrs, 1780 &errinfo); 1781 if (err) { 1782 fprintf(stderr, "Cannot set up transaction events\n"); 1783 parse_events_print_error(&errinfo, transaction_attrs); 1784 return -1; 1785 } 1786 return 0; 1787 } 1788 1789 if (smi_cost) { 1790 int smi; 1791 1792 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 1793 fprintf(stderr, "freeze_on_smi is not supported.\n"); 1794 return -1; 1795 } 1796 1797 if (!smi) { 1798 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 1799 fprintf(stderr, "Failed to set freeze_on_smi.\n"); 1800 return -1; 1801 } 1802 smi_reset = true; 1803 } 1804 1805 if (pmu_have_event("msr", "aperf") && 1806 pmu_have_event("msr", "smi")) { 1807 if (!force_metric_only) 1808 stat_config.metric_only = true; 1809 err = parse_events(evsel_list, smi_cost_attrs, &errinfo); 1810 } else { 1811 fprintf(stderr, "To measure SMI cost, it needs " 1812 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n"); 1813 parse_events_print_error(&errinfo, smi_cost_attrs); 1814 return -1; 1815 } 1816 if (err) { 1817 parse_events_print_error(&errinfo, smi_cost_attrs); 1818 fprintf(stderr, "Cannot set up SMI cost events\n"); 1819 return -1; 1820 } 1821 return 0; 1822 } 1823 1824 if (topdown_run) { 1825 const char **metric_attrs = topdown_metric_attrs; 1826 unsigned int max_level = 1; 1827 char *str = NULL; 1828 bool warn = false; 1829 1830 if (!force_metric_only) 1831 stat_config.metric_only = true; 1832 1833 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) { 1834 metric_attrs = topdown_metric_L2_attrs; 1835 max_level = 2; 1836 } 1837 1838 if (stat_config.topdown_level > max_level) { 1839 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level); 1840 return -1; 1841 } else if (!stat_config.topdown_level) 1842 stat_config.topdown_level = max_level; 1843 1844 if (topdown_filter_events(metric_attrs, &str, 1) < 0) { 1845 pr_err("Out of memory\n"); 1846 return -1; 1847 } 1848 if (metric_attrs[0] && str) { 1849 if (!stat_config.interval && !stat_config.metric_only) { 1850 fprintf(stat_config.output, 1851 "Topdown accuracy may decrease when measuring long periods.\n" 1852 "Please print the result regularly, e.g. -I1000\n"); 1853 } 1854 goto setup_metrics; 1855 } 1856 1857 zfree(&str); 1858 1859 if (stat_config.aggr_mode != AGGR_GLOBAL && 1860 stat_config.aggr_mode != AGGR_CORE) { 1861 pr_err("top down event configuration requires --per-core mode\n"); 1862 return -1; 1863 } 1864 stat_config.aggr_mode = AGGR_CORE; 1865 if (nr_cgroups || !target__has_cpu(&target)) { 1866 pr_err("top down event configuration requires system-wide mode (-a)\n"); 1867 return -1; 1868 } 1869 1870 if (topdown_filter_events(topdown_attrs, &str, 1871 arch_topdown_check_group(&warn)) < 0) { 1872 pr_err("Out of memory\n"); 1873 return -1; 1874 } 1875 if (topdown_attrs[0] && str) { 1876 if (warn) 1877 arch_topdown_group_warn(); 1878 setup_metrics: 1879 err = parse_events(evsel_list, str, &errinfo); 1880 if (err) { 1881 fprintf(stderr, 1882 "Cannot set up top down events %s: %d\n", 1883 str, err); 1884 parse_events_print_error(&errinfo, str); 1885 free(str); 1886 return -1; 1887 } 1888 } else { 1889 fprintf(stderr, "System does not support topdown\n"); 1890 return -1; 1891 } 1892 free(str); 1893 } 1894 1895 if (!evsel_list->core.nr_entries) { 1896 if (perf_pmu__has_hybrid()) { 1897 const char *hybrid_str = "cycles,instructions,branches,branch-misses"; 1898 1899 if (target__has_cpu(&target)) 1900 default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK; 1901 1902 if (evlist__add_default_attrs(evsel_list, 1903 default_sw_attrs) < 0) { 1904 return -1; 1905 } 1906 1907 err = parse_events(evsel_list, hybrid_str, &errinfo); 1908 if (err) { 1909 fprintf(stderr, 1910 "Cannot set up hybrid events %s: %d\n", 1911 hybrid_str, err); 1912 parse_events_print_error(&errinfo, hybrid_str); 1913 return -1; 1914 } 1915 return err; 1916 } 1917 1918 if (target__has_cpu(&target)) 1919 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; 1920 1921 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0) 1922 return -1; 1923 if (pmu_have_event("cpu", "stalled-cycles-frontend")) { 1924 