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