1 // SPDX-License-Identifier: GPL-2.0 2 #include <errno.h> 3 #include <inttypes.h> 4 #include "builtin.h" 5 #include "perf.h" 6 7 #include "util/evlist.h" // for struct evsel_str_handler 8 #include "util/evsel.h" 9 #include "util/symbol.h" 10 #include "util/thread.h" 11 #include "util/header.h" 12 #include "util/target.h" 13 #include "util/cgroup.h" 14 #include "util/callchain.h" 15 #include "util/lock-contention.h" 16 #include "util/bpf_skel/lock_data.h" 17 18 #include <subcmd/pager.h> 19 #include <subcmd/parse-options.h> 20 #include "util/trace-event.h" 21 #include "util/tracepoint.h" 22 23 #include "util/debug.h" 24 #include "util/session.h" 25 #include "util/tool.h" 26 #include "util/data.h" 27 #include "util/string2.h" 28 #include "util/map.h" 29 #include "util/util.h" 30 31 #include <stdio.h> 32 #include <sys/types.h> 33 #include <sys/prctl.h> 34 #include <semaphore.h> 35 #include <math.h> 36 #include <limits.h> 37 #include <ctype.h> 38 39 #include <linux/list.h> 40 #include <linux/hash.h> 41 #include <linux/kernel.h> 42 #include <linux/zalloc.h> 43 #include <linux/err.h> 44 #include <linux/stringify.h> 45 46 static struct perf_session *session; 47 static struct target target; 48 49 static struct rb_root thread_stats; 50 51 static bool combine_locks; 52 static bool show_thread_stats; 53 static bool show_lock_addrs; 54 static bool show_lock_owner; 55 static bool show_lock_cgroups; 56 static bool use_bpf; 57 static unsigned long bpf_map_entries = MAX_ENTRIES; 58 static int max_stack_depth = CONTENTION_STACK_DEPTH; 59 static int stack_skip = CONTENTION_STACK_SKIP; 60 static int print_nr_entries = INT_MAX / 2; 61 static const char *output_name = NULL; 62 static FILE *lock_output; 63 64 static struct lock_filter filters; 65 static struct lock_delay *delays; 66 static int nr_delays; 67 68 static enum lock_aggr_mode aggr_mode = LOCK_AGGR_ADDR; 69 70 static struct thread_stat *thread_stat_find(u32 tid) 71 { 72 struct rb_node *node; 73 struct thread_stat *st; 74 75 node = thread_stats.rb_node; 76 while (node) { 77 st = container_of(node, struct thread_stat, rb); 78 if (st->tid == tid) 79 return st; 80 else if (tid < st->tid) 81 node = node->rb_left; 82 else 83 node = node->rb_right; 84 } 85 86 return NULL; 87 } 88 89 static void thread_stat_insert(struct thread_stat *new) 90 { 91 struct rb_node **rb = &thread_stats.rb_node; 92 struct rb_node *parent = NULL; 93 struct thread_stat *p; 94 95 while (*rb) { 96 p = container_of(*rb, struct thread_stat, rb); 97 parent = *rb; 98 99 if (new->tid < p->tid) 100 rb = &(*rb)->rb_left; 101 else if (new->tid > p->tid) 102 rb = &(*rb)->rb_right; 103 else 104 BUG_ON("inserting invalid thread_stat\n"); 105 } 106 107 rb_link_node(&new->rb, parent, rb); 108 rb_insert_color(&new->rb, &thread_stats); 109 } 110 111 static struct thread_stat *thread_stat_findnew_after_first(u32 tid) 112 { 113 struct thread_stat *st; 114 115 st = thread_stat_find(tid); 116 if (st) 117 return st; 118 119 st = zalloc(sizeof(struct thread_stat)); 120 if (!st) { 121 pr_err("memory allocation failed\n"); 122 return NULL; 123 } 124 125 st->tid = tid; 126 INIT_LIST_HEAD(&st->seq_list); 127 128 thread_stat_insert(st); 129 130 return st; 131 } 132 133 static struct thread_stat *thread_stat_findnew_first(u32 tid); 134 static struct thread_stat *(*thread_stat_findnew)(u32 tid) = 135 thread_stat_findnew_first; 136 137 static struct thread_stat *thread_stat_findnew_first(u32 tid) 138 { 139 struct thread_stat *st; 140 141 st = zalloc(sizeof(struct thread_stat)); 142 if (!st) { 143 pr_err("memory allocation failed\n"); 144 return NULL; 145 } 146 st->tid = tid; 147 INIT_LIST_HEAD(&st->seq_list); 148 149 rb_link_node(&st->rb, NULL, &thread_stats.rb_node); 150 rb_insert_color(&st->rb, &thread_stats); 151 152 thread_stat_findnew = thread_stat_findnew_after_first; 153 return st; 154 } 155 156 /* build simple key function one is bigger than two */ 157 #define SINGLE_KEY(member) \ 158 static int lock_stat_key_ ## member(struct lock_stat *one, \ 159 struct lock_stat *two) \ 160 { \ 161 return one->member > two->member; \ 162 } 163 164 SINGLE_KEY(nr_acquired) 165 SINGLE_KEY(nr_contended) 166 SINGLE_KEY(avg_wait_time) 167 SINGLE_KEY(wait_time_total) 168 SINGLE_KEY(wait_time_max) 169 170 static int lock_stat_key_wait_time_min(struct lock_stat *one, 171 struct lock_stat *two) 172 { 173 u64 s1 = one->wait_time_min; 174 u64 s2 = two->wait_time_min; 175 if (s1 == ULLONG_MAX) 176 s1 = 0; 177 if (s2 == ULLONG_MAX) 178 s2 = 0; 179 return s1 > s2; 180 } 181 182 struct lock_key { 183 /* 184 * name: the value for specify by user 185 * this should be simpler than raw name of member 186 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total 187 */ 188 const char *name; 189 /* header: the string printed on the header line */ 190 const char *header; 191 /* len: the printing width of the field */ 192 int len; 193 /* key: a pointer to function to compare two lock stats for sorting */ 194 int (*key)(struct lock_stat*, struct lock_stat*); 195 /* print: a pointer to function to print a given lock stats */ 196 void (*print)(struct lock_key*, struct lock_stat*); 197 /* list: list entry to link this */ 198 struct list_head list; 199 }; 200 201 static void lock_stat_key_print_time(unsigned long long nsec, int len) 202 { 203 static const struct { 204 float base; 205 const char *unit; 206 } table[] = { 207 { 1e9 * 3600, "h " }, 208 { 1e9 * 60, "m " }, 209 { 1e9, "s " }, 210 { 1e6, "ms" }, 211 { 1e3, "us" }, 212 { 0, NULL }, 213 }; 214 215 /* for CSV output */ 216 if (len == 0) { 217 fprintf(lock_output, "%llu", nsec); 218 return; 219 } 220 221 for (int i = 0; table[i].unit; i++) { 222 if (nsec < table[i].base) 223 continue; 224 225 fprintf(lock_output, "%*.2f %s", len - 3, nsec / table[i].base, table[i].unit); 226 return; 227 } 228 229 fprintf(lock_output, "%*llu %s", len - 3, nsec, "ns"); 230 } 231 232 #define PRINT_KEY(member) \ 233 static void lock_stat_key_print_ ## member(struct lock_key *key, \ 234 struct lock_stat *ls) \ 235 { \ 236 fprintf(lock_output, "%*llu", key->len, (unsigned long long)ls->member);\ 237 } 238 239 #define PRINT_TIME(member) \ 240 static void lock_stat_key_print_ ## member(struct lock_key *key, \ 241 struct lock_stat *ls) \ 242 { \ 243 lock_stat_key_print_time((unsigned long long)ls->member, key->len); \ 244 } 245 246 PRINT_KEY(nr_acquired) 247 PRINT_KEY(nr_contended) 248 PRINT_TIME(avg_wait_time) 249 PRINT_TIME(wait_time_total) 250 PRINT_TIME(wait_time_max) 251 252 static void lock_stat_key_print_wait_time_min(struct lock_key *key, 253 struct lock_stat *ls) 254 { 255 u64 wait_time = ls->wait_time_min; 256 257 if (wait_time == ULLONG_MAX) 258 wait_time = 0; 259 260 lock_stat_key_print_time(wait_time, key->len); 261 } 262 263 264 static const char *sort_key = "acquired"; 265 266 static int (*compare)(struct lock_stat *, struct lock_stat *); 267 268 static struct rb_root sorted; /* place to store intermediate data */ 269 static struct rb_root result; /* place to store sorted data */ 270 271 static LIST_HEAD(lock_keys); 272 static const char *output_fields; 273 274 #define DEF_KEY_LOCK(name, header, fn_suffix, len) \ 275 { #name, header, len, lock_stat_key_ ## fn_suffix, lock_stat_key_print_ ## fn_suffix, {} } 276 static struct lock_key report_keys[] = { 277 DEF_KEY_LOCK(acquired, "acquired", nr_acquired, 10), 278 DEF_KEY_LOCK(contended, "contended", nr_contended, 10), 279 DEF_KEY_LOCK(avg_wait, "avg wait", avg_wait_time, 12), 280 DEF_KEY_LOCK(wait_total, "total wait", wait_time_total, 12), 281 DEF_KEY_LOCK(wait_max, "max wait", wait_time_max, 12), 282 DEF_KEY_LOCK(wait_min, "min wait", wait_time_min, 12), 283 284 /* extra comparisons much complicated should be here */ 285 { } 286 }; 287 288 static struct lock_key contention_keys[] = { 289 DEF_KEY_LOCK(contended, "contended", nr_contended, 10), 290 DEF_KEY_LOCK(wait_total, "total wait", wait_time_total, 12), 291 DEF_KEY_LOCK(wait_max, "max wait", wait_time_max, 12), 292 DEF_KEY_LOCK(wait_min, "min wait", wait_time_min, 12), 293 DEF_KEY_LOCK(avg_wait, "avg wait", avg_wait_time, 12), 294 295 /* extra comparisons much complicated should be here */ 296 { } 297 }; 298 299 static int select_key(bool contention) 300 { 301 int i; 302 struct lock_key *keys = report_keys; 303 304 if (contention) 305 keys = contention_keys; 306 307 for (i = 0; keys[i].name; i++) { 308 if (!strcmp(keys[i].name, sort_key)) { 309 compare = keys[i].key; 310 311 /* selected key should be in the output fields */ 312 if (list_empty(&keys[i].list)) 313 list_add_tail(&keys[i].list, &lock_keys); 314 315 return 0; 316 } 317 } 318 319 pr_err("Unknown compare key: %s\n", sort_key); 320 return -1; 321 } 322 323 static int add_output_field(bool contention, char *name) 324 { 325 int i; 326 struct lock_key *keys = report_keys; 327 328 if (contention) 329 keys = contention_keys; 330 331 for (i = 0; keys[i].name; i++) { 332 if (strcmp(keys[i].name, name)) 333 continue; 334 335 /* prevent double link */ 336 if (list_empty(&keys[i].list)) 337 list_add_tail(&keys[i].list, &lock_keys); 338 339 return 0; 340 } 341 342 pr_err("Unknown output field: %s\n", name); 343 return -1; 344 } 345 346 static int setup_output_field(bool contention, const char *str) 347 { 348 char *tok, *tmp, *orig; 349 int i, ret = 0; 350 