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0) 1925 return -1; 1926 } 1927 if (pmu_have_event("cpu", "stalled-cycles-backend")) { 1928 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0) 1929 return -1; 1930 } 1931 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0) 1932 return -1; 1933 1934 if (arch_evlist__add_default_attrs(evsel_list) < 0) 1935 return -1; 1936 } 1937 1938 /* Detailed events get appended to the event list: */ 1939 1940 if (detailed_run < 1) 1941 return 0; 1942 1943 /* Append detailed run extra attributes: */ 1944 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) 1945 return -1; 1946 1947 if (detailed_run < 2) 1948 return 0; 1949 1950 /* Append very detailed run extra attributes: */ 1951 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) 1952 return -1; 1953 1954 if (detailed_run < 3) 1955 return 0; 1956 1957 /* Append very, very detailed run extra attributes: */ 1958 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); 1959 } 1960 1961 static const char * const stat_record_usage[] = { 1962 "perf stat record [<options>]", 1963 NULL, 1964 }; 1965 1966 static void init_features(struct perf_session *session) 1967 { 1968 int feat; 1969 1970 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 1971 perf_header__set_feat(&session->header, feat); 1972 1973 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 1974 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 1975 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 1976 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 1977 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 1978 } 1979 1980 static int __cmd_record(int argc, const char **argv) 1981 { 1982 struct perf_session *session; 1983 struct perf_data *data = &perf_stat.data; 1984 1985 argc = parse_options(argc, argv, stat_options, stat_record_usage, 1986 PARSE_OPT_STOP_AT_NON_OPTION); 1987 1988 if (output_name) 1989 data->path = output_name; 1990 1991 if (stat_config.run_count != 1 || forever) { 1992 pr_err("Cannot use -r option with perf stat record.\n"); 1993 return -1; 1994 } 1995 1996 session = perf_session__new(data, false, NULL); 1997 if (IS_ERR(session)) { 1998 pr_err("Perf session creation failed\n"); 1999 return PTR_ERR(session); 2000 } 2001 2002 init_features(session); 2003 2004 session->evlist = evsel_list; 2005 perf_stat.session = session; 2006 perf_stat.record = true; 2007 return argc; 2008 } 2009 2010 static int process_stat_round_event(struct perf_session *session, 2011 union perf_event *event) 2012 { 2013 struct perf_record_stat_round *stat_round = &event->stat_round; 2014 struct evsel *counter; 2015 struct timespec tsh, *ts = NULL; 2016 const char **argv = session->header.env.cmdline_argv; 2017 int argc = session->header.env.nr_cmdline; 2018 2019 evlist__for_each_entry(evsel_list, counter) 2020 perf_stat_process_counter(&stat_config, counter); 2021 2022 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 2023 update_stats(&walltime_nsecs_stats, stat_round->time); 2024 2025 if (stat_config.interval && stat_round->time) { 2026 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 2027 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 2028 ts = &tsh; 2029 } 2030 2031 print_counters(ts, argc, argv); 2032 return 0; 2033 } 2034 2035 static 2036 int process_stat_config_event(struct perf_session *session, 2037 union perf_event *event) 2038 { 2039 struct perf_tool *tool = session->tool; 2040 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2041 2042 perf_event__read_stat_config(&stat_config, &event->stat_config); 2043 2044 if (perf_cpu_map__empty(st->cpus)) { 2045 if (st->aggr_mode != AGGR_UNSET) 2046 pr_warning("warning: processing task data, aggregation mode not set\n"); 2047 return 0; 2048 } 2049 2050 if (st->aggr_mode != AGGR_UNSET) 2051 stat_config.aggr_mode = st->aggr_mode; 2052 2053 if (perf_stat.data.is_pipe) 2054 perf_stat_init_aggr_mode(); 2055 else 2056 perf_stat_init_aggr_mode_file(st); 2057 2058 return 0; 2059 } 2060 2061 static int set_maps(struct perf_stat *st) 2062 { 2063 if (!st->cpus || !st->threads) 2064 return 0; 2065 2066 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 2067 return -EINVAL; 2068 2069 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 2070 2071 if (evlist__alloc_stats(evsel_list, true)) 2072 return -ENOMEM; 2073 2074 st->maps_allocated = true; 2075 return 0; 2076 } 2077 2078 static 2079 int process_thread_map_event(struct perf_session *session, 2080 union perf_event *event) 2081 { 2082 struct perf_tool *tool = session->tool; 2083 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2084 2085 if (st->threads) { 2086 pr_warning("Extra thread map event, ignoring.