xref: /freebsd/sys/kern/subr_witness.c (revision cc68614da8232d8baaca0ae0d0dd8f890f06623e)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2008 Isilon Systems, Inc.
5  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
6  * Copyright (c) 1998 Berkeley Software Design, Inc.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Berkeley Software Design Inc's name may not be used to endorse or
18  *    promote products derived from this software without specific prior
19  *    written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
34  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
35  */
36 
37 /*
38  * Implementation of the `witness' lock verifier.  Originally implemented for
39  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
40  * classes in FreeBSD.
41  */
42 
43 /*
44  *	Main Entry: witness
45  *	Pronunciation: 'wit-n&s
46  *	Function: noun
47  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
48  *	    testimony, witness, from 2wit
49  *	Date: before 12th century
50  *	1 : attestation of a fact or event : TESTIMONY
51  *	2 : one that gives evidence; specifically : one who testifies in
52  *	    a cause or before a judicial tribunal
53  *	3 : one asked to be present at a transaction so as to be able to
54  *	    testify to its having taken place
55  *	4 : one who has personal knowledge of something
56  *	5 a : something serving as evidence or proof : SIGN
57  *	  b : public affirmation by word or example of usually
58  *	      religious faith or conviction <the heroic witness to divine
59  *	      life -- Pilot>
60  *	6 capitalized : a member of the Jehovah's Witnesses
61  */
62 
63 /*
64  * Special rules concerning Giant and lock orders:
65  *
66  * 1) Giant must be acquired before any other mutexes.  Stated another way,
67  *    no other mutex may be held when Giant is acquired.
68  *
69  * 2) Giant must be released when blocking on a sleepable lock.
70  *
71  * This rule is less obvious, but is a result of Giant providing the same
72  * semantics as spl().  Basically, when a thread sleeps, it must release
73  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
74  * 2).
75  *
76  * 3) Giant may be acquired before or after sleepable locks.
77  *
78  * This rule is also not quite as obvious.  Giant may be acquired after
79  * a sleepable lock because it is a non-sleepable lock and non-sleepable
80  * locks may always be acquired while holding a sleepable lock.  The second
81  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
82  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
83  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
84  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
85  * execute.  Thus, acquiring Giant both before and after a sleepable lock
86  * will not result in a lock order reversal.
87  */
88 
89 #include <sys/cdefs.h>
90 __FBSDID("$FreeBSD$");
91 
92 #include "opt_ddb.h"
93 #include "opt_hwpmc_hooks.h"
94 #include "opt_stack.h"
95 #include "opt_witness.h"
96 
97 #include <sys/param.h>
98 #include <sys/bus.h>
99 #include <sys/kdb.h>
100 #include <sys/kernel.h>
101 #include <sys/ktr.h>
102 #include <sys/lock.h>
103 #include <sys/malloc.h>
104 #include <sys/mutex.h>
105 #include <sys/priv.h>
106 #include <sys/proc.h>
107 #include <sys/sbuf.h>
108 #include <sys/sched.h>
109 #include <sys/stack.h>
110 #include <sys/sysctl.h>
111 #include <sys/syslog.h>
112 #include <sys/systm.h>
113 
114 #ifdef DDB
115 #include <ddb/ddb.h>
116 #endif
117 
118 #include <machine/stdarg.h>
119 
120 #if !defined(DDB) && !defined(STACK)
121 #error "DDB or STACK options are required for WITNESS"
122 #endif
123 
124 /* Note that these traces do not work with KTR_ALQ. */
125 #if 0
126 #define	KTR_WITNESS	KTR_SUBSYS
127 #else
128 #define	KTR_WITNESS	0
129 #endif
130 
131 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
132 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
133 #define	LI_NORELEASE	0x00020000	/* Lock not allowed to be released. */
134 #define	LI_SLEEPABLE	0x00040000	/* Lock may be held while sleeping. */
135 
136 #ifndef WITNESS_COUNT
137 #define	WITNESS_COUNT 		1536
138 #endif
139 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
140 #define	WITNESS_PENDLIST	(512 + (MAXCPU * 4))
141 
142 /* Allocate 256 KB of stack data space */
143 #define	WITNESS_LO_DATA_COUNT	2048
144 
145 /* Prime, gives load factor of ~2 at full load */
146 #define	WITNESS_LO_HASH_SIZE	1021
147 
148 /*
149  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
150  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
151  * probably be safe for the most part, but it's still a SWAG.
152  */
153 #define	LOCK_NCHILDREN	5
154 #define	LOCK_CHILDCOUNT	2048
155 
156 #define	MAX_W_NAME	64
157 
158 #define	FULLGRAPH_SBUF_SIZE	512
159 
160 /*
161  * These flags go in the witness relationship matrix and describe the
162  * relationship between any two struct witness objects.
163  */
164 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
165 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
166 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
167 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
168 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
169 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
170 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
171 #define	WITNESS_RELATED_MASK						\
172 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
173 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
174 					  * observed. */
175 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
176 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
177 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
178 
179 /* Descendant to ancestor flags */
180 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
181 
182 /* Ancestor to descendant flags */
183 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
184 
185 #define	WITNESS_INDEX_ASSERT(i)						\
186 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
187 
188 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189 
190 /*
191  * Lock instances.  A lock instance is the data associated with a lock while
192  * it is held by witness.  For example, a lock instance will hold the
193  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
194  * are held in a per-cpu list while sleep locks are held in per-thread list.
195  */
196 struct lock_instance {
197 	struct lock_object	*li_lock;
198 	const char		*li_file;
199 	int			li_line;
200 	u_int			li_flags;
201 };
202 
203 /*
204  * A simple list type used to build the list of locks held by a thread
205  * or CPU.  We can't simply embed the list in struct lock_object since a
206  * lock may be held by more than one thread if it is a shared lock.  Locks
207  * are added to the head of the list, so we fill up each list entry from
208  * "the back" logically.  To ease some of the arithmetic, we actually fill
209  * in each list entry the normal way (children[0] then children[1], etc.) but
210  * when we traverse the list we read children[count-1] as the first entry
211  * down to children[0] as the final entry.
212  */
213 struct lock_list_entry {
214 	struct lock_list_entry	*ll_next;
215 	struct lock_instance	ll_children[LOCK_NCHILDREN];
216 	u_int			ll_count;
217 };
218 
219 /*
220  * The main witness structure. One of these per named lock type in the system
221  * (for example, "vnode interlock").
222  */
223 struct witness {
224 	char  			w_name[MAX_W_NAME];
225 	uint32_t 		w_index;  /* Index in the relationship matrix */
226 	struct lock_class	*w_class;
227 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
228 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
229 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
230 	const char		*w_file; /* File where last acquired */
231 	uint32_t 		w_line; /* Line where last acquired */
232 	uint32_t 		w_refcount;
233 	uint16_t 		w_num_ancestors; /* direct/indirect
234 						  * ancestor count */
235 	uint16_t 		w_num_descendants; /* direct/indirect
236 						    * descendant count */
237 	int16_t 		w_ddb_level;
238 	unsigned		w_displayed:1;
239 	unsigned		w_reversed:1;
240 };
241 
242 STAILQ_HEAD(witness_list, witness);
243 
244 /*
245  * The witness hash table. Keys are witness names (const char *), elements are
246  * witness objects (struct witness *).
247  */
248 struct witness_hash {
249 	struct witness	*wh_array[WITNESS_HASH_SIZE];
250 	uint32_t	wh_size;
251 	uint32_t	wh_count;
252 };
253 
254 /*
255  * Key type for the lock order data hash table.
256  */
257 struct witness_lock_order_key {
258 	uint16_t	from;
259 	uint16_t	to;
260 };
261 
262 struct witness_lock_order_data {
263 	struct stack			wlod_stack;
264 	struct witness_lock_order_key	wlod_key;
265 	struct witness_lock_order_data	*wlod_next;
266 };
267 
268 /*
269  * The witness lock order data hash table. Keys are witness index tuples
270  * (struct witness_lock_order_key), elements are lock order data objects
271  * (struct witness_lock_order_data).
272  */
273 struct witness_lock_order_hash {
274 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
275 	u_int	wloh_size;
276 	u_int	wloh_count;
277 };
278 
279 struct witness_blessed {
280 	const char	*b_lock1;
281 	const char	*b_lock2;
282 };
283 
284 struct witness_pendhelp {
285 	const char		*wh_type;
286 	struct lock_object	*wh_lock;
287 };
288 
289 struct witness_order_list_entry {
290 	const char		*w_name;
291 	struct lock_class	*w_class;
292 };
293 
294 /*
295  * Returns 0 if one of the locks is a spin lock and the other is not.
296  * Returns 1 otherwise.
297  */
298 static __inline int
299 witness_lock_type_equal(struct witness *w1, struct witness *w2)
300 {
301 
302 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
303 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
304 }
305 
306 static __inline int
307 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
308     const struct witness_lock_order_key *b)
309 {
310 
311 	return (a->from == b->from && a->to == b->to);
312 }
313 
314 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
315 		    const char *fname);
316 static void	adopt(struct witness *parent, struct witness *child);
317 static int	blessed(struct witness *, struct witness *);
318 static void	depart(struct witness *w);
319 static struct witness	*enroll(const char *description,
320 			    struct lock_class *lock_class);
321 static struct lock_instance	*find_instance(struct lock_list_entry *list,
322 				    const struct lock_object *lock);
323 static int	isitmychild(struct witness *parent, struct witness *child);
324 static int	isitmydescendant(struct witness *parent, struct witness *child);
325 static void	itismychild(struct witness *parent, struct witness *child);
326 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
327 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
328 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
329 static int	sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
330 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
331 #ifdef DDB
332 static void	witness_ddb_compute_levels(void);
333 static void	witness_ddb_display(int(*)(const char *fmt, ...));
334 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
335 		    struct witness *, int indent);
336 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
337 		    struct witness_list *list);
338 static void	witness_ddb_level_descendants(struct witness *parent, int l);
339 static void	witness_ddb_list(struct thread *td);
340 #endif
341 static void	witness_enter_debugger(const char *msg);
342 static void	witness_debugger(int cond, const char *msg);
343 static void	witness_free(struct witness *m);
344 static struct witness	*witness_get(void);
345 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
346 static struct witness	*witness_hash_get(const char *key);
347 static void	witness_hash_put(struct witness *w);
348 static void	witness_init_hash_tables(void);
349 static void	witness_increment_graph_generation(void);
350 static void	witness_lock_list_free(struct lock_list_entry *lle);
351 static struct lock_list_entry	*witness_lock_list_get(void);
352 static int	witness_lock_order_add(struct witness *parent,
353 		    struct witness *child);
354 static int	witness_lock_order_check(struct witness *parent,
355 		    struct witness *child);
356 static struct witness_lock_order_data	*witness_lock_order_get(
357 					    struct witness *parent,
358 					    struct witness *child);
359 static void	witness_list_lock(struct lock_instance *instance,
360 		    int (*prnt)(const char *fmt, ...));
361 static int	witness_output(const char *fmt, ...) __printflike(1, 2);
362 static int	witness_output_drain(void *arg __unused, const char *data,
363 		    int len);
364 static int	witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
365 static void	witness_setflag(struct lock_object *lock, int flag, int set);
366 
367 FEATURE(witness, "kernel has witness(9) support");
368 
369 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
370     "Witness Locking");
371 
372 /*
373  * If set to 0, lock order checking is disabled.  If set to -1,
374  * witness is completely disabled.  Otherwise witness performs full
375  * lock order checking for all locks.  At runtime, lock order checking
376  * may be toggled.  However, witness cannot be reenabled once it is
377  * completely disabled.
378  */
379 static int witness_watch = 1;
380 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
381     CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
382     sysctl_debug_witness_watch, "I",
383     "witness is watching lock operations");
384 
385 #ifdef KDB
386 /*
387  * When KDB is enabled and witness_kdb is 1, it will cause the system
388  * to drop into kdebug() when:
389  *	- a lock hierarchy violation occurs
390  *	- locks are held when going to sleep.
391  */
392 #ifdef WITNESS_KDB
393 int	witness_kdb = 1;
394 #else
395 int	witness_kdb = 0;
396 #endif
397 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
398 #endif /* KDB */
399 
400 #if defined(DDB) || defined(KDB)
401 /*
402  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
403  * to print a stack trace:
404  *	- a lock hierarchy violation occurs
405  *	- locks are held when going to sleep.
