1 /*- 2 * Copyright (c) 2008 Isilon Systems, Inc. 3 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com> 4 * Copyright (c) 1998 Berkeley Software Design, Inc. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Berkeley Software Design Inc's name may not be used to endorse or 16 * promote products derived from this software without specific prior 17 * written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 32 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 33 */ 34 35 /* 36 * Implementation of the `witness' lock verifier. Originally implemented for 37 * mutexes in BSD/OS. Extended to handle generic lock objects and lock 38 * classes in FreeBSD. 39 */ 40 41 /* 42 * Main Entry: witness 43 * Pronunciation: 'wit-n&s 44 * Function: noun 45 * Etymology: Middle English witnesse, from Old English witnes knowledge, 46 * testimony, witness, from 2wit 47 * Date: before 12th century 48 * 1 : attestation of a fact or event : TESTIMONY 49 * 2 : one that gives evidence; specifically : one who testifies in 50 * a cause or before a judicial tribunal 51 * 3 : one asked to be present at a transaction so as to be able to 52 * testify to its having taken place 53 * 4 : one who has personal knowledge of something 54 * 5 a : something serving as evidence or proof : SIGN 55 * b : public affirmation by word or example of usually 56 * religious faith or conviction <the heroic witness to divine 57 * life -- Pilot> 58 * 6 capitalized : a member of the Jehovah's Witnesses 59 */ 60 61 /* 62 * Special rules concerning Giant and lock orders: 63 * 64 * 1) Giant must be acquired before any other mutexes. Stated another way, 65 * no other mutex may be held when Giant is acquired. 66 * 67 * 2) Giant must be released when blocking on a sleepable lock. 68 * 69 * This rule is less obvious, but is a result of Giant providing the same 70 * semantics as spl(). Basically, when a thread sleeps, it must release 71 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule 72 * 2). 73 * 74 * 3) Giant may be acquired before or after sleepable locks. 75 * 76 * This rule is also not quite as obvious. Giant may be acquired after 77 * a sleepable lock because it is a non-sleepable lock and non-sleepable 78 * locks may always be acquired while holding a sleepable lock. The second 79 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose 80 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 81 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and 82 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to 83 * execute. Thus, acquiring Giant both before and after a sleepable lock 84 * will not result in a lock order reversal. 85 */ 86 87 #include <sys/cdefs.h> 88 __FBSDID("$FreeBSD$"); 89 90 #include "opt_ddb.h" 91 #include "opt_hwpmc_hooks.h" 92 #include "opt_stack.h" 93 #include "opt_witness.h" 94 95 #include <sys/param.h> 96 #include <sys/bus.h> 97 #include <sys/kdb.h> 98 #include <sys/kernel.h> 99 #include <sys/ktr.h> 100 #include <sys/lock.h> 101 #include <sys/malloc.h> 102 #include <sys/mutex.h> 103 #include <sys/priv.h> 104 #include <sys/proc.h> 105 #include <sys/sbuf.h> 106 #include <sys/sched.h> 107 #include <sys/stack.h> 108 #include <sys/sysctl.h> 109 #include <sys/systm.h> 110 111 #ifdef DDB 112 #include <ddb/ddb.h> 113 #endif 114 115 #include <machine/stdarg.h> 116 117 #if !defined(DDB) && !defined(STACK) 118 #error "DDB or STACK options are required for WITNESS" 119 #endif 120 121 /* Note that these traces do not work with KTR_ALQ. */ 122 #if 0 123 #define KTR_WITNESS KTR_SUBSYS 124 #else 125 #define KTR_WITNESS 0 126 #endif 127 128 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */ 129 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */ 130 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */ 131 132 /* Define this to check for blessed mutexes */ 133 #undef BLESSING 134 135 #define WITNESS_COUNT 1024 136 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 137 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */ 138 #define WITNESS_PENDLIST 768 139 140 /* Allocate 256 KB of stack data space */ 141 #define WITNESS_LO_DATA_COUNT 2048 142 143 /* Prime, gives load factor of ~2 at full load */ 144 #define WITNESS_LO_HASH_SIZE 1021 145 146 /* 147 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads 148 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should 149 * probably be safe for the most part, but it's still a SWAG. 150 */ 151 #define LOCK_NCHILDREN 5 152 #define LOCK_CHILDCOUNT 2048 153 154 #define MAX_W_NAME 64 155 156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT) 157 #define FULLGRAPH_SBUF_SIZE 512 158 159 /* 160 * These flags go in the witness relationship matrix and describe the 161 * relationship between any two struct witness objects. 162 */ 163 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */ 164 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ 165 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ 166 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ 167 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ 168 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) 169 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) 170 #define WITNESS_RELATED_MASK \ 171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) 172 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been 173 * observed. */ 174 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ 175 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ 176 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ 177 178 /* Descendant to ancestor flags */ 179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) 180 181 /* Ancestor to descendant flags */ 182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) 183 184 #define WITNESS_INDEX_ASSERT(i) \ 185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT) 186 187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness"); 188 189 /* 190 * Lock instances. A lock instance is the data associated with a lock while 191 * it is held by witness. For example, a lock instance will hold the 192 * recursion count of a lock. Lock instances are held in lists. Spin locks 193 * are held in a per-cpu list while sleep locks are held in per-thread list. 194 */ 195 struct lock_instance { 196 struct lock_object *li_lock; 197 const char *li_file; 198 int li_line; 199 u_int li_flags; 200 }; 201 202 /* 203 * A simple list type used to build the list of locks held by a thread 204 * or CPU. We can't simply embed the list in struct lock_object since a 205 * lock may be held by more than one thread if it is a shared lock. Locks 206 * are added to the head of the list, so we fill up each list entry from 207 * "the back" logically. To ease some of the arithmetic, we actually fill 208 * in each list entry the normal way (children[0] then children[1], etc.) but 209 * when we traverse the list we read children[count-1] as the first entry 210 * down to children[0] as the final entry. 211 */ 212 struct lock_list_entry { 213 struct lock_list_entry *ll_next; 214 struct lock_instance ll_children[LOCK_NCHILDREN]; 215 u_int ll_count; 216 }; 217 218 /* 219 * The main witness structure. One of these per named lock type in the system 220 * (for example, "vnode interlock"). 221 */ 222 struct witness { 223 char w_name[MAX_W_NAME]; 224 uint32_t w_index; /* Index in the relationship matrix */ 225 struct lock_class *w_class; 226 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 227 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 228 struct witness *w_hash_next; /* Linked list in hash buckets. */ 229 const char *w_file; /* File where last acquired */ 230 uint32_t w_line; /* Line where last acquired */ 231 uint32_t w_refcount; 232 uint16_t w_num_ancestors; /* direct/indirect 233 * ancestor count */ 234 uint16_t w_num_descendants; /* direct/indirect 235 * descendant count */ 236 int16_t w_ddb_level; 237 unsigned w_displayed:1; 238 unsigned w_reversed:1; 239 }; 240 241 STAILQ_HEAD(witness_list, witness); 242 243 /* 244 * The witness hash table. Keys are witness names (const char *), elements are 245 * witness objects (struct witness *). 246 */ 247 struct witness_hash { 248 struct witness *wh_array[WITNESS_HASH_SIZE]; 249 uint32_t wh_size; 250 uint32_t wh_count; 251 }; 252 253 /* 254 * Key type for the lock order data hash table. 255 */ 256 struct witness_lock_order_key { 257 uint16_t from; 258 uint16_t to; 259 }; 260 261 struct witness_lock_order_data { 262 struct stack wlod_stack; 263 struct witness_lock_order_key wlod_key; 264 struct witness_lock_order_data *wlod_next; 265 }; 266 267 /* 268 * The witness lock order data hash table. Keys are witness index tuples 269 * (struct witness_lock_order_key), elements are lock order data objects 270 * (struct witness_lock_order_data). 271 */ 272 struct witness_lock_order_hash { 273 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE]; 274 u_int wloh_size; 275 u_int wloh_count; 276 }; 277 278 #ifdef BLESSING 279 struct witness_blessed { 280 const char *b_lock1; 281 const char *b_lock2; 282 }; 283 #endif 284 285 struct witness_pendhelp { 286 const char *wh_type; 287 struct lock_object *wh_lock; 288 }; 289 290 struct witness_order_list_entry { 291 const char *w_name; 292 struct lock_class *w_class; 293 }; 294 295 /* 296 * Returns 0 if one of the locks is a spin lock and the other is not. 297 * Returns 1 otherwise. 