1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * builtin-kwork.c 4 * 5 * Copyright (c) 2022 Huawei Inc, Yang Jihong <yangjihong1@huawei.com> 6 */ 7 8 #include "builtin.h" 9 10 #include "util/data.h" 11 #include "util/evlist.h" 12 #include "util/evsel.h" 13 #include "util/header.h" 14 #include "util/kwork.h" 15 #include "util/debug.h" 16 #include "util/session.h" 17 #include "util/symbol.h" 18 #include "util/thread.h" 19 #include "util/string2.h" 20 #include "util/callchain.h" 21 #include "util/evsel_fprintf.h" 22 #include "util/util.h" 23 24 #include <subcmd/pager.h> 25 #include <subcmd/parse-options.h> 26 #include <traceevent/event-parse.h> 27 28 #include <errno.h> 29 #include <inttypes.h> 30 #include <signal.h> 31 #include <linux/err.h> 32 #include <linux/time64.h> 33 #include <linux/zalloc.h> 34 35 /* 36 * report header elements width 37 */ 38 #define PRINT_CPU_WIDTH 4 39 #define PRINT_COUNT_WIDTH 9 40 #define PRINT_RUNTIME_WIDTH 10 41 #define PRINT_LATENCY_WIDTH 10 42 #define PRINT_TIMESTAMP_WIDTH 17 43 #define PRINT_KWORK_NAME_WIDTH 30 44 #define RPINT_DECIMAL_WIDTH 3 45 #define PRINT_BRACKETPAIR_WIDTH 2 46 #define PRINT_TIME_UNIT_SEC_WIDTH 2 47 #define PRINT_TIME_UNIT_MESC_WIDTH 3 48 #define PRINT_PID_WIDTH 7 49 #define PRINT_TASK_NAME_WIDTH 16 50 #define PRINT_CPU_USAGE_WIDTH 6 51 #define PRINT_CPU_USAGE_DECIMAL_WIDTH 2 52 #define PRINT_CPU_USAGE_HIST_WIDTH 30 53 #define PRINT_RUNTIME_HEADER_WIDTH (PRINT_RUNTIME_WIDTH + PRINT_TIME_UNIT_MESC_WIDTH) 54 #define PRINT_LATENCY_HEADER_WIDTH (PRINT_LATENCY_WIDTH + PRINT_TIME_UNIT_MESC_WIDTH) 55 #define PRINT_TIMEHIST_CPU_WIDTH (PRINT_CPU_WIDTH + PRINT_BRACKETPAIR_WIDTH) 56 #define PRINT_TIMESTAMP_HEADER_WIDTH (PRINT_TIMESTAMP_WIDTH + PRINT_TIME_UNIT_SEC_WIDTH) 57 58 struct sort_dimension { 59 const char *name; 60 int (*cmp)(struct kwork_work *l, struct kwork_work *r); 61 struct list_head list; 62 }; 63 64 static int id_cmp(struct kwork_work *l, struct kwork_work *r) 65 { 66 if (l->cpu > r->cpu) 67 return 1; 68 if (l->cpu < r->cpu) 69 return -1; 70 71 if (l->id > r->id) 72 return 1; 73 if (l->id < r->id) 74 return -1; 75 76 return 0; 77 } 78 79 static int count_cmp(struct kwork_work *l, struct kwork_work *r) 80 { 81 if (l->nr_atoms > r->nr_atoms) 82 return 1; 83 if (l->nr_atoms < r->nr_atoms) 84 return -1; 85 86 return 0; 87 } 88 89 static int runtime_cmp(struct kwork_work *l, struct kwork_work *r) 90 { 91 if (l->total_runtime > r->total_runtime) 92 return 1; 93 if (l->total_runtime < r->total_runtime) 94 return -1; 95 96 return 0; 97 } 98 99 static int max_runtime_cmp(struct kwork_work *l, struct kwork_work *r) 100 { 101 if (l->max_runtime > r->max_runtime) 102 return 1; 103 if (l->max_runtime < r->max_runtime) 104 return -1; 105 106 return 0; 107 } 108 109 static int avg_latency_cmp(struct kwork_work *l, struct kwork_work *r) 110 { 111 u64 avgl, avgr; 112 113 if (!r->nr_atoms) 114 return 1; 115 if (!l->nr_atoms) 116 return -1; 117 118 avgl = l->total_latency / l->nr_atoms; 119 avgr = r->total_latency / r->nr_atoms; 120 121 if (avgl > avgr) 122 return 1; 123 if (avgl < avgr) 124 return -1; 125 126 return 0; 127 } 128 129 static int max_latency_cmp(struct kwork_work *l, struct kwork_work *r) 130 { 131 if (l->max_latency > r->max_latency) 132 return 1; 133 if (l->max_latency < r->max_latency) 134 return -1; 135 136 return 0; 137 } 138 139 static int cpu_usage_cmp(struct kwork_work *l, struct kwork_work *r) 140 { 141 if (l->cpu_usage > r->cpu_usage) 142 return 1; 143 if (l->cpu_usage < r->cpu_usage) 144 return -1; 145 146 return 0; 147 } 148 149 static int id_or_cpu_r_cmp(struct kwork_work *l, struct kwork_work *r) 150 { 151 if (l->id < r->id) 152 return 1; 153 if (l->id > r->id) 154 return -1; 155 156 if (l->id != 0) 157 return 0; 158 159 if (l->cpu < r->cpu) 160 return 1; 161 if (l->cpu > r->cpu) 162 return -1; 163 164 return 0; 165 } 166 167 static int sort_dimension__add(struct perf_kwork *kwork __maybe_unused, 168 const char *tok, struct list_head *list) 169 { 170 size_t i; 171 static struct sort_dimension max_sort_dimension = { 172 .name = "max", 173 .cmp = max_runtime_cmp, 174 }; 175 static struct sort_dimension id_sort_dimension = { 176 .name = "id", 177 .cmp = id_cmp, 178 }; 179 static struct sort_dimension runtime_sort_dimension = { 180 .name = "runtime", 181 .cmp = runtime_cmp, 182 }; 183 static struct sort_dimension count_sort_dimension = { 184 .name = "count", 185 .cmp = count_cmp, 186 }; 187 static struct sort_dimension avg_sort_dimension = { 188 .name = "avg", 189 .cmp = avg_latency_cmp, 190 }; 191 static struct sort_dimension rate_sort_dimension = { 192 .name = "rate", 193 .cmp = cpu_usage_cmp, 194 }; 195 static struct sort_dimension tid_sort_dimension = { 196 .name = "tid", 197 .cmp = id_or_cpu_r_cmp, 198 }; 199 struct sort_dimension *available_sorts[] = { 200 &id_sort_dimension, 201 &max_sort_dimension, 202 &count_sort_dimension, 203 &runtime_sort_dimension, 204 &avg_sort_dimension, 205 &rate_sort_dimension, 206 &tid_sort_dimension, 207 }; 208 209 if (kwork->report == KWORK_REPORT_LATENCY) 210 max_sort_dimension.cmp = max_latency_cmp; 211 212 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) { 213 if (!strcmp(available_sorts[i]->name, tok)) { 214 list_add_tail(&available_sorts[i]->list, list); 215 return 0; 216 } 217 } 218 219 return -1; 220 } 221 222 static void setup_sorting(struct perf_kwork *kwork, 223 const struct option *options, 224 const char * const usage_msg[]) 225 { 226 char *tmp, *tok, *str = strdup(kwork->sort_order); 227 228 for (tok = strtok_r(str, ", ", &tmp); 229 tok; tok = strtok_r(NULL, ", ", &tmp)) { 230 if (sort_dimension__add(kwork, tok, &kwork->sort_list) < 0) 231 usage_with_options_msg(usage_msg, options, 232 "Unknown --sort key: `%s'", tok); 233 } 234 235 pr_debug("Sort order: %s\n", kwork->sort_order); 236 free(str); 237 } 238 239 static struct kwork_atom *atom_new(struct perf_kwork *kwork, 240 struct perf_sample *sample) 241 { 242 unsigned long i; 243 struct kwork_atom_page *page; 244 struct kwork_atom *atom = NULL; 245 246 list_for_each_entry(page, &kwork->atom_page_list, list) { 247 if (!bitmap_full(page->bitmap, NR_ATOM_PER_PAGE)) { 248 i = find_first_zero_bit(page->bitmap, NR_ATOM_PER_PAGE); 249 BUG_ON(i >= NR_ATOM_PER_PAGE); 250 atom = &page->atoms[i]; 251 goto found_atom; 252 } 253 } 254 255 /* 256 * new page 257 */ 258 page = zalloc(sizeof(*page)); 259 if (page == NULL) { 260 pr_err("Failed to zalloc kwork atom page\n"); 261 return NULL; 262 } 263 264 i = 0; 265 atom = &page->atoms[0]; 266 list_add_tail(&page->list, &kwork->atom_page_list); 267 268 found_atom: 269 __set_bit(i, page->bitmap); 270 atom->time = sample->time; 271 atom->prev = NULL; 272 atom->page_addr = page; 273 atom->bit_inpage = i; 274 return atom; 275 } 276 277 static void atom_free(struct kwork_atom *atom) 278 { 279 if (atom->prev != NULL) 280 atom_free(atom->prev); 281 282 __clear_bit(atom->bit_inpage, 283 ((struct kwork_atom_page *)atom->page_addr)->bitmap); 284 } 285 286 static void atom_del(struct kwork_atom *atom) 287 { 288 list_del(&atom->list); 289 atom_free(atom); 290 } 291 292 static int work_cmp(struct list_head *list, 293 struct kwork_work *l, struct kwork_work *r) 294 { 295 int ret = 0; 296 struct sort_dimension *sort; 297 298 BUG_ON(list_empty(list)); 299 300 list_for_each_entry(sort, list, list) { 301 ret = sort->cmp(l, r); 302 if (ret) 303 return ret; 304 } 305 306 return ret; 307 } 308 309 static struct kwork_work *work_search(struct rb_root_cached *root, 310 struct kwork_work *key, 311 struct list_head *sort_list) 312 { 313 int cmp; 314 struct kwork_work *work; 315 struct rb_node *node = root->rb_root.rb_node; 316 317 while (node) { 318 work = container_of(node, struct kwork_work, node); 319 cmp = work_cmp(sort_list, key, work); 320 if (cmp > 0) 321 node = node->rb_left; 322 else if (cmp < 0) 323 node = node->rb_right; 324 else { 325 if (work->name == NULL) 326 work->name = key->name; 327 return work; 328 } 329 } 330 return NULL; 331 } 332 333 static void work_insert(struct rb_root_cached *root, 334 struct kwork_work *key, struct list_head *sort_list) 335 { 336 int cmp; 337 bool leftmost = true; 338 struct kwork_work *cur; 339 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; 340 341 while (*new) { 342 cur = container_of(*new, struct kwork_work, node); 343 parent = *new; 344 cmp = work_cmp(sort_list, key, cur); 345 346 if (cmp > 0) 347 new = &((*new)->rb_left); 348 else { 349 new = &((*new)->rb_right); 350 leftmost = false; 351 } 352 } 353 354 rb_link_node(&key->node, parent, new); 355 rb_insert_color_cached(&key->node, root, leftmost); 356 } 357 358 static struct kwork_work *work_new(struct kwork_work *key) 359 { 360 int i; 361 struct kwork_work *work = zalloc(sizeof(*work)); 362 363 if (work == NULL) { 364 pr_err("Failed to zalloc kwork work\n"); 365 