\n"); 2087 return 0; 2088 } 2089 2090 st->threads = thread_map__new_event(&event->thread_map); 2091 if (!st->threads) 2092 return -ENOMEM; 2093 2094 return set_maps(st); 2095 } 2096 2097 static 2098 int process_cpu_map_event(struct perf_session *session, 2099 union perf_event *event) 2100 { 2101 struct perf_tool *tool = session->tool; 2102 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 2103 struct perf_cpu_map *cpus; 2104 2105 if (st->cpus) { 2106 pr_warning("Extra cpu map event, ignoring.\n"); 2107 return 0; 2108 } 2109 2110 cpus = cpu_map__new_data(&event->cpu_map.data); 2111 if (!cpus) 2112 return -ENOMEM; 2113 2114 st->cpus = cpus; 2115 return set_maps(st); 2116 } 2117 2118 static const char * const stat_report_usage[] = { 2119 "perf stat report [<options>]", 2120 NULL, 2121 }; 2122 2123 static struct perf_stat perf_stat = { 2124 .tool = { 2125 .attr = perf_event__process_attr, 2126 .event_update = perf_event__process_event_update, 2127 .thread_map = process_thread_map_event, 2128 .cpu_map = process_cpu_map_event, 2129 .stat_config = process_stat_config_event, 2130 .stat = perf_event__process_stat_event, 2131 .stat_round = process_stat_round_event, 2132 }, 2133 .aggr_mode = AGGR_UNSET, 2134 }; 2135 2136 static int __cmd_report(int argc, const char **argv) 2137 { 2138 struct perf_session *session; 2139 const struct option options[] = { 2140 OPT_STRING('i', "input", &input_name, "file", "input file name"), 2141 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 2142 "aggregate counts per processor socket", AGGR_SOCKET), 2143 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 2144 "aggregate counts per processor die", AGGR_DIE), 2145 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 2146 "aggregate counts per physical processor core", AGGR_CORE), 2147 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, 2148 "aggregate counts per numa node", AGGR_NODE), 2149 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 2150 "disable CPU count aggregation", AGGR_NONE), 2151 OPT_END() 2152 }; 2153 struct stat st; 2154 int ret; 2155 2156 argc = parse_options(argc, argv, options, stat_report_usage, 0); 2157 2158 if (!input_name || !strlen(input_name)) { 2159 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 2160 input_name = "-"; 2161 else 2162 input_name = "perf.data"; 2163 } 2164 2165 perf_stat.data.path = input_name; 2166 perf_stat.data.mode = PERF_DATA_MODE_READ; 2167 2168 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool); 2169 if (IS_ERR(session)) 2170 return PTR_ERR(session); 2171 2172 perf_stat.session = session; 2173 stat_config.output = stderr; 2174 evsel_list = session->evlist; 2175 2176 ret = perf_session__process_events(session); 2177 if (ret) 2178 return ret; 2179 2180 perf_session__delete(session); 2181 return 0; 2182 } 2183 2184 static void setup_system_wide(int forks) 2185 { 2186 /* 2187 * Make system wide (-a) the default target if 2188 * no target was specified and one of following 2189 * conditions is met: 2190 * 2191 * - there's no workload specified 2192 * - there is workload specified but all requested 2193 * events are system wide events 2194 */ 2195 if (!target__none(&target)) 2196 return; 2197 2198 if (!forks) 2199 target.system_wide = true; 2200 else { 2201 struct evsel *counter; 2202 2203 evlist__for_each_entry(evsel_list, counter) { 2204 if (!counter->core.system_wide && 2205 strcmp(counter->name, "duration_time")) { 2206 return; 2207 } 2208 } 2209 2210 if (evsel_list->core.nr_entries) 2211 target.system_wide = true; 2212 } 2213 } 