xref: /freebsd/sys/kern/subr_witness.c (revision 7899f917b1c0ea178f1d2be0cfb452086d079d23)
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 #include "opt_ddb.h"
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
94 
95 #include <sys/param.h>
96 #include <sys/bus.h>
97 #include <sys/kdb.h>
98 #include <sys/kernel.h>
99 #include <sys/ktr.h>
100 #include <sys/lock.h>
101 #include <sys/malloc.h>
102 #include <sys/mutex.h>
103 #include <sys/priv.h>
104 #include <sys/proc.h>
105 #include <sys/sbuf.h>
106 #include <sys/sched.h>
107 #include <sys/stack.h>
108 #include <sys/sysctl.h>
109 #include <sys/syslog.h>
110 #include <sys/systm.h>
111 
112 #ifdef DDB
113 #include <ddb/ddb.h>
114 #endif
115 
116 #include <machine/stdarg.h>
117 
118 #if !defined(DDB) && !defined(STACK)
119 #error "DDB or STACK options are required for WITNESS"
120 #endif
121 
122 /* Note that these traces do not work with KTR_ALQ. */
123 #if 0
124 #define	KTR_WITNESS	KTR_SUBSYS
125 #else
126 #define	KTR_WITNESS	0
127 #endif
128 
129 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
130 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
131 #define	LI_NORELEASE	0x00020000	/* Lock not allowed to be released. */
132 #define	LI_SLEEPABLE	0x00040000	/* Lock may be held while sleeping. */
133 
134 #ifndef WITNESS_COUNT
135 #define	WITNESS_COUNT 		1536
136 #endif
137 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
138 #define	WITNESS_PENDLIST	(512 + (MAXCPU * 4))
139 
140 /* Allocate 256 KB of stack data space */
141 #define	WITNESS_LO_DATA_COUNT	2048
142 
143 /* Prime, gives load factor of ~2 at full load */
144 #define	WITNESS_LO_HASH_SIZE	1021
145 
146 /*
147  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
148  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
149  * probably be safe for the most part, but it's still a SWAG.
150  */
151 #define	LOCK_NCHILDREN	5
152 #define	LOCK_CHILDCOUNT	2048
153 
154 #define	MAX_W_NAME	64
155 
156 #define	FULLGRAPH_SBUF_SIZE	512
157 
158 /*
159  * These flags go in the witness relationship matrix and describe the
160  * relationship between any two struct witness objects.
161  */
162 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
163 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
164 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
165 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
166 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
167 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
168 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
169 #define	WITNESS_RELATED_MASK						\
170 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
171 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
172 					  * observed. */
173 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
174 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
175 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
176 
177 /* Descendant to ancestor flags */
178 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
179 
180 /* Ancestor to descendant flags */
181 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
182 
183 #define	WITNESS_INDEX_ASSERT(i)						\
184 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
185 
186 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
187 
188 /*
189  * Lock instances.  A lock instance is the data associated with a lock while
190  * it is held by witness.  For example, a lock instance will hold the
191  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
192  * are held in a per-cpu list while sleep locks are held in per-thread list.
193  */
194 struct lock_instance {
195 	struct lock_object	*li_lock;
196 	const char		*li_file;
197 	int			li_line;
198 	u_int			li_flags;
199 };
200 
201 /*
202  * A simple list type used to build the list of locks held by a thread
203  * or CPU.  We can't simply embed the list in struct lock_object since a
204  * lock may be held by more than one thread if it is a shared lock.  Locks
205  * are added to the head of the list, so we fill up each list entry from
206  * "the back" logically.  To ease some of the arithmetic, we actually fill
207  * in each list entry the normal way (children[0] then children[1], etc.) but
208  * when we traverse the list we read children[count-1] as the first entry
209  * down to children[0] as the final entry.
210  */
211 struct lock_list_entry {
212 	struct lock_list_entry	*ll_next;
213 	struct lock_instance	ll_children[LOCK_NCHILDREN];
214 	u_int			ll_count;
215 };
216 
217 /*
218  * The main witness structure. One of these per named lock type in the system
219  * (for example, "vnode interlock").
220  */
221 struct witness {
222 	char  			w_name[MAX_W_NAME];
223 	uint32_t 		w_index;  /* Index in the relationship matrix */
224 	struct lock_class	*w_class;
225 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
226 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
227 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
228 	const char		*w_file; /* File where last acquired */
229 	uint32_t 		w_line; /* Line where last acquired */
230 	uint32_t 		w_refcount;
231 	uint16_t 		w_num_ancestors; /* direct/indirect
232 						  * ancestor count */
233 	uint16_t 		w_num_descendants; /* direct/indirect
234 						    * descendant count */
235 	int16_t 		w_ddb_level;
236 	unsigned		w_displayed:1;
237 	unsigned		w_reversed:1;
238 };
239 
240 STAILQ_HEAD(witness_list, witness);
241 
242 /*
243  * The witness hash table. Keys are witness names (const char *), elements are
244  * witness objects (struct witness *).
245  */
246 struct witness_hash {
247 	struct witness	*wh_array[WITNESS_HASH_SIZE];
248 	uint32_t	wh_size;
249 	uint32_t	wh_count;
250 };
251 
252 /*
253  * Key type for the lock order data hash table.
254  */
255 struct witness_lock_order_key {
256 	uint16_t	from;
257 	uint16_t	to;
258 };
259 
260 struct witness_lock_order_data {
261 	struct stack			wlod_stack;
262 	struct witness_lock_order_key	wlod_key;
263 	struct witness_lock_order_data	*wlod_next;
264 };
265 
266 /*
267  * The witness lock order data hash table. Keys are witness index tuples
268  * (struct witness_lock_order_key), elements are lock order data objects
269  * (struct witness_lock_order_data).
270  */
271 struct witness_lock_order_hash {
272 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
273 	u_int	wloh_size;
274 	u_int	wloh_count;
275 };
276 
277 struct witness_blessed {
278 	const char	*b_lock1;
279 	const char	*b_lock2;
280 };
281 
282 struct witness_pendhelp {
283 	const char		*wh_type;
284 	struct lock_object	*wh_lock;
285 };
286 
287 struct witness_order_list_entry {
288 	const char		*w_name;
289 	struct lock_class	*w_class;
290 };
291 
292 /*
293  * Returns 0 if one of the locks is a spin lock and the other is not.
294  * Returns 1 otherwise.
295  */
296 static __inline int
297 witness_lock_type_equal(struct witness *w1, struct witness *w2)
298 {
299 
300 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
301 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
302 }
303 
304 static __inline int
305 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
306     const struct witness_lock_order_key *b)
307 {
308 
309 	return (a->from == b->from && a->to == b->to);
310 }
311 
312 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
313 		    const char *fname);
314 static void	adopt(struct witness *parent, struct witness *child);
315 static int	blessed(struct witness *, struct witness *);
316 static void	depart(struct witness *w);
317 static struct witness	*enroll(const char *description,
318 			    struct lock_class *lock_class);
319 static struct lock_instance	*find_instance(struct lock_list_entry *list,
320 				    const struct lock_object *lock);
321 static int	isitmychild(struct witness *parent, struct witness *child);
322 static int	isitmydescendant(struct witness *parent, struct witness *child);
323 static void	itismychild(struct witness *parent, struct witness *child);
324 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
325 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
326 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
327 static int	sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
328 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
329 #ifdef DDB
330 static void	witness_ddb_compute_levels(void);
331 static void	witness_ddb_display(int(*)(const char *fmt, ...));
332 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
333 		    struct witness *, int indent);
334 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
335 		    struct witness_list *list);
336 static void	witness_ddb_level_descendants(struct witness *parent, int l);
337 static void	witness_ddb_list(struct thread *td);
338 #endif
339 static void	witness_enter_debugger(const char *msg);
340 static void	witness_debugger(int cond, const char *msg);
341 static void	witness_free(struct witness *m);
342 static struct witness	*witness_get(void);
343 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
344 static struct witness	*witness_hash_get(const char *key);
345 static void	witness_hash_put(struct witness *w);
346 static void	witness_init_hash_tables(void);
347 static void	witness_increment_graph_generation(void);
348 static void	witness_lock_list_free(struct lock_list_entry *lle);
349 static struct lock_list_entry	*witness_lock_list_get(void);
350 static int	witness_lock_order_add(struct witness *parent,
351 		    struct witness *child);
352 static int	witness_lock_order_check(struct witness *parent,
353 		    struct witness *child);
354 static struct witness_lock_order_data	*witness_lock_order_get(
355 					    struct witness *parent,
356 					    struct witness *child);
357 static void	witness_list_lock(struct lock_instance *instance,
358 		    int (*prnt)(const char *fmt, ...));
359 static int	witness_output(const char *fmt, ...) __printflike(1, 2);
360 static int	witness_output_drain(void *arg __unused, const char *data,
361 		    int len);
362 static int	witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
363 static void	witness_setflag(struct lock_object *lock, int flag, int set);
364 
365 FEATURE(witness, "kernel has witness(9) support");
366 
367 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
368     "Witness Locking");
369 
370 /*
371  * If set to 0, lock order checking is disabled.  If set to -1,
372  * witness is completely disabled.  Otherwise witness performs full
373  * lock order checking for all locks.  At runtime, lock order checking
374  * may be toggled.  However, witness cannot be reenabled once it is
375  * completely disabled.
376  */
377 static int witness_watch = 1;
378 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
379     CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
380     sysctl_debug_witness_watch, "I",
381     "witness is watching lock operations");
382 
383 #ifdef KDB
384 /*
385  * When KDB is enabled and witness_kdb is 1, it will cause the system
386  * to drop into kdebug() when:
387  *	- a lock hierarchy violation occurs
388  *	- locks are held when going to sleep.
389  */
390 #ifdef WITNESS_KDB
391 int	witness_kdb = 1;
392 #else
393 int	witness_kdb = 0;
394 #endif
395 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
396 #endif /* KDB */
397 
398 #if defined(DDB) || defined(KDB)
399 /*
400  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
401  * to print a stack trace:
402  *	- a lock hierarchy violation occurs
403  *	- locks are held when going to sleep.
404  */
405 int	witness_trace = 1;
406 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
407 #endif /* DDB || KDB */
408 
409 #ifdef WITNESS_SKIPSPIN
410 int	witness_skipspin = 1;
411 #else
412 int	witness_skipspin = 0;
413 #endif
414 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
415 
416 int badstack_sbuf_size;
417 
418 int witness_count = WITNESS_COUNT;
419 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
420     &witness_count, 0, "");
421 
422 /*
423  * Output channel for witness messages.  By default we print to the console.
424  */
425 enum witness_channel {
426 	WITNESS_CONSOLE,
427 	WITNESS_LOG,
428 	WITNESS_NONE,
429 };
430 
431 static enum witness_channel witness_channel = WITNESS_CONSOLE;
432 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
433     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
434     sysctl_debug_witness_channel, "A",
435     "Output channel for warnings");
436 
437 /*
438  * Call this to print out the relations between locks.
439  */
440 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
441     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
442     sysctl_debug_witness_fullgraph, "A",
443     "Show locks relation graphs");
444 
445 /*
446  * Call this to print out the witness faulty stacks.