298 */ 299 static __inline int 300 witness_lock_type_equal(struct witness *w1, struct witness *w2) 301 { 302 303 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) == 304 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))); 305 } 306 307 static __inline int 308 witness_lock_order_key_empty(const struct witness_lock_order_key *key) 309 { 310 311 return (key->from == 0 && key->to == 0); 312 } 313 314 static __inline int 315 witness_lock_order_key_equal(const struct witness_lock_order_key *a, 316 const struct witness_lock_order_key *b) 317 { 318 319 return (a->from == b->from && a->to == b->to); 320 } 321 322 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, 323 const char *fname); 324 #ifdef KDB 325 static void _witness_debugger(int cond, const char *msg); 326 #endif 327 static void adopt(struct witness *parent, struct witness *child); 328 #ifdef BLESSING 329 static int blessed(struct witness *, struct witness *); 330 #endif 331 static void depart(struct witness *w); 332 static struct witness *enroll(const char *description, 333 struct lock_class *lock_class); 334 static struct lock_instance *find_instance(struct lock_list_entry *list, 335 const struct lock_object *lock); 336 static int isitmychild(struct witness *parent, struct witness *child); 337 static int isitmydescendant(struct witness *parent, struct witness *child); 338 static void itismychild(struct witness *parent, struct witness *child); 339 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS); 340 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 341 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS); 342 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent); 343 #ifdef DDB 344 static void witness_ddb_compute_levels(void); 345 static void witness_ddb_display(int(*)(const char *fmt, ...)); 346 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...), 347 struct witness *, int indent); 348 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 349 struct witness_list *list); 350 static void witness_ddb_level_descendants(struct witness *parent, int l); 351 static void witness_ddb_list(struct thread *td); 352 #endif 353 static void witness_free(struct witness *m); 354 static struct witness *witness_get(void); 355 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size); 356 static struct witness *witness_hash_get(const char *key); 357 static void witness_hash_put(struct witness *w); 358 static void witness_init_hash_tables(void); 359 static void witness_increment_graph_generation(void); 360 static void witness_lock_list_free(struct lock_list_entry *lle); 361 static struct lock_list_entry *witness_lock_list_get(void); 362 static int witness_lock_order_add(struct witness *parent, 363 struct witness *child); 364 static int witness_lock_order_check(struct witness *parent, 365 struct witness *child); 366 static struct witness_lock_order_data *witness_lock_order_get( 367 struct witness *parent, 368 struct witness *child); 369 static void witness_list_lock(struct lock_instance *instance, 370 int (*prnt)(const char *fmt, ...)); 371 static void witness_setflag(struct lock_object *lock, int flag, int set); 372 373 #ifdef KDB 374 #define witness_debugger(c) _witness_debugger(c, __func__) 375 #else 376 #define witness_debugger(c) 377 #endif 378 379 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, 380 "Witness Locking"); 381 382 /* 383 * If set to 0, lock order checking is disabled. If set to -1, 384 * witness is completely disabled. Otherwise witness performs full 385 * lock order checking for all locks. At runtime, lock order checking 386 * may be toggled. However, witness cannot be reenabled once it is 387 * completely disabled. 388 */ 389 static int witness_watch = 1; 390 TUNABLE_INT("debug.witness.watch", &witness_watch); 391 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 392 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 393 394 #ifdef KDB 395 /* 396 * When KDB is enabled and witness_kdb is 1, it will cause the system 397 * to drop into kdebug() when: 398 * - a lock hierarchy violation occurs 399 * - locks are held when going to sleep. 400 */ 401 #ifdef WITNESS_KDB 402 int witness_kdb = 1; 403 #else 404 int witness_kdb = 0; 405 #endif 406 TUNABLE_INT("debug.witness.kdb", &witness_kdb); 407 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, ""); 408 409 /* 410 * When KDB is enabled and witness_trace is 1, it will cause the system 411 * to print a stack trace: 412 * - a lock hierarchy violation occurs 413 * - locks are held when going to sleep. 414 */ 415 int witness_trace = 1; 416 TUNABLE_INT("debug.witness.trace", &witness_trace); 417 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, ""); 418 #endif /* KDB */ 419 420 #ifdef WITNESS_SKIPSPIN 421 int witness_skipspin = 1; 422 #else 423 int witness_skipspin = 0; 424 #endif 425 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin); 426 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 427 0, ""); 428 429 /* 430 * Call this to print out the relations between locks. 431 */ 432 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD, 433 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs"); 434 435 /* 436 * Call this to print out the witness faulty stacks. 437 */ 438 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD, 439 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks"); 440 441 static struct mtx w_mtx; 442 443 /* w_list */ 444 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 445 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 446 447 /* w_typelist */ 448 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 449 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 450 451 /* lock list */ 452 static struct lock_list_entry *w_lock_list_free = NULL; 453 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; 454 static u_int pending_cnt; 455 456 static int w_free_cnt, w_spin_cnt, w_sleep_cnt; 457 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); 458 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); 459 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, 460 ""); 461 462 static struct witness *w_data; 463 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1]; 464 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 465 static struct witness_hash w_hash; /* The witness hash table. */ 466 467 /* The lock order data hash */ 468 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; 469 static struct witness_lock_order_data *w_lofree = NULL; 470 static struct witness_lock_order_hash w_lohash; 471 static int w_max_used_index = 0; 472 static unsigned int w_generation = 0; 473 static const char w_notrunning[] = "Witness not running\n"; 474 static const char w_stillcold[] = "Witness is still cold\n"; 475 476 477 static struct witness_order_list_entry order_lists[] = { 478 /* 479 * sx locks 480 */ 481 { "proctree", &lock_class_sx }, 482 { "allproc", &lock_class_sx }, 483 { "allprison", &lock_class_sx }, 484 { NULL, NULL }, 485 /* 486 * Various mutexes 487 */ 488 { "Giant", &lock_class_mtx_sleep }, 489 { "pipe mutex", &lock_class_mtx_sleep }, 490 { "sigio lock", &lock_class_mtx_sleep }, 491 { "process group", &lock_class_mtx_sleep }, 492 { "process lock", &lock_class_mtx_sleep }, 493 { "session", &lock_class_mtx_sleep }, 494 { "uidinfo hash", &lock_class_rw }, 495 #ifdef HWPMC_HOOKS 496 { "pmc-sleep", &lock_class_mtx_sleep }, 497 #endif 498 { "time lock", &lock_class_mtx_sleep }, 499 { NULL, NULL }, 500 /* 501 * Sockets 502 */ 503 { "accept", &lock_class_mtx_sleep }, 504 { "so_snd", &lock_class_mtx_sleep }, 505 { "so_rcv", &lock_class_mtx_sleep }, 506 { "sellck", &lock_class_mtx_sleep }, 507 { NULL, NULL }, 508 /* 509 * Routing 510 */ 511 { "so_rcv", &lock_class_mtx_sleep }, 512 { "radix node head", &lock_class_rw }, 513 { "rtentry", &lock_class_mtx_sleep }, 514 { "ifaddr", &lock_class_mtx_sleep }, 515 { NULL, NULL }, 516 /* 517 * IPv4 multicast: 518 * protocol locks before interface locks, after UDP locks. 519 */ 520 { "udpinp", &lock_class_rw }, 521 { "in_multi_mtx", &lock_class_mtx_sleep }, 522 { "igmp_mtx", &lock_class_mtx_sleep }, 523 { "if_addr_lock", &lock_class_rw }, 524 { NULL, NULL }, 525 /* 526 * IPv6 multicast: 527 * protocol locks before interface locks, after UDP locks. 528 */ 529 { "udpinp", &lock_class_rw }, 530 { "in6_multi_mtx", &lock_class_mtx_sleep }, 531 { "mld_mtx", &lock_class_mtx_sleep }, 532 { "if_addr_lock", &lock_class_rw }, 533 { NULL, NULL }, 534 /* 535 * UNIX Domain Sockets 536 */ 537 { "unp_global_rwlock", &lock_class_rw }, 538 { "unp_list_lock", &lock_class_mtx_sleep }, 539 { "unp", &lock_class_mtx_sleep }, 540 { "so_snd", &lock_class_mtx_sleep }, 541 { NULL, NULL }, 542 /* 543 * UDP/IP 544 */ 545 { "udp", &lock_class_rw }, 546 { "udpinp", &lock_class_rw }, 547 { "so_snd", &lock_class_mtx_sleep }, 548 { NULL, NULL }, 549 /* 550 * TCP/IP 551 */ 552 { "tcp", &lock_class_rw }, 553 { "tcpinp", &lock_class_rw }, 554 { "so_snd", &lock_class_mtx_sleep }, 555 { NULL, NULL }, 556 /* 557 * netatalk 558 */ 559 { "ddp_list_mtx", &lock_class_mtx_sleep }, 560 { "ddp_mtx", &lock_class_mtx_sleep }, 561 { NULL, NULL }, 562 /* 563 * BPF 564 */ 565 { "bpf global lock", &lock_class_mtx_sleep }, 566 { "bpf interface lock", &lock_class_rw }, 567 { "bpf cdev lock", &lock_class_mtx_sleep }, 568 { NULL, NULL }, 569 /* 570 * NFS server 571 */ 572 { "nfsd_mtx", &lock_class_mtx_sleep }, 573 { "so_snd", &lock_class_mtx_sleep }, 574 { NULL, NULL }, 575 576 /* 577 * IEEE 802.11 578 */ 579 { "802.11 com lock", &lock_class_mtx_sleep}, 580 { NULL, NULL }, 581 /* 582 * Network drivers 583 */ 584 { "network driver", &lock_class_mtx_sleep}, 585 { NULL, NULL }, 586 587 /* 588 * Netgraph 589 */ 590 { "ng_node", &lock_class_mtx_sleep }, 591 { "ng_worklist", &lock_class_mtx_sleep }, 592 { NULL, NULL }, 593 /* 594 * CDEV 595 */ 596 { "vm map (system)", &lock_class_mtx_sleep }, 597 { "vm page queue", &lock_class_mtx_sleep }, 598 { "vnode interlock", &lock_class_mtx_sleep }, 599 { "cdev", &lock_class_mtx_sleep }, 600 { NULL, NULL }, 601 /* 602 * VM 603 */ 604 { "vm map (user)", &lock_class_sx }, 605 { "vm object", &lock_class_rw }, 606 { "vm page", &lock_class_mtx_sleep }, 607 { "vm page queue", &lock_class_mtx_sleep }, 608 { "pmap pv global", &lock_class_rw }, 609 { "pmap", &lock_class_mtx_sleep }, 610 { "pmap pv list", &lock_class_rw }, 611 { "vm page free queue", &lock_class_mtx_sleep }, 612 { NULL, NULL }, 613 /* 614 * kqueue/VFS interaction 615 */ 616 { "kqueue", &lock_class_mtx_sleep }, 617 { "struct mount mtx", &lock_class_mtx_sleep }, 618 { "vnode interlock", &lock_class_mtx_sleep }, 619 { NULL, NULL }, 620 /* 621 * ZFS locking 622 */ 623 { "dn->dn_mtx", &lock_class_sx }, 624 { "dr->dt.di.dr_mtx", &lock_class_sx }, 625 { "db->db_mtx", &lock_class_sx }, 626 { NULL, NULL }, 627 /* 628 * spin locks 629 */ 630 #ifdef SMP 631 { "ap boot", &lock_class_mtx_spin }, 632 #endif 633 { "rm.mutex_mtx", &lock_class_mtx_spin }, 634 { "sio", &lock_class_mtx_spin }, 635 { "scrlock", &lock_class_mtx_spin }, 636 #ifdef __i386__ 637 { "cy", &lock_class_mtx_spin }, 638 #endif 639 #ifdef __sparc64__ 640 { "pcib_mtx", &lock_class_mtx_spin }, 641 { "rtc_mtx", &lock_class_mtx_spin }, 642 #endif 643 { "scc_hwmtx", &lock_class_mtx_spin }, 644 { "uart_hwmtx", &lock_class_mtx_spin }, 645 { "fast_taskqueue", &lock_class_mtx_spin }, 646 { "intr table", &lock_class_mtx_spin }, 647 #ifdef HWPMC_HOOKS 648 { "pmc-per-proc", &lock_class_mtx_spin }, 649 #endif 650 { "process slock", &lock_class_mtx_spin }, 651 { "sleepq chain", &lock_class_mtx_spin }, 652 { "umtx lock", &lock_class_mtx_spin }, 653 { "rm_spinlock", &lock_class_mtx_spin }, 654 { "turnstile chain", &lock_class_mtx_spin }, 655 { "turnstile lock", &lock_class_mtx_spin }, 656 { "sched lock", &lock_class_mtx_spin }, 657 { "td_contested", &lock_class_mtx_spin }, 658 { "callout", &lock_class_mtx_spin }, 659 { "entropy harvest mutex", &lock_class_mtx_spin }, 660 { "syscons video lock", &lock_class_mtx_spin }, 661 #ifdef SMP 662 { "smp rendezvous", &lock_class_mtx_spin }, 663 #endif 664 #ifdef __powerpc__ 665 { "tlb0", &lock_class_mtx_spin }, 666 #endif 667 /* 668 * leaf locks 669 */ 670 { "intrcnt", &lock_class_mtx_spin }, 671 { "icu", &lock_class_mtx_spin }, 672 #ifdef __i386__ 673 { "allpmaps", &lock_class_mtx_spin }, 674 { "descriptor tables", &lock_class_mtx_spin }, 675 #endif 676 { "clk", &lock_class_mtx_spin }, 677 { "cpuset", &lock_class_mtx_spin }, 678 { "mprof lock", &lock_class_mtx_spin }, 679 { "zombie lock", &lock_class_mtx_spin }, 680 { "ALD Queue", &lock_class_mtx_spin }, 681 #ifdef __ia64__ 682 { "MCA spin lock", &lock_class_mtx_spin }, 683 #endif 684 #if defined(__i386__) || defined(__amd64__) 685 { "pcicfg", &lock_class_mtx_spin }, 686 { "NDIS thread lock", &lock_class_mtx_spin }, 687 #endif 688 { "tw_osl_io_lock", &lock_class_mtx_spin }, 689 { "tw_osl_q_lock", &lock_class_mtx_spin }, 690 { "tw_cl_io_lock", &lock_class_mtx_spin }, 691 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 692 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 693 #ifdef HWPMC_HOOKS 694 { "pmc-leaf", &lock_class_mtx_spin }, 695 #endif 696 { "blocked lock", &lock_class_mtx_spin }, 697 { NULL, NULL }, 698 { NULL, NULL } 699 }; 700 701 #ifdef BLESSING 702 /* 703 * Pairs of locks which have been blessed 704 * Don't complain about order problems with blessed locks 705 */ 706 static struct witness_blessed blessed_list[] = { 707 }; 708 static int blessed_count = 709 sizeof(blessed_list) / sizeof(struct witness_blessed); 710 #endif 711 712 /* 713 * This global is set to 0 once it becomes safe to use the witness code. 714 */ 715 static int witness_cold = 1; 716 717 /* 718 * This global is set to 1 once the static lock orders have been enrolled 719 * so that a warning can be issued for any spin locks enrolled later. 