return NULL; 366 } 367 368 for (i = 0; i < KWORK_TRACE_MAX; i++) 369 INIT_LIST_HEAD(&work->atom_list[i]); 370 371 work->id = key->id; 372 work->cpu = key->cpu; 373 work->name = key->name; 374 work->class = key->class; 375 return work; 376 } 377 378 static struct kwork_work *work_findnew(struct rb_root_cached *root, 379 struct kwork_work *key, 380 struct list_head *sort_list) 381 { 382 struct kwork_work *work = work_search(root, key, sort_list); 383 384 if (work != NULL) 385 return work; 386 387 work = work_new(key); 388 if (work) 389 work_insert(root, work, sort_list); 390 391 return work; 392 } 393 394 static void profile_update_timespan(struct perf_kwork *kwork, 395 struct perf_sample *sample) 396 { 397 if (!kwork->summary) 398 return; 399 400 if ((kwork->timestart == 0) || (kwork->timestart > sample->time)) 401 kwork->timestart = sample->time; 402 403 if (kwork->timeend < sample->time) 404 kwork->timeend = sample->time; 405 } 406 407 static bool profile_name_match(struct perf_kwork *kwork, 408 struct kwork_work *work) 409 { 410 if (kwork->profile_name && work->name && 411 (strcmp(work->name, kwork->profile_name) != 0)) { 412 return false; 413 } 414 415 return true; 416 } 417 418 static bool profile_event_match(struct perf_kwork *kwork, 419 struct kwork_work *work, 420 struct perf_sample *sample) 421 { 422 int cpu = work->cpu; 423 u64 time = sample->time; 424 struct perf_time_interval *ptime = &kwork->ptime; 425 426 if ((kwork->cpu_list != NULL) && !test_bit(cpu, kwork->cpu_bitmap)) 427 return false; 428 429 if (((ptime->start != 0) && (ptime->start > time)) || 430 ((ptime->end != 0) && (ptime->end < time))) 431 return false; 432 433 /* 434 * report top needs to collect the runtime of all tasks to 435 * calculate the load of each core. 436 */ 437 if ((kwork->report != KWORK_REPORT_TOP) && 438 !profile_name_match(kwork, work)) { 439 return false; 440 } 441 442 profile_update_timespan(kwork, sample); 443 return true; 444 } 445 446 static int work_push_atom(struct perf_kwork *kwork, 447 struct kwork_class *class, 448 enum kwork_trace_type src_type, 449 enum kwork_trace_type dst_type, 450 struct evsel *evsel, 451 struct perf_sample *sample, 452 struct machine *machine, 453 struct kwork_work **ret_work, 454 bool overwrite) 455 { 456 struct kwork_atom *atom, *dst_atom, *last_atom; 457 struct kwork_work *work, key; 458 459 BUG_ON(class->work_init == NULL); 460 class->work_init(kwork, class, &key, src_type, evsel, sample, machine); 461 462 atom = atom_new(kwork, sample); 463 if (atom == NULL) 464 return -1; 465 466 work = work_findnew(&class->work_root, &key, &kwork->cmp_id); 467 if (work == NULL) { 468 atom_free(atom); 469 return -1; 470 } 471 472 if (!profile_event_match(kwork, work, sample)) { 473 atom_free(atom); 474 return 0; 475 } 476 477 if (dst_type < KWORK_TRACE_MAX) { 478 dst_atom = list_last_entry_or_null(&work->atom_list[dst_type], 479 struct kwork_atom, list); 480 if (dst_atom != NULL) { 481 atom->prev = dst_atom; 482 list_del(&dst_atom->list); 483 } 484 } 485 486 if (ret_work != NULL) 487 *ret_work = work; 488 489 if (overwrite) { 490 last_atom = list_last_entry_or_null(&work->atom_list[src_type], 491 struct kwork_atom, list); 492 if (last_atom) { 493 atom_del(last_atom); 494 495 kwork->nr_skipped_events[src_type]++; 496 kwork->nr_skipped_events[KWORK_TRACE_MAX]++; 497 } 498 } 499 500 list_add_tail(&atom->list, &work->atom_list[src_type]); 501 502 return 0; 503 } 504 505 static struct kwork_atom *work_pop_atom(struct perf_kwork *kwork, 506 struct kwork_class *class, 507 enum kwork_trace_type src_type, 508 enum kwork_trace_type dst_type, 509 struct evsel *evsel, 510 struct perf_sample *sample, 511 struct machine *machine, 512 struct kwork_work **ret_work) 513 { 514 struct kwork_atom *atom, *src_atom; 515 struct kwork_work *work, key; 516 517 BUG_ON(class->work_init == NULL); 518 class->work_init(kwork, class, &key, src_type, evsel, sample, machine); 519 520 work = work_findnew(&class->work_root, &key, &kwork->cmp_id); 521 if (ret_work != NULL) 522 *ret_work = work; 523 524 if (work == NULL) 525 return NULL; 526 527 if (!profile_event_match(kwork, work, sample)) 528 return NULL; 529 530 atom = list_last_entry_or_null(&work->atom_list[dst_type], 531 struct kwork_atom, list); 532 if (atom != NULL) 533 return atom; 534 535 src_atom = atom_new(kwork, sample); 536 if (src_atom != NULL) 537 list_add_tail(&src_atom->list, &work->atom_list[src_type]); 538 else { 539 if (ret_work != NULL) 540 *ret_work = NULL; 541 } 542 543 return NULL; 544 } 545 546 static struct kwork_work *find_work_by_id(struct rb_root_cached *root, 547 u64 id, int cpu) 548 { 549 struct rb_node *next; 550 struct kwork_work *work; 551 552 next = rb_first_cached(root); 553 while (next) { 554 work = rb_entry(next, struct kwork_work, node); 555 if ((cpu != -1 && work->id == id && work->cpu == cpu) || 556 (cpu == -1 && work->id == id)) 557 return work; 558 559 next = rb_next(next); 560 } 561 562 return NULL; 563 } 564 565 static struct kwork_class *get_kwork_class(struct perf_kwork *kwork, 566 enum kwork_class_type type) 567 { 568 struct kwork_class *class; 569 570 list_for_each_entry(class, &kwork->class_list, list) { 571 if (class->type == type) 572 return class; 573 } 574 575 return NULL; 576 } 577 578 static void report_update_exit_event(struct kwork_work *work, 579 struct kwork_atom *atom, 580 struct perf_sample *sample) 581 { 582 u64 delta; 583 u64 exit_time = sample->time; 584 u64 entry_time = atom->time; 585 586 if ((entry_time != 0) && (exit_time >= entry_time)) { 587 delta = exit_time - entry_time; 588 if ((delta > work->max_runtime) || 589 (work->max_runtime == 0)) { 590 work->max_runtime = delta; 591 work->max_runtime_start = entry_time; 592 work->max_runtime_end = exit_time; 593 } 594 work->total_runtime += delta; 595 work->nr_atoms++; 596 } 597 } 598 599 static int report_entry_event(struct perf_kwork *kwork, 600 struct kwork_class *class, 601 struct evsel *evsel, 602 struct perf_sample *sample, 603 struct machine *machine) 604 { 605 return work_push_atom(kwork, class, KWORK_TRACE_ENTRY, 606 KWORK_TRACE_MAX, evsel, sample, 607 machine, NULL, true); 608 } 609 610 static int report_exit_event(struct perf_kwork *kwork, 611 struct kwork_class *class, 612 struct evsel *evsel, 613 struct perf_sample *sample, 614 struct machine *machine) 615 { 616 struct kwork_atom *atom = NULL; 617 struct kwork_work *work = NULL; 618 619 atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT, 620 KWORK_TRACE_ENTRY, evsel, sample, 621 machine, &work); 622 if (work == NULL) 623 return -1; 624 625 if (atom != NULL) { 626 report_update_exit_event(work, atom, sample); 627 atom_del(atom); 628 } 629 630 return 0; 631 } 632 633 static void latency_update_entry_event(struct kwork_work *work, 634 struct kwork_atom *atom, 635 struct perf_sample *sample) 636 { 637 u64 delta; 638 u64 entry_time = sample->time; 639 u64 raise_time = atom->time; 640 641 if ((raise_time != 0) && (entry_time >= raise_time)) { 642 delta = entry_time - raise_time; 643 if ((delta > work->max_latency) || 644 (work->max_latency == 0)) { 645 work->max_latency = delta; 646 work->max_latency_start = raise_time; 647 work->max_latency_end = entry_time; 648 } 649 work->total_latency += delta; 650 work->nr_atoms++; 651 } 652 } 653 654 static int latency_raise_event(struct perf_kwork *kwork, 655 struct kwork_class *class, 656 struct evsel *evsel, 657 struct perf_sample *sample, 658 struct machine *machine) 659 { 660 return work_push_atom(kwork, class, KWORK_TRACE_RAISE, 661 KWORK_TRACE_MAX, evsel, sample, 662 machine, NULL, true); 663 } 664 665 static int latency_entry_event(struct perf_kwork *kwork, 666 struct kwork_class *class, 667 struct evsel *evsel, 668 struct perf_sample *sample, 669 struct machine *machine) 670 { 671 struct kwork_atom *atom = NULL; 672 struct kwork_work *work = NULL; 673 674 atom = work_pop_atom(kwork, class, KWORK_TRACE_ENTRY, 675 KWORK_TRACE_RAISE, evsel, sample, 676 machine, &work); 677 if (work == NULL) 678 return -1; 679 680 if (atom != NULL) { 681 latency_update_entry_event(work, atom, sample); 682 atom_del(atom); 683 } 684 685 return 0; 686 } 687 688 static void timehist_save_callchain(struct perf_kwork *kwork, 689 struct perf_sample *sample, 690 struct evsel *evsel, 691 struct machine *machine) 692 { 693 struct symbol *sym; 694 struct thread *thread; 695 struct callchain_cursor_node *node; 696 struct callchain_cursor *cursor; 697 698 if (!kwork->show_callchain || sample->callchain == NULL) 699 return; 700 701 /* want main thread for process - has maps */ 702 thread = machine__findnew_thread(machine, sample->pid, sample->pid); 703 if (thread == NULL) { 704 pr_debug("Failed to get thread for pid %d\n", sample->pid); 705 return; 706 } 707 708 cursor = get_tls_callchain_cursor(); 709 710 if (thread__resolve_callchain(thread, cursor, evsel, sample, 711 NULL, NULL, kwork->max_stack + 2) != 0) { 712 pr_debug("Failed to resolve callchain, skipping\n"); 713 goto out_put; 714 } 715 716 callchain_cursor_commit(cursor); 717 718 while (true) { 719 node = callchain_cursor_current(cursor); 720 if (node == NULL) 721 break; 722 723 sym = node->ms.sym; 724 if (sym) { 725 if (!strcmp(sym->name, "__softirqentry_text_start") || 726 !strcmp(sym->name, "__do_softirq")) 727 sym->ignore = 1; 728 } 729 730 callchain_cursor_advance(cursor); 731 } 732 733 out_put: 734 thread__put(thread); 735 } 736 737 static void timehist_print_event(struct perf_kwork *kwork, 738 struct kwork_work *work, 739 struct kwork_atom *atom, 740 struct perf_sample *sample, 741 struct addr_location *al) 742 { 743 char entrytime[32], exittime[32]; 744 char kwork_name[PRINT_KWORK_NAME_WIDTH]; 745 746 /* 747 * runtime start 748 */ 749 timestamp__scnprintf_usec(atom->time, 750 entrytime, sizeof(entrytime)); 751 printf(" %*s ", PRINT_TIMESTAMP_WIDTH, entrytime); 752 753 /* 754 * runtime end 755 */ 756 timestamp__scnprintf_usec(sample->time, 757 exittime, sizeof(exittime)); 758 printf(" %*s ", PRINT_TIMESTAMP_WIDTH, exittime); 759 760 /* 761 * cpu 762 */ 763 printf(" [%0*d] ", PRINT_CPU_WIDTH, work->cpu); 764 765 /* 766 * kwork name 767 */ 768 if (work->class && work->class->work_name) { 769 work->class->work_name(work, kwork_name, 770 PRINT_KWORK_NAME_WIDTH); 771 printf(" %-*s ", PRINT_KWORK_NAME_WIDTH, kwork_name); 772 } else 773 printf(" %-*s ", PRINT_KWORK_NAME_WIDTH, ""); 774 775 /* 776 *runtime 777 */ 778 printf(" %*.*f ", 779 PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH, 780 (double)(sample->time - atom->time) / NSEC_PER_MSEC); 781 782 /* 783 * delaytime 784 */ 785 if (atom->prev != NULL) 786 printf(" %*.*f ", PRINT_LATENCY_WIDTH, RPINT_DECIMAL_WIDTH, 787 (double)(atom->time - atom->prev->time) / NSEC_PER_MSEC); 788 else 789 printf(" %*s ", PRINT_LATENCY_WIDTH, " "); 790 791 /* 792 * callchain 793 */ 794 if (kwork->show_callchain) { 795 struct callchain_cursor *cursor = get_tls_callchain_cursor(); 796 797 if (cursor == NULL) 798 return; 799 800 printf(" "); 801 802 sample__fprintf_sym(sample, al, 0, 803 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE | 804 EVSEL__PRINT_CALLCHAIN_ARROW | 805 EVSEL__PRINT_SKIP_IGNORED, 806 cursor, symbol_conf.bt_stop_list, 807 stdout); 808 } 809 810 printf("\n"); 811 } 812 813 static int timehist_raise_event(struct perf_kwork *kwork, 814 struct kwork_class *class, 815 struct evsel *evsel, 816 struct perf_sample *sample, 817 struct machine *machine) 818 { 819 return work_push_atom(kwork, class, KWORK_TRACE_RAISE, 820 KWORK_TRACE_MAX, evsel, sample, 821 machine, NULL, true); 822 } 823 824 static int timehist_entry_event(struct perf_kwork *kwork, 825 struct kwork_class *class, 826 struct evsel *evsel, 827 struct perf_sample *sample, 828 struct machine *machine) 829 { 830 int ret; 831 struct kwork_work *work = NULL; 832 833 ret = work_push_atom(kwork, class, KWORK_TRACE_ENTRY, 834 KWORK_TRACE_RAISE, evsel, sample, 835 machine, &work, true); 836 if (ret) 837 return ret; 838 839 if (work != NULL) 840 timehist_save_callchain(kwork, sample, evsel, machine); 841 842 return 0; 843 } 844 845 static int timehist_exit_event(struct perf_kwork *kwork, 846 struct kwork_class *class, 847 struct evsel *evsel, 848 struct perf_sample *sample, 849 struct machine *machine) 850 { 851 struct kwork_atom *atom = NULL; 852 struct kwork_work *work = NULL; 853 struct addr_location al; 854 int ret = 0; 855 856 addr_location__init(&al); 857 if (machine__resolve(machine, &al, sample) < 0) { 858 pr_debug("Problem processing event, skipping it\n"); 859 ret = -1; 860 goto out; 861 } 862 863 atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT, 864 KWORK_TRACE_ENTRY, evsel, sample, 865 machine, &work); 866 if (work == NULL) { 867 ret = -1; 868 goto out; 869 } 870 871 if (atom != NULL) { 872 work->nr_atoms++; 873 timehist_print_event(kwork, work, atom, sample, &al); 874 atom_del(atom); 875 } 876 877 out: 878 addr_location__exit(&al); 879 return ret; 880 } 881 882 static void top_update_runtime(struct kwork_work *work, 883 struct kwork_atom *atom, 884 struct perf_sample *sample) 885 { 886 u64 delta; 887 u64 exit_time = sample->time; 888 u64 entry_time = atom->time; 889 890 if ((entry_time != 0) && (exit_time >= entry_time)) { 891 delta = exit_time - entry_time; 892 work->total_runtime += delta; 893 } 894 } 895 896 static int top_entry_event(struct perf_kwork *kwork, 897 struct kwork_class *class, 898 struct evsel *evsel, 899 struct perf_sample *sample, 900 struct machine *machine) 901 { 902 return work_push_atom(kwork, class, KWORK_TRACE_ENTRY, 903 KWORK_TRACE_MAX, evsel, sample, 904 machine, NULL, true); 905 } 906 907 static int top_exit_event(struct perf_kwork *kwork, 908 struct kwork_class *class, 909 struct evsel *evsel, 910 struct perf_sample *sample, 911 struct machine *machine) 912 { 913 struct kwork_work *work, *sched_work; 914 struct kwork_class *sched_class; 915 struct kwork_atom *atom; 916 917 atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT, 918 KWORK_TRACE_ENTRY, evsel, sample, 919 machine, &work); 920 if (!work) 921 return -1; 922 923 if (atom) { 924 sched_class = get_kwork_class(kwork, KWORK_CLASS_SCHED); 925 if (sched_class) { 926 sched_work = find_work_by_id(&sched_class->work_root, 927 work->id, work->cpu); 928 if (sched_work) 929 top_update_runtime(work, atom, sample); 930 } 931 atom_del(atom); 932 } 933 934 return 0; 935 } 936 937 static int top_sched_switch_event(struct perf_kwork *kwork, 938 struct kwork_class *class, 939 struct evsel *evsel, 940 struct perf_sample *sample, 941 struct machine *machine) 942 { 943 struct kwork_atom *atom; 944 struct kwork_work *work; 945 946 atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT, 947 KWORK_TRACE_ENTRY, evsel, sample, 948 machine, &work); 949 if (!work) 950 return -1; 951 952 if (atom) { 953 top_update_runtime(work, atom, sample); 954 atom_del(atom); 955 } 956 957 return top_entry_event(kwork, class, evsel, sample, machine); 958 } 959 960 static struct kwork_class kwork_irq; 961 static int process_irq_handler_entry_event(const struct perf_tool *tool, 962 struct evsel *evsel, 963 struct perf_sample *sample, 964 struct machine *machine) 965 { 966 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 967 968 if (kwork->tp_handler->entry_event) 969 return kwork->tp_handler->entry_event(kwork, &kwork_irq, 970 evsel, sample, machine); 971 return 0; 972 } 973 974 static int process_irq_handler_exit_event(const struct perf_tool *tool, 975 struct evsel *evsel, 976 struct perf_sample *sample, 977 struct machine *machine) 978 { 979 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 980 981 if (kwork->tp_handler->exit_event) 982 return kwork->tp_handler->exit_event(kwork, &kwork_irq, 983 evsel, sample, machine); 984 return 0; 985 } 986 987 const struct evsel_str_handler irq_tp_handlers[] = { 988 { "irq:irq_handler_entry", process_irq_handler_entry_event, }, 989 { "irq:irq_handler_exit", process_irq_handler_exit_event, }, 990 }; 991 992 static int irq_class_init(struct kwork_class *class, 993 struct perf_session *session) 994 { 995 if (perf_session__set_tracepoints_handlers(session, irq_tp_handlers)) { 996 pr_err("Failed to set irq tracepoints handlers\n"); 997 return -1; 998 } 999 1000 class->work_root = RB_ROOT_CACHED; 1001 return 0; 1002 } 1003 1004 static void irq_work_init(struct perf_kwork *kwork, 1005 struct kwork_class *class, 1006 struct kwork_work *work, 1007 enum kwork_trace_type src_type __maybe_unused, 1008 struct evsel *evsel, 1009 struct perf_sample *sample, 1010 struct machine *machine __maybe_unused) 1011 { 1012 work->class = class; 1013 work->cpu = sample->cpu; 1014 1015 if (kwork->report == KWORK_REPORT_TOP) { 1016 work->id = evsel__intval_common(evsel, sample, "common_pid"); 1017 work->name = NULL; 1018 } else { 1019 work->id = evsel__intval(evsel, sample, "irq"); 1020 work->name = evsel__strval(evsel, sample, "name"); 1021 } 1022 } 1023 1024 static void irq_work_name(struct kwork_work *work, char *buf, int len) 1025 { 1026 snprintf(buf, len, "%s:%" PRIu64 "", work->name, work->id); 1027 } 1028 1029 static struct kwork_class kwork_irq = { 1030 .name = "irq", 1031 .type = KWORK_CLASS_IRQ, 1032 .nr_tracepoints = 2, 1033 .tp_handlers = irq_tp_handlers, 1034 .class_init = irq_class_init, 1035 .work_init = irq_work_init, 1036 .work_name = irq_work_name, 1037 }; 1038 1039 static struct kwork_class kwork_softirq; 1040 static int process_softirq_raise_event(const struct perf_tool *tool, 1041 struct evsel *evsel, 1042 struct perf_sample *sample, 1043 struct machine *machine) 1044 { 1045 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1046 1047 if (kwork->tp_handler->raise_event) 1048 return