406  */
407 int	witness_trace = 1;
408 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
409 #endif /* DDB || KDB */
410 
411 #ifdef WITNESS_SKIPSPIN
412 int	witness_skipspin = 1;
413 #else
414 int	witness_skipspin = 0;
415 #endif
416 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
417 
418 int badstack_sbuf_size;
419 
420 int witness_count = WITNESS_COUNT;
421 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
422     &witness_count, 0, "");
423 
424 /*
425  * Output channel for witness messages.  By default we print to the console.
426  */
427 enum witness_channel {
428 	WITNESS_CONSOLE,
429 	WITNESS_LOG,
430 	WITNESS_NONE,
431 };
432 
433 static enum witness_channel witness_channel = WITNESS_CONSOLE;
434 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
435     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
436     sysctl_debug_witness_channel, "A",
437     "Output channel for warnings");
438 
439 /*
440  * Call this to print out the relations between locks.
441  */
442 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
443     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
444     sysctl_debug_witness_fullgraph, "A",
445     "Show locks relation graphs");
446 
447 /*
448  * Call this to print out the witness faulty stacks.
449  */
450 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
451     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
452     sysctl_debug_witness_badstacks, "A",
453     "Show bad witness stacks");
454 
455 static struct mtx w_mtx;
456 
457 /* w_list */
458 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
459 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
460 
461 /* w_typelist */
462 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
463 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
464 
465 /* lock list */
466 static struct lock_list_entry *w_lock_list_free = NULL;
467 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
468 static u_int pending_cnt;
469 
470 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
471 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
472 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
473 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
474     "");
475 
476 static struct witness *w_data;
477 static uint8_t **w_rmatrix;
478 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
479 static struct witness_hash w_hash;	/* The witness hash table. */
480 
481 /* The lock order data hash */
482 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
483 static struct witness_lock_order_data *w_lofree = NULL;
484 static struct witness_lock_order_hash w_lohash;
485 static int w_max_used_index = 0;
486 static unsigned int w_generation = 0;
487 static const char w_notrunning[] = "Witness not running\n";
488 static const char w_stillcold[] = "Witness is still cold\n";
489 #ifdef __i386__
490 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
491 #endif
492 
493 static struct witness_order_list_entry order_lists[] = {
494 	/*
495 	 * sx locks
496 	 */
497 	{ "proctree", &lock_class_sx },
498 	{ "allproc", &lock_class_sx },
499 	{ "allprison", &lock_class_sx },
500 	{ NULL, NULL },
501 	/*
502 	 * Various mutexes
503 	 */
504 	{ "Giant", &lock_class_mtx_sleep },
505 	{ "pipe mutex", &lock_class_mtx_sleep },
506 	{ "sigio lock", &lock_class_mtx_sleep },
507 	{ "process group", &lock_class_mtx_sleep },
508 #ifdef	HWPMC_HOOKS
509 	{ "pmc-sleep", &lock_class_mtx_sleep },
510 #endif
511 	{ "process lock", &lock_class_mtx_sleep },
512 	{ "session", &lock_class_mtx_sleep },
513 	{ "uidinfo hash", &lock_class_rw },
514 	{ "time lock", &lock_class_mtx_sleep },
515 	{ NULL, NULL },
516 	/*
517 	 * umtx
518 	 */
519 	{ "umtx lock", &lock_class_mtx_sleep },
520 	{ NULL, NULL },
521 	/*
522 	 * Sockets
523 	 */
524 	{ "accept", &lock_class_mtx_sleep },
525 	{ "so_snd", &lock_class_mtx_sleep },
526 	{ "so_rcv", &lock_class_mtx_sleep },
527 	{ "sellck", &lock_class_mtx_sleep },
528 	{ NULL, NULL },
529 	/*
530 	 * Routing
531 	 */
532 	{ "so_rcv", &lock_class_mtx_sleep },
533 	{ "radix node head", &lock_class_rm },
534 	{ "ifaddr", &lock_class_mtx_sleep },
535 	{ NULL, NULL },
536 	/*
537 	 * IPv4 multicast:
538 	 * protocol locks before interface locks, after UDP locks.
539 	 */
540 	{ "in_multi_sx", &lock_class_sx },
541 	{ "udpinp", &lock_class_rw },
542 	{ "in_multi_list_mtx", &lock_class_mtx_sleep },
543 	{ "igmp_mtx", &lock_class_mtx_sleep },
544 	{ "if_addr_lock", &lock_class_mtx_sleep },
545 	{ NULL, NULL },
546 	/*
547 	 * IPv6 multicast:
548 	 * protocol locks before interface locks, after UDP locks.
549 	 */
550 	{ "in6_multi_sx", &lock_class_sx },
551 	{ "udpinp", &lock_class_rw },
552 	{ "in6_multi_list_mtx", &lock_class_mtx_sleep },
553 	{ "mld_mtx", &lock_class_mtx_sleep },
554 	{ "if_addr_lock", &lock_class_mtx_sleep },
555 	{ NULL, NULL },
556 	/*
557 	 * UNIX Domain Sockets
558 	 */
559 	{ "unp_link_rwlock", &lock_class_rw },
560 	{ "unp_list_lock", &lock_class_mtx_sleep },
561 	{ "unp", &lock_class_mtx_sleep },
562 	{ "so_snd", &lock_class_mtx_sleep },
563 	{ NULL, NULL },
564 	/*
565 	 * UDP/IP
566 	 */
567 	{ "udpinp", &lock_class_rw },
568 	{ "udp", &lock_class_mtx_sleep },
569 	{ "so_snd", &lock_class_mtx_sleep },
570 	{ NULL, NULL },
571 	/*
572 	 * TCP/IP
573 	 */
574 	{ "tcpinp", &lock_class_rw },
575 	{ "tcp", &lock_class_mtx_sleep },
576 	{ "so_snd", &lock_class_mtx_sleep },
577 	{ NULL, NULL },
578 	/*
579 	 * BPF
580 	 */
581 	{ "bpf global lock", &lock_class_sx },
582 	{ "bpf cdev lock", &lock_class_mtx_sleep },
583 	{ NULL, NULL },
584 	/*
585 	 * NFS server
586 	 */
587 	{ "nfsd_mtx", &lock_class_mtx_sleep },
588 	{ "so_snd", &lock_class_mtx_sleep },
589 	{ NULL, NULL },
590 
591 	/*
592 	 * IEEE 802.11
593 	 */
594 	{ "802.11 com lock", &lock_class_mtx_sleep},
595 	{ NULL, NULL },
596 	/*
597 	 * Network drivers
598 	 */
599 	{ "network driver", &lock_class_mtx_sleep},
600 	{ NULL, NULL },
601 
602 	/*
603 	 * Netgraph
604 	 */
605 	{ "ng_node", &lock_class_mtx_sleep },
606 	{ "ng_worklist", &lock_class_mtx_sleep },
607 	{ NULL, NULL },
608 	/*
609 	 * CDEV
610 	 */
611 	{ "vm map (system)", &lock_class_mtx_sleep },
612 	{ "vnode interlock", &lock_class_mtx_sleep },
613 	{ "cdev", &lock_class_mtx_sleep },
614 	{ "devthrd", &lock_class_mtx_sleep },
615 	{ NULL, NULL },
616 	/*
617 	 * VM
618 	 */
619 	{ "vm map (user)", &lock_class_sx },
620 	{ "vm object", &lock_class_rw },
621 	{ "vm page", &lock_class_mtx_sleep },
622 	{ "pmap pv global", &lock_class_rw },
623 	{ "pmap", &lock_class_mtx_sleep },
624 	{ "pmap pv list", &lock_class_rw },
625 	{ "vm page free queue", &lock_class_mtx_sleep },
626 	{ "vm pagequeue", &lock_class_mtx_sleep },
627 	{ NULL, NULL },
628 	/*
629 	 * kqueue/VFS interaction
630 	 */
631 	{ "kqueue", &lock_class_mtx_sleep },
632 	{ "struct mount mtx", &lock_class_mtx_sleep },
633 	{ "vnode interlock", &lock_class_mtx_sleep },
634 	{ NULL, NULL },
635 	/*
636 	 * VFS namecache
637 	 */
638 	{ "ncvn", &lock_class_mtx_sleep },
639 	{ "ncbuc", &lock_class_mtx_sleep },
640 	{ "vnode interlock", &lock_class_mtx_sleep },
641 	{ "ncneg", &lock_class_mtx_sleep },
642 	{ NULL, NULL },
643 	/*
644 	 * ZFS locking
645 	 */
646 	{ "dn->dn_mtx", &lock_class_sx },
647 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
648 	{ "db->db_mtx", &lock_class_sx },
649 	{ NULL, NULL },
650 	/*
651 	 * TCP log locks
652 	 */
653 	{ "TCP ID tree", &lock_class_rw },
654 	{ "tcp log id bucket", &lock_class_mtx_sleep },
655 	{ "tcpinp", &lock_class_rw },
656 	{ "TCP log expireq", &lock_class_mtx_sleep },
657 	{ NULL, NULL },
658 	/*
659 	 * spin locks
660 	 */
661 #ifdef SMP
662 	{ "ap boot", &lock_class_mtx_spin },
663 #endif
664 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
665 #ifdef __i386__
666 	{ "cy", &lock_class_mtx_spin },
667 #endif
668 	{ "scc_hwmtx", &lock_class_mtx_spin },
669 	{ "uart_hwmtx", &lock_class_mtx_spin },
670 	{ "fast_taskqueue", &lock_class_mtx_spin },
671 	{ "intr table", &lock_class_mtx_spin },
672 	{ "process slock", &lock_class_mtx_spin },
673 	{ "syscons video lock", &lock_class_mtx_spin },
674 	{ "sleepq chain", &lock_class_mtx_spin },
675 	{ "rm_spinlock", &lock_class_mtx_spin },
676 	{ "turnstile chain", &lock_class_mtx_spin },
677 	{ "turnstile lock", &lock_class_mtx_spin },
678 	{ "sched lock", &lock_class_mtx_spin },
679 	{ "td_contested", &lock_class_mtx_spin },
680 	{ "callout", &lock_class_mtx_spin },
681 	{ "entropy harvest mutex", &lock_class_mtx_spin },
682 #ifdef SMP
683 	{ "smp rendezvous", &lock_class_mtx_spin },
684 #endif
685 #ifdef __powerpc__
686 	{ "tlb0", &lock_class_mtx_spin },
687 #endif
688 	{ NULL, NULL },
689 	{ "sched lock", &lock_class_mtx_spin },
690 #ifdef	HWPMC_HOOKS
691 	{ "pmc-per-proc", &lock_class_mtx_spin },
692 #endif
693 	{ NULL, NULL },
694 	/*
695 	 * leaf locks
696 	 */
697 	{ "intrcnt", &lock_class_mtx_spin },
698 	{ "icu", &lock_class_mtx_spin },
699 #ifdef __i386__
700 	{ "allpmaps", &lock_class_mtx_spin },
701 	{ "descriptor tables", &lock_class_mtx_spin },
702 #endif
703 	{ "clk", &lock_class_mtx_spin },
704 	{ "cpuset", &lock_class_mtx_spin },
705 	{ "mprof lock", &lock_class_mtx_spin },
706 	{ "zombie lock", &lock_class_mtx_spin },
707 	{ "ALD Queue", &lock_class_mtx_spin },
708 #if defined(__i386__) || defined(__amd64__)
709 	{ "pcicfg", &lock_class_mtx_spin },
710 	{ "NDIS thread lock", &lock_class_mtx_spin },
711 #endif
712 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
713 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
714 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
715 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
716 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
717 #ifdef	HWPMC_HOOKS
718 	{ "pmc-leaf", &lock_class_mtx_spin },
719 #endif
720 	{ "blocked lock", &lock_class_mtx_spin },
721 	{ NULL, NULL },
722 	{ NULL, NULL }
723 };
724 
725 /*
726  * Pairs of locks which have been blessed.  Witness does not complain about
727  * order problems with blessed lock pairs.  Please do not add an entry to the
728  * table without an explanatory comment.
729  */
730 static struct witness_blessed blessed_list[] = {
731 	/*
732 	 * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
733 	 * both lock orders, so a deadlock cannot happen as a result of this
734 	 * LOR.
735 	 */
736 	{ "dirhash",	"bufwait" },
737 
738 	/*
739 	 * A UFS vnode may be locked in vget() while a buffer belonging to the
740 	 * parent directory vnode is locked.
741 	 */
742 	{ "ufs",	"bufwait" },
743 };
744 
745 /*
746  * This global is set to 0 once it becomes safe to use the witness code.
747  */
748 static int witness_cold = 1;
749 
750 /*
751  * This global is set to 1 once the static lock orders have been enrolled
752  * so that a warning can be issued for any spin locks enrolled later.