struct lock_key *keys = report_keys; 351 352 if (contention) 353 keys = contention_keys; 354 355 /* no output field given: use all of them */ 356 if (str == NULL) { 357 for (i = 0; keys[i].name; i++) 358 list_add_tail(&keys[i].list, &lock_keys); 359 return 0; 360 } 361 362 for (i = 0; keys[i].name; i++) 363 INIT_LIST_HEAD(&keys[i].list); 364 365 orig = tmp = strdup(str); 366 if (orig == NULL) 367 return -ENOMEM; 368 369 while ((tok = strsep(&tmp, ",")) != NULL){ 370 ret = add_output_field(contention, tok); 371 if (ret < 0) 372 break; 373 } 374 free(orig); 375 376 return ret; 377 } 378 379 static void combine_lock_stats(struct lock_stat *st) 380 { 381 struct rb_node **rb = &sorted.rb_node; 382 struct rb_node *parent = NULL; 383 struct lock_stat *p; 384 int ret; 385 386 while (*rb) { 387 p = container_of(*rb, struct lock_stat, rb); 388 parent = *rb; 389 390 if (st->name && p->name) 391 ret = strcmp(st->name, p->name); 392 else 393 ret = !!st->name - !!p->name; 394 395 if (ret == 0) { 396 p->nr_acquired += st->nr_acquired; 397 p->nr_contended += st->nr_contended; 398 p->wait_time_total += st->wait_time_total; 399 400 if (p->nr_contended) 401 p->avg_wait_time = p->wait_time_total / p->nr_contended; 402 403 if (p->wait_time_min > st->wait_time_min) 404 p->wait_time_min = st->wait_time_min; 405 if (p->wait_time_max < st->wait_time_max) 406 p->wait_time_max = st->wait_time_max; 407 408 p->broken |= st->broken; 409 st->combined = 1; 410 return; 411 } 412 413 if (ret < 0) 414 rb = &(*rb)->rb_left; 415 else 416 rb = &(*rb)->rb_right; 417 } 418 419 rb_link_node(&st->rb, parent, rb); 420 rb_insert_color(&st->rb, &sorted); 421 } 422 423 static void insert_to(struct rb_root *rr, struct lock_stat *st, 424 int (*bigger)(struct lock_stat *, struct lock_stat *)) 425 { 426 struct rb_node **rb = &rr->rb_node; 427 struct rb_node *parent = NULL; 428 struct lock_stat *p; 429 430 while (*rb) { 431 p = container_of(*rb, struct lock_stat, rb); 432 parent = *rb; 433 434 if (bigger(st, p)) 435 rb = &(*rb)->rb_left; 436 else 437 rb = &(*rb)->rb_right; 438 } 439 440 rb_link_node(&st->rb, parent, rb); 441 rb_insert_color(&st->rb, rr); 442 } 443 444 static inline void insert_to_result(struct lock_stat *st, 445 int (*bigger)(struct lock_stat *, 446 struct lock_stat *)) 447 { 448 if (combine_locks && st->combined) 449 return; 450 insert_to(&result, st, bigger); 451 } 452 453 static inline struct lock_stat *pop_from(struct rb_root *rr) 454 { 455 struct rb_node *node = rr->rb_node; 456 457 if (!node) 458 return NULL; 459 460 while (node->rb_left) 461 node = node->rb_left; 462 463 rb_erase(node, rr); 464 return container_of(node, struct lock_stat, rb); 465 466 } 467 468 /* returns left most element of result, and erase it */ 469 static struct lock_stat *pop_from_result(void) 470 { 471 return pop_from(&result); 472 } 473 474 struct trace_lock_handler { 475 /* it's used on CONFIG_LOCKDEP */ 476 int (*acquire_event)(struct evsel *evsel, 477 struct perf_sample *sample); 478 479 /* it's used on CONFIG_LOCKDEP && CONFIG_LOCK_STAT */ 480 int (*acquired_event)(struct evsel *evsel, 481 struct perf_sample *sample); 482 483 /* it's used on CONFIG_LOCKDEP && CONFIG_LOCK_STAT */ 484 int (*contended_event)(struct evsel *evsel, 485 struct perf_sample *sample); 486 487 /* it's used on CONFIG_LOCKDEP */ 488 int (*release_event)(struct evsel *evsel, 489 struct perf_sample *sample); 490 491 /* it's used when CONFIG_LOCKDEP is off */ 492 int (*contention_begin_event)(struct evsel *evsel, 493 struct perf_sample *sample); 494 495 /* it's used when CONFIG_LOCKDEP is off */ 496 int (*contention_end_event)(struct evsel *evsel, 497 struct perf_sample *sample); 498 }; 499 500 static struct lock_seq_stat *get_seq(struct thread_stat *ts, u64 addr) 501 { 502 struct lock_seq_stat *seq; 503 504 list_for_each_entry(seq, &ts->seq_list, list) { 505 if (seq->addr == addr) 506 return seq; 507 } 508 509 seq = zalloc(sizeof(struct lock_seq_stat)); 510 if (!seq) { 511 pr_err("memory allocation failed\n"); 512 return NULL; 513 } 514 seq->state = SEQ_STATE_UNINITIALIZED; 515 seq->addr = addr; 516 517 list_add(&seq->list, &ts->seq_list); 518 return seq; 519 } 520 521 enum broken_state { 522 BROKEN_ACQUIRE, 523 BROKEN_ACQUIRED, 524 BROKEN_CONTENDED, 525 BROKEN_RELEASE, 526 BROKEN_MAX, 527 }; 528 529 static int bad_hist[BROKEN_MAX]; 530 531 enum acquire_flags { 532 TRY_LOCK = 1, 533 READ_LOCK = 2, 534 }; 535 536 static int get_key_by_aggr_mode_simple(u64 *key, u64 addr, u32 tid) 537 { 538 switch (aggr_mode) { 539 case LOCK_AGGR_ADDR: 540 *key = addr; 541 break; 542 case LOCK_AGGR_TASK: 543 *key = tid; 544 break; 545 case LOCK_AGGR_CALLER: 546 case LOCK_AGGR_CGROUP: 547 default: 548 pr_err("Invalid aggregation mode: %d\n", aggr_mode); 549 return -EINVAL; 550 } 551 return 0; 552 } 553 554 static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample); 555 556 static int get_key_by_aggr_mode(u64 *key, u64 addr, struct evsel *evsel, 557 struct perf_sample *sample) 558 { 559 if (aggr_mode == LOCK_AGGR_CALLER) { 560 *key = callchain_id(evsel, sample); 561 return 0; 562 } 563 return get_key_by_aggr_mode_simple(key, addr, sample->tid); 564 } 565 566 static int report_lock_acquire_event(struct evsel *evsel, 567 struct perf_sample *sample) 568 { 569 struct lock_stat *ls; 570 struct thread_stat *ts; 571 struct lock_seq_stat *seq; 572 const char *name = evsel__strval(evsel, sample, "name"); 573 u64 addr = evsel__intval(evsel, sample, "lockdep_addr"); 574 int flag = evsel__intval(evsel, sample, "flags"); 575 u64 key; 576 int ret; 577 578 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid); 579 if (ret < 0) 580 return ret; 581 582 ls = lock_stat_findnew(key, name, 0); 583 if (!ls) 584 return -ENOMEM; 585 586 ts = thread_stat_findnew(sample->tid); 587 if (!ts) 588 return -ENOMEM; 589 590 seq = get_seq(ts, addr); 591 if (!seq) 592 return -ENOMEM; 593 594 switch (seq->state) { 595 case SEQ_STATE_UNINITIALIZED: 596 case SEQ_STATE_RELEASED: 597 if (!flag) { 598 seq->state = SEQ_STATE_ACQUIRING; 599 } else { 600 if (flag & TRY_LOCK) 601 ls->nr_trylock++; 602 if (flag & READ_LOCK) 603 ls->nr_readlock++; 604 seq->state = SEQ_STATE_READ_ACQUIRED; 605 seq->read_count = 1; 606 ls->nr_acquired++; 607 } 608 break; 609 case SEQ_STATE_READ_ACQUIRED: 610 if (flag & READ_LOCK) { 611 seq->read_count++; 612 ls->nr_acquired++; 613 goto end; 614 } else { 615 goto broken; 616 } 617 break; 618 case SEQ_STATE_ACQUIRED: 619 case SEQ_STATE_ACQUIRING: 620 case SEQ_STATE_CONTENDED: 621 broken: 622 /* broken lock sequence */ 623 if (!ls->broken) { 624 ls->broken = 1; 625 bad_hist[BROKEN_ACQUIRE]++; 626 } 627 list_del_init(&seq->list); 628 free(seq); 629 goto end; 630 default: 631 BUG_ON("Unknown state of lock sequence found!\n"); 632 break; 633 } 634 635 ls->nr_acquire++; 636 seq->prev_event_time = sample->time; 637 end: 638 return 0; 639 } 640 641 static int report_lock_acquired_event(struct evsel *evsel, 642 struct perf_sample *sample) 643 { 644 struct lock_stat *ls; 645 struct thread_stat *ts; 646 struct lock_seq_stat *seq; 647 u64 contended_term; 648 const char *name = evsel__strval(evsel, sample, "name"); 649 u64 addr = evsel__intval(evsel, sample, "lockdep_addr"); 650 u64 key; 651 int ret; 652 653 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid); 654 if (ret < 0) 655 return ret; 656 657 ls = lock_stat_findnew(key, name, 0); 658 if (!ls) 659 return -ENOMEM; 660 661 ts = thread_stat_findnew(sample->tid); 662 if (!ts) 663 return -ENOMEM; 664 665 seq = get_seq(ts, addr); 666 if (!seq) 667 return -ENOMEM; 668 669 switch (seq->state) { 670 case SEQ_STATE_UNINITIALIZED: 671 /* orphan event, do nothing */ 672 return 0; 673 case SEQ_STATE_ACQUIRING: 674 break; 675 case SEQ_STATE_CONTENDED: 676 contended_term = sample->time - seq->prev_event_time; 677 ls->wait_time_total += contended_term; 678 if (contended_term < ls->wait_time_min) 679 ls->wait_time_min = contended_term; 680 if (ls->wait_time_max < contended_term) 681 ls->wait_time_max = contended_term; 682 break; 683 case SEQ_STATE_RELEASED: 684 case SEQ_STATE_ACQUIRED: 685 case SEQ_STATE_READ_ACQUIRED: 686 /* broken lock sequence */ 687 if (!ls->broken) { 688 ls->broken = 1; 689 bad_hist[BROKEN_ACQUIRED]++; 690 } 691 list_del_init(&seq->list); 692 free(seq); 693 goto end; 694 default: 695 BUG_ON("Unknown state of lock sequence found!\n"); 696 break; 697 } 698 699 seq->state = SEQ_STATE_ACQUIRED; 700 ls->nr_acquired++; 701 ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0; 702 seq->prev_event_time = sample->time; 703 end: 704 return 0; 705 } 706 707 static int report_lock_contended_event(struct evsel *evsel, 708 struct perf_sample *sample) 709 { 710 struct lock_stat *ls; 711 struct thread_stat *ts; 712 struct lock_seq_stat *seq; 713 const char *name = evsel__strval(evsel, sample, "name"); 714 u64 addr = evsel__intval(evsel, sample, "lockdep_addr"); 715 u64 key; 716 int ret; 717 718 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid); 719 if (ret < 0) 720 return ret; 721 722 ls = lock_stat_findnew(key, name, 0); 723 if (!ls) 724 return -ENOMEM; 725 726 ts = thread_stat_findnew(sample->tid); 727 if (!ts) 728 return -ENOMEM; 729 730 seq = get_seq(ts, addr); 731 if (!seq) 732 return -ENOMEM; 733 734 switch (seq->state) { 735 case SEQ_STATE_UNINITIALIZED: 736 /* orphan event, do nothing */ 737 return 0; 738 case SEQ_STATE_ACQUIRING: 739 break; 740 case SEQ_STATE_RELEASED: 741 case SEQ_STATE_ACQUIRED: 742 case SEQ_STATE_READ_ACQUIRED: 743 case SEQ_STATE_CONTENDED: 744 /* broken lock sequence */ 745 if (!ls->broken) { 746 ls->broken = 1; 747 bad_hist[BROKEN_CONTENDED]++; 748 } 749 list_del_init(&seq->list); 750 free(seq); 751 goto end; 752 default: 753 BUG_ON("Unknown state of lock sequence found!