kwork->tp_handler->raise_event(kwork, &kwork_softirq, 1049 evsel, sample, machine); 1050 1051 return 0; 1052 } 1053 1054 static int process_softirq_entry_event(const struct perf_tool *tool, 1055 struct evsel *evsel, 1056 struct perf_sample *sample, 1057 struct machine *machine) 1058 { 1059 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1060 1061 if (kwork->tp_handler->entry_event) 1062 return kwork->tp_handler->entry_event(kwork, &kwork_softirq, 1063 evsel, sample, machine); 1064 1065 return 0; 1066 } 1067 1068 static int process_softirq_exit_event(const struct perf_tool *tool, 1069 struct evsel *evsel, 1070 struct perf_sample *sample, 1071 struct machine *machine) 1072 { 1073 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1074 1075 if (kwork->tp_handler->exit_event) 1076 return kwork->tp_handler->exit_event(kwork, &kwork_softirq, 1077 evsel, sample, machine); 1078 1079 return 0; 1080 } 1081 1082 const struct evsel_str_handler softirq_tp_handlers[] = { 1083 { "irq:softirq_raise", process_softirq_raise_event, }, 1084 { "irq:softirq_entry", process_softirq_entry_event, }, 1085 { "irq:softirq_exit", process_softirq_exit_event, }, 1086 }; 1087 1088 static int softirq_class_init(struct kwork_class *class, 1089 struct perf_session *session) 1090 { 1091 if (perf_session__set_tracepoints_handlers(session, 1092 softirq_tp_handlers)) { 1093 pr_err("Failed to set softirq tracepoints handlers\n"); 1094 return -1; 1095 } 1096 1097 class->work_root = RB_ROOT_CACHED; 1098 return 0; 1099 } 1100 1101 static char *evsel__softirq_name(struct evsel *evsel, u64 num) 1102 { 1103 char *name = NULL; 1104 bool found = false; 1105 struct tep_print_flag_sym *sym = NULL; 1106 struct tep_print_arg *args = evsel->tp_format->print_fmt.args; 1107 1108 if ((args == NULL) || (args->next == NULL)) 1109 return NULL; 1110 1111 /* skip softirq field: "REC->vec" */ 1112 for (sym = args->next->symbol.symbols; sym != NULL; sym = sym->next) { 1113 if ((eval_flag(sym->value) == (unsigned long long)num) && 1114 (strlen(sym->str) != 0)) { 1115 found = true; 1116 break; 1117 } 1118 } 1119 1120 if (!found) 1121 return NULL; 1122 1123 name = strdup(sym->str); 1124 if (name == NULL) { 1125 pr_err("Failed to copy symbol name\n"); 1126 return NULL; 1127 } 1128 return name; 1129 } 1130 1131 static void softirq_work_init(struct perf_kwork *kwork, 1132 struct kwork_class *class, 1133 struct kwork_work *work, 1134 enum kwork_trace_type src_type __maybe_unused, 1135 struct evsel *evsel, 1136 struct perf_sample *sample, 1137 struct machine *machine __maybe_unused) 1138 { 1139 u64 num; 1140 1141 work->class = class; 1142 work->cpu = sample->cpu; 1143 1144 if (kwork->report == KWORK_REPORT_TOP) { 1145 work->id = evsel__intval_common(evsel, sample, "common_pid"); 1146 work->name = NULL; 1147 } else { 1148 num = evsel__intval(evsel, sample, "vec"); 1149 work->id = num; 1150 work->name = evsel__softirq_name(evsel, num); 1151 } 1152 } 1153 1154 static void softirq_work_name(struct kwork_work *work, char *buf, int len) 1155 { 1156 snprintf(buf, len, "(s)%s:%" PRIu64 "", work->name, work->id); 1157 } 1158 1159 static struct kwork_class kwork_softirq = { 1160 .name = "softirq", 1161 .type = KWORK_CLASS_SOFTIRQ, 1162 .nr_tracepoints = 3, 1163 .tp_handlers = softirq_tp_handlers, 1164 .class_init = softirq_class_init, 1165 .work_init = softirq_work_init, 1166 .work_name = softirq_work_name, 1167 }; 1168 1169 static struct kwork_class kwork_workqueue; 1170 static int process_workqueue_activate_work_event(const struct perf_tool *tool, 1171 struct evsel *evsel, 1172 struct perf_sample *sample, 1173 struct machine *machine) 1174 { 1175 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1176 1177 if (kwork->tp_handler->raise_event) 1178 return kwork->tp_handler->raise_event(kwork, &kwork_workqueue, 1179 evsel, sample, machine); 1180 1181 return 0; 1182 } 1183 1184 static int process_workqueue_execute_start_event(const struct perf_tool *tool, 1185 struct evsel *evsel, 1186 struct perf_sample *sample, 1187 struct machine *machine) 1188 { 1189 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1190 1191 if (kwork->tp_handler->entry_event) 1192 return kwork->tp_handler->entry_event(kwork, &kwork_workqueue, 1193 evsel, sample, machine); 1194 1195 return 0; 1196 } 1197 1198 static int process_workqueue_execute_end_event(const struct perf_tool *tool, 1199 struct evsel *evsel, 1200 struct perf_sample *sample, 1201 struct machine *machine) 1202 { 1203 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1204 1205 if (kwork->tp_handler->exit_event) 1206 return kwork->tp_handler->exit_event(kwork, &kwork_workqueue, 1207 evsel, sample, machine); 1208 1209 return 0; 1210 } 1211 1212 const struct evsel_str_handler workqueue_tp_handlers[] = { 1213 { "workqueue:workqueue_activate_work", process_workqueue_activate_work_event, }, 1214 { "workqueue:workqueue_execute_start", process_workqueue_execute_start_event, }, 1215 { "workqueue:workqueue_execute_end", process_workqueue_execute_end_event, }, 1216 }; 1217 1218 static int workqueue_class_init(struct kwork_class *class, 1219 struct perf_session *session) 1220 { 1221 if (perf_session__set_tracepoints_handlers(session, 1222 workqueue_tp_handlers)) { 1223 pr_err("Failed to set workqueue tracepoints handlers\n"); 1224 return -1; 1225 } 1226 1227 class->work_root = RB_ROOT_CACHED; 1228 return 0; 1229 } 1230 1231 static void workqueue_work_init(struct perf_kwork *kwork __maybe_unused, 1232 struct kwork_class *class, 1233 struct kwork_work *work, 1234 enum kwork_trace_type src_type __maybe_unused, 1235 struct evsel *evsel, 1236 struct perf_sample *sample, 1237 struct machine *machine) 1238 { 1239 char *modp = NULL; 1240 unsigned long long function_addr = evsel__intval(evsel, 1241 sample, "function"); 1242 1243 work->class = class; 1244 work->cpu = sample->cpu; 1245 work->id = evsel__intval(evsel, sample, "work"); 1246 work->name = function_addr == 0 ? NULL : 1247 machine__resolve_kernel_addr(machine, &function_addr, &modp); 1248 } 1249 1250 static void workqueue_work_name(struct kwork_work *work, char *buf, int len) 1251 { 1252 if (work->name != NULL) 1253 snprintf(buf, len, "(w)%s", work->name); 1254 else 1255 snprintf(buf, len, "(w)0x%" PRIx64, work->id); 1256 } 1257 1258 static struct kwork_class kwork_workqueue = { 1259 .name = "workqueue", 1260 .type = KWORK_CLASS_WORKQUEUE, 1261 .nr_tracepoints = 3, 1262 .tp_handlers = workqueue_tp_handlers, 1263 .class_init = workqueue_class_init, 1264 .work_init = workqueue_work_init, 1265 .work_name = workqueue_work_name, 1266 }; 1267 1268 static struct kwork_class kwork_sched; 1269 static int process_sched_switch_event(const struct perf_tool *tool, 1270 struct evsel *evsel, 1271 struct perf_sample *sample, 1272 struct machine *machine) 1273 { 1274 struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool); 1275 1276 if (kwork->tp_handler->sched_switch_event) 1277 return kwork->tp_handler->sched_switch_event(kwork, &kwork_sched, 1278 evsel, sample, machine); 1279 return 0; 1280 } 1281 1282 const struct evsel_str_handler sched_tp_handlers[] = { 1283 { "sched:sched_switch", process_sched_switch_event, }, 1284 }; 1285 1286 static int sched_class_init(struct kwork_class *class, 1287 struct perf_session *session) 1288 { 1289 if (perf_session__set_tracepoints_handlers(session, 1290 sched_tp_handlers)) { 1291 pr_err("Failed to set sched tracepoints handlers\n"); 1292 return -1; 1293 } 1294 1295 class->work_root = RB_ROOT_CACHED; 1296 return 0; 1297 } 1298 1299 static void sched_work_init(struct perf_kwork *kwork __maybe_unused, 1300 struct kwork_class *class, 1301 struct kwork_work *work, 1302 enum kwork_trace_type src_type, 1303 struct evsel *evsel, 1304 struct perf_sample *sample, 1305 struct machine *machine __maybe_unused) 1306 { 1307 work->class = class; 1308 work->cpu = sample->cpu; 1309 1310 if (src_type == KWORK_TRACE_EXIT) { 1311 work->id = evsel__intval(evsel, sample, "prev_pid"); 1312 work->name = strdup(evsel__strval(evsel, sample, "prev_comm")); 1313 } else if (src_type == KWORK_TRACE_ENTRY) { 1314 work->id = evsel__intval(evsel, sample, "next_pid"); 1315 work->name = strdup(evsel__strval(evsel, sample, "next_comm")); 1316 } 1317 } 1318 1319 static void sched_work_name(struct kwork_work *work, char *buf, int len) 1320 { 1321 snprintf(buf, len, "%s", work->name); 1322 } 1323 1324 static struct kwork_class kwork_sched = { 1325 .name = "sched", 1326 .type = KWORK_CLASS_SCHED, 1327 .nr_tracepoints = ARRAY_SIZE(sched_tp_handlers), 1328 .tp_handlers = sched_tp_handlers, 1329 .class_init = sched_class_init, 1330 .work_init = sched_work_init, 1331 .work_name = sched_work_name, 1332 }; 1333 1334 static struct kwork_class *kwork_class_supported_list[KWORK_CLASS_MAX] = { 1335 [KWORK_CLASS_IRQ] = &kwork_irq, 1336 [KWORK_CLASS_SOFTIRQ] = &kwork_softirq, 1337 [KWORK_CLASS_WORKQUEUE] = &kwork_workqueue, 1338 [KWORK_CLASS_SCHED] = &kwork_sched, 1339 }; 1340 1341 static void print_separator(int len) 1342 { 1343 printf(" %.