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("$FreeBSD$"); 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/kdb.h> 93 #include <sys/kernel.h> 94 #include <sys/ktr.h> 95 #include <sys/lock.h> 96 #include <sys/malloc.h> 97 #include <sys/mutex.h> 98 #include <sys/proc.h> 99 #include <sys/sysctl.h> 100 #include <sys/systm.h> 101 102 #include <ddb/ddb.h> 103 104 #include <machine/stdarg.h> 105 106 /* Define this to check for blessed mutexes */ 107 #undef BLESSING 108 109 #define WITNESS_COUNT 200 110 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 111 /* 112 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads 113 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should 114 * probably be safe for the most part, but it's still a SWAG. 115 */ 116 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2 117 118 #define WITNESS_NCHILDREN 6 119 120 struct witness_child_list_entry; 121 122 struct witness { 123 const char *w_name; 124 struct lock_class *w_class; 125 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 126 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 127 struct witness_child_list_entry *w_children; /* Great evilness... */ 128 const char *w_file; 129 int w_line; 130 u_int w_level; 131 u_int w_refcount; 132 u_char w_Giant_squawked:1; 133 u_char w_other_squawked:1; 134 u_char w_same_squawked:1; 135 u_char w_displayed:1; 136 }; 137 138 struct witness_child_list_entry { 139 struct witness_child_list_entry *wcl_next; 140 struct witness *wcl_children[WITNESS_NCHILDREN]; 141 u_int wcl_count; 142 }; 143 144 STAILQ_HEAD(witness_list, witness); 145 146 #ifdef BLESSING 147 struct witness_blessed { 148 const char *b_lock1; 149 const char *b_lock2; 150 }; 151 #endif 152 153 struct witness_order_list_entry { 154 const char *w_name; 155 struct lock_class *w_class; 156 }; 157 158 #ifdef BLESSING 159 static int blessed(struct witness *, struct witness *); 160 #endif 161 static int depart(struct witness *w); 162 static struct witness *enroll(const char *description, 163 struct lock_class *lock_class); 164 static int insertchild(struct witness *parent, struct witness *child); 165 static int isitmychild(struct witness *parent, struct witness *child); 166 static int isitmydescendant(struct witness *parent, struct witness *child); 167 static int itismychild(struct witness *parent, struct witness *child); 168 static int rebalancetree(struct witness_list *list); 169 static void removechild(struct witness *parent, struct witness *child); 170 static int reparentchildren(struct witness *newparent, 171 struct witness *oldparent); 172 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 173 static void witness_displaydescendants(void(*)(const char *fmt, ...), 174 struct witness *, int indent); 175 static const char *fixup_filename(const char *file); 176 static void witness_leveldescendents(struct witness *parent, int level); 177 static void witness_levelall(void); 178 static struct witness *witness_get(void); 179 static void witness_free(struct witness *m); 180 static struct witness_child_list_entry *witness_child_get(void); 181 static void witness_child_free(struct witness_child_list_entry *wcl); 182 static struct lock_list_entry *witness_lock_list_get(void); 183 static void witness_lock_list_free(struct lock_list_entry *lle); 184 static struct lock_instance *find_instance(struct lock_list_entry *lock_list, 185 struct lock_object *lock); 186 static void witness_list_lock(struct lock_instance *instance); 187 #ifdef DDB 188 static void witness_list(struct thread *td); 189 static void witness_display_list(void(*prnt)(const char *fmt, ...), 190 struct witness_list *list); 191 static void witness_display(void(*)(const char *fmt, ...)); 192 #endif 193 194 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure"); 195 196 /* 197 * If set to 0, witness is disabled. If set to 1, witness performs full lock 198 * order checking for all locks. If set to 2 or higher, then witness skips 199 * the full lock order check if the lock being acquired is at a higher level 200 * (i.e. farther down in the tree) than the current lock. This last mode is 201 * somewhat experimental and not considered fully safe. At runtime, this 202 * value may be set to 0 to turn off witness. witness is not allowed be 203 * turned on once it is turned off, however. 204 */ 205 static int witness_watch = 1; 206 TUNABLE_INT("debug.witness_watch", &witness_watch); 207 SYSCTL_PROC(_debug, OID_AUTO, witness_watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 208 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 209 210 #ifdef KDB 211 /* 212 * When KDB is enabled and witness_kdb is set to 1, it will cause the system 213 * to drop into kdebug() when: 214 * - a lock heirarchy violation occurs 215 * - locks are held when going to sleep. 216 */ 217 #ifdef WITNESS_KDB 218 int witness_kdb = 1; 219 #else 220 int witness_kdb = 0; 221 #endif 222 TUNABLE_INT("debug.witness_kdb", &witness_kdb); 223 SYSCTL_INT(_debug, OID_AUTO, witness_kdb, CTLFLAG_RW, &witness_kdb, 0, ""); 224 225 /* 226 * When KDB is enabled and witness_trace is set to 1, it will cause the system 227 * to print a stack trace: 228 * - a lock heirarchy violation occurs 229 * - locks are held when going to sleep. 230 */ 231 int witness_trace = 1; 232 TUNABLE_INT("debug.witness_trace", &witness_trace); 233 SYSCTL_INT(_debug, OID_AUTO, witness_trace, CTLFLAG_RW, &witness_trace, 0, ""); 234 #endif /* KDB */ 235 236 #ifdef WITNESS_SKIPSPIN 237 int witness_skipspin = 1; 238 #else 239 int witness_skipspin = 0; 240 #endif 241 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin); 242 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, 243 ""); 244 245 static struct mtx w_mtx; 246 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 247 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 248 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 249 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 250 static struct witness_child_list_entry *w_child_free = NULL; 251 static struct lock_list_entry *w_lock_list_free = NULL; 252 253 static struct witness w_data[WITNESS_COUNT]; 254 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT]; 255 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 256 257 static struct witness_order_list_entry order_lists[] = { 258 { "proctree", &lock_class_sx }, 259 { "allproc", &lock_class_sx }, 260 { "Giant", &lock_class_mtx_sleep }, 261 { "filedesc structure", &lock_class_mtx_sleep }, 262 { "pipe mutex", &lock_class_mtx_sleep }, 263 { "sigio lock", &lock_class_mtx_sleep }, 264 { "process group", &lock_class_mtx_sleep }, 265 { "process lock", &lock_class_mtx_sleep }, 266 { "session", &lock_class_mtx_sleep }, 267 { "uidinfo hash", &lock_class_mtx_sleep }, 268 { "uidinfo struct", &lock_class_mtx_sleep }, 269 { "allprison", &lock_class_mtx_sleep }, 270 { NULL, NULL }, 271 /* 272 * Sockets 273 */ 274 { "filedesc structure", &lock_class_mtx_sleep }, 275 { "accept", &lock_class_mtx_sleep }, 276 { "so_snd", &lock_class_mtx_sleep }, 277 { "so_rcv", &lock_class_mtx_sleep }, 278 { "sellck", &lock_class_mtx_sleep }, 279 { NULL, NULL }, 280 /* 281 * Routing 282 */ 283 { "so_rcv", &lock_class_mtx_sleep }, 284 { "radix node head", &lock_class_mtx_sleep }, 285 { "rtentry", &lock_class_mtx_sleep }, 286 { "ifaddr", &lock_class_mtx_sleep }, 287 { NULL, NULL }, 288 /* 289 * UNIX Domain Sockets 290 */ 291 { "unp", &lock_class_mtx_sleep }, 292 { "so_snd", &lock_class_mtx_sleep }, 293 { NULL, NULL }, 294 /* 295 * UDP/IP 296 */ 297 { "udp", &lock_class_mtx_sleep }, 298 { "udpinp", &lock_class_mtx_sleep }, 299 { "so_snd", &lock_class_mtx_sleep }, 300 { NULL, NULL }, 301 /* 302 * TCP/IP 303 */ 304 { "tcp", &lock_class_mtx_sleep }, 305 { "tcpinp", &lock_class_mtx_sleep }, 306 { "so_snd", &lock_class_mtx_sleep }, 307 { NULL, NULL }, 308 /* 309 * SLIP 310 */ 311 { "slip_mtx", &lock_class_mtx_sleep }, 312 { "slip sc_mtx", &lock_class_mtx_sleep }, 313 { NULL, NULL }, 314 /* 315 * spin locks 316 */ 317 #ifdef SMP 318 { "ap boot", &lock_class_mtx_spin }, 319 #endif 320 { "sio", &lock_class_mtx_spin }, 321 #ifdef __i386__ 322 { "cy", &lock_class_mtx_spin }, 323 #endif 324 { "uart_hwmtx", &lock_class_mtx_spin }, 325 { "sabtty", &lock_class_mtx_spin }, 326 { "zstty", &lock_class_mtx_spin }, 327 { "ng_node", &lock_class_mtx_spin }, 328 { "ng_worklist", &lock_class_mtx_spin }, 329 { "taskqueue_fast", &lock_class_mtx_spin }, 330 { "intr table", &lock_class_mtx_spin }, 331 { "ithread table lock", &lock_class_mtx_spin }, 332 { "sleepq chain", &lock_class_mtx_spin }, 333 { "sched lock", &lock_class_mtx_spin }, 334 { "turnstile chain", &lock_class_mtx_spin }, 335 { "td_contested", &lock_class_mtx_spin }, 336 { "callout", &lock_class_mtx_spin }, 337 { "entropy harvest", &lock_class_mtx_spin }, 338 { "entropy harvest buffers", &lock_class_mtx_spin }, 339 /* 340 * leaf locks 341 */ 342 { "allpmaps", &lock_class_mtx_spin }, 343 { "vm page queue free mutex", &lock_class_mtx_spin }, 344 { "icu", &lock_class_mtx_spin }, 345 #ifdef SMP 346 { "smp rendezvous", &lock_class_mtx_spin }, 347 #if defined(__i386__) || defined(__amd64__) 348 { "tlb", &lock_class_mtx_spin }, 349 { "lazypmap", &lock_class_mtx_spin }, 350 #endif 351 #ifdef __sparc64__ 352 { "ipi", &lock_class_mtx_spin }, 353 #endif 354 #endif 355 { "clk", &lock_class_mtx_spin }, 356 { "mutex profiling lock", &lock_class_mtx_spin }, 357 { "kse zombie lock", &lock_class_mtx_spin }, 358 { "ALD Queue", &lock_class_mtx_spin }, 359 #ifdef __ia64__ 360 { "MCA spin lock", &lock_class_mtx_spin }, 361 #endif 362 #if defined(__i386__) || defined(__amd64__) 363 { "pcicfg", &lock_class_mtx_spin }, 364 #endif 365 { NULL, NULL }, 366 { NULL, NULL } 367 }; 368 369 #ifdef BLESSING 370 /* 371 * Pairs of locks which have been blessed 372 * Don't complain about order problems with blessed locks 373 */ 374 static struct witness_blessed blessed_list[] = { 375 }; 376 static int blessed_count = 377 sizeof(blessed_list) / sizeof(struct witness_blessed); 378 #endif 379 380 /* 381 * List of all locks in the system. 382 */ 383 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks); 384 385 static struct mtx all_mtx = { 386 { &lock_class_mtx_sleep, /* mtx_object.lo_class */ 387 "All locks list", /* mtx_object.lo_name */ 388 "All locks list", /* mtx_object.lo_type */ 389 LO_INITIALIZED, /* mtx_object.lo_flags */ 390 { NULL, NULL }, /* mtx_object.lo_list */ 391 NULL }, /* mtx_object.lo_witness */ 392 MTX_UNOWNED, 0 /* mtx_lock, mtx_recurse */ 393 }; 394 395 /* 396 * This global is set to 0 once it becomes safe to use the witness code. 397 */ 398 static int witness_cold = 1; 399 400 /* 401 * Global variables for book keeping. 402 */ 403 static int lock_cur_cnt; 404 static int lock_max_cnt; 405 406 /* 407 * The WITNESS-enabled diagnostic code. 408 */ 409 static void 410 witness_initialize(void *dummy __unused) 411 { 412 struct lock_object *lock; 413 struct witness_order_list_entry *order; 414 struct witness *w, *w1; 415 int i; 416 417 /* 418 * We have to release Giant before initializing its witness 419 * structure so that WITNESS doesn't get confused. 420 */ 421 mtx_unlock(&Giant); 422 mtx_assert(&Giant, MA_NOTOWNED); 423 424 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 425 TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list); 426 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 427 MTX_NOWITNESS); 428 for (i = 0; i < WITNESS_COUNT; i++) 429 witness_free(&w_data[i]); 430 for (i = 0; i < WITNESS_CHILDCOUNT; i++) 431 witness_child_free(&w_childdata[i]); 432 for (i = 0; i < LOCK_CHILDCOUNT; i++) 433 witness_lock_list_free(&w_locklistdata[i]); 434 435 /* First add in all the specified order lists. */ 436 for (order = order_lists; order->w_name != NULL; order++) { 437 w = enroll(order->w_name, order->w_class); 438 if (w == NULL) 439 continue; 440 w->w_file = "order list"; 441 for (order++; order->w_name != NULL; order++) { 442 w1 = enroll(order->w_name, order->w_class); 443 if (w1 == NULL) 444 continue; 445 w1->w_file = "order list"; 446 if (!itismychild(w, w1)) 447 panic("Not enough memory for static orders!"); 448 w = w1; 449 } 450 } 451 452 /* Iterate through all locks and add them to witness. */ 453 mtx_lock(&all_mtx); 454 TAILQ_FOREACH(lock, &all_locks, lo_list) { 455 if (lock->lo_flags & LO_WITNESS) 456 lock->lo_witness = enroll(lock->lo_type, 457 lock->lo_class); 458 else 459 lock->lo_witness = NULL; 460 } 461 mtx_unlock(&all_mtx); 462 463 /* Mark the witness code as being ready for use. */ 464 atomic_store_rel_int(&witness_cold, 0); 465 466 mtx_lock(&Giant); 467 } 468 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL) 469 470 static int 471 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 472 { 473 int error, value; 474 475 value = witness_watch; 476 error = sysctl_handle_int(oidp, &value, 0, req); 477 if (error != 0 || req->newptr == NULL) 478 return (error); 479 error = suser(req->td); 480 if (error != 0) 481 return (error); 482 if (value == witness_watch) 483 return (0); 484 if (value != 0) 485 return (EINVAL); 486 witness_watch = 0; 487 return (0); 488 } 489 490 void 491 witness_init(struct lock_object *lock) 492 { 493 struct lock_class *class; 494 495 class = lock->lo_class; 496 if (lock->lo_flags & LO_INITIALIZED) 497 panic("%s: lock (%s) %s is already initialized", __func__, 498 class->lc_name, lock->lo_name); 499 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 500 (class->lc_flags & LC_RECURSABLE) == 0) 501 panic("%s: lock (%s) %s can not be recursable", __func__, 502 class->lc_name, lock->lo_name); 503 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 504 (class->lc_flags & LC_SLEEPABLE) == 0) 505 panic("%s: lock (%s) %s can not be sleepable", __func__, 506 class->lc_name, lock->lo_name); 