\n"); 754 break; 755 } 756 757 seq->state = SEQ_STATE_CONTENDED; 758 ls->nr_contended++; 759 ls->avg_wait_time = ls->wait_time_total/ls->nr_contended; 760 seq->prev_event_time = sample->time; 761 end: 762 return 0; 763 } 764 765 static int report_lock_release_event(struct evsel *evsel, 766 struct perf_sample *sample) 767 { 768 struct lock_stat *ls; 769 struct thread_stat *ts; 770 struct lock_seq_stat *seq; 771 const char *name = evsel__strval(evsel, sample, "name"); 772 u64 addr = evsel__intval(evsel, sample, "lockdep_addr"); 773 u64 key; 774 int ret; 775 776 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid); 777 if (ret < 0) 778 return ret; 779 780 ls = lock_stat_findnew(key, name, 0); 781 if (!ls) 782 return -ENOMEM; 783 784 ts = thread_stat_findnew(sample->tid); 785 if (!ts) 786 return -ENOMEM; 787 788 seq = get_seq(ts, addr); 789 if (!seq) 790 return -ENOMEM; 791 792 switch (seq->state) { 793 case SEQ_STATE_UNINITIALIZED: 794 goto end; 795 case SEQ_STATE_ACQUIRED: 796 break; 797 case SEQ_STATE_READ_ACQUIRED: 798 seq->read_count--; 799 BUG_ON(seq->read_count < 0); 800 if (seq->read_count) { 801 ls->nr_release++; 802 goto end; 803 } 804 break; 805 case SEQ_STATE_ACQUIRING: 806 case SEQ_STATE_CONTENDED: 807 case SEQ_STATE_RELEASED: 808 /* broken lock sequence */ 809 if (!ls->broken) { 810 ls->broken = 1; 811 bad_hist[BROKEN_RELEASE]++; 812 } 813 goto free_seq; 814 default: 815 BUG_ON("Unknown state of lock sequence found!\n"); 816 break; 817 } 818 819 ls->nr_release++; 820 free_seq: 821 list_del_init(&seq->list); 822 free(seq); 823 end: 824 return 0; 825 } 826 827 static int get_symbol_name_offset(struct map *map, struct symbol *sym, u64 ip, 828 char *buf, int size) 829 { 830 u64 offset; 831 832 if (map == NULL || sym == NULL) { 833 buf[0] = '\0'; 834 return 0; 835 } 836 837 offset = map__map_ip(map, ip) - sym->start; 838 839 if (offset) 840 return scnprintf(buf, size, "%s+%#lx", sym->name, offset); 841 else 842 return strlcpy(buf, sym->name, size); 843 } 844 static int lock_contention_caller(struct evsel *evsel, struct perf_sample *sample, 845 char *buf, int size) 846 { 847 struct thread *thread; 848 struct callchain_cursor *cursor; 849 struct machine *machine = &session->machines.host; 850 struct symbol *sym; 851 int skip = 0; 852 int ret; 853 854 /* lock names will be replaced to task name later */ 855 if (show_thread_stats) 856 return -1; 857 858 thread = machine__findnew_thread(machine, -1, sample->pid); 859 if (thread == NULL) 860 return -1; 861 862 cursor = get_tls_callchain_cursor(); 863 864 /* use caller function name from the callchain */ 865 ret = thread__resolve_callchain(thread, cursor, evsel, sample, 866 NULL, NULL, max_stack_depth); 867 if (ret != 0) { 868 thread__put(thread); 869 return -1; 870 } 871 872 callchain_cursor_commit(cursor); 873 thread__put(thread); 874 875 while (true) { 876 struct callchain_cursor_node *node; 877 878 node = callchain_cursor_current(cursor); 879 if (node == NULL) 880 break; 881 882 /* skip first few entries - for lock functions */ 883 if (++skip <= stack_skip) 884 goto next; 885 886 sym = node->ms.sym; 887 if (sym && !machine__is_lock_function(machine, node->ip)) { 888 get_symbol_name_offset(node->ms.map, sym, node->ip, 889 buf, size); 890 return 0; 891 } 892 893 next: 894 callchain_cursor_advance(cursor); 895 } 896 return -1; 897 } 898 899 static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample) 900 { 901 struct callchain_cursor *cursor; 902 struct machine *machine = &session->machines.host; 903 struct thread *thread; 904 u64 hash = 0; 905 int skip = 0; 906 int ret; 907 908 thread = machine__findnew_thread(machine, -1, sample->pid); 909 if (thread == NULL) 910 return -1; 911 912 cursor = get_tls_callchain_cursor(); 913 /* use caller function name from the callchain */ 914 ret = thread__resolve_callchain(thread, cursor, evsel, sample, 915 NULL, NULL, max_stack_depth); 916 thread__put(thread); 917 918 if (ret != 0) 919 return -1; 920 921 callchain_cursor_commit(cursor); 922 923 while (true) { 924 struct callchain_cursor_node *node; 925 926 node = callchain_cursor_current(cursor); 927 if (node == NULL) 928 break; 929 930 /* skip first few entries - for lock functions */ 931 if (++skip <= stack_skip) 932 goto next; 933 934 if (node->ms.sym && machine__is_lock_function(machine, node->ip)) 935 goto next; 936 937 hash ^= hash_long((unsigned long)node->ip, 64); 938 939 next: 940 callchain_cursor_advance(cursor); 941 } 942 return hash; 943 } 944 945 static u64 *get_callstack(struct perf_sample *sample, int max_stack) 946 { 947 u64 *callstack; 948 u64 i; 949 int c; 950 951 callstack = calloc(max_stack, sizeof(*callstack)); 952 if (callstack == NULL) 953 return NULL; 954 955 for (i = 0, c = 0; i < sample->callchain->nr && c < max_stack; i++) { 956 u64 ip = sample->callchain->ips[i]; 957 958 if (ip >= PERF_CONTEXT_MAX) 959 continue; 960 961 callstack[c++] = ip; 962 } 963 return callstack; 964 } 965 966 static int report_lock_contention_begin_event(struct evsel *evsel, 967 struct perf_sample *sample) 968 { 969 struct lock_stat *ls; 970 struct thread_stat *ts; 971 struct lock_seq_stat *seq; 972 u64 addr = evsel__intval(evsel, sample, "lock_addr"); 973 unsigned int flags = evsel__intval(evsel, sample, "flags"); 974 u64 key; 975 int i, ret; 976 static bool kmap_loaded; 977 struct machine *machine = &session->machines.host; 978 struct map *kmap; 979 struct symbol *sym; 980 981 ret = get_key_by_aggr_mode(&key, addr, evsel, sample); 982 if (ret < 0) 983 return ret; 984 985 if (!kmap_loaded) { 986 unsigned long *addrs; 987 988 /* make sure it loads the kernel map to find lock symbols */ 989 map__load(machine__kernel_map(machine)); 990 kmap_loaded = true; 991 992 /* convert (kernel) symbols to addresses */ 993 for (i = 0; i < filters.nr_syms; i++) { 994 sym = machine__find_kernel_symbol_by_name(machine, 995 filters.syms[i], 996 &kmap); 997 if (sym == NULL) { 998 pr_warning("ignore unknown symbol: %s\n", 999 filters.syms[i]); 1000 continue; 1001 } 1002 1003 addrs = realloc(filters.addrs, 1004 (filters.nr_addrs + 1) * sizeof(*addrs)); 1005 if (addrs == NULL) { 1006 pr_warning("memory allocation failure\n"); 1007 return -ENOMEM; 1008 } 1009 1010 addrs[filters.nr_addrs++] = map__unmap_ip(kmap, sym->start); 1011 filters.addrs = addrs; 1012 } 1013 } 1014 1015 ls = lock_stat_find(key); 1016 if (!ls) { 1017 char buf[128]; 1018 const char *name = ""; 1019 1020 switch (aggr_mode) { 1021 case LOCK_AGGR_ADDR: 1022 sym = machine__find_kernel_symbol(machine, key, &kmap); 1023 if (sym) 1024 name = sym->name; 1025 break; 1026 case LOCK_AGGR_CALLER: 1027 name = buf; 1028 if (lock_contention_caller(evsel, sample, buf, sizeof(buf)) < 0) 1029 name = "Unknown"; 1030 break; 1031 case LOCK_AGGR_CGROUP: 1032 case LOCK_AGGR_TASK: 1033 default: 1034 break; 1035 } 1036 1037 ls = lock_stat_findnew(key, name, flags); 1038 if (!ls) 1039 return -ENOMEM; 1040 } 1041 1042 if (filters.nr_types) { 1043 bool found = false; 1044 1045 for (i = 0; i < filters.nr_types; i++) { 1046 if (flags == filters.types[i]) { 1047 found = true; 1048 break; 1049 } 1050 } 1051 1052 if (!found) 1053 return 0; 1054 } 1055 1056 if (filters.nr_addrs) { 1057 bool found = false; 1058 1059 for (i = 0; i < filters.nr_addrs; i++) { 1060 if (addr == filters.addrs[i]) { 1061 found = true; 1062 break; 1063 } 1064 } 1065 1066 if (!found) 1067 return 0; 1068 } 1069 1070 if (needs_callstack()) { 1071 u64 *callstack = get_callstack(sample, max_stack_depth); 1072 if (callstack == NULL) 1073 return -ENOMEM; 1074 1075 if (!match_callstack_filter(machine, callstack, max_stack_depth)) { 1076 free(callstack); 1077 return 0; 1078 } 1079 1080 if (ls->callstack == NULL) 1081 ls->callstack = callstack; 1082 else 1083 free(callstack); 1084 } 1085 1086 ts = thread_stat_findnew(sample->tid); 1087 if (!ts) 1088 return -ENOMEM; 1089 1090 seq = get_seq(ts, addr); 1091 if (!seq) 1092 return -ENOMEM; 1093 1094 switch (seq->state) { 1095 case SEQ_STATE_UNINITIALIZED: 1096 case SEQ_STATE_ACQUIRED: 1097 break; 1098 case SEQ_STATE_CONTENDED: 1099 /* 1100 * It can have nested contention begin with mutex spinning, 1101 * then we would use the original contention begin event and 1102 * ignore the second one. 1103 */ 1104 goto end; 1105 case SEQ_STATE_ACQUIRING: 1106 case SEQ_STATE_READ_ACQUIRED: 1107 case SEQ_STATE_RELEASED: 1108 /* broken lock sequence */ 1109 if (!ls->broken) { 1110 ls->broken = 1; 1111 bad_hist[BROKEN_CONTENDED]++; 1112 } 1113 list_del_init(&seq->list); 1114 free(seq); 1115 goto end; 1116 default: 1117 BUG_ON("Unknown state of lock sequence found!