720 */ 721 static int witness_spin_warn = 0; 722 723 /* Trim useless garbage from filenames. */ 724 static const char * 725 fixup_filename(const char *file) 726 { 727 728 if (file == NULL) 729 return (NULL); 730 while (strncmp(file, "../", 3) == 0) 731 file += 3; 732 return (file); 733 } 734 735 /* 736 * The WITNESS-enabled diagnostic code. Note that the witness code does 737 * assume that the early boot is single-threaded at least until after this 738 * routine is completed. 739 */ 740 static void 741 witness_initialize(void *dummy __unused) 742 { 743 struct lock_object *lock; 744 struct witness_order_list_entry *order; 745 struct witness *w, *w1; 746 int i; 747 748 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS, 749 M_NOWAIT | M_ZERO); 750 751 /* 752 * We have to release Giant before initializing its witness 753 * structure so that WITNESS doesn't get confused. 754 */ 755 mtx_unlock(&Giant); 756 mtx_assert(&Giant, MA_NOTOWNED); 757 758 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 759 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 760 MTX_NOWITNESS | MTX_NOPROFILE); 761 for (i = WITNESS_COUNT - 1; i >= 0; i--) { 762 w = &w_data[i]; 763 memset(w, 0, sizeof(*w)); 764 w_data[i].w_index = i; /* Witness index never changes. */ 765 witness_free(w); 766 } 767 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, 768 ("%s: Invalid list of free witness objects", __func__)); 769 770 /* Witness with index 0 is not used to aid in debugging. */ 771 STAILQ_REMOVE_HEAD(&w_free, w_list); 772 w_free_cnt--; 773 774 memset(w_rmatrix, 0, 775 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1))); 776 777 for (i = 0; i < LOCK_CHILDCOUNT; i++) 778 witness_lock_list_free(&w_locklistdata[i]); 779 witness_init_hash_tables(); 780 781 /* First add in all the specified order lists. */ 782 for (order = order_lists; order->w_name != NULL; order++) { 783 w = enroll(order->w_name, order->w_class); 784 if (w == NULL) 785 continue; 786 w->w_file = "order list"; 787 for (order++; order->w_name != NULL; order++) { 788 w1 = enroll(order->w_name, order->w_class); 789 if (w1 == NULL) 790 continue; 791 w1->w_file = "order list"; 792 itismychild(w, w1); 793 w = w1; 794 } 795 } 796 witness_spin_warn = 1; 797 798 /* Iterate through all locks and add them to witness. */ 799 for (i = 0; pending_locks[i].wh_lock != NULL; i++) { 800 lock = pending_locks[i].wh_lock; 801 KASSERT(lock->lo_flags & LO_WITNESS, 802 ("%s: lock %s is on pending list but not LO_WITNESS", 803 __func__, lock->lo_name)); 804 lock->lo_witness = enroll(pending_locks[i].wh_type, 805 LOCK_CLASS(lock)); 806 } 807 808 /* Mark the witness code as being ready for use. */ 809 witness_cold = 0; 810 811 mtx_lock(&Giant); 812 } 813 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, 814 NULL); 815 816 void 817 witness_init(struct lock_object *lock, const char *type) 818 { 819 struct lock_class *class; 820 821 /* Various sanity checks. */ 822 class = LOCK_CLASS(lock); 823 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 824 (class->lc_flags & LC_RECURSABLE) == 0) 825 kassert_panic("%s: lock (%s) %s can not be recursable", 826 __func__, class->lc_name, lock->lo_name); 827 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 828 (class->lc_flags & LC_SLEEPABLE) == 0) 829 kassert_panic("%s: lock (%s) %s can not be sleepable", 830 __func__, class->lc_name, lock->lo_name); 831 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 832 (class->lc_flags & LC_UPGRADABLE) == 0) 833 kassert_panic("%s: lock (%s) %s can not be upgradable", 834 __func__, class->lc_name, lock->lo_name); 835 836 /* 837 * If we shouldn't watch this lock, then just clear lo_witness. 838 * Otherwise, if witness_cold is set, then it is too early to 839 * enroll this lock, so defer it to witness_initialize() by adding 840 * it to the pending_locks list. If it is not too early, then enroll 841 * the lock now. 842 */ 843 if (witness_watch < 1 || panicstr != NULL || 844 (lock->lo_flags & LO_WITNESS) == 0) 845 lock->lo_witness = NULL; 846 else if (witness_cold) { 847 pending_locks[pending_cnt].wh_lock = lock; 848 pending_locks[pending_cnt++].wh_type = type; 849 if (pending_cnt > WITNESS_PENDLIST) 850 panic("%s: pending locks list is too small, " 851 "increase WITNESS_PENDLIST\n", 852 __func__); 853 } else 854 lock->lo_witness = enroll(type, class); 855 } 856 857 void 858 witness_destroy(struct lock_object *lock) 859 { 860 struct lock_class *class; 861 struct witness *w; 862 863 class = LOCK_CLASS(lock); 864 865 if (witness_cold) 866 panic("lock (%s) %s destroyed while witness_cold", 867 class->lc_name, lock->lo_name); 868 869 /* XXX: need to verify that no one holds the lock */ 870 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) 871 return; 872 w = lock->lo_witness; 873 874 mtx_lock_spin(&w_mtx); 875 MPASS(w->w_refcount > 0); 876 w->w_refcount--; 877 878 if (w->w_refcount == 0) 879 depart(w); 880 mtx_unlock_spin(&w_mtx); 881 } 882 883 #ifdef DDB 884 static void 885 witness_ddb_compute_levels(void) 886 { 887 struct witness *w; 888 889 /* 890 * First clear all levels. 891 */ 892 STAILQ_FOREACH(w, &w_all, w_list) 893 w->w_ddb_level = -1; 894 895 /* 896 * Look for locks with no parents and level all their descendants. 897 */ 898 STAILQ_FOREACH(w, &w_all, w_list) { 899 900 /* If the witness has ancestors (is not a root), skip it. */ 901 if (w->w_num_ancestors > 0) 902 continue; 903 witness_ddb_level_descendants(w, 0); 904 } 905 } 906 907 static void 908 witness_ddb_level_descendants(struct witness *w, int l) 909 { 910 int i; 911 912 if (w->w_ddb_level >= l) 913 return; 914 915 w->w_ddb_level = l; 916 l++; 917 918 for (i = 1; i <= w_max_used_index; i++) { 919 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 920 witness_ddb_level_descendants(&w_data[i], l); 921 } 922 } 923 924 static void 925 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...), 926 struct witness *w, int indent) 927 { 928 int i; 929 930 for (i = 0; i < indent; i++) 931 prnt(" "); 932 prnt("%s (type: %s, depth: %d, active refs: %d)", 933 w->w_name, w->w_class->lc_name, 934 w->w_ddb_level, w->w_refcount); 935 if (w->w_displayed) { 936 prnt(" -- (already displayed)\n"); 937 return; 938 } 939 w->w_displayed = 1; 940 if (w->w_file != NULL && w->w_line != 0) 941 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file), 942 w->w_line); 943 else 944 prnt(" -- never acquired\n"); 945 indent++; 946 WITNESS_INDEX_ASSERT(w->w_index); 947 for (i = 1; i <= w_max_used_index; i++) { 948 if (db_pager_quit) 949 return; 950 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 951 witness_ddb_display_descendants(prnt, &w_data[i], 952 indent); 953 } 954 } 955 956 static void 957 witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 958 struct witness_list *list) 959 { 960 struct witness *w; 961 962 STAILQ_FOREACH(w, list, w_typelist) { 963 if (w->w_file == NULL || w->w_ddb_level > 0) 964 continue; 965 966 /* This lock has no anscestors - display its descendants. */ 967 witness_ddb_display_descendants(prnt, w, 0); 968 if (db_pager_quit) 969 return; 970 } 971 } 972 973 static void 974 witness_ddb_display(int(*prnt)(const char *fmt, ...)) 975 { 976 struct witness *w; 977 978 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 979 witness_ddb_compute_levels(); 980 981 /* Clear all the displayed flags. */ 982 STAILQ_FOREACH(w, &w_all, w_list) 983 w->w_displayed = 0; 984 985 /* 986 * First, handle sleep locks which have been acquired at least 987 * once. 988 */ 989 prnt("Sleep locks:\n"); 990 witness_ddb_display_list(prnt, &w_sleep); 991 if (db_pager_quit) 992 return; 993 994 /* 995 * Now do spin locks which have been acquired at least once. 996 */ 997 prnt("\nSpin locks:\n"); 998 witness_ddb_display_list(prnt, &w_spin); 999 if (db_pager_quit) 1000 return; 1001 1002 /* 1003 * Finally, any locks which have not been acquired yet. 1004 */ 1005 prnt("\nLocks which were never acquired:\n"); 1006 STAILQ_FOREACH(w, &w_all, w_list) { 1007 if (w->w_file != NULL || w->w_refcount == 0) 1008 continue; 1009 prnt("%s (type: %s, depth: %d)\n", w->w_name, 1010 w->w_class->lc_name, w->w_ddb_level); 1011 if (db_pager_quit) 1012 return; 1013 } 1014 } 1015 #endif /* DDB */ 1016 1017 int 1018 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 1019 { 1020 1021 if (witness_watch == -1 || panicstr != NULL) 1022 return (0); 1023 1024 /* Require locks that witness knows about. */ 1025 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 1026 lock2->lo_witness == NULL) 1027 return (EINVAL); 1028 1029 mtx_assert(&w_mtx, MA_NOTOWNED); 1030 mtx_lock_spin(&w_mtx); 1031 1032 /* 1033 * If we already have either an explicit or implied lock order that 1034 * is the other way around, then return an error. 1035 */ 1036 if (witness_watch && 1037 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 1038 mtx_unlock_spin(&w_mtx); 1039 return (EDOOFUS); 1040 } 1041 1042 /* Try to add the new order. */ 1043 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1044 lock2->lo_witness->w_name, lock1->lo_witness->w_name); 1045 itismychild(lock1->lo_witness, lock2->lo_witness); 1046 mtx_unlock_spin(&w_mtx); 1047 return (0); 1048 } 1049 1050 void 1051 witness_checkorder(struct lock_object *lock, int flags, const char *file, 1052 int line, struct lock_object *interlock) 1053 { 1054 struct lock_list_entry *lock_list, *lle; 1055 struct lock_instance *lock1, *lock2, *plock; 1056 struct lock_class *class, *iclass; 1057 struct witness *w, *w1; 1058 struct thread *td; 1059 int i, j; 1060 1061 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || 1062 panicstr != NULL) 1063 return; 1064 1065 w = lock->lo_witness; 1066 class = LOCK_CLASS(lock); 1067 td = curthread; 1068 1069 if (class->lc_flags & LC_SLEEPLOCK) { 1070 1071 /* 1072 * Since spin locks include a critical section, this check 1073 * implicitly enforces a lock order of all sleep locks before 1074 * all spin locks. 