447  */
448 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
449     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
450     sysctl_debug_witness_badstacks, "A",
451     "Show bad witness stacks");
452 
453 static struct mtx w_mtx;
454 
455 /* w_list */
456 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
457 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
458 
459 /* w_typelist */
460 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
461 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
462 
463 /* lock list */
464 static struct lock_list_entry *w_lock_list_free = NULL;
465 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
466 static u_int pending_cnt;
467 
468 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
469 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
470 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
471 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
472     "");
473 
474 static struct witness *w_data;
475 static uint8_t **w_rmatrix;
476 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
477 static struct witness_hash w_hash;	/* The witness hash table. */
478 
479 /* The lock order data hash */
480 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
481 static struct witness_lock_order_data *w_lofree = NULL;
482 static struct witness_lock_order_hash w_lohash;
483 static int w_max_used_index = 0;
484 static unsigned int w_generation = 0;
485 static const char w_notrunning[] = "Witness not running\n";
486 static const char w_stillcold[] = "Witness is still cold\n";
487 #ifdef __i386__
488 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
489 #endif
490 
491 static struct witness_order_list_entry order_lists[] = {
492 	/*
493 	 * sx locks
494 	 */
495 	{ "proctree", &lock_class_sx },
496 	{ "allproc", &lock_class_sx },
497 	{ "allprison", &lock_class_sx },
498 	{ NULL, NULL },
499 	/*
500 	 * Various mutexes
501 	 */
502 	{ "Giant", &lock_class_mtx_sleep },
503 	{ "pipe mutex", &lock_class_mtx_sleep },
504 	{ "sigio lock", &lock_class_mtx_sleep },
505 	{ "process group", &lock_class_mtx_sleep },
506 #ifdef	HWPMC_HOOKS
507 	{ "pmc-sleep", &lock_class_mtx_sleep },
508 #endif
509 	{ "process lock", &lock_class_mtx_sleep },
510 	{ "session", &lock_class_mtx_sleep },
511 	{ "uidinfo hash", &lock_class_rw },
512 	{ "time lock", &lock_class_mtx_sleep },
513 	{ NULL, NULL },
514 	/*
515 	 * umtx
516 	 */
517 	{ "umtx lock", &lock_class_mtx_sleep },
518 	{ NULL, NULL },
519 	/*
520 	 * Sockets
521 	 */
522 	{ "accept", &lock_class_mtx_sleep },
523 	{ "so_snd", &lock_class_mtx_sleep },
524 	{ "so_rcv", &lock_class_mtx_sleep },
525 	{ "sellck", &lock_class_mtx_sleep },
526 	{ NULL, NULL },
527 	/*
528 	 * Routing
529 	 */
530 	{ "so_rcv", &lock_class_mtx_sleep },
531 	{ "radix node head", &lock_class_rm },
532 	{ "ifaddr", &lock_class_mtx_sleep },
533 	{ NULL, NULL },
534 	/*
535 	 * IPv4 multicast:
536 	 * protocol locks before interface locks, after UDP locks.
537 	 */
538 	{ "in_multi_sx", &lock_class_sx },
539 	{ "udpinp", &lock_class_rw },
540 	{ "in_multi_list_mtx", &lock_class_mtx_sleep },
541 	{ "igmp_mtx", &lock_class_mtx_sleep },
542 	{ "if_addr_lock", &lock_class_mtx_sleep },
543 	{ NULL, NULL },
544 	/*
545 	 * IPv6 multicast:
546 	 * protocol locks before interface locks, after UDP locks.
547 	 */
548 	{ "in6_multi_sx", &lock_class_sx },
549 	{ "udpinp", &lock_class_rw },
550 	{ "in6_multi_list_mtx", &lock_class_mtx_sleep },
551 	{ "mld_mtx", &lock_class_mtx_sleep },
552 	{ "if_addr_lock", &lock_class_mtx_sleep },
553 	{ NULL, NULL },
554 	/*
555 	 * UNIX Domain Sockets
556 	 */
557 	{ "unp_link_rwlock", &lock_class_rw },
558 	{ "unp_list_lock", &lock_class_mtx_sleep },
559 	{ "unp", &lock_class_mtx_sleep },
560 	{ "so_snd", &lock_class_mtx_sleep },
561 	{ NULL, NULL },
562 	/*
563 	 * UDP/IP
564 	 */
565 	{ "udpinp", &lock_class_rw },
566 	{ "udp", &lock_class_mtx_sleep },
567 	{ "so_snd", &lock_class_mtx_sleep },
568 	{ NULL, NULL },
569 	/*
570 	 * TCP/IP
571 	 */
572 	{ "tcpinp", &lock_class_rw },
573 	{ "tcp", &lock_class_mtx_sleep },
574 	{ "so_snd", &lock_class_mtx_sleep },
575 	{ NULL, NULL },
576 	/*
577 	 * BPF
578 	 */
579 	{ "bpf global lock", &lock_class_sx },
580 	{ "bpf cdev lock", &lock_class_mtx_sleep },
581 	{ NULL, NULL },
582 	/*
583 	 * NFS server
584 	 */
585 	{ "nfsd_mtx", &lock_class_mtx_sleep },
586 	{ "so_snd", &lock_class_mtx_sleep },
587 	{ NULL, NULL },
588 
589 	/*
590 	 * IEEE 802.11
591 	 */
592 	{ "802.11 com lock", &lock_class_mtx_sleep},
593 	{ NULL, NULL },
594 	/*
595 	 * Network drivers
596 	 */
597 	{ "network driver", &lock_class_mtx_sleep},
598 	{ NULL, NULL },
599 
600 	/*
601 	 * Netgraph
602 	 */
603 	{ "ng_node", &lock_class_mtx_sleep },
604 	{ "ng_worklist", &lock_class_mtx_sleep },
605 	{ NULL, NULL },
606 	/*
607 	 * CDEV
608 	 */
609 	{ "vm map (system)", &lock_class_mtx_sleep },
610 	{ "vnode interlock", &lock_class_mtx_sleep },
611 	{ "cdev", &lock_class_mtx_sleep },
612 	{ "devthrd", &lock_class_mtx_sleep },
613 	{ NULL, NULL },
614 	/*
615 	 * VM
616 	 */
617 	{ "vm map (user)", &lock_class_sx },
618 	{ "vm object", &lock_class_rw },
619 	{ "vm page", &lock_class_mtx_sleep },
620 	{ "pmap pv global", &lock_class_rw },
621 	{ "pmap", &lock_class_mtx_sleep },
622 	{ "pmap pv list", &lock_class_rw },
623 	{ "vm page free queue", &lock_class_mtx_sleep },
624 	{ "vm pagequeue", &lock_class_mtx_sleep },
625 	{ NULL, NULL },
626 	/*
627 	 * kqueue/VFS interaction
628 	 */
629 	{ "kqueue", &lock_class_mtx_sleep },
630 	{ "struct mount mtx", &lock_class_mtx_sleep },
631 	{ "vnode interlock", &lock_class_mtx_sleep },
632 	{ NULL, NULL },
633 	/*
634 	 * VFS namecache
635 	 */
636 	{ "ncvn", &lock_class_mtx_sleep },
637 	{ "ncbuc", &lock_class_mtx_sleep },
638 	{ "vnode interlock", &lock_class_mtx_sleep },
639 	{ "ncneg", &lock_class_mtx_sleep },
640 	{ NULL, NULL },
641 	/*
642 	 * ZFS locking
643 	 */
644 	{ "dn->dn_mtx", &lock_class_sx },
645 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
646 	{ "db->db_mtx", &lock_class_sx },
647 	{ NULL, NULL },
648 	/*
649 	 * TCP log locks
650 	 */
651 	{ "TCP ID tree", &lock_class_rw },
652 	{ "tcp log id bucket", &lock_class_mtx_sleep },
653 	{ "tcpinp", &lock_class_rw },
654 	{ "TCP log expireq", &lock_class_mtx_sleep },
655 	{ NULL, NULL },
656 	/*
657 	 * spin locks
658 	 */
659 #ifdef SMP
660 	{ "ap boot", &lock_class_mtx_spin },
661 #endif
662 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
663 #ifdef __i386__
664 	{ "cy", &lock_class_mtx_spin },
665 #endif
666 	{ "scc_hwmtx", &lock_class_mtx_spin },
667 	{ "uart_hwmtx", &lock_class_mtx_spin },
668 	{ "fast_taskqueue", &lock_class_mtx_spin },
669 	{ "intr table", &lock_class_mtx_spin },
670 	{ "process slock", &lock_class_mtx_spin },
671 	{ "syscons video lock", &lock_class_mtx_spin },
672 	{ "sleepq chain", &lock_class_mtx_spin },
673 	{ "rm_spinlock", &lock_class_mtx_spin },
674 	{ "turnstile chain", &lock_class_mtx_spin },
675 	{ "turnstile lock", &lock_class_mtx_spin },
676 	{ "sched lock", &lock_class_mtx_spin },
677 	{ "td_contested", &lock_class_mtx_spin },
678 	{ "callout", &lock_class_mtx_spin },
679 	{ "entropy harvest mutex", &lock_class_mtx_spin },
680 #ifdef SMP
681 	{ "smp rendezvous", &lock_class_mtx_spin },
682 #endif
683 #ifdef __powerpc__
684 	{ "tlb0", &lock_class_mtx_spin },
685 #endif
686 	{ NULL, NULL },
687 	{ "sched lock", &lock_class_mtx_spin },
688 #ifdef	HWPMC_HOOKS
689 	{ "pmc-per-proc", &lock_class_mtx_spin },
690 #endif
691 	{ NULL, NULL },
692 	/*
693 	 * leaf locks
694 	 */
695 	{ "intrcnt", &lock_class_mtx_spin },
696 	{ "icu", &lock_class_mtx_spin },
697 #ifdef __i386__
698 	{ "allpmaps", &lock_class_mtx_spin },
699 	{ "descriptor tables", &lock_class_mtx_spin },
700 #endif
701 	{ "clk", &lock_class_mtx_spin },
702 	{ "cpuset", &lock_class_mtx_spin },
703 	{ "mprof lock", &lock_class_mtx_spin },
704 	{ "zombie lock", &lock_class_mtx_spin },
705 	{ "ALD Queue", &lock_class_mtx_spin },
706 #if defined(__i386__) || defined(__amd64__)
707 	{ "pcicfg", &lock_class_mtx_spin },
708 	{ "NDIS thread lock", &lock_class_mtx_spin },
709 #endif
710 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
711 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
712 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
713 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
714 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
715 #ifdef	HWPMC_HOOKS
716 	{ "pmc-leaf", &lock_class_mtx_spin },
717 #endif
718 	{ "blocked lock", &lock_class_mtx_spin },
719 	{ NULL, NULL },
720 	{ NULL, NULL }
721 };
722 
723 /*
724  * Pairs of locks which have been blessed.  Witness does not complain about
725  * order problems with blessed lock pairs.  Please do not add an entry to the
726  * table without an explanatory comment.
727  */
728 static struct witness_blessed blessed_list[] = {
729 	/*
730 	 * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
731 	 * both lock orders, so a deadlock cannot happen as a result of this
732 	 * LOR.
733 	 */
734 	{ "dirhash",	"bufwait" },
735 
736 	/*
737 	 * A UFS vnode may be locked in vget() while a buffer belonging to the
738 	 * parent directory vnode is locked.
739 	 */
740 	{ "ufs",	"bufwait" },
741 
742 	/*
743 	 * The tarfs decompression stream vnode may be locked while a
744 	 * buffer belonging to a tarfs data vnode is locked.
745 	 */
746 	{ "tarfs",	"bufwait" },
747 };
748 
749 /*
750  * This global is set to 0 once it becomes safe to use the witness code.
751  */
752 static int witness_cold = 1;
753 
754 /*
755  * This global is set to 1 once the static lock orders have been enrolled
756  * so that a warning can be issued for any spin locks enrolled later.
757  */
758 static int witness_spin_warn = 0;
759 
760 /* Trim useless garbage from filenames. */
761 static const char *
762 fixup_filename(const char *file)
763 {
764 
765 	if (file == NULL)
766 		return (NULL);
767 	while (strncmp(file, "../", 3) == 0)
768 		file += 3;
769 	return (file);
770 }
771 
772 /*
773  * Calculate the size of early witness structures.
774  */
775 int
776 witness_startup_count(void)
777 {
778 	int sz;
779 
780 	sz = sizeof(struct witness) * witness_count;
781 	sz += sizeof(*w_rmatrix) * (witness_count + 1);
782 	sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
783 	    (witness_count + 1);
784 
785 	return (sz);
786 }
787 
788 /*
789  * The WITNESS-enabled diagnostic code.  Note that the witness code does
790  * assume that the early boot is single-threaded at least until after this
791  * routine is completed.
792  */
793 void
794 witness_startup(void *mem)
795 {
796 	struct lock_object *lock;
797 	struct witness_order_list_entry *order;
798 	struct witness *w, *w1;
799 	uintptr_t p;
800 	int i;
801 
802 	p = (uintptr_t)mem;
803 	w_data = (void *)p;
804 	p += sizeof(struct witness) * witness_count;
805 
806 	w_rmatrix = (void *)p;
807 	p += sizeof(*w_rmatrix) * (witness_count + 1);
808 
809 	for (i = 0; i < witness_count + 1; i++) {
810 		w_rmatrix[i] = (void *)p;
811 		p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
812 	}
813 	badstack_sbuf_size = witness_count * 256;
814 
815 	/*
816 	 * We have to release Giant before initializing its witness
817 	 * structure so that WITNESS doesn't get confused.