2214 2215 int cmd_stat(int argc, const char **argv) 2216 { 2217 const char * const stat_usage[] = { 2218 "perf stat [<options>] [<command>]", 2219 NULL 2220 }; 2221 int status = -EINVAL, run_idx, err; 2222 const char *mode; 2223 FILE *output = stderr; 2224 unsigned int interval, timeout; 2225 const char * const stat_subcommands[] = { "record", "report" }; 2226 char errbuf[BUFSIZ]; 2227 2228 setlocale(LC_ALL, ""); 2229 2230 evsel_list = evlist__new(); 2231 if (evsel_list == NULL) 2232 return -ENOMEM; 2233 2234 parse_events__shrink_config_terms(); 2235 2236 /* String-parsing callback-based options would segfault when negated */ 2237 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 2238 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 2239 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 2240 2241 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 2242 (const char **) stat_usage, 2243 PARSE_OPT_STOP_AT_NON_OPTION); 2244 perf_stat__collect_metric_expr(evsel_list); 2245 perf_stat__init_shadow_stats(); 2246 2247 if (stat_config.csv_sep) { 2248 stat_config.csv_output = true; 2249 if (!strcmp(stat_config.csv_sep, "\\t")) 2250 stat_config.csv_sep = "\t"; 2251 } else 2252 stat_config.csv_sep = DEFAULT_SEPARATOR; 2253 2254 if (argc && !strncmp(argv[0], "rec", 3)) { 2255 argc = __cmd_record(argc, argv); 2256 if (argc < 0) 2257 return -1; 2258 } else if (argc && !strncmp(argv[0], "rep", 3)) 2259 return __cmd_report(argc, argv); 2260 2261 interval = stat_config.interval; 2262 timeout = stat_config.timeout; 2263 2264 /* 2265 * For record command the -o is already taken care of. 2266 */ 2267 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 2268 output = NULL; 2269 2270 if (output_name && output_fd) { 2271 fprintf(stderr, "cannot use both --output and --log-fd\n"); 2272 parse_options_usage(stat_usage, stat_options, "o", 1); 2273 parse_options_usage(NULL, stat_options, "log-fd", 0); 2274 goto out; 2275 } 2276 2277 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 2278 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 2279 goto out; 2280 } 2281 2282 if (stat_config.metric_only && stat_config.run_count > 1) { 2283 fprintf(stderr, "--metric-only is not supported with -r\n"); 2284 goto out; 2285 } 2286 2287 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 2288 fprintf(stderr, "--table is only supported with -r\n"); 2289 parse_options_usage(stat_usage, stat_options, "r", 1); 2290 parse_options_usage(NULL, stat_options, "table", 0); 2291 goto out; 2292 } 2293 2294 if (output_fd < 0) { 2295 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 2296 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 2297 goto out; 2298 } 2299 2300 if (!output && !stat_config.quiet) { 2301 struct timespec tm; 2302 mode = append_file ? "a" : "w"; 2303 2304 output = fopen(output_name, mode); 2305 if (!output) { 2306 perror("failed to create output file"); 2307 return -1; 2308 } 2309 clock_gettime(CLOCK_REALTIME, &tm); 2310 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 2311 } else if (output_fd > 0) { 2312 mode = append_file ? "a" : "w"; 2313 output = fdopen(output_fd, mode); 2314 if (!output) { 2315 perror("Failed opening logfd"); 2316 return -errno; 2317 } 2318 } 2319 2320 stat_config.output = output; 2321 2322 /* 2323 * let the spreadsheet do the pretty-printing 2324 */ 2325 if (stat_config.csv_output) { 2326 /* User explicitly passed -B? */ 2327 if (big_num_opt == 1) { 2328 fprintf(stderr, "-B option not supported with -x\n"); 2329 parse_options_usage(stat_usage, stat_options, "B", 1); 2330 parse_options_usage(NULL, stat_options, "x", 1); 2331 goto out; 2332 } else /* Nope, so disable big number formatting */ 2333 stat_config.big_num = false; 2334 } else if (big_num_opt == 0) /* User passed --no-big-num */ 2335 stat_config.big_num = false; 2336 2337 err = target__validate(&target); 2338 if (err) { 2339 target__strerror(&target, err, errbuf, BUFSIZ); 2340 pr_warning("%s\n", errbuf); 2341 } 2342 2343 setup_system_wide(argc); 2344 2345 /* 2346 * Display user/system times only for single 2347 * run and when there's specified tracee. 