xref: /freebsd/sys/kern/subr_witness.c (revision 3b8f08459569bf0faa21473e5cec2491e95c9349)
1 /*-
2  * Copyright (c) 2008 Isilon Systems, Inc.
3  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
4  * Copyright (c) 1998 Berkeley Software Design, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Berkeley Software Design Inc's name may not be used to endorse or
16  *    promote products derived from this software without specific prior
17  *    written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
32  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
33  */
34 
35 /*
36  * Implementation of the `witness' lock verifier.  Originally implemented for
37  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
38  * classes in FreeBSD.
39  */
40 
41 /*
42  *	Main Entry: witness
43  *	Pronunciation: 'wit-n&s
44  *	Function: noun
45  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
46  *	    testimony, witness, from 2wit
47  *	Date: before 12th century
48  *	1 : attestation of a fact or event : TESTIMONY
49  *	2 : one that gives evidence; specifically : one who testifies in
50  *	    a cause or before a judicial tribunal
51  *	3 : one asked to be present at a transaction so as to be able to
52  *	    testify to its having taken place
53  *	4 : one who has personal knowledge of something
54  *	5 a : something serving as evidence or proof : SIGN
55  *	  b : public affirmation by word or example of usually
56  *	      religious faith or conviction <the heroic witness to divine
57  *	      life -- Pilot>
58  *	6 capitalized : a member of the Jehovah's Witnesses
59  */
60 
61 /*
62  * Special rules concerning Giant and lock orders:
63  *
64  * 1) Giant must be acquired before any other mutexes.  Stated another way,
65  *    no other mutex may be held when Giant is acquired.
66  *
67  * 2) Giant must be released when blocking on a sleepable lock.
68  *
69  * This rule is less obvious, but is a result of Giant providing the same
70  * semantics as spl().  Basically, when a thread sleeps, it must release
71  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
72  * 2).
73  *
74  * 3) Giant may be acquired before or after sleepable locks.
75  *
76  * This rule is also not quite as obvious.  Giant may be acquired after
77  * a sleepable lock because it is a non-sleepable lock and non-sleepable
78  * locks may always be acquired while holding a sleepable lock.  The second
79  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
80  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
81  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
82  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
83  * execute.  Thus, acquiring Giant both before and after a sleepable lock
84  * will not result in a lock order reversal.
85  */
86 
87 #include <sys/cdefs.h>
88 __FBSDID("$FreeBSD$");
89 
90 #include "opt_ddb.h"
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
94 
95 #include <sys/param.h>
96 #include <sys/bus.h>
97 #include <sys/kdb.h>
98 #include <sys/kernel.h>
99 #include <sys/ktr.h>
100 #include <sys/lock.h>
101 #include <sys/malloc.h>
102 #include <sys/mutex.h>
103 #include <sys/priv.h>
104 #include <sys/proc.h>
105 #include <sys/sbuf.h>
106 #include <sys/sched.h>
107 #include <sys/stack.h>
108 #include <sys/sysctl.h>
109 #include <sys/systm.h>
110 
111 #ifdef DDB
112 #include <ddb/ddb.h>
113 #endif
114 
115 #include <machine/stdarg.h>
116 
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
119 #endif
120 
121 /* Note that these traces do not work with KTR_ALQ. */
122 #if 0
123 #define	KTR_WITNESS	KTR_SUBSYS
124 #else
125 #define	KTR_WITNESS	0
126 #endif
127 
128 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
129 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
130 #define	LI_NORELEASE	0x00020000	/* Lock not allowed to be released. */
131 
132 /* Define this to check for blessed mutexes */
133 #undef BLESSING
134 
135 #define	WITNESS_COUNT 		1536
136 #define	WITNESS_CHILDCOUNT 	(WITNESS_COUNT * 4)
137 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
138 #define	WITNESS_PENDLIST	1024
139 
140 /* Allocate 256 KB of stack data space */
141 #define	WITNESS_LO_DATA_COUNT	2048
142 
143 /* Prime, gives load factor of ~2 at full load */
144 #define	WITNESS_LO_HASH_SIZE	1021
145 
146 /*
147  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
148  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
149  * probably be safe for the most part, but it's still a SWAG.
150  */
151 #define	LOCK_NCHILDREN	5
152 #define	LOCK_CHILDCOUNT	2048
153 
154 #define	MAX_W_NAME	64
155 
156 #define	BADSTACK_SBUF_SIZE	(256 * WITNESS_COUNT)
157 #define	FULLGRAPH_SBUF_SIZE	512
158 
159 /*
160  * These flags go in the witness relationship matrix and describe the
161  * relationship between any two struct witness objects.
162  */
163 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
164 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
165 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
166 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
167 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
168 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define	WITNESS_RELATED_MASK						\
171 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
173 					  * observed. */
174 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
175 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
176 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
177 
178 /* Descendant to ancestor flags */
179 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
180 
181 /* Ancestor to descendant flags */
182 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
183 
184 #define	WITNESS_INDEX_ASSERT(i)						\
185 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
186 
187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
188 
189 /*
190  * Lock instances.  A lock instance is the data associated with a lock while
191  * it is held by witness.  For example, a lock instance will hold the
192  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
193  * are held in a per-cpu list while sleep locks are held in per-thread list.
194  */
195 struct lock_instance {
196 	struct lock_object	*li_lock;
197 	const char		*li_file;
198 	int			li_line;
199 	u_int			li_flags;
200 };
201 
202 /*
203  * A simple list type used to build the list of locks held by a thread
204  * or CPU.  We can't simply embed the list in struct lock_object since a
205  * lock may be held by more than one thread if it is a shared lock.  Locks
206  * are added to the head of the list, so we fill up each list entry from
207  * "the back" logically.  To ease some of the arithmetic, we actually fill
208  * in each list entry the normal way (children[0] then children[1], etc.) but
209  * when we traverse the list we read children[count-1] as the first entry
210  * down to children[0] as the final entry.
211  */
212 struct lock_list_entry {
213 	struct lock_list_entry	*ll_next;
214 	struct lock_instance	ll_children[LOCK_NCHILDREN];
215 	u_int			ll_count;
216 };
217 
218 /*
219  * The main witness structure. One of these per named lock type in the system
220  * (for example, "vnode interlock").
221  */
222 struct witness {
223 	char  			w_name[MAX_W_NAME];
224 	uint32_t 		w_index;  /* Index in the relationship matrix */
225 	struct lock_class	*w_class;
226 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
227 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
228 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
229 	const char		*w_file; /* File where last acquired */
230 	uint32_t 		w_line; /* Line where last acquired */
231 	uint32_t 		w_refcount;
232 	uint16_t 		w_num_ancestors; /* direct/indirect
233 						  * ancestor count */
234 	uint16_t 		w_num_descendants; /* direct/indirect
235 						    * descendant count */
236 	int16_t 		w_ddb_level;
237 	unsigned		w_displayed:1;
238 	unsigned		w_reversed:1;
239 };
240 
241 STAILQ_HEAD(witness_list, witness);
242 
243 /*
244  * The witness hash table. Keys are witness names (const char *), elements are
245  * witness objects (struct witness *).
246  */
247 struct witness_hash {
248 	struct witness	*wh_array[WITNESS_HASH_SIZE];
249 	uint32_t	wh_size;
250 	uint32_t	wh_count;
251 };
252 
253 /*
254  * Key type for the lock order data hash table.
255  */
256 struct witness_lock_order_key {
257 	uint16_t	from;
258 	uint16_t	to;
259 };
260 
261 struct witness_lock_order_data {
262 	struct stack			wlod_stack;
263 	struct witness_lock_order_key	wlod_key;
264 	struct witness_lock_order_data	*wlod_next;
265 };
266 
267 /*
268  * The witness lock order data hash table. Keys are witness index tuples
269  * (struct witness_lock_order_key), elements are lock order data objects
270  * (struct witness_lock_order_data).
271  */
272 struct witness_lock_order_hash {
273 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
274 	u_int	wloh_size;
275 	u_int	wloh_count;
276 };
277 
278 #ifdef BLESSING
279 struct witness_blessed {
280 	const char	*b_lock1;
281 	const char	*b_lock2;
282 };
283 #endif
284 
285 struct witness_pendhelp {
286 	const char		*wh_type;
287 	struct lock_object	*wh_lock;
288 };
289 
290 struct witness_order_list_entry {
291 	const char		*w_name;
292 	struct lock_class	*w_class;
293 };
294 
295 /*
296  * Returns 0 if one of the locks is a spin lock and the other is not.
297  * Returns 1 otherwise.
298  */
299 static __inline int
300 witness_lock_type_equal(struct witness *w1, struct witness *w2)
301 {
302 
303 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
304 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
305 }
306 
307 static __inline int
308 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
309     const struct witness_lock_order_key *b)
310 {
311 
312 	return (a->from == b->from && a->to == b->to);
313 }
314 
315 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
316 		    const char *fname);
317 #ifdef KDB
318 static void	_witness_debugger(int cond, const char *msg);
319 #endif
320 static void	adopt(struct witness *parent, struct witness *child);
321 #ifdef BLESSING
322 static int	blessed(struct witness *, struct witness *);
323 #endif
324 static void	depart(struct witness *w);
325 static struct witness	*enroll(const char *description,
326 			    struct lock_class *lock_class);
327 static struct lock_instance	*find_instance(struct lock_list_entry *list,
328 				    const struct lock_object *lock);
329 static int	isitmychild(struct witness *parent, struct witness *child);
330 static int	isitmydescendant(struct witness *parent, struct witness *child);
331 static void	itismychild(struct witness *parent, struct witness *child);
332 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
333 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
334 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
335 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
336 #ifdef DDB
337 static void	witness_ddb_compute_levels(void);
338 static void	witness_ddb_display(int(*)(const char *fmt, ...));
339 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
340 		    struct witness *, int indent);
341 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
342 		    struct witness_list *list);
343 static void	witness_ddb_level_descendants(struct witness *parent, int l);
344 static void	witness_ddb_list(struct thread *td);
345 #endif
346 static void	witness_free(struct witness *m);
347 static struct witness	*witness_get(void);
348 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
349 static struct witness	*witness_hash_get(const char *key);
350 static void	witness_hash_put(struct witness *w);
351 static void	witness_init_hash_tables(void);
352 static void	witness_increment_graph_generation(void);
353 static void	witness_lock_list_free(struct lock_list_entry *lle);
354 static struct lock_list_entry	*witness_lock_list_get(void);
355 static int	witness_lock_order_add(struct witness *parent,
356 		    struct witness *child);
357 static int	witness_lock_order_check(struct witness *parent,
358 		    struct witness *child);
359 static struct witness_lock_order_data	*witness_lock_order_get(
360 					    struct witness *parent,
361 					    struct witness *child);
362 static void	witness_list_lock(struct lock_instance *instance,
363 		    int (*prnt)(const char *fmt, ...));
364 static void	witness_setflag(struct lock_object *lock, int flag, int set);
365 
366 #ifdef KDB
367 #define	witness_debugger(c)	_witness_debugger(c, __func__)
368 #else
369 #define	witness_debugger(c)
370 #endif
371 
372 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
373     "Witness Locking");
374 
375 /*
376  * If set to 0, lock order checking is disabled.  If set to -1,
377  * witness is completely disabled.  Otherwise witness performs full
378  * lock order checking for all locks.  At runtime, lock order checking
379  * may be toggled.  However, witness cannot be reenabled once it is
380  * completely disabled.
