subr_witness.c revision 118441
1/*- 2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 3. Berkeley Software Design Inc's name may not be used to endorse or 13 * promote products derived from this software without specific prior 14 * written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 29 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 30 */ 31 32/* 33 * Implementation of the `witness' lock verifier. Originally implemented for 34 * mutexes in BSD/OS. Extended to handle generic lock objects and lock 35 * classes in FreeBSD. 36 */ 37 38/* 39 * Main Entry: witness 40 * Pronunciation: 'wit-n&s 41 * Function: noun 42 * Etymology: Middle English witnesse, from Old English witnes knowledge, 43 * testimony, witness, from 2wit 44 * Date: before 12th century 45 * 1 : attestation of a fact or event : TESTIMONY 46 * 2 : one that gives evidence; specifically : one who testifies in 47 * a cause or before a judicial tribunal 48 * 3 : one asked to be present at a transaction so as to be able to 49 * testify to its having taken place 50 * 4 : one who has personal knowledge of something 51 * 5 a : something serving as evidence or proof : SIGN 52 * b : public affirmation by word or example of usually 53 * religious faith or conviction <the heroic witness to divine 54 * life -- Pilot> 55 * 6 capitalized : a member of the Jehovah's Witnesses 56 */ 57 58/* 59 * Special rules concerning Giant and lock orders: 60 * 61 * 1) Giant must be acquired before any other mutexes. Stated another way, 62 * no other mutex may be held when Giant is acquired. 63 * 64 * 2) Giant must be released when blocking on a sleepable lock. 65 * 66 * This rule is less obvious, but is a result of Giant providing the same 67 * semantics as spl(). Basically, when a thread sleeps, it must release 68 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule 69 * 2). 70 * 71 * 3) Giant may be acquired before or after sleepable locks. 72 * 73 * This rule is also not quite as obvious. Giant may be acquired after 74 * a sleepable lock because it is a non-sleepable lock and non-sleepable 75 * locks may always be acquired while holding a sleepable lock. The second 76 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose 77 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 78 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and 79 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to 80 * execute. Thus, acquiring Giant both before and after a sleepable lock 81 * will not result in a lock order reversal. 82 */ 83 84#include <sys/cdefs.h> 85__FBSDID("$FreeBSD: head/sys/kern/subr_witness.c 118441 2003-08-04 19:24:25Z jhb $"); 86 87#include "opt_ddb.h" 88#include "opt_witness.h" 89 90#include <sys/param.h> 91#include <sys/bus.h> 92#include <sys/kernel.h> 93#include <sys/ktr.h> 94#include <sys/lock.h> 95#include <sys/malloc.h> 96#include <sys/mutex.h> 97#include <sys/proc.h> 98#include <sys/sysctl.h> 99#include <sys/systm.h> 100 101#include <ddb/ddb.h> 102 103#include <machine/stdarg.h> 104 105/* Define this to check for blessed mutexes */ 106#undef BLESSING 107 108#define WITNESS_COUNT 200 109#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 110/* 111 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads 112 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should 113 * probably be safe for the most part, but it's still a SWAG. 114 */ 115#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2 116 117#define WITNESS_NCHILDREN 6 118 119struct witness_child_list_entry; 120 121struct witness { 122 const char *w_name; 123 struct lock_class *w_class; 124 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 125 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 126 struct witness_child_list_entry *w_children; /* Great evilness... */ 127 const char *w_file; 128 int w_line; 129 u_int w_level; 130 u_int w_refcount; 131 u_char w_Giant_squawked:1; 132 u_char w_other_squawked:1; 133 u_char w_same_squawked:1; 134 u_char w_displayed:1; 135}; 136 137struct witness_child_list_entry { 138 struct witness_child_list_entry *wcl_next; 139 struct witness *wcl_children[WITNESS_NCHILDREN]; 140 u_int wcl_count; 141}; 142 143STAILQ_HEAD(witness_list, witness); 144 145#ifdef BLESSING 146struct witness_blessed { 147 const char *b_lock1; 148 const char *b_lock2; 149}; 150#endif 151 152struct witness_order_list_entry { 153 const char *w_name; 154 struct lock_class *w_class; 155}; 156 157#ifdef BLESSING 158static int blessed(struct witness *, struct witness *); 159#endif 160static int depart(struct witness *w); 161static struct witness *enroll(const char *description, 162 struct lock_class *lock_class); 163static int insertchild(struct witness *parent, struct witness *child); 164static int isitmychild(struct witness *parent, struct witness *child); 165static int isitmydescendant(struct witness *parent, struct witness *child); 166static int itismychild(struct witness *parent, struct witness *child); 167static int rebalancetree(struct witness_list *list); 168static void removechild(struct witness *parent, struct witness *child); 169static int reparentchildren(struct witness *newparent, 170 struct witness *oldparent); 171static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 172static void witness_displaydescendants(void(*)(const char *fmt, ...), 173 struct witness *, int indent); 174static const char *fixup_filename(const char *file); 175static void witness_leveldescendents(struct witness *parent, int level); 176static void witness_levelall(void); 177static struct witness *witness_get(void); 178static void witness_free(struct witness *m); 179static struct witness_child_list_entry *witness_child_get(void); 180static void witness_child_free(struct witness_child_list_entry *wcl); 181static struct lock_list_entry *witness_lock_list_get(void); 182static void witness_lock_list_free(struct lock_list_entry *lle); 183static struct lock_instance *find_instance(struct lock_list_entry *lock_list, 184 struct lock_object *lock); 185static void witness_list_lock(struct lock_instance *instance); 186#ifdef DDB 187static void witness_list(struct thread *td); 188static void witness_display_list(void(*prnt)(const char *fmt, ...), 189 struct witness_list *list); 190static void witness_display(void(*)(const char *fmt, ...)); 191#endif 192 193MALLOC_DEFINE(M_WITNESS, "witness", "witness structure"); 194 195/* 196 * If set to 0, witness is disabled. If set to 1, witness performs full lock 197 * order checking for all locks. If set to 2 or higher, then witness skips 198 * the full lock order check if the lock being acquired is at a higher level 199 * (i.e. farther down in the tree) than the current lock. This last mode is 200 * somewhat experimental and not considered fully safe. At runtime, this 201 * value may be set to 0 to turn off witness. witness is not allowed be 202 * turned on once it is turned off, however. 