xref: /freebsd/sys/kern/subr_witness.c (revision b2d48be1bc7df45ddd13b143a160d0acb5a383c5)
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 #ifndef WITNESS_COUNT
136 #define	WITNESS_COUNT 		1536
137 #endif
138 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
139 #define	WITNESS_PENDLIST	(1024 + MAXCPU)
140 
141 /* Allocate 256 KB of stack data space */
142 #define	WITNESS_LO_DATA_COUNT	2048
143 
144 /* Prime, gives load factor of ~2 at full load */
145 #define	WITNESS_LO_HASH_SIZE	1021
146 
147 /*
148  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
149  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
150  * probably be safe for the most part, but it's still a SWAG.
151  */
152 #define	LOCK_NCHILDREN	5
153 #define	LOCK_CHILDCOUNT	2048
154 
155 #define	MAX_W_NAME	64
156 
157 #define	FULLGRAPH_SBUF_SIZE	512
158 
159 /*
160  * These flags go in the witness relationship matrix and describe the
161  * relationship between any two struct witness objects.
162  */
163 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
164 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
165 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
166 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
167 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
168 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define	WITNESS_RELATED_MASK						\
171 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
173 					  * observed. */
174 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
175 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
176 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
177 
178 /* Descendant to ancestor flags */
179 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
180 
181 /* Ancestor to descendant flags */
182 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
183 
184 #define	WITNESS_INDEX_ASSERT(i)						\
185 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
186 
187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
188 
189 /*
190  * Lock instances.  A lock instance is the data associated with a lock while
191  * it is held by witness.  For example, a lock instance will hold the
192  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
193  * are held in a per-cpu list while sleep locks are held in per-thread list.
194  */
195 struct lock_instance {
196 	struct lock_object	*li_lock;
197 	const char		*li_file;
198 	int			li_line;
199 	u_int			li_flags;
200 };
201 
202 /*
203  * A simple list type used to build the list of locks held by a thread
204  * or CPU.  We can't simply embed the list in struct lock_object since a
205  * lock may be held by more than one thread if it is a shared lock.  Locks
206  * are added to the head of the list, so we fill up each list entry from
207  * "the back" logically.  To ease some of the arithmetic, we actually fill
208  * in each list entry the normal way (children[0] then children[1], etc.) but
209  * when we traverse the list we read children[count-1] as the first entry
210  * down to children[0] as the final entry.
211  */
212 struct lock_list_entry {
213 	struct lock_list_entry	*ll_next;
214 	struct lock_instance	ll_children[LOCK_NCHILDREN];
215 	u_int			ll_count;
216 };
217 
218 /*
219  * The main witness structure. One of these per named lock type in the system
220  * (for example, "vnode interlock").
221  */
222 struct witness {
223 	char  			w_name[MAX_W_NAME];
224 	uint32_t 		w_index;  /* Index in the relationship matrix */
225 	struct lock_class	*w_class;
226 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
227 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
228 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
229 	const char		*w_file; /* File where last acquired */
230 	uint32_t 		w_line; /* Line where last acquired */
231 	uint32_t 		w_refcount;
232 	uint16_t 		w_num_ancestors; /* direct/indirect
233 						  * ancestor count */
234 	uint16_t 		w_num_descendants; /* direct/indirect
235 						    * descendant count */
236 	int16_t 		w_ddb_level;
237 	unsigned		w_displayed:1;
238 	unsigned		w_reversed:1;
239 };
240 
241 STAILQ_HEAD(witness_list, witness);
242 
243 /*
244  * The witness hash table. Keys are witness names (const char *), elements are
245  * witness objects (struct witness *).
246  */
247 struct witness_hash {
248 	struct witness	*wh_array[WITNESS_HASH_SIZE];
249 	uint32_t	wh_size;
250 	uint32_t	wh_count;
251 };
252 
253 /*
254  * Key type for the lock order data hash table.