753  */
754 static int witness_spin_warn = 0;
755 
756 /* Trim useless garbage from filenames. */
757 static const char *
758 fixup_filename(const char *file)
759 {
760 
761 	if (file == NULL)
762 		return (NULL);
763 	while (strncmp(file, "../", 3) == 0)
764 		file += 3;
765 	return (file);
766 }
767 
768 /*
769  * Calculate the size of early witness structures.
770  */
771 int
772 witness_startup_count(void)
773 {
774 	int sz;
775 
776 	sz = sizeof(struct witness) * witness_count;
777 	sz += sizeof(*w_rmatrix) * (witness_count + 1);
778 	sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
779 	    (witness_count + 1);
780 
781 	return (sz);
782 }
783 
784 /*
785  * The WITNESS-enabled diagnostic code.  Note that the witness code does
786  * assume that the early boot is single-threaded at least until after this
787  * routine is completed.
788  */
789 void
790 witness_startup(void *mem)
791 {
792 	struct lock_object *lock;
793 	struct witness_order_list_entry *order;
794 	struct witness *w, *w1;
795 	uintptr_t p;
796 	int i;
797 
798 	p = (uintptr_t)mem;
799 	w_data = (void *)p;
800 	p += sizeof(struct witness) * witness_count;
801 
802 	w_rmatrix = (void *)p;
803 	p += sizeof(*w_rmatrix) * (witness_count + 1);
804 
805 	for (i = 0; i < witness_count + 1; i++) {
806 		w_rmatrix[i] = (void *)p;
807 		p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
808 	}
809 	badstack_sbuf_size = witness_count * 256;
810 
811 	/*
812 	 * We have to release Giant before initializing its witness
813 	 * structure so that WITNESS doesn't get confused.
814 	 */
815 	mtx_unlock(&Giant);
816 	mtx_assert(&Giant, MA_NOTOWNED);
817 
818 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
819 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
820 	    MTX_NOWITNESS | MTX_NOPROFILE);
821 	for (i = witness_count - 1; i >= 0; i--) {
822 		w = &w_data[i];
823 		memset(w, 0, sizeof(*w));
824 		w_data[i].w_index = i;	/* Witness index never changes. */
825 		witness_free(w);
826 	}
827 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
828 	    ("%s: Invalid list of free witness objects", __func__));
829 
830 	/* Witness with index 0 is not used to aid in debugging. */
831 	STAILQ_REMOVE_HEAD(&w_free, w_list);
832 	w_free_cnt--;
833 
834 	for (i = 0; i < witness_count; i++) {
835 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
836 		    (witness_count + 1));
837 	}
838 
839 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
840 		witness_lock_list_free(&w_locklistdata[i]);
841 	witness_init_hash_tables();
842 
843 	/* First add in all the specified order lists. */
844 	for (order = order_lists; order->w_name != NULL; order++) {
845 		w = enroll(order->w_name, order->w_class);
846 		if (w == NULL)
847 			continue;
848 		w->w_file = "order list";
849 		for (order++; order->w_name != NULL; order++) {
850 			w1 = enroll(order->w_name, order->w_class);
851 			if (w1 == NULL)
852 				continue;
853 			w1->w_file = "order list";
854 			itismychild(w, w1);
855 			w = w1;
856 		}
857 	}
858 	witness_spin_warn = 1;
859 
860 	/* Iterate through all locks and add them to witness. */
861 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
862 		lock = pending_locks[i].wh_lock;
863 		KASSERT(lock->lo_flags & LO_WITNESS,
864 		    ("%s: lock %s is on pending list but not LO_WITNESS",
865 		    __func__, lock->lo_name));
866 		lock->lo_witness = enroll(pending_locks[i].wh_type,
867 		    LOCK_CLASS(lock));
868 	}
869 
870 	/* Mark the witness code as being ready for use. */
871 	witness_cold = 0;
872 
873 	mtx_lock(&Giant);
874 }
875 
876 void
877 witness_init(struct lock_object *lock, const char *type)
878 {
879 	struct lock_class *class;
880 
881 	/* Various sanity checks. */
882 	class = LOCK_CLASS(lock);
883 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
884 	    (class->lc_flags & LC_RECURSABLE) == 0)
885 		kassert_panic("%s: lock (%s) %s can not be recursable",
886 		    __func__, class->lc_name, lock->lo_name);
887 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
888 	    (class->lc_flags & LC_SLEEPABLE) == 0)
889 		kassert_panic("%s: lock (%s) %s can not be sleepable",
890 		    __func__, class->lc_name, lock->lo_name);
891 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
892 	    (class->lc_flags & LC_UPGRADABLE) == 0)
893 		kassert_panic("%s: lock (%s) %s can not be upgradable",
894 		    __func__, class->lc_name, lock->lo_name);
895 
896 	/*
897 	 * If we shouldn't watch this lock, then just clear lo_witness.
898 	 * Otherwise, if witness_cold is set, then it is too early to
899 	 * enroll this lock, so defer it to witness_initialize() by adding
900 	 * it to the pending_locks list.  If it is not too early, then enroll
901 	 * the lock now.
902 	 */
903 	if (witness_watch < 1 || KERNEL_PANICKED() ||
904 	    (lock->lo_flags & LO_WITNESS) == 0)
905 		lock->lo_witness = NULL;
906 	else if (witness_cold) {
907 		pending_locks[pending_cnt].wh_lock = lock;
908 		pending_locks[pending_cnt++].wh_type = type;
909 		if (pending_cnt > WITNESS_PENDLIST)
910 			panic("%s: pending locks list is too small, "
911 			    "increase WITNESS_PENDLIST\n",
912 			    __func__);
913 	} else
914 		lock->lo_witness = enroll(type, class);
915 }
916 
917 void
918 witness_destroy(struct lock_object *lock)
919 {
920 	struct lock_class *class;
921 	struct witness *w;
922 
923 	class = LOCK_CLASS(lock);
924 
925 	if (witness_cold)
926 		panic("lock (%s) %s destroyed while witness_cold",
927 		    class->lc_name, lock->lo_name);
928 
929 	/* XXX: need to verify that no one holds the lock */
930 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
931 		return;
932 	w = lock->lo_witness;
933 
934 	mtx_lock_spin(&w_mtx);
935 	MPASS(w->w_refcount > 0);
936 	w->w_refcount--;
937 
938 	if (w->w_refcount == 0)
939 		depart(w);
940 	mtx_unlock_spin(&w_mtx);
941 }
942 
943 #ifdef DDB
944 static void
945 witness_ddb_compute_levels(void)
946 {
947 	struct witness *w;
948 
949 	/*
950 	 * First clear all levels.
951 	 */
952 	STAILQ_FOREACH(w, &w_all, w_list)
953 		w->w_ddb_level = -1;
954 
955 	/*
956 	 * Look for locks with no parents and level all their descendants.
957 	 */
958 	STAILQ_FOREACH(w, &w_all, w_list) {
959 		/* If the witness has ancestors (is not a root), skip it. */
960 		if (w->w_num_ancestors > 0)
961 			continue;
962 		witness_ddb_level_descendants(w, 0);
963 	}
964 }
965 
966 static void
967 witness_ddb_level_descendants(struct witness *w, int l)
968 {
969 	int i;
970 
971 	if (w->w_ddb_level >= l)
972 		return;
973 
974 	w->w_ddb_level = l;
975 	l++;
976 
977 	for (i = 1; i <= w_max_used_index; i++) {
978 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
979 			witness_ddb_level_descendants(&w_data[i], l);
980 	}
981 }
982 
983 static void
984 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
985     struct witness *w, int indent)
986 {
987 	int i;
988 
989  	for (i = 0; i < indent; i++)
990  		prnt(" ");
991 	prnt("%s (type: %s, depth: %d, active refs: %d)",
992 	     w->w_name, w->w_class->lc_name,
993 	     w->w_ddb_level, w->w_refcount);
994  	if (w->w_displayed) {
995  		prnt(" -- (already displayed)\n");
996  		return;
997  	}
998  	w->w_displayed = 1;
999 	if (w->w_file != NULL && w->w_line != 0)
1000 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1001 		    w->w_line);
1002 	else
1003 		prnt(" -- never acquired\n");
1004 	indent++;
1005 	WITNESS_INDEX_ASSERT(w->w_index);
1006 	for (i = 1; i <= w_max_used_index; i++) {
1007 		if (db_pager_quit)
1008 			return;
1009 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1010 			witness_ddb_display_descendants(prnt, &w_data[i],
1011 			    indent);
1012 	}
1013 }
1014 
1015 static void
1016 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1017     struct witness_list *list)
1018 {
1019 	struct witness *w;
1020 
1021 	STAILQ_FOREACH(w, list, w_typelist) {
1022 		if (w->w_file == NULL || w->w_ddb_level > 0)
1023 			continue;
1024 
1025 		/* This lock has no anscestors - display its descendants. */
1026 		witness_ddb_display_descendants(prnt, w, 0);
1027 		if (db_pager_quit)
1028 			return;
1029 	}
1030 }
1031 
1032 static void
1033 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1034 {
1035 	struct witness *w;
1036 
1037 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1038 	witness_ddb_compute_levels();
1039 
1040 	/* Clear all the displayed flags. */
1041 	STAILQ_FOREACH(w, &w_all, w_list)
1042 		w->w_displayed = 0;
1043 
1044 	/*
1045 	 * First, handle sleep locks which have been acquired at least
1046 	 * once.
1047 	 */
1048 	prnt("Sleep locks:\n");
1049 	witness_ddb_display_list(prnt, &w_sleep);
1050 	if (db_pager_quit)
1051 		return;
1052 
1053 	/*
1054 	 * Now do spin locks which have been acquired at least once.
1055 	 */
1056 	prnt("\nSpin locks:\n");
1057 	witness_ddb_display_list(prnt, &w_spin);
1058 	if (db_pager_quit)
1059 		return;
1060 
1061 	/*
1062 	 * Finally, any locks which have not been acquired yet.
1063 	 */
1064 	prnt("\nLocks which were never acquired:\n");
1065 	STAILQ_FOREACH(w, &w_all, w_list) {
1066 		if (w->w_file != NULL || w->w_refcount == 0)
1067 			continue;
1068 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1069 		    w->w_class->lc_name, w->w_ddb_level);
1070 		if (db_pager_quit)
1071 			return;
1072 	}
1073 }
1074 #endif /* DDB */
1075 
1076 int
1077 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1078 {
1079 
1080 	if (witness_watch == -1 || KERNEL_PANICKED())
1081 		return (0);
1082 
1083 	/* Require locks that witness knows about. */
1084 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1085 	    lock2->lo_witness == NULL)
1086 		return (EINVAL);
1087 
1088 	mtx_assert(&w_mtx, MA_NOTOWNED);
1089 	mtx_lock_spin(&w_mtx);
1090 
1091 	/*
1092 	 * If we already have either an explicit or implied lock order that
1093 	 * is the other way around, then return an error.
1094 	 */
1095 	if (witness_watch &&
1096 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1097 		mtx_unlock_spin(&w_mtx);
1098 		return (EDOOFUS);
1099 	}
1100 
1101 	/* Try to add the new order. */
1102 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1103 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1104 	itismychild(lock1->lo_witness, lock2->lo_witness);
1105 	mtx_unlock_spin(&w_mtx);
1106 	return (0);
1107 }
1108 
1109 void
1110 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1111     int line, struct lock_object *interlock)
1112 {
1113 	struct lock_list_entry *lock_list, *lle;
1114 	struct lock_instance *lock1, *lock2, *plock;
1115 	struct lock_class *class, *iclass;
1116 	struct witness *w, *w1;
1117 	struct thread *td;
1118 	int i, j;
1119 
1120 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1121 	    KERNEL_PANICKED())
1122 		return;
1123 
1124 	w = lock->lo_witness;
1125 	class = LOCK_CLASS(lock);
1126 	td = curthread;
1127 
1128 	if (class->lc_flags & LC_SLEEPLOCK) {
1129 		/*
1130 		 * Since spin locks include a critical section, this check
1131 		 * implicitly enforces a lock order of all sleep locks before
1132 		 * all spin locks.
1133 		 */
1134 		if (td->td_critnest != 0 && !kdb_active)
1135 			kassert_panic("acquiring blockable sleep lock with "
1136 			    "spinlock or critical section held (%s) %s @ %s:%d",
1137 			    class->lc_name, lock->lo_name,
1138 			    fixup_filename(file), line);
1139 
1140 		/*
1141 		 * If this is the first lock acquired then just return as
1142 		 * no order checking is needed.