203 */ 204static int witness_watch = 1; 205TUNABLE_INT("debug.witness_watch", &witness_watch); 206SYSCTL_PROC(_debug, OID_AUTO, witness_watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 207 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 208 209#ifdef DDB 210/* 211 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to 212 * drop into kdebug() when: 213 * - a lock heirarchy violation occurs 214 * - locks are held when going to sleep. 215 */ 216#ifdef WITNESS_DDB 217int witness_ddb = 1; 218#else 219int witness_ddb = 0; 220#endif 221TUNABLE_INT("debug.witness_ddb", &witness_ddb); 222SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, ""); 223 224/* 225 * When DDB is enabled and witness_trace is set to 1, it will cause the system 226 * to print a stack trace: 227 * - a lock heirarchy violation occurs 228 * - locks are held when going to sleep. 229 */ 230int witness_trace = 1; 231TUNABLE_INT("debug.witness_trace", &witness_trace); 232SYSCTL_INT(_debug, OID_AUTO, witness_trace, CTLFLAG_RW, &witness_trace, 0, ""); 233#endif /* DDB */ 234 235#ifdef WITNESS_SKIPSPIN 236int witness_skipspin = 1; 237#else 238int witness_skipspin = 0; 239#endif 240TUNABLE_INT("debug.witness_skipspin", &witness_skipspin); 241SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0, 242 ""); 243 244static struct mtx w_mtx; 245static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 246static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 247static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 248static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 249static struct witness_child_list_entry *w_child_free = NULL; 250static struct lock_list_entry *w_lock_list_free = NULL; 251 252static struct witness w_data[WITNESS_COUNT]; 253static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT]; 254static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 255 256static struct witness_order_list_entry order_lists[] = { 257 { "proctree", &lock_class_sx }, 258 { "allproc", &lock_class_sx }, 259 { "Giant", &lock_class_mtx_sleep }, 260 { "filedesc structure", &lock_class_mtx_sleep }, 261 { "pipe mutex", &lock_class_mtx_sleep }, 262 { "sigio lock", &lock_class_mtx_sleep }, 263 { "process group", &lock_class_mtx_sleep }, 264 { "process lock", &lock_class_mtx_sleep }, 265 { "session", &lock_class_mtx_sleep }, 266 { "uidinfo hash", &lock_class_mtx_sleep }, 267 { "uidinfo struct", &lock_class_mtx_sleep }, 268 { "allprison", &lock_class_mtx_sleep }, 269 { NULL, NULL }, 270 /* 271 * spin locks 272 */ 273#ifdef SMP 274 { "ap boot", &lock_class_mtx_spin }, 275#ifdef __i386__ 276 { "com", &lock_class_mtx_spin }, 277#endif 278#endif 279 { "sio", &lock_class_mtx_spin }, 280#ifdef __i386__ 281 { "cy", &lock_class_mtx_spin }, 282#endif 283 { "sabtty", &lock_class_mtx_spin }, 284 { "zstty", &lock_class_mtx_spin }, 285 { "ng_node", &lock_class_mtx_spin }, 286 { "ng_worklist", &lock_class_mtx_spin }, 287 { "ithread table lock", &lock_class_mtx_spin }, 288 { "sched lock", &lock_class_mtx_spin }, 289 { "callout", &lock_class_mtx_spin }, 290 /* 291 * leaf locks 292 */ 293 { "allpmaps", &lock_class_mtx_spin }, 294 { "vm page queue free mutex", &lock_class_mtx_spin }, 295 { "icu", &lock_class_mtx_spin }, 296#ifdef SMP 297 { "smp rendezvous", &lock_class_mtx_spin }, 298#if defined(__i386__) && defined(APIC_IO) 299 { "tlb", &lock_class_mtx_spin }, 300#endif 301#ifdef __i386__ 302 { "lazypmap", &lock_class_mtx_spin }, 303#endif 304#ifdef __sparc64__ 305 { "ipi", &lock_class_mtx_spin }, 306#endif 307#endif 308 { "clk", &lock_class_mtx_spin }, 309 { "mutex profiling lock", &lock_class_mtx_spin }, 310 { "kse zombie lock", &lock_class_mtx_spin }, 311 { "ALD Queue", &lock_class_mtx_spin }, 312#ifdef __ia64__ 313 { "MCA spin lock", &lock_class_mtx_spin }, 314#endif 315#if defined(__i386__) || defined(__amd64__) 316 { "pcicfg", &lock_class_mtx_spin }, 317#endif 318 { NULL, NULL }, 319 { NULL, NULL } 320}; 321 322#ifdef BLESSING 323/* 324 * Pairs of locks which have been blessed 325 * Don't complain about order problems with blessed locks 326 */ 327static struct witness_blessed blessed_list[] = { 328}; 329static int blessed_count = 330 sizeof(blessed_list) / sizeof(struct witness_blessed); 331#endif 332 333/* 334 * List of all locks in the system. 335 */ 336TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks); 337 338static struct mtx all_mtx = { 339 { &lock_class_mtx_sleep, /* mtx_object.lo_class */ 340 "All locks list", /* mtx_object.lo_name */ 341 "All locks list", /* mtx_object.lo_type */ 342 LO_INITIALIZED, /* mtx_object.lo_flags */ 343 { NULL, NULL }, /* mtx_object.lo_list */ 344 NULL }, /* mtx_object.lo_witness */ 345 MTX_UNOWNED, 0, /* mtx_lock, mtx_recurse */ 346 TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked), 347 { NULL, NULL } /* mtx_contested */ 348}; 349 350/* 351 * This global is set to 0 once it becomes safe to use the witness code. 352 */ 353static int witness_cold = 1; 354 355/* 356 * Global variables for book keeping. 357 */ 358static int lock_cur_cnt; 359static int lock_max_cnt; 360 361/* 362 * The WITNESS-enabled diagnostic code. 363 */ 364static void 365witness_initialize(void *dummy __unused) 366{ 367 struct lock_object *lock; 368 struct witness_order_list_entry *order; 369 struct witness *w, *w1; 370 int i; 371 372 /* 373 * We have to release Giant before initializing its witness 374 * structure so that WITNESS doesn't get confused. 