381  */
382 static int witness_watch = 1;
383 TUNABLE_INT("debug.witness.watch", &witness_watch);
384 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
385     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
386 
387 #ifdef KDB
388 /*
389  * When KDB is enabled and witness_kdb is 1, it will cause the system
390  * to drop into kdebug() when:
391  *	- a lock hierarchy violation occurs
392  *	- locks are held when going to sleep.
393  */
394 #ifdef WITNESS_KDB
395 int	witness_kdb = 1;
396 #else
397 int	witness_kdb = 0;
398 #endif
399 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
400 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
401 
402 /*
403  * When KDB is enabled and witness_trace is 1, it will cause the system
404  * to print a stack trace:
405  *	- a lock hierarchy violation occurs
406  *	- locks are held when going to sleep.
407  */
408 int	witness_trace = 1;
409 TUNABLE_INT("debug.witness.trace", &witness_trace);
410 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
411 #endif /* KDB */
412 
413 #ifdef WITNESS_SKIPSPIN
414 int	witness_skipspin = 1;
415 #else
416 int	witness_skipspin = 0;
417 #endif
418 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
419 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
420     0, "");
421 
422 /*
423  * Call this to print out the relations between locks.
424  */
425 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
426     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
427 
428 /*
429  * Call this to print out the witness faulty stacks.
430  */
431 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
432     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
433 
434 static struct mtx w_mtx;
435 
436 /* w_list */
437 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
438 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
439 
440 /* w_typelist */
441 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
442 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
443 
444 /* lock list */
445 static struct lock_list_entry *w_lock_list_free = NULL;
446 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
447 static u_int pending_cnt;
448 
449 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
450 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
451 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
452 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
453     "");
454 
455 static struct witness *w_data;
456 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
457 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
458 static struct witness_hash w_hash;	/* The witness hash table. */
459 
460 /* The lock order data hash */
461 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
462 static struct witness_lock_order_data *w_lofree = NULL;
463 static struct witness_lock_order_hash w_lohash;
464 static int w_max_used_index = 0;
465 static unsigned int w_generation = 0;
466 static const char w_notrunning[] = "Witness not running\n";
467 static const char w_stillcold[] = "Witness is still cold\n";
468 
469 
470 static struct witness_order_list_entry order_lists[] = {
471 	/*
472 	 * sx locks
473 	 */
474 	{ "proctree", &lock_class_sx },
475 	{ "allproc", &lock_class_sx },
476 	{ "allprison", &lock_class_sx },
477 	{ NULL, NULL },
478 	/*
479 	 * Various mutexes
480 	 */
481 	{ "Giant", &lock_class_mtx_sleep },
482 	{ "pipe mutex", &lock_class_mtx_sleep },
483 	{ "sigio lock", &lock_class_mtx_sleep },
484 	{ "process group", &lock_class_mtx_sleep },
485 	{ "process lock", &lock_class_mtx_sleep },
486 	{ "session", &lock_class_mtx_sleep },
487 	{ "uidinfo hash", &lock_class_rw },
488 #ifdef	HWPMC_HOOKS
489 	{ "pmc-sleep", &lock_class_mtx_sleep },
490 #endif
491 	{ "time lock", &lock_class_mtx_sleep },
492 	{ NULL, NULL },
493 	/*
494 	 * Sockets
495 	 */
496 	{ "accept", &lock_class_mtx_sleep },
497 	{ "so_snd", &lock_class_mtx_sleep },
498 	{ "so_rcv", &lock_class_mtx_sleep },
499 	{ "sellck", &lock_class_mtx_sleep },
500 	{ NULL, NULL },
501 	/*
502 	 * Routing
503 	 */
504 	{ "so_rcv", &lock_class_mtx_sleep },
505 	{ "radix node head", &lock_class_rw },
506 	{ "rtentry", &lock_class_mtx_sleep },
507 	{ "ifaddr", &lock_class_mtx_sleep },
508 	{ NULL, NULL },
509 	/*
510 	 * IPv4 multicast:
511 	 * protocol locks before interface locks, after UDP locks.
512 	 */
513 	{ "udpinp", &lock_class_rw },
514 	{ "in_multi_mtx", &lock_class_mtx_sleep },
515 	{ "igmp_mtx", &lock_class_mtx_sleep },
516 	{ "if_addr_lock", &lock_class_rw },
517 	{ NULL, NULL },
518 	/*
519 	 * IPv6 multicast:
520 	 * protocol locks before interface locks, after UDP locks.
521 	 */
522 	{ "udpinp", &lock_class_rw },
523 	{ "in6_multi_mtx", &lock_class_mtx_sleep },
524 	{ "mld_mtx", &lock_class_mtx_sleep },
525 	{ "if_addr_lock", &lock_class_rw },
526 	{ NULL, NULL },
527 	/*
528 	 * UNIX Domain Sockets
529 	 */
530 	{ "unp_global_rwlock", &lock_class_rw },
531 	{ "unp_list_lock", &lock_class_mtx_sleep },
532 	{ "unp", &lock_class_mtx_sleep },
533 	{ "so_snd", &lock_class_mtx_sleep },
534 	{ NULL, NULL },
535 	/*
536 	 * UDP/IP
537 	 */
538 	{ "udp", &lock_class_rw },
539 	{ "udpinp", &lock_class_rw },
540 	{ "so_snd", &lock_class_mtx_sleep },
541 	{ NULL, NULL },
542 	/*
543 	 * TCP/IP
544 	 */
545 	{ "tcp", &lock_class_rw },
546 	{ "tcpinp", &lock_class_rw },
547 	{ "so_snd", &lock_class_mtx_sleep },
548 	{ NULL, NULL },
549 	/*
550 	 * BPF
551 	 */
552 	{ "bpf global lock", &lock_class_mtx_sleep },
553 	{ "bpf interface lock", &lock_class_rw },
554 	{ "bpf cdev lock", &lock_class_mtx_sleep },
555 	{ NULL, NULL },
556 	/*
557 	 * NFS server
558 	 */
559 	{ "nfsd_mtx", &lock_class_mtx_sleep },
560 	{ "so_snd", &lock_class_mtx_sleep },
561 	{ NULL, NULL },
562 
563 	/*
564 	 * IEEE 802.11
565 	 */
566 	{ "802.11 com lock", &lock_class_mtx_sleep},
567 	{ NULL, NULL },
568 	/*
569 	 * Network drivers
570 	 */
571 	{ "network driver", &lock_class_mtx_sleep},
572 	{ NULL, NULL },
573 
574 	/*
575 	 * Netgraph
576 	 */
577 	{ "ng_node", &lock_class_mtx_sleep },
578 	{ "ng_worklist", &lock_class_mtx_sleep },
579 	{ NULL, NULL },
580 	/*
581 	 * CDEV
582 	 */
583 	{ "vm map (system)", &lock_class_mtx_sleep },
584 	{ "vm page queue", &lock_class_mtx_sleep },
585 	{ "vnode interlock", &lock_class_mtx_sleep },
586 	{ "cdev", &lock_class_mtx_sleep },
587 	{ NULL, NULL },
588 	/*
589 	 * VM
590 	 */
591 	{ "vm map (user)", &lock_class_sx },
592 	{ "vm object", &lock_class_rw },
593 	{ "vm page", &lock_class_mtx_sleep },
594 	{ "vm page queue", &lock_class_mtx_sleep },
595 	{ "pmap pv global", &lock_class_rw },
596 	{ "pmap", &lock_class_mtx_sleep },
597 	{ "pmap pv list", &lock_class_rw },
598 	{ "vm page free queue", &lock_class_mtx_sleep },
599 	{ NULL, NULL },
600 	/*
601 	 * kqueue/VFS interaction
602 	 */
603 	{ "kqueue", &lock_class_mtx_sleep },
604 	{ "struct mount mtx", &lock_class_mtx_sleep },
605 	{ "vnode interlock", &lock_class_mtx_sleep },
606 	{ NULL, NULL },
607 	/*
608 	 * ZFS locking
609 	 */
610 	{ "dn->dn_mtx", &lock_class_sx },
611 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
612 	{ "db->db_mtx", &lock_class_sx },
613 	{ NULL, NULL },
614 	/*
615 	 * spin locks
616 	 */
617 #ifdef SMP
618 	{ "ap boot", &lock_class_mtx_spin },
619 #endif
620 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
621 	{ "sio", &lock_class_mtx_spin },
622 	{ "scrlock", &lock_class_mtx_spin },
623 #ifdef __i386__
624 	{ "cy", &lock_class_mtx_spin },
625 #endif
626 #ifdef __sparc64__
627 	{ "pcib_mtx", &lock_class_mtx_spin },
628 	{ "rtc_mtx", &lock_class_mtx_spin },
629 #endif
630 	{ "scc_hwmtx", &lock_class_mtx_spin },
631 	{ "uart_hwmtx", &lock_class_mtx_spin },
632 	{ "fast_taskqueue", &lock_class_mtx_spin },
633 	{ "intr table", &lock_class_mtx_spin },
634 #ifdef	HWPMC_HOOKS
635 	{ "pmc-per-proc", &lock_class_mtx_spin },
636 #endif
637 	{ "process slock", &lock_class_mtx_spin },
638 	{ "sleepq chain", &lock_class_mtx_spin },
639 	{ "umtx lock", &lock_class_mtx_spin },
640 	{ "rm_spinlock", &lock_class_mtx_spin },
641 	{ "turnstile chain", &lock_class_mtx_spin },
642 	{ "turnstile lock", &lock_class_mtx_spin },
643 	{ "sched lock", &lock_class_mtx_spin },
644 	{ "td_contested", &lock_class_mtx_spin },
645 	{ "callout", &lock_class_mtx_spin },
646 	{ "entropy harvest mutex", &lock_class_mtx_spin },
647 	{ "syscons video lock", &lock_class_mtx_spin },
648 #ifdef SMP
649 	{ "smp rendezvous", &lock_class_mtx_spin },
650 #endif
651 #ifdef __powerpc__
652 	{ "tlb0", &lock_class_mtx_spin },
653 #endif
654 	/*
655 	 * leaf locks
656 	 */
657 	{ "intrcnt", &lock_class_mtx_spin },
658 	{ "icu", &lock_class_mtx_spin },
659 #ifdef __i386__
660 	{ "allpmaps", &lock_class_mtx_spin },
661 	{ "descriptor tables", &lock_class_mtx_spin },
662 #endif
663 	{ "clk", &lock_class_mtx_spin },
664 	{ "cpuset", &lock_class_mtx_spin },
665 	{ "mprof lock", &lock_class_mtx_spin },
666 	{ "zombie lock", &lock_class_mtx_spin },
667 	{ "ALD Queue", &lock_class_mtx_spin },
668 #ifdef __ia64__
669 	{ "MCA spin lock", &lock_class_mtx_spin },
670 #endif
671 #if defined(__i386__) || defined(__amd64__)
672 	{ "pcicfg", &lock_class_mtx_spin },
673 	{ "NDIS thread lock", &lock_class_mtx_spin },
674 #endif
675 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
676 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
677 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
678 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
679 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
680 #ifdef	HWPMC_HOOKS
681 	{ "pmc-leaf", &lock_class_mtx_spin },
682 #endif
683 	{ "blocked lock", &lock_class_mtx_spin },
684 	{ NULL, NULL },
685 	{ NULL, NULL }
686 };
687 
688 #ifdef BLESSING
689 /*
690  * Pairs of locks which have been blessed
691  * Don't complain about order problems with blessed locks
692  */
693 static struct witness_blessed blessed_list[] = {
694 };
695 static int blessed_count =
696 	sizeof(blessed_list) / sizeof(struct witness_blessed);
697 #endif
698 
699 /*
700  * This global is set to 0 once it becomes safe to use the witness code.