818 	 */
819 	mtx_unlock(&Giant);
820 	mtx_assert(&Giant, MA_NOTOWNED);
821 
822 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
823 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
824 	    MTX_NOWITNESS | MTX_NOPROFILE);
825 	for (i = witness_count - 1; i >= 0; i--) {
826 		w = &w_data[i];
827 		memset(w, 0, sizeof(*w));
828 		w_data[i].w_index = i;	/* Witness index never changes. */
829 		witness_free(w);
830 	}
831 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
832 	    ("%s: Invalid list of free witness objects", __func__));
833 
834 	/* Witness with index 0 is not used to aid in debugging. */
835 	STAILQ_REMOVE_HEAD(&w_free, w_list);
836 	w_free_cnt--;
837 
838 	for (i = 0; i < witness_count; i++) {
839 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
840 		    (witness_count + 1));
841 	}
842 
843 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
844 		witness_lock_list_free(&w_locklistdata[i]);
845 	witness_init_hash_tables();
846 
847 	/* First add in all the specified order lists. */
848 	for (order = order_lists; order->w_name != NULL; order++) {
849 		w = enroll(order->w_name, order->w_class);
850 		if (w == NULL)
851 			continue;
852 		w->w_file = "order list";
853 		for (order++; order->w_name != NULL; order++) {
854 			w1 = enroll(order->w_name, order->w_class);
855 			if (w1 == NULL)
856 				continue;
857 			w1->w_file = "order list";
858 			itismychild(w, w1);
859 			w = w1;
860 		}
861 	}
862 	witness_spin_warn = 1;
863 
864 	/* Iterate through all locks and add them to witness. */
865 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
866 		lock = pending_locks[i].wh_lock;
867 		KASSERT(lock->lo_flags & LO_WITNESS,
868 		    ("%s: lock %s is on pending list but not LO_WITNESS",
869 		    __func__, lock->lo_name));
870 		lock->lo_witness = enroll(pending_locks[i].wh_type,
871 		    LOCK_CLASS(lock));
872 	}
873 
874 	/* Mark the witness code as being ready for use. */
875 	witness_cold = 0;
876 
877 	mtx_lock(&Giant);
878 }
879 
880 void
881 witness_init(struct lock_object *lock, const char *type)
882 {
883 	struct lock_class *class;
884 
885 	/* Various sanity checks. */
886 	class = LOCK_CLASS(lock);
887 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
888 	    (class->lc_flags & LC_RECURSABLE) == 0)
889 		kassert_panic("%s: lock (%s) %s can not be recursable",
890 		    __func__, class->lc_name, lock->lo_name);
891 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
892 	    (class->lc_flags & LC_SLEEPABLE) == 0)
893 		kassert_panic("%s: lock (%s) %s can not be sleepable",
894 		    __func__, class->lc_name, lock->lo_name);
895 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
896 	    (class->lc_flags & LC_UPGRADABLE) == 0)
897 		kassert_panic("%s: lock (%s) %s can not be upgradable",
898 		    __func__, class->lc_name, lock->lo_name);
899 
900 	/*
901 	 * If we shouldn't watch this lock, then just clear lo_witness.
902 	 * Otherwise, if witness_cold is set, then it is too early to
903 	 * enroll this lock, so defer it to witness_initialize() by adding
904 	 * it to the pending_locks list.  If it is not too early, then enroll
905 	 * the lock now.
906 	 */
907 	if (witness_watch < 1 || KERNEL_PANICKED() ||
908 	    (lock->lo_flags & LO_WITNESS) == 0)
909 		lock->lo_witness = NULL;
910 	else if (witness_cold) {
911 		pending_locks[pending_cnt].wh_lock = lock;
912 		pending_locks[pending_cnt++].wh_type = type;
913 		if (pending_cnt > WITNESS_PENDLIST)
914 			panic("%s: pending locks list is too small, "
915 			    "increase WITNESS_PENDLIST\n",
916 			    __func__);
917 	} else
918 		lock->lo_witness = enroll(type, class);
919 }
920 
921 void
922 witness_destroy(struct lock_object *lock)
923 {
924 	struct lock_class *class;
925 	struct witness *w;
926 
927 	class = LOCK_CLASS(lock);
928 
929 	if (witness_cold)
930 		panic("lock (%s) %s destroyed while witness_cold",
931 		    class->lc_name, lock->lo_name);
932 
933 	/* XXX: need to verify that no one holds the lock */
934 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
935 		return;
936 	w = lock->lo_witness;
937 
938 	mtx_lock_spin(&w_mtx);
939 	MPASS(w->w_refcount > 0);
940 	w->w_refcount--;
941 
942 	if (w->w_refcount == 0)
943 		depart(w);
944 	mtx_unlock_spin(&w_mtx);
945 }
946 
947 #ifdef DDB
948 static void
949 witness_ddb_compute_levels(void)
950 {
951 	struct witness *w;
952 
953 	/*
954 	 * First clear all levels.
955 	 */
956 	STAILQ_FOREACH(w, &w_all, w_list)
957 		w->w_ddb_level = -1;
958 
959 	/*
960 	 * Look for locks with no parents and level all their descendants.
961 	 */
962 	STAILQ_FOREACH(w, &w_all, w_list) {
963 		/* If the witness has ancestors (is not a root), skip it. */
964 		if (w->w_num_ancestors > 0)
965 			continue;
966 		witness_ddb_level_descendants(w, 0);
967 	}
968 }
969 
970 static void
971 witness_ddb_level_descendants(struct witness *w, int l)
972 {
973 	int i;
974 
975 	if (w->w_ddb_level >= l)
976 		return;
977 
978 	w->w_ddb_level = l;
979 	l++;
980 
981 	for (i = 1; i <= w_max_used_index; i++) {
982 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
983 			witness_ddb_level_descendants(&w_data[i], l);
984 	}
985 }
986 
987 static void
988 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
989     struct witness *w, int indent)
990 {
991 	int i;
992 
993  	for (i = 0; i < indent; i++)
994  		prnt(" ");
995 	prnt("%s (type: %s, depth: %d, active refs: %d)",
996 	     w->w_name, w->w_class->lc_name,
997 	     w->w_ddb_level, w->w_refcount);
998  	if (w->w_displayed) {
999  		prnt(" -- (already displayed)\n");
1000  		return;
1001  	}
1002  	w->w_displayed = 1;
1003 	if (w->w_file != NULL && w->w_line != 0)
1004 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1005 		    w->w_line);
1006 	else
1007 		prnt(" -- never acquired\n");
1008 	indent++;
1009 	WITNESS_INDEX_ASSERT(w->w_index);
1010 	for (i = 1; i <= w_max_used_index; i++) {
1011 		if (db_pager_quit)
1012 			return;
1013 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1014 			witness_ddb_display_descendants(prnt, &w_data[i],
1015 			    indent);
1016 	}
1017 }
1018 
1019 static void
1020 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1021     struct witness_list *list)
1022 {
1023 	struct witness *w;
1024 
1025 	STAILQ_FOREACH(w, list, w_typelist) {
1026 		if (w->w_file == NULL || w->w_ddb_level > 0)
1027 			continue;
1028 
1029 		/* This lock has no anscestors - display its descendants. */
1030 		witness_ddb_display_descendants(prnt, w, 0);
1031 		if (db_pager_quit)
1032 			return;
1033 	}
1034 }
1035 
1036 static void
1037 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1038 {
1039 	struct witness *w;
1040 
1041 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1042 	witness_ddb_compute_levels();
1043 
1044 	/* Clear all the displayed flags. */
1045 	STAILQ_FOREACH(w, &w_all, w_list)
1046 		w->w_displayed = 0;
1047 
1048 	/*
1049 	 * First, handle sleep locks which have been acquired at least
1050 	 * once.
1051 	 */
1052 	prnt("Sleep locks:\n");
1053 	witness_ddb_display_list(prnt, &w_sleep);
1054 	if (db_pager_quit)
1055 		return;
1056 
1057 	/*
1058 	 * Now do spin locks which have been acquired at least once.
1059 	 */
1060 	prnt("\nSpin locks:\n");
1061 	witness_ddb_display_list(prnt, &w_spin);
1062 	if (db_pager_quit)
1063 		return;
1064 
1065 	/*
1066 	 * Finally, any locks which have not been acquired yet.
1067 	 */
1068 	prnt("\nLocks which were never acquired:\n");
1069 	STAILQ_FOREACH(w, &w_all, w_list) {
1070 		if (w->w_file != NULL || w->w_refcount == 0)
1071 			continue;
1072 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1073 		    w->w_class->lc_name, w->w_ddb_level);
1074 		if (db_pager_quit)
1075 			return;
1076 	}
1077 }
1078 #endif /* DDB */
1079 
1080 int
1081 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1082 {
1083 
1084 	if (witness_watch == -1 || KERNEL_PANICKED())
1085 		return (0);
1086 
1087 	/* Require locks that witness knows about. */
1088 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1089 	    lock2->lo_witness == NULL)
1090 		return (EINVAL);
1091 
1092 	mtx_assert(&w_mtx, MA_NOTOWNED);
1093 	mtx_lock_spin(&w_mtx);
1094 
1095 	/*
1096 	 * If we already have either an explicit or implied lock order that
1097 	 * is the other way around, then return an error.
1098 	 */
1099 	if (witness_watch &&
1100 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1101 		mtx_unlock_spin(&w_mtx);
1102 		return (EDOOFUS);
1103 	}
1104 
1105 	/* Try to add the new order. */
1106 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1107 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1108 	itismychild(lock1->lo_witness, lock2->lo_witness);
1109 	mtx_unlock_spin(&w_mtx);
1110 	return (0);
1111 }
1112 
1113 void
1114 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1115     int line, struct lock_object *interlock)
1116 {
1117 	struct lock_list_entry *lock_list, *lle;
1118 	struct lock_instance *lock1, *lock2, *plock;
1119 	struct lock_class *class, *iclass;
1120 	struct witness *w, *w1;
1121 	struct thread *td;
1122 	int i, j;
1123 
1124 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1125 	    KERNEL_PANICKED())
1126 		return;
1127 
1128 	w = lock->lo_witness;
1129 	class = LOCK_CLASS(lock);
1130 	td = curthread;
1131 
1132 	if (class->lc_flags & LC_SLEEPLOCK) {
1133 		/*
1134 		 * Since spin locks include a critical section, this check
1135 		 * implicitly enforces a lock order of all sleep locks before
1136 		 * all spin locks.
1137 		 */
1138 		if (td->td_critnest != 0 && !kdb_active)
1139 			kassert_panic("acquiring blockable sleep lock with "
1140 			    "spinlock or critical section held (%s) %s @ %s:%d",
1141 			    class->lc_name, lock->lo_name,
1142 			    fixup_filename(file), line);
1143 
1144 		/*
1145 		 * If this is the first lock acquired then just return as
1146 		 * no order checking is needed.
1147 		 */
1148 		lock_list = td->td_sleeplocks;
1149 		if (lock_list == NULL || lock_list->ll_count == 0)
1150 			return;
1151 	} else {
1152 		/*
1153 		 * If this is the first lock, just return as no order
1154 		 * checking is needed.  Avoid problems with thread
1155 		 * migration pinning the thread while checking if
1156 		 * spinlocks are held.  If at least one spinlock is held
1157 		 * the thread is in a safe path and it is allowed to
1158 		 * unpin it.
1159 		 */
1160 		sched_pin();
1161 		lock_list = PCPU_GET(spinlocks);
1162 		if (lock_list == NULL || lock_list->ll_count == 0) {
1163 			sched_unpin();
1164 			return;
1165 		}
1166 		sched_unpin();
1167 	}
1168 
1169 	/*
1170 	 * Check to see if we are recursing on a lock we already own.  If
1171 	 * so, make sure that we don't mismatch exclusive and shared lock
1172 	 * acquires.
1173 	 */
1174 	lock1 = find_instance(lock_list, lock);
1175 	if (lock1 != NULL) {
1176 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1177 		    (flags & LOP_EXCLUSIVE) == 0) {
1178 			witness_output("shared lock of (%s) %s @ %s:%d\n",
1179 			    class->lc_name, lock->lo_name,
1180 			    fixup_filename(file), line);
1181 			witness_output("while exclusively locked from %s:%d\n",
1182 			    fixup_filename(lock1->li_file), lock1->li_line);
1183 			kassert_panic("excl->share");
1184 		}
1185 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1186 		    (flags & LOP_EXCLUSIVE) != 0) {
1187 			witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1188 			    class->lc_name, lock->lo_name,
1189 			    fixup_filename(file), line);
1190 			witness_output("while share locked from %s:%d\n",
1191 			    fixup_filename(lock1->li_file), lock1->li_line);
1192 			kassert_panic("share->excl");
1193 		}
1194 		return;
1195 	}
1196 
1197 	/* Warn if the interlock is not locked exactly once. */
1198 	if (interlock != NULL) {
1199 		iclass = LOCK_CLASS(interlock);
1200 		lock1 = find_instance(lock_list, interlock);
1201 		if (lock1 == NULL)
1202 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1203 			    iclass->lc_name, interlock->lo_name,
1204 			    fixup_filename(file), line);
1205 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1206 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1207 			    iclass->lc_name, interlock->lo_name,
1208 			    fixup_filename(file), line);
1209 	}
1210 
1211 	/*
1212 	 * Find the previously acquired lock, but ignore interlocks.