375 */ 376 mtx_unlock(&Giant); 377 mtx_assert(&Giant, MA_NOTOWNED); 378 379 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 380 TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list); 381 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 382 MTX_NOWITNESS); 383 for (i = 0; i < WITNESS_COUNT; i++) 384 witness_free(&w_data[i]); 385 for (i = 0; i < WITNESS_CHILDCOUNT; i++) 386 witness_child_free(&w_childdata[i]); 387 for (i = 0; i < LOCK_CHILDCOUNT; i++) 388 witness_lock_list_free(&w_locklistdata[i]); 389 390 /* First add in all the specified order lists. */ 391 for (order = order_lists; order->w_name != NULL; order++) { 392 w = enroll(order->w_name, order->w_class); 393 if (w == NULL) 394 continue; 395 w->w_file = "order list"; 396 for (order++; order->w_name != NULL; order++) { 397 w1 = enroll(order->w_name, order->w_class); 398 if (w1 == NULL) 399 continue; 400 w1->w_file = "order list"; 401 if (!itismychild(w, w1)) 402 panic("Not enough memory for static orders!"); 403 w = w1; 404 } 405 } 406 407 /* Iterate through all locks and add them to witness. */ 408 mtx_lock(&all_mtx); 409 TAILQ_FOREACH(lock, &all_locks, lo_list) { 410 if (lock->lo_flags & LO_WITNESS) 411 lock->lo_witness = enroll(lock->lo_type, 412 lock->lo_class); 413 else 414 lock->lo_witness = NULL; 415 } 416 mtx_unlock(&all_mtx); 417 418 /* Mark the witness code as being ready for use. */ 419 atomic_store_rel_int(&witness_cold, 0); 420 421 mtx_lock(&Giant); 422} 423SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL) 424 425static int 426sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 427{ 428 int error, value; 429 430 value = witness_watch; 431 error = sysctl_handle_int(oidp, &value, 0, req); 432 if (error != 0 || req->newptr == NULL) 433 return (error); 434 error = suser(req->td); 435 if (error != 0) 436 return (error); 437 if (value == witness_watch) 438 return (0); 439 if (value != 0) 440 return (EINVAL); 441 witness_watch = 0; 442 return (0); 443} 444 445void 446witness_init(struct lock_object *lock) 447{ 448 struct lock_class *class; 449 450 class = lock->lo_class; 451 if (lock->lo_flags & LO_INITIALIZED) 452 panic("%s: lock (%s) %s is already initialized", __func__, 453 class->lc_name, lock->lo_name); 454 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 455 (class->lc_flags & LC_RECURSABLE) == 0) 456 panic("%s: lock (%s) %s can not be recursable", __func__, 457 class->lc_name, lock->lo_name); 458 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 459 (class->lc_flags & LC_SLEEPABLE) == 0) 460 panic("%s: lock (%s) %s can not be sleepable", __func__, 461 class->lc_name, lock->lo_name); 462 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 463 (class->lc_flags & LC_UPGRADABLE) == 0) 464 panic("%s: lock (%s) %s can not be upgradable", __func__, 465 class->lc_name, lock->lo_name); 466 467 mtx_lock(&all_mtx); 468 TAILQ_INSERT_TAIL(&all_locks, lock, lo_list); 469 lock->lo_flags |= LO_INITIALIZED; 470 lock_cur_cnt++; 471 if (lock_cur_cnt > lock_max_cnt) 472 lock_max_cnt = lock_cur_cnt; 473 mtx_unlock(&all_mtx); 474 if (!witness_cold && witness_watch != 0 && panicstr == NULL && 475 (lock->lo_flags & LO_WITNESS) != 0) 476 lock->lo_witness = enroll(lock->lo_type, class); 477 else 478 lock->lo_witness = NULL; 479} 480 481void 482witness_destroy(struct lock_object *lock) 483{ 484 struct witness *w; 485 486 if (witness_cold) 487 panic("lock (%s) %s destroyed while witness_cold", 488 lock->lo_class->lc_name, lock->lo_name); 489 if ((lock->lo_flags & LO_INITIALIZED) == 0) 490 panic("%s: lock (%s) %s is not initialized", __func__, 491 lock->lo_class->lc_name, lock->lo_name); 492 493 /* XXX: need to verify that no one holds the lock */ 494 w = lock->lo_witness; 495 if (w != NULL) { 496 mtx_lock_spin(&w_mtx); 497 MPASS(w->w_refcount > 0); 498 w->w_refcount--; 499 500 /* 501 * Lock is already released if we have an allocation failure 502 * and depart() fails. 503 */ 504 if (w->w_refcount != 0 || depart(w)) 505 mtx_unlock_spin(&w_mtx); 506 } 507 508 mtx_lock(&all_mtx); 509 lock_cur_cnt--; 510 TAILQ_REMOVE(&all_locks, lock, lo_list); 511 lock->lo_flags &= ~LO_INITIALIZED; 512 mtx_unlock(&all_mtx); 513} 514 515#ifdef DDB 516static void 517witness_display_list(void(*prnt)(const char *fmt, ...), 518 struct witness_list *list) 519{ 520 struct witness *w; 521 522 STAILQ_FOREACH(w, list, w_typelist) { 523 if (w->w_file == NULL || w->w_level > 0) 524 continue; 525 /* 526 * This lock has no anscestors, display its descendants. 527 */ 528 witness_displaydescendants(prnt, w, 0); 529 } 530} 531 532static void 533witness_display(void(*prnt)(const char *fmt, ...)) 534{ 535 struct witness *w; 536 537 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 538 witness_levelall(); 539 540 /* Clear all the displayed flags. */ 541 STAILQ_FOREACH(w, &w_all, w_list) { 542 w->w_displayed = 0; 543 } 544 545 /* 546 * First, handle sleep locks which have been acquired at least 547 * once. 548 */ 549 prnt("Sleep locks:\n"); 550 witness_display_list(prnt, &w_sleep); 551 552 /* 553 * Now do spin locks which have been acquired at least once. 554 */ 555 prnt("\nSpin locks:\n"); 556 witness_display_list(prnt, &w_spin); 557 558 /* 559 * Finally, any locks which have not been acquired yet. 560 */ 561 prnt("\nLocks which were never acquired:\n"); 562 STAILQ_FOREACH(w, &w_all, w_list) { 563 if (w->w_file != NULL || w->w_refcount == 0) 564 continue; 565 prnt("%s\n", w->w_name); 566 } 567} 568#endif /* DDB */ 569 570/* Trim useless garbage from filenames. */ 571static const char * 572fixup_filename(const char *file) 573{ 574 575 if (file == NULL) 576 return (NULL); 577 while (strncmp(file, "../", 3) == 0) 578 file += 3; 579 return (file); 580} 581 582void 583witness_lock(struct lock_object *lock, int flags, const char *file, int line) 584{ 585 struct lock_list_entry **lock_list, *lle; 586 struct lock_instance *lock1, *lock2; 587 struct lock_class *class; 588 struct witness *w, *w1; 589 struct thread *td; 590 int i, j; 591#ifdef DDB 592 int go_into_ddb = 0; 593#endif 594 595 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 596 panicstr != NULL) 597 return; 598 w = lock->lo_witness; 599 class = lock->lo_class; 600 td = curthread; 601 file = fixup_filename(file); 602 603 if (class->lc_flags & LC_SLEEPLOCK) { 604 /* 605 * Since spin locks include a critical section, this check 606 * impliclty enforces a lock order of all sleep locks before 607 * all spin locks. 608 */ 609 if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0) 610 panic("blockable sleep lock (%s) %s @ %s:%d", 611 class->lc_name, lock->lo_name, file, line); 612 lock_list = &td->td_sleeplocks; 613 } else 614 lock_list = PCPU_PTR(spinlocks); 615 616 /* 617 * Is this the first lock acquired? If so, then no order checking 618 * is needed. 619 */ 620 if (*lock_list == NULL) 621 goto out; 622 623 /* 624 * Check to see if we are recursing on a lock we already own. 625 */ 626 lock1 = find_instance(*lock_list, lock); 627 if (lock1 != NULL) { 628 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 629 (flags & LOP_EXCLUSIVE) == 0) { 630 printf("shared lock of (%s) %s @ %s:%d\n", 631 class->lc_name, lock->lo_name, file, line); 632 printf("while exclusively locked from %s:%d\n", 633 lock1->li_file, lock1->li_line); 634 panic("share->excl"); 635 } 636 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 637 (flags & LOP_EXCLUSIVE) != 0) { 638 printf("exclusive lock of (%s) %s @ %s:%d\n", 639 class->lc_name, lock->lo_name, file, line); 640 printf("while share locked from %s:%d\n", 641 lock1->li_file, lock1->li_line); 642 panic("excl->share"); 643 } 644 lock1->li_flags++; 645 if ((lock->lo_flags & LO_RECURSABLE) == 0) { 646 printf( 647 "recursed on non-recursive lock (%s) %s @ %s:%d\n", 648 class->lc_name, lock->lo_name, file, line); 649 printf("first acquired @ %s:%d\n", lock1->li_file, 650 lock1->li_line); 651 panic("recurse"); 652 } 653 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 654 td->td_proc->p_pid, lock->lo_name, 655 lock1->li_flags & LI_RECURSEMASK); 656 lock1->li_file = file; 657 lock1->li_line = line; 658 return; 659 } 660 661 /* 662 * Try locks do not block if they fail to acquire the lock, thus 663 * there is no danger of deadlocks or of switching while holding a 664 * spin lock if we acquire a lock via a try operation. 