1075 */ 1076 if (td->td_critnest != 0 && !kdb_active) 1077 kassert_panic("acquiring blockable sleep lock with " 1078 "spinlock or critical section held (%s) %s @ %s:%d", 1079 class->lc_name, lock->lo_name, 1080 fixup_filename(file), line); 1081 1082 /* 1083 * If this is the first lock acquired then just return as 1084 * no order checking is needed. 1085 */ 1086 lock_list = td->td_sleeplocks; 1087 if (lock_list == NULL || lock_list->ll_count == 0) 1088 return; 1089 } else { 1090 1091 /* 1092 * If this is the first lock, just return as no order 1093 * checking is needed. Avoid problems with thread 1094 * migration pinning the thread while checking if 1095 * spinlocks are held. If at least one spinlock is held 1096 * the thread is in a safe path and it is allowed to 1097 * unpin it. 1098 */ 1099 sched_pin(); 1100 lock_list = PCPU_GET(spinlocks); 1101 if (lock_list == NULL || lock_list->ll_count == 0) { 1102 sched_unpin(); 1103 return; 1104 } 1105 sched_unpin(); 1106 } 1107 1108 /* 1109 * Check to see if we are recursing on a lock we already own. If 1110 * so, make sure that we don't mismatch exclusive and shared lock 1111 * acquires. 1112 */ 1113 lock1 = find_instance(lock_list, lock); 1114 if (lock1 != NULL) { 1115 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 1116 (flags & LOP_EXCLUSIVE) == 0) { 1117 printf("shared lock of (%s) %s @ %s:%d\n", 1118 class->lc_name, lock->lo_name, 1119 fixup_filename(file), line); 1120 printf("while exclusively locked from %s:%d\n", 1121 fixup_filename(lock1->li_file), lock1->li_line); 1122 kassert_panic("excl->share"); 1123 } 1124 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 1125 (flags & LOP_EXCLUSIVE) != 0) { 1126 printf("exclusive lock of (%s) %s @ %s:%d\n", 1127 class->lc_name, lock->lo_name, 1128 fixup_filename(file), line); 1129 printf("while share locked from %s:%d\n", 1130 fixup_filename(lock1->li_file), lock1->li_line); 1131 kassert_panic("share->excl"); 1132 } 1133 return; 1134 } 1135 1136 /* Warn if the interlock is not locked exactly once. */ 1137 if (interlock != NULL) { 1138 iclass = LOCK_CLASS(interlock); 1139 lock1 = find_instance(lock_list, interlock); 1140 if (lock1 == NULL) 1141 kassert_panic("interlock (%s) %s not locked @ %s:%d", 1142 iclass->lc_name, interlock->lo_name, 1143 flags & LOP_EXCLUSIVE ? "exclusive" : "shared", 1144 fixup_filename(file), line); 1145 else if ((lock1->li_flags & LI_RECURSEMASK) != 0) 1146 kassert_panic("interlock (%s) %s recursed @ %s:%d", 1147 iclass->lc_name, interlock->lo_name, 1148 flags & LOP_EXCLUSIVE ? "exclusive" : "shared", 1149 fixup_filename(file), line); 1150 } 1151 1152 /* 1153 * Find the previously acquired lock, but ignore interlocks. 1154 */ 1155 plock = &lock_list->ll_children[lock_list->ll_count - 1]; 1156 if (interlock != NULL && plock->li_lock == interlock) { 1157 if (lock_list->ll_count > 1) 1158 plock = 1159 &lock_list->ll_children[lock_list->ll_count - 2]; 1160 else { 1161 lle = lock_list->ll_next; 1162 1163 /* 1164 * The interlock is the only lock we hold, so 1165 * simply return. 1166 */ 1167 if (lle == NULL) 1168 return; 1169 plock = &lle->ll_children[lle->ll_count - 1]; 1170 } 1171 } 1172 1173 /* 1174 * Try to perform most checks without a lock. If this succeeds we 1175 * can skip acquiring the lock and return success. 1176 */ 1177 w1 = plock->li_lock->lo_witness; 1178 if (witness_lock_order_check(w1, w)) 1179 return; 1180 1181 /* 1182 * Check for duplicate locks of the same type. Note that we only 1183 * have to check for this on the last lock we just acquired. Any 1184 * other cases will be caught as lock order violations. 1185 */ 1186 mtx_lock_spin(&w_mtx); 1187 witness_lock_order_add(w1, w); 1188 if (w1 == w) { 1189 i = w->w_index; 1190 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && 1191 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { 1192 w_rmatrix[i][i] |= WITNESS_REVERSAL; 1193 w->w_reversed = 1; 1194 mtx_unlock_spin(&w_mtx); 1195 printf( 1196 "acquiring duplicate lock of same type: \"%s\"\n", 1197 w->w_name); 1198 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, 1199 fixup_filename(plock->li_file), plock->li_line); 1200 printf(" 2nd %s @ %s:%d\n", lock->lo_name, 1201 fixup_filename(file), line); 1202 witness_debugger(1); 1203 } else 1204 mtx_unlock_spin(&w_mtx); 1205 return; 1206 } 1207 mtx_assert(&w_mtx, MA_OWNED); 1208 1209 /* 1210 * If we know that the lock we are acquiring comes after 1211 * the lock we most recently acquired in the lock order tree, 1212 * then there is no need for any further checks. 1213 */ 1214 if (isitmychild(w1, w)) 1215 goto out; 1216 1217 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { 1218 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 1219 1220 MPASS(j < WITNESS_COUNT); 1221 lock1 = &lle->ll_children[i]; 1222 1223 /* 1224 * Ignore the interlock. 1225 */ 1226 if (interlock == lock1->li_lock) 1227 continue; 1228 1229 /* 1230 * If this lock doesn't undergo witness checking, 1231 * then skip it. 1232 */ 1233 w1 = lock1->li_lock->lo_witness; 1234 if (w1 == NULL) { 1235 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 1236 ("lock missing witness structure")); 1237 continue; 1238 } 1239 1240 /* 1241 * If we are locking Giant and this is a sleepable 1242 * lock, then skip it. 1243 */ 1244 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 1245 lock == &Giant.lock_object) 1246 continue; 1247 1248 /* 1249 * If we are locking a sleepable lock and this lock 1250 * is Giant, then skip it. 1251 */ 1252 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1253 lock1->li_lock == &Giant.lock_object) 1254 continue; 1255 1256 /* 1257 * If we are locking a sleepable lock and this lock 1258 * isn't sleepable, we want to treat it as a lock 1259 * order violation to enfore a general lock order of 1260 * sleepable locks before non-sleepable locks. 1261 */ 1262 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1263 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1264 goto reversal; 1265 1266 /* 1267 * If we are locking Giant and this is a non-sleepable 1268 * lock, then treat it as a reversal. 1269 */ 1270 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && 1271 lock == &Giant.lock_object) 1272 goto reversal; 1273 1274 /* 1275 * Check the lock order hierarchy for a reveresal. 1276 */ 1277 if (!isitmydescendant(w, w1)) 1278 continue; 1279 reversal: 1280 1281 /* 1282 * We have a lock order violation, check to see if it 1283 * is allowed or has already been yelled about. 1284 */ 1285 #ifdef BLESSING 1286 1287 /* 1288 * If the lock order is blessed, just bail. We don't 1289 * look for other lock order violations though, which 1290 * may be a bug. 1291 */ 1292 if (blessed(w, w1)) 1293 goto out; 1294 #endif 1295 1296 /* Bail if this violation is known */ 1297 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) 1298 goto out; 1299 1300 /* Record this as a violation */ 1301 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; 1302 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; 1303 w->w_reversed = w1->w_reversed = 1; 1304 witness_increment_graph_generation(); 1305 mtx_unlock_spin(&w_mtx); 1306 1307 #ifdef WITNESS_NO_VNODE 1308 /* 1309 * There are known LORs between VNODE locks. They are 1310 * not an indication of a bug. VNODE locks are flagged 1311 * as such (LO_IS_VNODE) and we don't yell if the LOR 1312 * is between 2 VNODE locks. 1313 */ 1314 if ((lock->lo_flags & LO_IS_VNODE) != 0 && 1315 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0) 1316 return; 1317 #endif 1318 1319 /* 1320 * Ok, yell about it. 1321 */ 1322 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1323 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1324 printf( 1325 "lock order reversal: (sleepable after non-sleepable)\n"); 1326 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 1327 && lock == &Giant.lock_object) 1328 printf( 1329 "lock order reversal: (Giant after non-sleepable)\n"); 1330 else 1331 printf("lock order reversal:\n"); 1332 1333 /* 1334 * Try to locate an earlier lock with 1335 * witness w in our list. 1336 */ 1337 do { 1338 lock2 = &lle->ll_children[i]; 1339 MPASS(lock2->li_lock != NULL); 1340 if (lock2->li_lock->lo_witness == w) 1341 break; 1342 if (i == 0 && lle->ll_next != NULL) { 1343 lle = lle->ll_next; 1344 i = lle->ll_count - 1; 1345 MPASS(i >= 0 && i < LOCK_NCHILDREN); 1346 } else 1347 i--; 1348 } while (i >= 0); 1349 if (i < 0) { 1350 printf(" 1st %p %s (%s) @ %s:%d\n", 1351 lock1->li_lock, lock1->li_lock->lo_name, 1352 w1->w_name, fixup_filename(lock1->li_file), 1353 lock1->li_line); 1354 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 1355 lock->lo_name, w->w_name, 1356 fixup_filename(file), line); 1357 } else { 1358 printf(" 1st %p %s (%s) @ %s:%d\n", 1359 lock2->li_lock, lock2->li_lock->lo_name, 1360 lock2->li_lock->lo_witness->w_name, 1361 fixup_filename(lock2->li_file), 1362 lock2->li_line); 1363 printf(" 2nd %p %s (%s) @ %s:%d\n", 1364 lock1->li_lock, lock1->li_lock->lo_name, 1365 w1->w_name, fixup_filename(lock1->li_file), 1366 lock1->li_line); 1367 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 1368 lock->lo_name, w->w_name, 1369 fixup_filename(file), line); 1370 } 1371 witness_debugger(1); 1372 return; 1373 } 1374 } 1375 1376 /* 1377 * If requested, build a new lock order. However, don't build a new 1378 * relationship between a sleepable lock and Giant if it is in the 1379 * wrong direction. The correct lock order is that sleepable locks 1380 * always come before Giant. 1381 */ 1382 if (flags & LOP_NEWORDER && 1383 !