1143 		 */
1144 		lock_list = td->td_sleeplocks;
1145 		if (lock_list == NULL || lock_list->ll_count == 0)
1146 			return;
1147 	} else {
1148 		/*
1149 		 * If this is the first lock, just return as no order
1150 		 * checking is needed.  Avoid problems with thread
1151 		 * migration pinning the thread while checking if
1152 		 * spinlocks are held.  If at least one spinlock is held
1153 		 * the thread is in a safe path and it is allowed to
1154 		 * unpin it.
1155 		 */
1156 		sched_pin();
1157 		lock_list = PCPU_GET(spinlocks);
1158 		if (lock_list == NULL || lock_list->ll_count == 0) {
1159 			sched_unpin();
1160 			return;
1161 		}
1162 		sched_unpin();
1163 	}
1164 
1165 	/*
1166 	 * Check to see if we are recursing on a lock we already own.  If
1167 	 * so, make sure that we don't mismatch exclusive and shared lock
1168 	 * acquires.
1169 	 */
1170 	lock1 = find_instance(lock_list, lock);
1171 	if (lock1 != NULL) {
1172 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1173 		    (flags & LOP_EXCLUSIVE) == 0) {
1174 			witness_output("shared lock of (%s) %s @ %s:%d\n",
1175 			    class->lc_name, lock->lo_name,
1176 			    fixup_filename(file), line);
1177 			witness_output("while exclusively locked from %s:%d\n",
1178 			    fixup_filename(lock1->li_file), lock1->li_line);
1179 			kassert_panic("excl->share");
1180 		}
1181 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1182 		    (flags & LOP_EXCLUSIVE) != 0) {
1183 			witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1184 			    class->lc_name, lock->lo_name,
1185 			    fixup_filename(file), line);
1186 			witness_output("while share locked from %s:%d\n",
1187 			    fixup_filename(lock1->li_file), lock1->li_line);
1188 			kassert_panic("share->excl");
1189 		}
1190 		return;
1191 	}
1192 
1193 	/* Warn if the interlock is not locked exactly once. */
1194 	if (interlock != NULL) {
1195 		iclass = LOCK_CLASS(interlock);
1196 		lock1 = find_instance(lock_list, interlock);
1197 		if (lock1 == NULL)
1198 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1199 			    iclass->lc_name, interlock->lo_name,
1200 			    fixup_filename(file), line);
1201 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1202 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1203 			    iclass->lc_name, interlock->lo_name,
1204 			    fixup_filename(file), line);
1205 	}
1206 
1207 	/*
1208 	 * Find the previously acquired lock, but ignore interlocks.
1209 	 */
1210 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1211 	if (interlock != NULL && plock->li_lock == interlock) {
1212 		if (lock_list->ll_count > 1)
1213 			plock =
1214 			    &lock_list->ll_children[lock_list->ll_count - 2];
1215 		else {
1216 			lle = lock_list->ll_next;
1217 
1218 			/*
1219 			 * The interlock is the only lock we hold, so
1220 			 * simply return.
1221 			 */
1222 			if (lle == NULL)
1223 				return;
1224 			plock = &lle->ll_children[lle->ll_count - 1];
1225 		}
1226 	}
1227 
1228 	/*
1229 	 * Try to perform most checks without a lock.  If this succeeds we
1230 	 * can skip acquiring the lock and return success.  Otherwise we redo
1231 	 * the check with the lock held to handle races with concurrent updates.
1232 	 */
1233 	w1 = plock->li_lock->lo_witness;
1234 	if (witness_lock_order_check(w1, w))
1235 		return;
1236 
1237 	mtx_lock_spin(&w_mtx);
1238 	if (witness_lock_order_check(w1, w)) {
1239 		mtx_unlock_spin(&w_mtx);
1240 		return;
1241 	}
1242 	witness_lock_order_add(w1, w);
1243 
1244 	/*
1245 	 * Check for duplicate locks of the same type.  Note that we only
1246 	 * have to check for this on the last lock we just acquired.  Any
1247 	 * other cases will be caught as lock order violations.
1248 	 */
1249 	if (w1 == w) {
1250 		i = w->w_index;
1251 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1252 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1253 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1254 			w->w_reversed = 1;
1255 			mtx_unlock_spin(&w_mtx);
1256 			witness_output(
1257 			    "acquiring duplicate lock of same type: \"%s\"\n",
1258 			    w->w_name);
1259 			witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1260 			    fixup_filename(plock->li_file), plock->li_line);
1261 			witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1262 			    fixup_filename(file), line);
1263 			witness_debugger(1, __func__);
1264 		} else
1265 			mtx_unlock_spin(&w_mtx);
1266 		return;
1267 	}
1268 	mtx_assert(&w_mtx, MA_OWNED);
1269 
1270 	/*
1271 	 * If we know that the lock we are acquiring comes after
1272 	 * the lock we most recently acquired in the lock order tree,
1273 	 * then there is no need for any further checks.
1274 	 */
1275 	if (isitmychild(w1, w))
1276 		goto out;
1277 
1278 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1279 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1280 			struct stack pstack;
1281 			bool pstackv, trace;
1282 
1283 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1284 			lock1 = &lle->ll_children[i];
1285 
1286 			/*
1287 			 * Ignore the interlock.
1288 			 */
1289 			if (interlock == lock1->li_lock)
1290 				continue;
1291 
1292 			/*
1293 			 * If this lock doesn't undergo witness checking,
1294 			 * then skip it.
1295 			 */
1296 			w1 = lock1->li_lock->lo_witness;
1297 			if (w1 == NULL) {
1298 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1299 				    ("lock missing witness structure"));
1300 				continue;
1301 			}
1302 
1303 			/*
1304 			 * If we are locking Giant and this is a sleepable
1305 			 * lock, then skip it.
1306 			 */
1307 			if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1308 			    lock == &Giant.lock_object)
1309 				continue;
1310 
1311 			/*
1312 			 * If we are locking a sleepable lock and this lock
1313 			 * is Giant, then skip it.
1314 			 */
1315 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1316 			    (flags & LOP_NOSLEEP) == 0 &&
1317 			    lock1->li_lock == &Giant.lock_object)
1318 				continue;
1319 
1320 			/*
1321 			 * If we are locking a sleepable lock and this lock
1322 			 * isn't sleepable, we want to treat it as a lock
1323 			 * order violation to enfore a general lock order of
1324 			 * sleepable locks before non-sleepable locks.
1325 			 */
1326 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1327 			    (flags & LOP_NOSLEEP) == 0 &&
1328 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1329 				goto reversal;
1330 
1331 			/*
1332 			 * If we are locking Giant and this is a non-sleepable
1333 			 * lock, then treat it as a reversal.
1334 			 */
1335 			if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1336 			    lock == &Giant.lock_object)
1337 				goto reversal;
1338 
1339 			/*
1340 			 * Check the lock order hierarchy for a reveresal.
1341 			 */
1342 			if (!isitmydescendant(w, w1))
1343 				continue;
1344 		reversal:
1345 
1346 			/*
1347 			 * We have a lock order violation, check to see if it
1348 			 * is allowed or has already been yelled about.
1349 			 */
1350 
1351 			/* Bail if this violation is known */
1352 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1353 				goto out;
1354 
1355 			/* Record this as a violation */
1356 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358 			w->w_reversed = w1->w_reversed = 1;
1359 			witness_increment_graph_generation();
1360 
1361 			/*
1362 			 * If the lock order is blessed, bail before logging
1363 			 * anything.  We don't look for other lock order
1364 			 * violations though, which may be a bug.
1365 			 */
1366 			if (blessed(w, w1))
1367 				goto out;
1368 
1369 			trace = atomic_load_int(&witness_trace);
1370 			if (trace) {
1371 				struct witness_lock_order_data *data;
1372 
1373 				pstackv = false;
1374 				data = witness_lock_order_get(w, w1);
1375 				if (data != NULL) {
1376 					stack_copy(&data->wlod_stack,
1377 					    &pstack);
1378 					pstackv = true;
1379 				}
1380 			}
1381 			mtx_unlock_spin(&w_mtx);
1382 
1383 #ifdef WITNESS_NO_VNODE
1384 			/*
1385 			 * There are known LORs between VNODE locks. They are
1386 			 * not an indication of a bug. VNODE locks are flagged
1387 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1388 			 * is between 2 VNODE locks.
1389 			 */
1390 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1391 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1392 				return;
1393 #endif
1394 
1395 			/*
1396 			 * Ok, yell about it.
1397 			 */
1398 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1399 			    (flags & LOP_NOSLEEP) == 0 &&
1400 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1401 				witness_output(
1402 		"lock order reversal: (sleepable after non-sleepable)\n");
1403 			else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1404 			    && lock == &Giant.lock_object)
1405 				witness_output(
1406 		"lock order reversal: (Giant after non-sleepable)\n");
1407 			else
1408 				witness_output("lock order reversal:\n");
1409 
1410 			/*
1411 			 * Try to locate an earlier lock with
1412 			 * witness w in our list.
1413 			 */
1414 			do {
1415 				lock2 = &lle->ll_children[i];
1416 				MPASS(lock2->li_lock != NULL);
1417 				if (lock2->li_lock->lo_witness == w)
1418 					break;
1419 				if (i == 0 && lle->ll_next != NULL) {
1420 					lle = lle->ll_next;
1421 					i = lle->ll_count - 1;
1422 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1423 				} else
1424 					i--;
1425 			} while (i >= 0);
1426 			if (i < 0) {
1427 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1428 				    lock1->li_lock, lock1->li_lock->lo_name,
1429 				    w1->w_name, w1->w_class->lc_name,
1430 				    fixup_filename(lock1->li_file),
1431 				    lock1->li_line);
1432 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1433 				    lock, lock->lo_name, w->w_name,
1434 				    w->w_class->lc_name, fixup_filename(file),
1435 				    line);
1436 			} else {
1437 				struct witness *w2 = lock2->li_lock->lo_witness;
1438 
1439 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1440 				    lock2->li_lock, lock2->li_lock->lo_name,
1441 				    w2->w_name, w2->w_class->lc_name,
1442 				    fixup_filename(lock2->li_file),
1443 				    lock2->li_line);
1444 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1445 				    lock1->li_lock, lock1->li_lock->lo_name,
1446 				    w1->w_name, w1->w_class->lc_name,
1447 				    fixup_filename(lock1->li_file),
1448 				    lock1->li_line);
1449 				witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1450 				    lock->lo_name, w->w_name,
1451 				    w->w_class->lc_name, fixup_filename(file),
1452 				    line);
1453 			}
1454 			if (trace) {
1455 				char buf[64];
1456 				struct sbuf sb;
1457 
1458 				sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1459 				sbuf_set_drain(&sb, witness_output_drain,
1460 				    NULL);
1461 
1462 				if (pstackv) {
1463 					sbuf_printf(&sb,
1464 				    "lock order %s -> %s established at:\n",
1465 					    w->w_name, w1->w_name);
1466 					stack_sbuf_print_flags(&sb, &pstack,
1467 					    M_NOWAIT, STACK_SBUF_FMT_LONG);
1468 				}
1469 
1470 				sbuf_printf(&sb,
1471 				    "lock order %s -> %s attempted at:\n",
1472 				    w1->w_name, w->w_name);
1473 				stack_save(&pstack);
1474 				stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1475 				    STACK_SBUF_FMT_LONG);
1476 
1477 				sbuf_finish(&sb);
1478 				sbuf_delete(&sb);
1479 			}
1480 			witness_enter_debugger(__func__);
1481 			return;
1482 		}
1483 	}
1484 
1485 	/*
1486 	 * If requested, build a new lock order.  However, don't build a new
1487 	 * relationship between a sleepable lock and Giant if it is in the
1488 	 * wrong direction.  The correct lock order is that sleepable locks
1489 	 * always come before Giant.