1213 	 */
1214 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1215 	if (interlock != NULL && plock->li_lock == interlock) {
1216 		if (lock_list->ll_count > 1)
1217 			plock =
1218 			    &lock_list->ll_children[lock_list->ll_count - 2];
1219 		else {
1220 			lle = lock_list->ll_next;
1221 
1222 			/*
1223 			 * The interlock is the only lock we hold, so
1224 			 * simply return.
1225 			 */
1226 			if (lle == NULL)
1227 				return;
1228 			plock = &lle->ll_children[lle->ll_count - 1];
1229 		}
1230 	}
1231 
1232 	/*
1233 	 * Try to perform most checks without a lock.  If this succeeds we
1234 	 * can skip acquiring the lock and return success.  Otherwise we redo
1235 	 * the check with the lock held to handle races with concurrent updates.
1236 	 */
1237 	w1 = plock->li_lock->lo_witness;
1238 	if (witness_lock_order_check(w1, w))
1239 		return;
1240 
1241 	mtx_lock_spin(&w_mtx);
1242 	if (witness_lock_order_check(w1, w)) {
1243 		mtx_unlock_spin(&w_mtx);
1244 		return;
1245 	}
1246 	witness_lock_order_add(w1, w);
1247 
1248 	/*
1249 	 * Check for duplicate locks of the same type.  Note that we only
1250 	 * have to check for this on the last lock we just acquired.  Any
1251 	 * other cases will be caught as lock order violations.
1252 	 */
1253 	if (w1 == w) {
1254 		i = w->w_index;
1255 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1256 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1257 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1258 			w->w_reversed = 1;
1259 			mtx_unlock_spin(&w_mtx);
1260 			witness_output(
1261 			    "acquiring duplicate lock of same type: \"%s\"\n",
1262 			    w->w_name);
1263 			witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1264 			    fixup_filename(plock->li_file), plock->li_line);
1265 			witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1266 			    fixup_filename(file), line);
1267 			witness_debugger(1, __func__);
1268 		} else
1269 			mtx_unlock_spin(&w_mtx);
1270 		return;
1271 	}
1272 	mtx_assert(&w_mtx, MA_OWNED);
1273 
1274 	/*
1275 	 * If we know that the lock we are acquiring comes after
1276 	 * the lock we most recently acquired in the lock order tree,
1277 	 * then there is no need for any further checks.
1278 	 */
1279 	if (isitmychild(w1, w))
1280 		goto out;
1281 
1282 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1283 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1284 			struct stack pstack;
1285 			bool pstackv, trace;
1286 
1287 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1288 			lock1 = &lle->ll_children[i];
1289 
1290 			/*
1291 			 * Ignore the interlock.
1292 			 */
1293 			if (interlock == lock1->li_lock)
1294 				continue;
1295 
1296 			/*
1297 			 * If this lock doesn't undergo witness checking,
1298 			 * then skip it.
1299 			 */
1300 			w1 = lock1->li_lock->lo_witness;
1301 			if (w1 == NULL) {
1302 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1303 				    ("lock missing witness structure"));
1304 				continue;
1305 			}
1306 
1307 			/*
1308 			 * If we are locking Giant and this is a sleepable
1309 			 * lock, then skip it.
1310 			 */
1311 			if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1312 			    lock == &Giant.lock_object)
1313 				continue;
1314 
1315 			/*
1316 			 * If we are locking a sleepable lock and this lock
1317 			 * is Giant, then skip it.
1318 			 */
1319 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1320 			    (flags & LOP_NOSLEEP) == 0 &&
1321 			    lock1->li_lock == &Giant.lock_object)
1322 				continue;
1323 
1324 			/*
1325 			 * If we are locking a sleepable lock and this lock
1326 			 * isn't sleepable, we want to treat it as a lock
1327 			 * order violation to enfore a general lock order of
1328 			 * sleepable locks before non-sleepable locks.
1329 			 */
1330 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1331 			    (flags & LOP_NOSLEEP) == 0 &&
1332 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1333 				goto reversal;
1334 
1335 			/*
1336 			 * If we are locking Giant and this is a non-sleepable
1337 			 * lock, then treat it as a reversal.
1338 			 */
1339 			if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1340 			    lock == &Giant.lock_object)
1341 				goto reversal;
1342 
1343 			/*
1344 			 * Check the lock order hierarchy for a reveresal.
1345 			 */
1346 			if (!isitmydescendant(w, w1))
1347 				continue;
1348 		reversal:
1349 
1350 			/*
1351 			 * We have a lock order violation, check to see if it
1352 			 * is allowed or has already been yelled about.
1353 			 */
1354 
1355 			/* Bail if this violation is known */
1356 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1357 				goto out;
1358 
1359 			/* Record this as a violation */
1360 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1361 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1362 			w->w_reversed = w1->w_reversed = 1;
1363 			witness_increment_graph_generation();
1364 
1365 			/*
1366 			 * If the lock order is blessed, bail before logging
1367 			 * anything.  We don't look for other lock order
1368 			 * violations though, which may be a bug.
1369 			 */
1370 			if (blessed(w, w1))
1371 				goto out;
1372 
1373 			trace = atomic_load_int(&witness_trace);
1374 			if (trace) {
1375 				struct witness_lock_order_data *data;
1376 
1377 				pstackv = false;
1378 				data = witness_lock_order_get(w, w1);
1379 				if (data != NULL) {
1380 					stack_copy(&data->wlod_stack,
1381 					    &pstack);
1382 					pstackv = true;
1383 				}
1384 			}
1385 			mtx_unlock_spin(&w_mtx);
1386 
1387 #ifdef WITNESS_NO_VNODE
1388 			/*
1389 			 * There are known LORs between VNODE locks. They are
1390 			 * not an indication of a bug. VNODE locks are flagged
1391 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1392 			 * is between 2 VNODE locks.
1393 			 */
1394 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1395 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1396 				return;
1397 #endif
1398 
1399 			/*
1400 			 * Ok, yell about it.
1401 			 */
1402 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1403 			    (flags & LOP_NOSLEEP) == 0 &&
1404 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1405 				witness_output(
1406 		"lock order reversal: (sleepable after non-sleepable)\n");
1407 			else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1408 			    && lock == &Giant.lock_object)
1409 				witness_output(
1410 		"lock order reversal: (Giant after non-sleepable)\n");
1411 			else
1412 				witness_output("lock order reversal:\n");
1413 
1414 			/*
1415 			 * Try to locate an earlier lock with
1416 			 * witness w in our list.
1417 			 */
1418 			do {
1419 				lock2 = &lle->ll_children[i];
1420 				MPASS(lock2->li_lock != NULL);
1421 				if (lock2->li_lock->lo_witness == w)
1422 					break;
1423 				if (i == 0 && lle->ll_next != NULL) {
1424 					lle = lle->ll_next;
1425 					i = lle->ll_count - 1;
1426 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1427 				} else
1428 					i--;
1429 			} while (i >= 0);
1430 			if (i < 0) {
1431 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1432 				    lock1->li_lock, lock1->li_lock->lo_name,
1433 				    w1->w_name, w1->w_class->lc_name,
1434 				    fixup_filename(lock1->li_file),
1435 				    lock1->li_line);
1436 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1437 				    lock, lock->lo_name, w->w_name,
1438 				    w->w_class->lc_name, fixup_filename(file),
1439 				    line);
1440 			} else {
1441 				struct witness *w2 = lock2->li_lock->lo_witness;
1442 
1443 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1444 				    lock2->li_lock, lock2->li_lock->lo_name,
1445 				    w2->w_name, w2->w_class->lc_name,
1446 				    fixup_filename(lock2->li_file),
1447 				    lock2->li_line);
1448 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1449 				    lock1->li_lock, lock1->li_lock->lo_name,
1450 				    w1->w_name, w1->w_class->lc_name,
1451 				    fixup_filename(lock1->li_file),
1452 				    lock1->li_line);
1453 				witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1454 				    lock->lo_name, w->w_name,
1455 				    w->w_class->lc_name, fixup_filename(file),
1456 				    line);
1457 			}
1458 			if (trace) {
1459 				char buf[64];
1460 				struct sbuf sb;
1461 
1462 				sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1463 				sbuf_set_drain(&sb, witness_output_drain,
1464 				    NULL);
1465 
1466 				if (pstackv) {
1467 					sbuf_printf(&sb,
1468 				    "lock order %s -> %s established at:\n",
1469 					    w->w_name, w1->w_name);
1470 					stack_sbuf_print_flags(&sb, &pstack,
1471 					    M_NOWAIT, STACK_SBUF_FMT_LONG);
1472 				}
1473 
1474 				sbuf_printf(&sb,
1475 				    "lock order %s -> %s attempted at:\n",
1476 				    w1->w_name, w->w_name);
1477 				stack_save(&pstack);
1478 				stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1479 				    STACK_SBUF_FMT_LONG);
1480 
1481 				sbuf_finish(&sb);
1482 				sbuf_delete(&sb);
1483 			}
1484 			witness_enter_debugger(__func__);
1485 			return;
1486 		}
1487 	}
1488 
1489 	/*
1490 	 * If requested, build a new lock order.  However, don't build a new
1491 	 * relationship between a sleepable lock and Giant if it is in the
1492 	 * wrong direction.  The correct lock order is that sleepable locks
1493 	 * always come before Giant.