(plock->li_lock == &Giant.lock_object && 1384 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 1385 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1386 w->w_name, plock->li_lock->lo_witness->w_name); 1387 itismychild(plock->li_lock->lo_witness, w); 1388 } 1389 out: 1390 mtx_unlock_spin(&w_mtx); 1391 } 1392 1393 void 1394 witness_lock(struct lock_object *lock, int flags, const char *file, int line) 1395 { 1396 struct lock_list_entry **lock_list, *lle; 1397 struct lock_instance *instance; 1398 struct witness *w; 1399 struct thread *td; 1400 1401 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || 1402 panicstr != NULL) 1403 return; 1404 w = lock->lo_witness; 1405 td = curthread; 1406 1407 /* Determine lock list for this lock. */ 1408 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) 1409 lock_list = &td->td_sleeplocks; 1410 else 1411 lock_list = PCPU_PTR(spinlocks); 1412 1413 /* Check to see if we are recursing on a lock we already own. */ 1414 instance = find_instance(*lock_list, lock); 1415 if (instance != NULL) { 1416 instance->li_flags++; 1417 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1418 td->td_proc->p_pid, lock->lo_name, 1419 instance->li_flags & LI_RECURSEMASK); 1420 instance->li_file = file; 1421 instance->li_line = line; 1422 return; 1423 } 1424 1425 /* Update per-witness last file and line acquire. */ 1426 w->w_file = file; 1427 w->w_line = line; 1428 1429 /* Find the next open lock instance in the list and fill it. */ 1430 lle = *lock_list; 1431 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1432 lle = witness_lock_list_get(); 1433 if (lle == NULL) 1434 return; 1435 lle->ll_next = *lock_list; 1436 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1437 td->td_proc->p_pid, lle); 1438 *lock_list = lle; 1439 } 1440 instance = &lle->ll_children[lle->ll_count++]; 1441 instance->li_lock = lock; 1442 instance->li_line = line; 1443 instance->li_file = file; 1444 if ((flags & LOP_EXCLUSIVE) != 0) 1445 instance->li_flags = LI_EXCLUSIVE; 1446 else 1447 instance->li_flags = 0; 1448 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1449 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1450 } 1451 1452 void 1453 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1454 { 1455 struct lock_instance *instance; 1456 struct lock_class *class; 1457 1458 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1459 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1460 return; 1461 class = LOCK_CLASS(lock); 1462 if (witness_watch) { 1463 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1464 kassert_panic( 1465 "upgrade of non-upgradable lock (%s) %s @ %s:%d", 1466 class->lc_name, lock->lo_name, 1467 fixup_filename(file), line); 1468 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1469 kassert_panic( 1470 "upgrade of non-sleep lock (%s) %s @ %s:%d", 1471 class->lc_name, lock->lo_name, 1472 fixup_filename(file), line); 1473 } 1474 instance = find_instance(curthread->td_sleeplocks, lock); 1475 if (instance == NULL) { 1476 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1477 class->lc_name, lock->lo_name, 1478 fixup_filename(file), line); 1479 return; 1480 } 1481 if (witness_watch) { 1482 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1483 kassert_panic( 1484 "upgrade of exclusive lock (%s) %s @ %s:%d", 1485 class->lc_name, lock->lo_name, 1486 fixup_filename(file), line); 1487 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1488 kassert_panic( 1489 "upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1490 class->lc_name, lock->lo_name, 1491 instance->li_flags & LI_RECURSEMASK, 1492 fixup_filename(file), line); 1493 } 1494 instance->li_flags |= LI_EXCLUSIVE; 1495 } 1496 1497 void 1498 witness_downgrade(struct lock_object *lock, int flags, const char *file, 1499 int line) 1500 { 1501 struct lock_instance *instance; 1502 struct lock_class *class; 1503 1504 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1505 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1506 return; 1507 class = LOCK_CLASS(lock); 1508 if (witness_watch) { 1509 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1510 kassert_panic( 1511 "downgrade of non-upgradable lock (%s) %s @ %s:%d", 1512 class->lc_name, lock->lo_name, 1513 fixup_filename(file), line); 1514 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1515 kassert_panic( 1516 "downgrade of non-sleep lock (%s) %s @ %s:%d", 1517 class->lc_name, lock->lo_name, 1518 fixup_filename(file), line); 1519 } 1520 instance = find_instance(curthread->td_sleeplocks, lock); 1521 if (instance == NULL) { 1522 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1523 class->lc_name, lock->lo_name, 1524 fixup_filename(file), line); 1525 return; 1526 } 1527 if (witness_watch) { 1528 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1529 kassert_panic( 1530 "downgrade of shared lock (%s) %s @ %s:%d", 1531 class->lc_name, lock->lo_name, 1532 fixup_filename(file), line); 1533 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1534 kassert_panic( 1535 "downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1536 class->lc_name, lock->lo_name, 1537 instance->li_flags & LI_RECURSEMASK, 1538 fixup_filename(file), line); 1539 } 1540 instance->li_flags &= ~LI_EXCLUSIVE; 1541 } 1542 1543 void 1544 witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1545 { 1546 struct lock_list_entry **lock_list, *lle; 1547 struct lock_instance *instance; 1548 struct lock_class *class; 1549 struct thread *td; 1550 register_t s; 1551 int i, j; 1552 1553 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL) 1554 return; 1555 td = curthread; 1556 class = LOCK_CLASS(lock); 1557 1558 /* Find lock instance associated with this lock. */ 1559 if (class->lc_flags & LC_SLEEPLOCK) 1560 lock_list = &td->td_sleeplocks; 1561 else 1562 lock_list = PCPU_PTR(spinlocks); 1563 lle = *lock_list; 1564 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1565 for (i = 0; i < (*lock_list)->ll_count; i++) { 1566 instance = &(*lock_list)->ll_children[i]; 1567 if (instance->li_lock == lock) 1568 goto found; 1569 } 1570 1571 /* 1572 * When disabling WITNESS through witness_watch we could end up in 1573 * having registered locks in the td_sleeplocks queue. 1574 * We have to make sure we flush these queues, so just search for 1575 * eventual register locks and remove them. 1576 */ 1577 if (witness_watch > 0) { 1578 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name, 1579 lock->lo_name, fixup_filename(file), line); 1580 return; 1581 } else { 1582 return; 1583 } 1584 found: 1585 1586 /* First, check for shared/exclusive mismatches. */ 1587 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && 1588 (flags & LOP_EXCLUSIVE) == 0) { 1589 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1590 lock->lo_name, fixup_filename(file), line); 1591 printf("while exclusively locked from %s:%d\n", 1592 fixup_filename(instance->li_file), instance->li_line); 1593 kassert_panic("excl->ushare"); 1594 } 1595 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && 1596 (flags & LOP_EXCLUSIVE) != 0) { 1597 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1598 lock->lo_name, fixup_filename(file), line); 1599 printf("while share locked from %s:%d\n", 1600 fixup_filename(instance->li_file), 1601 instance->li_line); 1602 kassert_panic("share->uexcl"); 1603 } 1604 /* If we are recursed, unrecurse. */ 1605 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1606 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1607 td->td_proc->p_pid, instance->li_lock->lo_name, 1608 instance->li_flags); 1609 instance->li_flags--; 1610 return; 1611 } 1612 /* The lock is now being dropped, check for NORELEASE flag */ 1613 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { 1614 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name, 1615 lock->lo_name, fixup_filename(file), line); 1616 kassert_panic("lock marked norelease"); 1617 } 1618 1619 /* Otherwise, remove this item from the list. */ 1620 s = intr_disable(); 1621 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1622 td->td_proc->p_pid, instance->li_lock->lo_name, 1623 (*lock_list)->ll_count - 1); 1624 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1625 (*lock_list)->ll_children[j] = 1626 (*lock_list)->ll_children[j + 1]; 1627 (*lock_list)->ll_count--; 1628 intr_restore(s); 1629 1630 /* 1631 * In order to reduce contention on w_mtx, we want to keep always an 1632 * head object into lists so that frequent allocation from the 1633 * free witness pool (and subsequent locking) is avoided. 1634 * In order to maintain the current code simple, when the head 1635 * object is totally unloaded it means also that we do not have 1636 * further objects in the list, so the list ownership needs to be 1637 * hand over to another object if the current head needs to be freed. 1638 */ 1639 if ((*lock_list)->ll_count == 0) { 1640 if (*lock_list == lle) { 1641 if (lle->ll_next == NULL) 1642 return; 1643 } else 1644 lle = *lock_list; 1645 *lock_list = lle->ll_next; 1646 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1647 td->td_proc->p_pid, lle); 1648 witness_lock_list_free(lle); 1649 } 1650 } 1651 1652 void 1653 witness_thread_exit(struct thread *td) 1654 { 1655 struct lock_list_entry *lle; 1656 int i, n; 1657 1658 lle = td->td_sleeplocks; 1659 if (lle == NULL || panicstr != NULL) 1660 return; 1661 if (lle->ll_count != 0) { 1662 for (n = 0; lle != NULL; lle = lle->ll_next) 1663 for (i = lle->ll_count - 1; i >= 0; i--) { 1664 if (n == 0) 1665 printf("Thread %p exiting with the following locks held:\n", 1666 td); 1667 n++; 1668 witness_list_lock(&lle->ll_children[i], printf); 1669 1670 } 1671 kassert_panic( 1672 "Thread %p cannot exit while holding sleeplocks\n", td); 1673 } 1674 witness_lock_list_free(lle); 1675 } 1676 1677 /* 1678 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1679 * exempt Giant and sleepable locks from the checks as well. If any 1680 * non-exempt locks are held, then a supplied message is printed to the 1681 * console along with a list of the offending locks. If indicated in the 1682 * flags then a failure results in a panic as well. 1683 */ 1684 int 1685 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1686 { 1687 struct lock_list_entry *lock_list, *lle; 1688 struct lock_instance *lock1; 1689 struct thread *td; 1690 va_list ap; 1691 int i, n; 1692 1693 if (witness_cold || witness_watch < 1 || panicstr != NULL) 1694 return (0); 1695 n = 0; 1696 td = curthread; 1697 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1698 for (i = lle->ll_count - 1; i >= 0; i--) { 1699 lock1 = &lle->ll_children[i]; 1700 if (lock1->li_lock == lock) 1701 continue; 1702 if (flags & WARN_GIANTOK && 1703 lock1->li_lock == &Giant.lock_object) 1704 continue; 1705 if (flags & WARN_SLEEPOK && 1706 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1707 continue; 1708 if (n == 0) { 1709 va_start(ap, fmt); 1710 vprintf(fmt, ap); 1711 va_end(ap); 1712 printf(" with the following"); 1713 if (flags & WARN_SLEEPOK) 1714 printf(" non-sleepable"); 1715 printf(" locks held:\n"); 1716 } 1717 n++; 1718 witness_list_lock(lock1, printf); 1719 } 1720 1721 /* 1722 * Pin the thread in order to avoid problems with thread migration. 1723 * Once that all verifies are passed about spinlocks ownership, 1724 * the thread is in a safe path and it can be unpinned. 