701  */
702 static int witness_cold = 1;
703 
704 /*
705  * This global is set to 1 once the static lock orders have been enrolled
706  * so that a warning can be issued for any spin locks enrolled later.
707  */
708 static int witness_spin_warn = 0;
709 
710 /* Trim useless garbage from filenames. */
711 static const char *
712 fixup_filename(const char *file)
713 {
714 
715 	if (file == NULL)
716 		return (NULL);
717 	while (strncmp(file, "../", 3) == 0)
718 		file += 3;
719 	return (file);
720 }
721 
722 /*
723  * The WITNESS-enabled diagnostic code.  Note that the witness code does
724  * assume that the early boot is single-threaded at least until after this
725  * routine is completed.
726  */
727 static void
728 witness_initialize(void *dummy __unused)
729 {
730 	struct lock_object *lock;
731 	struct witness_order_list_entry *order;
732 	struct witness *w, *w1;
733 	int i;
734 
735 	w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
736 	    M_NOWAIT | M_ZERO);
737 
738 	/*
739 	 * We have to release Giant before initializing its witness
740 	 * structure so that WITNESS doesn't get confused.
741 	 */
742 	mtx_unlock(&Giant);
743 	mtx_assert(&Giant, MA_NOTOWNED);
744 
745 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
746 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
747 	    MTX_NOWITNESS | MTX_NOPROFILE);
748 	for (i = WITNESS_COUNT - 1; i >= 0; i--) {
749 		w = &w_data[i];
750 		memset(w, 0, sizeof(*w));
751 		w_data[i].w_index = i;	/* Witness index never changes. */
752 		witness_free(w);
753 	}
754 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
755 	    ("%s: Invalid list of free witness objects", __func__));
756 
757 	/* Witness with index 0 is not used to aid in debugging. */
758 	STAILQ_REMOVE_HEAD(&w_free, w_list);
759 	w_free_cnt--;
760 
761 	memset(w_rmatrix, 0,
762 	    (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
763 
764 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
765 		witness_lock_list_free(&w_locklistdata[i]);
766 	witness_init_hash_tables();
767 
768 	/* First add in all the specified order lists. */
769 	for (order = order_lists; order->w_name != NULL; order++) {
770 		w = enroll(order->w_name, order->w_class);
771 		if (w == NULL)
772 			continue;
773 		w->w_file = "order list";
774 		for (order++; order->w_name != NULL; order++) {
775 			w1 = enroll(order->w_name, order->w_class);
776 			if (w1 == NULL)
777 				continue;
778 			w1->w_file = "order list";
779 			itismychild(w, w1);
780 			w = w1;
781 		}
782 	}
783 	witness_spin_warn = 1;
784 
785 	/* Iterate through all locks and add them to witness. */
786 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
787 		lock = pending_locks[i].wh_lock;
788 		KASSERT(lock->lo_flags & LO_WITNESS,
789 		    ("%s: lock %s is on pending list but not LO_WITNESS",
790 		    __func__, lock->lo_name));
791 		lock->lo_witness = enroll(pending_locks[i].wh_type,
792 		    LOCK_CLASS(lock));
793 	}
794 
795 	/* Mark the witness code as being ready for use. */
796 	witness_cold = 0;
797 
798 	mtx_lock(&Giant);
799 }
800 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
801     NULL);
802 
803 void
804 witness_init(struct lock_object *lock, const char *type)
805 {
806 	struct lock_class *class;
807 
808 	/* Various sanity checks. */
809 	class = LOCK_CLASS(lock);
810 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
811 	    (class->lc_flags & LC_RECURSABLE) == 0)
812 		kassert_panic("%s: lock (%s) %s can not be recursable",
813 		    __func__, class->lc_name, lock->lo_name);
814 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
815 	    (class->lc_flags & LC_SLEEPABLE) == 0)
816 		kassert_panic("%s: lock (%s) %s can not be sleepable",
817 		    __func__, class->lc_name, lock->lo_name);
818 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
819 	    (class->lc_flags & LC_UPGRADABLE) == 0)
820 		kassert_panic("%s: lock (%s) %s can not be upgradable",
821 		    __func__, class->lc_name, lock->lo_name);
822 
823 	/*
824 	 * If we shouldn't watch this lock, then just clear lo_witness.
825 	 * Otherwise, if witness_cold is set, then it is too early to
826 	 * enroll this lock, so defer it to witness_initialize() by adding
827 	 * it to the pending_locks list.  If it is not too early, then enroll
828 	 * the lock now.
829 	 */
830 	if (witness_watch < 1 || panicstr != NULL ||
831 	    (lock->lo_flags & LO_WITNESS) == 0)
832 		lock->lo_witness = NULL;
833 	else if (witness_cold) {
834 		pending_locks[pending_cnt].wh_lock = lock;
835 		pending_locks[pending_cnt++].wh_type = type;
836 		if (pending_cnt > WITNESS_PENDLIST)
837 			panic("%s: pending locks list is too small, "
838 			    "increase WITNESS_PENDLIST\n",
839 			    __func__);
840 	} else
841 		lock->lo_witness = enroll(type, class);
842 }
843 
844 void
845 witness_destroy(struct lock_object *lock)
846 {
847 	struct lock_class *class;
848 	struct witness *w;
849 
850 	class = LOCK_CLASS(lock);
851 
852 	if (witness_cold)
853 		panic("lock (%s) %s destroyed while witness_cold",
854 		    class->lc_name, lock->lo_name);
855 
856 	/* XXX: need to verify that no one holds the lock */
857 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
858 		return;
859 	w = lock->lo_witness;
860 
861 	mtx_lock_spin(&w_mtx);
862 	MPASS(w->w_refcount > 0);
863 	w->w_refcount--;
864 
865 	if (w->w_refcount == 0)
866 		depart(w);
867 	mtx_unlock_spin(&w_mtx);
868 }
869 
870 #ifdef DDB
871 static void
872 witness_ddb_compute_levels(void)
873 {
874 	struct witness *w;
875 
876 	/*
877 	 * First clear all levels.
878 	 */
879 	STAILQ_FOREACH(w, &w_all, w_list)
880 		w->w_ddb_level = -1;
881 
882 	/*
883 	 * Look for locks with no parents and level all their descendants.
884 	 */
885 	STAILQ_FOREACH(w, &w_all, w_list) {
886 
887 		/* If the witness has ancestors (is not a root), skip it. */
888 		if (w->w_num_ancestors > 0)
889 			continue;
890 		witness_ddb_level_descendants(w, 0);
891 	}
892 }
893 
894 static void
895 witness_ddb_level_descendants(struct witness *w, int l)
896 {
897 	int i;
898 
899 	if (w->w_ddb_level >= l)
900 		return;
901 
902 	w->w_ddb_level = l;
903 	l++;
904 
905 	for (i = 1; i <= w_max_used_index; i++) {
906 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
907 			witness_ddb_level_descendants(&w_data[i], l);
908 	}
909 }
910 
911 static void
912 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
913     struct witness *w, int indent)
914 {
915 	int i;
916 
917  	for (i = 0; i < indent; i++)
918  		prnt(" ");
919 	prnt("%s (type: %s, depth: %d, active refs: %d)",
920 	     w->w_name, w->w_class->lc_name,
921 	     w->w_ddb_level, w->w_refcount);
922  	if (w->w_displayed) {
923  		prnt(" -- (already displayed)\n");
924  		return;
925  	}
926  	w->w_displayed = 1;
927 	if (w->w_file != NULL && w->w_line != 0)
928 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
929 		    w->w_line);
930 	else
931 		prnt(" -- never acquired\n");
932 	indent++;
933 	WITNESS_INDEX_ASSERT(w->w_index);
934 	for (i = 1; i <= w_max_used_index; i++) {
935 		if (db_pager_quit)
936 			return;
937 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
938 			witness_ddb_display_descendants(prnt, &w_data[i],
939 			    indent);
940 	}
941 }
942 
943 static void
944 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
945     struct witness_list *list)
946 {
947 	struct witness *w;
948 
949 	STAILQ_FOREACH(w, list, w_typelist) {
950 		if (w->w_file == NULL || w->w_ddb_level > 0)
951 			continue;
952 
953 		/* This lock has no anscestors - display its descendants. */
954 		witness_ddb_display_descendants(prnt, w, 0);
955 		if (db_pager_quit)
956 			return;
957 	}
958 }
959 
960 static void
961 witness_ddb_display(int(*prnt)(const char *fmt, ...))
962 {
963 	struct witness *w;
964 
965 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
966 	witness_ddb_compute_levels();
967 
968 	/* Clear all the displayed flags. */
969 	STAILQ_FOREACH(w, &w_all, w_list)
970 		w->w_displayed = 0;
971 
972 	/*
973 	 * First, handle sleep locks which have been acquired at least
974 	 * once.
975 	 */
976 	prnt("Sleep locks:\n");
977 	witness_ddb_display_list(prnt, &w_sleep);
978 	if (db_pager_quit)
979 		return;
980 
981 	/*
982 	 * Now do spin locks which have been acquired at least once.
983 	 */
984 	prnt("\nSpin locks:\n");
985 	witness_ddb_display_list(prnt, &w_spin);
986 	if (db_pager_quit)
987 		return;
988 
989 	/*
990 	 * Finally, any locks which have not been acquired yet.
991 	 */
992 	prnt("\nLocks which were never acquired:\n");
993 	STAILQ_FOREACH(w, &w_all, w_list) {
994 		if (w->w_file != NULL || w->w_refcount == 0)
995 			continue;
996 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
997 		    w->w_class->lc_name, w->w_ddb_level);
998 		if (db_pager_quit)
999 			return;
1000 	}
1001 }
1002 #endif /* DDB */
1003 
1004 int
1005 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1006 {
1007 
1008 	if (witness_watch == -1 || panicstr != NULL)
1009 		return (0);
1010 
1011 	/* Require locks that witness knows about. */
1012 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1013 	    lock2->lo_witness == NULL)
1014 		return (EINVAL);
1015 
1016 	mtx_assert(&w_mtx, MA_NOTOWNED);
1017 	mtx_lock_spin(&w_mtx);
1018 
1019 	/*
1020 	 * If we already have either an explicit or implied lock order that
1021 	 * is the other way around, then return an error.
1022 	 */
1023 	if (witness_watch &&
1024 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1025 		mtx_unlock_spin(&w_mtx);
1026 		return (EDOOFUS);
1027 	}
1028 
1029 	/* Try to add the new order. */
1030 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1031 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1032 	itismychild(lock1->lo_witness, lock2->lo_witness);
1033 	mtx_unlock_spin(&w_mtx);
1034 	return (0);
1035 }
1036 
1037 void
1038 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1039     int line, struct lock_object *interlock)
1040 {
1041 	struct lock_list_entry *lock_list, *lle;
1042 	struct lock_instance *lock1, *lock2, *plock;
1043 	struct lock_class *class, *iclass;
1044 	struct witness *w, *w1;
1045 	struct thread *td;
1046 	int i, j;
1047 
1048 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1049 	    panicstr != NULL)
1050 		return;
1051 
1052 	w = lock->lo_witness;
1053 	class = LOCK_CLASS(lock);
1054 	td = curthread;
1055 
1056 	if (class->lc_flags & LC_SLEEPLOCK) {
1057 
1058 		/*
1059 		 * Since spin locks include a critical section, this check
1060 		 * implicitly enforces a lock order of all sleep locks before
1061 		 * all spin locks.