2348 */ 2349 if ((stat_config.run_count == 1) && target__none(&target)) 2350 stat_config.ru_display = true; 2351 2352 if (stat_config.run_count < 0) { 2353 pr_err("Run count must be a positive number\n"); 2354 parse_options_usage(stat_usage, stat_options, "r", 1); 2355 goto out; 2356 } else if (stat_config.run_count == 0) { 2357 forever = true; 2358 stat_config.run_count = 1; 2359 } 2360 2361 if (stat_config.walltime_run_table) { 2362 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 2363 if (!stat_config.walltime_run) { 2364 pr_err("failed to setup -r option"); 2365 goto out; 2366 } 2367 } 2368 2369 if ((stat_config.aggr_mode == AGGR_THREAD) && 2370 !target__has_task(&target)) { 2371 if (!target.system_wide || target.cpu_list) { 2372 fprintf(stderr, "The --per-thread option is only " 2373 "available when monitoring via -p -t -a " 2374 "options or only --per-thread.\n"); 2375 parse_options_usage(NULL, stat_options, "p", 1); 2376 parse_options_usage(NULL, stat_options, "t", 1); 2377 goto out; 2378 } 2379 } 2380 2381 /* 2382 * no_aggr, cgroup are for system-wide only 2383 * --per-thread is aggregated per thread, we dont mix it with cpu mode 2384 */ 2385 if (((stat_config.aggr_mode != AGGR_GLOBAL && 2386 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && 2387 !target__has_cpu(&target)) { 2388 fprintf(stderr, "both cgroup and no-aggregation " 2389 "modes only available in system-wide mode\n"); 2390 2391 parse_options_usage(stat_usage, stat_options, "G", 1); 2392 parse_options_usage(NULL, stat_options, "A", 1); 2393 parse_options_usage(NULL, stat_options, "a", 1); 2394 goto out; 2395 } 2396 2397 if (stat_config.iostat_run) { 2398 status = iostat_prepare(evsel_list, &stat_config); 2399 if (status) 2400 goto out; 2401 if (iostat_mode == IOSTAT_LIST) { 2402 iostat_list(evsel_list, &stat_config); 2403 goto out; 2404 } else if (verbose) 2405 iostat_list(evsel_list, &stat_config); 2406 } 2407 2408 if (add_default_attributes()) 2409 goto out; 2410 2411 if (stat_config.cgroup_list) { 2412 if (nr_cgroups > 0) { 2413 pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); 2414 parse_options_usage(stat_usage, stat_options, "G", 1); 2415 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); 2416 goto out; 2417 } 2418 2419 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, 2420 &stat_config.metric_events, true) < 0) { 2421 parse_options_usage(stat_usage, stat_options, 2422 "for-each-cgroup", 0); 2423 goto out; 2424 } 2425 } 2426 2427 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 2428 target.per_thread = true; 2429 2430 if (evlist__create_maps(evsel_list, &target) < 0) { 2431 if (target__has_task(&target)) { 2432 pr_err("Problems finding threads of monitor\n"); 2433 parse_options_usage(stat_usage, stat_options, "p", 1); 2434 parse_options_usage(NULL, stat_options, "t", 1); 2435 } else if (target__has_cpu(&target)) { 2436 perror("failed to parse CPUs map"); 2437 parse_options_usage(stat_usage, stat_options, "C", 1); 2438 parse_options_usage(NULL, stat_options, "a", 1); 2439 } 2440 goto out; 2441 } 2442 2443 evlist__check_cpu_maps(evsel_list); 2444 2445 if (perf_pmu__has_hybrid()) 2446 stat_config.no_merge = true; 2447 2448 /* 2449 * Initialize thread_map with comm names, 2450 * so we could print it out on output. 2451 */ 2452 if (stat_config.aggr_mode == AGGR_THREAD) { 2453 thread_map__read_comms(evsel_list->core.threads); 2454 if (target.system_wide) { 2455 if (runtime_stat_new(&stat_config, 2456 perf_thread_map__nr(evsel_list->core.threads))) { 2457 goto out; 2458 } 2459 } 2460 } 2461 2462 if (stat_config.aggr_mode == AGGR_NODE) 2463 cpu__setup_cpunode_map(); 2464 2465 if (stat_config.times && interval) 2466 interval_count = true; 2467 else if (stat_config.times && !interval) { 2468 pr_err("interval-count option should be used together with " 2469 "interval-print.\n"); 2470 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 2471 parse_options_usage(stat_usage, stat_options, "I", 1); 2472 goto out; 2473 } 2474 2475 if (timeout && timeout < 100) { 2476 if (timeout < 10) { 2477 pr_err("timeout must be >= 10ms.