1725 */ 1726 sched_pin(); 1727 lock_list = PCPU_GET(spinlocks); 1728 if (lock_list != NULL && lock_list->ll_count != 0) { 1729 sched_unpin(); 1730 1731 /* 1732 * We should only have one spinlock and as long as 1733 * the flags cannot match for this locks class, 1734 * check if the first spinlock is the one curthread 1735 * should hold. 1736 */ 1737 lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; 1738 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && 1739 lock1->li_lock == lock && n == 0) 1740 return (0); 1741 1742 va_start(ap, fmt); 1743 vprintf(fmt, ap); 1744 va_end(ap); 1745 printf(" with the following"); 1746 if (flags & WARN_SLEEPOK) 1747 printf(" non-sleepable"); 1748 printf(" locks held:\n"); 1749 n += witness_list_locks(&lock_list, printf); 1750 } else 1751 sched_unpin(); 1752 if (flags & WARN_PANIC && n) 1753 kassert_panic("%s", __func__); 1754 else 1755 witness_debugger(n); 1756 return (n); 1757 } 1758 1759 const char * 1760 witness_file(struct lock_object *lock) 1761 { 1762 struct witness *w; 1763 1764 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1765 return ("?"); 1766 w = lock->lo_witness; 1767 return (w->w_file); 1768 } 1769 1770 int 1771 witness_line(struct lock_object *lock) 1772 { 1773 struct witness *w; 1774 1775 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1776 return (0); 1777 w = lock->lo_witness; 1778 return (w->w_line); 1779 } 1780 1781 static struct witness * 1782 enroll(const char *description, struct lock_class *lock_class) 1783 { 1784 struct witness *w; 1785 struct witness_list *typelist; 1786 1787 MPASS(description != NULL); 1788 1789 if (witness_watch == -1 || panicstr != NULL) 1790 return (NULL); 1791 if ((lock_class->lc_flags & LC_SPINLOCK)) { 1792 if (witness_skipspin) 1793 return (NULL); 1794 else 1795 typelist = &w_spin; 1796 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) { 1797 typelist = &w_sleep; 1798 } else { 1799 kassert_panic("lock class %s is not sleep or spin", 1800 lock_class->lc_name); 1801 return (NULL); 1802 } 1803 1804 mtx_lock_spin(&w_mtx); 1805 w = witness_hash_get(description); 1806 if (w) 1807 goto found; 1808 if ((w = witness_get()) == NULL) 1809 return (NULL); 1810 MPASS(strlen(description) < MAX_W_NAME); 1811 strcpy(w->w_name, description); 1812 w->w_class = lock_class; 1813 w->w_refcount = 1; 1814 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1815 if (lock_class->lc_flags & LC_SPINLOCK) { 1816 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1817 w_spin_cnt++; 1818 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1819 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1820 w_sleep_cnt++; 1821 } 1822 1823 /* Insert new witness into the hash */ 1824 witness_hash_put(w); 1825 witness_increment_graph_generation(); 1826 mtx_unlock_spin(&w_mtx); 1827 return (w); 1828 found: 1829 w->w_refcount++; 1830 mtx_unlock_spin(&w_mtx); 1831 if (lock_class != w->w_class) 1832 kassert_panic( 1833 "lock (%s) %s does not match earlier (%s) lock", 1834 description, lock_class->lc_name, 1835 w->w_class->lc_name); 1836 return (w); 1837 } 1838 1839 static void 1840 depart(struct witness *w) 1841 { 1842 struct witness_list *list; 1843 1844 MPASS(w->w_refcount == 0); 1845 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1846 list = &w_sleep; 1847 w_sleep_cnt--; 1848 } else { 1849 list = &w_spin; 1850 w_spin_cnt--; 1851 } 1852 /* 1853 * Set file to NULL as it may point into a loadable module. 1854 */ 1855 w->w_file = NULL; 1856 w->w_line = 0; 1857 witness_increment_graph_generation(); 1858 } 1859 1860 1861 static void 1862 adopt(struct witness *parent, struct witness *child) 1863 { 1864 int pi, ci, i, j; 1865 1866 if (witness_cold == 0) 1867 mtx_assert(&w_mtx, MA_OWNED); 1868 1869 /* If the relationship is already known, there's no work to be done. */ 1870 if (isitmychild(parent, child)) 1871 return; 1872 1873 /* When the structure of the graph changes, bump up the generation. */ 1874 witness_increment_graph_generation(); 1875 1876 /* 1877 * The hard part ... create the direct relationship, then propagate all 1878 * indirect relationships. 1879 */ 1880 pi = parent->w_index; 1881 ci = child->w_index; 1882 WITNESS_INDEX_ASSERT(pi); 1883 WITNESS_INDEX_ASSERT(ci); 1884 MPASS(pi != ci); 1885 w_rmatrix[pi][ci] |= WITNESS_PARENT; 1886 w_rmatrix[ci][pi] |= WITNESS_CHILD; 1887 1888 /* 1889 * If parent was not already an ancestor of child, 1890 * then we increment the descendant and ancestor counters. 1891 */ 1892 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { 1893 parent->w_num_descendants++; 1894 child->w_num_ancestors++; 1895 } 1896 1897 /* 1898 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 1899 * an ancestor of 'pi' during this loop. 1900 */ 1901 for (i = 1; i <= w_max_used_index; i++) { 1902 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 1903 (i != pi)) 1904 continue; 1905 1906 /* Find each descendant of 'i' and mark it as a descendant. */ 1907 for (j = 1; j <= w_max_used_index; j++) { 1908 1909 /* 1910 * Skip children that are already marked as 1911 * descendants of 'i'. 1912 */ 1913 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) 1914 continue; 1915 1916 /* 1917 * We are only interested in descendants of 'ci'. Note 1918 * that 'ci' itself is counted as a descendant of 'ci'. 1919 */ 1920 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 1921 (j != ci)) 1922 continue; 1923 w_rmatrix[i][j] |= WITNESS_ANCESTOR; 1924 w_rmatrix[j][i] |= WITNESS_DESCENDANT; 1925 w_data[i].w_num_descendants++; 1926 w_data[j].w_num_ancestors++; 1927 1928 /* 1929 * Make sure we aren't marking a node as both an 1930 * ancestor and descendant. We should have caught 1931 * this as a lock order reversal earlier. 1932 */ 1933 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && 1934 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { 1935 printf("witness rmatrix paradox! [%d][%d]=%d " 1936 "both ancestor and descendant\n", 1937 i, j, w_rmatrix[i][j]); 1938 kdb_backtrace(); 1939 printf("Witness disabled.\n"); 1940 witness_watch = -1; 1941 } 1942 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && 1943 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { 1944 printf("witness rmatrix paradox! [%d][%d]=%d " 1945 "both ancestor and descendant\n", 1946 j, i, w_rmatrix[j][i]); 1947 kdb_backtrace(); 1948 printf("Witness disabled.\n"); 1949 witness_watch = -1; 1950 } 1951 } 1952 } 1953 } 1954 1955 static void 1956 itismychild(struct witness *parent, struct witness *child) 1957 { 1958 int unlocked; 1959 1960 MPASS(child != NULL && parent != NULL); 1961 if (witness_cold == 0) 1962 mtx_assert(&w_mtx, MA_OWNED); 1963 1964 if (!witness_lock_type_equal(parent, child)) { 1965 if (witness_cold == 0) { 1966 unlocked = 1; 1967 mtx_unlock_spin(&w_mtx); 1968 } else { 1969 unlocked = 0; 1970 } 1971 kassert_panic( 1972 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " 1973 "the same lock type", __func__, parent->w_name, 1974 parent->w_class->lc_name, child->w_name, 1975 child->w_class->lc_name); 1976 if (unlocked) 1977 mtx_lock_spin(&w_mtx); 1978 } 1979 adopt(parent, child); 1980 } 1981 1982 /* 1983 * Generic code for the isitmy*() functions. The rmask parameter is the 1984 * expected relationship of w1 to w2. 1985 */ 1986 static int 1987 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) 1988 { 1989 unsigned char r1, r2; 1990 int i1, i2; 1991 1992 i1 = w1->w_index; 1993 i2 = w2->w_index; 1994 WITNESS_INDEX_ASSERT(i1); 1995 WITNESS_INDEX_ASSERT(i2); 1996 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; 1997 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; 1998 1999 /* The flags on one better be the inverse of the flags on the other */ 2000 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || 2001 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { 2002 printf("%s: rmatrix mismatch between %s (index %d) and %s " 2003 "(index %d): w_rmatrix[%d][%d] == %hhx but " 2004 "w_rmatrix[%d][%d] == %hhx\n", 2005 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, 2006 i2, i1, r2); 2007 kdb_backtrace(); 2008 printf("Witness disabled.\n"); 2009 witness_watch = -1; 2010 } 2011 return (r1 & rmask); 2012 } 2013 2014 /* 2015 * Checks if @child is a direct child of @parent. 2016 */ 2017 static int 2018 isitmychild(struct witness *parent, struct witness *child) 2019 { 2020 2021 return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); 2022 } 2023 2024 /* 2025 * Checks if @descendant is a direct or inderect descendant of @ancestor. 2026 */ 2027 static int 2028 isitmydescendant(struct witness *ancestor, struct witness *descendant) 2029 { 2030 2031 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, 2032 __func__)); 2033 } 2034 2035 #ifdef BLESSING 2036 static int 2037 blessed(struct witness *w1, struct witness *w2) 2038 { 2039 int i; 2040 struct witness_blessed *b; 2041 2042 for (i = 0; i < blessed_count; i++) { 2043 b = &blessed_list[i]; 2044 if (strcmp(w1->w_name, b->b_lock1) == 0) { 2045 if (strcmp(w2->w_name, b->b_lock2) == 0) 2046 return (1); 2047 continue; 2048 } 2049 if (strcmp(w1->w_name, b->b_lock2) == 0) 2050 if (strcmp(w2->w_name, b->b_lock1) == 0) 2051 return (1); 2052 } 2053 return (0); 2054 } 2055 #endif 2056 2057 static struct witness * 2058 witness_get(void) 2059 { 2060 struct witness *w; 2061 int index; 2062 2063 if (witness_cold == 0) 2064 mtx_assert(&w_mtx, MA_OWNED); 2065 2066 if (witness_watch == -1) { 2067 mtx_unlock_spin(&w_mtx); 2068 return (NULL); 2069 } 2070 if (STAILQ_EMPTY(&w_free)) { 2071 witness_watch = -1; 2072 mtx_unlock_spin(&w_mtx); 2073 printf("WITNESS: unable to allocate a new witness object\n"); 2074 return (NULL); 2075 } 2076 w = STAILQ_FIRST(&w_free); 2077 STAILQ_REMOVE_HEAD(&w_free, w_list); 2078 w_free_cnt--; 2079 index = w->w_index; 2080 MPASS(index > 0 && index == w_max_used_index+1 && 2081 index < WITNESS_COUNT); 2082 bzero(w, sizeof(*w)); 2083 w->w_index = index; 2084 if (index > w_max_used_index) 2085 w_max_used_index = index; 2086 return (w); 2087 } 2088 2089 static void 2090 witness_free(struct witness *w) 2091 { 2092 2093 STAILQ_INSERT_HEAD(&w_free, w, w_list); 2094 w_free_cnt++; 2095 } 2096 2097 static struct lock_list_entry * 2098 witness_lock_list_get(void) 2099 { 2100 struct lock_list_entry *lle; 2101 2102 if (witness_watch == -1) 2103 return (NULL); 2104 mtx_lock_spin(&w_mtx); 2105 lle = w_lock_list_free; 2106 if (lle == NULL) { 2107 witness_watch = -1; 2108 mtx_unlock_spin(&w_mtx); 2109 printf("%s: witness exhausted\n", __func__); 2110 return (NULL); 2111 } 2112 w_lock_list_free = lle->ll_next; 2113 mtx_unlock_spin(&w_mtx); 2114 bzero(lle, sizeof(*lle)); 2115 return (lle); 2116 } 2117 2118 static void 2119 witness_lock_list_free(struct lock_list_entry *lle) 2120 { 2121 2122 mtx_lock_spin(&w_mtx); 2123 lle->ll_next = w_lock_list_free; 2124 w_lock_list_free = lle; 2125 mtx_unlock_spin(&w_mtx); 2126 } 2127 2128 static struct lock_instance * 2129 find_instance(struct lock_list_entry *list, const struct lock_object *lock) 2130 { 2131 struct lock_list_entry *lle; 2132 struct lock_instance *instance; 2133 int i; 2134 2135 for (lle = list; lle != NULL; lle = lle->ll_next) 2136 for (i = lle->ll_count - 1; i >= 0; i--) { 2137 instance = &lle->ll_children[i]; 2138 if (instance->li_lock == lock) 2139 return (instance); 2140 } 2141 return (NULL); 2142 } 2143 2144 static void 2145 witness_list_lock(struct lock_instance *instance, 2146 int (*prnt)(const char *fmt, ...)) 