\n"); 1118 break; 1119 } 1120 1121 if (seq->state != SEQ_STATE_CONTENDED) { 1122 seq->state = SEQ_STATE_CONTENDED; 1123 seq->prev_event_time = sample->time; 1124 ls->nr_contended++; 1125 } 1126 end: 1127 return 0; 1128 } 1129 1130 static int report_lock_contention_end_event(struct evsel *evsel, 1131 struct perf_sample *sample) 1132 { 1133 struct lock_stat *ls; 1134 struct thread_stat *ts; 1135 struct lock_seq_stat *seq; 1136 u64 contended_term; 1137 u64 addr = evsel__intval(evsel, sample, "lock_addr"); 1138 u64 key; 1139 int ret; 1140 1141 ret = get_key_by_aggr_mode(&key, addr, evsel, sample); 1142 if (ret < 0) 1143 return ret; 1144 1145 ls = lock_stat_find(key); 1146 if (!ls) 1147 return 0; 1148 1149 ts = thread_stat_find(sample->tid); 1150 if (!ts) 1151 return 0; 1152 1153 seq = get_seq(ts, addr); 1154 if (!seq) 1155 return -ENOMEM; 1156 1157 switch (seq->state) { 1158 case SEQ_STATE_UNINITIALIZED: 1159 goto end; 1160 case SEQ_STATE_CONTENDED: 1161 contended_term = sample->time - seq->prev_event_time; 1162 ls->wait_time_total += contended_term; 1163 if (contended_term < ls->wait_time_min) 1164 ls->wait_time_min = contended_term; 1165 if (ls->wait_time_max < contended_term) 1166 ls->wait_time_max = contended_term; 1167 break; 1168 case SEQ_STATE_ACQUIRING: 1169 case SEQ_STATE_ACQUIRED: 1170 case SEQ_STATE_READ_ACQUIRED: 1171 case SEQ_STATE_RELEASED: 1172 /* broken lock sequence */ 1173 if (!ls->broken) { 1174 ls->broken = 1; 1175 bad_hist[BROKEN_ACQUIRED]++; 1176 } 1177 list_del_init(&seq->list); 1178 free(seq); 1179 goto end; 1180 default: 1181 BUG_ON("Unknown state of lock sequence found!\n"); 1182 break; 1183 } 1184 1185 seq->state = SEQ_STATE_ACQUIRED; 1186 ls->nr_acquired++; 1187 ls->avg_wait_time = ls->wait_time_total/ls->nr_acquired; 1188 end: 1189 return 0; 1190 } 1191 1192 /* lock oriented handlers */ 1193 /* TODO: handlers for CPU oriented, thread oriented */ 1194 static struct trace_lock_handler report_lock_ops = { 1195 .acquire_event = report_lock_acquire_event, 1196 .acquired_event = report_lock_acquired_event, 1197 .contended_event = report_lock_contended_event, 1198 .release_event = report_lock_release_event, 1199 .contention_begin_event = report_lock_contention_begin_event, 1200 .contention_end_event = report_lock_contention_end_event, 1201 }; 1202 1203 static struct trace_lock_handler contention_lock_ops = { 1204 .contention_begin_event = report_lock_contention_begin_event, 1205 .contention_end_event = report_lock_contention_end_event, 1206 }; 1207 1208 1209 static struct trace_lock_handler *trace_handler; 1210 1211 static int evsel__process_lock_acquire(struct evsel *evsel, struct perf_sample *sample) 1212 { 1213 if (trace_handler->acquire_event) 1214 return trace_handler->acquire_event(evsel, sample); 1215 return 0; 1216 } 1217 1218 static int evsel__process_lock_acquired(struct evsel *evsel, struct perf_sample *sample) 1219 { 1220 if (trace_handler->acquired_event) 1221 return trace_handler->acquired_event(evsel, sample); 1222 return 0; 1223 } 1224 1225 static int evsel__process_lock_contended(struct evsel *evsel, struct perf_sample *sample) 1226 { 1227 if (trace_handler->contended_event) 1228 return trace_handler->contended_event(evsel, sample); 1229 return 0; 1230 } 1231 1232 static int evsel__process_lock_release(struct evsel *evsel, struct perf_sample *sample) 1233 { 1234 if (trace_handler->release_event) 1235 return trace_handler->release_event(evsel, sample); 1236 return 0; 1237 } 1238 1239 static int evsel__process_contention_begin(struct evsel *evsel, struct perf_sample *sample) 1240 { 1241 if (trace_handler->contention_begin_event) 1242 return trace_handler->contention_begin_event(evsel, sample); 1243 return 0; 1244 } 1245 1246 static int evsel__process_contention_end(struct evsel *evsel, struct perf_sample *sample) 1247 { 1248 if (trace_handler->contention_end_event) 1249 return trace_handler->contention_end_event(evsel, sample); 1250 return 0; 1251 } 1252 1253 static void print_bad_events(int bad, int total) 1254 { 1255 /* Output for debug, this have to be removed */ 1256 int i; 1257 int broken = 0; 1258 const char *name[4] = 1259 { "acquire", "acquired", "contended", "release" }; 1260 1261 for (i = 0; i < BROKEN_MAX; i++) 1262 broken += bad_hist[i]; 1263 1264 if (quiet || total == 0 || (broken == 0 && verbose <= 0)) 1265 return; 1266 1267 fprintf(lock_output, "\n=== output for debug ===\n\n"); 1268 fprintf(lock_output, "bad: %d, total: %d\n", bad, total); 1269 fprintf(lock_output, "bad rate: %.2f %%\n", (double)bad / (double)total * 100); 1270 fprintf(lock_output, "histogram of events caused bad sequence\n"); 1271 for (i = 0; i < BROKEN_MAX; i++) 1272 fprintf(lock_output, " %10s: %d\n", name[i], bad_hist[i]); 1273 } 1274 1275 /* TODO: various way to print, coloring, nano or milli sec */ 1276 static void print_result(void) 1277 { 1278 struct lock_stat *st; 1279 struct lock_key *key; 1280 char cut_name[20]; 1281 int bad, total, printed; 1282 1283 if (!quiet) { 1284 fprintf(lock_output, "%20s ", "Name"); 1285 list_for_each_entry(key, &lock_keys, list) 1286 fprintf(lock_output, "%*s ", key->len, key->header); 1287 fprintf(lock_output, "\n\n"); 1288 } 1289 1290 bad = total = printed = 0; 1291 while ((st = pop_from_result())) { 1292 total++; 1293 if (st->broken) 1294 bad++; 1295 if (!st->nr_acquired) 1296 continue; 1297 1298 bzero(cut_name, 20); 1299 1300 if (strlen(st->name) < 20) { 1301 /* output raw name */ 1302 const char *name = st->name; 1303 1304 if (show_thread_stats) { 1305 struct thread *t; 1306 1307 /* st->addr contains tid of thread */ 1308 t = perf_session__findnew(session, st->addr); 1309 name = thread__comm_str(t); 1310 } 1311 1312 fprintf(lock_output, "%20s ", name); 1313 } else { 1314 strncpy(cut_name, st->name, 16); 1315 cut_name[16] = '.'; 1316 cut_name[17] = '.'; 1317 cut_name[18] = '.'; 1318 cut_name[19] = '\0'; 1319 /* cut off name for saving output style */ 1320 fprintf(lock_output, "%20s ", cut_name); 1321 } 1322 1323 list_for_each_entry(key, &lock_keys, list) { 1324 key->print(key, st); 1325 fprintf(lock_output, " "); 1326 } 1327 fprintf(lock_output, "\n"); 1328 1329 if (++printed >= print_nr_entries) 1330 break; 1331 } 1332 1333 print_bad_events(bad, total); 1334 } 1335 1336 static bool info_threads, info_map; 1337 1338 static void dump_threads(void) 1339 { 1340 struct thread_stat *st; 1341 struct rb_node *node; 1342 struct thread *t; 1343 1344 fprintf(lock_output, "%10s: comm\n", "Thread ID"); 1345 1346 node = rb_first(&thread_stats); 1347 while (node) { 1348 st = container_of(node, struct thread_stat, rb); 1349 t = perf_session__findnew(session, st->tid); 1350 fprintf(lock_output, "%10d: %s\n", st->tid, thread__comm_str(t)); 1351 node = rb_next(node); 1352 thread__put(t); 1353 } 1354 } 1355 1356 static int compare_maps(struct lock_stat *a, struct lock_stat *b) 1357 { 1358 int ret; 1359 1360 if (a->name && b->name) 1361 ret = strcmp(a->name, b->name); 1362 else 1363 ret = !!a->name - !!b->name; 1364 1365 if (!ret) 1366 return a->addr < b->addr; 1367 else 1368 return ret < 0; 1369 } 1370 1371 static void dump_map(void) 1372 { 1373 unsigned int i; 1374 struct lock_stat *st; 1375 1376 fprintf(lock_output, "Address of instance: name of class\n"); 1377 for (i = 0; i < LOCKHASH_SIZE; i++) { 1378 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) { 1379 insert_to_result(st, compare_maps); 1380 } 1381 } 1382 1383 while ((st = pop_from_result())) 1384 fprintf(lock_output, " %#llx: %s\n", (unsigned long long)st->addr, st->name); 1385 } 1386 1387 static void dump_info(void) 1388 { 1389 if (info_threads) 1390 dump_threads(); 1391 1392 if (info_map) { 1393 if (info_threads) 1394 fputc('\n', lock_output); 1395 dump_map(); 1396 } 1397 } 1398 1399 static const struct evsel_str_handler lock_tracepoints[] = { 1400 { "lock:lock_acquire", evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */ 1401 { "lock:lock_acquired", evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */ 1402 { "lock:lock_contended", evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */ 1403 { "lock:lock_release", evsel__process_lock_release, }, /* CONFIG_LOCKDEP */ 1404 }; 1405 1406 static const struct evsel_str_handler contention_tracepoints[] = { 1407 { "lock:contention_begin", evsel__process_contention_begin, }, 1408 { "lock:contention_end", evsel__process_contention_end, }, 1409 }; 1410 1411 static int process_event_update(const struct perf_tool *tool, 1412 union perf_event *event, 1413 struct evlist **pevlist) 1414 { 1415 int ret; 1416 1417 ret = perf_event__process_event_update(tool, event, pevlist); 1418 if (ret < 0) 1419 return ret; 1420 1421 /* this can return -EEXIST since we call it for each evsel */ 1422 perf_session__set_tracepoints_handlers(session, lock_tracepoints); 1423 perf_session__set_tracepoints_handlers(session, contention_tracepoints); 1424 return 0; 1425 } 1426 1427 typedef int (*tracepoint_handler)(struct evsel *evsel, 1428 struct perf_sample *sample); 1429 1430 static int process_sample_event(const struct perf_tool *tool __maybe_unused, 1431 union perf_event *event, 1432 struct perf_sample *sample, 1433 struct evsel *evsel, 1434 struct machine *machine) 1435 { 1436 int err = 0; 1437 struct thread *thread = machine__findnew_thread(machine, sample->pid, 1438 sample->tid); 1439 1440 if (thread == NULL) { 1441 pr_debug("problem processing %d event, skipping it.