665 */ 666 if (flags & LOP_TRYLOCK) 667 goto out; 668 669 /* 670 * Check for duplicate locks of the same type. Note that we only 671 * have to check for this on the last lock we just acquired. Any 672 * other cases will be caught as lock order violations. 673 */ 674 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 675 w1 = lock1->li_lock->lo_witness; 676 if (w1 == w) { 677 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK)) 678 goto out; 679 w->w_same_squawked = 1; 680 printf("acquiring duplicate lock of same type: \"%s\"\n", 681 lock->lo_type); 682 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name, 683 lock1->li_file, lock1->li_line); 684 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line); 685#ifdef DDB 686 go_into_ddb = 1; 687#endif 688 goto out; 689 } 690 MPASS(!mtx_owned(&w_mtx)); 691 mtx_lock_spin(&w_mtx); 692 /* 693 * If we have a known higher number just say ok 694 */ 695 if (witness_watch > 1 && w->w_level > w1->w_level) { 696 mtx_unlock_spin(&w_mtx); 697 goto out; 698 } 699 /* 700 * If we know that the the lock we are acquiring comes after 701 * the lock we most recently acquired in the lock order tree, 702 * then there is no need for any further checks. 703 */ 704 if (isitmydescendant(w1, w)) { 705 mtx_unlock_spin(&w_mtx); 706 goto out; 707 } 708 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) { 709 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 710 711 MPASS(j < WITNESS_COUNT); 712 lock1 = &lle->ll_children[i]; 713 w1 = lock1->li_lock->lo_witness; 714 715 /* 716 * If this lock doesn't undergo witness checking, 717 * then skip it. 718 */ 719 if (w1 == NULL) { 720 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 721 ("lock missing witness structure")); 722 continue; 723 } 724 /* 725 * If we are locking Giant and this is a sleepable 726 * lock, then skip it. 727 */ 728 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 729 lock == &Giant.mtx_object) 730 continue; 731 /* 732 * If we are locking a sleepable lock and this lock 733 * is Giant, then skip it. 734 */ 735 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 736 lock1->li_lock == &Giant.mtx_object) 737 continue; 738 /* 739 * If we are locking a sleepable lock and this lock 740 * isn't sleepable, we want to treat it as a lock 741 * order violation to enfore a general lock order of 742 * sleepable locks before non-sleepable locks. 743 */ 744 if (!((lock->lo_flags & LO_SLEEPABLE) != 0 && 745 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 746 /* 747 * Check the lock order hierarchy for a reveresal. 748 */ 749 if (!isitmydescendant(w, w1)) 750 continue; 751 /* 752 * We have a lock order violation, check to see if it 753 * is allowed or has already been yelled about. 754 */ 755 mtx_unlock_spin(&w_mtx); 756#ifdef BLESSING 757 if (blessed(w, w1)) 758 goto out; 759#endif 760 if (lock1->li_lock == &Giant.mtx_object) { 761 if (w1->w_Giant_squawked) 762 goto out; 763 else 764 w1->w_Giant_squawked = 1; 765 } else { 766 if (w1->w_other_squawked) 767 goto out; 768 else 769 w1->w_other_squawked = 1; 770 } 771 /* 772 * Ok, yell about it. 773 */ 774 printf("lock order reversal\n"); 775 /* 776 * Try to locate an earlier lock with 777 * witness w in our list. 778 */ 779 do { 780 lock2 = &lle->ll_children[i]; 781 MPASS(lock2->li_lock != NULL); 782 if (lock2->li_lock->lo_witness == w) 783 break; 784 i--; 785 if (i == 0 && lle->ll_next != NULL) { 786 lle = lle->ll_next; 787 i = lle->ll_count - 1; 788 MPASS(i >= 0 && i < LOCK_NCHILDREN); 789 } 790 } while (i >= 0); 791 if (i < 0) { 792 printf(" 1st %p %s (%s) @ %s:%d\n", 793 lock1->li_lock, lock1->li_lock->lo_name, 794 lock1->li_lock->lo_type, lock1->li_file, 795 lock1->li_line); 796 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 797 lock->lo_name, lock->lo_type, file, line); 798 } else { 799 printf(" 1st %p %s (%s) @ %s:%d\n", 800 lock2->li_lock, lock2->li_lock->lo_name, 801 lock2->li_lock->lo_type, lock2->li_file, 802 lock2->li_line); 803 printf(" 2nd %p %s (%s) @ %s:%d\n", 804 lock1->li_lock, lock1->li_lock->lo_name, 805 lock1->li_lock->lo_type, lock1->li_file, 806 lock1->li_line); 807 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 808 lock->lo_name, lock->lo_type, file, line); 809 } 810#ifdef DDB 811 go_into_ddb = 1; 812#endif 813 goto out; 814 } 815 } 816 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 817 /* 818 * Don't build a new relationship between a sleepable lock and 819 * Giant if it is the wrong direction. The real lock order is that 820 * sleepable locks come before Giant. 821 */ 822 if (!(lock1->li_lock == &Giant.mtx_object && 823 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 824 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 825 lock->lo_type, lock1->li_lock->lo_type); 826 if (!itismychild(lock1->li_lock->lo_witness, w)) 827 /* Witness is dead. */ 828 return; 829 } 830 mtx_unlock_spin(&w_mtx); 831 832out: 833#ifdef DDB 834 if (go_into_ddb) { 835 if (witness_trace) 836 backtrace(); 837 if (witness_ddb) 838 Debugger(__func__); 839 } 840#endif 841 w->w_file = file; 842 w->w_line = line; 843 844 lle = *lock_list; 845 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 846 lle = witness_lock_list_get(); 847 if (lle == NULL) 848 return; 849 lle->ll_next = *lock_list; 850 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 851 td->td_proc->p_pid, lle); 852 *lock_list = lle; 853 } 854 lock1 = &lle->ll_children[lle->ll_count++]; 855 lock1->li_lock = lock; 856 lock1->li_line = line; 857 lock1->li_file = file; 858 if ((flags & LOP_EXCLUSIVE) != 0) 859 lock1->li_flags = LI_EXCLUSIVE; 860 else 861 lock1->li_flags = 0; 862 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 863 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 864} 865 866void 867witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 868{ 869 struct lock_instance *instance; 870 struct lock_class *class; 871 872 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 873 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 874 return; 875 class = lock->lo_class; 876 file = fixup_filename(file); 877 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 878 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d", 879 class->lc_name, lock->lo_name, file, line); 880 if ((flags & LOP_TRYLOCK) == 0) 881 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name, 882 lock->lo_name, file, line); 883 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 884 panic("upgrade of non-sleep lock (%s) %s @ %s:%d", 885 class->lc_name, lock->lo_name, file, line); 886 instance = find_instance(curthread->td_sleeplocks, lock); 887 if (instance == NULL) 888 panic("upgrade of unlocked lock (%s) %s @ %s:%d", 889 class->lc_name, lock->lo_name, file, line); 890 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 891 panic("upgrade of exclusive lock (%s) %s @ %s:%d", 892 class->lc_name, lock->lo_name, file, line); 893 if ((instance->li_flags & LI_RECURSEMASK) != 0) 894 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d", 895 class->lc_name, lock->lo_name, 896 instance->li_flags & LI_RECURSEMASK, file, line); 897 instance->li_flags |= LI_EXCLUSIVE; 898} 899 900void 901witness_downgrade(struct lock_object *lock, int flags, const char *file, 902 int line) 903{ 904 struct lock_instance *instance; 905 struct lock_class *class; 906 907 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 908 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 909 return; 910 class = lock->lo_class; 911 file = fixup_filename(file); 912 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 913 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d", 914 class->lc_name, lock->lo_name, file, line); 915 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 916 panic("downgrade of non-sleep lock (%s) %s @ %s:%d", 917 class->lc_name, lock->lo_name, file, line); 918 instance = find_instance(curthread->td_sleeplocks, lock); 919 if (instance == NULL) 920 panic("downgrade of unlocked lock (%s) %s @ %s:%d", 921 class->lc_name, lock->lo_name, file, line); 922 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 923 panic("downgrade of shared lock (%s) %s @ %s:%d", 924 class->lc_name, lock->lo_name, file, line); 925 if ((instance->li_flags & LI_RECURSEMASK) != 0) 926 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d", 927 class->lc_name, lock->lo_name, 928 instance->li_flags & LI_RECURSEMASK, file, line); 929 instance->li_flags &= ~LI_EXCLUSIVE; 930} 931 932void 933witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 934{ 935 struct lock_list_entry **lock_list, *lle; 936 struct lock_instance *instance; 937 struct lock_class *class; 938 struct thread *td; 939 register_t s; 940 int i, j; 941 942 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 943 panicstr != NULL) 944 return; 945 td = curthread; 946 class = lock->lo_class; 947 file = fixup_filename(file); 948 if (class->lc_flags & LC_SLEEPLOCK) 949 lock_list = &td->td_sleeplocks; 950 else 951 lock_list = PCPU_PTR(spinlocks); 952 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 953 for (i = 0; i < (*lock_list)->ll_count; i++) { 954 instance = &(*lock_list)->ll_children[i]; 955 if (instance->li_lock == lock) { 956 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && 957 (flags & LOP_EXCLUSIVE) == 0) { 958 printf( 959 "shared unlock of (%s) %s @ %s:%d\n", 960 class->lc_name, lock->lo_name, 961 file, line); 962 printf( 963 "while exclusively locked from %s:%d\n", 964 instance->li_file, 965 instance->li_line); 966 panic("excl->ushare"); 967 } 968 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && 969 (flags & LOP_EXCLUSIVE) != 0) { 970 printf( 971 "exclusive unlock of (%s) %s @ %s:%d\n", 972 class->lc_name, lock->lo_name, 973 file, line); 974 printf( 975 "while share locked from %s:%d\n", 976 instance->li_file, 977 instance->li_line); 978 panic("share->uexcl"); 979 } 980 /* If we are recursed, unrecurse. */ 981 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 982 CTR4(KTR_WITNESS, 983 "%s: pid %d unrecursed on %s r=%d", __func__, 984 td->td_proc->p_pid, 985 instance->li_lock->lo_name, 986 instance->li_flags); 987 instance->li_flags--; 988 return; 989 } 990 s = intr_disable(); 991 CTR4(KTR_WITNESS, 992 "%s: pid %d removed %s from lle[%d]", __func__, 993 td->td_proc->p_pid, 994 instance->li_lock->lo_name, 995 (*lock_list)->ll_count - 1); 996 for (j = i; j < (*lock_list)->ll_count - 1; j++) 997 (*lock_list)->ll_children[j] = 998 (*lock_list)->ll_children[j + 1]; 999 (*lock_list)->ll_count--; 1000 intr_restore(s); 1001 if ((*lock_list)->ll_count == 0) { 1002 lle = *lock_list; 1003 *lock_list = lle->ll_next; 1004 CTR3(KTR_WITNESS, 1005 "%s: pid %d removed lle %p", __func__, 1006 td->td_proc->p_pid, lle); 1007 witness_lock_list_free(lle); 1008 } 1009 return; 1010 } 1011 } 1012 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, 1013 file, line); 1014} 1015 1016/* 1017 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1018 * exempt Giant and sleepable locks from the checks as well. If any 1019 * non-exempt locks are held, then a supplied message is printed to the 1020 * console along with a list of the offending locks. If indicated in the 1021 * flags then a failure results in a panic as well. 1022 */ 1023int 1024witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1025{ 1026 struct lock_list_entry *lle; 1027 struct lock_instance *lock1; 1028 struct thread *td; 1029 va_list ap; 1030 int i, n; 1031 1032 if (witness_cold || witness_watch == 0 || panicstr != NULL) 1033 return (0); 1034 n = 0; 1035 td = curthread; 1036 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1037 for (i = lle->ll_count - 1; i >= 0; i--) { 1038 lock1 = &lle->ll_children[i]; 1039 if (lock1->li_lock == lock) 1040 continue; 1041 if (flags & WARN_GIANTOK && 1042 lock1->li_lock == &Giant.mtx_object) 1043 continue; 1044 if (flags & WARN_SLEEPOK && 1045 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1046 continue; 1047 if (n == 0) { 1048 va_start(ap, fmt); 1049 vprintf(fmt, ap); 1050 va_end(ap); 1051 printf(" with the following"); 1052 if (flags & WARN_SLEEPOK) 1053 printf(" non-sleepable"); 1054 printf(" locks held:\n"); 1055 } 1056 n++; 1057 witness_list_lock(lock1); 1058 } 1059 if (PCPU_GET(spinlocks) != NULL) { 1060 /* 1061 * Since we already hold a spinlock preemption is 1062 * already blocked. 