*s\n", len, graph_dotted_line); 1344 } 1345 1346 static int report_print_work(struct perf_kwork *kwork, struct kwork_work *work) 1347 { 1348 int ret = 0; 1349 char kwork_name[PRINT_KWORK_NAME_WIDTH]; 1350 char max_runtime_start[32], max_runtime_end[32]; 1351 char max_latency_start[32], max_latency_end[32]; 1352 1353 printf(" "); 1354 1355 /* 1356 * kwork name 1357 */ 1358 if (work->class && work->class->work_name) { 1359 work->class->work_name(work, kwork_name, 1360 PRINT_KWORK_NAME_WIDTH); 1361 ret += printf(" %-*s |", PRINT_KWORK_NAME_WIDTH, kwork_name); 1362 } else { 1363 ret += printf(" %-*s |", PRINT_KWORK_NAME_WIDTH, ""); 1364 } 1365 1366 /* 1367 * cpu 1368 */ 1369 ret += printf(" %0*d |", PRINT_CPU_WIDTH, work->cpu); 1370 1371 /* 1372 * total runtime 1373 */ 1374 if (kwork->report == KWORK_REPORT_RUNTIME) { 1375 ret += printf(" %*.*f ms |", 1376 PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH, 1377 (double)work->total_runtime / NSEC_PER_MSEC); 1378 } else if (kwork->report == KWORK_REPORT_LATENCY) { // avg delay 1379 ret += printf(" %*.*f ms |", 1380 PRINT_LATENCY_WIDTH, RPINT_DECIMAL_WIDTH, 1381 (double)work->total_latency / 1382 work->nr_atoms / NSEC_PER_MSEC); 1383 } 1384 1385 /* 1386 * count 1387 */ 1388 ret += printf(" %*" PRIu64 " |", PRINT_COUNT_WIDTH, work->nr_atoms); 1389 1390 /* 1391 * max runtime, max runtime start, max runtime end 1392 */ 1393 if (kwork->report == KWORK_REPORT_RUNTIME) { 1394 timestamp__scnprintf_usec(work->max_runtime_start, 1395 max_runtime_start, 1396 sizeof(max_runtime_start)); 1397 timestamp__scnprintf_usec(work->max_runtime_end, 1398 max_runtime_end, 1399 sizeof(max_runtime_end)); 1400 ret += printf(" %*.*f ms | %*s s | %*s s |", 1401 PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH, 1402 (double)work->max_runtime / NSEC_PER_MSEC, 1403 PRINT_TIMESTAMP_WIDTH, max_runtime_start, 1404 PRINT_TIMESTAMP_WIDTH, max_runtime_end); 1405 } 1406 /* 1407 * max delay, max delay start, max delay end 1408 */ 1409 else if (kwork->report == KWORK_REPORT_LATENCY) { 1410 timestamp__scnprintf_usec(work->max_latency_start, 1411 max_latency_start, 1412 sizeof(max_latency_start)); 1413 timestamp__scnprintf_usec(work->max_latency_end, 1414 max_latency_end, 1415 sizeof(max_latency_end)); 1416 ret += printf(" %*.*f ms | %*s s | %*s s |", 1417 PRINT_LATENCY_WIDTH, RPINT_DECIMAL_WIDTH, 1418 (double)work->max_latency / NSEC_PER_MSEC, 1419 PRINT_TIMESTAMP_WIDTH, max_latency_start, 1420 PRINT_TIMESTAMP_WIDTH, max_latency_end); 1421 } 1422 1423 printf("\n"); 1424 return ret; 1425 } 1426 1427 static int report_print_header(struct perf_kwork *kwork) 1428 { 1429 int ret; 1430 1431 printf("\n "); 1432 ret = printf(" %-*s | %-*s |", 1433 PRINT_KWORK_NAME_WIDTH, "Kwork Name", 1434 PRINT_CPU_WIDTH, "Cpu"); 1435 1436 if (kwork->report == KWORK_REPORT_RUNTIME) { 1437 ret += printf(" %-*s |", 1438 PRINT_RUNTIME_HEADER_WIDTH, "Total Runtime"); 1439 } else if (kwork->report == KWORK_REPORT_LATENCY) { 1440 ret += printf(" %-*s |", 1441 PRINT_LATENCY_HEADER_WIDTH, "Avg delay"); 1442 } 1443 1444 ret += printf(" %-*s |", PRINT_COUNT_WIDTH, "Count"); 1445 1446 if (kwork->report == KWORK_REPORT_RUNTIME) { 1447 ret += printf(" %-*s | %-*s | %-*s |", 1448 PRINT_RUNTIME_HEADER_WIDTH, "Max runtime", 1449 PRINT_TIMESTAMP_HEADER_WIDTH, "Max runtime start", 1450 PRINT_TIMESTAMP_HEADER_WIDTH, "Max runtime end"); 1451 } else if (kwork->report == KWORK_REPORT_LATENCY) { 1452 ret += printf(" %-*s | %-*s | %-*s |", 1453 PRINT_LATENCY_HEADER_WIDTH, "Max delay", 1454 PRINT_TIMESTAMP_HEADER_WIDTH, "Max delay start", 1455 PRINT_TIMESTAMP_HEADER_WIDTH, "Max delay end"); 1456 } 1457 1458 printf("\n"); 1459 print_separator(ret); 1460 return ret; 1461 } 1462 1463 static void timehist_print_header(void) 1464 { 1465 /* 1466 * header row 1467 */ 1468 printf(" %-*s %-*s %-*s %-*s %-*s %-*s\n", 1469 PRINT_TIMESTAMP_WIDTH, "Runtime start", 1470 PRINT_TIMESTAMP_WIDTH, "Runtime end", 1471 PRINT_TIMEHIST_CPU_WIDTH, "Cpu", 1472 PRINT_KWORK_NAME_WIDTH, "Kwork name", 1473 PRINT_RUNTIME_WIDTH, "Runtime", 1474 PRINT_RUNTIME_WIDTH, "Delaytime"); 1475 1476 /* 1477 * units row 1478 */ 1479 printf(" %-*s %-*s %-*s %-*s %-*s %-*s\n", 1480 PRINT_TIMESTAMP_WIDTH, "", 1481 PRINT_TIMESTAMP_WIDTH, "", 1482 PRINT_TIMEHIST_CPU_WIDTH, "", 1483 PRINT_KWORK_NAME_WIDTH, "(TYPE)NAME:NUM", 1484 PRINT_RUNTIME_WIDTH, "(msec)", 1485 PRINT_RUNTIME_WIDTH, "(msec)"); 1486 1487 /* 1488 * separator 1489 */ 1490 printf(" %.*s %.*s %.*s %.*s %.*s %.*s\n", 1491 PRINT_TIMESTAMP_WIDTH, graph_dotted_line, 1492 PRINT_TIMESTAMP_WIDTH, graph_dotted_line, 1493 PRINT_TIMEHIST_CPU_WIDTH, graph_dotted_line, 1494 PRINT_KWORK_NAME_WIDTH, graph_dotted_line, 1495 PRINT_RUNTIME_WIDTH, graph_dotted_line, 1496 PRINT_RUNTIME_WIDTH, graph_dotted_line); 1497 } 1498 1499 static void print_summary(struct perf_kwork *kwork) 1500 { 1501 u64 time = kwork->timeend - kwork->timestart; 1502 1503 printf(" Total count : %9" PRIu64 "\n", kwork->all_count); 1504 printf(" Total runtime (msec) : %9.3f (%.3f%% load average)\n", 1505 (double)kwork->all_runtime / NSEC_PER_MSEC, 1506 time == 0 ? 0 : (double)kwork->all_runtime / time); 1507 printf(" Total time span (msec) : %9.3f\n", 1508 (double)time / NSEC_PER_MSEC); 1509 } 1510 1511 static unsigned long long nr_list_entry(struct list_head *head) 1512 { 1513 struct list_head *pos; 1514 unsigned long long n = 0; 1515 1516 list_for_each(pos, head) 1517 n++; 1518 1519 return n; 1520 } 1521 1522 static void print_skipped_events(struct perf_kwork *kwork) 1523 { 1524 int i; 1525 const char *const kwork_event_str[] = { 1526 [KWORK_TRACE_RAISE] = "raise", 1527 [KWORK_TRACE_ENTRY] = "entry", 1528 [KWORK_TRACE_EXIT] = "exit", 1529 }; 1530 1531 if ((kwork->nr_skipped_events[KWORK_TRACE_MAX] != 0) && 1532 (kwork->nr_events != 0)) { 1533 printf(" INFO: %.3f%% skipped events (%" PRIu64 " including ", 1534 (double)kwork->nr_skipped_events[KWORK_TRACE_MAX] / 1535 (double)kwork->nr_events * 100.0, 1536 kwork->nr_skipped_events[KWORK_TRACE_MAX]); 1537 1538 for (i = 0; i < KWORK_TRACE_MAX; i++) { 1539 printf("%" PRIu64 " %s%s", 1540 kwork->nr_skipped_events[i], 1541 kwork_event_str[i], 1542 (i == KWORK_TRACE_MAX - 1) ? ")\n" : ", "); 1543 } 1544 } 1545 1546 if (verbose > 0) 1547 printf(" INFO: use %lld atom pages\n", 1548 nr_list_entry(&kwork->atom_page_list)); 1549 } 1550 1551 static void print_bad_events(struct perf_kwork *kwork) 1552 { 1553 if ((kwork->nr_lost_events != 0) && (kwork->nr_events != 0)) { 1554 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n", 1555 (double)kwork->nr_lost_events / 1556 (double)kwork->nr_events * 100.0, 1557 kwork->nr_lost_events, kwork->nr_events, 1558 kwork->nr_lost_chunks); 1559 } 1560 } 1561 1562 const char *graph_load = "||||||||||||||||||||||||||||||||||||||||||||||||"; 1563 const char *graph_idle = " "; 1564 static void top_print_per_cpu_load(struct perf_kwork *kwork) 1565 { 1566 int i, load_width; 1567 u64 total, load, load_ratio; 1568 struct kwork_top_stat *stat = &kwork->top_stat; 1569 1570 for (i = 0; i < MAX_NR_CPUS; i++) { 1571 total = stat->cpus_runtime[i].total; 1572 load = stat->cpus_runtime[i].load; 1573 if (test_bit(i, stat->all_cpus_bitmap) && total) { 1574 load_ratio = load * 10000 / total; 1575 load_width = PRINT_CPU_USAGE_HIST_WIDTH * 1576 load_ratio / 10000; 1577 1578 printf("%%Cpu%-*d[%.*s%.*s %*.*f%%]\n", 1579 PRINT_CPU_WIDTH, i, 1580 load_width, graph_load, 1581 PRINT_CPU_USAGE_HIST_WIDTH - load_width, 1582 graph_idle, 1583 PRINT_CPU_USAGE_WIDTH, 1584 PRINT_CPU_USAGE_DECIMAL_WIDTH, 1585 (double)load_ratio / 100); 1586 } 1587 } 1588 } 1589 1590 static void top_print_cpu_usage(struct perf_kwork *kwork) 1591 { 1592 struct kwork_top_stat *stat = &kwork->top_stat; 1593 u64 idle_time = stat->cpus_runtime[MAX_NR_CPUS].idle; 1594 u64 hardirq_time = stat->cpus_runtime[MAX_NR_CPUS].irq; 1595 u64 softirq_time = stat->cpus_runtime[MAX_NR_CPUS].softirq; 1596 int cpus_nr = bitmap_weight(stat->all_cpus_bitmap, MAX_NR_CPUS); 1597 u64 cpus_total_time = stat->cpus_runtime[MAX_NR_CPUS].total; 1598 1599 printf("Total : %*.*f ms, %d cpus\n", 1600 PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH, 1601 (double)cpus_total_time / NSEC_PER_MSEC, 1602 cpus_nr); 1603 1604 printf("%%Cpu(s): %*.*f%% id, %*.*f%% hi, %*.