1062 		 */
1063 		if (td->td_critnest != 0 && !kdb_active)
1064 			kassert_panic("acquiring blockable sleep lock with "
1065 			    "spinlock or critical section held (%s) %s @ %s:%d",
1066 			    class->lc_name, lock->lo_name,
1067 			    fixup_filename(file), line);
1068 
1069 		/*
1070 		 * If this is the first lock acquired then just return as
1071 		 * no order checking is needed.
1072 		 */
1073 		lock_list = td->td_sleeplocks;
1074 		if (lock_list == NULL || lock_list->ll_count == 0)
1075 			return;
1076 	} else {
1077 
1078 		/*
1079 		 * If this is the first lock, just return as no order
1080 		 * checking is needed.  Avoid problems with thread
1081 		 * migration pinning the thread while checking if
1082 		 * spinlocks are held.  If at least one spinlock is held
1083 		 * the thread is in a safe path and it is allowed to
1084 		 * unpin it.
1085 		 */
1086 		sched_pin();
1087 		lock_list = PCPU_GET(spinlocks);
1088 		if (lock_list == NULL || lock_list->ll_count == 0) {
1089 			sched_unpin();
1090 			return;
1091 		}
1092 		sched_unpin();
1093 	}
1094 
1095 	/*
1096 	 * Check to see if we are recursing on a lock we already own.  If
1097 	 * so, make sure that we don't mismatch exclusive and shared lock
1098 	 * acquires.
1099 	 */
1100 	lock1 = find_instance(lock_list, lock);
1101 	if (lock1 != NULL) {
1102 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1103 		    (flags & LOP_EXCLUSIVE) == 0) {
1104 			printf("shared lock of (%s) %s @ %s:%d\n",
1105 			    class->lc_name, lock->lo_name,
1106 			    fixup_filename(file), line);
1107 			printf("while exclusively locked from %s:%d\n",
1108 			    fixup_filename(lock1->li_file), lock1->li_line);
1109 			kassert_panic("excl->share");
1110 		}
1111 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1112 		    (flags & LOP_EXCLUSIVE) != 0) {
1113 			printf("exclusive lock of (%s) %s @ %s:%d\n",
1114 			    class->lc_name, lock->lo_name,
1115 			    fixup_filename(file), line);
1116 			printf("while share locked from %s:%d\n",
1117 			    fixup_filename(lock1->li_file), lock1->li_line);
1118 			kassert_panic("share->excl");
1119 		}
1120 		return;
1121 	}
1122 
1123 	/* Warn if the interlock is not locked exactly once. */
1124 	if (interlock != NULL) {
1125 		iclass = LOCK_CLASS(interlock);
1126 		lock1 = find_instance(lock_list, interlock);
1127 		if (lock1 == NULL)
1128 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1129 			    iclass->lc_name, interlock->lo_name,
1130 			    fixup_filename(file), line);
1131 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1132 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1133 			    iclass->lc_name, interlock->lo_name,
1134 			    fixup_filename(file), line);
1135 	}
1136 
1137 	/*
1138 	 * Find the previously acquired lock, but ignore interlocks.
1139 	 */
1140 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1141 	if (interlock != NULL && plock->li_lock == interlock) {
1142 		if (lock_list->ll_count > 1)
1143 			plock =
1144 			    &lock_list->ll_children[lock_list->ll_count - 2];
1145 		else {
1146 			lle = lock_list->ll_next;
1147 
1148 			/*
1149 			 * The interlock is the only lock we hold, so
1150 			 * simply return.
1151 			 */
1152 			if (lle == NULL)
1153 				return;
1154 			plock = &lle->ll_children[lle->ll_count - 1];
1155 		}
1156 	}
1157 
1158 	/*
1159 	 * Try to perform most checks without a lock.  If this succeeds we
1160 	 * can skip acquiring the lock and return success.
1161 	 */
1162 	w1 = plock->li_lock->lo_witness;
1163 	if (witness_lock_order_check(w1, w))
1164 		return;
1165 
1166 	/*
1167 	 * Check for duplicate locks of the same type.  Note that we only
1168 	 * have to check for this on the last lock we just acquired.  Any
1169 	 * other cases will be caught as lock order violations.
1170 	 */
1171 	mtx_lock_spin(&w_mtx);
1172 	witness_lock_order_add(w1, w);
1173 	if (w1 == w) {
1174 		i = w->w_index;
1175 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1176 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1177 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1178 			w->w_reversed = 1;
1179 			mtx_unlock_spin(&w_mtx);
1180 			printf(
1181 			    "acquiring duplicate lock of same type: \"%s\"\n",
1182 			    w->w_name);
1183 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1184 			    fixup_filename(plock->li_file), plock->li_line);
1185 			printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1186 			    fixup_filename(file), line);
1187 			witness_debugger(1);
1188 		} else
1189 			mtx_unlock_spin(&w_mtx);
1190 		return;
1191 	}
1192 	mtx_assert(&w_mtx, MA_OWNED);
1193 
1194 	/*
1195 	 * If we know that the lock we are acquiring comes after
1196 	 * the lock we most recently acquired in the lock order tree,
1197 	 * then there is no need for any further checks.
1198 	 */
1199 	if (isitmychild(w1, w))
1200 		goto out;
1201 
1202 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1203 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1204 
1205 			MPASS(j < WITNESS_COUNT);
1206 			lock1 = &lle->ll_children[i];
1207 
1208 			/*
1209 			 * Ignore the interlock.
1210 			 */
1211 			if (interlock == lock1->li_lock)
1212 				continue;
1213 
1214 			/*
1215 			 * If this lock doesn't undergo witness checking,
1216 			 * then skip it.
1217 			 */
1218 			w1 = lock1->li_lock->lo_witness;
1219 			if (w1 == NULL) {
1220 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1221 				    ("lock missing witness structure"));
1222 				continue;
1223 			}
1224 
1225 			/*
1226 			 * If we are locking Giant and this is a sleepable
1227 			 * lock, then skip it.
1228 			 */
1229 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1230 			    lock == &Giant.lock_object)
1231 				continue;
1232 
1233 			/*
1234 			 * If we are locking a sleepable lock and this lock
1235 			 * is Giant, then skip it.
1236 			 */
1237 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1238 			    lock1->li_lock == &Giant.lock_object)
1239 				continue;
1240 
1241 			/*
1242 			 * If we are locking a sleepable lock and this lock
1243 			 * isn't sleepable, we want to treat it as a lock
1244 			 * order violation to enfore a general lock order of
1245 			 * sleepable locks before non-sleepable locks.
1246 			 */
1247 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1248 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1249 				goto reversal;
1250 
1251 			/*
1252 			 * If we are locking Giant and this is a non-sleepable
1253 			 * lock, then treat it as a reversal.
1254 			 */
1255 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1256 			    lock == &Giant.lock_object)
1257 				goto reversal;
1258 
1259 			/*
1260 			 * Check the lock order hierarchy for a reveresal.
1261 			 */
1262 			if (!isitmydescendant(w, w1))
1263 				continue;
1264 		reversal:
1265 
1266 			/*
1267 			 * We have a lock order violation, check to see if it
1268 			 * is allowed or has already been yelled about.
1269 			 */
1270 #ifdef BLESSING
1271 
1272 			/*
1273 			 * If the lock order is blessed, just bail.  We don't
1274 			 * look for other lock order violations though, which
1275 			 * may be a bug.
1276 			 */
1277 			if (blessed(w, w1))
1278 				goto out;
1279 #endif
1280 
1281 			/* Bail if this violation is known */
1282 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1283 				goto out;
1284 
1285 			/* Record this as a violation */
1286 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1287 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1288 			w->w_reversed = w1->w_reversed = 1;
1289 			witness_increment_graph_generation();
1290 			mtx_unlock_spin(&w_mtx);
1291 
1292 #ifdef WITNESS_NO_VNODE
1293 			/*
1294 			 * There are known LORs between VNODE locks. They are
1295 			 * not an indication of a bug. VNODE locks are flagged
1296 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1297 			 * is between 2 VNODE locks.
1298 			 */
1299 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1300 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1301 				return;
1302 #endif
1303 
1304 			/*
1305 			 * Ok, yell about it.
1306 			 */
1307 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1308 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1309 				printf(
1310 		"lock order reversal: (sleepable after non-sleepable)\n");
1311 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1312 			    && lock == &Giant.lock_object)
1313 				printf(
1314 		"lock order reversal: (Giant after non-sleepable)\n");
1315 			else
1316 				printf("lock order reversal:\n");
1317 
1318 			/*
1319 			 * Try to locate an earlier lock with
1320 			 * witness w in our list.
1321 			 */
1322 			do {
1323 				lock2 = &lle->ll_children[i];
1324 				MPASS(lock2->li_lock != NULL);
1325 				if (lock2->li_lock->lo_witness == w)
1326 					break;
1327 				if (i == 0 && lle->ll_next != NULL) {
1328 					lle = lle->ll_next;
1329 					i = lle->ll_count - 1;
1330 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1331 				} else
1332 					i--;
1333 			} while (i >= 0);
1334 			if (i < 0) {
1335 				printf(" 1st %p %s (%s) @ %s:%d\n",
1336 				    lock1->li_lock, lock1->li_lock->lo_name,
1337 				    w1->w_name, fixup_filename(lock1->li_file),
1338 				    lock1->li_line);
1339 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1340 				    lock->lo_name, w->w_name,
1341 				    fixup_filename(file), line);
1342 			} else {
1343 				printf(" 1st %p %s (%s) @ %s:%d\n",
1344 				    lock2->li_lock, lock2->li_lock->lo_name,
1345 				    lock2->li_lock->lo_witness->w_name,
1346 				    fixup_filename(lock2->li_file),
1347 				    lock2->li_line);
1348 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1349 				    lock1->li_lock, lock1->li_lock->lo_name,
1350 				    w1->w_name, fixup_filename(lock1->li_file),
1351 				    lock1->li_line);
1352 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1353 				    lock->lo_name, w->w_name,
1354 				    fixup_filename(file), line);
1355 			}
1356 			witness_debugger(1);
1357 			return;
1358 		}
1359 	}
1360 
1361 	/*
1362 	 * If requested, build a new lock order.  However, don't build a new
1363 	 * relationship between a sleepable lock and Giant if it is in the
1364 	 * wrong direction.  The correct lock order is that sleepable locks
1365 	 * always come before Giant.