1063 */ 1064 if (n == 0) { 1065 va_start(ap, fmt); 1066 vprintf(fmt, ap); 1067 va_end(ap); 1068 printf(" with the following"); 1069 if (flags & WARN_SLEEPOK) 1070 printf(" non-sleepable"); 1071 printf(" locks held:\n"); 1072 } 1073 n += witness_list_locks(PCPU_PTR(spinlocks)); 1074 } 1075 if (flags & WARN_PANIC && n) 1076 panic("witness_warn"); 1077#ifdef DDB 1078 else if (witness_ddb && n) 1079 Debugger(__func__); 1080#endif 1081 return (n); 1082} 1083 1084const char * 1085witness_file(struct lock_object *lock) 1086{ 1087 struct witness *w; 1088 1089 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1090 return ("?"); 1091 w = lock->lo_witness; 1092 return (w->w_file); 1093} 1094 1095int 1096witness_line(struct lock_object *lock) 1097{ 1098 struct witness *w; 1099 1100 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1101 return (0); 1102 w = lock->lo_witness; 1103 return (w->w_line); 1104} 1105 1106static struct witness * 1107enroll(const char *description, struct lock_class *lock_class) 1108{ 1109 struct witness *w; 1110 1111 if (!witness_watch || witness_watch == 0 || panicstr != NULL) 1112 return (NULL); 1113 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin) 1114 return (NULL); 1115 mtx_lock_spin(&w_mtx); 1116 STAILQ_FOREACH(w, &w_all, w_list) { 1117 if (w->w_name == description || (w->w_refcount > 0 && 1118 strcmp(description, w->w_name) == 0)) { 1119 w->w_refcount++; 1120 mtx_unlock_spin(&w_mtx); 1121 if (lock_class != w->w_class) 1122 panic( 1123 "lock (%s) %s does not match earlier (%s) lock", 1124 description, lock_class->lc_name, 1125 w->w_class->lc_name); 1126 return (w); 1127 } 1128 } 1129 /* 1130 * This isn't quite right, as witness_cold is still 0 while we 1131 * enroll all the locks initialized before witness_initialize(). 1132 */ 1133 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) { 1134 mtx_unlock_spin(&w_mtx); 1135 panic("spin lock %s not in order list", description); 1136 } 1137 if ((w = witness_get()) == NULL) 1138 return (NULL); 1139 w->w_name = description; 1140 w->w_class = lock_class; 1141 w->w_refcount = 1; 1142 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1143 if (lock_class->lc_flags & LC_SPINLOCK) 1144 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1145 else if (lock_class->lc_flags & LC_SLEEPLOCK) 1146 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1147 else { 1148 mtx_unlock_spin(&w_mtx); 1149 panic("lock class %s is not sleep or spin", 1150 lock_class->lc_name); 1151 } 1152 mtx_unlock_spin(&w_mtx); 1153 return (w); 1154} 1155 1156/* Don't let the door bang you on the way out... */ 1157static int 1158depart(struct witness *w) 1159{ 1160 struct witness_child_list_entry *wcl, *nwcl; 1161 struct witness_list *list; 1162 struct witness *parent; 1163 1164 MPASS(w->w_refcount == 0); 1165 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1166 list = &w_sleep; 1167 else 1168 list = &w_spin; 1169 /* 1170 * First, we run through the entire tree looking for any 1171 * witnesses that the outgoing witness is a child of. For 1172 * each parent that we find, we reparent all the direct 1173 * children of the outgoing witness to its parent. 1174 */ 1175 STAILQ_FOREACH(parent, list, w_typelist) { 1176 if (!isitmychild(parent, w)) 1177 continue; 1178 removechild(parent, w); 1179 if (!reparentchildren(parent, w)) 1180 return (0); 1181 } 1182 1183 /* 1184 * Now we go through and free up the child list of the 1185 * outgoing witness. 1186 */ 1187 for (wcl = w->w_children; wcl != NULL; wcl = nwcl) { 1188 nwcl = wcl->wcl_next; 1189 witness_child_free(wcl); 1190 } 1191 1192 /* 1193 * Detach from various lists and free. 1194 */ 1195 STAILQ_REMOVE(list, w, witness, w_typelist); 1196 STAILQ_REMOVE(&w_all, w, witness, w_list); 1197 witness_free(w); 1198 1199 /* Finally, fixup the tree. */ 1200 return (rebalancetree(list)); 1201} 1202 1203/* 1204 * Prune an entire lock order tree. We look for cases where a lock 1205 * is now both a descendant and a direct child of a given lock. In 1206 * that case, we want to remove the direct child link from the tree. 1207 * 1208 * Returns false if insertchild() fails. 1209 */ 1210static int 1211rebalancetree(struct witness_list *list) 1212{ 1213 struct witness *child, *parent; 1214 1215 STAILQ_FOREACH(child, list, w_typelist) { 1216 STAILQ_FOREACH(parent, list, w_typelist) { 1217 if (!isitmychild(parent, child)) 1218 continue; 1219 removechild(parent, child); 1220 if (isitmydescendant(parent, child)) 1221 continue; 1222 if (!insertchild(parent, child)) 1223 return (0); 1224 } 1225 } 1226 witness_levelall(); 1227 return (1); 1228} 1229 1230/* 1231 * Add "child" as a direct child of "parent". Returns false if 1232 * we fail due to out of memory. 1233 */ 1234static int 1235insertchild(struct witness *parent, struct witness *child) 1236{ 1237 struct witness_child_list_entry **wcl; 1238 1239 MPASS(child != NULL && parent != NULL); 1240 1241 /* 1242 * Insert "child" after "parent" 1243 */ 1244 wcl = &parent->w_children; 1245 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN) 1246 wcl = &(*wcl)->wcl_next; 1247 if (*wcl == NULL) { 1248 *wcl = witness_child_get(); 1249 if (*wcl == NULL) 1250 return (0); 1251 } 1252 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child; 1253 1254 return (1); 1255} 1256 1257/* 1258 * Make all the direct descendants of oldparent be direct descendants 1259 * of newparent. 1260 */ 1261static int 1262reparentchildren(struct witness *newparent, struct witness *oldparent) 1263{ 1264 struct witness_child_list_entry *wcl; 1265 int i; 1266 1267 /* Avoid making a witness a child of itself. */ 1268 MPASS(!isitmychild(oldparent, newparent)); 1269 1270 for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1271 for (i = 0; i < wcl->wcl_count; i++) 1272 if (!insertchild(newparent, wcl->wcl_children[i])) 1273 return (0); 1274 return (1); 1275} 1276 1277static int 1278itismychild(struct witness *parent, struct witness *child) 1279{ 1280 struct witness_list *list; 1281 1282 MPASS(child != NULL && parent != NULL); 1283 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) != 1284 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))) 1285 panic( 1286 "%s: parent (%s) and child (%s) are not the same lock type", 1287 __func__, parent->w_class->lc_name, 1288 child->w_class->lc_name); 1289 1290 if (!insertchild(parent, child)) 1291 return (0); 1292 1293 if (parent->w_class->lc_flags & LC_SLEEPLOCK) 1294 list = &w_sleep; 1295 else 1296 list = &w_spin; 1297 return (rebalancetree(list)); 1298} 1299 1300static void 1301removechild(struct witness *parent, struct witness *child) 1302{ 1303 struct witness_child_list_entry **wcl, *wcl1; 1304 int i; 1305 1306 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next) 1307 for (i = 0; i < (*wcl)->wcl_count; i++) 1308 if ((*wcl)->wcl_children[i] == child) 