507 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 508 (class->lc_flags & LC_UPGRADABLE) == 0) 509 panic("%s: lock (%s) %s can not be upgradable", __func__, 510 class->lc_name, lock->lo_name); 511 512 mtx_lock(&all_mtx); 513 TAILQ_INSERT_TAIL(&all_locks, lock, lo_list); 514 lock->lo_flags |= LO_INITIALIZED; 515 lock_cur_cnt++; 516 if (lock_cur_cnt > lock_max_cnt) 517 lock_max_cnt = lock_cur_cnt; 518 mtx_unlock(&all_mtx); 519 if (!witness_cold && witness_watch != 0 && panicstr == NULL && 520 (lock->lo_flags & LO_WITNESS) != 0) 521 lock->lo_witness = enroll(lock->lo_type, class); 522 else 523 lock->lo_witness = NULL; 524 } 525 526 void 527 witness_destroy(struct lock_object *lock) 528 { 529 struct witness *w; 530 531 if (witness_cold) 532 panic("lock (%s) %s destroyed while witness_cold", 533 lock->lo_class->lc_name, lock->lo_name); 534 if ((lock->lo_flags & LO_INITIALIZED) == 0) 535 panic("%s: lock (%s) %s is not initialized", __func__, 536 lock->lo_class->lc_name, lock->lo_name); 537 538 /* XXX: need to verify that no one holds the lock */ 539 w = lock->lo_witness; 540 if (w != NULL) { 541 mtx_lock_spin(&w_mtx); 542 MPASS(w->w_refcount > 0); 543 w->w_refcount--; 544 545 /* 546 * Lock is already released if we have an allocation failure 547 * and depart() fails. 548 */ 549 if (w->w_refcount != 0 || depart(w)) 550 mtx_unlock_spin(&w_mtx); 551 } 552 553 mtx_lock(&all_mtx); 554 lock_cur_cnt--; 555 TAILQ_REMOVE(&all_locks, lock, lo_list); 556 lock->lo_flags &= ~LO_INITIALIZED; 557 mtx_unlock(&all_mtx); 558 } 559 560 #ifdef DDB 561 static void 562 witness_display_list(void(*prnt)(const char *fmt, ...), 563 struct witness_list *list) 564 { 565 struct witness *w; 566 567 STAILQ_FOREACH(w, list, w_typelist) { 568 if (w->w_file == NULL || w->w_level > 0) 569 continue; 570 /* 571 * This lock has no anscestors, display its descendants. 572 */ 573 witness_displaydescendants(prnt, w, 0); 574 } 575 } 576 577 static void 578 witness_display(void(*prnt)(const char *fmt, ...)) 579 { 580 struct witness *w; 581 582 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 583 witness_levelall(); 584 585 /* Clear all the displayed flags. */ 586 STAILQ_FOREACH(w, &w_all, w_list) { 587 w->w_displayed = 0; 588 } 589 590 /* 591 * First, handle sleep locks which have been acquired at least 592 * once. 593 */ 594 prnt("Sleep locks:\n"); 595 witness_display_list(prnt, &w_sleep); 596 597 /* 598 * Now do spin locks which have been acquired at least once. 599 */ 600 prnt("\nSpin locks:\n"); 601 witness_display_list(prnt, &w_spin); 602 603 /* 604 * Finally, any locks which have not been acquired yet. 605 */ 606 prnt("\nLocks which were never acquired:\n"); 607 STAILQ_FOREACH(w, &w_all, w_list) { 608 if (w->w_file != NULL || w->w_refcount == 0) 609 continue; 610 prnt("%s\n", w->w_name); 611 } 612 } 613 #endif /* DDB */ 614 615 /* Trim useless garbage from filenames. */ 616 static const char * 617 fixup_filename(const char *file) 618 { 619 620 if (file == NULL) 621 return (NULL); 622 while (strncmp(file, "../", 3) == 0) 623 file += 3; 624 return (file); 625 } 626 627 int 628 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 629 { 630 631 if (witness_watch == 0 || panicstr != NULL) 632 return (0); 633 634 /* Require locks that witness knows about. */ 635 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 636 lock2->lo_witness == NULL) 637 return (EINVAL); 638 639 MPASS(!mtx_owned(&w_mtx)); 640 mtx_lock_spin(&w_mtx); 641 642 /* 643 * If we already have either an explicit or implied lock order that 644 * is the other way around, then return an error. 645 */ 646 if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 647 mtx_unlock_spin(&w_mtx); 648 return (EDOOFUS); 649 } 650 651 /* Try to add the new order. */ 652 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 653 lock2->lo_type, lock1->lo_type); 654 if (!itismychild(lock1->lo_witness, lock2->lo_witness)) 655 return (ENOMEM); 656 mtx_unlock_spin(&w_mtx); 657 return (0); 658 } 659 660 void 661 witness_checkorder(struct lock_object *lock, int flags, const char *file, 662 int line) 663 { 664 struct lock_list_entry **lock_list, *lle; 665 struct lock_instance *lock1, *lock2; 666 struct lock_class *class; 667 struct witness *w, *w1; 668 struct thread *td; 669 int i, j; 670 671 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 672 panicstr != NULL) 673 return; 674 675 /* 676 * Try locks do not block if they fail to acquire the lock, thus 677 * there is no danger of deadlocks or of switching while holding a 678 * spin lock if we acquire a lock via a try operation. This 679 * function shouldn't even be called for try locks, so panic if 680 * that happens. 681 */ 682 if (flags & LOP_TRYLOCK) 683 panic("%s should not be called for try lock operations", 684 __func__); 685 686 w = lock->lo_witness; 687 class = lock->lo_class; 688 td = curthread; 689 file = fixup_filename(file); 690 691 if (class->lc_flags & LC_SLEEPLOCK) { 692 /* 693 * Since spin locks include a critical section, this check 694 * implicitly enforces a lock order of all sleep locks before 695 * all spin locks. 696 */ 697 if (td->td_critnest != 0) 698 panic("blockable sleep lock (%s) %s @ %s:%d", 699 class->lc_name, lock->lo_name, file, line); 700 701 /* 702 * If this is the first lock acquired then just return as 703 * no order checking is needed. 704 */ 705 if (td->td_sleeplocks == NULL) 706 return; 707 lock_list = &td->td_sleeplocks; 708 } else { 709 /* 710 * If this is the first lock, just return as no order 711 * checking is needed. We check this in both if clauses 712 * here as unifying the check would require us to use a 713 * critical section to ensure we don't migrate while doing 714 * the check. Note that if this is not the first lock, we 715 * are already in a critical section and are safe for the 716 * rest of the check. 717 */ 718 if (PCPU_GET(spinlocks) == NULL) 719 return; 720 lock_list = PCPU_PTR(spinlocks); 721 } 722 723 /* 724 * Check to see if we are recursing on a lock we already own. If 725 * so, make sure that we don't mismatch exclusive and shared lock 726 * acquires. 727 */ 728 lock1 = find_instance(*lock_list, lock); 729 if (lock1 != NULL) { 730 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 731 (flags & LOP_EXCLUSIVE) == 0) { 732 printf("shared lock of (%s) %s @ %s:%d\n", 733 class->lc_name, lock->lo_name, file, line); 734 printf("while exclusively locked from %s:%d\n", 735 lock1->li_file, lock1->li_line); 736 panic("share->excl"); 737 } 738 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 739 (flags & LOP_EXCLUSIVE) != 0) { 740 printf("exclusive lock of (%s) %s @ %s:%d\n", 741 class->lc_name, lock->lo_name, file, line); 742 printf("while share locked from %s:%d\n", 743 lock1->li_file, lock1->li_line); 744 panic("excl->share"); 745 } 746 return; 747 } 748 749 /* 750 * Try locks do not block if they fail to acquire the lock, thus 751 * there is no danger of deadlocks or of switching while holding a 752 * spin lock if we acquire a lock via a try operation. 