\n", 1442 event->header.type); 1443 return -1; 1444 } 1445 1446 if (evsel->handler != NULL) { 1447 tracepoint_handler f = evsel->handler; 1448 err = f(evsel, sample); 1449 } 1450 1451 thread__put(thread); 1452 1453 return err; 1454 } 1455 1456 static void combine_result(void) 1457 { 1458 unsigned int i; 1459 struct lock_stat *st; 1460 1461 if (!combine_locks) 1462 return; 1463 1464 for (i = 0; i < LOCKHASH_SIZE; i++) { 1465 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) { 1466 combine_lock_stats(st); 1467 } 1468 } 1469 } 1470 1471 static void sort_result(void) 1472 { 1473 unsigned int i; 1474 struct lock_stat *st; 1475 1476 for (i = 0; i < LOCKHASH_SIZE; i++) { 1477 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) { 1478 insert_to_result(st, compare); 1479 } 1480 } 1481 } 1482 1483 static const struct { 1484 unsigned int flags; 1485 /* 1486 * Name of the lock flags (access), with delimeter ':'. 1487 * For example, rwsem:R of rwsem:W. 1488 */ 1489 const char *flags_name; 1490 /* Name of the lock (type), for example, rwlock or rwsem. */ 1491 const char *lock_name; 1492 } lock_type_table[] = { 1493 { 0, "semaphore", "semaphore" }, 1494 { LCB_F_SPIN, "spinlock", "spinlock" }, 1495 { LCB_F_SPIN | LCB_F_READ, "rwlock:R", "rwlock" }, 1496 { LCB_F_SPIN | LCB_F_WRITE, "rwlock:W", "rwlock" }, 1497 { LCB_F_READ, "rwsem:R", "rwsem" }, 1498 { LCB_F_WRITE, "rwsem:W", "rwsem" }, 1499 { LCB_F_RT, "rt-mutex", "rt-mutex" }, 1500 { LCB_F_RT | LCB_F_READ, "rwlock-rt:R", "rwlock-rt" }, 1501 { LCB_F_RT | LCB_F_WRITE, "rwlock-rt:W", "rwlock-rt" }, 1502 { LCB_F_PERCPU | LCB_F_READ, "pcpu-sem:R", "percpu-rwsem" }, 1503 { LCB_F_PERCPU | LCB_F_WRITE, "pcpu-sem:W", "percpu-rwsem" }, 1504 { LCB_F_MUTEX, "mutex", "mutex" }, 1505 { LCB_F_MUTEX | LCB_F_SPIN, "mutex", "mutex" }, 1506 /* alias for optimistic spinning only */ 1507 { LCB_F_MUTEX | LCB_F_SPIN, "mutex:spin", "mutex-spin" }, 1508 }; 1509 1510 static const char *get_type_flags_name(unsigned int flags) 1511 { 1512 flags &= LCB_F_TYPE_MASK; 1513 1514 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) { 1515 if (lock_type_table[i].flags == flags) 1516 return lock_type_table[i].flags_name; 1517 } 1518 return "unknown"; 1519 } 1520 1521 static const char *get_type_lock_name(unsigned int flags) 1522 { 1523 flags &= LCB_F_TYPE_MASK; 1524 1525 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) { 1526 if (lock_type_table[i].flags == flags) 1527 return lock_type_table[i].lock_name; 1528 } 1529 return "unknown"; 1530 } 1531 1532 static void lock_filter_finish(void) 1533 { 1534 zfree(&filters.types); 1535 filters.nr_types = 0; 1536 1537 zfree(&filters.addrs); 1538 filters.nr_addrs = 0; 1539 1540 for (int i = 0; i < filters.nr_syms; i++) 1541 free(filters.syms[i]); 1542 1543 zfree(&filters.syms); 1544 filters.nr_syms = 0; 1545 1546 zfree(&filters.cgrps); 1547 filters.nr_cgrps = 0; 1548 1549 for (int i = 0; i < filters.nr_slabs; i++) 1550 free(filters.slabs[i]); 1551 1552 zfree(&filters.slabs); 1553 filters.nr_slabs = 0; 1554 } 1555 1556 static void sort_contention_result(void) 1557 { 1558 sort_result(); 1559 } 1560 1561 static void print_header_stdio(void) 1562 { 1563 struct lock_key *key; 1564 1565 list_for_each_entry(key, &lock_keys, list) 1566 fprintf(lock_output, "%*s ", key->len, key->header); 1567 1568 switch (aggr_mode) { 1569 case LOCK_AGGR_TASK: 1570 fprintf(lock_output, " %10s %s\n\n", "pid", 1571 show_lock_owner ? "owner" : "comm"); 1572 break; 1573 case LOCK_AGGR_CALLER: 1574 fprintf(lock_output, " %10s %s\n\n", "type", "caller"); 1575 break; 1576 case LOCK_AGGR_ADDR: 1577 fprintf(lock_output, " %16s %s\n\n", "address", "symbol"); 1578 break; 1579 case LOCK_AGGR_CGROUP: 1580 fprintf(lock_output, " %s\n\n", "cgroup"); 1581 break; 1582 default: 1583 break; 1584 } 1585 } 1586 1587 static void print_header_csv(const char *sep) 1588 { 1589 struct lock_key *key; 1590 1591 fprintf(lock_output, "# output: "); 1592 list_for_each_entry(key, &lock_keys, list) 1593 fprintf(lock_output, "%s%s ", key->header, sep); 1594 1595 switch (aggr_mode) { 1596 case LOCK_AGGR_TASK: 1597 fprintf(lock_output, "%s%s %s\n", "pid", sep, 1598 show_lock_owner ? "owner" : "comm"); 1599 break; 1600 case LOCK_AGGR_CALLER: 1601 fprintf(lock_output, "%s%s %s", "type", sep, "caller"); 1602 if (verbose > 0) 1603 fprintf(lock_output, "%s %s", sep, "stacktrace"); 1604 fprintf(lock_output, "\n"); 1605 break; 1606 case LOCK_AGGR_ADDR: 1607 fprintf(lock_output, "%s%s %s%s %s\n", "address", sep, "symbol", sep, "type"); 1608 break; 1609 case LOCK_AGGR_CGROUP: 1610 fprintf(lock_output, "%s\n", "cgroup"); 1611 break; 1612 default: 1613 break; 1614 } 1615 } 1616 1617 static void print_header(void) 1618 { 1619 if (!quiet) { 1620 if (symbol_conf.field_sep) 1621 print_header_csv(symbol_conf.field_sep); 1622 else 1623 print_header_stdio(); 1624 } 1625 } 1626 1627 static void print_lock_stat_stdio(struct lock_contention *con, struct lock_stat *st) 1628 { 1629 struct lock_key *key; 1630 struct thread *t; 1631 int pid; 1632 1633 list_for_each_entry(key, &lock_keys, list) { 1634 key->print(key, st); 1635 fprintf(lock_output, " "); 1636 } 1637 1638 switch (aggr_mode) { 1639 case LOCK_AGGR_CALLER: 1640 fprintf(lock_output, " %10s %s\n", get_type_flags_name(st->flags), st->name); 1641 break; 1642 case LOCK_AGGR_TASK: 1643 pid = st->addr; 1644 t = perf_session__findnew(session, pid); 1645 fprintf(lock_output, " %10d %s\n", 1646 pid, pid == -1 ? "Unknown" : thread__comm_str(t)); 1647 break; 1648 case LOCK_AGGR_ADDR: 1649 fprintf(lock_output, " %016llx %s (%s)\n", (unsigned long long)st->addr, 1650 st->name, get_type_lock_name(st->flags)); 1651 break; 1652 case LOCK_AGGR_CGROUP: 1653 fprintf(lock_output, " %s\n", st->name); 1654 break; 1655 default: 1656 break; 1657 } 1658 1659 if (aggr_mode == LOCK_AGGR_CALLER && verbose > 0) { 1660 struct map *kmap; 1661 struct symbol *sym; 1662 char buf[128]; 1663 u64 ip; 1664 1665 for (int i = 0; i < max_stack_depth; i++) { 1666 if (!st->callstack || !st->callstack[i]) 1667 break; 1668 1669 ip = st->callstack[i]; 1670 sym = machine__find_kernel_symbol(con->machine, ip, &kmap); 1671 get_symbol_name_offset(kmap, sym, ip, buf, sizeof(buf)); 1672 fprintf(lock_output, "\t\t\t%#lx %s\n", (unsigned long)ip, buf); 1673 } 1674 } 1675 } 1676 1677 static void print_lock_stat_csv(struct lock_contention *con, struct lock_stat *st, 1678 const char *sep) 1679 { 1680 struct lock_key *key; 1681 struct thread *t; 1682 int pid; 1683 1684 list_for_each_entry(key, &lock_keys, list) { 1685 key->print(key, st); 1686 fprintf(lock_output, "%s ", sep); 1687 } 1688 1689 switch (aggr_mode) { 1690 case LOCK_AGGR_CALLER: 1691 fprintf(lock_output, "%s%s %s", get_type_flags_name(st->flags), sep, st->name); 1692 if (verbose <= 0) 1693 fprintf(lock_output, "\n"); 1694 break; 1695 case LOCK_AGGR_TASK: 1696 pid = st->addr; 1697 t = perf_session__findnew(session, pid); 1698 fprintf(lock_output, "%d%s %s\n", pid, sep, 1699 pid == -1 ? "Unknown" : thread__comm_str(t)); 1700 break; 1701 case LOCK_AGGR_ADDR: 1702 fprintf(lock_output, "%llx%s %s%s %s\n", (unsigned long long)st->addr, sep, 1703 st->name, sep, get_type_lock_name(st->flags)); 1704 break; 1705 case LOCK_AGGR_CGROUP: 1706 fprintf(lock_output, "%s\n",st->name); 1707 break; 1708 default: 1709 break; 1710 } 1711 1712 if (aggr_mode == LOCK_AGGR_CALLER && verbose > 0) { 1713 struct map *kmap; 1714 struct symbol *sym; 1715 char buf[128]; 1716 u64 ip; 1717 1718 for (int i = 0; i < max_stack_depth; i++) { 1719 if (!st->callstack || !st->callstack[i]) 1720 break; 1721 1722 ip = st->callstack[i]; 1723 sym = machine__find_kernel_symbol(con->machine, ip, &kmap); 1724 get_symbol_name_offset(kmap, sym, ip, buf, sizeof(buf)); 1725 fprintf(lock_output, "%s %#lx %s", i ? ":" : sep, (unsigned long) ip, buf); 1726 } 1727 fprintf(lock_output, "\n"); 1728 } 1729 } 1730 1731 static void print_lock_stat(struct lock_contention *con, struct lock_stat *st) 1732 { 1733 if (symbol_conf.field_sep) 1734 print_lock_stat_csv(con, st, symbol_conf.field_sep); 1735 else 1736 print_lock_stat_stdio(con, st); 1737 } 1738 1739 static void print_footer_stdio(int total, int bad, struct lock_contention_fails *fails) 1740 { 1741 /* Output for debug, this have to be removed */ 1742 int broken = fails->task + fails->stack + fails->time + fails->data; 1743 1744 if (!use_bpf) 1745 print_bad_events(bad, total); 1746 1747 if (quiet || total == 0 || (broken == 0 && verbose <= 0)) 1748 return; 1749 1750 total += broken; 1751 fprintf(lock_output, "\n=== output for debug ===\n\n"); 1752 fprintf(lock_output, "bad: %d, total: %d\n", broken, total); 1753 fprintf(lock_output, "bad rate: %.2f %%\n", 100.0 * broken / total); 1754 1755 fprintf(lock_output, "histogram of failure reasons\n"); 1756 fprintf(lock_output, " %10s: %d\n", "task", fails->task); 1757 fprintf(lock_output, " %10s: %d\n", "stack", fails->stack); 1758 fprintf(lock_output, " %10s: %d\n", "time", fails->time); 1759 fprintf(lock_output, " %10s: %d\n", "data", fails->data); 1760 } 1761 1762 static void print_footer_csv(int total, int bad, struct lock_contention_fails *fails, 1763 const char *sep) 1764 { 1765 /* Output for debug, this have to be removed */ 1766 if (use_bpf) 1767 bad = fails->task + fails->stack + fails->time + fails->data; 1768 1769 if (quiet || total == 0 || (bad == 0 && verbose <= 0)) 1770 return; 1771 1772 total += bad; 1773 fprintf(lock_output, "# debug: total=%d%s bad=%d", total, sep, bad); 1774 1775 if (use_bpf) { 1776 fprintf(lock_output, "%s bad_%s=%d", sep, "task", fails->task); 1777 fprintf(lock_output, "%s bad_%s=%d", sep, "stack", fails->stack); 1778 fprintf(lock_output, "%s bad_%s=%d", sep, "time", fails->time); 1779 fprintf(lock_output, "%s bad_%s=%d", sep, "data", fails->data); 1780 } else { 1781 int i; 1782 const char *name[4] = { "acquire", "acquired", "contended", "release" }; 1783 1784 for (i = 0; i < BROKEN_MAX; i++) 1785 fprintf(lock_output, "%s bad_%s=%d", sep, name[i], bad_hist[i]); 1786 } 1787 fprintf(lock_output, "\n"); 1788 } 1789 1790 static void print_footer(int total, int bad, struct lock_contention_fails *fails) 1791 { 1792 if (symbol_conf.field_sep) 1793 print_footer_csv(total, bad, fails, symbol_conf.field_sep); 1794 else 1795 print_footer_stdio(total, bad, fails); 1796 } 1797 1798 static void print_contention_result(struct lock_contention *con) 1799 { 1800 struct lock_stat *st; 1801 int bad, total, printed; 1802 1803 if (!quiet) 1804 print_header(); 1805 1806 bad = total = printed = 0; 1807 1808 while ((st = pop_from_result())) { 1809 total += use_bpf ? st->nr_contended : 1; 1810 if (st->broken) 1811 bad++; 1812 1813 if (!st->wait_time_total) 1814 continue; 1815 1816 print_lock_stat(con, st); 1817 1818 if (++printed >= print_nr_entries) 1819 break; 1820 } 1821 1822 if (con->owner && con->save_callstack && verbose > 0) { 1823 struct rb_root root = RB_ROOT; 1824 1825 if (symbol_conf.field_sep) 1826 fprintf(lock_output, "# owner stack trace:\n"); 1827 else 1828 fprintf(lock_output, "\n=== owner stack trace ===\n\n"); 1829 while ((st = pop_owner_stack_trace(con))) 1830 insert_to(&root, st, compare); 1831 1832 while ((st = pop_from(&root))) { 1833 print_lock_stat(con, st); 1834 free(st); 1835 } 1836 } 1837 1838 if (print_nr_entries) { 1839 /* update the total/bad stats */ 1840 while ((st = pop_from_result())) { 1841 total += use_bpf ? st->nr_contended : 1; 1842 if (st->broken) 1843 bad++; 1844 } 1845 } 1846 /* some entries are collected but hidden by the callstack filter */ 1847 total += con->nr_filtered; 1848 1849 print_footer(total, bad, &con->fails); 1850 } 1851 1852 static bool force; 1853 1854 static int __cmd_report(bool display_info) 1855 { 1856 int err = -EINVAL; 1857 struct perf_tool eops; 1858 struct perf_data data = { 1859 .path = input_name, 1860 .mode = PERF_DATA_MODE_READ, 1861 .force = force, 1862 }; 1863 1864 perf_tool__init(&eops, /*ordered_events=*/true); 1865 eops.attr = perf_event__process_attr; 1866 eops.event_update = process_event_update; 1867 eops.sample = process_sample_event; 1868 eops.comm = perf_event__process_comm; 1869 eops.mmap = perf_event__process_mmap; 1870 eops.namespaces = perf_event__process_namespaces; 1871 eops.tracing_data = perf_event__process_tracing_data; 1872 session = perf_session__new(&data, &eops); 1873 if (IS_ERR(session)) { 1874 pr_err("Initializing perf session failed\n"); 1875 return PTR_ERR(session); 1876 } 1877 1878 symbol_conf.allow_aliases = true; 1879 symbol__init(perf_session__env(session)); 1880 1881 if (!data.is_pipe) { 1882 if (!perf_session__has_traces(session, "lock record")) 1883 goto out_delete; 1884 1885 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) { 1886 pr_err("Initializing perf session tracepoint handlers failed\n"); 1887 goto out_delete; 1888 } 1889 1890 if (perf_session__set_tracepoints_handlers(session, contention_tracepoints)) { 1891 pr_err("Initializing perf session tracepoint handlers failed\n"); 1892 goto out_delete; 1893 } 1894 } 1895 1896 if (setup_output_field(false, output_fields)) 1897 goto out_delete; 1898 1899 if (select_key(false)) 1900 goto out_delete; 1901 1902 if (show_thread_stats) 1903 aggr_mode = LOCK_AGGR_TASK; 1904 1905 err = perf_session__process_events(session); 1906 if (err) 1907 goto out_delete; 1908 1909 setup_pager(); 1910 if (display_info) /* used for info subcommand */ 1911 dump_info(); 1912 else { 1913 combine_result(); 1914 sort_result(); 1915 print_result(); 1916 } 1917 1918 out_delete: 1919 perf_session__delete(session); 1920 return err; 1921 } 1922 1923 static void sighandler(int sig __maybe_unused) 1924 { 1925 } 1926 1927 static int check_lock_contention_options(const struct option *options, 1928 const char * const *usage) 1929 1930 { 1931 if (show_thread_stats && show_lock_addrs) { 1932 pr_err("Cannot use thread and addr mode together\n"); 1933 parse_options_usage(usage, options, "threads", 0); 1934 parse_options_usage(NULL, options, "lock-addr", 0); 1935 return -1; 1936 } 1937 1938 if (show_lock_owner && !use_bpf) { 1939 pr_err("Lock owners are available only with BPF\n"); 1940 parse_options_usage(usage, options, "lock-owner", 0); 1941 parse_options_usage(NULL, options, "use-bpf", 0); 1942 return -1; 1943 } 1944 1945 if (show_lock_owner && show_lock_addrs) { 1946 pr_err("Cannot use owner and addr mode together\n"); 1947 parse_options_usage(usage, options, "lock-owner", 0); 1948 parse_options_usage(NULL, options, "lock-addr", 0); 1949 return -1; 1950 } 1951 1952 if (show_lock_cgroups && !use_bpf) { 1953 pr_err("Cgroups are available only with BPF\n"); 1954 parse_options_usage(usage, options, "lock-cgroup", 0); 1955 parse_options_usage(NULL, options, "use-bpf", 0); 1956 return -1; 1957 } 1958 1959 if (show_lock_cgroups && show_lock_addrs) { 1960 pr_err("Cannot use cgroup and addr mode together\n"); 1961 parse_options_usage(usage, options, "lock-cgroup", 0); 1962 parse_options_usage(NULL, options, "lock-addr", 0); 1963 return -1; 1964 } 1965 1966 if (show_lock_cgroups && show_thread_stats) { 1967 pr_err("Cannot use cgroup and thread mode together\n"); 1968 parse_options_usage(usage, options, "lock-cgroup", 0); 1969 parse_options_usage(NULL, options, "threads", 0); 1970 return -1; 1971 } 1972 1973 if (symbol_conf.field_sep) { 1974 if (strstr(symbol_conf.field_sep, ":") || /* part of type flags */ 1975 strstr(symbol_conf.field_sep, "+") || /* part of caller offset */ 1976 strstr(symbol_conf.field_sep, ".")) { /* can be in a symbol name */ 1977 pr_err("Cannot use the separator that is already used\n"); 1978 parse_options_usage(usage, options, "x", 1); 1979 return -1; 1980 } 1981 } 1982 1983 if (show_lock_owner && !show_thread_stats) { 1984 pr_warning("Now -o try to show owner's callstack instead of pid and comm.\n"); 1985 pr_warning("Please use -t option too to keep the old behavior.\n"); 1986 } 1987 1988 return 0; 1989 } 1990 1991 static int __cmd_contention(int argc, const char **argv) 1992 { 1993 int err = -EINVAL; 1994 struct perf_tool eops; 1995 struct perf_data data = { 1996 .path = input_name, 1997 .mode = PERF_DATA_MODE_READ, 1998 .force = force, 1999 }; 2000 struct lock_contention con = { 2001 .target = &target, 2002 .map_nr_entries = bpf_map_entries, 2003 .max_stack = max_stack_depth, 2004 .stack_skip = stack_skip, 2005 .filters = &filters, 2006 .delays = delays, 2007 .nr_delays = nr_delays, 2008 .save_callstack = needs_callstack(), 2009 .owner = show_lock_owner, 2010 .cgroups = RB_ROOT, 2011 }; 2012 2013 lockhash_table = calloc(LOCKHASH_SIZE, sizeof(*lockhash_table)); 2014 if (!lockhash_table) 2015 return -ENOMEM; 2016 2017 con.result = &lockhash_table[0]; 2018 2019 perf_tool__init(&eops, /*ordered_events=*/true); 2020 eops.attr = perf_event__process_attr; 2021 eops.event_update = process_event_update; 2022 eops.sample = process_sample_event; 2023 eops.comm = perf_event__process_comm; 2024 eops.mmap = perf_event__process_mmap; 2025 eops.tracing_data = perf_event__process_tracing_data; 2026 2027 session = perf_session__new(use_bpf ? NULL : &data, &eops); 2028 if (IS_ERR(session)) { 2029 pr_err("Initializing perf session failed\n"); 2030 err = PTR_ERR(session); 2031 session = NULL; 2032 goto out_delete; 2033 } 2034 2035 con.machine = &session->machines.host; 2036 2037 con.aggr_mode = aggr_mode = show_thread_stats ? LOCK_AGGR_TASK : 2038 show_lock_addrs ? LOCK_AGGR_ADDR : 2039 show_lock_cgroups ? LOCK_AGGR_CGROUP : LOCK_AGGR_CALLER; 2040 2041 if (con.aggr_mode == LOCK_AGGR_CALLER) 2042 con.save_callstack = true; 2043 2044 symbol_conf.allow_aliases = true; 2045 symbol__init(perf_session__env(session)); 2046 2047 if (use_bpf) { 2048 err = target__validate(&target); 2049 if (err) { 2050 char errbuf[512]; 2051 2052 target__strerror(&target, err, errbuf, 512); 2053 pr_err("%s\n", errbuf); 2054 goto out_delete; 2055 } 2056 2057 signal(SIGINT, sighandler); 2058 signal(SIGCHLD, sighandler); 2059 signal(SIGTERM, sighandler); 2060 2061 con.evlist = evlist__new(); 2062 if (con.evlist == NULL) { 2063 err = -ENOMEM; 2064 goto out_delete; 2065 } 2066 2067 err = evlist__create_maps(con.evlist, &target); 2068 if (err < 0) 2069 goto out_delete; 2070 2071 if (argc) { 2072 err = evlist__prepare_workload(con.evlist, &target, 2073 argv, false, NULL); 2074 if (err < 0) 2075 goto out_delete; 2076 } 2077 2078 err = lock_contention_prepare(&con); 2079 if (err < 0) { 2080 pr_err("lock contention BPF setup failed\n"); 2081 goto out_delete; 2082 } 2083 } else if (!data.is_pipe) { 2084 if (!perf_session__has_traces(session, "lock record")) 2085 goto out_delete; 2086 2087 if (!evlist__find_evsel_by_str(session->evlist, 2088 "lock:contention_begin")) { 2089 pr_err("lock contention evsel not found\n"); 2090 goto out_delete; 