2147 { 2148 struct lock_object *lock; 2149 2150 lock = instance->li_lock; 2151 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 2152 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 2153 if (lock->lo_witness->w_name != lock->lo_name) 2154 prnt(" (%s)", lock->lo_witness->w_name); 2155 prnt(" r = %d (%p) locked @ %s:%d\n", 2156 instance->li_flags & LI_RECURSEMASK, lock, 2157 fixup_filename(instance->li_file), instance->li_line); 2158 } 2159 2160 #ifdef DDB 2161 static int 2162 witness_thread_has_locks(struct thread *td) 2163 { 2164 2165 if (td->td_sleeplocks == NULL) 2166 return (0); 2167 return (td->td_sleeplocks->ll_count != 0); 2168 } 2169 2170 static int 2171 witness_proc_has_locks(struct proc *p) 2172 { 2173 struct thread *td; 2174 2175 FOREACH_THREAD_IN_PROC(p, td) { 2176 if (witness_thread_has_locks(td)) 2177 return (1); 2178 } 2179 return (0); 2180 } 2181 #endif 2182 2183 int 2184 witness_list_locks(struct lock_list_entry **lock_list, 2185 int (*prnt)(const char *fmt, ...)) 2186 { 2187 struct lock_list_entry *lle; 2188 int i, nheld; 2189 2190 nheld = 0; 2191 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 2192 for (i = lle->ll_count - 1; i >= 0; i--) { 2193 witness_list_lock(&lle->ll_children[i], prnt); 2194 nheld++; 2195 } 2196 return (nheld); 2197 } 2198 2199 /* 2200 * This is a bit risky at best. We call this function when we have timed 2201 * out acquiring a spin lock, and we assume that the other CPU is stuck 2202 * with this lock held. So, we go groveling around in the other CPU's 2203 * per-cpu data to try to find the lock instance for this spin lock to 2204 * see when it was last acquired. 2205 */ 2206 void 2207 witness_display_spinlock(struct lock_object *lock, struct thread *owner, 2208 int (*prnt)(const char *fmt, ...)) 2209 { 2210 struct lock_instance *instance; 2211 struct pcpu *pc; 2212 2213 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 2214 return; 2215 pc = pcpu_find(owner->td_oncpu); 2216 instance = find_instance(pc->pc_spinlocks, lock); 2217 if (instance != NULL) 2218 witness_list_lock(instance, prnt); 2219 } 2220 2221 void 2222 witness_save(struct lock_object *lock, const char **filep, int *linep) 2223 { 2224 struct lock_list_entry *lock_list; 2225 struct lock_instance *instance; 2226 struct lock_class *class; 2227 2228 /* 2229 * This function is used independently in locking code to deal with 2230 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2231 * is gone. 2232 */ 2233 if (SCHEDULER_STOPPED()) 2234 return; 2235 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2236 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2237 return; 2238 class = LOCK_CLASS(lock); 2239 if (class->lc_flags & LC_SLEEPLOCK) 2240 lock_list = curthread->td_sleeplocks; 2241 else { 2242 if (witness_skipspin) 2243 return; 2244 lock_list = PCPU_GET(spinlocks); 2245 } 2246 instance = find_instance(lock_list, lock); 2247 if (instance == NULL) { 2248 kassert_panic("%s: lock (%s) %s not locked", __func__, 2249 class->lc_name, lock->lo_name); 2250 return; 2251 } 2252 *filep = instance->li_file; 2253 *linep = instance->li_line; 2254 } 2255 2256 void 2257 witness_restore(struct lock_object *lock, const char *file, int line) 2258 { 2259 struct lock_list_entry *lock_list; 2260 struct lock_instance *instance; 2261 struct lock_class *class; 2262 2263 /* 2264 * This function is used independently in locking code to deal with 2265 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2266 * is gone. 2267 */ 2268 if (SCHEDULER_STOPPED()) 2269 return; 2270 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2271 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2272 return; 2273 class = LOCK_CLASS(lock); 2274 if (class->lc_flags & LC_SLEEPLOCK) 2275 lock_list = curthread->td_sleeplocks; 2276 else { 2277 if (witness_skipspin) 2278 return; 2279 lock_list = PCPU_GET(spinlocks); 2280 } 2281 instance = find_instance(lock_list, lock); 2282 if (instance == NULL) 2283 kassert_panic("%s: lock (%s) %s not locked", __func__, 2284 class->lc_name, lock->lo_name); 2285 lock->lo_witness->w_file = file; 2286 lock->lo_witness->w_line = line; 2287 if (instance == NULL) 2288 return; 2289 instance->li_file = file; 2290 instance->li_line = line; 2291 } 2292 2293 void 2294 witness_assert(const struct lock_object *lock, int flags, const char *file, 2295 int line) 2296 { 2297 #ifdef INVARIANT_SUPPORT 2298 struct lock_instance *instance; 2299 struct lock_class *class; 2300 2301 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL) 2302 return; 2303 class = LOCK_CLASS(lock); 2304 if ((class->lc_flags & LC_SLEEPLOCK) != 0) 2305 instance = find_instance(curthread->td_sleeplocks, lock); 2306 else if ((class->lc_flags & LC_SPINLOCK) != 0) 2307 instance = find_instance(PCPU_GET(spinlocks), lock); 2308 else { 2309 kassert_panic("Lock (%s) %s is not sleep or spin!", 2310 class->lc_name, lock->lo_name); 2311 return; 2312 } 2313 switch (flags) { 2314 case LA_UNLOCKED: 2315 if (instance != NULL) 2316 kassert_panic("Lock (%s) %s locked @ %s:%d.", 2317 class->lc_name, lock->lo_name, 2318 fixup_filename(file), line); 2319 break; 2320 case LA_LOCKED: 2321 case LA_LOCKED | LA_RECURSED: 2322 case LA_LOCKED | LA_NOTRECURSED: 2323 case LA_SLOCKED: 2324 case LA_SLOCKED | LA_RECURSED: 2325 case LA_SLOCKED | LA_NOTRECURSED: 2326 case LA_XLOCKED: 2327 case LA_XLOCKED | LA_RECURSED: 2328 case LA_XLOCKED | LA_NOTRECURSED: 2329 if (instance == NULL) { 2330 kassert_panic("Lock (%s) %s not locked @ %s:%d.", 2331 class->lc_name, lock->lo_name, 2332 fixup_filename(file), line); 2333 break; 2334 } 2335 if ((flags & LA_XLOCKED) != 0 && 2336 (instance->li_flags & LI_EXCLUSIVE) == 0) 2337 kassert_panic( 2338 "Lock (%s) %s not exclusively locked @ %s:%d.", 2339 class->lc_name, lock->lo_name, 2340 fixup_filename(file), line); 2341 if ((flags & LA_SLOCKED) != 0 && 2342 (instance->li_flags & LI_EXCLUSIVE) != 0) 2343 kassert_panic( 2344 "Lock (%s) %s exclusively locked @ %s:%d.", 2345 class->lc_name, lock->lo_name, 2346 fixup_filename(file), line); 2347 if ((flags & LA_RECURSED) != 0 && 2348 (instance->li_flags & LI_RECURSEMASK) == 0) 2349 kassert_panic("Lock (%s) %s not recursed @ %s:%d.", 2350 class->lc_name, lock->lo_name, 2351 fixup_filename(file), line); 2352 if ((flags & LA_NOTRECURSED) != 0 && 2353 (instance->li_flags & LI_RECURSEMASK) != 0) 2354 kassert_panic("Lock (%s) %s recursed @ %s:%d.", 2355 class->lc_name, lock->lo_name, 2356 fixup_filename(file), line); 2357 break; 2358 default: 2359 kassert_panic("Invalid lock assertion at %s:%d.", 2360 fixup_filename(file), line); 2361 2362 } 2363 #endif /* INVARIANT_SUPPORT */ 2364 } 2365 2366 static void 2367 witness_setflag(struct lock_object *lock, int flag, int set) 2368 { 2369 struct lock_list_entry *lock_list; 2370 struct lock_instance *instance; 2371 struct lock_class *class; 2372 2373 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2374 return; 2375 class = LOCK_CLASS(lock); 2376 if (class->lc_flags & LC_SLEEPLOCK) 2377 lock_list = curthread->td_sleeplocks; 2378 else { 2379 if (witness_skipspin) 2380 return; 2381 lock_list = PCPU_GET(spinlocks); 2382 } 2383 instance = find_instance(lock_list, lock); 2384 if (instance == NULL) { 2385 kassert_panic("%s: lock (%s) %s not locked", __func__, 2386 class->lc_name, lock->lo_name); 2387 return; 2388 } 2389 2390 if (set) 2391 instance->li_flags |= flag; 2392 else 2393 instance->li_flags &= ~flag; 2394 } 2395 2396 void 2397 witness_norelease(struct lock_object *lock) 2398 { 2399 2400 witness_setflag(lock, LI_NORELEASE, 1); 2401 } 2402 2403 void 2404 witness_releaseok(struct lock_object *lock) 2405 { 2406 2407 witness_setflag(lock, LI_NORELEASE, 0); 2408 } 2409 2410 #ifdef DDB 2411 static void 2412 witness_ddb_list(struct thread *td) 2413 { 2414 2415 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2416 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2417 2418 if (witness_watch < 1) 2419 return; 2420 2421 witness_list_locks(&td->td_sleeplocks, db_printf); 2422 2423 /* 2424 * We only handle spinlocks if td == curthread. This is somewhat broken 2425 * if td is currently executing on some other CPU and holds spin locks 2426 * as we won't display those locks. If we had a MI way of getting 2427 * the per-cpu data for a given cpu then we could use 2428 * td->td_oncpu to get the list of spinlocks for this thread 2429 * and "fix" this. 2430 * 2431 * That still wouldn't really fix this unless we locked the scheduler 2432 * lock or stopped the other CPU to make sure it wasn't changing the 2433 * list out from under us. It is probably best to just not try to 2434 * handle threads on other CPU's for now. 2435 */ 2436 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2437 witness_list_locks(PCPU_PTR(spinlocks), db_printf); 2438 } 2439 2440 DB_SHOW_COMMAND(locks, db_witness_list) 2441 { 2442 struct thread *td; 2443 2444 if (have_addr) 2445 td = db_lookup_thread(addr, TRUE); 2446 else 2447 td = kdb_thread; 2448 witness_ddb_list(td); 2449 } 2450 2451 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) 2452 { 2453 struct thread *td; 2454 struct proc *p; 2455 2456 /* 2457 * It would be nice to list only threads and processes that actually 2458 * held sleep locks, but that information is currently not exported 2459 * by WITNESS. 