1366 	 */
1367 	if (flags & LOP_NEWORDER &&
1368 	    !(plock->li_lock == &Giant.lock_object &&
1369 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1370 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1371 		    w->w_name, plock->li_lock->lo_witness->w_name);
1372 		itismychild(plock->li_lock->lo_witness, w);
1373 	}
1374 out:
1375 	mtx_unlock_spin(&w_mtx);
1376 }
1377 
1378 void
1379 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1380 {
1381 	struct lock_list_entry **lock_list, *lle;
1382 	struct lock_instance *instance;
1383 	struct witness *w;
1384 	struct thread *td;
1385 
1386 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1387 	    panicstr != NULL)
1388 		return;
1389 	w = lock->lo_witness;
1390 	td = curthread;
1391 
1392 	/* Determine lock list for this lock. */
1393 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1394 		lock_list = &td->td_sleeplocks;
1395 	else
1396 		lock_list = PCPU_PTR(spinlocks);
1397 
1398 	/* Check to see if we are recursing on a lock we already own. */
1399 	instance = find_instance(*lock_list, lock);
1400 	if (instance != NULL) {
1401 		instance->li_flags++;
1402 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1403 		    td->td_proc->p_pid, lock->lo_name,
1404 		    instance->li_flags & LI_RECURSEMASK);
1405 		instance->li_file = file;
1406 		instance->li_line = line;
1407 		return;
1408 	}
1409 
1410 	/* Update per-witness last file and line acquire. */
1411 	w->w_file = file;
1412 	w->w_line = line;
1413 
1414 	/* Find the next open lock instance in the list and fill it. */
1415 	lle = *lock_list;
1416 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1417 		lle = witness_lock_list_get();
1418 		if (lle == NULL)
1419 			return;
1420 		lle->ll_next = *lock_list;
1421 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1422 		    td->td_proc->p_pid, lle);
1423 		*lock_list = lle;
1424 	}
1425 	instance = &lle->ll_children[lle->ll_count++];
1426 	instance->li_lock = lock;
1427 	instance->li_line = line;
1428 	instance->li_file = file;
1429 	if ((flags & LOP_EXCLUSIVE) != 0)
1430 		instance->li_flags = LI_EXCLUSIVE;
1431 	else
1432 		instance->li_flags = 0;
1433 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1434 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1435 }
1436 
1437 void
1438 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1439 {
1440 	struct lock_instance *instance;
1441 	struct lock_class *class;
1442 
1443 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1444 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1445 		return;
1446 	class = LOCK_CLASS(lock);
1447 	if (witness_watch) {
1448 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1449 			kassert_panic(
1450 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1451 			    class->lc_name, lock->lo_name,
1452 			    fixup_filename(file), line);
1453 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1454 			kassert_panic(
1455 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1456 			    class->lc_name, lock->lo_name,
1457 			    fixup_filename(file), line);
1458 	}
1459 	instance = find_instance(curthread->td_sleeplocks, lock);
1460 	if (instance == NULL) {
1461 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1462 		    class->lc_name, lock->lo_name,
1463 		    fixup_filename(file), line);
1464 		return;
1465 	}
1466 	if (witness_watch) {
1467 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1468 			kassert_panic(
1469 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1470 			    class->lc_name, lock->lo_name,
1471 			    fixup_filename(file), line);
1472 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1473 			kassert_panic(
1474 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1475 			    class->lc_name, lock->lo_name,
1476 			    instance->li_flags & LI_RECURSEMASK,
1477 			    fixup_filename(file), line);
1478 	}
1479 	instance->li_flags |= LI_EXCLUSIVE;
1480 }
1481 
1482 void
1483 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1484     int line)
1485 {
1486 	struct lock_instance *instance;
1487 	struct lock_class *class;
1488 
1489 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1490 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1491 		return;
1492 	class = LOCK_CLASS(lock);
1493 	if (witness_watch) {
1494 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1495 			kassert_panic(
1496 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1497 			    class->lc_name, lock->lo_name,
1498 			    fixup_filename(file), line);
1499 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1500 			kassert_panic(
1501 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1502 			    class->lc_name, lock->lo_name,
1503 			    fixup_filename(file), line);
1504 	}
1505 	instance = find_instance(curthread->td_sleeplocks, lock);
1506 	if (instance == NULL) {
1507 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1508 		    class->lc_name, lock->lo_name,
1509 		    fixup_filename(file), line);
1510 		return;
1511 	}
1512 	if (witness_watch) {
1513 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1514 			kassert_panic(
1515 			    "downgrade of shared lock (%s) %s @ %s:%d",
1516 			    class->lc_name, lock->lo_name,
1517 			    fixup_filename(file), line);
1518 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1519 			kassert_panic(
1520 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1521 			    class->lc_name, lock->lo_name,
1522 			    instance->li_flags & LI_RECURSEMASK,
1523 			    fixup_filename(file), line);
1524 	}
1525 	instance->li_flags &= ~LI_EXCLUSIVE;
1526 }
1527 
1528 void
1529 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1530 {
1531 	struct lock_list_entry **lock_list, *lle;
1532 	struct lock_instance *instance;
1533 	struct lock_class *class;
1534 	struct thread *td;
1535 	register_t s;
1536 	int i, j;
1537 
1538 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1539 		return;
1540 	td = curthread;
1541 	class = LOCK_CLASS(lock);
1542 
1543 	/* Find lock instance associated with this lock. */
1544 	if (class->lc_flags & LC_SLEEPLOCK)
1545 		lock_list = &td->td_sleeplocks;
1546 	else
1547 		lock_list = PCPU_PTR(spinlocks);
1548 	lle = *lock_list;
1549 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1550 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1551 			instance = &(*lock_list)->ll_children[i];
1552 			if (instance->li_lock == lock)
1553 				goto found;
1554 		}
1555 
1556 	/*
1557 	 * When disabling WITNESS through witness_watch we could end up in
1558 	 * having registered locks in the td_sleeplocks queue.
1559 	 * We have to make sure we flush these queues, so just search for
1560 	 * eventual register locks and remove them.
1561 	 */
1562 	if (witness_watch > 0) {
1563 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1564 		    lock->lo_name, fixup_filename(file), line);
1565 		return;
1566 	} else {
1567 		return;
1568 	}
1569 found:
1570 
1571 	/* First, check for shared/exclusive mismatches. */
1572 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1573 	    (flags & LOP_EXCLUSIVE) == 0) {
1574 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1575 		    lock->lo_name, fixup_filename(file), line);
1576 		printf("while exclusively locked from %s:%d\n",
1577 		    fixup_filename(instance->li_file), instance->li_line);
1578 		kassert_panic("excl->ushare");
1579 	}
1580 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1581 	    (flags & LOP_EXCLUSIVE) != 0) {
1582 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1583 		    lock->lo_name, fixup_filename(file), line);
1584 		printf("while share locked from %s:%d\n",
1585 		    fixup_filename(instance->li_file),
1586 		    instance->li_line);
1587 		kassert_panic("share->uexcl");
1588 	}
1589 	/* If we are recursed, unrecurse. */
1590 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1591 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1592 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1593 		    instance->li_flags);
1594 		instance->li_flags--;
1595 		return;
1596 	}
1597 	/* The lock is now being dropped, check for NORELEASE flag */
1598 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1599 		printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1600 		    lock->lo_name, fixup_filename(file), line);
1601 		kassert_panic("lock marked norelease");
1602 	}
1603 
1604 	/* Otherwise, remove this item from the list. */
1605 	s = intr_disable();
1606 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1607 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1608 	    (*lock_list)->ll_count - 1);
1609 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1610 		(*lock_list)->ll_children[j] =
1611 		    (*lock_list)->ll_children[j + 1];
1612 	(*lock_list)->ll_count--;
1613 	intr_restore(s);
1614 
1615 	/*
1616 	 * In order to reduce contention on w_mtx, we want to keep always an
1617 	 * head object into lists so that frequent allocation from the
1618 	 * free witness pool (and subsequent locking) is avoided.
1619 	 * In order to maintain the current code simple, when the head
1620 	 * object is totally unloaded it means also that we do not have
1621 	 * further objects in the list, so the list ownership needs to be
1622 	 * hand over to another object if the current head needs to be freed.
1623 	 */
1624 	if ((*lock_list)->ll_count == 0) {
1625 		if (*lock_list == lle) {
1626 			if (lle->ll_next == NULL)
1627 				return;
1628 		} else
1629 			lle = *lock_list;
1630 		*lock_list = lle->ll_next;
1631 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1632 		    td->td_proc->p_pid, lle);
1633 		witness_lock_list_free(lle);
1634 	}
1635 }
1636 
1637 void
1638 witness_thread_exit(struct thread *td)
1639 {
1640 	struct lock_list_entry *lle;
1641 	int i, n;
1642 
1643 	lle = td->td_sleeplocks;
1644 	if (lle == NULL || panicstr != NULL)
1645 		return;
1646 	if (lle->ll_count != 0) {
1647 		for (n = 0; lle != NULL; lle = lle->ll_next)
1648 			for (i = lle->ll_count - 1; i >= 0; i--) {
1649 				if (n == 0)
1650 		printf("Thread %p exiting with the following locks held:\n",
1651 					    td);
1652 				n++;
1653 				witness_list_lock(&lle->ll_children[i], printf);
1654 
1655 			}
1656 		kassert_panic(
1657 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1658 	}
1659 	witness_lock_list_free(lle);
1660 }
1661 
1662 /*
1663  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1664  * exempt Giant and sleepable locks from the checks as well.  If any
1665  * non-exempt locks are held, then a supplied message is printed to the
1666  * console along with a list of the offending locks.  If indicated in the
1667  * flags then a failure results in a panic as well.
1668  */
1669 int
1670 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1671 {
1672 	struct lock_list_entry *lock_list, *lle;
1673 	struct lock_instance *lock1;
1674 	struct thread *td;
1675 	va_list ap;
1676 	int i, n;
1677 
1678 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1679 		return (0);
1680 	n = 0;
1681 	td = curthread;
1682 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1683 		for (i = lle->ll_count - 1; i >= 0; i--) {
1684 			lock1 = &lle->ll_children[i];
1685 			if (lock1->li_lock == lock)
1686 				continue;
1687 			if (flags & WARN_GIANTOK &&
1688 			    lock1->li_lock == &Giant.lock_object)
1689 				continue;
1690 			if (flags & WARN_SLEEPOK &&
1691 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1692 				continue;
1693 			if (n == 0) {
1694 				va_start(ap, fmt);
1695 				vprintf(fmt, ap);
1696 				va_end(ap);
1697 				printf(" with the following");
1698 				if (flags & WARN_SLEEPOK)
1699 					printf(" non-sleepable");
1700 				printf(" locks held:\n");
1701 			}
1702 			n++;
1703 			witness_list_lock(lock1, printf);
1704 		}
1705 
1706 	/*
1707 	 * Pin the thread in order to avoid problems with thread migration.
1708 	 * Once that all verifies are passed about spinlocks ownership,
1709 	 * the thread is in a safe path and it can be unpinned.
1710 	 */
1711 	sched_pin();
1712 	lock_list = PCPU_GET(spinlocks);
1713 	if (lock_list != NULL && lock_list->ll_count != 0) {
1714 		sched_unpin();
1715 
1716 		/*
1717 		 * We should only have one spinlock and as long as
1718 		 * the flags cannot match for this locks class,
1719 		 * check if the first spinlock is the one curthread
1720 		 * should hold.