1494 	 */
1495 	if (flags & LOP_NEWORDER &&
1496 	    !(plock->li_lock == &Giant.lock_object &&
1497 	    (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1498 	    (flags & LOP_NOSLEEP) == 0)) {
1499 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1500 		    w->w_name, plock->li_lock->lo_witness->w_name);
1501 		itismychild(plock->li_lock->lo_witness, w);
1502 	}
1503 out:
1504 	mtx_unlock_spin(&w_mtx);
1505 }
1506 
1507 void
1508 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1509 {
1510 	struct lock_list_entry **lock_list, *lle;
1511 	struct lock_instance *instance;
1512 	struct witness *w;
1513 	struct thread *td;
1514 
1515 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1516 	    KERNEL_PANICKED())
1517 		return;
1518 	w = lock->lo_witness;
1519 	td = curthread;
1520 
1521 	/* Determine lock list for this lock. */
1522 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1523 		lock_list = &td->td_sleeplocks;
1524 	else
1525 		lock_list = PCPU_PTR(spinlocks);
1526 
1527 	/* Check to see if we are recursing on a lock we already own. */
1528 	instance = find_instance(*lock_list, lock);
1529 	if (instance != NULL) {
1530 		instance->li_flags++;
1531 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1532 		    td->td_proc->p_pid, lock->lo_name,
1533 		    instance->li_flags & LI_RECURSEMASK);
1534 		instance->li_file = file;
1535 		instance->li_line = line;
1536 		return;
1537 	}
1538 
1539 	/* Update per-witness last file and line acquire. */
1540 	w->w_file = file;
1541 	w->w_line = line;
1542 
1543 	/* Find the next open lock instance in the list and fill it. */
1544 	lle = *lock_list;
1545 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1546 		lle = witness_lock_list_get();
1547 		if (lle == NULL)
1548 			return;
1549 		lle->ll_next = *lock_list;
1550 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1551 		    td->td_proc->p_pid, lle);
1552 		*lock_list = lle;
1553 	}
1554 	instance = &lle->ll_children[lle->ll_count++];
1555 	instance->li_lock = lock;
1556 	instance->li_line = line;
1557 	instance->li_file = file;
1558 	instance->li_flags = 0;
1559 	if ((flags & LOP_EXCLUSIVE) != 0)
1560 		instance->li_flags |= LI_EXCLUSIVE;
1561 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1562 		instance->li_flags |= LI_SLEEPABLE;
1563 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1564 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1565 }
1566 
1567 void
1568 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1569 {
1570 	struct lock_instance *instance;
1571 	struct lock_class *class;
1572 
1573 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1574 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1575 		return;
1576 	class = LOCK_CLASS(lock);
1577 	if (witness_watch) {
1578 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1579 			kassert_panic(
1580 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1581 			    class->lc_name, lock->lo_name,
1582 			    fixup_filename(file), line);
1583 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1584 			kassert_panic(
1585 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1586 			    class->lc_name, lock->lo_name,
1587 			    fixup_filename(file), line);
1588 	}
1589 	instance = find_instance(curthread->td_sleeplocks, lock);
1590 	if (instance == NULL) {
1591 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1592 		    class->lc_name, lock->lo_name,
1593 		    fixup_filename(file), line);
1594 		return;
1595 	}
1596 	if (witness_watch) {
1597 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1598 			kassert_panic(
1599 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1600 			    class->lc_name, lock->lo_name,
1601 			    fixup_filename(file), line);
1602 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1603 			kassert_panic(
1604 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1605 			    class->lc_name, lock->lo_name,
1606 			    instance->li_flags & LI_RECURSEMASK,
1607 			    fixup_filename(file), line);
1608 	}
1609 	instance->li_flags |= LI_EXCLUSIVE;
1610 }
1611 
1612 void
1613 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1614     int line)
1615 {
1616 	struct lock_instance *instance;
1617 	struct lock_class *class;
1618 
1619 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1620 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1621 		return;
1622 	class = LOCK_CLASS(lock);
1623 	if (witness_watch) {
1624 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1625 			kassert_panic(
1626 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1627 			    class->lc_name, lock->lo_name,
1628 			    fixup_filename(file), line);
1629 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1630 			kassert_panic(
1631 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1632 			    class->lc_name, lock->lo_name,
1633 			    fixup_filename(file), line);
1634 	}
1635 	instance = find_instance(curthread->td_sleeplocks, lock);
1636 	if (instance == NULL) {
1637 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1638 		    class->lc_name, lock->lo_name,
1639 		    fixup_filename(file), line);
1640 		return;
1641 	}
1642 	if (witness_watch) {
1643 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1644 			kassert_panic(
1645 			    "downgrade of shared lock (%s) %s @ %s:%d",
1646 			    class->lc_name, lock->lo_name,
1647 			    fixup_filename(file), line);
1648 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1649 			kassert_panic(
1650 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1651 			    class->lc_name, lock->lo_name,
1652 			    instance->li_flags & LI_RECURSEMASK,
1653 			    fixup_filename(file), line);
1654 	}
1655 	instance->li_flags &= ~LI_EXCLUSIVE;
1656 }
1657 
1658 void
1659 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1660 {
1661 	struct lock_list_entry **lock_list, *lle;
1662 	struct lock_instance *instance;
1663 	struct lock_class *class;
1664 	struct thread *td;
1665 	register_t s;
1666 	int i, j;
1667 
1668 	if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1669 		return;
1670 	td = curthread;
1671 	class = LOCK_CLASS(lock);
1672 
1673 	/* Find lock instance associated with this lock. */
1674 	if (class->lc_flags & LC_SLEEPLOCK)
1675 		lock_list = &td->td_sleeplocks;
1676 	else
1677 		lock_list = PCPU_PTR(spinlocks);
1678 	lle = *lock_list;
1679 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1680 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1681 			instance = &(*lock_list)->ll_children[i];
1682 			if (instance->li_lock == lock)
1683 				goto found;
1684 		}
1685 
1686 	/*
1687 	 * When disabling WITNESS through witness_watch we could end up in
1688 	 * having registered locks in the td_sleeplocks queue.
1689 	 * We have to make sure we flush these queues, so just search for
1690 	 * eventual register locks and remove them.
1691 	 */
1692 	if (witness_watch > 0) {
1693 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1694 		    lock->lo_name, fixup_filename(file), line);
1695 		return;
1696 	} else {
1697 		return;
1698 	}
1699 found:
1700 
1701 	/* First, check for shared/exclusive mismatches. */
1702 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1703 	    (flags & LOP_EXCLUSIVE) == 0) {
1704 		witness_output("shared unlock of (%s) %s @ %s:%d\n",
1705 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1706 		witness_output("while exclusively locked from %s:%d\n",
1707 		    fixup_filename(instance->li_file), instance->li_line);
1708 		kassert_panic("excl->ushare");
1709 	}
1710 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1711 	    (flags & LOP_EXCLUSIVE) != 0) {
1712 		witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1713 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1714 		witness_output("while share locked from %s:%d\n",
1715 		    fixup_filename(instance->li_file),
1716 		    instance->li_line);
1717 		kassert_panic("share->uexcl");
1718 	}
1719 	/* If we are recursed, unrecurse. */
1720 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1721 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1722 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1723 		    instance->li_flags);
1724 		instance->li_flags--;
1725 		return;
1726 	}
1727 	/* The lock is now being dropped, check for NORELEASE flag */
1728 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1729 		witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1730 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1731 		kassert_panic("lock marked norelease");
1732 	}
1733 
1734 	/* Otherwise, remove this item from the list. */
1735 	s = intr_disable();
1736 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1737 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1738 	    (*lock_list)->ll_count - 1);
1739 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1740 		(*lock_list)->ll_children[j] =
1741 		    (*lock_list)->ll_children[j + 1];
1742 	(*lock_list)->ll_count--;
1743 	intr_restore(s);
1744 
1745 	/*
1746 	 * In order to reduce contention on w_mtx, we want to keep always an
1747 	 * head object into lists so that frequent allocation from the
1748 	 * free witness pool (and subsequent locking) is avoided.
1749 	 * In order to maintain the current code simple, when the head
1750 	 * object is totally unloaded it means also that we do not have
1751 	 * further objects in the list, so the list ownership needs to be
1752 	 * hand over to another object if the current head needs to be freed.
1753 	 */
1754 	if ((*lock_list)->ll_count == 0) {
1755 		if (*lock_list == lle) {
1756 			if (lle->ll_next == NULL)
1757 				return;
1758 		} else
1759 			lle = *lock_list;
1760 		*lock_list = lle->ll_next;
1761 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1762 		    td->td_proc->p_pid, lle);
1763 		witness_lock_list_free(lle);
1764 	}
1765 }
1766 
1767 void
1768 witness_thread_exit(struct thread *td)
1769 {
1770 	struct lock_list_entry *lle;
1771 	int i, n;
1772 
1773 	lle = td->td_sleeplocks;
1774 	if (lle == NULL || KERNEL_PANICKED())
1775 		return;
1776 	if (lle->ll_count != 0) {
1777 		for (n = 0; lle != NULL; lle = lle->ll_next)
1778 			for (i = lle->ll_count - 1; i >= 0; i--) {
1779 				if (n == 0)
1780 					witness_output(
1781 		    "Thread %p exiting with the following locks held:\n", td);
1782 				n++;
1783 				witness_list_lock(&lle->ll_children[i],
1784 				    witness_output);
1785 
1786 			}
1787 		kassert_panic(
1788 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1789 	}
1790 	witness_lock_list_free(lle);
1791 }
1792 
1793 /*
1794  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1795  * exempt Giant and sleepable locks from the checks as well.  If any
1796  * non-exempt locks are held, then a supplied message is printed to the
1797  * output channel along with a list of the offending locks.  If indicated in the
1798  * flags then a failure results in a panic as well.
1799  */
1800 int
1801 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1802 {
1803 	struct lock_list_entry *lock_list, *lle;
1804 	struct lock_instance *lock1;
1805 	struct thread *td;
1806 	va_list ap;
1807 	int i, n;
1808 
1809 	if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1810 		return (0);
1811 	n = 0;
1812 	td = curthread;
1813 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1814 		for (i = lle->ll_count - 1; i >= 0; i--) {
1815 			lock1 = &lle->ll_children[i];
1816 			if (lock1->li_lock == lock)
1817 				continue;
1818 			if (flags & WARN_GIANTOK &&
1819 			    lock1->li_lock == &Giant.lock_object)
1820 				continue;
1821 			if (flags & WARN_SLEEPOK &&
1822 			    (lock1->li_flags & LI_SLEEPABLE) != 0)
1823 				continue;
1824 			if (n == 0) {
1825 				va_start(ap, fmt);
1826 				vprintf(fmt, ap);
1827 				va_end(ap);
1828 				printf(" with the following %slocks held:\n",
1829 				    (flags & WARN_SLEEPOK) != 0 ?
1830 				    "non-sleepable " : "");
1831 			}
1832 			n++;
1833 			witness_list_lock(lock1, printf);
1834 		}
1835 
1836 	/*
1837 	 * Pin the thread in order to avoid problems with thread migration.
1838 	 * Once that all verifies are passed about spinlocks ownership,
1839 	 * the thread is in a safe path and it can be unpinned.
1840 	 */
1841 	sched_pin();
1842 	lock_list = PCPU_GET(spinlocks);
1843 	if (lock_list != NULL && lock_list->ll_count != 0) {
1844 		sched_unpin();
1845 
1846 		/*
1847 		 * We should only have one spinlock and as long as
1848 		 * the flags cannot match for this locks class,
1849 		 * check if the first spinlock is the one curthread
1850 		 * should hold.
1851 		 */
1852 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1853 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1854 		    lock1->li_lock == lock && n == 0)
1855 			return (0);
1856 
1857 		va_start(ap, fmt);
1858 		vprintf(fmt, ap);
1859 		va_end(ap);
1860 		printf(" with the following %slocks held:\n",
1861 		    (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1862 		n += witness_list_locks(&lock_list, printf);
1863 	} else
1864 		sched_unpin();
1865 	if (flags & WARN_PANIC && n)
1866 		kassert_panic("%s", __func__);
1867 	else
1868 		witness_debugger(n, __func__);
1869 	return (n);
1870 }
1871 
1872 const char *
1873 witness_file(struct lock_object *lock)
1874 {
1875 	struct witness *w;
1876 
1877 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1878 		return ("?");
1879 	w = lock->lo_witness;
1880 	return (w->w_file);
1881 }
1882 
1883 int
1884 witness_line(struct lock_object *lock)
1885 {
1886 	struct witness *w;
1887 
1888 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1889 		return (0);
1890 	w = lock->lo_witness;
1891 	return (w->w_line);
1892 }
1893 
1894 static struct witness *
1895 enroll(const char *description, struct lock_class *lock_class)
1896 {
1897 	struct witness *w;
1898 
1899 	MPASS(description != NULL);
1900 
1901 	if (witness_watch == -1 || KERNEL_PANICKED())
1902 		return (NULL);
1903 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1904 		if (witness_skipspin)
1905 			return (NULL);
1906 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1907 		kassert_panic("lock class %s is not sleep or spin",
1908 		    lock_class->lc_name);
1909 		return (NULL);
1910 	}
1911 
1912 	mtx_lock_spin(&w_mtx);
1913 	w = witness_hash_get(description);
1914 	if (w)
1915 		goto found;
1916 	if ((w = witness_get()) == NULL)
1917 		return (NULL);
1918 	MPASS(strlen(description) < MAX_W_NAME);
1919 	strcpy(w->w_name, description);
1920 	w->w_class = lock_class;
1921 	w->w_refcount = 1;
1922 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1923 	if (lock_class->lc_flags & LC_SPINLOCK) {
1924 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1925 		w_spin_cnt++;
1926 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1927 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1928 		w_sleep_cnt++;
1929 	}
1930 
1931 	/* Insert new witness into the hash */
1932 	witness_hash_put(w);
1933 	witness_increment_graph_generation();
1934 	mtx_unlock_spin(&w_mtx);
1935 	return (w);
1936 found:
1937 	w->w_refcount++;
1938 	if (w->w_refcount == 1)
1939 		w->w_class = lock_class;
1940 	mtx_unlock_spin(&w_mtx);
1941 	if (lock_class != w->w_class)
1942 		kassert_panic(
1943 		    "lock (%s) %s does not match earlier (%s) lock",
1944 		    description, lock_class->lc_name,
1945 		    w->w_class->lc_name);
1946 	return (w);
1947 }
1948 
1949 static void
1950 depart(struct witness *w)
1951 {
1952 
1953 	MPASS(w->w_refcount == 0);
1954 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1955 		w_sleep_cnt--;
1956 	} else {
1957 		w_spin_cnt--;
1958 	}
1959 	/*
1960 	 * Set file to NULL as it may point into a loadable module.