2091 } 2092 2093 if (perf_session__set_tracepoints_handlers(session, 2094 contention_tracepoints)) { 2095 pr_err("Initializing perf session tracepoint handlers failed\n"); 2096 goto out_delete; 2097 } 2098 } 2099 2100 err = setup_output_field(true, output_fields); 2101 if (err) { 2102 pr_err("Failed to setup output field\n"); 2103 goto out_delete; 2104 } 2105 2106 err = select_key(true); 2107 if (err) 2108 goto out_delete; 2109 2110 if (symbol_conf.field_sep) { 2111 int i; 2112 struct lock_key *keys = contention_keys; 2113 2114 /* do not align output in CSV format */ 2115 for (i = 0; keys[i].name; i++) 2116 keys[i].len = 0; 2117 } 2118 2119 if (use_bpf) { 2120 lock_contention_start(); 2121 if (argc) 2122 evlist__start_workload(con.evlist); 2123 2124 /* wait for signal */ 2125 pause(); 2126 2127 lock_contention_stop(); 2128 lock_contention_read(&con); 2129 } else { 2130 err = perf_session__process_events(session); 2131 if (err) 2132 goto out_delete; 2133 } 2134 2135 setup_pager(); 2136 2137 sort_contention_result(); 2138 print_contention_result(&con); 2139 2140 out_delete: 2141 lock_filter_finish(); 2142 evlist__delete(con.evlist); 2143 lock_contention_finish(&con); 2144 perf_session__delete(session); 2145 zfree(&lockhash_table); 2146 return err; 2147 } 2148 2149 2150 static int __cmd_record(int argc, const char **argv) 2151 { 2152 const char *record_args[] = { 2153 "record", "-R", "-m", "1024", "-c", "1", "--synth", "task", 2154 }; 2155 const char *callgraph_args[] = { 2156 "--call-graph", "fp," __stringify(CONTENTION_STACK_DEPTH), 2157 }; 2158 unsigned int rec_argc, i, j, ret; 2159 unsigned int nr_tracepoints; 2160 unsigned int nr_callgraph_args = 0; 2161 const char **rec_argv; 2162 bool has_lock_stat = true; 2163 2164 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) { 2165 if (!is_valid_tracepoint(lock_tracepoints[i].name)) { 2166 pr_debug("tracepoint %s is not enabled. " 2167 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n", 2168 lock_tracepoints[i].name); 2169 has_lock_stat = false; 2170 break; 2171 } 2172 } 2173 2174 if (has_lock_stat) 2175 goto setup_args; 2176 2177 for (i = 0; i < ARRAY_SIZE(contention_tracepoints); i++) { 2178 if (!is_valid_tracepoint(contention_tracepoints[i].name)) { 2179 pr_err("tracepoint %s is not enabled.\n", 2180 contention_tracepoints[i].name); 2181 return 1; 2182 } 2183 } 2184 2185 nr_callgraph_args = ARRAY_SIZE(callgraph_args); 2186 2187 setup_args: 2188 rec_argc = ARRAY_SIZE(record_args) + nr_callgraph_args + argc - 1; 2189 2190 if (has_lock_stat) 2191 nr_tracepoints = ARRAY_SIZE(lock_tracepoints); 2192 else 2193 nr_tracepoints = ARRAY_SIZE(contention_tracepoints); 2194 2195 /* factor of 2 is for -e in front of each tracepoint */ 2196 rec_argc += 2 * nr_tracepoints; 2197 2198 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 2199 if (!rec_argv) 2200 return -ENOMEM; 2201 2202 for (i = 0; i < ARRAY_SIZE(record_args); i++) 2203 rec_argv[i] = record_args[i]; 2204 2205 for (j = 0; j < nr_tracepoints; j++) { 2206 rec_argv[i++] = "-e"; 2207 rec_argv[i++] = has_lock_stat 2208 ? lock_tracepoints[j].name 2209 : contention_tracepoints[j].name; 2210 } 2211 2212 for (j = 0; j < nr_callgraph_args; j++, i++) 2213 rec_argv[i] = callgraph_args[j]; 2214 2215 for (j = 1; j < (unsigned int)argc; j++, i++) 2216 rec_argv[i] = argv[j]; 2217 2218 BUG_ON(i != rec_argc); 2219 2220 ret = cmd_record(i, rec_argv); 2221 free(rec_argv); 2222 return ret; 2223 } 2224 2225 static int parse_map_entry(const struct option *opt, const char *str, 2226 int unset __maybe_unused) 2227 { 2228 unsigned long *len = (unsigned long *)opt->value; 2229 unsigned long val; 2230 char *endptr; 2231 2232 errno = 0; 2233 val = strtoul(str, &endptr, 0); 2234 if (*endptr != '\0' || errno != 0) { 2235 pr_err("invalid BPF map length: %s\n", str); 2236 return -1; 2237 } 2238 2239 *len = val; 2240 return 0; 2241 } 2242 2243 static int parse_max_stack(const struct option *opt, const char *str, 2244 int unset __maybe_unused) 2245 { 2246 unsigned long *len = (unsigned long *)opt->value; 2247 long val; 2248 char *endptr; 2249 2250 errno = 0; 2251 val = strtol(str, &endptr, 0); 2252 if (*endptr != '\0' || errno != 0) { 2253 pr_err("invalid max stack depth: %s\n", str); 2254 return -1; 2255 } 2256 2257 if (val < 0 || val > sysctl__max_stack()) { 2258 pr_err("invalid max stack depth: %ld\n", val); 2259 return -1; 2260 } 2261 2262 *len = val; 2263 return 0; 2264 } 2265 2266 static bool add_lock_type(unsigned int flags) 2267 { 2268 unsigned int *tmp; 2269 2270 tmp = realloc(filters.types, (filters.nr_types + 1) * sizeof(*filters.types)); 2271 if (tmp == NULL) 2272 return false; 2273 2274 tmp[filters.nr_types++] = flags; 2275 filters.types = tmp; 2276 return true; 2277 } 2278 2279 static int parse_lock_type(const struct option *opt __maybe_unused, const char *str, 2280 int unset __maybe_unused) 2281 { 2282 char *s, *tmp, *tok; 2283 2284 s = strdup(str); 2285 if (s == NULL) 2286 return -1; 2287 2288 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) { 2289 bool found = false; 2290 2291 /* `tok` is a flags name if it contains ':'. */ 2292 if (strchr(tok, ':')) { 2293 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) { 2294 if (!strcmp(lock_type_table[i].flags_name, tok) && 2295 add_lock_type(lock_type_table[i].flags)) { 2296 found = true; 2297 break; 2298 } 2299 } 2300 2301 if (!found) { 2302 pr_err("Unknown lock flags name: %s\n", tok); 2303 free(s); 2304 return -1; 2305 } 2306 2307 continue; 2308 } 2309 2310 /* 2311 * Otherwise `tok` is a lock name. 2312 * Single lock name could contain multiple flags. 2313 * Replace alias `pcpu-sem` with actual name `percpu-rwsem. 2314 */ 2315 if (!strcmp(tok, "pcpu-sem")) 2316 tok = (char *)"percpu-rwsem"; 2317 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) { 2318 if (!strcmp(lock_type_table[i].lock_name, tok)) { 2319 if (add_lock_type(lock_type_table[i].flags)) { 2320 found = true; 2321 } else { 2322 free(s); 2323 return -1; 2324 } 2325 } 2326 } 2327 2328 if (!found) { 2329 pr_err("Unknown lock name: %s\n", tok); 2330 free(s); 2331 return -1; 2332 } 2333 2334 } 2335 2336 free(s); 2337 return 0; 2338 } 2339 2340 static bool add_lock_addr(unsigned long addr) 2341 { 2342 unsigned long *tmp; 2343 2344 tmp = realloc(filters.addrs, (filters.nr_addrs + 1) * sizeof(*filters.addrs)); 2345 if (tmp == NULL) { 2346 pr_err("Memory allocation failure\n"); 2347 return false; 2348 } 2349 2350 tmp[filters.nr_addrs++] = addr; 2351 filters.addrs = tmp; 2352 return true; 2353 } 2354 2355 static bool add_lock_sym(char *name) 2356 { 2357 char **tmp; 2358 char *sym = strdup(name); 2359 2360 if (sym == NULL) { 2361 pr_err("Memory allocation failure\n"); 2362 return false; 2363 } 2364 2365 tmp = realloc(filters.syms, (filters.nr_syms + 1) * sizeof(*filters.syms)); 2366 if (tmp == NULL) { 2367 pr_err("Memory allocation failure\n"); 2368 free(sym); 2369 return false; 2370 } 2371 2372 tmp[filters.nr_syms++] = sym; 2373 filters.syms = tmp; 2374 return true; 2375 } 2376 2377 static bool add_lock_slab(char *name) 2378 { 2379 char **tmp; 2380 char *sym = strdup(name); 2381 2382 if (sym == NULL) { 2383 pr_err("Memory allocation failure\n"); 2384 return false; 2385 } 2386 2387 tmp = realloc(filters.slabs, (filters.nr_slabs + 1) * sizeof(*filters.slabs)); 2388 if (tmp == NULL) { 2389 pr_err("Memory allocation failure\n"); 2390 return false; 2391 } 2392 2393 tmp[filters.nr_slabs++] = sym; 2394 filters.slabs = tmp; 2395 return true; 2396 } 2397 2398 static int parse_lock_addr(const struct option *opt __maybe_unused, const char *str, 2399 int unset __maybe_unused) 2400 { 2401 char *s, *tmp, *tok; 2402 int ret = 0; 2403 u64 addr; 2404 2405 s = strdup(str); 2406 if (s == NULL) 2407 return -1; 2408 2409 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) { 2410 char *end; 2411 2412 addr = strtoul(tok, &end, 16); 2413 if (*end == '\0') { 2414 if (!add_lock_addr(addr)) { 2415 ret = -1; 2416 break; 2417 } 2418 continue; 2419 } 2420 2421 if (*tok == '&') { 2422 if (!add_lock_slab(tok + 1)) { 2423 ret = -1; 2424 break; 2425 } 2426 continue; 2427 } 2428 2429 /* 2430 * At this moment, we don't have kernel symbols. Save the symbols 2431 * in a separate list and resolve them to addresses later. 