1721 		 */
1722 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1723 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1724 		    lock1->li_lock == lock && n == 0)
1725 			return (0);
1726 
1727 		va_start(ap, fmt);
1728 		vprintf(fmt, ap);
1729 		va_end(ap);
1730 		printf(" with the following");
1731 		if (flags & WARN_SLEEPOK)
1732 			printf(" non-sleepable");
1733 		printf(" locks held:\n");
1734 		n += witness_list_locks(&lock_list, printf);
1735 	} else
1736 		sched_unpin();
1737 	if (flags & WARN_PANIC && n)
1738 		kassert_panic("%s", __func__);
1739 	else
1740 		witness_debugger(n);
1741 	return (n);
1742 }
1743 
1744 const char *
1745 witness_file(struct lock_object *lock)
1746 {
1747 	struct witness *w;
1748 
1749 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1750 		return ("?");
1751 	w = lock->lo_witness;
1752 	return (w->w_file);
1753 }
1754 
1755 int
1756 witness_line(struct lock_object *lock)
1757 {
1758 	struct witness *w;
1759 
1760 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1761 		return (0);
1762 	w = lock->lo_witness;
1763 	return (w->w_line);
1764 }
1765 
1766 static struct witness *
1767 enroll(const char *description, struct lock_class *lock_class)
1768 {
1769 	struct witness *w;
1770 	struct witness_list *typelist;
1771 
1772 	MPASS(description != NULL);
1773 
1774 	if (witness_watch == -1 || panicstr != NULL)
1775 		return (NULL);
1776 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1777 		if (witness_skipspin)
1778 			return (NULL);
1779 		else
1780 			typelist = &w_spin;
1781 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1782 		typelist = &w_sleep;
1783 	} else {
1784 		kassert_panic("lock class %s is not sleep or spin",
1785 		    lock_class->lc_name);
1786 		return (NULL);
1787 	}
1788 
1789 	mtx_lock_spin(&w_mtx);
1790 	w = witness_hash_get(description);
1791 	if (w)
1792 		goto found;
1793 	if ((w = witness_get()) == NULL)
1794 		return (NULL);
1795 	MPASS(strlen(description) < MAX_W_NAME);
1796 	strcpy(w->w_name, description);
1797 	w->w_class = lock_class;
1798 	w->w_refcount = 1;
1799 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1800 	if (lock_class->lc_flags & LC_SPINLOCK) {
1801 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1802 		w_spin_cnt++;
1803 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1804 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1805 		w_sleep_cnt++;
1806 	}
1807 
1808 	/* Insert new witness into the hash */
1809 	witness_hash_put(w);
1810 	witness_increment_graph_generation();
1811 	mtx_unlock_spin(&w_mtx);
1812 	return (w);
1813 found:
1814 	w->w_refcount++;
1815 	mtx_unlock_spin(&w_mtx);
1816 	if (lock_class != w->w_class)
1817 		kassert_panic(
1818 			"lock (%s) %s does not match earlier (%s) lock",
1819 			description, lock_class->lc_name,
1820 			w->w_class->lc_name);
1821 	return (w);
1822 }
1823 
1824 static void
1825 depart(struct witness *w)
1826 {
1827 	struct witness_list *list;
1828 
1829 	MPASS(w->w_refcount == 0);
1830 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1831 		list = &w_sleep;
1832 		w_sleep_cnt--;
1833 	} else {
1834 		list = &w_spin;
1835 		w_spin_cnt--;
1836 	}
1837 	/*
1838 	 * Set file to NULL as it may point into a loadable module.
1839 	 */
1840 	w->w_file = NULL;
1841 	w->w_line = 0;
1842 	witness_increment_graph_generation();
1843 }
1844 
1845 
1846 static void
1847 adopt(struct witness *parent, struct witness *child)
1848 {
1849 	int pi, ci, i, j;
1850 
1851 	if (witness_cold == 0)
1852 		mtx_assert(&w_mtx, MA_OWNED);
1853 
1854 	/* If the relationship is already known, there's no work to be done. */
1855 	if (isitmychild(parent, child))
1856 		return;
1857 
1858 	/* When the structure of the graph changes, bump up the generation. */
1859 	witness_increment_graph_generation();
1860 
1861 	/*
1862 	 * The hard part ... create the direct relationship, then propagate all
1863 	 * indirect relationships.
1864 	 */
1865 	pi = parent->w_index;
1866 	ci = child->w_index;
1867 	WITNESS_INDEX_ASSERT(pi);
1868 	WITNESS_INDEX_ASSERT(ci);
1869 	MPASS(pi != ci);
1870 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1871 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1872 
1873 	/*
1874 	 * If parent was not already an ancestor of child,
1875 	 * then we increment the descendant and ancestor counters.
1876 	 */
1877 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1878 		parent->w_num_descendants++;
1879 		child->w_num_ancestors++;
1880 	}
1881 
1882 	/*
1883 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1884 	 * an ancestor of 'pi' during this loop.
1885 	 */
1886 	for (i = 1; i <= w_max_used_index; i++) {
1887 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1888 		    (i != pi))
1889 			continue;
1890 
1891 		/* Find each descendant of 'i' and mark it as a descendant. */
1892 		for (j = 1; j <= w_max_used_index; j++) {
1893 
1894 			/*
1895 			 * Skip children that are already marked as
1896 			 * descendants of 'i'.
1897 			 */
1898 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1899 				continue;
1900 
1901 			/*
1902 			 * We are only interested in descendants of 'ci'. Note
1903 			 * that 'ci' itself is counted as a descendant of 'ci'.
1904 			 */
1905 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1906 			    (j != ci))
1907 				continue;
1908 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1909 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1910 			w_data[i].w_num_descendants++;
1911 			w_data[j].w_num_ancestors++;
1912 
1913 			/*
1914 			 * Make sure we aren't marking a node as both an
1915 			 * ancestor and descendant. We should have caught
1916 			 * this as a lock order reversal earlier.
1917 			 */
1918 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1919 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1920 				printf("witness rmatrix paradox! [%d][%d]=%d "
1921 				    "both ancestor and descendant\n",
1922 				    i, j, w_rmatrix[i][j]);
1923 				kdb_backtrace();
1924 				printf("Witness disabled.\n");
1925 				witness_watch = -1;
1926 			}
1927 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1928 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1929 				printf("witness rmatrix paradox! [%d][%d]=%d "
1930 				    "both ancestor and descendant\n",
1931 				    j, i, w_rmatrix[j][i]);
1932 				kdb_backtrace();
1933 				printf("Witness disabled.\n");
1934 				witness_watch = -1;
1935 			}
1936 		}
1937 	}
1938 }
1939 
1940 static void
1941 itismychild(struct witness *parent, struct witness *child)
1942 {
1943 	int unlocked;
1944 
1945 	MPASS(child != NULL && parent != NULL);
1946 	if (witness_cold == 0)
1947 		mtx_assert(&w_mtx, MA_OWNED);
1948 
1949 	if (!witness_lock_type_equal(parent, child)) {
1950 		if (witness_cold == 0) {
1951 			unlocked = 1;
1952 			mtx_unlock_spin(&w_mtx);
1953 		} else {
1954 			unlocked = 0;
1955 		}
1956 		kassert_panic(
1957 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1958 		    "the same lock type", __func__, parent->w_name,
1959 		    parent->w_class->lc_name, child->w_name,
1960 		    child->w_class->lc_name);
1961 		if (unlocked)
1962 			mtx_lock_spin(&w_mtx);
1963 	}
1964 	adopt(parent, child);
1965 }
1966 
1967 /*
1968  * Generic code for the isitmy*() functions. The rmask parameter is the
1969  * expected relationship of w1 to w2.
1970  */
1971 static int
1972 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1973 {
1974 	unsigned char r1, r2;
1975 	int i1, i2;
1976 
1977 	i1 = w1->w_index;
1978 	i2 = w2->w_index;
1979 	WITNESS_INDEX_ASSERT(i1);
1980 	WITNESS_INDEX_ASSERT(i2);
1981 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1982 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1983 
1984 	/* The flags on one better be the inverse of the flags on the other */
1985 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1986 		(WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1987 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
1988 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
1989 		    "w_rmatrix[%d][%d] == %hhx\n",
1990 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1991 		    i2, i1, r2);
1992 		kdb_backtrace();
1993 		printf("Witness disabled.\n");
1994 		witness_watch = -1;
1995 	}
1996 	return (r1 & rmask);
1997 }
1998 
1999 /*
2000  * Checks if @child is a direct child of @parent.
2001  */
2002 static int
2003 isitmychild(struct witness *parent, struct witness *child)
2004 {
2005 
2006 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2007 }
2008 
2009 /*
2010  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2011  */
2012 static int
2013 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2014 {
2015 
2016 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2017 	    __func__));
2018 }
2019 
2020 #ifdef BLESSING
2021 static int
2022 blessed(struct witness *w1, struct witness *w2)
2023 {
2024 	int i;
2025 	struct witness_blessed *b;
2026 
2027 	for (i = 0; i < blessed_count; i++) {
2028 		b = &blessed_list[i];
2029 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2030 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2031 				return (1);
2032 			continue;
2033 		}
2034 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2035 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2036 				return (1);
2037 	}
2038 	return (0);
2039 }
2040 #endif
2041 
2042 static struct witness *
2043 witness_get(void)
2044 {
2045 	struct witness *w;
2046 	int index;
2047 
2048 	if (witness_cold == 0)
2049 		mtx_assert(&w_mtx, MA_OWNED);
2050 
2051 	if (witness_watch == -1) {
2052 		mtx_unlock_spin(&w_mtx);
2053 		return (NULL);
2054 	}
2055 	if (STAILQ_EMPTY(&w_free)) {
2056 		witness_watch = -1;
2057 		mtx_unlock_spin(&w_mtx);
2058 		printf("WITNESS: unable to allocate a new witness object\n");
2059 		return (NULL);
2060 	}
2061 	w = STAILQ_FIRST(&w_free);
2062 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2063 	w_free_cnt--;
2064 	index = w->w_index;
2065 	MPASS(index > 0 && index == w_max_used_index+1 &&
2066 	    index < WITNESS_COUNT);
2067 	bzero(w, sizeof(*w));
2068 	w->w_index = index;
2069 	if (index > w_max_used_index)
2070 		w_max_used_index = index;
2071 	return (w);
2072 }
2073 
2074 static void
2075 witness_free(struct witness *w)
2076 {
2077 
2078 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2079 	w_free_cnt++;
2080 }
2081 
2082 static struct lock_list_entry *
2083 witness_lock_list_get(void)
2084 {
2085 	struct lock_list_entry *lle;
2086 
2087 	if (witness_watch == -1)
2088 		return (NULL);
2089 	mtx_lock_spin(&w_mtx);
2090 	lle = w_lock_list_free;
2091 	if (lle == NULL) {
2092 		witness_watch = -1;
2093 		mtx_unlock_spin(&w_mtx);
2094 		printf("%s: witness exhausted\n", __func__);
2095 		return (NULL);
2096 	}
2097 	w_lock_list_free = lle->ll_next;
2098 	mtx_unlock_spin(&w_mtx);
2099 	bzero(lle, sizeof(*lle));
2100 	return (lle);
2101 }
2102 
2103 static void
2104 witness_lock_list_free(struct lock_list_entry *lle)
2105 {
2106 
2107 	mtx_lock_spin(&w_mtx);
2108 	lle->ll_next = w_lock_list_free;
2109 	w_lock_list_free = lle;
2110 	mtx_unlock_spin(&w_mtx);
2111 }
2112 
2113 static struct lock_instance *
2114 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2115 {
2116 	struct lock_list_entry *lle;
2117 	struct lock_instance *instance;
2118 	int i;
2119 
2120 	for (lle = list; lle != NULL; lle = lle->ll_next)
2121 		for (i = lle->ll_count - 1; i >= 0; i--) {
2122 			instance = &lle->ll_children[i];
2123 			if (instance->li_lock == lock)
2124 				return (instance);
2125 		}
2126 	return (NULL);
2127 }
2128 
2129 static void
2130 witness_list_lock(struct lock_instance *instance,
2131     int (*prnt)(const char *fmt, ...))
2132 {
2133 	struct lock_object *lock;
2134 
2135 	lock = instance->li_lock;
2136 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2137 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2138 	if (lock->lo_witness->w_name != lock->lo_name)
2139 		prnt(" (%s)", lock->lo_witness->w_name);
2140 	prnt(" r = %d (%p) locked @ %s:%d\n",
2141 	    instance->li_flags & LI_RECURSEMASK, lock,
2142 	    fixup_filename(instance->li_file), instance->li_line);
2143 }
2144 
2145 #ifdef DDB
2146 static int
2147 witness_thread_has_locks(struct thread *td)
2148 {
2149 
2150 	if (td->td_sleeplocks == NULL)
2151 		return (0);
2152 	return (td->td_sleeplocks->ll_count != 0);
2153 }
2154 
2155 static int
2156 witness_proc_has_locks(struct proc *p)
2157 {
2158 	struct thread *td;
2159 
2160 	FOREACH_THREAD_IN_PROC(p, td) {
2161 		if (witness_thread_has_locks(td))
2162 			return (1);
2163 	}
2164 	return (0);
2165 }
2166 #endif
2167 
2168 int
2169 witness_list_locks(struct lock_list_entry **lock_list,
2170     int (*prnt)(const char *fmt, ...))