1961 	 */
1962 	w->w_file = NULL;
1963 	w->w_line = 0;
1964 	witness_increment_graph_generation();
1965 }
1966 
1967 static void
1968 adopt(struct witness *parent, struct witness *child)
1969 {
1970 	int pi, ci, i, j;
1971 
1972 	if (witness_cold == 0)
1973 		mtx_assert(&w_mtx, MA_OWNED);
1974 
1975 	/* If the relationship is already known, there's no work to be done. */
1976 	if (isitmychild(parent, child))
1977 		return;
1978 
1979 	/* When the structure of the graph changes, bump up the generation. */
1980 	witness_increment_graph_generation();
1981 
1982 	/*
1983 	 * The hard part ... create the direct relationship, then propagate all
1984 	 * indirect relationships.
1985 	 */
1986 	pi = parent->w_index;
1987 	ci = child->w_index;
1988 	WITNESS_INDEX_ASSERT(pi);
1989 	WITNESS_INDEX_ASSERT(ci);
1990 	MPASS(pi != ci);
1991 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1992 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1993 
1994 	/*
1995 	 * If parent was not already an ancestor of child,
1996 	 * then we increment the descendant and ancestor counters.
1997 	 */
1998 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1999 		parent->w_num_descendants++;
2000 		child->w_num_ancestors++;
2001 	}
2002 
2003 	/*
2004 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2005 	 * an ancestor of 'pi' during this loop.
2006 	 */
2007 	for (i = 1; i <= w_max_used_index; i++) {
2008 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2009 		    (i != pi))
2010 			continue;
2011 
2012 		/* Find each descendant of 'i' and mark it as a descendant. */
2013 		for (j = 1; j <= w_max_used_index; j++) {
2014 			/*
2015 			 * Skip children that are already marked as
2016 			 * descendants of 'i'.
2017 			 */
2018 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2019 				continue;
2020 
2021 			/*
2022 			 * We are only interested in descendants of 'ci'. Note
2023 			 * that 'ci' itself is counted as a descendant of 'ci'.
2024 			 */
2025 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2026 			    (j != ci))
2027 				continue;
2028 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2029 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2030 			w_data[i].w_num_descendants++;
2031 			w_data[j].w_num_ancestors++;
2032 
2033 			/*
2034 			 * Make sure we aren't marking a node as both an
2035 			 * ancestor and descendant. We should have caught
2036 			 * this as a lock order reversal earlier.
2037 			 */
2038 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2039 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2040 				printf("witness rmatrix paradox! [%d][%d]=%d "
2041 				    "both ancestor and descendant\n",
2042 				    i, j, w_rmatrix[i][j]);
2043 				kdb_backtrace();
2044 				printf("Witness disabled.\n");
2045 				witness_watch = -1;
2046 			}
2047 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2048 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2049 				printf("witness rmatrix paradox! [%d][%d]=%d "
2050 				    "both ancestor and descendant\n",
2051 				    j, i, w_rmatrix[j][i]);
2052 				kdb_backtrace();
2053 				printf("Witness disabled.\n");
2054 				witness_watch = -1;
2055 			}
2056 		}
2057 	}
2058 }
2059 
2060 static void
2061 itismychild(struct witness *parent, struct witness *child)
2062 {
2063 	int unlocked;
2064 
2065 	MPASS(child != NULL && parent != NULL);
2066 	if (witness_cold == 0)
2067 		mtx_assert(&w_mtx, MA_OWNED);
2068 
2069 	if (!witness_lock_type_equal(parent, child)) {
2070 		if (witness_cold == 0) {
2071 			unlocked = 1;
2072 			mtx_unlock_spin(&w_mtx);
2073 		} else {
2074 			unlocked = 0;
2075 		}
2076 		kassert_panic(
2077 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2078 		    "the same lock type", __func__, parent->w_name,
2079 		    parent->w_class->lc_name, child->w_name,
2080 		    child->w_class->lc_name);
2081 		if (unlocked)
2082 			mtx_lock_spin(&w_mtx);
2083 	}
2084 	adopt(parent, child);
2085 }
2086 
2087 /*
2088  * Generic code for the isitmy*() functions. The rmask parameter is the
2089  * expected relationship of w1 to w2.
2090  */
2091 static int
2092 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2093 {
2094 	unsigned char r1, r2;
2095 	int i1, i2;
2096 
2097 	i1 = w1->w_index;
2098 	i2 = w2->w_index;
2099 	WITNESS_INDEX_ASSERT(i1);
2100 	WITNESS_INDEX_ASSERT(i2);
2101 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2102 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2103 
2104 	/* The flags on one better be the inverse of the flags on the other */
2105 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2106 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2107 		/* Don't squawk if we're potentially racing with an update. */
2108 		if (!mtx_owned(&w_mtx))
2109 			return (0);
2110 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2111 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2112 		    "w_rmatrix[%d][%d] == %hhx\n",
2113 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2114 		    i2, i1, r2);
2115 		kdb_backtrace();
2116 		printf("Witness disabled.\n");
2117 		witness_watch = -1;
2118 	}
2119 	return (r1 & rmask);
2120 }
2121 
2122 /*
2123  * Checks if @child is a direct child of @parent.
2124  */
2125 static int
2126 isitmychild(struct witness *parent, struct witness *child)
2127 {
2128 
2129 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2130 }
2131 
2132 /*
2133  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2134  */
2135 static int
2136 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2137 {
2138 
2139 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2140 	    __func__));
2141 }
2142 
2143 static int
2144 blessed(struct witness *w1, struct witness *w2)
2145 {
2146 	int i;
2147 	struct witness_blessed *b;
2148 
2149 	for (i = 0; i < nitems(blessed_list); i++) {
2150 		b = &blessed_list[i];
2151 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2152 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2153 				return (1);
2154 			continue;
2155 		}
2156 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2157 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2158 				return (1);
2159 	}
2160 	return (0);
2161 }
2162 
2163 static struct witness *
2164 witness_get(void)
2165 {
2166 	struct witness *w;
2167 	int index;
2168 
2169 	if (witness_cold == 0)
2170 		mtx_assert(&w_mtx, MA_OWNED);
2171 
2172 	if (witness_watch == -1) {
2173 		mtx_unlock_spin(&w_mtx);
2174 		return (NULL);
2175 	}
2176 	if (STAILQ_EMPTY(&w_free)) {
2177 		witness_watch = -1;
2178 		mtx_unlock_spin(&w_mtx);
2179 		printf("WITNESS: unable to allocate a new witness object\n");
2180 		return (NULL);
2181 	}
2182 	w = STAILQ_FIRST(&w_free);
2183 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2184 	w_free_cnt--;
2185 	index = w->w_index;
2186 	MPASS(index > 0 && index == w_max_used_index+1 &&
2187 	    index < witness_count);
2188 	bzero(w, sizeof(*w));
2189 	w->w_index = index;
2190 	if (index > w_max_used_index)
2191 		w_max_used_index = index;
2192 	return (w);
2193 }
2194 
2195 static void
2196 witness_free(struct witness *w)
2197 {
2198 
2199 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2200 	w_free_cnt++;
2201 }
2202 
2203 static struct lock_list_entry *
2204 witness_lock_list_get(void)
2205 {
2206 	struct lock_list_entry *lle;
2207 
2208 	if (witness_watch == -1)
2209 		return (NULL);
2210 	mtx_lock_spin(&w_mtx);
2211 	lle = w_lock_list_free;
2212 	if (lle == NULL) {
2213 		witness_watch = -1;
2214 		mtx_unlock_spin(&w_mtx);
2215 		printf("%s: witness exhausted\n", __func__);
2216 		return (NULL);
2217 	}
2218 	w_lock_list_free = lle->ll_next;
2219 	mtx_unlock_spin(&w_mtx);
2220 	bzero(lle, sizeof(*lle));
2221 	return (lle);
2222 }
2223 
2224 static void
2225 witness_lock_list_free(struct lock_list_entry *lle)
2226 {
2227 
2228 	mtx_lock_spin(&w_mtx);
2229 	lle->ll_next = w_lock_list_free;
2230 	w_lock_list_free = lle;
2231 	mtx_unlock_spin(&w_mtx);
2232 }
2233 
2234 static struct lock_instance *
2235 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2236 {
2237 	struct lock_list_entry *lle;
2238 	struct lock_instance *instance;
2239 	int i;
2240 
2241 	for (lle = list; lle != NULL; lle = lle->ll_next)
2242 		for (i = lle->ll_count - 1; i >= 0; i--) {
2243 			instance = &lle->ll_children[i];
2244 			if (instance->li_lock == lock)
2245 				return (instance);
2246 		}
2247 	return (NULL);
2248 }
2249 
2250 static void
2251 witness_list_lock(struct lock_instance *instance,
2252     int (*prnt)(const char *fmt, ...))
2253 {
2254 	struct lock_object *lock;
2255 
2256 	lock = instance->li_lock;
2257 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2258 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2259 	if (lock->lo_witness->w_name != lock->lo_name)
2260 		prnt(" (%s)", lock->lo_witness->w_name);
2261 	prnt(" r = %d (%p) locked @ %s:%d\n",
2262 	    instance->li_flags & LI_RECURSEMASK, lock,
2263 	    fixup_filename(instance->li_file), instance->li_line);
2264 }
2265 
2266 static int
2267 witness_output(const char *fmt, ...)
2268 {
2269 	va_list ap;
2270 	int ret;
2271 
2272 	va_start(ap, fmt);
2273 	ret = witness_voutput(fmt, ap);
2274 	va_end(ap);
2275 	return (ret);
2276 }
2277 
2278 static int
2279 witness_voutput(const char *fmt, va_list ap)
2280 {
2281 	int ret;
2282 
2283 	ret = 0;
2284 	switch (witness_channel) {
2285 	case WITNESS_CONSOLE:
2286 		ret = vprintf(fmt, ap);
2287 		break;
2288 	case WITNESS_LOG:
2289 		vlog(LOG_NOTICE, fmt, ap);
2290 		break;
2291 	case WITNESS_NONE:
2292 		break;
2293 	}
2294 	return (ret);
2295 }
2296 
2297 #ifdef DDB
2298 static int
2299 witness_thread_has_locks(struct thread *td)
2300 {
2301 
2302 	if (td->td_sleeplocks == NULL)
2303 		return (0);
2304 	return (td->td_sleeplocks->ll_count != 0);
2305 }
2306 
2307 static int
2308 witness_proc_has_locks(struct proc *p)
2309 {
2310 	struct thread *td;
2311 
2312 	FOREACH_THREAD_IN_PROC(p, td) {
2313 		if (witness_thread_has_locks(td))
2314 			return (1);
2315 	}
2316 	return (0);
2317 }
2318 #endif
2319 
2320 int
2321 witness_list_locks(struct lock_list_entry **lock_list,
2322     int (*prnt)(const char *fmt, ...))
2323 {
2324 	struct lock_list_entry *lle;
2325 	int i, nheld;
2326 
2327 	nheld = 0;
2328 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2329 		for (i = lle->ll_count - 1; i >= 0; i--) {
2330 			witness_list_lock(&lle->ll_children[i], prnt);
2331 			nheld++;
2332 		}
2333 	return (nheld);
2334 }
2335 
2336 /*
2337  * This is a bit risky at best.  We call this function when we have timed
2338  * out acquiring a spin lock, and we assume that the other CPU is stuck
2339  * with this lock held.  So, we go groveling around in the other CPU's
2340  * per-cpu data to try to find the lock instance for this spin lock to
2341  * see when it was last acquired.
2342  */
2343 void
2344 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2345     int (*prnt)(const char *fmt, ...))
2346 {
2347 	struct lock_instance *instance;
2348 	struct pcpu *pc;
2349 
2350 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2351 		return;
2352 	pc = pcpu_find(owner->td_oncpu);
2353 	instance = find_instance(pc->pc_spinlocks, lock);
2354 	if (instance != NULL)
2355 		witness_list_lock(instance, prnt);
2356 }
2357 
2358 void
2359 witness_save(struct lock_object *lock, const char **filep, int *linep)
2360 {
2361 	struct lock_list_entry *lock_list;
2362 	struct lock_instance *instance;
2363 	struct lock_class *class;
2364 
2365 	/* Initialize for KMSAN's benefit. */
2366 	*filep = NULL;
2367 	*linep = 0;
2368 
2369 	/*
2370 	 * This function is used independently in locking code to deal with
2371 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2372 	 * is gone.