*f%% si\n", 1605 PRINT_CPU_USAGE_WIDTH, PRINT_CPU_USAGE_DECIMAL_WIDTH, 1606 cpus_total_time ? (double)idle_time * 100 / cpus_total_time : 0, 1607 1608 PRINT_CPU_USAGE_WIDTH, PRINT_CPU_USAGE_DECIMAL_WIDTH, 1609 cpus_total_time ? (double)hardirq_time * 100 / cpus_total_time : 0, 1610 1611 PRINT_CPU_USAGE_WIDTH, PRINT_CPU_USAGE_DECIMAL_WIDTH, 1612 cpus_total_time ? (double)softirq_time * 100 / cpus_total_time : 0); 1613 1614 top_print_per_cpu_load(kwork); 1615 } 1616 1617 static void top_print_header(struct perf_kwork *kwork __maybe_unused) 1618 { 1619 int ret; 1620 1621 printf("\n "); 1622 ret = printf(" %*s %s%*s%s %*s %*s %-*s", 1623 PRINT_PID_WIDTH, "PID", 1624 1625 kwork->use_bpf ? " " : "", 1626 kwork->use_bpf ? PRINT_PID_WIDTH : 0, 1627 kwork->use_bpf ? "SPID" : "", 1628 kwork->use_bpf ? " " : "", 1629 1630 PRINT_CPU_USAGE_WIDTH, "%CPU", 1631 PRINT_RUNTIME_HEADER_WIDTH + RPINT_DECIMAL_WIDTH, "RUNTIME", 1632 PRINT_TASK_NAME_WIDTH, "COMMAND"); 1633 printf("\n "); 1634 print_separator(ret); 1635 } 1636 1637 static int top_print_work(struct perf_kwork *kwork __maybe_unused, struct kwork_work *work) 1638 { 1639 int ret = 0; 1640 1641 printf(" "); 1642 1643 /* 1644 * pid 1645 */ 1646 ret += printf(" %*" PRIu64 " ", PRINT_PID_WIDTH, work->id); 1647 1648 /* 1649 * tgid 1650 */ 1651 if (kwork->use_bpf) 1652 ret += printf(" %*d ", PRINT_PID_WIDTH, work->tgid); 1653 1654 /* 1655 * cpu usage 1656 */ 1657 ret += printf(" %*.*f ", 1658 PRINT_CPU_USAGE_WIDTH, PRINT_CPU_USAGE_DECIMAL_WIDTH, 1659 (double)work->cpu_usage / 100); 1660 1661 /* 1662 * total runtime 1663 */ 1664 ret += printf(" %*.*f ms ", 1665 PRINT_RUNTIME_WIDTH + RPINT_DECIMAL_WIDTH, RPINT_DECIMAL_WIDTH, 1666 (double)work->total_runtime / NSEC_PER_MSEC); 1667 1668 /* 1669 * command 1670 */ 1671 if (kwork->use_bpf) 1672 ret += printf(" %s%s%s", 1673 work->is_kthread ? "[" : "", 1674 work->name, 1675 work->is_kthread ? "]" : ""); 1676 else 1677 ret += printf(" %-*s", PRINT_TASK_NAME_WIDTH, work->name); 1678 1679 printf("\n"); 1680 return ret; 1681 } 1682 1683 static void work_sort(struct perf_kwork *kwork, 1684 struct kwork_class *class, struct rb_root_cached *root) 1685 { 1686 struct rb_node *node; 1687 struct kwork_work *data; 1688 1689 pr_debug("Sorting %s ...\n", class->name); 1690 for (;;) { 1691 node = rb_first_cached(root); 1692 if (!node) 1693 break; 1694 1695 rb_erase_cached(node, root); 1696 data = rb_entry(node, struct kwork_work, node); 1697 work_insert(&kwork->sorted_work_root, 1698 data, &kwork->sort_list); 1699 } 1700 } 1701 1702 static void perf_kwork__sort(struct perf_kwork *kwork) 1703 { 1704 struct kwork_class *class; 1705 1706 list_for_each_entry(class, &kwork->class_list, list) 1707 work_sort(kwork, class, &class->work_root); 1708 } 1709 1710 static int perf_kwork__check_config(struct perf_kwork *kwork, 1711 struct perf_session *session) 1712 { 1713 int ret; 1714 struct evsel *evsel; 1715 struct kwork_class *class; 1716 1717 static struct trace_kwork_handler report_ops = { 1718 .entry_event = report_entry_event, 1719 .exit_event = report_exit_event, 1720 }; 1721 static struct trace_kwork_handler latency_ops = { 1722 .raise_event = latency_raise_event, 1723 .entry_event = latency_entry_event, 1724 }; 1725 static struct trace_kwork_handler timehist_ops = { 1726 .raise_event = timehist_raise_event, 1727 .entry_event = timehist_entry_event, 1728 .exit_event = timehist_exit_event, 1729 }; 1730 static struct trace_kwork_handler top_ops = { 1731 .entry_event = timehist_entry_event, 1732 .exit_event = top_exit_event, 1733 .sched_switch_event = top_sched_switch_event, 1734 }; 1735 1736 switch (kwork->report) { 1737 case KWORK_REPORT_RUNTIME: 1738 kwork->tp_handler = &report_ops; 1739 break; 1740 case KWORK_REPORT_LATENCY: 1741 kwork->tp_handler = &latency_ops; 1742 break; 1743 case KWORK_REPORT_TIMEHIST: 1744 kwork->tp_handler = &timehist_ops; 1745 break; 1746 case KWORK_REPORT_TOP: 1747 kwork->tp_handler = &top_ops; 1748 break; 1749 default: 1750 pr_debug("Invalid report type %d\n", kwork->report); 1751 return -1; 1752 } 1753 1754 list_for_each_entry(class, &kwork->class_list, list) 1755 if ((class->class_init != NULL) && 1756 (class->class_init(class, session) != 0)) 1757 return -1; 1758 1759 if (kwork->cpu_list != NULL) { 1760 ret = perf_session__cpu_bitmap(session, 1761 kwork->cpu_list, 1762 kwork->cpu_bitmap); 1763 if (ret < 0) { 1764 pr_err("Invalid cpu bitmap\n"); 1765 return -1; 1766 } 1767 } 1768 1769 if (kwork->time_str != NULL) { 1770 ret = perf_time__parse_str(&kwork->ptime, kwork->time_str); 1771 if (ret != 0) { 1772 pr_err("Invalid time span\n"); 1773 return -1; 1774 } 1775 } 1776 1777 list_for_each_entry(evsel, &session->evlist->core.entries, core.node) { 1778 if (kwork->show_callchain && !evsel__has_callchain(evsel)) { 1779 pr_debug("Samples do not have callchains\n"); 1780 kwork->show_callchain = 0; 1781 symbol_conf.use_callchain = 0; 1782 } 1783 } 1784 1785 return 0; 1786 } 1787 1788 static int perf_kwork__read_events(struct perf_kwork *kwork) 1789 { 1790 int ret = -1; 1791 struct perf_session *session = NULL; 1792 1793 struct perf_data data = { 1794 .path = input_name, 1795 .mode = PERF_DATA_MODE_READ, 1796 .force = kwork->force, 1797 }; 1798 1799 session = perf_session__new(&data, &kwork->tool); 1800 if (IS_ERR(session)) { 1801 pr_debug("Error creating perf session\n"); 1802 return PTR_ERR(session); 1803 } 1804 1805 symbol__init(&session->header.env); 1806 1807 if (perf_kwork__check_config(kwork, session) != 0) 1808 goto out_delete; 1809 1810 if (session->tevent.pevent && 1811 tep_set_function_resolver(session->tevent.pevent, 1812 machine__resolve_kernel_addr, 1813 &session->machines.host) < 0) { 1814 pr_err("Failed to set libtraceevent function resolver\n"); 1815 goto out_delete; 1816 } 1817 1818 if (kwork->report == KWORK_REPORT_TIMEHIST) 1819 timehist_print_header(); 1820 1821 ret = perf_session__process_events(session); 1822 if (ret) { 1823 pr_debug("Failed to process events, error %d\n", ret); 1824 goto out_delete; 1825 } 1826 1827 kwork->nr_events = session->evlist->stats.nr_events[0]; 1828 kwork->nr_lost_events = session->evlist->stats.total_lost; 1829 kwork->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST]; 1830 1831 out_delete: 1832 perf_session__delete(session); 1833 return ret; 1834 } 1835 1836 static void process_skipped_events(struct perf_kwork *kwork, 1837 struct kwork_work *work) 1838 { 1839 int i; 1840 unsigned long long count; 1841 1842 for (i = 0; i < KWORK_TRACE_MAX; i++) { 1843 count = nr_list_entry(&work->atom_list[i]); 1844 kwork->nr_skipped_events[i] += count; 1845 kwork->nr_skipped_events[KWORK_TRACE_MAX] += count; 1846 } 1847 } 1848 1849 struct kwork_work *perf_kwork_add_work(struct perf_kwork *kwork, 1850 struct kwork_class *class, 1851 struct kwork_work *key) 1852 { 1853 struct kwork_work *work = NULL; 1854 1855 work = work_new(key); 1856 if (work == NULL) 1857 return NULL; 1858 1859 work_insert(&class->work_root, work, &kwork->cmp_id); 1860 return work; 1861 } 1862 1863 static void sig_handler(int sig) 1864 { 1865 /* 1866 * Simply capture termination signal so that 1867 * the program can continue after pause returns 1868 */ 1869 pr_debug("Capture signal %d\n", sig); 1870 } 1871 1872 static int perf_kwork__report_bpf(struct perf_kwork *kwork) 1873 { 1874 int ret; 1875 1876 signal(SIGINT, sig_handler); 1877 signal(SIGTERM, sig_handler); 1878 1879 ret = perf_kwork__trace_prepare_bpf(kwork); 1880 if (ret) 1881 return -1; 1882 1883 printf("Starting trace, Hit <Ctrl+C> to stop and report\n"); 1884 1885 perf_kwork__trace_start(); 1886 1887 /* 1888 * a simple pause, wait here for stop signal 1889 */ 1890 pause(); 1891 1892 perf_kwork__trace_finish(); 1893 1894 perf_kwork__report_read_bpf(kwork); 1895 1896 perf_kwork__report_cleanup_bpf(); 1897 1898 return 0; 1899 } 1900 1901 static int perf_kwork__report(struct perf_kwork *kwork) 1902 { 1903 int ret; 1904 struct rb_node *next; 1905 struct kwork_work *work; 1906 1907 if (kwork->use_bpf) 1908 ret = perf_kwork__report_bpf(kwork); 1909 else 1910 ret = perf_kwork__read_events(kwork); 1911 1912 if (ret != 0) 1913 return -1; 1914 1915 perf_kwork__sort(kwork); 1916 1917 setup_pager(); 1918 1919 ret = report_print_header(kwork); 1920 next = rb_first_cached(&kwork->sorted_work_root); 1921 while (next) { 1922 work = rb_entry(next, struct kwork_work, node); 1923 process_skipped_events(kwork, work); 1924 1925 if (work->nr_atoms != 0) { 1926 report_print_work(kwork, work); 1927 if (kwork->summary) { 1928 kwork->all_runtime += work->total_runtime; 1929 kwork->all_count += work->nr_atoms; 1930 } 1931 } 1932 next = rb_next(next); 1933 } 1934 print_separator(ret); 1935 1936 if (kwork->summary) { 1937 print_summary(kwork); 1938 print_separator(ret); 1939 } 1940 1941 print_bad_events(kwork); 1942 print_skipped_events(kwork); 1943 printf("\n"); 1944 1945 return 0; 1946 } 1947 1948 typedef int (*tracepoint_handler)(const struct perf_tool *tool, 1949 struct evsel *evsel, 1950 struct perf_sample *sample, 1951 struct machine *machine); 1952 1953 static int perf_kwork__process_tracepoint_sample(const struct perf_tool *tool, 1954 union perf_event *event __maybe_unused, 1955 struct perf_sample *sample, 1956 struct