2432 */ 2433 if (!add_lock_sym(tok)) { 2434 ret = -1; 2435 break; 2436 } 2437 } 2438 2439 free(s); 2440 return ret; 2441 } 2442 2443 static int parse_output(const struct option *opt __maybe_unused, const char *str, 2444 int unset __maybe_unused) 2445 { 2446 const char **name = (const char **)opt->value; 2447 2448 if (str == NULL) 2449 return -1; 2450 2451 lock_output = fopen(str, "w"); 2452 if (lock_output == NULL) { 2453 pr_err("Cannot open %s\n", str); 2454 return -1; 2455 } 2456 2457 *name = str; 2458 return 0; 2459 } 2460 2461 static bool add_lock_cgroup(char *name) 2462 { 2463 u64 *tmp; 2464 struct cgroup *cgrp; 2465 2466 cgrp = cgroup__new(name, /*do_open=*/false); 2467 if (cgrp == NULL) { 2468 pr_err("Failed to create cgroup: %s\n", name); 2469 return false; 2470 } 2471 2472 if (read_cgroup_id(cgrp) < 0) { 2473 pr_err("Failed to read cgroup id for %s\n", name); 2474 cgroup__put(cgrp); 2475 return false; 2476 } 2477 2478 tmp = realloc(filters.cgrps, (filters.nr_cgrps + 1) * sizeof(*filters.cgrps)); 2479 if (tmp == NULL) { 2480 pr_err("Memory allocation failure\n"); 2481 return false; 2482 } 2483 2484 tmp[filters.nr_cgrps++] = cgrp->id; 2485 filters.cgrps = tmp; 2486 cgroup__put(cgrp); 2487 return true; 2488 } 2489 2490 static int parse_cgroup_filter(const struct option *opt __maybe_unused, const char *str, 2491 int unset __maybe_unused) 2492 { 2493 char *s, *tmp, *tok; 2494 int ret = 0; 2495 2496 s = strdup(str); 2497 if (s == NULL) 2498 return -1; 2499 2500 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) { 2501 if (!add_lock_cgroup(tok)) { 2502 ret = -1; 2503 break; 2504 } 2505 } 2506 2507 free(s); 2508 return ret; 2509 } 2510 2511 static bool add_lock_delay(char *spec) 2512 { 2513 char *at, *pos; 2514 struct lock_delay *tmp; 2515 unsigned long duration; 2516 2517 at = strchr(spec, '@'); 2518 if (at == NULL) { 2519 pr_err("lock delay should have '@' sign: %s\n", spec); 2520 return false; 2521 } 2522 if (at == spec) { 2523 pr_err("lock delay should have time before '@': %s\n", spec); 2524 return false; 2525 } 2526 2527 *at = '\0'; 2528 duration = strtoul(spec, &pos, 0); 2529 if (!strcmp(pos, "ns")) 2530 duration *= 1; 2531 else if (!strcmp(pos, "us")) 2532 duration *= 1000; 2533 else if (!strcmp(pos, "ms")) 2534 duration *= 1000 * 1000; 2535 else if (*pos) { 2536 pr_err("invalid delay time: %s@%s\n", spec, at + 1); 2537 return false; 2538 } 2539 2540 if (duration > 10 * 1000 * 1000) { 2541 pr_err("lock delay is too long: %s (> 10ms)\n", spec); 2542 return false; 2543 } 2544 2545 tmp = realloc(delays, (nr_delays + 1) * sizeof(*delays)); 2546 if (tmp == NULL) { 2547 pr_err("Memory allocation failure\n"); 2548 return false; 2549 } 2550 delays = tmp; 2551 2552 delays[nr_delays].sym = strdup(at + 1); 2553 if (delays[nr_delays].sym == NULL) { 2554 pr_err("Memory allocation failure\n"); 2555 return false; 2556 } 2557 delays[nr_delays].time = duration; 2558 2559 nr_delays++; 2560 return true; 2561 } 2562 2563 static int parse_lock_delay(const struct option *opt __maybe_unused, const char *str, 2564 int unset __maybe_unused) 2565 { 2566 char *s, *tmp, *tok; 2567 int ret = 0; 2568 2569 s = strdup(str); 2570 if (s == NULL) 2571 return -1; 2572 2573 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) { 2574 if (!add_lock_delay(tok)) { 2575 ret = -1; 2576 break; 2577 } 2578 } 2579 2580 free(s); 2581 return ret; 2582 } 2583 2584 int cmd_lock(int argc, const char **argv) 2585 { 2586 const struct option lock_options[] = { 2587 OPT_STRING('i', "input", &input_name, "file", "input file name"), 2588 OPT_CALLBACK(0, "output", &output_name, "file", "output file name", parse_output), 2589 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"), 2590 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"), 2591 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"), 2592 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name, 2593 "file", "vmlinux pathname"), 2594 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 2595 "file", "kallsyms pathname"), 2596 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"), 2597 OPT_END() 2598 }; 2599 2600 const struct option info_options[] = { 2601 OPT_BOOLEAN('t', "threads", &info_threads, 2602 "dump the thread list in perf.data"), 2603 OPT_BOOLEAN('m', "map", &info_map, 2604 "dump the map of lock instances (address:name table)"), 2605 OPT_PARENT(lock_options) 2606 }; 2607 2608 const struct option report_options[] = { 2609 OPT_STRING('k', "key", &sort_key, "acquired", 2610 "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"), 2611 OPT_STRING('F', "field", &output_fields, NULL, 2612 "output fields (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"), 2613 /* TODO: type */ 2614 OPT_BOOLEAN('c', "combine-locks", &combine_locks, 2615 "combine locks in the same class"), 2616 OPT_BOOLEAN('t', "threads", &show_thread_stats, 2617 "show per-thread lock stats"), 2618 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"), 2619 OPT_PARENT(lock_options) 2620 }; 2621 2622 struct option contention_options[] = { 2623 OPT_STRING('k', "key", &sort_key, "wait_total", 2624 "key for sorting (contended / wait_total / wait_max / wait_min / avg_wait)"), 2625 OPT_STRING('F', "field", &output_fields, "contended,wait_total,wait_max,avg_wait", 2626 "output fields (contended / wait_total / wait_max / wait_min / avg_wait)"), 2627 OPT_BOOLEAN('t', "threads", &show_thread_stats, 2628 "show per-thread lock stats"), 2629 OPT_BOOLEAN('b', "use-bpf", &use_bpf, "use BPF program to collect lock contention stats"), 2630 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 2631 "System-wide collection from all CPUs"), 2632 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 2633 "List of cpus to monitor"), 2634 OPT_STRING('p', "pid", &target.pid, "pid", 2635 "Trace on existing process id"), 2636 OPT_STRING(0, "tid", &target.tid, "tid", 2637 "Trace on existing thread id (exclusive to --pid)"), 2638 OPT_CALLBACK('M', "map-nr-entries", &bpf_map_entries, "num", 2639 "Max number of BPF map entries", parse_map_entry), 2640 OPT_CALLBACK(0, "max-stack", &max_stack_depth, "num", 2641 "Set the maximum stack depth when collecting lock contention, " 2642 "Default: " __stringify(CONTENTION_STACK_DEPTH), parse_max_stack), 2643 OPT_INTEGER(0, "stack-skip", &stack_skip, 2644 "Set the number of stack depth to skip when finding a lock caller, " 2645 "Default: " __stringify(CONTENTION_STACK_SKIP)), 2646 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"), 2647 OPT_BOOLEAN('l', "lock-addr", &show_lock_addrs, "show lock stats by address"), 2648 OPT_CALLBACK('Y', "type-filter", NULL, "FLAGS", 2649 "Filter specific type of locks", parse_lock_type), 2650 OPT_CALLBACK('L', "lock-filter", NULL, "ADDRS/NAMES", 2651 "Filter specific address/symbol of locks", parse_lock_addr), 2652 OPT_CALLBACK('S', "callstack-filter", NULL, "NAMES", 2653 "Filter specific function in the callstack", parse_call_stack), 2654 OPT_BOOLEAN('o', "lock-owner", &show_lock_owner, "show lock owners instead of waiters"), 2655 OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep, "separator", 2656 "print result in CSV format with custom separator"), 2657 OPT_BOOLEAN(0, "lock-cgroup", &show_lock_cgroups, "show lock stats by cgroup"), 2658 OPT_CALLBACK('G', "cgroup-filter", NULL, "CGROUPS", 2659 "Filter specific cgroups", parse_cgroup_filter), 2660 OPT_CALLBACK('J', "inject-delay", NULL, "TIME@FUNC", 2661 "Inject delays to specific locks", parse_lock_delay), 2662 OPT_PARENT(lock_options) 2663 }; 2664 2665 const char * const info_usage[] = { 2666 "perf lock info [<options>]", 2667 NULL 2668 }; 2669 const char *const lock_subcommands[] = { "record", "report", "script", 2670 "info", "contention", NULL }; 2671 const char *lock_usage[] = { 2672 NULL, 2673 NULL 2674 }; 2675 const char * const report_usage[] = { 2676 "perf lock report [<options>]", 2677 NULL 2678 }; 2679 const char * const contention_usage[] = { 2680 "perf lock contention [<options>]", 2681 NULL 2682 }; 2683 unsigned int i; 2684 int rc = 0; 2685 2686 lockhash_table = calloc(LOCKHASH_SIZE, sizeof(*lockhash_table)); 2687 if (!lockhash_table) 2688 return -ENOMEM; 2689 2690 for (i = 0; i < LOCKHASH_SIZE; i++) 2691 INIT_HLIST_HEAD(lockhash_table + i); 2692 2693 lock_output = stderr; 2694 argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands, 2695 lock_usage, PARSE_OPT_STOP_AT_NON_OPTION); 2696 if (!argc) 2697 usage_with_options(lock_usage, lock_options); 2698 2699 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2700 return __cmd_record(argc, argv); 2701 } else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) { 2702 trace_handler = &report_lock_ops; 2703 if (argc) { 2704 argc = parse_options(argc, argv, 2705 report_options, report_usage, 0); 2706 if (argc) 2707 usage_with_options(report_usage, report_options); 2708 } 2709 rc = __cmd_report(false); 2710 } else if (!strcmp(argv[0], "script")) { 2711 /* Aliased to 'perf script' */ 2712 rc = cmd_script(argc, argv); 2713 } else if (!strcmp(argv[0], "info")) { 2714 if (argc) { 2715 argc = parse_options(argc, argv, 2716 info_options, info_usage, 0); 2717 if (argc) 2718 usage_with_options(info_usage, info_options); 2719 } 2720 2721 /* If neither threads nor map requested, display both */ 2722 if (!info_threads && !info_map) { 2723 info_threads = true; 2724 info_map = true; 2725 } 2726 2727 /* recycling report_lock_ops */ 2728 trace_handler = &report_lock_ops; 2729 rc = __cmd_report(true); 2730 } else if (strlen(argv[0]) > 2 && strstarts("contention", argv[0])) { 2731 trace_handler = &contention_lock_ops; 2732 sort_key = "wait_total"; 2733 output_fields = "contended,wait_total,wait_max,avg_wait"; 2734 2735 #ifndef HAVE_BPF_SKEL 2736 set_option_nobuild(contention_options, 'b', "use-bpf", 2737 "no BUILD_BPF_SKEL=1", false); 2738 #endif 2739 if (argc) { 2740 argc = parse_options(argc, argv, contention_options, 2741 contention_usage, 0); 2742 } 2743 2744 if (check_lock_contention_options(contention_options, 2745 contention_usage) < 0) 2746 return -1; 2747 2748 rc = __cmd_contention(argc, argv); 2749 } else { 2750 usage_with_options(lock_usage, lock_options); 2751 } 2752 2753 /* free usage string allocated by parse_options_subcommand */ 2754 free((void *)lock_usage[0]); 2755 2756 zfree(&lockhash_table); 2757 return rc; 2758 } 2759