2171 {
2172 	struct lock_list_entry *lle;
2173 	int i, nheld;
2174 
2175 	nheld = 0;
2176 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2177 		for (i = lle->ll_count - 1; i >= 0; i--) {
2178 			witness_list_lock(&lle->ll_children[i], prnt);
2179 			nheld++;
2180 		}
2181 	return (nheld);
2182 }
2183 
2184 /*
2185  * This is a bit risky at best.  We call this function when we have timed
2186  * out acquiring a spin lock, and we assume that the other CPU is stuck
2187  * with this lock held.  So, we go groveling around in the other CPU's
2188  * per-cpu data to try to find the lock instance for this spin lock to
2189  * see when it was last acquired.
2190  */
2191 void
2192 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2193     int (*prnt)(const char *fmt, ...))
2194 {
2195 	struct lock_instance *instance;
2196 	struct pcpu *pc;
2197 
2198 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2199 		return;
2200 	pc = pcpu_find(owner->td_oncpu);
2201 	instance = find_instance(pc->pc_spinlocks, lock);
2202 	if (instance != NULL)
2203 		witness_list_lock(instance, prnt);
2204 }
2205 
2206 void
2207 witness_save(struct lock_object *lock, const char **filep, int *linep)
2208 {
2209 	struct lock_list_entry *lock_list;
2210 	struct lock_instance *instance;
2211 	struct lock_class *class;
2212 
2213 	/*
2214 	 * This function is used independently in locking code to deal with
2215 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2216 	 * is gone.
2217 	 */
2218 	if (SCHEDULER_STOPPED())
2219 		return;
2220 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2221 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2222 		return;
2223 	class = LOCK_CLASS(lock);
2224 	if (class->lc_flags & LC_SLEEPLOCK)
2225 		lock_list = curthread->td_sleeplocks;
2226 	else {
2227 		if (witness_skipspin)
2228 			return;
2229 		lock_list = PCPU_GET(spinlocks);
2230 	}
2231 	instance = find_instance(lock_list, lock);
2232 	if (instance == NULL) {
2233 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2234 		    class->lc_name, lock->lo_name);
2235 		return;
2236 	}
2237 	*filep = instance->li_file;
2238 	*linep = instance->li_line;
2239 }
2240 
2241 void
2242 witness_restore(struct lock_object *lock, const char *file, int line)
2243 {
2244 	struct lock_list_entry *lock_list;
2245 	struct lock_instance *instance;
2246 	struct lock_class *class;
2247 
2248 	/*
2249 	 * This function is used independently in locking code to deal with
2250 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2251 	 * is gone.
2252 	 */
2253 	if (SCHEDULER_STOPPED())
2254 		return;
2255 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2256 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2257 		return;
2258 	class = LOCK_CLASS(lock);
2259 	if (class->lc_flags & LC_SLEEPLOCK)
2260 		lock_list = curthread->td_sleeplocks;
2261 	else {
2262 		if (witness_skipspin)
2263 			return;
2264 		lock_list = PCPU_GET(spinlocks);
2265 	}
2266 	instance = find_instance(lock_list, lock);
2267 	if (instance == NULL)
2268 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2269 		    class->lc_name, lock->lo_name);
2270 	lock->lo_witness->w_file = file;
2271 	lock->lo_witness->w_line = line;
2272 	if (instance == NULL)
2273 		return;
2274 	instance->li_file = file;
2275 	instance->li_line = line;
2276 }
2277 
2278 void
2279 witness_assert(const struct lock_object *lock, int flags, const char *file,
2280     int line)
2281 {
2282 #ifdef INVARIANT_SUPPORT
2283 	struct lock_instance *instance;
2284 	struct lock_class *class;
2285 
2286 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2287 		return;
2288 	class = LOCK_CLASS(lock);
2289 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2290 		instance = find_instance(curthread->td_sleeplocks, lock);
2291 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2292 		instance = find_instance(PCPU_GET(spinlocks), lock);
2293 	else {
2294 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2295 		    class->lc_name, lock->lo_name);
2296 		return;
2297 	}
2298 	switch (flags) {
2299 	case LA_UNLOCKED:
2300 		if (instance != NULL)
2301 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2302 			    class->lc_name, lock->lo_name,
2303 			    fixup_filename(file), line);
2304 		break;
2305 	case LA_LOCKED:
2306 	case LA_LOCKED | LA_RECURSED:
2307 	case LA_LOCKED | LA_NOTRECURSED:
2308 	case LA_SLOCKED:
2309 	case LA_SLOCKED | LA_RECURSED:
2310 	case LA_SLOCKED | LA_NOTRECURSED:
2311 	case LA_XLOCKED:
2312 	case LA_XLOCKED | LA_RECURSED:
2313 	case LA_XLOCKED | LA_NOTRECURSED:
2314 		if (instance == NULL) {
2315 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2316 			    class->lc_name, lock->lo_name,
2317 			    fixup_filename(file), line);
2318 			break;
2319 		}
2320 		if ((flags & LA_XLOCKED) != 0 &&
2321 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2322 			kassert_panic(
2323 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2324 			    class->lc_name, lock->lo_name,
2325 			    fixup_filename(file), line);
2326 		if ((flags & LA_SLOCKED) != 0 &&
2327 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2328 			kassert_panic(
2329 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2330 			    class->lc_name, lock->lo_name,
2331 			    fixup_filename(file), line);
2332 		if ((flags & LA_RECURSED) != 0 &&
2333 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2334 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2335 			    class->lc_name, lock->lo_name,
2336 			    fixup_filename(file), line);
2337 		if ((flags & LA_NOTRECURSED) != 0 &&
2338 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2339 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2340 			    class->lc_name, lock->lo_name,
2341 			    fixup_filename(file), line);
2342 		break;
2343 	default:
2344 		kassert_panic("Invalid lock assertion at %s:%d.",
2345 		    fixup_filename(file), line);
2346 
2347 	}
2348 #endif	/* INVARIANT_SUPPORT */
2349 }
2350 
2351 static void
2352 witness_setflag(struct lock_object *lock, int flag, int set)
2353 {
2354 	struct lock_list_entry *lock_list;
2355 	struct lock_instance *instance;
2356 	struct lock_class *class;
2357 
2358 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2359 		return;
2360 	class = LOCK_CLASS(lock);
2361 	if (class->lc_flags & LC_SLEEPLOCK)
2362 		lock_list = curthread->td_sleeplocks;
2363 	else {
2364 		if (witness_skipspin)
2365 			return;
2366 		lock_list = PCPU_GET(spinlocks);
2367 	}
2368 	instance = find_instance(lock_list, lock);
2369 	if (instance == NULL) {
2370 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2371 		    class->lc_name, lock->lo_name);
2372 		return;
2373 	}
2374 
2375 	if (set)
2376 		instance->li_flags |= flag;
2377 	else
2378 		instance->li_flags &= ~flag;
2379 }
2380 
2381 void
2382 witness_norelease(struct lock_object *lock)
2383 {
2384 
2385 	witness_setflag(lock, LI_NORELEASE, 1);
2386 }
2387 
2388 void
2389 witness_releaseok(struct lock_object *lock)
2390 {
2391 
2392 	witness_setflag(lock, LI_NORELEASE, 0);
2393 }
2394 
2395 #ifdef DDB
2396 static void
2397 witness_ddb_list(struct thread *td)
2398 {
2399 
2400 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2401 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2402 
2403 	if (witness_watch < 1)
2404 		return;
2405 
2406 	witness_list_locks(&td->td_sleeplocks, db_printf);
2407 
2408 	/*
2409 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2410 	 * if td is currently executing on some other CPU and holds spin locks
2411 	 * as we won't display those locks.  If we had a MI way of getting
2412 	 * the per-cpu data for a given cpu then we could use
2413 	 * td->td_oncpu to get the list of spinlocks for this thread
2414 	 * and "fix" this.
2415 	 *
2416 	 * That still wouldn't really fix this unless we locked the scheduler
2417 	 * lock or stopped the other CPU to make sure it wasn't changing the
2418 	 * list out from under us.  It is probably best to just not try to
2419 	 * handle threads on other CPU's for now.
2420 	 */
2421 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2422 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2423 }
2424 
2425 DB_SHOW_COMMAND(locks, db_witness_list)
2426 {
2427 	struct thread *td;
2428 
2429 	if (have_addr)
2430 		td = db_lookup_thread(addr, TRUE);
2431 	else
2432 		td = kdb_thread;
2433 	witness_ddb_list(td);
2434 }
2435 
2436 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2437 {
2438 	struct thread *td;
2439 	struct proc *p;
2440 
2441 	/*
2442 	 * It would be nice to list only threads and processes that actually
2443 	 * held sleep locks, but that information is currently not exported
2444 	 * by WITNESS.
2445 	 */
2446 	FOREACH_PROC_IN_SYSTEM(p) {
2447 		if (!witness_proc_has_locks(p))
2448 			continue;
2449 		FOREACH_THREAD_IN_PROC(p, td) {
2450 			if (!witness_thread_has_locks(td))
2451 				continue;
2452 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2453 			    p->p_comm, td, td->td_tid);
2454 			witness_ddb_list(td);
2455 			if (db_pager_quit)
2456 				return;
2457 		}
2458 	}
2459 }
2460 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2461 
2462 DB_SHOW_COMMAND(witness, db_witness_display)
2463 {
2464 
2465 	witness_ddb_display(db_printf);
2466 }
2467 #endif
2468 
2469 static int
2470 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2471 {
2472 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2473 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2474 	struct sbuf *sb;
2475 	u_int w_rmatrix1, w_rmatrix2;
2476 	int error, generation, i, j;
2477 
2478 	tmp_data1 = NULL;
2479 	tmp_data2 = NULL;
2480 	tmp_w1 = NULL;
2481 	tmp_w2 = NULL;
2482 	if (witness_watch < 1) {
2483 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2484 		return (error);
2485 	}
2486 	if (witness_cold) {
2487 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2488 		return (error);
2489 	}
2490 	error = 0;
2491 	sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2492 	if (sb == NULL)
2493 		return (ENOMEM);
2494 
2495 	/* Allocate and init temporary storage space. */
2496 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2497 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2498 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2499 	    M_WAITOK | M_ZERO);
2500 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2501 	    M_WAITOK | M_ZERO);
2502 	stack_zero(&tmp_data1->wlod_stack);
2503 	stack_zero(&tmp_data2->wlod_stack);
2504 
2505 restart:
2506 	mtx_lock_spin(&w_mtx);
2507 	generation = w_generation;
2508 	mtx_unlock_spin(&w_mtx);
2509 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2510 	    w_lohash.wloh_count);
2511 	for (i = 1; i < w_max_used_index; i++) {
2512 		mtx_lock_spin(&w_mtx);
2513 		if (generation != w_generation) {
2514 			mtx_unlock_spin(&w_mtx);
2515 
2516 			/* The graph has changed, try again. */
2517 			req->oldidx = 0;
2518 			sbuf_clear(sb);
2519 			goto restart;
2520 		}
2521 
2522 		w1 = &w_data[i];
2523 		if (w1->w_reversed == 0) {
2524 			mtx_unlock_spin(&w_mtx);
2525 			continue;
2526 		}
2527 
2528 		/* Copy w1 locally so we can release the spin lock. */
2529 		*tmp_w1 = *w1;
2530 		mtx_unlock_spin(&w_mtx);
2531 
2532 		if (tmp_w1->w_reversed == 0)
2533 			continue;
2534 		for (j = 1; j < w_max_used_index; j++) {
2535 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2536 				continue;
2537 
2538 			mtx_lock_spin(&w_mtx);
2539 			if (generation != w_generation) {
2540 				mtx_unlock_spin(&w_mtx);
2541 
2542 				/* The graph has changed, try again. */
2543 				req->oldidx = 0;
2544 				sbuf_clear(sb);
2545 				goto restart;
2546 			}
2547 
2548 			w2 = &w_data[j];
2549 			data1 = witness_lock_order_get(w1, w2);
2550 			data2 = witness_lock_order_get(w2, w1);
2551 
2552 			/*
2553 			 * Copy information locally so we can release the
2554 			 * spin lock.