1309 goto found; 1310 return; 1311found: 1312 (*wcl)->wcl_count--; 1313 if ((*wcl)->wcl_count > i) 1314 (*wcl)->wcl_children[i] = 1315 (*wcl)->wcl_children[(*wcl)->wcl_count]; 1316 MPASS((*wcl)->wcl_children[i] != NULL); 1317 if ((*wcl)->wcl_count != 0) 1318 return; 1319 wcl1 = *wcl; 1320 *wcl = wcl1->wcl_next; 1321 witness_child_free(wcl1); 1322} 1323 1324static int 1325isitmychild(struct witness *parent, struct witness *child) 1326{ 1327 struct witness_child_list_entry *wcl; 1328 int i; 1329 1330 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1331 for (i = 0; i < wcl->wcl_count; i++) { 1332 if (wcl->wcl_children[i] == child) 1333 return (1); 1334 } 1335 } 1336 return (0); 1337} 1338 1339static int 1340isitmydescendant(struct witness *parent, struct witness *child) 1341{ 1342 struct witness_child_list_entry *wcl; 1343 int i, j; 1344 1345 if (isitmychild(parent, child)) 1346 return (1); 1347 j = 0; 1348 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1349 MPASS(j < 1000); 1350 for (i = 0; i < wcl->wcl_count; i++) { 1351 if (isitmydescendant(wcl->wcl_children[i], child)) 1352 return (1); 1353 } 1354 j++; 1355 } 1356 return (0); 1357} 1358 1359static void 1360witness_levelall (void) 1361{ 1362 struct witness_list *list; 1363 struct witness *w, *w1; 1364 1365 /* 1366 * First clear all levels. 1367 */ 1368 STAILQ_FOREACH(w, &w_all, w_list) { 1369 w->w_level = 0; 1370 } 1371 1372 /* 1373 * Look for locks with no parent and level all their descendants. 1374 */ 1375 STAILQ_FOREACH(w, &w_all, w_list) { 1376 /* 1377 * This is just an optimization, technically we could get 1378 * away just walking the all list each time. 1379 */ 1380 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1381 list = &w_sleep; 1382 else 1383 list = &w_spin; 1384 STAILQ_FOREACH(w1, list, w_typelist) { 1385 if (isitmychild(w1, w)) 1386 goto skip; 1387 } 1388 witness_leveldescendents(w, 0); 1389 skip: 1390 ; /* silence GCC 3.x */ 1391 } 1392} 1393 1394static void 1395witness_leveldescendents(struct witness *parent, int level) 1396{ 1397 struct witness_child_list_entry *wcl; 1398 int i; 1399 1400 if (parent->w_level < level) 1401 parent->w_level = level; 1402 level++; 1403 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1404 for (i = 0; i < wcl->wcl_count; i++) 1405 witness_leveldescendents(wcl->wcl_children[i], level); 1406} 1407 1408static void 1409witness_displaydescendants(void(*prnt)(const char *fmt, ...), 1410 struct witness *parent, int indent) 1411{ 1412 struct witness_child_list_entry *wcl; 1413 int i, level; 1414 1415 level = parent->w_level; 1416 prnt("%-2d", level); 1417 for (i = 0; i < indent; i++) 1418 prnt(" "); 1419 if (parent->w_refcount > 0) 1420 prnt("%s", parent->w_name); 1421 else 1422 prnt("(dead)"); 1423 if (parent->w_displayed) { 1424 prnt(" -- (already displayed)\n"); 1425 return; 1426 } 1427 parent->w_displayed = 1; 1428 if (parent->w_refcount > 0) { 1429 if (parent->w_file != NULL) 1430 prnt(" -- last acquired @ %s:%d", parent->w_file, 1431 parent->w_line); 1432 } 1433 prnt("\n"); 1434 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1435 for (i = 0; i < wcl->wcl_count; i++) 1436 witness_displaydescendants(prnt, 1437 wcl->wcl_children[i], indent + 1); 1438} 1439 1440#ifdef BLESSING 1441static int 1442blessed(struct witness *w1, struct witness *w2) 1443{ 1444 int i; 1445 struct witness_blessed *b; 1446 1447 for (i = 0; i < blessed_count; i++) { 1448 b = &blessed_list[i]; 1449 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1450 if (strcmp(w2->w_name, b->b_lock2) == 0) 1451 return (1); 1452 continue; 1453 } 1454 if (strcmp(w1->w_name, b->b_lock2) == 0) 1455 if (strcmp(w2->w_name, b->b_lock1) == 0) 1456 return (1); 1457 } 1458 return (0); 1459} 1460#endif 1461 1462static struct witness * 1463witness_get(void) 1464{ 1465 struct witness *w; 1466 1467 if (witness_watch == 0) { 1468 mtx_unlock_spin(&w_mtx); 1469 return (NULL); 1470 } 1471 if (STAILQ_EMPTY(&w_free)) { 1472 witness_watch = 0; 1473 mtx_unlock_spin(&w_mtx); 1474 printf("%s: witness exhausted\n", __func__); 1475 return (NULL); 1476 } 1477 w = STAILQ_FIRST(&w_free); 1478 STAILQ_REMOVE_HEAD(&w_free, w_list); 1479 bzero(w, sizeof(*w)); 1480 return (w); 1481} 1482 1483static void 1484witness_free(struct witness *w) 1485{ 1486 1487 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1488} 1489 1490static struct witness_child_list_entry * 1491witness_child_get(void) 1492{ 1493 struct witness_child_list_entry *wcl; 1494 1495 if (witness_watch == 0) { 1496 mtx_unlock_spin(&w_mtx); 1497 return (NULL); 1498 } 1499 wcl = w_child_free; 1500 if (wcl == NULL) { 1501 witness_watch = 0; 1502 mtx_unlock_spin(&w_mtx); 1503 printf("%s: witness exhausted\n", __func__); 1504 return (NULL); 1505 } 1506 w_child_free = wcl->wcl_next; 1507 bzero(wcl, sizeof(*wcl)); 1508 return (wcl); 1509} 1510 1511static void 1512witness_child_free(struct witness_child_list_entry *wcl) 1513{ 1514 1515 wcl->wcl_next = w_child_free; 1516 w_child_free = wcl; 1517} 1518 1519static struct lock_list_entry * 1520witness_lock_list_get(void) 1521{ 1522 struct lock_list_entry *lle; 1523 1524 if (witness_watch == 0) 1525 return (NULL); 1526 mtx_lock_spin(&w_mtx); 1527 lle = w_lock_list_free; 1528 if (lle == NULL) { 1529 witness_watch = 0; 1530 mtx_unlock_spin(&w_mtx); 1531 printf("%s: witness exhausted\n", __func__); 1532 return (NULL); 1533 } 1534 w_lock_list_free = lle->ll_next; 1535 mtx_unlock_spin(&w_mtx); 1536 bzero(lle, sizeof(*lle)); 1537 return (lle); 1538} 1539 1540static void 1541witness_lock_list_free(struct lock_list_entry *lle) 1542{ 1543 1544 mtx_lock_spin(&w_mtx); 1545 lle->ll_next = w_lock_list_free; 1546 w_lock_list_free = lle; 1547 mtx_unlock_spin(&w_mtx); 1548} 1549 1550static struct lock_instance * 1551find_instance(struct lock_list_entry *lock_list, struct lock_object *lock) 1552{ 1553 struct lock_list_entry *lle; 1554 struct lock_instance *instance; 1555 int i; 1556 1557 for (lle = lock_list; lle != NULL; lle = lle->ll_next) 1558 for (i = lle->ll_count - 1; i >= 0; i--) { 1559 instance = &lle->ll_children[i]; 1560 if (instance->li_lock == lock) 1561 return (instance); 1562 } 1563 return (NULL); 1564} 1565 1566static void 1567witness_list_lock(struct lock_instance *instance) 1568{ 1569 struct lock_object *lock; 1570 1571 lock = instance->li_lock; 1572 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 1573 "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name); 1574 if (lock->lo_type != lock->lo_name) 1575 printf(" (%s)", lock->lo_type); 1576 printf(" r = %d (%p) locked @ %s:%d\n", 1577 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file, 1578 instance->li_line); 1579} 1580 1581int 1582witness_list_locks(struct lock_list_entry **lock_list) 1583{ 1584 struct lock_list_entry *lle; 1585 int i, nheld; 1586 1587 nheld = 0; 1588 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 1589 for (i = lle->ll_count - 1; i >= 0; i--) { 1590 witness_list_lock(&lle->ll_children[i]); 1591 nheld++; 1592 } 1593 return (nheld); 1594} 1595 1596/* 1597 * This is a bit risky at best. We call this function when we have timed 1598 * out acquiring a spin lock, and we assume that the other CPU is stuck 1599 * with this lock held. So, we go groveling around in the other CPU's 1600 * per-cpu data to try to find the lock instance for this spin lock to 1601 * see when it was last acquired. 