753 */ 754 if (flags & LOP_TRYLOCK) 755 return; 756 757 /* 758 * Check for duplicate locks of the same type. Note that we only 759 * have to check for this on the last lock we just acquired. Any 760 * other cases will be caught as lock order violations. 761 */ 762 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 763 w1 = lock1->li_lock->lo_witness; 764 if (w1 == w) { 765 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK)) 766 return; 767 w->w_same_squawked = 1; 768 printf("acquiring duplicate lock of same type: \"%s\"\n", 769 lock->lo_type); 770 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name, 771 lock1->li_file, lock1->li_line); 772 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line); 773 #ifdef KDB 774 goto debugger; 775 #else 776 return; 777 #endif 778 } 779 MPASS(!mtx_owned(&w_mtx)); 780 mtx_lock_spin(&w_mtx); 781 /* 782 * If we have a known higher number just say ok 783 */ 784 if (witness_watch > 1 && w->w_level > w1->w_level) { 785 mtx_unlock_spin(&w_mtx); 786 return; 787 } 788 /* 789 * If we know that the the lock we are acquiring comes after 790 * the lock we most recently acquired in the lock order tree, 791 * then there is no need for any further checks. 792 */ 793 if (isitmydescendant(w1, w)) { 794 mtx_unlock_spin(&w_mtx); 795 return; 796 } 797 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) { 798 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 799 800 MPASS(j < WITNESS_COUNT); 801 lock1 = &lle->ll_children[i]; 802 w1 = lock1->li_lock->lo_witness; 803 804 /* 805 * If this lock doesn't undergo witness checking, 806 * then skip it. 807 */ 808 if (w1 == NULL) { 809 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 810 ("lock missing witness structure")); 811 continue; 812 } 813 /* 814 * If we are locking Giant and this is a sleepable 815 * lock, then skip it. 816 */ 817 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 818 lock == &Giant.mtx_object) 819 continue; 820 /* 821 * If we are locking a sleepable lock and this lock 822 * is Giant, then skip it. 823 */ 824 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 825 lock1->li_lock == &Giant.mtx_object) 826 continue; 827 /* 828 * If we are locking a sleepable lock and this lock 829 * isn't sleepable, we want to treat it as a lock 830 * order violation to enfore a general lock order of 831 * sleepable locks before non-sleepable locks. 832 */ 833 if (!((lock->lo_flags & LO_SLEEPABLE) != 0 && 834 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 835 /* 836 * Check the lock order hierarchy for a reveresal. 837 */ 838 if (!isitmydescendant(w, w1)) 839 continue; 840 /* 841 * We have a lock order violation, check to see if it 842 * is allowed or has already been yelled about. 843 */ 844 mtx_unlock_spin(&w_mtx); 845 #ifdef BLESSING 846 /* 847 * If the lock order is blessed, just bail. We don't 848 * look for other lock order violations though, which 849 * may be a bug. 850 */ 851 if (blessed(w, w1)) 852 return; 853 #endif 854 if (lock1->li_lock == &Giant.mtx_object) { 855 if (w1->w_Giant_squawked) 856 return; 857 else 858 w1->w_Giant_squawked = 1; 859 } else { 860 if (w1->w_other_squawked) 861 return; 862 else 863 w1->w_other_squawked = 1; 864 } 865 /* 866 * Ok, yell about it. 867 */ 868 printf("lock order reversal\n"); 869 /* 870 * Try to locate an earlier lock with 871 * witness w in our list. 872 */ 873 do { 874 lock2 = &lle->ll_children[i]; 875 MPASS(lock2->li_lock != NULL); 876 if (lock2->li_lock->lo_witness == w) 877 break; 878 if (i == 0 && lle->ll_next != NULL) { 879 lle = lle->ll_next; 880 i = lle->ll_count - 1; 881 MPASS(i >= 0 && i < LOCK_NCHILDREN); 882 } else 883 i--; 884 } while (i >= 0); 885 if (i < 0) { 886 printf(" 1st %p %s (%s) @ %s:%d\n", 887 lock1->li_lock, lock1->li_lock->lo_name, 888 lock1->li_lock->lo_type, lock1->li_file, 889 lock1->li_line); 890 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 891 lock->lo_name, lock->lo_type, file, line); 892 } else { 893 printf(" 1st %p %s (%s) @ %s:%d\n", 894 lock2->li_lock, lock2->li_lock->lo_name, 895 lock2->li_lock->lo_type, lock2->li_file, 896 lock2->li_line); 897 printf(" 2nd %p %s (%s) @ %s:%d\n", 898 lock1->li_lock, lock1->li_lock->lo_name, 899 lock1->li_lock->lo_type, lock1->li_file, 900 lock1->li_line); 901 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 902 lock->lo_name, lock->lo_type, file, line); 903 } 904 #ifdef KDB 905 goto debugger; 906 #else 907 return; 908 #endif 909 } 910 } 911 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 912 /* 913 * If requested, build a new lock order. However, don't build a new 914 * relationship between a sleepable lock and Giant if it is in the 915 * wrong direction. The correct lock order is that sleepable locks 916 * always come before Giant. 917 */ 918 if (flags & LOP_NEWORDER && 919 !(lock1->li_lock == &Giant.mtx_object && 920 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 921 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 922 lock->lo_type, lock1->li_lock->lo_type); 923 if (!itismychild(lock1->li_lock->lo_witness, w)) 924 /* Witness is dead. */ 925 return; 926 } 927 mtx_unlock_spin(&w_mtx); 928 return; 929 930 #ifdef KDB 931 debugger: 932 if (witness_trace) 933 kdb_backtrace(); 934 if (witness_kdb) 935 kdb_enter(__func__); 936 #endif 937 } 938 939 void 940 witness_lock(struct lock_object *lock, int flags, const char *file, int line) 941 { 942 struct lock_list_entry **lock_list, *lle; 943 struct lock_instance *instance; 944 struct witness *w; 945 struct thread *td; 946 947 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 948 panicstr != NULL) 949 return; 950 w = lock->lo_witness; 951 td = curthread; 952 file = fixup_filename(file); 953 954 /* Determine lock list for this lock. */ 955 if (lock->lo_class->lc_flags & LC_SLEEPLOCK) 956 lock_list = &td->td_sleeplocks; 957 else 958 lock_list = PCPU_PTR(spinlocks); 959 960 /* Check to see if we are recursing on a lock we already own. */ 961 instance = find_instance(*lock_list, lock); 962 if (instance != NULL) { 963 instance->li_flags++; 964 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 965 td->td_proc->p_pid, lock->lo_name, 966 instance->li_flags & LI_RECURSEMASK); 967 instance->li_file = file; 968 instance->li_line = line; 969 return; 970 } 971 972 /* Update per-witness last file and line acquire. */ 973 w->w_file = file; 974 w->w_line = line; 975 976 /* Find the next open lock instance in the list and fill it. */ 977 lle = *lock_list; 978 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 979 lle = witness_lock_list_get(); 980 if (lle == NULL) 981 return; 982 lle->ll_next = *lock_list; 983 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 984 td->td_proc->p_pid, lle); 985 *lock_list = lle; 986 } 987 instance = &lle->ll_children[lle->ll_count++]; 988 instance->li_lock = lock; 989 instance->li_line = line; 990 instance->li_file = file; 991 if ((flags & LOP_EXCLUSIVE) != 0) 992 instance->li_flags = LI_EXCLUSIVE; 993 else 994 instance->li_flags = 0; 995 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 996 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 997 } 998 999 void 1000 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1001 { 1002 struct lock_instance *instance; 1003 struct lock_class *class; 1004 1005 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1006 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1007 return; 1008 class = lock->lo_class; 1009 file = fixup_filename(file); 1010 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1011 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d", 1012 class->lc_name, lock->lo_name, file, line); 1013 if ((flags & LOP_TRYLOCK) == 0) 1014 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name, 1015 lock->lo_name, file, line); 1016 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1017 panic("upgrade of non-sleep lock (%s) %s @ %s:%d", 1018 class->lc_name, lock->lo_name, file, line); 1019 instance = find_instance(curthread->td_sleeplocks, lock); 1020 if (instance == NULL) 1021 panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1022 class->lc_name, lock->lo_name, file, line); 1023 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1024 panic("upgrade of exclusive lock (%s) %s @ %s:%d", 1025 class->lc_name, lock->lo_name, file, line); 1026 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1027 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1028 class->lc_name, lock->lo_name, 1029 instance->li_flags & LI_RECURSEMASK, file, line); 1030 instance->li_flags |= LI_EXCLUSIVE; 1031 } 1032 1033 void 1034 witness_downgrade(struct lock_object *lock, int flags, const char *file, 1035 int line) 1036 { 1037 struct lock_instance *instance; 1038 struct lock_class *class; 1039 1040 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1041 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1042 return; 1043 class = lock->lo_class; 1044 file = fixup_filename(file); 1045 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1046 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d", 1047 class->lc_name, lock->lo_name, file, line); 1048 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1049 panic("downgrade of non-sleep lock (%s) %s @ %s:%d", 1050 class->lc_name, lock->lo_name, file, line); 1051 instance = find_instance(curthread->td_sleeplocks, lock); 1052 if (instance == NULL) 1053 panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1054 class->lc_name, lock->lo_name, file, line); 1055 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1056 panic("downgrade of shared lock (%s) %s @ %s:%d", 1057 class->lc_name, lock->lo_name, file, line); 1058 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1059 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1060 class->lc_name, lock->lo_name, 1061 instance->li_flags & LI_RECURSEMASK, file, line); 1062 instance->li_flags &= ~LI_EXCLUSIVE; 1063 } 1064 1065 void 1066 witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1067 { 1068 struct lock_list_entry **lock_list, *lle; 1069 struct lock_instance *instance; 1070 struct lock_class *class; 1071 struct thread *td; 1072 register_t s; 1073 int i, j; 1074 1075 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 1076 panicstr != NULL) 1077 return; 1078 td = curthread; 1079 class = lock->lo_class; 1080 file = fixup_filename(file); 1081 1082 /* Find lock instance associated with this lock. */ 1083 if (class->lc_flags & LC_SLEEPLOCK) 1084 lock_list = &td->td_sleeplocks; 1085 else 1086 lock_list = PCPU_PTR(spinlocks); 1087 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1088 for (i = 0; i < (*lock_list)->ll_count; i++) { 1089 instance = &(*lock_list)->ll_children[i]; 1090 if (instance->li_lock == lock) 1091 goto found; 1092 } 1093 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, 1094 file, line); 1095 found: 1096 1097 /* First, check for shared/exclusive mismatches. */ 1098 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && 1099 (flags & LOP_EXCLUSIVE) == 0) { 1100 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1101 lock->lo_name, file, line); 1102 printf("while exclusively locked from %s:%d\n", 1103 instance->li_file, instance->li_line); 1104 panic("excl->ushare"); 1105 } 1106 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && 1107 (flags & LOP_EXCLUSIVE) != 0) { 1108 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1109 lock->lo_name, file, line); 1110 printf("while share locked from %s:%d\n", instance->li_file, 1111 instance->li_line); 1112 panic("share->uexcl"); 1113 } 1114 1115 /* If we are recursed, unrecurse. */ 1116 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1117 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1118 td->td_proc->p_pid, instance->li_lock->lo_name, 1119 instance->li_flags); 1120 instance->li_flags--; 1121 return; 1122 } 1123 1124 /* Otherwise, remove this item from the list. */ 1125 s = intr_disable(); 1126 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1127 td->td_proc->p_pid, instance->li_lock->lo_name, 1128 (*lock_list)->ll_count - 1); 1129 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1130 (*lock_list)->ll_children[j] = 1131 (*lock_list)->ll_children[j + 1]; 1132 (*lock_list)->ll_count--; 1133 intr_restore(s); 1134 1135 /* If this lock list entry is now empty, free it. */ 1136 if ((*lock_list)->ll_count == 0) { 1137 lle = *lock_list; 1138 *lock_list = lle->ll_next; 1139 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1140 td->td_proc->p_pid, lle); 1141 witness_lock_list_free(lle); 1142 } 1143 } 1144 1145 /* 1146 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1147 * exempt Giant and sleepable locks from the checks as well. If any 1148 * non-exempt locks are held, then a supplied message is printed to the 1149 * console along with a list of the offending locks. If indicated in the 1150 * flags then a failure results in a panic as well. 1151 */ 1152 int 1153 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1154 { 1155 struct lock_list_entry *lle; 1156 struct lock_instance *lock1; 1157 struct thread *td; 1158 va_list ap; 1159 int i, n; 1160 1161 if (witness_cold || witness_watch == 0 || panicstr != NULL) 1162 return (0); 1163 n = 0; 1164 td = curthread; 1165 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1166 for (i = lle->ll_count - 1; i >= 0; i--) { 1167 lock1 = &lle->ll_children[i]; 1168 if (lock1->li_lock == lock) 1169 continue; 1170 if (flags & WARN_GIANTOK && 1171 lock1->li_lock == &Giant.mtx_object) 1172 continue; 1173 if (flags & WARN_SLEEPOK && 1174 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1175 continue; 1176 if (n == 0) { 1177 va_start(ap, fmt); 1178 vprintf(fmt, ap); 1179 va_end(ap); 1180 printf(" with the following"); 1181 if (flags & WARN_SLEEPOK) 1182 printf(" non-sleepable"); 1183 printf(" locks held:\n"); 1184 } 1185 n++; 1186 witness_list_lock(lock1); 1187 } 1188 if (PCPU_GET(spinlocks) != NULL) { 1189 /* 1190 * Since we already hold a spinlock preemption is 1191 * already blocked. 