2460 */ 2461 FOREACH_PROC_IN_SYSTEM(p) { 2462 if (!witness_proc_has_locks(p)) 2463 continue; 2464 FOREACH_THREAD_IN_PROC(p, td) { 2465 if (!witness_thread_has_locks(td)) 2466 continue; 2467 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2468 p->p_comm, td, td->td_tid); 2469 witness_ddb_list(td); 2470 if (db_pager_quit) 2471 return; 2472 } 2473 } 2474 } 2475 DB_SHOW_ALIAS(alllocks, db_witness_list_all) 2476 2477 DB_SHOW_COMMAND(witness, db_witness_display) 2478 { 2479 2480 witness_ddb_display(db_printf); 2481 } 2482 #endif 2483 2484 static int 2485 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) 2486 { 2487 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; 2488 struct witness *tmp_w1, *tmp_w2, *w1, *w2; 2489 struct sbuf *sb; 2490 u_int w_rmatrix1, w_rmatrix2; 2491 int error, generation, i, j; 2492 2493 tmp_data1 = NULL; 2494 tmp_data2 = NULL; 2495 tmp_w1 = NULL; 2496 tmp_w2 = NULL; 2497 if (witness_watch < 1) { 2498 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2499 return (error); 2500 } 2501 if (witness_cold) { 2502 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2503 return (error); 2504 } 2505 error = 0; 2506 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND); 2507 if (sb == NULL) 2508 return (ENOMEM); 2509 2510 /* Allocate and init temporary storage space. */ 2511 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2512 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2513 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2514 M_WAITOK | M_ZERO); 2515 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2516 M_WAITOK | M_ZERO); 2517 stack_zero(&tmp_data1->wlod_stack); 2518 stack_zero(&tmp_data2->wlod_stack); 2519 2520 restart: 2521 mtx_lock_spin(&w_mtx); 2522 generation = w_generation; 2523 mtx_unlock_spin(&w_mtx); 2524 sbuf_printf(sb, "Number of known direct relationships is %d\n", 2525 w_lohash.wloh_count); 2526 for (i = 1; i < w_max_used_index; i++) { 2527 mtx_lock_spin(&w_mtx); 2528 if (generation != w_generation) { 2529 mtx_unlock_spin(&w_mtx); 2530 2531 /* The graph has changed, try again. */ 2532 req->oldidx = 0; 2533 sbuf_clear(sb); 2534 goto restart; 2535 } 2536 2537 w1 = &w_data[i]; 2538 if (w1->w_reversed == 0) { 2539 mtx_unlock_spin(&w_mtx); 2540 continue; 2541 } 2542 2543 /* Copy w1 locally so we can release the spin lock. */ 2544 *tmp_w1 = *w1; 2545 mtx_unlock_spin(&w_mtx); 2546 2547 if (tmp_w1->w_reversed == 0) 2548 continue; 2549 for (j = 1; j < w_max_used_index; j++) { 2550 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) 2551 continue; 2552 2553 mtx_lock_spin(&w_mtx); 2554 if (generation != w_generation) { 2555 mtx_unlock_spin(&w_mtx); 2556 2557 /* The graph has changed, try again. */ 2558 req->oldidx = 0; 2559 sbuf_clear(sb); 2560 goto restart; 2561 } 2562 2563 w2 = &w_data[j]; 2564 data1 = witness_lock_order_get(w1, w2); 2565 data2 = witness_lock_order_get(w2, w1); 2566 2567 /* 2568 * Copy information locally so we can release the 2569 * spin lock. 2570 */ 2571 *tmp_w2 = *w2; 2572 w_rmatrix1 = (unsigned int)w_rmatrix[i][j]; 2573 w_rmatrix2 = (unsigned int)w_rmatrix[j][i]; 2574 2575 if (data1) { 2576 stack_zero(&tmp_data1->wlod_stack); 2577 stack_copy(&data1->wlod_stack, 2578 &tmp_data1->wlod_stack); 2579 } 2580 if (data2 && data2 != data1) { 2581 stack_zero(&tmp_data2->wlod_stack); 2582 stack_copy(&data2->wlod_stack, 2583 &tmp_data2->wlod_stack); 2584 } 2585 mtx_unlock_spin(&w_mtx); 2586 2587 sbuf_printf(sb, 2588 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", 2589 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2590 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2591 #if 0 2592 sbuf_printf(sb, 2593 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n", 2594 tmp_w1->name, tmp_w2->w_name, w_rmatrix1, 2595 tmp_w2->name, tmp_w1->w_name, w_rmatrix2); 2596 #endif 2597 if (data1) { 2598 sbuf_printf(sb, 2599 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2600 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2601 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2602 stack_sbuf_print(sb, &tmp_data1->wlod_stack); 2603 sbuf_printf(sb, "\n"); 2604 } 2605 if (data2 && data2 != data1) { 2606 sbuf_printf(sb, 2607 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2608 tmp_w2->w_name, tmp_w2->w_class->lc_name, 2609 tmp_w1->w_name, tmp_w1->w_class->lc_name); 2610 stack_sbuf_print(sb, &tmp_data2->wlod_stack); 2611 sbuf_printf(sb, "\n"); 2612 } 2613 } 2614 } 2615 mtx_lock_spin(&w_mtx); 2616 if (generation != w_generation) { 2617 mtx_unlock_spin(&w_mtx); 2618 2619 /* 2620 * The graph changed while we were printing stack data, 2621 * try again. 2622 */ 2623 req->oldidx = 0; 2624 sbuf_clear(sb); 2625 goto restart; 2626 } 2627 mtx_unlock_spin(&w_mtx); 2628 2629 /* Free temporary storage space. */ 2630 free(tmp_data1, M_TEMP); 2631 free(tmp_data2, M_TEMP); 2632 free(tmp_w1, M_TEMP); 2633 free(tmp_w2, M_TEMP); 2634 2635 sbuf_finish(sb); 2636 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2637 sbuf_delete(sb); 2638 2639 return (error); 2640 } 2641 2642 static int 2643 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) 2644 { 2645 struct witness *w; 2646 struct sbuf *sb; 2647 int error; 2648 2649 if (witness_watch < 1) { 2650 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2651 return (error); 2652 } 2653 if (witness_cold) { 2654 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2655 return (error); 2656 } 2657 error = 0; 2658 2659 error = sysctl_wire_old_buffer(req, 0); 2660 if (error != 0) 2661 return (error); 2662 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); 2663 if (sb == NULL) 2664 return (ENOMEM); 2665 sbuf_printf(sb, "\n"); 2666 2667 mtx_lock_spin(&w_mtx); 2668 STAILQ_FOREACH(w, &w_all, w_list) 2669 w->w_displayed = 0; 2670 STAILQ_FOREACH(w, &w_all, w_list) 2671 witness_add_fullgraph(sb, w); 2672 mtx_unlock_spin(&w_mtx); 2673 2674 /* 2675 * Close the sbuf and return to userland. 2676 */ 2677 error = sbuf_finish(sb); 2678 sbuf_delete(sb); 2679 2680 return (error); 2681 } 2682 2683 static int 2684 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 2685 { 2686 int error, value; 2687 2688 value = witness_watch; 2689 error = sysctl_handle_int(oidp, &value, 0, req); 2690 if (error != 0 || req->newptr == NULL) 2691 return (error); 2692 if (value > 1 || value < -1 || 2693 (witness_watch == -1 && value != witness_watch)) 2694 return (EINVAL); 2695 witness_watch = value; 2696 return (0); 2697 } 2698 2699 static void 2700 witness_add_fullgraph(struct sbuf *sb, struct witness *w) 2701 { 2702 int i; 2703 2704 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) 2705 return; 2706 w->w_displayed = 1; 2707 2708 WITNESS_INDEX_ASSERT(w->w_index); 2709 for (i = 1; i <= w_max_used_index; i++) { 2710 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { 2711 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, 2712 w_data[i].w_name); 2713 witness_add_fullgraph(sb, &w_data[i]); 2714 } 2715 } 2716 } 2717 2718 /* 2719 * A simple hash function. Takes a key pointer and a key size. If size == 0, 2720 * interprets the key as a string and reads until the null 2721 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit 2722 * hash value computed from the key. 2723 */ 2724 static uint32_t 2725 witness_hash_djb2(const uint8_t *key, uint32_t size) 2726 { 2727 unsigned int hash = 5381; 2728 int i; 2729 2730 /* hash = hash * 33 + key[i] */ 2731 if (size) 2732 for (i = 0; i < size; i++) 2733 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2734 else 2735 for (i = 0; key[i] != 0; i++) 2736 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2737 2738 return (hash); 2739 } 2740 2741 2742 /* 2743 * Initializes the two witness hash tables. Called exactly once from 2744 * witness_initialize(). 2745 */ 2746 static void 2747 witness_init_hash_tables(void) 2748 { 2749 int i; 2750 2751 MPASS(witness_cold); 2752 2753 /* Initialize the hash tables. */ 2754 for (i = 0; i < WITNESS_HASH_SIZE; i++) 2755 w_hash.wh_array[i] = NULL; 2756 2757 w_hash.wh_size = WITNESS_HASH_SIZE; 2758 w_hash.wh_count = 0; 2759 2760 /* Initialize the lock order data hash. */ 2761 w_lofree = NULL; 2762 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { 2763 memset(&w_lodata[i], 0, sizeof(w_lodata[i])); 2764 w_lodata[i].wlod_next = w_lofree; 2765 w_lofree = &w_lodata[i]; 2766 } 2767 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; 2768 w_lohash.wloh_count = 0; 2769 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) 2770 w_lohash.wloh_array[i] = NULL; 2771 } 2772 2773 static struct witness * 2774 witness_hash_get(const char *key) 2775 { 2776 struct witness *w; 2777 uint32_t hash; 2778 2779 MPASS(key != NULL); 2780 if (witness_cold == 0) 2781 mtx_assert(&w_mtx, MA_OWNED); 2782 hash = witness_hash_djb2(key, 0) % w_hash.wh_size; 2783 w = w_hash.wh_array[hash]; 2784 while (w != NULL) { 2785 if (strcmp(w->w_name, key) == 0) 2786 goto out; 2787 w = w->w_hash_next; 2788 } 2789 2790 out: 2791 return (w); 2792 } 2793 2794 static void 2795 witness_hash_put(struct witness *w) 2796 { 2797 uint32_t hash; 2798 2799 MPASS(w != NULL); 2800 MPASS(w->w_name != NULL); 2801 if (witness_cold == 0) 2802 mtx_assert(&w_mtx, MA_OWNED); 2803 KASSERT(witness_hash_get(w->w_name) == NULL, 2804 ("%s: trying to add a hash entry that already exists!", __func__)); 2805 KASSERT(w->w_hash_next == NULL, 2806 ("%s: w->w_hash_next != NULL", __func__)); 2807 2808 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; 2809 w->w_hash_next = w_hash.wh_array[hash]; 2810 w_hash.wh_array[hash] = w; 2811 w_hash.wh_count++; 2812 } 2813 2814 2815 static struct witness_lock_order_data * 2816 witness_lock_order_get(struct witness *parent, struct witness *child) 2817 { 2818 struct witness_lock_order_data *data = NULL; 2819 struct witness_lock_order_key key; 2820 unsigned int hash; 2821 2822 MPASS(parent != NULL && child != NULL); 2823 key.from = parent->w_index; 2824 key.to = child->w_index; 2825 WITNESS_INDEX_ASSERT(key.from); 2826 WITNESS_INDEX_ASSERT(key.to); 2827 if ((w_rmatrix[parent->w_index][child->w_index] 2828 & WITNESS_LOCK_ORDER_KNOWN) == 0) 2829 goto out; 2830 2831 hash = witness_hash_djb2((const char*)&key, 2832 sizeof(key)) % w_lohash.wloh_size; 2833 data = w_lohash.wloh_array[hash]; 2834 while (data != NULL) { 2835 if (witness_lock_order_key_equal(&data->wlod_key, &key)) 2836 break; 2837 data = data->wlod_next; 2838 } 2839 2840 out: 2841 return (data); 2842 } 2843 2844 /* 2845 * Verify that parent and child have a known relationship, are not the same, 2846 * and child is actually a child of parent. This is done without w_mtx 2847 * to avoid contention in the common case. 2848 */ 2849 static int 2850 witness_lock_order_check(struct witness *parent, struct witness *child) 2851 { 2852 2853 if (parent != child && 2854 w_rmatrix[parent->w_index][child->w_index] 2855 & WITNESS_LOCK_ORDER_KNOWN && 2856 isitmychild(parent, child)) 2857 return (1); 2858 2859 return (0); 2860 } 2861 2862 static int 2863 witness_lock_order_add(struct witness *parent, struct witness *child) 2864 { 2865 struct witness_lock_order_data *data = NULL; 2866 struct witness_lock_order_key key; 2867 unsigned int hash; 2868 2869 MPASS(parent != NULL && child != NULL); 2870 key.from = parent->w_index; 2871 key.to = child->w_index; 2872 WITNESS_INDEX_ASSERT(key.from); 2873 WITNESS_INDEX_ASSERT(key.to); 2874 if (w_rmatrix[parent->w_index][child->w_index] 2875 & WITNESS_LOCK_ORDER_KNOWN) 2876 return (1); 2877 2878 hash = witness_hash_djb2((const char*)&key, 2879 sizeof(key)) % w_lohash.wloh_size; 2880 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; 2881 data = w_lofree; 2882 if (data == NULL) 2883 return (0); 2884 w_lofree = data->wlod_next; 2885 data->wlod_next = w_lohash.wloh_array[hash]; 2886 data->wlod_key = key; 2887 w_lohash.wloh_array[hash] = data; 2888 w_lohash.wloh_count++; 2889 stack_zero(&data->wlod_stack); 2890 stack_save(&data->wlod_stack); 2891 return (1); 2892 } 2893 2894 /* Call this whenver the structure of the witness graph changes. */ 2895 static void 2896 witness_increment_graph_generation(void) 2897 { 2898 2899 if (witness_cold == 0) 2900 mtx_assert(&w_mtx, MA_OWNED); 2901 w_generation++; 2902 } 2903 2904 #ifdef KDB 2905 static void 2906 _witness_debugger(int cond, const char *msg) 2907 { 2908 2909 if (witness_trace && cond) 2910 kdb_backtrace(); 2911 if (witness_kdb && cond) 2912 kdb_enter(KDB_WHY_WITNESS, msg); 2913 } 2914 #endif 2915