\n"); 2478 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2479 goto out; 2480 } else 2481 pr_warning("timeout < 100ms. " 2482 "The overhead percentage could be high in some cases. " 2483 "Please proceed with caution.\n"); 2484 } 2485 if (timeout && interval) { 2486 pr_err("timeout option is not supported with interval-print.\n"); 2487 parse_options_usage(stat_usage, stat_options, "timeout", 0); 2488 parse_options_usage(stat_usage, stat_options, "I", 1); 2489 goto out; 2490 } 2491 2492 if (evlist__alloc_stats(evsel_list, interval)) 2493 goto out; 2494 2495 if (perf_stat_init_aggr_mode()) 2496 goto out; 2497 2498 /* 2499 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 2500 * while avoiding that older tools show confusing messages. 2501 * 2502 * However for pipe sessions we need to keep it zero, 2503 * because script's perf_evsel__check_attr is triggered 2504 * by attr->sample_type != 0, and we can't run it on 2505 * stat sessions. 2506 */ 2507 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 2508 2509 /* 2510 * We dont want to block the signals - that would cause 2511 * child tasks to inherit that and Ctrl-C would not work. 2512 * What we want is for Ctrl-C to work in the exec()-ed 2513 * task, but being ignored by perf stat itself: 2514 */ 2515 atexit(sig_atexit); 2516 if (!forever) 2517 signal(SIGINT, skip_signal); 2518 signal(SIGCHLD, skip_signal); 2519 signal(SIGALRM, skip_signal); 2520 signal(SIGABRT, skip_signal); 2521 2522 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack)) 2523 goto out; 2524 2525 status = 0; 2526 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 2527 if (stat_config.run_count != 1 && verbose > 0) 2528 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 2529 run_idx + 1); 2530 2531 if (run_idx != 0) 2532 evlist__reset_prev_raw_counts(evsel_list); 2533 2534 status = run_perf_stat(argc, argv, run_idx); 2535 if (forever && status != -1 && !interval) { 2536 print_counters(NULL, argc, argv); 2537 perf_stat__reset_stats(); 2538 } 2539 } 2540 2541 if (!forever && status != -1 && (!interval || stat_config.summary)) 2542 print_counters(NULL, argc, argv); 2543 2544 evlist__finalize_ctlfd(evsel_list); 2545 2546 if (STAT_RECORD) { 2547 /* 2548 * We synthesize the kernel mmap record just so that older tools 2549 * don't emit warnings about not being able to resolve symbols 2550 * due to /proc/sys/kernel/kptr_restrict settings and instead provide 2551 * a saner message about no samples being in the perf.data file. 2552 * 2553 * This also serves to suppress a warning about f_header.data.size == 0 2554 * in header.c at the moment 'perf stat record' gets introduced, which 2555 * is not really needed once we start adding the stat specific PERF_RECORD_ 2556 * records, but the need to suppress the kptr_restrict messages in older 2557 * tools remain -acme 2558 */ 2559 int fd = perf_data__fd(&perf_stat.data); 2560 2561 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 2562 process_synthesized_event, 2563 &perf_stat.session->machines.host); 2564 if (err) { 2565 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 2566 "older tools may produce warnings about this file\n."); 2567 } 2568 2569 if (!interval) { 2570 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) 2571 pr_err("failed to write stat round event\n"); 2572 } 2573 2574 if (!perf_stat.data.is_pipe) { 2575 perf_stat.session->header.data_size += perf_stat.bytes_written; 2576 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 2577 } 2578 2579 evlist__close(evsel_list); 2580 perf_session__delete(perf_stat.session); 2581 } 2582 2583 perf_stat__exit_aggr_mode(); 2584 evlist__free_stats(evsel_list); 2585 out: 2586 if (stat_config.iostat_run) 2587 iostat_release(evsel_list); 2588 2589 zfree(&stat_config.walltime_run); 2590 2591 if (smi_cost && smi_reset) 2592 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 2593 2594 evlist__delete(evsel_list); 2595 2596 metricgroup__rblist_exit(&stat_config.metric_events); 2597 runtime_stat_delete(&stat_config); 2598 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close); 2599 2600 return status; 2601 } 2602