1192 */ 1193 if (n == 0) { 1194 va_start(ap, fmt); 1195 vprintf(fmt, ap); 1196 va_end(ap); 1197 printf(" with the following"); 1198 if (flags & WARN_SLEEPOK) 1199 printf(" non-sleepable"); 1200 printf(" locks held:\n"); 1201 } 1202 n += witness_list_locks(PCPU_PTR(spinlocks)); 1203 } 1204 if (flags & WARN_PANIC && n) 1205 panic("witness_warn"); 1206 #ifdef KDB 1207 else if (witness_kdb && n) 1208 kdb_enter(__func__); 1209 else if (witness_trace && n) 1210 kdb_backtrace(); 1211 #endif 1212 return (n); 1213 } 1214 1215 const char * 1216 witness_file(struct lock_object *lock) 1217 { 1218 struct witness *w; 1219 1220 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1221 return ("?"); 1222 w = lock->lo_witness; 1223 return (w->w_file); 1224 } 1225 1226 int 1227 witness_line(struct lock_object *lock) 1228 { 1229 struct witness *w; 1230 1231 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1232 return (0); 1233 w = lock->lo_witness; 1234 return (w->w_line); 1235 } 1236 1237 static struct witness * 1238 enroll(const char *description, struct lock_class *lock_class) 1239 { 1240 struct witness *w; 1241 1242 if (witness_watch == 0 || panicstr != NULL) 1243 return (NULL); 1244 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin) 1245 return (NULL); 1246 mtx_lock_spin(&w_mtx); 1247 STAILQ_FOREACH(w, &w_all, w_list) { 1248 if (w->w_name == description || (w->w_refcount > 0 && 1249 strcmp(description, w->w_name) == 0)) { 1250 w->w_refcount++; 1251 mtx_unlock_spin(&w_mtx); 1252 if (lock_class != w->w_class) 1253 panic( 1254 "lock (%s) %s does not match earlier (%s) lock", 1255 description, lock_class->lc_name, 1256 w->w_class->lc_name); 1257 return (w); 1258 } 1259 } 1260 /* 1261 * This isn't quite right, as witness_cold is still 0 while we 1262 * enroll all the locks initialized before witness_initialize(). 1263 */ 1264 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) { 1265 mtx_unlock_spin(&w_mtx); 1266 panic("spin lock %s not in order list", description); 1267 } 1268 if ((w = witness_get()) == NULL) 1269 return (NULL); 1270 w->w_name = description; 1271 w->w_class = lock_class; 1272 w->w_refcount = 1; 1273 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1274 if (lock_class->lc_flags & LC_SPINLOCK) 1275 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1276 else if (lock_class->lc_flags & LC_SLEEPLOCK) 1277 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1278 else { 1279 mtx_unlock_spin(&w_mtx); 1280 panic("lock class %s is not sleep or spin", 1281 lock_class->lc_name); 1282 } 1283 mtx_unlock_spin(&w_mtx); 1284 return (w); 1285 } 1286 1287 /* Don't let the door bang you on the way out... */ 1288 static int 1289 depart(struct witness *w) 1290 { 1291 struct witness_child_list_entry *wcl, *nwcl; 1292 struct witness_list *list; 1293 struct witness *parent; 1294 1295 MPASS(w->w_refcount == 0); 1296 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1297 list = &w_sleep; 1298 else 1299 list = &w_spin; 1300 /* 1301 * First, we run through the entire tree looking for any 1302 * witnesses that the outgoing witness is a child of. For 1303 * each parent that we find, we reparent all the direct 1304 * children of the outgoing witness to its parent. 1305 */ 1306 STAILQ_FOREACH(parent, list, w_typelist) { 1307 if (!isitmychild(parent, w)) 1308 continue; 1309 removechild(parent, w); 1310 if (!reparentchildren(parent, w)) 1311 return (0); 1312 } 1313 1314 /* 1315 * Now we go through and free up the child list of the 1316 * outgoing witness. 1317 */ 1318 for (wcl = w->w_children; wcl != NULL; wcl = nwcl) { 1319 nwcl = wcl->wcl_next; 1320 witness_child_free(wcl); 1321 } 1322 1323 /* 1324 * Detach from various lists and free. 1325 */ 1326 STAILQ_REMOVE(list, w, witness, w_typelist); 1327 STAILQ_REMOVE(&w_all, w, witness, w_list); 1328 witness_free(w); 1329 1330 /* Finally, fixup the tree. */ 1331 return (rebalancetree(list)); 1332 } 1333 1334 /* 1335 * Prune an entire lock order tree. We look for cases where a lock 1336 * is now both a descendant and a direct child of a given lock. In 1337 * that case, we want to remove the direct child link from the tree. 1338 * 1339 * Returns false if insertchild() fails. 1340 */ 1341 static int 1342 rebalancetree(struct witness_list *list) 1343 { 1344 struct witness *child, *parent; 1345 1346 STAILQ_FOREACH(child, list, w_typelist) { 1347 STAILQ_FOREACH(parent, list, w_typelist) { 1348 if (!isitmychild(parent, child)) 1349 continue; 1350 removechild(parent, child); 1351 if (isitmydescendant(parent, child)) 1352 continue; 1353 if (!insertchild(parent, child)) 1354 return (0); 1355 } 1356 } 1357 witness_levelall(); 1358 return (1); 1359 } 1360 1361 /* 1362 * Add "child" as a direct child of "parent". Returns false if 1363 * we fail due to out of memory. 1364 */ 1365 static int 1366 insertchild(struct witness *parent, struct witness *child) 1367 { 1368 struct witness_child_list_entry **wcl; 1369 1370 MPASS(child != NULL && parent != NULL); 1371 1372 /* 1373 * Insert "child" after "parent" 1374 */ 1375 wcl = &parent->w_children; 1376 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN) 1377 wcl = &(*wcl)->wcl_next; 1378 if (*wcl == NULL) { 1379 *wcl = witness_child_get(); 1380 if (*wcl == NULL) 1381 return (0); 1382 } 1383 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child; 1384 1385 return (1); 1386 } 1387 1388 /* 1389 * Make all the direct descendants of oldparent be direct descendants 1390 * of newparent. 1391 */ 1392 static int 1393 reparentchildren(struct witness *newparent, struct witness *oldparent) 1394 { 1395 struct witness_child_list_entry *wcl; 1396 int i; 1397 1398 /* Avoid making a witness a child of itself. */ 1399 MPASS(!isitmychild(oldparent, newparent)); 1400 1401 for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1402 for (i = 0; i < wcl->wcl_count; i++) 1403 if (!insertchild(newparent, wcl->wcl_children[i])) 1404 return (0); 1405 return (1); 1406 } 1407 1408 static int 1409 itismychild(struct witness *parent, struct witness *child) 1410 { 1411 struct witness_list *list; 1412 1413 MPASS(child != NULL && parent != NULL); 1414 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) != 1415 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))) 1416 panic( 1417 "%s: parent (%s) and child (%s) are not the same lock type", 1418 __func__, parent->w_class->lc_name, 1419 child->w_class->lc_name); 1420 1421 if (!insertchild(parent, child)) 1422 return (0); 1423 1424 if (parent->w_class->lc_flags & LC_SLEEPLOCK) 