255  */
256 struct witness_lock_order_key {
257 	uint16_t	from;
258 	uint16_t	to;
259 };
260 
261 struct witness_lock_order_data {
262 	struct stack			wlod_stack;
263 	struct witness_lock_order_key	wlod_key;
264 	struct witness_lock_order_data	*wlod_next;
265 };
266 
267 /*
268  * The witness lock order data hash table. Keys are witness index tuples
269  * (struct witness_lock_order_key), elements are lock order data objects
270  * (struct witness_lock_order_data).
271  */
272 struct witness_lock_order_hash {
273 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
274 	u_int	wloh_size;
275 	u_int	wloh_count;
276 };
277 
278 #ifdef BLESSING
279 struct witness_blessed {
280 	const char	*b_lock1;
281 	const char	*b_lock2;
282 };
283 #endif
284 
285 struct witness_pendhelp {
286 	const char		*wh_type;
287 	struct lock_object	*wh_lock;
288 };
289 
290 struct witness_order_list_entry {
291 	const char		*w_name;
292 	struct lock_class	*w_class;
293 };
294 
295 /*
296  * Returns 0 if one of the locks is a spin lock and the other is not.
297  * Returns 1 otherwise.
298  */
299 static __inline int
300 witness_lock_type_equal(struct witness *w1, struct witness *w2)
301 {
302 
303 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
304 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
305 }
306 
307 static __inline int
308 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
309     const struct witness_lock_order_key *b)
310 {
311 
312 	return (a->from == b->from && a->to == b->to);
313 }
314 
315 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
316 		    const char *fname);
317 #ifdef KDB
318 static void	_witness_debugger(int cond, const char *msg);
319 #endif
320 static void	adopt(struct witness *parent, struct witness *child);
321 #ifdef BLESSING
322 static int	blessed(struct witness *, struct witness *);
323 #endif
324 static void	depart(struct witness *w);
325 static struct witness	*enroll(const char *description,
326 			    struct lock_class *lock_class);
327 static struct lock_instance	*find_instance(struct lock_list_entry *list,
328 				    const struct lock_object *lock);
329 static int	isitmychild(struct witness *parent, struct witness *child);
330 static int	isitmydescendant(struct witness *parent, struct witness *child);
331 static void	itismychild(struct witness *parent, struct witness *child);
332 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
333 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
334 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
335 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
336 #ifdef DDB
337 static void	witness_ddb_compute_levels(void);
338 static void	witness_ddb_display(int(*)(const char *fmt, ...));
339 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
340 		    struct witness *, int indent);
341 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
342 		    struct witness_list *list);
343 static void	witness_ddb_level_descendants(struct witness *parent, int l);
344 static void	witness_ddb_list(struct thread *td);
345 #endif
346 static void	witness_free(struct witness *m);
347 static struct witness	*witness_get(void);
348 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
349 static struct witness	*witness_hash_get(const char *key);
350 static void	witness_hash_put(struct witness *w);
351 static void	witness_init_hash_tables(void);
352 static void	witness_increment_graph_generation(void);
353 static void	witness_lock_list_free(struct lock_list_entry *lle);
354 static struct lock_list_entry	*witness_lock_list_get(void);
355 static int	witness_lock_order_add(struct witness *parent,
356 		    struct witness *child);
357 static int	witness_lock_order_check(struct witness *parent,
358 		    struct witness *child);
359 static struct witness_lock_order_data	*witness_lock_order_get(
360 					    struct witness *parent,
361 					    struct witness *child);
362 static void	witness_list_lock(struct lock_instance *instance,
363 		    int (*prnt)(const char *fmt, ...));
364 static void	witness_setflag(struct lock_object *lock, int flag, int set);
365 
366 #ifdef KDB
367 #define	witness_debugger(c)	_witness_debugger(c, __func__)
368 #else
369 #define	witness_debugger(c)
370 #endif
371 
372 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
373     "Witness Locking");
374 
375 /*
376  * If set to 0, lock order checking is disabled.  If set to -1,
377  * witness is completely disabled.  Otherwise witness performs full
378  * lock order checking for all locks.  At runtime, lock order checking
379  * may be toggled.  However, witness cannot be reenabled once it is
380  * completely disabled.