1490 	 */
1491 	if (flags & LOP_NEWORDER &&
1492 	    !(plock->li_lock == &Giant.lock_object &&
1493 	    (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1494 	    (flags & LOP_NOSLEEP) == 0)) {
1495 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1496 		    w->w_name, plock->li_lock->lo_witness->w_name);
1497 		itismychild(plock->li_lock->lo_witness, w);
1498 	}
1499 out:
1500 	mtx_unlock_spin(&w_mtx);
1501 }
1502 
1503 void
1504 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1505 {
1506 	struct lock_list_entry **lock_list, *lle;
1507 	struct lock_instance *instance;
1508 	struct witness *w;
1509 	struct thread *td;
1510 
1511 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1512 	    KERNEL_PANICKED())
1513 		return;
1514 	w = lock->lo_witness;
1515 	td = curthread;
1516 
1517 	/* Determine lock list for this lock. */
1518 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1519 		lock_list = &td->td_sleeplocks;
1520 	else
1521 		lock_list = PCPU_PTR(spinlocks);
1522 
1523 	/* Check to see if we are recursing on a lock we already own. */
1524 	instance = find_instance(*lock_list, lock);
1525 	if (instance != NULL) {
1526 		instance->li_flags++;
1527 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1528 		    td->td_proc->p_pid, lock->lo_name,
1529 		    instance->li_flags & LI_RECURSEMASK);
1530 		instance->li_file = file;
1531 		instance->li_line = line;
1532 		return;
1533 	}
1534 
1535 	/* Update per-witness last file and line acquire. */
1536 	w->w_file = file;
1537 	w->w_line = line;
1538 
1539 	/* Find the next open lock instance in the list and fill it. */
1540 	lle = *lock_list;
1541 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1542 		lle = witness_lock_list_get();
1543 		if (lle == NULL)
1544 			return;
1545 		lle->ll_next = *lock_list;
1546 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1547 		    td->td_proc->p_pid, lle);
1548 		*lock_list = lle;
1549 	}
1550 	instance = &lle->ll_children[lle->ll_count++];
1551 	instance->li_lock = lock;
1552 	instance->li_line = line;
1553 	instance->li_file = file;
1554 	instance->li_flags = 0;
1555 	if ((flags & LOP_EXCLUSIVE) != 0)
1556 		instance->li_flags |= LI_EXCLUSIVE;
1557 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1558 		instance->li_flags |= LI_SLEEPABLE;
1559 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1560 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1561 }
1562 
1563 void
1564 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1565 {
1566 	struct lock_instance *instance;
1567 	struct lock_class *class;
1568 
1569 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1570 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1571 		return;
1572 	class = LOCK_CLASS(lock);
1573 	if (witness_watch) {
1574 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1575 			kassert_panic(
1576 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1577 			    class->lc_name, lock->lo_name,
1578 			    fixup_filename(file), line);
1579 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1580 			kassert_panic(
1581 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1582 			    class->lc_name, lock->lo_name,
1583 			    fixup_filename(file), line);
1584 	}
1585 	instance = find_instance(curthread->td_sleeplocks, lock);
1586 	if (instance == NULL) {
1587 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1588 		    class->lc_name, lock->lo_name,
1589 		    fixup_filename(file), line);
1590 		return;
1591 	}
1592 	if (witness_watch) {
1593 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1594 			kassert_panic(
1595 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1596 			    class->lc_name, lock->lo_name,
1597 			    fixup_filename(file), line);
1598 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1599 			kassert_panic(
1600 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1601 			    class->lc_name, lock->lo_name,
1602 			    instance->li_flags & LI_RECURSEMASK,
1603 			    fixup_filename(file), line);
1604 	}
1605 	instance->li_flags |= LI_EXCLUSIVE;
1606 }
1607 
1608 void
1609 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1610     int line)
1611 {
1612 	struct lock_instance *instance;
1613 	struct lock_class *class;
1614 
1615 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1616 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1617 		return;
1618 	class = LOCK_CLASS(lock);
1619 	if (witness_watch) {
1620 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1621 			kassert_panic(
1622 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1623 			    class->lc_name, lock->lo_name,
1624 			    fixup_filename(file), line);
1625 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1626 			kassert_panic(
1627 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1628 			    class->lc_name, lock->lo_name,
1629 			    fixup_filename(file), line);
1630 	}
1631 	instance = find_instance(curthread->td_sleeplocks, lock);
1632 	if (instance == NULL) {
1633 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1634 		    class->lc_name, lock->lo_name,
1635 		    fixup_filename(file), line);
1636 		return;
1637 	}
1638 	if (witness_watch) {
1639 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1640 			kassert_panic(
1641 			    "downgrade of shared lock (%s) %s @ %s:%d",
1642 			    class->lc_name, lock->lo_name,
1643 			    fixup_filename(file), line);
1644 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1645 			kassert_panic(
1646 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1647 			    class->lc_name, lock->lo_name,
1648 			    instance->li_flags & LI_RECURSEMASK,
1649 			    fixup_filename(file), line);
1650 	}
1651 	instance->li_flags &= ~LI_EXCLUSIVE;
1652 }
1653 
1654 void
1655 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1656 {
1657 	struct lock_list_entry **lock_list, *lle;
1658 	struct lock_instance *instance;
1659 	struct lock_class *class;
1660 	struct thread *td;
1661 	register_t s;
1662 	int i, j;
1663 
1664 	if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1665 		return;
1666 	td = curthread;
1667 	class = LOCK_CLASS(lock);
1668 
1669 	/* Find lock instance associated with this lock. */
1670 	if (class->lc_flags & LC_SLEEPLOCK)
1671 		lock_list = &td->td_sleeplocks;
1672 	else
1673 		lock_list = PCPU_PTR(spinlocks);
1674 	lle = *lock_list;
1675 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1676 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1677 			instance = &(*lock_list)->ll_children[i];
1678 			if (instance->li_lock == lock)
1679 				goto found;
1680 		}
1681 
1682 	/*
1683 	 * When disabling WITNESS through witness_watch we could end up in
1684 	 * having registered locks in the td_sleeplocks queue.
1685 	 * We have to make sure we flush these queues, so just search for
1686 	 * eventual register locks and remove them.
1687 	 */
1688 	if (witness_watch > 0) {
1689 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1690 		    lock->lo_name, fixup_filename(file), line);
1691 		return;
1692 	} else {
1693 		return;
1694 	}
1695 found:
1696 
1697 	/* First, check for shared/exclusive mismatches. */
1698 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1699 	    (flags & LOP_EXCLUSIVE) == 0) {
1700 		witness_output("shared unlock of (%s) %s @ %s:%d\n",
1701 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1702 		witness_output("while exclusively locked from %s:%d\n",
1703 		    fixup_filename(instance->li_file), instance->li_line);
1704 		kassert_panic("excl->ushare");
1705 	}
1706 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1707 	    (flags & LOP_EXCLUSIVE) != 0) {
1708 		witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1709 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1710 		witness_output("while share locked from %s:%d\n",
1711 		    fixup_filename(instance->li_file),
1712 		    instance->li_line);
1713 		kassert_panic("share->uexcl");
1714 	}
1715 	/* If we are recursed, unrecurse. */
1716 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1717 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1718 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1719 		    instance->li_flags);
1720 		instance->li_flags--;
1721 		return;
1722 	}
1723 	/* The lock is now being dropped, check for NORELEASE flag */
1724 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1725 		witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1726 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1727 		kassert_panic("lock marked norelease");
1728 	}
1729 
1730 	/* Otherwise, remove this item from the list. */
1731 	s = intr_disable();
1732 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1733 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1734 	    (*lock_list)->ll_count - 1);
1735 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1736 		(*lock_list)->ll_children[j] =
1737 		    (*lock_list)->ll_children[j + 1];
1738 	(*lock_list)->ll_count--;
1739 	intr_restore(s);
1740 
1741 	/*
1742 	 * In order to reduce contention on w_mtx, we want to keep always an
1743 	 * head object into lists so that frequent allocation from the
1744 	 * free witness pool (and subsequent locking) is avoided.
1745 	 * In order to maintain the current code simple, when the head
1746 	 * object is totally unloaded it means also that we do not have
1747 	 * further objects in the list, so the list ownership needs to be
1748 	 * hand over to another object if the current head needs to be freed.
1749 	 */
1750 	if ((*lock_list)->ll_count == 0) {
1751 		if (*lock_list == lle) {
1752 			if (lle->ll_next == NULL)
1753 				return;
1754 		} else
1755 			lle = *lock_list;
1756 		*lock_list = lle->ll_next;
1757 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1758 		    td->td_proc->p_pid, lle);
1759 		witness_lock_list_free(lle);
1760 	}
1761 }
1762 
1763 void
1764 witness_thread_exit(struct thread *td)
1765 {
1766 	struct lock_list_entry *lle;
1767 	int i, n;
1768 
1769 	lle = td->td_sleeplocks;
1770 	if (lle == NULL || KERNEL_PANICKED())
1771 		return;
1772 	if (lle->ll_count != 0) {
1773 		for (n = 0; lle != NULL; lle = lle->ll_next)
1774 			for (i = lle->ll_count - 1; i >= 0; i--) {
1775 				if (n == 0)
1776 					witness_output(
1777 		    "Thread %p exiting with the following locks held:\n", td);
1778 				n++;
1779 				witness_list_lock(&lle->ll_children[i],
1780 				    witness_output);
1781 
1782 			}
1783 		kassert_panic(
1784 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1785 	}
1786 	witness_lock_list_free(lle);
1787 }
1788 
1789 /*
1790  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1791  * exempt Giant and sleepable locks from the checks as well.  If any
1792  * non-exempt locks are held, then a supplied message is printed to the
1793  * output channel along with a list of the offending locks.  If indicated in the
1794  * flags then a failure results in a panic as well.
1795  */
1796 int
1797 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1798 {
1799 	struct lock_list_entry *lock_list, *lle;
1800 	struct lock_instance *lock1;
1801 	struct thread *td;
1802 	va_list ap;
1803 	int i, n;
1804 
1805 	if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1806 		return (0);
1807 	n = 0;
1808 	td = curthread;
1809 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1810 		for (i = lle->ll_count - 1; i >= 0; i--) {
1811 			lock1 = &lle->ll_children[i];
1812 			if (lock1->li_lock == lock)
1813 				continue;
1814 			if (flags & WARN_GIANTOK &&
1815 			    lock1->li_lock == &Giant.lock_object)
1816 				continue;
1817 			if (flags & WARN_SLEEPOK &&
1818 			    (lock1->li_flags & LI_SLEEPABLE) != 0)
1819 				continue;
1820 			if (n == 0) {
1821 				va_start(ap, fmt);
1822 				vprintf(fmt, ap);
1823 				va_end(ap);
1824 				printf(" with the following %slocks held:\n",
1825 				    (flags & WARN_SLEEPOK) != 0 ?
1826 				    "non-sleepable " : "");
1827 			}
1828 			n++;
1829 			witness_list_lock(lock1, printf);
1830 		}
1831 
1832 	/*
1833 	 * Pin the thread in order to avoid problems with thread migration.
1834 	 * Once that all verifies are passed about spinlocks ownership,
1835 	 * the thread is in a safe path and it can be unpinned.
1836 	 */
1837 	sched_pin();
1838 	lock_list = PCPU_GET(spinlocks);
1839 	if (lock_list != NULL && lock_list->ll_count != 0) {
1840 		sched_unpin();
1841 
1842 		/*
1843 		 * We should only have one spinlock and as long as
1844 		 * the flags cannot match for this locks class,
1845 		 * check if the first spinlock is the one curthread
1846 		 * should hold.