1425 list = &w_sleep; 1426 else 1427 list = &w_spin; 1428 return (rebalancetree(list)); 1429 } 1430 1431 static void 1432 removechild(struct witness *parent, struct witness *child) 1433 { 1434 struct witness_child_list_entry **wcl, *wcl1; 1435 int i; 1436 1437 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next) 1438 for (i = 0; i < (*wcl)->wcl_count; i++) 1439 if ((*wcl)->wcl_children[i] == child) 1440 goto found; 1441 return; 1442 found: 1443 (*wcl)->wcl_count--; 1444 if ((*wcl)->wcl_count > i) 1445 (*wcl)->wcl_children[i] = 1446 (*wcl)->wcl_children[(*wcl)->wcl_count]; 1447 MPASS((*wcl)->wcl_children[i] != NULL); 1448 if ((*wcl)->wcl_count != 0) 1449 return; 1450 wcl1 = *wcl; 1451 *wcl = wcl1->wcl_next; 1452 witness_child_free(wcl1); 1453 } 1454 1455 static int 1456 isitmychild(struct witness *parent, struct witness *child) 1457 { 1458 struct witness_child_list_entry *wcl; 1459 int i; 1460 1461 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1462 for (i = 0; i < wcl->wcl_count; i++) { 1463 if (wcl->wcl_children[i] == child) 1464 return (1); 1465 } 1466 } 1467 return (0); 1468 } 1469 1470 static int 1471 isitmydescendant(struct witness *parent, struct witness *child) 1472 { 1473 struct witness_child_list_entry *wcl; 1474 int i, j; 1475 1476 if (isitmychild(parent, child)) 1477 return (1); 1478 j = 0; 1479 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1480 MPASS(j < 1000); 1481 for (i = 0; i < wcl->wcl_count; i++) { 1482 if (isitmydescendant(wcl->wcl_children[i], child)) 1483 return (1); 1484 } 1485 j++; 1486 } 1487 return (0); 1488 } 1489 1490 static void 1491 witness_levelall (void) 1492 { 1493 struct witness_list *list; 1494 struct witness *w, *w1; 1495 1496 /* 1497 * First clear all levels. 1498 */ 1499 STAILQ_FOREACH(w, &w_all, w_list) { 1500 w->w_level = 0; 1501 } 1502 1503 /* 1504 * Look for locks with no parent and level all their descendants. 1505 */ 1506 STAILQ_FOREACH(w, &w_all, w_list) { 1507 /* 1508 * This is just an optimization, technically we could get 1509 * away just walking the all list each time. 1510 */ 1511 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1512 list = &w_sleep; 1513 else 1514 list = &w_spin; 1515 STAILQ_FOREACH(w1, list, w_typelist) { 1516 if (isitmychild(w1, w)) 1517 goto skip; 1518 } 1519 witness_leveldescendents(w, 0); 1520 skip: 1521 ; /* silence GCC 3.x */ 1522 } 1523 } 1524 1525 static void 1526 witness_leveldescendents(struct witness *parent, int level) 1527 { 1528 struct witness_child_list_entry *wcl; 1529 int i; 1530 1531 if (parent->w_level < level) 1532 parent->w_level = level; 1533 level++; 1534 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1535 for (i = 0; i < wcl->wcl_count; i++) 1536 witness_leveldescendents(wcl->wcl_children[i], level); 1537 } 1538 1539 static void 1540 witness_displaydescendants(void(*prnt)(const char *fmt, ...), 1541 struct witness *parent, int indent) 1542 { 1543 struct witness_child_list_entry *wcl; 1544 int i, level; 1545 1546 level = parent->w_level; 1547 prnt("%-2d", level); 1548 for (i = 0; i < indent; i++) 1549 prnt(" "); 1550 if (parent->w_refcount > 0) 1551 prnt("%s", parent->w_name); 1552 else 1553 prnt("(dead)"); 1554 if (parent->w_displayed) { 1555 prnt(" -- (already displayed)\n"); 1556 return; 1557 } 1558 parent->w_displayed = 1; 1559 if (parent->w_refcount > 0) { 1560 if (parent->w_file != NULL) 1561 prnt(" -- last acquired @ %s:%d", parent->w_file, 1562 parent->w_line); 1563 } 1564 prnt("\n"); 1565 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1566 for (i = 0; i < wcl->wcl_count; i++) 1567 witness_displaydescendants(prnt, 1568 wcl->wcl_children[i], indent + 1); 1569 } 1570 1571 #ifdef BLESSING 1572 static int 1573 blessed(struct witness *w1, struct witness *w2) 1574 { 1575 int i; 1576 struct witness_blessed *b; 1577 1578 for (i = 0; i < blessed_count; i++) { 1579 b = &blessed_list[i]; 1580 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1581 if (strcmp(w2->w_name, b->b_lock2) == 0) 1582 return (1); 1583 continue; 1584 } 1585 if (strcmp(w1->w_name, b->b_lock2) == 0) 1586 if (strcmp(w2->w_name, b->b_lock1) == 0) 1587 return (1); 1588 } 1589 return (0); 1590 } 1591 #endif 1592 1593 static struct witness * 1594 witness_get(void) 1595 { 1596 struct witness *w; 1597 1598 if (witness_watch == 0) { 1599 mtx_unlock_spin(&w_mtx); 1600 return (NULL); 1601 } 1602 if (STAILQ_EMPTY(&w_free)) { 1603 witness_watch = 0; 1604 mtx_unlock_spin(&w_mtx); 1605 printf("%s: witness exhausted\n", __func__); 1606 return (NULL); 1607 } 1608 w = STAILQ_FIRST(&w_free); 1609 STAILQ_REMOVE_HEAD(&w_free, w_list); 1610 bzero(w, sizeof(*w)); 1611 return (w); 1612 } 1613 1614 static void 1615 witness_free(struct witness *w) 1616 { 1617 1618 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1619 } 1620 1621 static struct witness_child_list_entry * 1622 witness_child_get(void) 1623 { 1624 struct witness_child_list_entry *wcl; 1625 1626 if (witness_watch == 0) { 1627 mtx_unlock_spin(&w_mtx); 1628 return (NULL); 1629 } 1630 wcl = w_child_free; 1631 if (wcl == NULL) { 1632 witness_watch = 0; 1633 mtx_unlock_spin(&w_mtx); 1634 printf("%s: witness exhausted\n", __func__); 1635 return (NULL); 1636 } 1637 w_child_free = wcl->wcl_next; 1638 bzero(wcl, sizeof(*wcl)); 1639 return (wcl); 1640 } 1641 1642 static void 1643 witness_child_free(struct witness_child_list_entry *wcl) 1644 { 1645 1646 wcl->wcl_next = w_child_free; 1647 w_child_free = wcl; 1648 } 1649 1650 static struct lock_list_entry * 1651 witness_lock_list_get(void) 1652 { 1653 struct lock_list_entry *lle; 1654 1655 if (witness_watch == 0) 1656 return (NULL); 1657 mtx_lock_spin(&w_mtx); 1658 lle = w_lock_list_free; 1659 if (lle == NULL) { 1660 witness_watch = 0; 1661 mtx_unlock_spin(&w_mtx); 1662 printf("%s: witness exhausted\n", __func__); 1663 return (NULL); 1664 } 1665 w_lock_list_free = lle->ll_next; 1666 mtx_unlock_spin(&w_mtx); 1667 bzero(lle, sizeof(*lle)); 1668 return (lle); 1669 } 1670 1671 static void 1672 witness_lock_list_free(struct lock_list_entry *lle) 1673 { 1674 1675 mtx_lock_spin(&w_mtx); 1676 lle->ll_next = w_lock_list_free; 1677 w_lock_list_free = lle; 1678 mtx_unlock_spin(&w_mtx); 1679 } 1680 1681 static struct lock_instance * 1682 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock) 1683 { 1684 struct lock_list_entry *lle; 1685 struct lock_instance *instance; 1686 int i; 1687 1688 for (lle = lock_list; lle != NULL; lle = lle->ll_next) 1689 for (i = lle->ll_count - 1; i >= 0; i--) { 1690 instance = &lle->ll_children[i]; 1691 if (instance->li_lock == lock) 1692 return (instance); 1693 } 1694 return (NULL); 1695 } 1696 1697 static void 1698 witness_list_lock(struct lock_instance *instance) 1699 { 1700 struct lock_object *lock; 1701 1702 lock = instance->li_lock; 1703 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 1704 "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name); 1705 if (lock->lo_type != lock->lo_name) 1706 printf(" (%s)", lock->lo_type); 1707 printf(" r = %d (%p) locked @ %s:%d\n", 1708 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file, 1709 instance->li_line); 1710 } 1711 1712 int 1713 witness_list_locks(struct lock_list_entry **lock_list) 1714 { 1715 struct lock_list_entry *lle; 1716 int i, nheld; 1717 1718 nheld = 0; 1719 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 1720 for (i = lle->ll_count - 1; i >= 0; i--) { 1721 witness_list_lock(&lle->ll_children[i]); 1722 nheld++; 1723 } 1724 return (nheld); 1725 } 1726 1727 /* 1728 * This is a bit risky at best. We call this function when we have timed 1729 * out acquiring a spin lock, and we assume that the other CPU is stuck 1730 * with this lock held. So, we go groveling around in the other CPU's 1731 * per-cpu data to try to find the lock instance for this spin lock to 1732 * see when it was last acquired. 