2555 			 */
2556 			*tmp_w2 = *w2;
2557 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2558 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2559 
2560 			if (data1) {
2561 				stack_zero(&tmp_data1->wlod_stack);
2562 				stack_copy(&data1->wlod_stack,
2563 				    &tmp_data1->wlod_stack);
2564 			}
2565 			if (data2 && data2 != data1) {
2566 				stack_zero(&tmp_data2->wlod_stack);
2567 				stack_copy(&data2->wlod_stack,
2568 				    &tmp_data2->wlod_stack);
2569 			}
2570 			mtx_unlock_spin(&w_mtx);
2571 
2572 			sbuf_printf(sb,
2573 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2574 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2575 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2576 #if 0
2577  			sbuf_printf(sb,
2578 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2579  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2580  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2581 #endif
2582 			if (data1) {
2583 				sbuf_printf(sb,
2584 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2585 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2586 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2587 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2588 				sbuf_printf(sb, "\n");
2589 			}
2590 			if (data2 && data2 != data1) {
2591 				sbuf_printf(sb,
2592 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2593 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2594 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2595 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2596 				sbuf_printf(sb, "\n");
2597 			}
2598 		}
2599 	}
2600 	mtx_lock_spin(&w_mtx);
2601 	if (generation != w_generation) {
2602 		mtx_unlock_spin(&w_mtx);
2603 
2604 		/*
2605 		 * The graph changed while we were printing stack data,
2606 		 * try again.
2607 		 */
2608 		req->oldidx = 0;
2609 		sbuf_clear(sb);
2610 		goto restart;
2611 	}
2612 	mtx_unlock_spin(&w_mtx);
2613 
2614 	/* Free temporary storage space. */
2615 	free(tmp_data1, M_TEMP);
2616 	free(tmp_data2, M_TEMP);
2617 	free(tmp_w1, M_TEMP);
2618 	free(tmp_w2, M_TEMP);
2619 
2620 	sbuf_finish(sb);
2621 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2622 	sbuf_delete(sb);
2623 
2624 	return (error);
2625 }
2626 
2627 static int
2628 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2629 {
2630 	struct witness *w;
2631 	struct sbuf *sb;
2632 	int error;
2633 
2634 	if (witness_watch < 1) {
2635 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2636 		return (error);
2637 	}
2638 	if (witness_cold) {
2639 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2640 		return (error);
2641 	}
2642 	error = 0;
2643 
2644 	error = sysctl_wire_old_buffer(req, 0);
2645 	if (error != 0)
2646 		return (error);
2647 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2648 	if (sb == NULL)
2649 		return (ENOMEM);
2650 	sbuf_printf(sb, "\n");
2651 
2652 	mtx_lock_spin(&w_mtx);
2653 	STAILQ_FOREACH(w, &w_all, w_list)
2654 		w->w_displayed = 0;
2655 	STAILQ_FOREACH(w, &w_all, w_list)
2656 		witness_add_fullgraph(sb, w);
2657 	mtx_unlock_spin(&w_mtx);
2658 
2659 	/*
2660 	 * Close the sbuf and return to userland.
2661 	 */
2662 	error = sbuf_finish(sb);
2663 	sbuf_delete(sb);
2664 
2665 	return (error);
2666 }
2667 
2668 static int
2669 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2670 {
2671 	int error, value;
2672 
2673 	value = witness_watch;
2674 	error = sysctl_handle_int(oidp, &value, 0, req);
2675 	if (error != 0 || req->newptr == NULL)
2676 		return (error);
2677 	if (value > 1 || value < -1 ||
2678 	    (witness_watch == -1 && value != witness_watch))
2679 		return (EINVAL);
2680 	witness_watch = value;
2681 	return (0);
2682 }
2683 
2684 static void
2685 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2686 {
2687 	int i;
2688 
2689 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2690 		return;
2691 	w->w_displayed = 1;
2692 
2693 	WITNESS_INDEX_ASSERT(w->w_index);
2694 	for (i = 1; i <= w_max_used_index; i++) {
2695 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2696 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2697 			    w_data[i].w_name);
2698 			witness_add_fullgraph(sb, &w_data[i]);
2699 		}
2700 	}
2701 }
2702 
2703 /*
2704  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2705  * interprets the key as a string and reads until the null
2706  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2707  * hash value computed from the key.
2708  */
2709 static uint32_t
2710 witness_hash_djb2(const uint8_t *key, uint32_t size)
2711 {
2712 	unsigned int hash = 5381;
2713 	int i;
2714 
2715 	/* hash = hash * 33 + key[i] */
2716 	if (size)
2717 		for (i = 0; i < size; i++)
2718 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2719 	else
2720 		for (i = 0; key[i] != 0; i++)
2721 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2722 
2723 	return (hash);
2724 }
2725 
2726 
2727 /*
2728  * Initializes the two witness hash tables. Called exactly once from
2729  * witness_initialize().
2730  */
2731 static void
2732 witness_init_hash_tables(void)
2733 {
2734 	int i;
2735 
2736 	MPASS(witness_cold);
2737 
2738 	/* Initialize the hash tables. */
2739 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2740 		w_hash.wh_array[i] = NULL;
2741 
2742 	w_hash.wh_size = WITNESS_HASH_SIZE;
2743 	w_hash.wh_count = 0;
2744 
2745 	/* Initialize the lock order data hash. */
2746 	w_lofree = NULL;
2747 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2748 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2749 		w_lodata[i].wlod_next = w_lofree;
2750 		w_lofree = &w_lodata[i];
2751 	}
2752 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2753 	w_lohash.wloh_count = 0;
2754 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2755 		w_lohash.wloh_array[i] = NULL;
2756 }
2757 
2758 static struct witness *
2759 witness_hash_get(const char *key)
2760 {
2761 	struct witness *w;
2762 	uint32_t hash;
2763 
2764 	MPASS(key != NULL);
2765 	if (witness_cold == 0)
2766 		mtx_assert(&w_mtx, MA_OWNED);
2767 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2768 	w = w_hash.wh_array[hash];
2769 	while (w != NULL) {
2770 		if (strcmp(w->w_name, key) == 0)
2771 			goto out;
2772 		w = w->w_hash_next;
2773 	}
2774 
2775 out:
2776 	return (w);
2777 }
2778 
2779 static void
2780 witness_hash_put(struct witness *w)
2781 {
2782 	uint32_t hash;
2783 
2784 	MPASS(w != NULL);
2785 	MPASS(w->w_name != NULL);
2786 	if (witness_cold == 0)
2787 		mtx_assert(&w_mtx, MA_OWNED);
2788 	KASSERT(witness_hash_get(w->w_name) == NULL,
2789 	    ("%s: trying to add a hash entry that already exists!", __func__));
2790 	KASSERT(w->w_hash_next == NULL,
2791 	    ("%s: w->w_hash_next != NULL", __func__));
2792 
2793 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2794 	w->w_hash_next = w_hash.wh_array[hash];
2795 	w_hash.wh_array[hash] = w;
2796 	w_hash.wh_count++;
2797 }
2798 
2799 
2800 static struct witness_lock_order_data *
2801 witness_lock_order_get(struct witness *parent, struct witness *child)
2802 {
2803 	struct witness_lock_order_data *data = NULL;
2804 	struct witness_lock_order_key key;
2805 	unsigned int hash;
2806 
2807 	MPASS(parent != NULL && child != NULL);
2808 	key.from = parent->w_index;
2809 	key.to = child->w_index;
2810 	WITNESS_INDEX_ASSERT(key.from);
2811 	WITNESS_INDEX_ASSERT(key.to);
2812 	if ((w_rmatrix[parent->w_index][child->w_index]
2813 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2814 		goto out;
2815 
2816 	hash = witness_hash_djb2((const char*)&key,
2817 	    sizeof(key)) % w_lohash.wloh_size;
2818 	data = w_lohash.wloh_array[hash];
2819 	while (data != NULL) {
2820 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2821 			break;
2822 		data = data->wlod_next;
2823 	}
2824 
2825 out:
2826 	return (data);
2827 }
2828 
2829 /*
2830  * Verify that parent and child have a known relationship, are not the same,
2831  * and child is actually a child of parent.  This is done without w_mtx
2832  * to avoid contention in the common case.
2833  */
2834 static int
2835 witness_lock_order_check(struct witness *parent, struct witness *child)
2836 {
2837 
2838 	if (parent != child &&
2839 	    w_rmatrix[parent->w_index][child->w_index]
2840 	    & WITNESS_LOCK_ORDER_KNOWN &&
2841 	    isitmychild(parent, child))
2842 		return (1);
2843 
2844 	return (0);
2845 }
2846 
2847 static int
2848 witness_lock_order_add(struct witness *parent, struct witness *child)
2849 {
2850 	struct witness_lock_order_data *data = NULL;
2851 	struct witness_lock_order_key key;
2852 	unsigned int hash;
2853 
2854 	MPASS(parent != NULL && child != NULL);
2855 	key.from = parent->w_index;
2856 	key.to = child->w_index;
2857 	WITNESS_INDEX_ASSERT(key.from);
2858 	WITNESS_INDEX_ASSERT(key.to);
2859 	if (w_rmatrix[parent->w_index][child->w_index]
2860 	    & WITNESS_LOCK_ORDER_KNOWN)
2861 		return (1);
2862 
2863 	hash = witness_hash_djb2((const char*)&key,
2864 	    sizeof(key)) % w_lohash.wloh_size;
2865 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2866 	data = w_lofree;
2867 	if (data == NULL)
2868 		return (0);
2869 	w_lofree = data->wlod_next;
2870 	data->wlod_next = w_lohash.wloh_array[hash];
2871 	data->wlod_key = key;
2872 	w_lohash.wloh_array[hash] = data;
2873 	w_lohash.wloh_count++;
2874 	stack_zero(&data->wlod_stack);
2875 	stack_save(&data->wlod_stack);
2876 	return (1);
2877 }
2878 
2879 /* Call this whenver the structure of the witness graph changes. */
2880 static void
2881 witness_increment_graph_generation(void)
2882 {
2883 
2884 	if (witness_cold == 0)
2885 		mtx_assert(&w_mtx, MA_OWNED);
2886 	w_generation++;
2887 }
2888 
2889 #ifdef KDB
2890 static void
2891 _witness_debugger(int cond, const char *msg)
2892 {
2893 
2894 	if (witness_trace && cond)
2895 		kdb_backtrace();
2896 	if (witness_kdb && cond)
2897 		kdb_enter(KDB_WHY_WITNESS, msg);
2898 }
2899 #endif
2900