2373 	 */
2374 	if (SCHEDULER_STOPPED())
2375 		return;
2376 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2377 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2378 		return;
2379 	class = LOCK_CLASS(lock);
2380 	if (class->lc_flags & LC_SLEEPLOCK)
2381 		lock_list = curthread->td_sleeplocks;
2382 	else {
2383 		if (witness_skipspin)
2384 			return;
2385 		lock_list = PCPU_GET(spinlocks);
2386 	}
2387 	instance = find_instance(lock_list, lock);
2388 	if (instance == NULL) {
2389 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2390 		    class->lc_name, lock->lo_name);
2391 		return;
2392 	}
2393 	*filep = instance->li_file;
2394 	*linep = instance->li_line;
2395 }
2396 
2397 void
2398 witness_restore(struct lock_object *lock, const char *file, int line)
2399 {
2400 	struct lock_list_entry *lock_list;
2401 	struct lock_instance *instance;
2402 	struct lock_class *class;
2403 
2404 	/*
2405 	 * This function is used independently in locking code to deal with
2406 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2407 	 * is gone.
2408 	 */
2409 	if (SCHEDULER_STOPPED())
2410 		return;
2411 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2412 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2413 		return;
2414 	class = LOCK_CLASS(lock);
2415 	if (class->lc_flags & LC_SLEEPLOCK)
2416 		lock_list = curthread->td_sleeplocks;
2417 	else {
2418 		if (witness_skipspin)
2419 			return;
2420 		lock_list = PCPU_GET(spinlocks);
2421 	}
2422 	instance = find_instance(lock_list, lock);
2423 	if (instance == NULL)
2424 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2425 		    class->lc_name, lock->lo_name);
2426 	lock->lo_witness->w_file = file;
2427 	lock->lo_witness->w_line = line;
2428 	if (instance == NULL)
2429 		return;
2430 	instance->li_file = file;
2431 	instance->li_line = line;
2432 }
2433 
2434 static bool
2435 witness_find_instance(const struct lock_object *lock,
2436     struct lock_instance **instance)
2437 {
2438 #ifdef INVARIANT_SUPPORT
2439 	struct lock_class *class;
2440 
2441 	if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2442 		return (false);
2443 	class = LOCK_CLASS(lock);
2444 	if ((class->lc_flags & LC_SLEEPLOCK) != 0) {
2445 		*instance = find_instance(curthread->td_sleeplocks, lock);
2446 		return (true);
2447 	} else if ((class->lc_flags & LC_SPINLOCK) != 0) {
2448 		*instance = find_instance(PCPU_GET(spinlocks), lock);
2449 		return (true);
2450 	} else {
2451 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2452 		    class->lc_name, lock->lo_name);
2453 		return (false);
2454 	}
2455 #else
2456 	return (false);
2457 #endif
2458 }
2459 
2460 void
2461 witness_assert(const struct lock_object *lock, int flags, const char *file,
2462     int line)
2463 {
2464 #ifdef INVARIANT_SUPPORT
2465 	struct lock_instance *instance;
2466 	struct lock_class *class;
2467 
2468 	if (!witness_find_instance(lock, &instance))
2469 		return;
2470 	class = LOCK_CLASS(lock);
2471 	switch (flags) {
2472 	case LA_UNLOCKED:
2473 		if (instance != NULL)
2474 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2475 			    class->lc_name, lock->lo_name,
2476 			    fixup_filename(file), line);
2477 		break;
2478 	case LA_LOCKED:
2479 	case LA_LOCKED | LA_RECURSED:
2480 	case LA_LOCKED | LA_NOTRECURSED:
2481 	case LA_SLOCKED:
2482 	case LA_SLOCKED | LA_RECURSED:
2483 	case LA_SLOCKED | LA_NOTRECURSED:
2484 	case LA_XLOCKED:
2485 	case LA_XLOCKED | LA_RECURSED:
2486 	case LA_XLOCKED | LA_NOTRECURSED:
2487 		if (instance == NULL) {
2488 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2489 			    class->lc_name, lock->lo_name,
2490 			    fixup_filename(file), line);
2491 			break;
2492 		}
2493 		if ((flags & LA_XLOCKED) != 0 &&
2494 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2495 			kassert_panic(
2496 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2497 			    class->lc_name, lock->lo_name,
2498 			    fixup_filename(file), line);
2499 		if ((flags & LA_SLOCKED) != 0 &&
2500 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2501 			kassert_panic(
2502 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2503 			    class->lc_name, lock->lo_name,
2504 			    fixup_filename(file), line);
2505 		if ((flags & LA_RECURSED) != 0 &&
2506 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2507 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2508 			    class->lc_name, lock->lo_name,
2509 			    fixup_filename(file), line);
2510 		if ((flags & LA_NOTRECURSED) != 0 &&
2511 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2512 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2513 			    class->lc_name, lock->lo_name,
2514 			    fixup_filename(file), line);
2515 		break;
2516 	default:
2517 		kassert_panic("Invalid lock assertion at %s:%d.",
2518 		    fixup_filename(file), line);
2519 	}
2520 #endif	/* INVARIANT_SUPPORT */
2521 }
2522 
2523 /*
2524  * Checks the ownership of the lock by curthread, consulting the witness list.
2525  * Returns:
2526  *   0  if witness is disabled or did not work
2527  *   -1 if not owned
2528  *   1  if owned
2529  */
2530 int
2531 witness_is_owned(const struct lock_object *lock)
2532 {
2533 #ifdef INVARIANT_SUPPORT
2534 	struct lock_instance *instance;
2535 
2536 	if (!witness_find_instance(lock, &instance))
2537 		return (0);
2538 	return (instance == NULL ? -1 : 1);
2539 #else
2540 	return (0);
2541 #endif
2542 }
2543 
2544 static void
2545 witness_setflag(struct lock_object *lock, int flag, int set)
2546 {
2547 	struct lock_list_entry *lock_list;
2548 	struct lock_instance *instance;
2549 	struct lock_class *class;
2550 
2551 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2552 		return;
2553 	class = LOCK_CLASS(lock);
2554 	if (class->lc_flags & LC_SLEEPLOCK)
2555 		lock_list = curthread->td_sleeplocks;
2556 	else {
2557 		if (witness_skipspin)
2558 			return;
2559 		lock_list = PCPU_GET(spinlocks);
2560 	}
2561 	instance = find_instance(lock_list, lock);
2562 	if (instance == NULL) {
2563 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2564 		    class->lc_name, lock->lo_name);
2565 		return;
2566 	}
2567 
2568 	if (set)
2569 		instance->li_flags |= flag;
2570 	else
2571 		instance->li_flags &= ~flag;
2572 }
2573 
2574 void
2575 witness_norelease(struct lock_object *lock)
2576 {
2577 
2578 	witness_setflag(lock, LI_NORELEASE, 1);
2579 }
2580 
2581 void
2582 witness_releaseok(struct lock_object *lock)
2583 {
2584 
2585 	witness_setflag(lock, LI_NORELEASE, 0);
2586 }
2587 
2588 #ifdef DDB
2589 static void
2590 witness_ddb_list(struct thread *td)
2591 {
2592 
2593 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2594 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2595 
2596 	if (witness_watch < 1)
2597 		return;
2598 
2599 	witness_list_locks(&td->td_sleeplocks, db_printf);
2600 
2601 	/*
2602 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2603 	 * if td is currently executing on some other CPU and holds spin locks
2604 	 * as we won't display those locks.  If we had a MI way of getting
2605 	 * the per-cpu data for a given cpu then we could use
2606 	 * td->td_oncpu to get the list of spinlocks for this thread
2607 	 * and "fix" this.
2608 	 *
2609 	 * That still wouldn't really fix this unless we locked the scheduler
2610 	 * lock or stopped the other CPU to make sure it wasn't changing the
2611 	 * list out from under us.  It is probably best to just not try to
2612 	 * handle threads on other CPU's for now.
2613 	 */
2614 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2615 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2616 }
2617 
2618 DB_SHOW_COMMAND(locks, db_witness_list)
2619 {
2620 	struct thread *td;
2621 
2622 	if (have_addr)
2623 		td = db_lookup_thread(addr, true);
2624 	else
2625 		td = kdb_thread;
2626 	witness_ddb_list(td);
2627 }
2628 
2629 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2630 {
2631 	struct thread *td;
2632 	struct proc *p;
2633 
2634 	/*
2635 	 * It would be nice to list only threads and processes that actually
2636 	 * held sleep locks, but that information is currently not exported
2637 	 * by WITNESS.
2638 	 */
2639 	FOREACH_PROC_IN_SYSTEM(p) {
2640 		if (!witness_proc_has_locks(p))
2641 			continue;
2642 		FOREACH_THREAD_IN_PROC(p, td) {
2643 			if (!witness_thread_has_locks(td))
2644 				continue;
2645 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2646 			    p->p_comm, td, td->td_tid);
2647 			witness_ddb_list(td);
2648 			if (db_pager_quit)
2649 				return;
2650 		}
2651 	}
2652 }
2653 DB_SHOW_ALIAS_FLAGS(alllocks, db_witness_list_all, DB_CMD_MEMSAFE);
2654 
2655 DB_SHOW_COMMAND_FLAGS(witness, db_witness_display, DB_CMD_MEMSAFE)
2656 {
2657 
2658 	witness_ddb_display(db_printf);
2659 }
2660 #endif
2661 
2662 static void
2663 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2664 {
2665 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2666 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2667 	int generation, i, j;
2668 
2669 	tmp_data1 = NULL;
2670 	tmp_data2 = NULL;
2671 	tmp_w1 = NULL;
2672 	tmp_w2 = NULL;
2673 
2674 	/* Allocate and init temporary storage space. */
2675 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2676 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2677 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2678 	    M_WAITOK | M_ZERO);
2679 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2680 	    M_WAITOK | M_ZERO);
2681 	stack_zero(&tmp_data1->wlod_stack);
2682 	stack_zero(&tmp_data2->wlod_stack);
2683 
2684 restart:
2685 	mtx_lock_spin(&w_mtx);
2686 	generation = w_generation;
2687 	mtx_unlock_spin(&w_mtx);
2688 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2689 	    w_lohash.wloh_count);
2690 	for (i = 1; i < w_max_used_index; i++) {
2691 		mtx_lock_spin(&w_mtx);
2692 		if (generation != w_generation) {
2693 			mtx_unlock_spin(&w_mtx);
2694 
2695 			/* The graph has changed, try again. */
2696 			*oldidx = 0;
2697 			sbuf_clear(sb);
2698 			goto restart;
2699 		}
2700 
2701 		w1 = &w_data[i];
2702 		if (w1->w_reversed == 0) {
2703 			mtx_unlock_spin(&w_mtx);
2704 			continue;
2705 		}
2706 
2707 		/* Copy w1 locally so we can release the spin lock. */
2708 		*tmp_w1 = *w1;
2709 		mtx_unlock_spin(&w_mtx);
2710 
2711 		if (tmp_w1->w_reversed == 0)
2712 			continue;
2713 		for (j = 1; j < w_max_used_index; j++) {
2714 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2715 				continue;
2716 
2717 			mtx_lock_spin(&w_mtx);
2718 			if (generation != w_generation) {
2719 				mtx_unlock_spin(&w_mtx);
2720 
2721 				/* The graph has changed, try again. */
2722 				*oldidx = 0;
2723 				sbuf_clear(sb);
2724 				goto restart;
2725 			}
2726 
2727 			w2 = &w_data[j];
2728 			data1 = witness_lock_order_get(w1, w2);
2729 			data2 = witness_lock_order_get(w2, w1);
2730 
2731 			/*
2732 			 * Copy information locally so we can release the
2733 			 * spin lock.
2734 			 */
2735 			*tmp_w2 = *w2;
2736 
2737 			if (data1) {
2738 				stack_zero(&tmp_data1->wlod_stack);
2739 				stack_copy(&data1->wlod_stack,
2740 				    &tmp_data1->wlod_stack);
2741 			}
2742 			if (data2 && data2 != data1) {
2743 				stack_zero(&tmp_data2->wlod_stack);
2744 				stack_copy(&data2->wlod_stack,
2745 				    &tmp_data2->wlod_stack);
2746 			}
2747 			mtx_unlock_spin(&w_mtx);
2748 
2749 			if (blessed(tmp_w1, tmp_w2))
2750 				continue;
2751 
2752 			sbuf_printf(sb,
2753 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2754 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2755 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2756 			if (data1) {
2757 				sbuf_printf(sb,
2758 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2759 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2760 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2761 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2762 				sbuf_putc(sb, '\n');
2763 			}
2764 			if (data2 && data2 != data1) {
2765 				sbuf_printf(sb,
2766 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2767 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2768 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2769 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2770 				sbuf_putc(sb, '\n');
2771 			}
2772 		}
2773 	}
2774 	mtx_lock_spin(&w_mtx);
2775 	if (generation != w_generation) {
2776 		mtx_unlock_spin(&w_mtx);
2777 
2778 		/*
2779 		 * The graph changed while we were printing stack data,
2780 		 * try again.