1847 		 */
1848 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1849 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1850 		    lock1->li_lock == lock && n == 0)
1851 			return (0);
1852 
1853 		va_start(ap, fmt);
1854 		vprintf(fmt, ap);
1855 		va_end(ap);
1856 		printf(" with the following %slocks held:\n",
1857 		    (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1858 		n += witness_list_locks(&lock_list, printf);
1859 	} else
1860 		sched_unpin();
1861 	if (flags & WARN_PANIC && n)
1862 		kassert_panic("%s", __func__);
1863 	else
1864 		witness_debugger(n, __func__);
1865 	return (n);
1866 }
1867 
1868 const char *
1869 witness_file(struct lock_object *lock)
1870 {
1871 	struct witness *w;
1872 
1873 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1874 		return ("?");
1875 	w = lock->lo_witness;
1876 	return (w->w_file);
1877 }
1878 
1879 int
1880 witness_line(struct lock_object *lock)
1881 {
1882 	struct witness *w;
1883 
1884 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1885 		return (0);
1886 	w = lock->lo_witness;
1887 	return (w->w_line);
1888 }
1889 
1890 static struct witness *
1891 enroll(const char *description, struct lock_class *lock_class)
1892 {
1893 	struct witness *w;
1894 
1895 	MPASS(description != NULL);
1896 
1897 	if (witness_watch == -1 || KERNEL_PANICKED())
1898 		return (NULL);
1899 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1900 		if (witness_skipspin)
1901 			return (NULL);
1902 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1903 		kassert_panic("lock class %s is not sleep or spin",
1904 		    lock_class->lc_name);
1905 		return (NULL);
1906 	}
1907 
1908 	mtx_lock_spin(&w_mtx);
1909 	w = witness_hash_get(description);
1910 	if (w)
1911 		goto found;
1912 	if ((w = witness_get()) == NULL)
1913 		return (NULL);
1914 	MPASS(strlen(description) < MAX_W_NAME);
1915 	strcpy(w->w_name, description);
1916 	w->w_class = lock_class;
1917 	w->w_refcount = 1;
1918 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1919 	if (lock_class->lc_flags & LC_SPINLOCK) {
1920 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1921 		w_spin_cnt++;
1922 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1923 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1924 		w_sleep_cnt++;
1925 	}
1926 
1927 	/* Insert new witness into the hash */
1928 	witness_hash_put(w);
1929 	witness_increment_graph_generation();
1930 	mtx_unlock_spin(&w_mtx);
1931 	return (w);
1932 found:
1933 	w->w_refcount++;
1934 	if (w->w_refcount == 1)
1935 		w->w_class = lock_class;
1936 	mtx_unlock_spin(&w_mtx);
1937 	if (lock_class != w->w_class)
1938 		kassert_panic(
1939 		    "lock (%s) %s does not match earlier (%s) lock",
1940 		    description, lock_class->lc_name,
1941 		    w->w_class->lc_name);
1942 	return (w);
1943 }
1944 
1945 static void
1946 depart(struct witness *w)
1947 {
1948 
1949 	MPASS(w->w_refcount == 0);
1950 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1951 		w_sleep_cnt--;
1952 	} else {
1953 		w_spin_cnt--;
1954 	}
1955 	/*
1956 	 * Set file to NULL as it may point into a loadable module.
1957 	 */
1958 	w->w_file = NULL;
1959 	w->w_line = 0;
1960 	witness_increment_graph_generation();
1961 }
1962 
1963 static void
1964 adopt(struct witness *parent, struct witness *child)
1965 {
1966 	int pi, ci, i, j;
1967 
1968 	if (witness_cold == 0)
1969 		mtx_assert(&w_mtx, MA_OWNED);
1970 
1971 	/* If the relationship is already known, there's no work to be done. */
1972 	if (isitmychild(parent, child))
1973 		return;
1974 
1975 	/* When the structure of the graph changes, bump up the generation. */
1976 	witness_increment_graph_generation();
1977 
1978 	/*
1979 	 * The hard part ... create the direct relationship, then propagate all
1980 	 * indirect relationships.
1981 	 */
1982 	pi = parent->w_index;
1983 	ci = child->w_index;
1984 	WITNESS_INDEX_ASSERT(pi);
1985 	WITNESS_INDEX_ASSERT(ci);
1986 	MPASS(pi != ci);
1987 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1988 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1989 
1990 	/*
1991 	 * If parent was not already an ancestor of child,
1992 	 * then we increment the descendant and ancestor counters.
1993 	 */
1994 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1995 		parent->w_num_descendants++;
1996 		child->w_num_ancestors++;
1997 	}
1998 
1999 	/*
2000 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2001 	 * an ancestor of 'pi' during this loop.
2002 	 */
2003 	for (i = 1; i <= w_max_used_index; i++) {
2004 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2005 		    (i != pi))
2006 			continue;
2007 
2008 		/* Find each descendant of 'i' and mark it as a descendant. */
2009 		for (j = 1; j <= w_max_used_index; j++) {
2010 			/*
2011 			 * Skip children that are already marked as
2012 			 * descendants of 'i'.
2013 			 */
2014 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2015 				continue;
2016 
2017 			/*
2018 			 * We are only interested in descendants of 'ci'. Note
2019 			 * that 'ci' itself is counted as a descendant of 'ci'.
2020 			 */
2021 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2022 			    (j != ci))
2023 				continue;
2024 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2025 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2026 			w_data[i].w_num_descendants++;
2027 			w_data[j].w_num_ancestors++;
2028 
2029 			/*
2030 			 * Make sure we aren't marking a node as both an
2031 			 * ancestor and descendant. We should have caught
2032 			 * this as a lock order reversal earlier.
2033 			 */
2034 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2035 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2036 				printf("witness rmatrix paradox! [%d][%d]=%d "
2037 				    "both ancestor and descendant\n",
2038 				    i, j, w_rmatrix[i][j]);
2039 				kdb_backtrace();
2040 				printf("Witness disabled.\n");
2041 				witness_watch = -1;
2042 			}
2043 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2044 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2045 				printf("witness rmatrix paradox! [%d][%d]=%d "
2046 				    "both ancestor and descendant\n",
2047 				    j, i, w_rmatrix[j][i]);
2048 				kdb_backtrace();
2049 				printf("Witness disabled.\n");
2050 				witness_watch = -1;
2051 			}
2052 		}
2053 	}
2054 }
2055 
2056 static void
2057 itismychild(struct witness *parent, struct witness *child)
2058 {
2059 	int unlocked;
2060 
2061 	MPASS(child != NULL && parent != NULL);
2062 	if (witness_cold == 0)
2063 		mtx_assert(&w_mtx, MA_OWNED);
2064 
2065 	if (!witness_lock_type_equal(parent, child)) {
2066 		if (witness_cold == 0) {
2067 			unlocked = 1;
2068 			mtx_unlock_spin(&w_mtx);
2069 		} else {
2070 			unlocked = 0;
2071 		}
2072 		kassert_panic(
2073 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2074 		    "the same lock type", __func__, parent->w_name,
2075 		    parent->w_class->lc_name, child->w_name,
2076 		    child->w_class->lc_name);
2077 		if (unlocked)
2078 			mtx_lock_spin(&w_mtx);
2079 	}
2080 	adopt(parent, child);
2081 }
2082 
2083 /*
2084  * Generic code for the isitmy*() functions. The rmask parameter is the
2085  * expected relationship of w1 to w2.
2086  */
2087 static int
2088 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2089 {
2090 	unsigned char r1, r2;
2091 	int i1, i2;
2092 
2093 	i1 = w1->w_index;
2094 	i2 = w2->w_index;
2095 	WITNESS_INDEX_ASSERT(i1);
2096 	WITNESS_INDEX_ASSERT(i2);
2097 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2098 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2099 
2100 	/* The flags on one better be the inverse of the flags on the other */
2101 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2102 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2103 		/* Don't squawk if we're potentially racing with an update. */
2104 		if (!mtx_owned(&w_mtx))
2105 			return (0);
2106 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2107 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2108 		    "w_rmatrix[%d][%d] == %hhx\n",
2109 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2110 		    i2, i1, r2);
2111 		kdb_backtrace();
2112 		printf("Witness disabled.\n");
2113 		witness_watch = -1;
2114 	}
2115 	return (r1 & rmask);
2116 }
2117 
2118 /*
2119  * Checks if @child is a direct child of @parent.
2120  */
2121 static int
2122 isitmychild(struct witness *parent, struct witness *child)
2123 {
2124 
2125 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2126 }
2127 
2128 /*
2129  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2130  */
2131 static int
2132 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2133 {
2134 
2135 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2136 	    __func__));
2137 }
2138 
2139 static int
2140 blessed(struct witness *w1, struct witness *w2)
2141 {
2142 	int i;
2143 	struct witness_blessed *b;
2144 
2145 	for (i = 0; i < nitems(blessed_list); i++) {
2146 		b = &blessed_list[i];
2147 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2148 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2149 				return (1);
2150 			continue;
2151 		}
2152 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2153 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2154 				return (1);
2155 	}
2156 	return (0);
2157 }
2158 
2159 static struct witness *
2160 witness_get(void)
2161 {
2162 	struct witness *w;
2163 	int index;
2164 
2165 	if (witness_cold == 0)
2166 		mtx_assert(&w_mtx, MA_OWNED);
2167 
2168 	if (witness_watch == -1) {
2169 		mtx_unlock_spin(&w_mtx);
2170 		return (NULL);
2171 	}
2172 	if (STAILQ_EMPTY(&w_free)) {
2173 		witness_watch = -1;
2174 		mtx_unlock_spin(&w_mtx);
2175 		printf("WITNESS: unable to allocate a new witness object\n");
2176 		return (NULL);
2177 	}
2178 	w = STAILQ_FIRST(&w_free);
2179 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2180 	w_free_cnt--;
2181 	index = w->w_index;
2182 	MPASS(index > 0 && index == w_max_used_index+1 &&
2183 	    index < witness_count);
2184 	bzero(w, sizeof(*w));
2185 	w->w_index = index;
2186 	if (index > w_max_used_index)
2187 		w_max_used_index = index;
2188 	return (w);
2189 }
2190 
2191 static void
2192 witness_free(struct witness *w)
2193 {
2194 
2195 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2196 	w_free_cnt++;
2197 }
2198 
2199 static struct lock_list_entry *
2200 witness_lock_list_get(void)
2201 {
2202 	struct lock_list_entry *lle;
2203 
2204 	if (witness_watch == -1)
2205 		return (NULL);
2206 	mtx_lock_spin(&w_mtx);
2207 	lle = w_lock_list_free;
2208 	if (lle == NULL) {
2209 		witness_watch = -1;
2210 		mtx_unlock_spin(&w_mtx);
2211 		printf("%s: witness exhausted\n", __func__);
2212 		return (NULL);
2213 	}
2214 	w_lock_list_free = lle->ll_next;
2215 	mtx_unlock_spin(&w_mtx);
2216 	bzero(lle, sizeof(*lle));
2217 	return (lle);
2218 }
2219 
2220 static void
2221 witness_lock_list_free(struct lock_list_entry *lle)
2222 {
2223 
2224 	mtx_lock_spin(&w_mtx);
2225 	lle->ll_next = w_lock_list_free;
2226 	w_lock_list_free = lle;
2227 	mtx_unlock_spin(&w_mtx);
2228 }
2229 
2230 static struct lock_instance *
2231 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2232 {
2233 	struct lock_list_entry *lle;
2234 	struct lock_instance *instance;
2235 	int i;
2236 
2237 	for (lle = list; lle != NULL; lle = lle->ll_next)
2238 		for (i = lle->ll_count - 1; i >= 0; i--) {
2239 			instance = &lle->ll_children[i];
2240 			if (instance->li_lock == lock)
2241 				return (instance);
2242 		}
2243 	return (NULL);
2244 }
2245 
2246 static void
2247 witness_list_lock(struct lock_instance *instance,
2248     int (*prnt)(const char *fmt, ...))
2249 {
2250 	struct lock_object *lock;
2251 
2252 	lock = instance->li_lock;
2253 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2254 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2255 	if (lock->lo_witness->w_name != lock->lo_name)
2256 		prnt(" (%s)", lock->lo_witness->w_name);
2257 	prnt(" r = %d (%p) locked @ %s:%d\n",
2258 	    instance->li_flags & LI_RECURSEMASK, lock,
2259 	    fixup_filename(instance->li_file), instance->li_line);
2260 }
2261 
2262 static int
2263 witness_output(const char *fmt, ...)
2264 {
2265 	va_list ap;
2266 	int ret;
2267 
2268 	va_start(ap, fmt);
2269 	ret = witness_voutput(fmt, ap);
2270 	va_end(ap);
2271 	return (ret);
2272 }
2273 
2274 static int
2275 witness_voutput(const char *fmt, va_list ap)
2276 {
2277 	int ret;
2278 
2279 	ret = 0;
2280 	switch (witness_channel) {
2281 	case WITNESS_CONSOLE:
2282 		ret = vprintf(fmt, ap);
2283 		break;
2284 	case WITNESS_LOG:
2285 		vlog(LOG_NOTICE, fmt, ap);
2286 		break;
2287 	case WITNESS_NONE:
2288 		break;
2289 	}
2290 	return (ret);
2291 }
2292 
2293 #ifdef DDB
2294 static int
2295 witness_thread_has_locks(struct thread *td)
2296 {
2297 
2298 	if (td->td_sleeplocks == NULL)
2299 		return (0);
2300 	return (td->td_sleeplocks->ll_count != 0);
2301 }
2302 
2303 static int
2304 witness_proc_has_locks(struct proc *p)
2305 {
2306 	struct thread *td;
2307 
2308 	FOREACH_THREAD_IN_PROC(p, td) {
2309 		if (witness_thread_has_locks(td))
2310 			return (1);
2311 	}
2312 	return (0);
2313 }
2314 #endif
2315 
2316 int
2317 witness_list_locks(struct lock_list_entry **lock_list,
2318     int (*prnt)(const char *fmt, ...))