1602 */ 1603void 1604witness_display_spinlock(struct lock_object *lock, struct thread *owner) 1605{ 1606 struct lock_instance *instance; 1607 struct pcpu *pc; 1608 1609 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 1610 return; 1611 pc = pcpu_find(owner->td_oncpu); 1612 instance = find_instance(pc->pc_spinlocks, lock); 1613 if (instance != NULL) 1614 witness_list_lock(instance); 1615} 1616 1617void 1618witness_save(struct lock_object *lock, const char **filep, int *linep) 1619{ 1620 struct lock_instance *instance; 1621 1622 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1623 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1624 return; 1625 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1626 panic("%s: lock (%s) %s is not a sleep lock", __func__, 1627 lock->lo_class->lc_name, lock->lo_name); 1628 instance = find_instance(curthread->td_sleeplocks, lock); 1629 if (instance == NULL) 1630 panic("%s: lock (%s) %s not locked", __func__, 1631 lock->lo_class->lc_name, lock->lo_name); 1632 *filep = instance->li_file; 1633 *linep = instance->li_line; 1634} 1635 1636void 1637witness_restore(struct lock_object *lock, const char *file, int line) 1638{ 1639 struct lock_instance *instance; 1640 1641 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1642 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1643 return; 1644 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1645 panic("%s: lock (%s) %s is not a sleep lock", __func__, 1646 lock->lo_class->lc_name, lock->lo_name); 1647 instance = find_instance(curthread->td_sleeplocks, lock); 1648 if (instance == NULL) 1649 panic("%s: lock (%s) %s not locked", __func__, 1650 lock->lo_class->lc_name, lock->lo_name); 1651 lock->lo_witness->w_file = file; 1652 lock->lo_witness->w_line = line; 1653 instance->li_file = file; 1654 instance->li_line = line; 1655} 1656 1657void 1658witness_assert(struct lock_object *lock, int flags, const char *file, int line) 1659{ 1660#ifdef INVARIANT_SUPPORT 1661 struct lock_instance *instance; 1662 1663 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1664 return; 1665 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0) 1666 instance = find_instance(curthread->td_sleeplocks, lock); 1667 else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0) 1668 instance = find_instance(PCPU_GET(spinlocks), lock); 1669 else { 1670 panic("Lock (%s) %s is not sleep or spin!", 1671 lock->lo_class->lc_name, lock->lo_name); 1672 } 1673 file = fixup_filename(file); 1674 switch (flags) { 1675 case LA_UNLOCKED: 1676 if (instance != NULL) 1677 panic("Lock (%s) %s locked @ %s:%d.", 1678 lock->lo_class->lc_name, lock->lo_name, file, line); 1679 break; 1680 case LA_LOCKED: 1681 case LA_LOCKED | LA_RECURSED: 1682 case LA_LOCKED | LA_NOTRECURSED: 1683 case LA_SLOCKED: 1684 case LA_SLOCKED | LA_RECURSED: 1685 case LA_SLOCKED | LA_NOTRECURSED: 1686 case LA_XLOCKED: 1687 case LA_XLOCKED | LA_RECURSED: 1688 case LA_XLOCKED | LA_NOTRECURSED: 1689 if (instance == NULL) { 1690 panic("Lock (%s) %s not locked @ %s:%d.", 1691 lock->lo_class->lc_name, lock->lo_name, file, line); 1692 break; 1693 } 1694 if ((flags & LA_XLOCKED) != 0 && 1695 (instance->li_flags & LI_EXCLUSIVE) == 0) 1696 panic("Lock (%s) %s not exclusively locked @ %s:%d.", 1697 lock->lo_class->lc_name, lock->lo_name, file, line); 1698 if ((flags & LA_SLOCKED) != 0 && 1699 (instance->li_flags & LI_EXCLUSIVE) != 0) 1700 panic("Lock (%s) %s exclusively locked @ %s:%d.", 1701 lock->lo_class->lc_name, lock->lo_name, file, line); 1702 if ((flags & LA_RECURSED) != 0 && 1703 (instance->li_flags & LI_RECURSEMASK) == 0) 1704 panic("Lock (%s) %s not recursed @ %s:%d.", 1705 lock->lo_class->lc_name, lock->lo_name, file, line); 1706 if ((flags & LA_NOTRECURSED) != 0 && 1707 (instance->li_flags & LI_RECURSEMASK) != 0) 1708 panic("Lock (%s) %s recursed @ %s:%d.", 1709 lock->lo_class->lc_name, lock->lo_name, file, line); 1710 break; 1711 default: 1712 panic("Invalid lock assertion at %s:%d.", file, line); 1713 1714 } 1715#endif /* INVARIANT_SUPPORT */ 1716} 1717 1718#ifdef DDB 1719static void 1720witness_list(struct thread *td) 1721{ 1722 1723 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1724 KASSERT(db_active, ("%s: not in the debugger", __func__)); 1725 1726 if (witness_watch == 0) 1727 return; 1728 1729 witness_list_locks(&td->td_sleeplocks); 1730 1731 /* 1732 * We only handle spinlocks if td == curthread. This is somewhat broken 1733 * if td is currently executing on some other CPU and holds spin locks 1734 * as we won't display those locks. If we had a MI way of getting 1735 * the per-cpu data for a given cpu then we could use 1736 * td->td_oncpu to get the list of spinlocks for this thread 1737 * and "fix" this. 1738 * 1739 * That still wouldn't really fix this unless we locked sched_lock 1740 * or stopped the other CPU to make sure it wasn't changing the list 1741 * out from under us. It is probably best to just not try to handle 1742 * threads on other CPU's for now. 1743 */ 1744 if (td == curthread && PCPU_GET(spinlocks) != NULL) 1745 witness_list_locks(PCPU_PTR(spinlocks)); 1746} 1747 1748DB_SHOW_COMMAND(locks, db_witness_list) 1749{ 1750 struct thread *td; 1751 pid_t pid; 1752 struct proc *p; 1753 1754 if (have_addr) { 1755 pid = (addr % 16) + ((addr >> 4) % 16) * 10 + 1756 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 + 1757 ((addr >> 16) % 16) * 10000; 1758 /* sx_slock(&allproc_lock); */ 1759 FOREACH_PROC_IN_SYSTEM(p) { 1760 if (p->p_pid == pid) 1761 break; 1762 } 1763 /* sx_sunlock(&allproc_lock); */ 1764 if (p == NULL) { 1765 db_printf("pid %d not found\n", pid); 1766 return; 1767 } 1768 FOREACH_THREAD_IN_PROC(p, td) { 1769 witness_list(td); 1770 } 1771 } else { 1772 td = curthread; 1773 witness_list(td); 1774 } 1775} 1776 1777DB_SHOW_COMMAND(witness, db_witness_display) 1778{ 1779 1780 witness_display(db_printf); 1781} 1782#endif 1783