2781 		 */
2782 		*oldidx = 0;
2783 		sbuf_clear(sb);
2784 		goto restart;
2785 	}
2786 	mtx_unlock_spin(&w_mtx);
2787 
2788 	/* Free temporary storage space. */
2789 	free(tmp_data1, M_TEMP);
2790 	free(tmp_data2, M_TEMP);
2791 	free(tmp_w1, M_TEMP);
2792 	free(tmp_w2, M_TEMP);
2793 }
2794 
2795 static int
2796 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2797 {
2798 	struct sbuf *sb;
2799 	int error;
2800 
2801 	if (witness_watch < 1) {
2802 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2803 		return (error);
2804 	}
2805 	if (witness_cold) {
2806 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2807 		return (error);
2808 	}
2809 	error = 0;
2810 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2811 	if (sb == NULL)
2812 		return (ENOMEM);
2813 
2814 	sbuf_print_witness_badstacks(sb, &req->oldidx);
2815 
2816 	sbuf_finish(sb);
2817 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2818 	sbuf_delete(sb);
2819 
2820 	return (error);
2821 }
2822 
2823 #ifdef DDB
2824 static int
2825 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2826 {
2827 
2828 	return (db_printf("%.*s", len, data));
2829 }
2830 
2831 DB_SHOW_COMMAND_FLAGS(badstacks, db_witness_badstacks, DB_CMD_MEMSAFE)
2832 {
2833 	struct sbuf sb;
2834 	char buffer[128];
2835 	size_t dummy;
2836 
2837 	sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2838 	sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2839 	sbuf_print_witness_badstacks(&sb, &dummy);
2840 	sbuf_finish(&sb);
2841 }
2842 #endif
2843 
2844 static int
2845 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2846 {
2847 	static const struct {
2848 		enum witness_channel channel;
2849 		const char *name;
2850 	} channels[] = {
2851 		{ WITNESS_CONSOLE, "console" },
2852 		{ WITNESS_LOG, "log" },
2853 		{ WITNESS_NONE, "none" },
2854 	};
2855 	char buf[16];
2856 	u_int i;
2857 	int error;
2858 
2859 	buf[0] = '\0';
2860 	for (i = 0; i < nitems(channels); i++)
2861 		if (witness_channel == channels[i].channel) {
2862 			snprintf(buf, sizeof(buf), "%s", channels[i].name);
2863 			break;
2864 		}
2865 
2866 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2867 	if (error != 0 || req->newptr == NULL)
2868 		return (error);
2869 
2870 	error = EINVAL;
2871 	for (i = 0; i < nitems(channels); i++)
2872 		if (strcmp(channels[i].name, buf) == 0) {
2873 			witness_channel = channels[i].channel;
2874 			error = 0;
2875 			break;
2876 		}
2877 	return (error);
2878 }
2879 
2880 static int
2881 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2882 {
2883 	struct witness *w;
2884 	struct sbuf *sb;
2885 	int error;
2886 
2887 #ifdef __i386__
2888 	error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2889 	return (error);
2890 #endif
2891 
2892 	if (witness_watch < 1) {
2893 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2894 		return (error);
2895 	}
2896 	if (witness_cold) {
2897 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2898 		return (error);
2899 	}
2900 	error = 0;
2901 
2902 	error = sysctl_wire_old_buffer(req, 0);
2903 	if (error != 0)
2904 		return (error);
2905 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2906 	if (sb == NULL)
2907 		return (ENOMEM);
2908 	sbuf_putc(sb, '\n');
2909 
2910 	mtx_lock_spin(&w_mtx);
2911 	STAILQ_FOREACH(w, &w_all, w_list)
2912 		w->w_displayed = 0;
2913 	STAILQ_FOREACH(w, &w_all, w_list)
2914 		witness_add_fullgraph(sb, w);
2915 	mtx_unlock_spin(&w_mtx);
2916 
2917 	/*
2918 	 * Close the sbuf and return to userland.
2919 	 */
2920 	error = sbuf_finish(sb);
2921 	sbuf_delete(sb);
2922 
2923 	return (error);
2924 }
2925 
2926 static int
2927 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2928 {
2929 	int error, value;
2930 
2931 	value = witness_watch;
2932 	error = sysctl_handle_int(oidp, &value, 0, req);
2933 	if (error != 0 || req->newptr == NULL)
2934 		return (error);
2935 	if (value > 1 || value < -1 ||
2936 	    (witness_watch == -1 && value != witness_watch))
2937 		return (EINVAL);
2938 	witness_watch = value;
2939 	return (0);
2940 }
2941 
2942 static void
2943 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2944 {
2945 	int i;
2946 
2947 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2948 		return;
2949 	w->w_displayed = 1;
2950 
2951 	WITNESS_INDEX_ASSERT(w->w_index);
2952 	for (i = 1; i <= w_max_used_index; i++) {
2953 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2954 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2955 			    w_data[i].w_name);
2956 			witness_add_fullgraph(sb, &w_data[i]);
2957 		}
2958 	}
2959 }
2960 
2961 /*
2962  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2963  * interprets the key as a string and reads until the null
2964  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2965  * hash value computed from the key.
2966  */
2967 static uint32_t
2968 witness_hash_djb2(const uint8_t *key, uint32_t size)
2969 {
2970 	unsigned int hash = 5381;
2971 	int i;
2972 
2973 	/* hash = hash * 33 + key[i] */
2974 	if (size)
2975 		for (i = 0; i < size; i++)
2976 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2977 	else
2978 		for (i = 0; key[i] != 0; i++)
2979 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2980 
2981 	return (hash);
2982 }
2983 
2984 /*
2985  * Initializes the two witness hash tables. Called exactly once from
2986  * witness_initialize().
2987  */
2988 static void
2989 witness_init_hash_tables(void)
2990 {
2991 	int i;
2992 
2993 	MPASS(witness_cold);
2994 
2995 	/* Initialize the hash tables. */
2996 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2997 		w_hash.wh_array[i] = NULL;
2998 
2999 	w_hash.wh_size = WITNESS_HASH_SIZE;
3000 	w_hash.wh_count = 0;
3001 
3002 	/* Initialize the lock order data hash. */
3003 	w_lofree = NULL;
3004 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
3005 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
3006 		w_lodata[i].wlod_next = w_lofree;
3007 		w_lofree = &w_lodata[i];
3008 	}
3009 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
3010 	w_lohash.wloh_count = 0;
3011 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
3012 		w_lohash.wloh_array[i] = NULL;
3013 }
3014 
3015 static struct witness *
3016 witness_hash_get(const char *key)
3017 {
3018 	struct witness *w;
3019 	uint32_t hash;
3020 
3021 	MPASS(key != NULL);
3022 	if (witness_cold == 0)
3023 		mtx_assert(&w_mtx, MA_OWNED);
3024 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
3025 	w = w_hash.wh_array[hash];
3026 	while (w != NULL) {
3027 		if (strcmp(w->w_name, key) == 0)
3028 			goto out;
3029 		w = w->w_hash_next;
3030 	}
3031 
3032 out:
3033 	return (w);
3034 }
3035 
3036 static void
3037 witness_hash_put(struct witness *w)
3038 {
3039 	uint32_t hash;
3040 
3041 	MPASS(w != NULL);
3042 	MPASS(w->w_name != NULL);
3043 	if (witness_cold == 0)
3044 		mtx_assert(&w_mtx, MA_OWNED);
3045 	KASSERT(witness_hash_get(w->w_name) == NULL,
3046 	    ("%s: trying to add a hash entry that already exists!", __func__));
3047 	KASSERT(w->w_hash_next == NULL,
3048 	    ("%s: w->w_hash_next != NULL", __func__));
3049 
3050 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3051 	w->w_hash_next = w_hash.wh_array[hash];
3052 	w_hash.wh_array[hash] = w;
3053 	w_hash.wh_count++;
3054 }
3055 
3056 static struct witness_lock_order_data *
3057 witness_lock_order_get(struct witness *parent, struct witness *child)
3058 {
3059 	struct witness_lock_order_data *data = NULL;
3060 	struct witness_lock_order_key key;
3061 	unsigned int hash;
3062 
3063 	MPASS(parent != NULL && child != NULL);
3064 	key.from = parent->w_index;
3065 	key.to = child->w_index;
3066 	WITNESS_INDEX_ASSERT(key.from);
3067 	WITNESS_INDEX_ASSERT(key.to);
3068 	if ((w_rmatrix[parent->w_index][child->w_index]
3069 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
3070 		goto out;
3071 
3072 	hash = witness_hash_djb2((const char*)&key,
3073 	    sizeof(key)) % w_lohash.wloh_size;
3074 	data = w_lohash.wloh_array[hash];
3075 	while (data != NULL) {
3076 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
3077 			break;
3078 		data = data->wlod_next;
3079 	}
3080 
3081 out:
3082 	return (data);
3083 }
3084 
3085 /*
3086  * Verify that parent and child have a known relationship, are not the same,
3087  * and child is actually a child of parent.  This is done without w_mtx
3088  * to avoid contention in the common case.
3089  */
3090 static int
3091 witness_lock_order_check(struct witness *parent, struct witness *child)
3092 {
3093 
3094 	if (parent != child &&
3095 	    w_rmatrix[parent->w_index][child->w_index]
3096 	    & WITNESS_LOCK_ORDER_KNOWN &&
3097 	    isitmychild(parent, child))
3098 		return (1);
3099 
3100 	return (0);
3101 }
3102 
3103 static int
3104 witness_lock_order_add(struct witness *parent, struct witness *child)
3105 {
3106 	struct witness_lock_order_data *data = NULL;
3107 	struct witness_lock_order_key key;
3108 	unsigned int hash;
3109 
3110 	MPASS(parent != NULL && child != NULL);
3111 	key.from = parent->w_index;
3112 	key.to = child->w_index;
3113 	WITNESS_INDEX_ASSERT(key.from);
3114 	WITNESS_INDEX_ASSERT(key.to);
3115 	if (w_rmatrix[parent->w_index][child->w_index]
3116 	    & WITNESS_LOCK_ORDER_KNOWN)
3117 		return (1);
3118 
3119 	hash = witness_hash_djb2((const char*)&key,
3120 	    sizeof(key)) % w_lohash.wloh_size;
3121 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3122 	data = w_lofree;
3123 	if (data == NULL)
3124 		return (0);
3125 	w_lofree = data->wlod_next;
3126 	data->wlod_next = w_lohash.wloh_array[hash];
3127 	data->wlod_key = key;
3128 	w_lohash.wloh_array[hash] = data;
3129 	w_lohash.wloh_count++;
3130 	stack_save(&data->wlod_stack);
3131 	return (1);
3132 }
3133 
3134 /* Call this whenever the structure of the witness graph changes. */
3135 static void
3136 witness_increment_graph_generation(void)
3137 {
3138 
3139 	if (witness_cold == 0)
3140 		mtx_assert(&w_mtx, MA_OWNED);
3141 	w_generation++;
3142 }
3143 
3144 static int
3145 witness_output_drain(void *arg __unused, const char *data, int len)
3146 {
3147 
3148 	witness_output("%.*s", len, data);
3149 	return (len);
3150 }
3151 
3152 static void
3153 witness_debugger(int cond, const char *msg)
3154 {
3155 	char buf[32];
3156 	struct sbuf sb;
3157 	struct stack st;
3158 
3159 	if (!cond)
3160 		return;
3161 
3162 	if (witness_trace) {
3163 		sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3164 		sbuf_set_drain(&sb, witness_output_drain, NULL);
3165 
3166 		stack_save(&st);
3167 		witness_output("stack backtrace:\n");
3168 		stack_sbuf_print_ddb(&sb, &st);
3169 
3170 		sbuf_finish(&sb);
3171 	}
3172 
3173 	witness_enter_debugger(msg);
3174 }
3175 
3176 static void
3177 witness_enter_debugger(const char *msg)
3178 {
3179 #ifdef KDB
3180 	if (witness_kdb)
3181 		kdb_enter(KDB_WHY_WITNESS, msg);
3182 #endif
3183 }
3184