2319 {
2320 	struct lock_list_entry *lle;
2321 	int i, nheld;
2322 
2323 	nheld = 0;
2324 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2325 		for (i = lle->ll_count - 1; i >= 0; i--) {
2326 			witness_list_lock(&lle->ll_children[i], prnt);
2327 			nheld++;
2328 		}
2329 	return (nheld);
2330 }
2331 
2332 /*
2333  * This is a bit risky at best.  We call this function when we have timed
2334  * out acquiring a spin lock, and we assume that the other CPU is stuck
2335  * with this lock held.  So, we go groveling around in the other CPU's
2336  * per-cpu data to try to find the lock instance for this spin lock to
2337  * see when it was last acquired.
2338  */
2339 void
2340 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2341     int (*prnt)(const char *fmt, ...))
2342 {
2343 	struct lock_instance *instance;
2344 	struct pcpu *pc;
2345 
2346 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2347 		return;
2348 	pc = pcpu_find(owner->td_oncpu);
2349 	instance = find_instance(pc->pc_spinlocks, lock);
2350 	if (instance != NULL)
2351 		witness_list_lock(instance, prnt);
2352 }
2353 
2354 void
2355 witness_save(struct lock_object *lock, const char **filep, int *linep)
2356 {
2357 	struct lock_list_entry *lock_list;
2358 	struct lock_instance *instance;
2359 	struct lock_class *class;
2360 
2361 	/*
2362 	 * This function is used independently in locking code to deal with
2363 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2364 	 * is gone.
2365 	 */
2366 	if (SCHEDULER_STOPPED())
2367 		return;
2368 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2369 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2370 		return;
2371 	class = LOCK_CLASS(lock);
2372 	if (class->lc_flags & LC_SLEEPLOCK)
2373 		lock_list = curthread->td_sleeplocks;
2374 	else {
2375 		if (witness_skipspin)
2376 			return;
2377 		lock_list = PCPU_GET(spinlocks);
2378 	}
2379 	instance = find_instance(lock_list, lock);
2380 	if (instance == NULL) {
2381 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2382 		    class->lc_name, lock->lo_name);
2383 		return;
2384 	}
2385 	*filep = instance->li_file;
2386 	*linep = instance->li_line;
2387 }
2388 
2389 void
2390 witness_restore(struct lock_object *lock, const char *file, int line)
2391 {
2392 	struct lock_list_entry *lock_list;
2393 	struct lock_instance *instance;
2394 	struct lock_class *class;
2395 
2396 	/*
2397 	 * This function is used independently in locking code to deal with
2398 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2399 	 * is gone.
2400 	 */
2401 	if (SCHEDULER_STOPPED())
2402 		return;
2403 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2404 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2405 		return;
2406 	class = LOCK_CLASS(lock);
2407 	if (class->lc_flags & LC_SLEEPLOCK)
2408 		lock_list = curthread->td_sleeplocks;
2409 	else {
2410 		if (witness_skipspin)
2411 			return;
2412 		lock_list = PCPU_GET(spinlocks);
2413 	}
2414 	instance = find_instance(lock_list, lock);
2415 	if (instance == NULL)
2416 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2417 		    class->lc_name, lock->lo_name);
2418 	lock->lo_witness->w_file = file;
2419 	lock->lo_witness->w_line = line;
2420 	if (instance == NULL)
2421 		return;
2422 	instance->li_file = file;
2423 	instance->li_line = line;
2424 }
2425 
2426 void
2427 witness_assert(const struct lock_object *lock, int flags, const char *file,
2428     int line)
2429 {
2430 #ifdef INVARIANT_SUPPORT
2431 	struct lock_instance *instance;
2432 	struct lock_class *class;
2433 
2434 	if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2435 		return;
2436 	class = LOCK_CLASS(lock);
2437 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2438 		instance = find_instance(curthread->td_sleeplocks, lock);
2439 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2440 		instance = find_instance(PCPU_GET(spinlocks), lock);
2441 	else {
2442 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2443 		    class->lc_name, lock->lo_name);
2444 		return;
2445 	}
2446 	switch (flags) {
2447 	case LA_UNLOCKED:
2448 		if (instance != NULL)
2449 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2450 			    class->lc_name, lock->lo_name,
2451 			    fixup_filename(file), line);
2452 		break;
2453 	case LA_LOCKED:
2454 	case LA_LOCKED | LA_RECURSED:
2455 	case LA_LOCKED | LA_NOTRECURSED:
2456 	case LA_SLOCKED:
2457 	case LA_SLOCKED | LA_RECURSED:
2458 	case LA_SLOCKED | LA_NOTRECURSED:
2459 	case LA_XLOCKED:
2460 	case LA_XLOCKED | LA_RECURSED:
2461 	case LA_XLOCKED | LA_NOTRECURSED:
2462 		if (instance == NULL) {
2463 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2464 			    class->lc_name, lock->lo_name,
2465 			    fixup_filename(file), line);
2466 			break;
2467 		}
2468 		if ((flags & LA_XLOCKED) != 0 &&
2469 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2470 			kassert_panic(
2471 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2472 			    class->lc_name, lock->lo_name,
2473 			    fixup_filename(file), line);
2474 		if ((flags & LA_SLOCKED) != 0 &&
2475 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2476 			kassert_panic(
2477 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2478 			    class->lc_name, lock->lo_name,
2479 			    fixup_filename(file), line);
2480 		if ((flags & LA_RECURSED) != 0 &&
2481 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2482 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2483 			    class->lc_name, lock->lo_name,
2484 			    fixup_filename(file), line);
2485 		if ((flags & LA_NOTRECURSED) != 0 &&
2486 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2487 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2488 			    class->lc_name, lock->lo_name,
2489 			    fixup_filename(file), line);
2490 		break;
2491 	default:
2492 		kassert_panic("Invalid lock assertion at %s:%d.",
2493 		    fixup_filename(file), line);
2494 	}
2495 #endif	/* INVARIANT_SUPPORT */
2496 }
2497 
2498 static void
2499 witness_setflag(struct lock_object *lock, int flag, int set)
2500 {
2501 	struct lock_list_entry *lock_list;
2502 	struct lock_instance *instance;
2503 	struct lock_class *class;
2504 
2505 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2506 		return;
2507 	class = LOCK_CLASS(lock);
2508 	if (class->lc_flags & LC_SLEEPLOCK)
2509 		lock_list = curthread->td_sleeplocks;
2510 	else {
2511 		if (witness_skipspin)
2512 			return;
2513 		lock_list = PCPU_GET(spinlocks);
2514 	}
2515 	instance = find_instance(lock_list, lock);
2516 	if (instance == NULL) {
2517 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2518 		    class->lc_name, lock->lo_name);
2519 		return;
2520 	}
2521 
2522 	if (set)
2523 		instance->li_flags |= flag;
2524 	else
2525 		instance->li_flags &= ~flag;
2526 }
2527 
2528 void
2529 witness_norelease(struct lock_object *lock)
2530 {
2531 
2532 	witness_setflag(lock, LI_NORELEASE, 1);
2533 }
2534 
2535 void
2536 witness_releaseok(struct lock_object *lock)
2537 {
2538 
2539 	witness_setflag(lock, LI_NORELEASE, 0);
2540 }
2541 
2542 #ifdef DDB
2543 static void
2544 witness_ddb_list(struct thread *td)
2545 {
2546 
2547 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2548 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2549 
2550 	if (witness_watch < 1)
2551 		return;
2552 
2553 	witness_list_locks(&td->td_sleeplocks, db_printf);
2554 
2555 	/*
2556 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2557 	 * if td is currently executing on some other CPU and holds spin locks
2558 	 * as we won't display those locks.  If we had a MI way of getting
2559 	 * the per-cpu data for a given cpu then we could use
2560 	 * td->td_oncpu to get the list of spinlocks for this thread
2561 	 * and "fix" this.
2562 	 *
2563 	 * That still wouldn't really fix this unless we locked the scheduler
2564 	 * lock or stopped the other CPU to make sure it wasn't changing the
2565 	 * list out from under us.  It is probably best to just not try to
2566 	 * handle threads on other CPU's for now.
2567 	 */
2568 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2569 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2570 }
2571 
2572 DB_SHOW_COMMAND(locks, db_witness_list)
2573 {
2574 	struct thread *td;
2575 
2576 	if (have_addr)
2577 		td = db_lookup_thread(addr, true);
2578 	else
2579 		td = kdb_thread;
2580 	witness_ddb_list(td);
2581 }
2582 
2583 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2584 {
2585 	struct thread *td;
2586 	struct proc *p;
2587 
2588 	/*
2589 	 * It would be nice to list only threads and processes that actually
2590 	 * held sleep locks, but that information is currently not exported
2591 	 * by WITNESS.
2592 	 */
2593 	FOREACH_PROC_IN_SYSTEM(p) {
2594 		if (!witness_proc_has_locks(p))
2595 			continue;
2596 		FOREACH_THREAD_IN_PROC(p, td) {
2597 			if (!witness_thread_has_locks(td))
2598 				continue;
2599 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2600 			    p->p_comm, td, td->td_tid);
2601 			witness_ddb_list(td);
2602 			if (db_pager_quit)
2603 				return;
2604 		}
2605 	}
2606 }
2607 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2608 
2609 DB_SHOW_COMMAND(witness, db_witness_display)
2610 {
2611 
2612 	witness_ddb_display(db_printf);
2613 }
2614 #endif
2615 
2616 static void
2617 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2618 {
2619 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2620 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2621 	int generation, i, j;
2622 
2623 	tmp_data1 = NULL;
2624 	tmp_data2 = NULL;
2625 	tmp_w1 = NULL;
2626 	tmp_w2 = NULL;
2627 
2628 	/* Allocate and init temporary storage space. */
2629 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2630 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2631 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2632 	    M_WAITOK | M_ZERO);
2633 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2634 	    M_WAITOK | M_ZERO);
2635 	stack_zero(&tmp_data1->wlod_stack);
2636 	stack_zero(&tmp_data2->wlod_stack);
2637 
2638 restart:
2639 	mtx_lock_spin(&w_mtx);
2640 	generation = w_generation;
2641 	mtx_unlock_spin(&w_mtx);
2642 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2643 	    w_lohash.wloh_count);
2644 	for (i = 1; i < w_max_used_index; i++) {
2645 		mtx_lock_spin(&w_mtx);
2646 		if (generation != w_generation) {
2647 			mtx_unlock_spin(&w_mtx);
2648 
2649 			/* The graph has changed, try again. */
2650 			*oldidx = 0;
2651 			sbuf_clear(sb);
2652 			goto restart;
2653 		}
2654 
2655 		w1 = &w_data[i];
2656 		if (w1->w_reversed == 0) {
2657 			mtx_unlock_spin(&w_mtx);
2658 			continue;
2659 		}
2660 
2661 		/* Copy w1 locally so we can release the spin lock. */
2662 		*tmp_w1 = *w1;
2663 		mtx_unlock_spin(&w_mtx);
2664 
2665 		if (tmp_w1->w_reversed == 0)
2666 			continue;
2667 		for (j = 1; j < w_max_used_index; j++) {
2668 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2669 				continue;
2670 
2671 			mtx_lock_spin(&w_mtx);
2672 			if (generation != w_generation) {
2673 				mtx_unlock_spin(&w_mtx);
2674 
2675 				/* The graph has changed, try again. */
2676 				*oldidx = 0;
2677 				sbuf_clear(sb);
2678 				goto restart;
2679 			}
2680 
2681 			w2 = &w_data[j];
2682 			data1 = witness_lock_order_get(w1, w2);
2683 			data2 = witness_lock_order_get(w2, w1);
2684 
2685 			/*
2686 			 * Copy information locally so we can release the
2687 			 * spin lock.
2688 			 */
2689 			*tmp_w2 = *w2;
2690 
2691 			if (data1) {
2692 				stack_zero(&tmp_data1->wlod_stack);
2693 				stack_copy(&data1->wlod_stack,
2694 				    &tmp_data1->wlod_stack);
2695 			}
2696 			if (data2 && data2 != data1) {
2697 				stack_zero(&tmp_data2->wlod_stack);
2698 				stack_copy(&data2->wlod_stack,
2699 				    &tmp_data2->wlod_stack);
2700 			}
2701 			mtx_unlock_spin(&w_mtx);
2702 
2703 			if (blessed(tmp_w1, tmp_w2))
2704 				continue;
2705 
2706 			sbuf_printf(sb,
2707 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2708 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2709 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2710 			if (data1) {
2711 				sbuf_printf(sb,
2712 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2713 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2714 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2715 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2716 				sbuf_printf(sb, "\n");
2717 			}
2718 			if (data2 && data2 != data1) {
2719 				sbuf_printf(sb,
2720 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2721 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2722 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2723 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2724 				sbuf_printf(sb, "\n");
2725 			}
2726 		}
2727 	}
2728 	mtx_lock_spin(&w_mtx);
2729 	if (generation != w_generation) {
2730 		mtx_unlock_spin(&w_mtx);
2731 
2732 		/*
2733 		 * The graph changed while we were printing stack data,
2734 		 * try again.