1733 */ 1734 void 1735 witness_display_spinlock(struct lock_object *lock, struct thread *owner) 1736 { 1737 struct lock_instance *instance; 1738 struct pcpu *pc; 1739 1740 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 1741 return; 1742 pc = pcpu_find(owner->td_oncpu); 1743 instance = find_instance(pc->pc_spinlocks, lock); 1744 if (instance != NULL) 1745 witness_list_lock(instance); 1746 } 1747 1748 void 1749 witness_save(struct lock_object *lock, const char **filep, int *linep) 1750 { 1751 struct lock_instance *instance; 1752 1753 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1754 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1755 return; 1756 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1757 panic("%s: lock (%s) %s is not a sleep lock", __func__, 1758 lock->lo_class->lc_name, lock->lo_name); 1759 instance = find_instance(curthread->td_sleeplocks, lock); 1760 if (instance == NULL) 1761 panic("%s: lock (%s) %s not locked", __func__, 1762 lock->lo_class->lc_name, lock->lo_name); 1763 *filep = instance->li_file; 1764 *linep = instance->li_line; 1765 } 1766 1767 void 1768 witness_restore(struct lock_object *lock, const char *file, int line) 1769 { 1770 struct lock_instance *instance; 1771 1772 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1773 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1774 return; 1775 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1776 panic("%s: lock (%s) %s is not a sleep lock", __func__, 1777 lock->lo_class->lc_name, lock->lo_name); 1778 instance = find_instance(curthread->td_sleeplocks, lock); 1779 if (instance == NULL) 1780 panic("%s: lock (%s) %s not locked", __func__, 1781 lock->lo_class->lc_name, lock->lo_name); 1782 lock->lo_witness->w_file = file; 1783 lock->lo_witness->w_line = line; 1784 instance->li_file = file; 1785 instance->li_line = line; 1786 } 1787 1788 void 1789 witness_assert(struct lock_object *lock, int flags, const char *file, int line) 1790 { 1791 #ifdef INVARIANT_SUPPORT 1792 struct lock_instance *instance; 1793 1794 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1795 return; 1796 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0) 1797 instance = find_instance(curthread->td_sleeplocks, lock); 1798 else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0) 1799 instance = find_instance(PCPU_GET(spinlocks), lock); 1800 else { 1801 panic("Lock (%s) %s is not sleep or spin!", 1802 lock->lo_class->lc_name, lock->lo_name); 1803 } 1804 file = fixup_filename(file); 1805 switch (flags) { 1806 case LA_UNLOCKED: 1807 if (instance != NULL) 1808 panic("Lock (%s) %s locked @ %s:%d.", 1809 lock->lo_class->lc_name, lock->lo_name, file, line); 1810 break; 1811 case LA_LOCKED: 1812 case LA_LOCKED | LA_RECURSED: 1813 case LA_LOCKED | LA_NOTRECURSED: 1814 case LA_SLOCKED: 1815 case LA_SLOCKED | LA_RECURSED: 1816 case LA_SLOCKED | LA_NOTRECURSED: 1817 case LA_XLOCKED: 1818 case LA_XLOCKED | LA_RECURSED: 1819 case LA_XLOCKED | LA_NOTRECURSED: 1820 if (instance == NULL) { 1821 panic("Lock (%s) %s not locked @ %s:%d.", 1822 lock->lo_class->lc_name, lock->lo_name, file, line); 1823 break; 1824 } 1825 if ((flags & LA_XLOCKED) != 0 && 1826 (instance->li_flags & LI_EXCLUSIVE) == 0) 1827 panic("Lock (%s) %s not exclusively locked @ %s:%d.", 1828 lock->lo_class->lc_name, lock->lo_name, file, line); 1829 if ((flags & LA_SLOCKED) != 0 && 1830 (instance->li_flags & LI_EXCLUSIVE) != 0) 1831 panic("Lock (%s) %s exclusively locked @ %s:%d.", 1832 lock->lo_class->lc_name, lock->lo_name, file, line); 1833 if ((flags & LA_RECURSED) != 0 && 1834 (instance->li_flags & LI_RECURSEMASK) == 0) 1835 panic("Lock (%s) %s not recursed @ %s:%d.", 1836 lock->lo_class->lc_name, lock->lo_name, file, line); 1837 if ((flags & LA_NOTRECURSED) != 0 && 1838 (instance->li_flags & LI_RECURSEMASK) != 0) 1839 panic("Lock (%s) %s recursed @ %s:%d.", 1840 lock->lo_class->lc_name, lock->lo_name, file, line); 1841 break; 1842 default: 1843 panic("Invalid lock assertion at %s:%d.", file, line); 1844 1845 } 1846 #endif /* INVARIANT_SUPPORT */ 1847 } 1848 1849 #ifdef DDB 1850 static void 1851 witness_list(struct thread *td) 1852 { 1853 1854 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1855 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 1856 1857 if (witness_watch == 0) 1858 return; 1859 1860 witness_list_locks(&td->td_sleeplocks); 1861 1862 /* 1863 * We only handle spinlocks if td == curthread. This is somewhat broken 1864 * if td is currently executing on some other CPU and holds spin locks 1865 * as we won't display those locks. If we had a MI way of getting 1866 * the per-cpu data for a given cpu then we could use 1867 * td->td_oncpu to get the list of spinlocks for this thread 1868 * and "fix" this. 1869 * 1870 * That still wouldn't really fix this unless we locked sched_lock 1871 * or stopped the other CPU to make sure it wasn't changing the list 1872 * out from under us. It is probably best to just not try to handle 1873 * threads on other CPU's for now. 1874 */ 1875 if (td == curthread && PCPU_GET(spinlocks) != NULL) 1876 witness_list_locks(PCPU_PTR(spinlocks)); 1877 } 1878 1879 DB_SHOW_COMMAND(locks, db_witness_list) 1880 { 1881 struct thread *td; 1882 pid_t pid; 1883 struct proc *p; 1884 1885 if (have_addr) { 1886 pid = (addr % 16) + ((addr >> 4) % 16) * 10 + 1887 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 + 1888 ((addr >> 16) % 16) * 10000; 1889 /* sx_slock(&allproc_lock); */ 1890 FOREACH_PROC_IN_SYSTEM(p) { 1891 if (p->p_pid == pid) 1892 break; 1893 } 1894 /* sx_sunlock(&allproc_lock); */ 1895 if (p == NULL) { 1896 db_printf("pid %d not found\n", pid); 1897 return; 1898 } 1899 FOREACH_THREAD_IN_PROC(p, td) { 1900 witness_list(td); 1901 } 1902 } else { 1903 td = curthread; 1904 witness_list(td); 1905 } 1906 } 1907 1908 DB_SHOW_COMMAND(witness, db_witness_display) 1909 { 1910 1911 witness_display(db_printf); 1912 } 1913 #endif 1914