evsel *evsel, 1957 struct machine *machine) 1958 { 1959 int err = 0; 1960 1961 if (evsel->handler != NULL) { 1962 tracepoint_handler f = evsel->handler; 1963 1964 err = f(tool, evsel, sample, machine); 1965 } 1966 1967 return err; 1968 } 1969 1970 static int perf_kwork__timehist(struct perf_kwork *kwork) 1971 { 1972 /* 1973 * event handlers for timehist option 1974 */ 1975 kwork->tool.comm = perf_event__process_comm; 1976 kwork->tool.exit = perf_event__process_exit; 1977 kwork->tool.fork = perf_event__process_fork; 1978 kwork->tool.attr = perf_event__process_attr; 1979 kwork->tool.tracing_data = perf_event__process_tracing_data; 1980 kwork->tool.build_id = perf_event__process_build_id; 1981 kwork->tool.ordered_events = true; 1982 kwork->tool.ordering_requires_timestamps = true; 1983 symbol_conf.use_callchain = kwork->show_callchain; 1984 1985 if (symbol__validate_sym_arguments()) { 1986 pr_err("Failed to validate sym arguments\n"); 1987 return -1; 1988 } 1989 1990 setup_pager(); 1991 1992 return perf_kwork__read_events(kwork); 1993 } 1994 1995 static void top_calc_total_runtime(struct perf_kwork *kwork) 1996 { 1997 struct kwork_class *class; 1998 struct kwork_work *work; 1999 struct rb_node *next; 2000 struct kwork_top_stat *stat = &kwork->top_stat; 2001 2002 class = get_kwork_class(kwork, KWORK_CLASS_SCHED); 2003 if (!class) 2004 return; 2005 2006 next = rb_first_cached(&class->work_root); 2007 while (next) { 2008 work = rb_entry(next, struct kwork_work, node); 2009 BUG_ON(work->cpu >= MAX_NR_CPUS); 2010 stat->cpus_runtime[work->cpu].total += work->total_runtime; 2011 stat->cpus_runtime[MAX_NR_CPUS].total += work->total_runtime; 2012 next = rb_next(next); 2013 } 2014 } 2015 2016 static void top_calc_idle_time(struct perf_kwork *kwork, 2017 struct kwork_work *work) 2018 { 2019 struct kwork_top_stat *stat = &kwork->top_stat; 2020 2021 if (work->id == 0) { 2022 stat->cpus_runtime[work->cpu].idle += work->total_runtime; 2023 stat->cpus_runtime[MAX_NR_CPUS].idle += work->total_runtime; 2024 } 2025 } 2026 2027 static void top_calc_irq_runtime(struct perf_kwork *kwork, 2028 enum kwork_class_type type, 2029 struct kwork_work *work) 2030 { 2031 struct kwork_top_stat *stat = &kwork->top_stat; 2032 2033 if (type == KWORK_CLASS_IRQ) { 2034 stat->cpus_runtime[work->cpu].irq += work->total_runtime; 2035 stat->cpus_runtime[MAX_NR_CPUS].irq += work->total_runtime; 2036 } else if (type == KWORK_CLASS_SOFTIRQ) { 2037 stat->cpus_runtime[work->cpu].softirq += work->total_runtime; 2038 stat->cpus_runtime[MAX_NR_CPUS].softirq += work->total_runtime; 2039 } 2040 } 2041 2042 static void top_subtract_irq_runtime(struct perf_kwork *kwork, 2043 struct kwork_work *work) 2044 { 2045 struct kwork_class *class; 2046 struct kwork_work *data; 2047 unsigned int i; 2048 int irq_class_list[] = {KWORK_CLASS_IRQ, KWORK_CLASS_SOFTIRQ}; 2049 2050 for (i = 0; i < ARRAY_SIZE(irq_class_list); i++) { 2051 class = get_kwork_class(kwork, irq_class_list[i]); 2052 if (!class) 2053 continue; 2054 2055 data = find_work_by_id(&class->work_root, 2056 work->id, work->cpu); 2057 if (!data) 2058 continue; 2059 2060 if (work->total_runtime > data->total_runtime) { 2061 work->total_runtime -= data->total_runtime; 2062 top_calc_irq_runtime(kwork, irq_class_list[i], data); 2063 } 2064 } 2065 } 2066 2067 static void top_calc_cpu_usage(struct perf_kwork *kwork) 2068 { 2069 struct kwork_class *class; 2070 struct kwork_work *work; 2071 struct rb_node *next; 2072 struct kwork_top_stat *stat = &kwork->top_stat; 2073 2074 class = get_kwork_class(kwork, KWORK_CLASS_SCHED); 2075 if (!class) 2076 return; 2077 2078 next = rb_first_cached(&class->work_root); 2079 while (next) { 2080 work = rb_entry(next, struct kwork_work, node); 2081 2082 if (work->total_runtime == 0) 2083 goto next; 2084 2085 __set_bit(work->cpu, stat->all_cpus_bitmap); 2086 2087 top_subtract_irq_runtime(kwork, work); 2088 2089 work->cpu_usage = work->total_runtime * 10000 / 2090 stat->cpus_runtime[work->cpu].total; 2091 2092 top_calc_idle_time(kwork, work); 2093 next: 2094 next = rb_next(next); 2095 } 2096 } 2097 2098 static void top_calc_load_runtime(struct perf_kwork *kwork, 2099 struct kwork_work *work) 2100 { 2101 struct kwork_top_stat *stat = &kwork->top_stat; 2102 2103 if (work->id != 0) { 2104 stat->cpus_runtime[work->cpu].load += work->total_runtime; 2105 stat->cpus_runtime[MAX_NR_CPUS].load += work->total_runtime; 2106 } 2107 } 2108 2109 static void top_merge_tasks(struct perf_kwork *kwork) 2110 { 2111 struct kwork_work *merged_work, *data; 2112 struct kwork_class *class; 2113 struct rb_node *node; 2114 int cpu; 2115 struct rb_root_cached merged_root = RB_ROOT_CACHED; 2116 2117 class = get_kwork_class(kwork, KWORK_CLASS_SCHED); 2118 if (!class) 2119 return; 2120 2121 for (;;) { 2122 node = rb_first_cached(&class->work_root); 2123 if (!node) 2124 break; 2125 2126 rb_erase_cached(node, &class->work_root); 2127 data = rb_entry(node, struct kwork_work, node); 2128 2129 if (!profile_name_match(kwork, data)) 2130 continue; 2131 2132 cpu = data->cpu; 2133 merged_work = find_work_by_id(&merged_root, data->id, 2134 data->id == 0 ? cpu : -1); 2135 if (!merged_work) { 2136 work_insert(&merged_root, data, &kwork->cmp_id); 2137 } else { 2138 merged_work->total_runtime += data->total_runtime; 2139 merged_work->cpu_usage += data->cpu_usage; 2140 } 2141 2142 top_calc_load_runtime(kwork, data); 2143 } 2144 2145 work_sort(kwork, class, &merged_root); 2146 } 2147 2148 static void perf_kwork__top_report(struct perf_kwork *kwork) 2149 { 2150 struct kwork_work *work; 2151 struct rb_node *next; 2152 2153 printf("\n"); 2154 2155 top_print_cpu_usage(kwork); 2156 top_print_header(kwork); 2157 next = rb_first_cached(&kwork->sorted_work_root); 2158 while (next) { 2159 work = rb_entry(next, struct kwork_work, node); 2160 process_skipped_events(kwork, work); 2161 2162 if (work->total_runtime == 0) 2163 goto next; 2164 2165 top_print_work(kwork, work); 2166 2167 next: 2168 next = rb_next(next); 2169 } 2170 2171 printf("\n"); 2172 } 2173 2174 static int perf_kwork__top_bpf(struct perf_kwork *kwork) 2175 { 2176 int ret; 2177 2178 signal(SIGINT, sig_handler); 2179 signal(SIGTERM, sig_handler); 2180 2181 ret = perf_kwork__top_prepare_bpf(kwork); 2182 if (ret) 2183 return -1; 2184 2185 printf("Starting trace, Hit <Ctrl+C> to stop and report\n"); 2186 2187 perf_kwork__top_start(); 2188 2189 /* 2190 * a simple pause, wait here for stop signal 2191 */ 2192 pause(); 2193 2194 perf_kwork__top_finish(); 2195 2196 perf_kwork__top_read_bpf(kwork); 2197 2198 perf_kwork__top_cleanup_bpf(); 2199 2200 return 0; 2201 2202 } 2203 2204 static int perf_kwork__top(struct perf_kwork *kwork) 2205 { 2206 struct __top_cpus_runtime *cpus_runtime; 2207 int ret = 0; 2208 2209 cpus_runtime = zalloc(sizeof(struct __top_cpus_runtime) * (MAX_NR_CPUS + 1)); 2210 if (!cpus_runtime) 2211 return -1; 2212 2213 kwork->top_stat.cpus_runtime = cpus_runtime; 2214 bitmap_zero(kwork->top_stat.all_cpus_bitmap, MAX_NR_CPUS); 2215 2216 if (kwork->use_bpf) 2217 ret = perf_kwork__top_bpf(kwork); 2218 else 2219 ret = perf_kwork__read_events(kwork); 2220 2221 if (ret) 2222 goto out; 2223 2224 top_calc_total_runtime(kwork); 2225 top_calc_cpu_usage(kwork); 2226 top_merge_tasks(kwork); 2227 2228 setup_pager(); 2229 2230 perf_kwork__top_report(kwork); 2231 2232 out: 2233 zfree(&kwork->top_stat.cpus_runtime); 2234 return ret; 2235 } 2236 2237 static void setup_event_list(struct perf_kwork *kwork, 2238 const struct option *options, 2239 const char * const usage_msg[]) 2240 { 2241 int i; 2242 struct kwork_class *class; 2243 char *tmp, *tok, *str; 2244 2245 /* 2246 * set default events list if not specified 2247 */ 2248 if (kwork->event_list_str == NULL) 2249 kwork->event_list_str = "irq, softirq, workqueue"; 2250 2251 str = strdup(kwork->event_list_str); 2252 for (tok = strtok_r(str, ", ", &tmp); 2253 tok; tok = strtok_r(NULL, ", ", &tmp)) { 2254 for (i = 0; i < KWORK_CLASS_MAX; i++) { 2255 class = kwork_class_supported_list[i]; 2256 if (strcmp(tok, class->name) == 0) { 2257 list_add_tail(&class->list, &kwork->class_list); 2258 break; 2259 } 2260 } 2261 if (i == KWORK_CLASS_MAX) { 2262 usage_with_options_msg(usage_msg, options, 2263 "Unknown --event key: `%s'", tok); 2264 } 2265 } 2266 free(str); 2267 2268 pr_debug("Config event list:"); 2269 list_for_each_entry(class, &kwork->class_list, list) 2270 pr_debug(" %s", class->name); 2271 pr_debug("\n"); 2272 } 2273 2274 static int perf_kwork__record(struct perf_kwork *kwork, 2275 int argc, const char **argv) 2276 { 2277 const char **rec_argv; 2278 unsigned int rec_argc, i, j; 2279 struct kwork_class *class; 2280 2281 const char *const record_args[] = { 2282 "record", 2283 "-a", 2284 "-R", 2285 "-m", "1024", 2286 "-c", "1", 2287 }; 2288 2289 rec_argc = ARRAY_SIZE(record_args) + argc - 1; 2290 2291 list_for_each_entry(class, &kwork->class_list, list) 2292 rec_argc += 2 * class->nr_tracepoints; 2293 2294 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 2295 if (rec_argv == NULL) 2296 return -ENOMEM; 2297 2298 for (i = 0; i < ARRAY_SIZE(record_args); i++) 2299 rec_argv[i] = strdup(record_args[i]); 2300 2301 list_for_each_entry(class, &kwork->class_list, list) { 2302 for (j = 0; j < class->nr_tracepoints; j++) { 2303 rec_argv[i++] = strdup("-e"); 2304 rec_argv[i++] = strdup(class->tp_handlers[j].name); 2305 } 2306 } 2307 2308 for (j = 1; j < (unsigned int)argc; j++, i++) 2309 rec_argv[i] = argv[j]; 2310 2311 BUG_ON(i != rec_argc); 2312 2313 pr_debug("record comm: "); 2314 for (j = 0; j < rec_argc; j++) 2315 pr_debug("%s ", rec_argv[j]); 2316 pr_debug("\n"); 2317 2318 return cmd_record(i, rec_argv); 2319 } 2320 2321 int cmd_kwork(int argc, const char **argv) 2322 { 2323 static struct perf_kwork kwork = { 2324 .class_list = LIST_HEAD_INIT(kwork.class_list), 2325 .atom_page_list = LIST_HEAD_INIT(kwork.atom_page_list), 2326 .sort_list = LIST_HEAD_INIT(kwork.sort_list), 2327 .cmp_id = LIST_HEAD_INIT(kwork.cmp_id), 2328 .sorted_work_root = RB_ROOT_CACHED, 2329 .tp_handler = NULL, 2330 .profile_name = NULL, 2331 .cpu_list = NULL, 2332 .time_str = NULL, 2333 .force = false, 2334 .event_list_str = NULL, 2335 .summary = false, 2336 .sort_order = NULL, 2337 .show_callchain = false, 2338 .max_stack = 5, 2339 .timestart = 0, 2340 .timeend = 0, 2341 .nr_events = 0, 2342 .nr_lost_chunks = 0, 2343 .nr_lost_events = 0, 2344 .all_runtime = 0, 2345 .all_count = 0, 2346 .nr_skipped_events = { 0 }, 2347 }; 2348 static const char default_report_sort_order[] = "runtime, max, count"; 2349 static const char default_latency_sort_order[] = "avg, max, count"; 2350 static const char default_top_sort_order[] = "rate, runtime"; 2351 const struct option kwork_options[] = { 2352 OPT_INCR('v', "verbose", &verbose, 2353 "be more verbose (show symbol address, etc)"), 2354 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, 2355 "dump raw trace in ASCII"), 2356 OPT_STRING('k', "kwork", &kwork.event_list_str, "kwork", 2357 "list of kwork to profile (irq, softirq, workqueue, sched, etc)"), 2358 OPT_BOOLEAN('f', "force", &kwork.force, "don't complain, do it"), 2359 OPT_END() 2360 }; 2361 const struct option report_options[] = { 2362 OPT_STRING('s', "sort", &kwork.sort_order, "key[,key2...]", 2363 "sort by key(s): runtime, max, count"), 2364 OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu", 2365 "list of cpus to profile"), 2366 OPT_STRING('n', "name", &kwork.profile_name, "name", 2367 "event name to profile"), 2368 OPT_STRING(0, "time", &kwork.time_str, "str", 2369 "Time span for analysis (start,stop)"), 2370 OPT_STRING('i', "input", &input_name, "file", 2371 "input file name"), 2372 OPT_BOOLEAN('S', "with-summary", &kwork.summary, 2373 "Show summary with statistics"), 2374 #ifdef HAVE_BPF_SKEL 2375 OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf, 2376 "Use BPF to measure kwork runtime"), 2377 #endif 2378 OPT_PARENT(kwork_options) 2379 }; 2380 const struct option latency_options[] = { 2381 OPT_STRING('s', "sort", &kwork.sort_order, "key[,key2...]", 2382 "sort by key(s): avg, max, count"), 2383 OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu", 2384 "list of cpus to profile"), 2385 OPT_STRING('n', "name", &kwork.profile_name, "name", 2386 "event name to profile"), 2387 OPT_STRING(0, "time", &kwork.time_str, "str", 2388 "Time span for analysis (start,stop)"), 2389 OPT_STRING('i', "input", &input_name, "file", 2390 "input file name"), 2391 #ifdef HAVE_BPF_SKEL 2392 OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf, 2393 "Use BPF to measure kwork latency"), 2394 #endif 2395 OPT_PARENT(kwork_options) 2396 }; 2397 const struct option timehist_options[] = { 2398 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 2399 "file", "vmlinux pathname"), 2400 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 2401 "file", "kallsyms pathname"), 2402 OPT_BOOLEAN('g', "call-graph", &kwork.show_callchain, 2403 "Display call chains if present"), 2404 OPT_UINTEGER(0, "max-stack", &kwork.max_stack, 2405 "Maximum number of functions to display backtrace."), 2406 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory", 2407 "Look for files with symbols relative to this directory"), 2408 OPT_STRING(0, "time", &kwork.time_str, "str", 2409 "Time span for analysis (start,stop)"), 2410 OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu", 2411 "list of cpus to profile"), 2412 OPT_STRING('n', "name", &kwork.profile_name, "name", 2413 "event name to profile"), 2414 OPT_STRING('i', "input", &input_name, "file", 2415 "input file name"), 2416 OPT_PARENT(kwork_options) 2417 }; 2418 const struct option top_options[] = { 2419 OPT_STRING('s', "sort", &kwork.sort_order, "key[,key2...]", 2420 "sort by key(s): rate, runtime, tid"), 2421 OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu", 2422 "list of cpus to profile"), 2423 OPT_STRING('n', "name", &kwork.profile_name, "name", 2424 "event name to profile"), 2425 OPT_STRING(0, "time", &kwork.time_str, "str", 2426 "Time span for analysis (start,stop)"), 2427 OPT_STRING('i', "input", &input_name, "file", 2428 "input file name"), 2429 #ifdef HAVE_BPF_SKEL 2430 OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf, 2431 "Use BPF to measure task cpu usage"), 2432 #endif 2433 OPT_PARENT(kwork_options) 2434 }; 2435 const char *kwork_usage[] = { 2436 NULL, 2437 NULL 2438 }; 2439 const char * const report_usage[] = { 2440 "perf kwork report [<options>]", 2441 NULL 2442 }; 2443 const char * const latency_usage[] = { 2444 "perf kwork latency [<options>]", 2445 NULL 2446 }; 2447 const char * const timehist_usage[] = { 2448 "perf kwork timehist [<options>]", 2449 NULL 2450 }; 2451 const char * const top_usage[] = { 2452 "perf kwork top [<options>]", 2453 NULL 2454 }; 2455 const char *const kwork_subcommands[] = { 2456 "record", "report", "latency", "timehist", "top", NULL 2457 }; 2458 2459 perf_tool__init(&kwork.tool, /*ordered_events=*/true); 2460 kwork.tool.mmap = perf_event__process_mmap; 2461 kwork.tool.mmap2 = perf_event__process_mmap2; 2462 kwork.tool.sample = perf_kwork__process_tracepoint_sample; 2463 2464 argc = parse_options_subcommand(argc, argv, kwork_options, 2465 kwork_subcommands, kwork_usage, 2466 PARSE_OPT_STOP_AT_NON_OPTION); 2467 if (!argc) 2468 usage_with_options(kwork_usage, kwork_options); 2469 2470 sort_dimension__add(&kwork, "id", &kwork.cmp_id); 2471 2472 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2473 setup_event_list(&kwork, kwork_options, kwork_usage); 2474 return perf_kwork__record(&kwork, argc, argv); 2475 } else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) { 2476 kwork.sort_order = default_report_sort_order; 2477 if (argc > 1) { 2478 argc = parse_options(argc, argv, report_options, report_usage, 0); 2479 if (argc) 2480 usage_with_options(report_usage, report_options); 2481 } 2482 kwork.report = KWORK_REPORT_RUNTIME; 2483 setup_sorting(&kwork, report_options, report_usage); 2484 setup_event_list(&kwork, kwork_options, kwork_usage); 2485 return perf_kwork__report(&kwork); 2486 } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) { 2487 kwork.sort_order = default_latency_sort_order; 2488 if (argc > 1) { 2489 argc = parse_options(argc, argv, latency_options, latency_usage, 0); 2490 if (argc) 2491 usage_with_options(latency_usage, latency_options); 2492 } 2493 kwork.report = KWORK_REPORT_LATENCY; 2494 setup_sorting(&kwork, latency_options, latency_usage); 2495 setup_event_list(&kwork, kwork_options, kwork_usage); 2496 return perf_kwork__report(&kwork); 2497 } else if (strlen(argv[0]) > 2 && strstarts("timehist", argv[0])) { 2498 if (argc > 1) { 2499 argc = parse_options(argc, argv, timehist_options, timehist_usage, 0); 2500 if (argc) 2501 usage_with_options(timehist_usage, timehist_options); 2502 } 2503 kwork.report = KWORK_REPORT_TIMEHIST; 2504 setup_event_list(&kwork, kwork_options, kwork_usage); 2505 return perf_kwork__timehist(&kwork); 2506 } else if (strlen(argv[0]) > 2 && strstarts("top", argv[0])) { 2507 kwork.sort_order = default_top_sort_order; 2508 if (argc > 1) { 2509 argc = parse_options(argc, argv, top_options, top_usage, 0); 2510 if (argc) 2511 usage_with_options(top_usage, top_options); 2512 } 2513 kwork.report = KWORK_REPORT_TOP; 2514 if (!kwork.event_list_str) 2515 kwork.event_list_str = "sched, irq, softirq"; 2516 setup_event_list(&kwork, kwork_options, kwork_usage); 2517 setup_sorting(&kwork, top_options, top_usage); 2518 return perf_kwork__top(&kwork); 2519 } else 2520 usage_with_options(kwork_usage, kwork_options); 2521 2522 /* free usage string allocated by parse_options_subcommand */ 2523 free((void *)kwork_usage[0]); 2524 2525 return 0; 2526 } 2527