2735 		 */
2736 		*oldidx = 0;
2737 		sbuf_clear(sb);
2738 		goto restart;
2739 	}
2740 	mtx_unlock_spin(&w_mtx);
2741 
2742 	/* Free temporary storage space. */
2743 	free(tmp_data1, M_TEMP);
2744 	free(tmp_data2, M_TEMP);
2745 	free(tmp_w1, M_TEMP);
2746 	free(tmp_w2, M_TEMP);
2747 }
2748 
2749 static int
2750 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2751 {
2752 	struct sbuf *sb;
2753 	int error;
2754 
2755 	if (witness_watch < 1) {
2756 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2757 		return (error);
2758 	}
2759 	if (witness_cold) {
2760 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2761 		return (error);
2762 	}
2763 	error = 0;
2764 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2765 	if (sb == NULL)
2766 		return (ENOMEM);
2767 
2768 	sbuf_print_witness_badstacks(sb, &req->oldidx);
2769 
2770 	sbuf_finish(sb);
2771 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2772 	sbuf_delete(sb);
2773 
2774 	return (error);
2775 }
2776 
2777 #ifdef DDB
2778 static int
2779 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2780 {
2781 
2782 	return (db_printf("%.*s", len, data));
2783 }
2784 
2785 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2786 {
2787 	struct sbuf sb;
2788 	char buffer[128];
2789 	size_t dummy;
2790 
2791 	sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2792 	sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2793 	sbuf_print_witness_badstacks(&sb, &dummy);
2794 	sbuf_finish(&sb);
2795 }
2796 #endif
2797 
2798 static int
2799 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2800 {
2801 	static const struct {
2802 		enum witness_channel channel;
2803 		const char *name;
2804 	} channels[] = {
2805 		{ WITNESS_CONSOLE, "console" },
2806 		{ WITNESS_LOG, "log" },
2807 		{ WITNESS_NONE, "none" },
2808 	};
2809 	char buf[16];
2810 	u_int i;
2811 	int error;
2812 
2813 	buf[0] = '\0';
2814 	for (i = 0; i < nitems(channels); i++)
2815 		if (witness_channel == channels[i].channel) {
2816 			snprintf(buf, sizeof(buf), "%s", channels[i].name);
2817 			break;
2818 		}
2819 
2820 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2821 	if (error != 0 || req->newptr == NULL)
2822 		return (error);
2823 
2824 	error = EINVAL;
2825 	for (i = 0; i < nitems(channels); i++)
2826 		if (strcmp(channels[i].name, buf) == 0) {
2827 			witness_channel = channels[i].channel;
2828 			error = 0;
2829 			break;
2830 		}
2831 	return (error);
2832 }
2833 
2834 static int
2835 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2836 {
2837 	struct witness *w;
2838 	struct sbuf *sb;
2839 	int error;
2840 
2841 #ifdef __i386__
2842 	error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2843 	return (error);
2844 #endif
2845 
2846 	if (witness_watch < 1) {
2847 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2848 		return (error);
2849 	}
2850 	if (witness_cold) {
2851 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2852 		return (error);
2853 	}
2854 	error = 0;
2855 
2856 	error = sysctl_wire_old_buffer(req, 0);
2857 	if (error != 0)
2858 		return (error);
2859 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2860 	if (sb == NULL)
2861 		return (ENOMEM);
2862 	sbuf_printf(sb, "\n");
2863 
2864 	mtx_lock_spin(&w_mtx);
2865 	STAILQ_FOREACH(w, &w_all, w_list)
2866 		w->w_displayed = 0;
2867 	STAILQ_FOREACH(w, &w_all, w_list)
2868 		witness_add_fullgraph(sb, w);
2869 	mtx_unlock_spin(&w_mtx);
2870 
2871 	/*
2872 	 * Close the sbuf and return to userland.
2873 	 */
2874 	error = sbuf_finish(sb);
2875 	sbuf_delete(sb);
2876 
2877 	return (error);
2878 }
2879 
2880 static int
2881 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2882 {
2883 	int error, value;
2884 
2885 	value = witness_watch;
2886 	error = sysctl_handle_int(oidp, &value, 0, req);
2887 	if (error != 0 || req->newptr == NULL)
2888 		return (error);
2889 	if (value > 1 || value < -1 ||
2890 	    (witness_watch == -1 && value != witness_watch))
2891 		return (EINVAL);
2892 	witness_watch = value;
2893 	return (0);
2894 }
2895 
2896 static void
2897 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2898 {
2899 	int i;
2900 
2901 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2902 		return;
2903 	w->w_displayed = 1;
2904 
2905 	WITNESS_INDEX_ASSERT(w->w_index);
2906 	for (i = 1; i <= w_max_used_index; i++) {
2907 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2908 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2909 			    w_data[i].w_name);
2910 			witness_add_fullgraph(sb, &w_data[i]);
2911 		}
2912 	}
2913 }
2914 
2915 /*
2916  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2917  * interprets the key as a string and reads until the null
2918  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2919  * hash value computed from the key.
2920  */
2921 static uint32_t
2922 witness_hash_djb2(const uint8_t *key, uint32_t size)
2923 {
2924 	unsigned int hash = 5381;
2925 	int i;
2926 
2927 	/* hash = hash * 33 + key[i] */
2928 	if (size)
2929 		for (i = 0; i < size; i++)
2930 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2931 	else
2932 		for (i = 0; key[i] != 0; i++)
2933 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2934 
2935 	return (hash);
2936 }
2937 
2938 /*
2939  * Initializes the two witness hash tables. Called exactly once from
2940  * witness_initialize().
2941  */
2942 static void
2943 witness_init_hash_tables(void)
2944 {
2945 	int i;
2946 
2947 	MPASS(witness_cold);
2948 
2949 	/* Initialize the hash tables. */
2950 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2951 		w_hash.wh_array[i] = NULL;
2952 
2953 	w_hash.wh_size = WITNESS_HASH_SIZE;
2954 	w_hash.wh_count = 0;
2955 
2956 	/* Initialize the lock order data hash. */
2957 	w_lofree = NULL;
2958 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2959 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2960 		w_lodata[i].wlod_next = w_lofree;
2961 		w_lofree = &w_lodata[i];
2962 	}
2963 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2964 	w_lohash.wloh_count = 0;
2965 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2966 		w_lohash.wloh_array[i] = NULL;
2967 }
2968 
2969 static struct witness *
2970 witness_hash_get(const char *key)
2971 {
2972 	struct witness *w;
2973 	uint32_t hash;
2974 
2975 	MPASS(key != NULL);
2976 	if (witness_cold == 0)
2977 		mtx_assert(&w_mtx, MA_OWNED);
2978 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2979 	w = w_hash.wh_array[hash];
2980 	while (w != NULL) {
2981 		if (strcmp(w->w_name, key) == 0)
2982 			goto out;
2983 		w = w->w_hash_next;
2984 	}
2985 
2986 out:
2987 	return (w);
2988 }
2989 
2990 static void
2991 witness_hash_put(struct witness *w)
2992 {
2993 	uint32_t hash;
2994 
2995 	MPASS(w != NULL);
2996 	MPASS(w->w_name != NULL);
2997 	if (witness_cold == 0)
2998 		mtx_assert(&w_mtx, MA_OWNED);
2999 	KASSERT(witness_hash_get(w->w_name) == NULL,
3000 	    ("%s: trying to add a hash entry that already exists!", __func__));
3001 	KASSERT(w->w_hash_next == NULL,
3002 	    ("%s: w->w_hash_next != NULL", __func__));
3003 
3004 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3005 	w->w_hash_next = w_hash.wh_array[hash];
3006 	w_hash.wh_array[hash] = w;
3007 	w_hash.wh_count++;
3008 }
3009 
3010 static struct witness_lock_order_data *
3011 witness_lock_order_get(struct witness *parent, struct witness *child)
3012 {
3013 	struct witness_lock_order_data *data = NULL;
3014 	struct witness_lock_order_key key;
3015 	unsigned int hash;
3016 
3017 	MPASS(parent != NULL && child != NULL);
3018 	key.from = parent->w_index;
3019 	key.to = child->w_index;
3020 	WITNESS_INDEX_ASSERT(key.from);
3021 	WITNESS_INDEX_ASSERT(key.to);
3022 	if ((w_rmatrix[parent->w_index][child->w_index]
3023 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
3024 		goto out;
3025 
3026 	hash = witness_hash_djb2((const char*)&key,
3027 	    sizeof(key)) % w_lohash.wloh_size;
3028 	data = w_lohash.wloh_array[hash];
3029 	while (data != NULL) {
3030 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
3031 			break;
3032 		data = data->wlod_next;
3033 	}
3034 
3035 out:
3036 	return (data);
3037 }
3038 
3039 /*
3040  * Verify that parent and child have a known relationship, are not the same,
3041  * and child is actually a child of parent.  This is done without w_mtx
3042  * to avoid contention in the common case.
3043  */
3044 static int
3045 witness_lock_order_check(struct witness *parent, struct witness *child)
3046 {
3047 
3048 	if (parent != child &&
3049 	    w_rmatrix[parent->w_index][child->w_index]
3050 	    & WITNESS_LOCK_ORDER_KNOWN &&
3051 	    isitmychild(parent, child))
3052 		return (1);
3053 
3054 	return (0);
3055 }
3056 
3057 static int
3058 witness_lock_order_add(struct witness *parent, struct witness *child)
3059 {
3060 	struct witness_lock_order_data *data = NULL;
3061 	struct witness_lock_order_key key;
3062 	unsigned int hash;
3063 
3064 	MPASS(parent != NULL && child != NULL);
3065 	key.from = parent->w_index;
3066 	key.to = child->w_index;
3067 	WITNESS_INDEX_ASSERT(key.from);
3068 	WITNESS_INDEX_ASSERT(key.to);
3069 	if (w_rmatrix[parent->w_index][child->w_index]
3070 	    & WITNESS_LOCK_ORDER_KNOWN)
3071 		return (1);
3072 
3073 	hash = witness_hash_djb2((const char*)&key,
3074 	    sizeof(key)) % w_lohash.wloh_size;
3075 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3076 	data = w_lofree;
3077 	if (data == NULL)
3078 		return (0);
3079 	w_lofree = data->wlod_next;
3080 	data->wlod_next = w_lohash.wloh_array[hash];
3081 	data->wlod_key = key;
3082 	w_lohash.wloh_array[hash] = data;
3083 	w_lohash.wloh_count++;
3084 	stack_zero(&data->wlod_stack);
3085 	stack_save(&data->wlod_stack);
3086 	return (1);
3087 }
3088 
3089 /* Call this whenever the structure of the witness graph changes. */
3090 static void
3091 witness_increment_graph_generation(void)
3092 {
3093 
3094 	if (witness_cold == 0)
3095 		mtx_assert(&w_mtx, MA_OWNED);
3096 	w_generation++;
3097 }
3098 
3099 static int
3100 witness_output_drain(void *arg __unused, const char *data, int len)
3101 {
3102 
3103 	witness_output("%.*s", len, data);
3104 	return (len);
3105 }
3106 
3107 static void
3108 witness_debugger(int cond, const char *msg)
3109 {
3110 	char buf[32];
3111 	struct sbuf sb;
3112 	struct stack st;
3113 
3114 	if (!cond)
3115 		return;
3116 
3117 	if (witness_trace) {
3118 		sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3119 		sbuf_set_drain(&sb, witness_output_drain, NULL);
3120 
3121 		stack_zero(&st);
3122 		stack_save(&st);
3123 		witness_output("stack backtrace:\n");
3124 		stack_sbuf_print_ddb(&sb, &st);
3125 
3126 		sbuf_finish(&sb);
3127 	}
3128 
3129 	witness_enter_debugger(msg);
3130 }
3131 
3132 static void
3133 witness_enter_debugger(const char *msg)
3134 {
3135 #ifdef KDB
3136 	if (witness_kdb)
3137 		kdb_enter(KDB_WHY_WITNESS, msg);
3138 #endif
3139 }
3140