381  */
382 static int witness_watch = 1;
383 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
384     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
385 
386 #ifdef KDB
387 /*
388  * When KDB is enabled and witness_kdb is 1, it will cause the system
389  * to drop into kdebug() when:
390  *	- a lock hierarchy violation occurs
391  *	- locks are held when going to sleep.
392  */
393 #ifdef WITNESS_KDB
394 int	witness_kdb = 1;
395 #else
396 int	witness_kdb = 0;
397 #endif
398 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
399 
400 /*
401  * When KDB is enabled and witness_trace is 1, it will cause the system
402  * to print a stack trace:
403  *	- a lock hierarchy violation occurs
404  *	- locks are held when going to sleep.
405  */
406 int	witness_trace = 1;
407 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
408 #endif /* KDB */
409 
410 #ifdef WITNESS_SKIPSPIN
411 int	witness_skipspin = 1;
412 #else
413 int	witness_skipspin = 0;
414 #endif
415 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
416 
417 int badstack_sbuf_size;
418 
419 int witness_count = WITNESS_COUNT;
420 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
421     &witness_count, 0, "");
422 
423 /*
424  * Call this to print out the relations between locks.
425  */
426 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
427     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
428 
429 /*
430  * Call this to print out the witness faulty stacks.
431  */
432 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
433     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
434 
435 static struct mtx w_mtx;
436 
437 /* w_list */
438 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
439 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
440 
441 /* w_typelist */
442 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
443 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
444 
445 /* lock list */
446 static struct lock_list_entry *w_lock_list_free = NULL;
447 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
448 static u_int pending_cnt;
449 
450 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
451 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
452 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
453 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
454     "");
455 
456 static struct witness *w_data;
457 static uint8_t **w_rmatrix;
458 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
459 static struct witness_hash w_hash;	/* The witness hash table. */
460 
461 /* The lock order data hash */
462 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
463 static struct witness_lock_order_data *w_lofree = NULL;
464 static struct witness_lock_order_hash w_lohash;
465 static int w_max_used_index = 0;
466 static unsigned int w_generation = 0;
467 static const char w_notrunning[] = "Witness not running\n";
468 static const char w_stillcold[] = "Witness is still cold\n";
469 
470 
471 static struct witness_order_list_entry order_lists[] = {
472 	/*
473 	 * sx locks
474 	 */
475 	{ "proctree", &lock_class_sx },
476 	{ "allproc", &lock_class_sx },
477 	{ "allprison", &lock_class_sx },
478 	{ NULL, NULL },
479 	/*
480 	 * Various mutexes
481 	 */
482 	{ "Giant", &lock_class_mtx_sleep },
483 	{ "pipe mutex", &lock_class_mtx_sleep },
484 	{ "sigio lock", &lock_class_mtx_sleep },
485 	{ "process group", &lock_class_mtx_sleep },
486 	{ "process lock", &lock_class_mtx_sleep },
487 	{ "session", &lock_class_mtx_sleep },
488 	{ "uidinfo hash", &lock_class_rw },
489 #ifdef	HWPMC_HOOKS
490 	{ "pmc-sleep", &lock_class_mtx_sleep },
491 #endif
492 	{ "time lock", &lock_class_mtx_sleep },
493 	{ NULL, NULL },
494 	/*
495 	 * umtx
496 	 */
497 	{ "umtx lock", &lock_class_mtx_sleep },
498 	{ NULL, NULL },
499 	/*
500 	 * Sockets
501 	 */
502 	{ "accept", &lock_class_mtx_sleep },
503 	{ "so_snd", &lock_class_mtx_sleep },
504 	{ "so_rcv", &lock_class_mtx_sleep },
505 	{ "sellck", &lock_class_mtx_sleep },
506 	{ NULL, NULL },
507 	/*
508 	 * Routing
509 	 */
510 	{ "so_rcv", &lock_class_mtx_sleep },
511 	{ "radix node head", &lock_class_rw },
512 	{ "rtentry", &lock_class_mtx_sleep },
513 	{ "ifaddr", &lock_class_mtx_sleep },
514 	{ NULL, NULL },
515 	/*
516 	 * IPv4 multicast:
517 	 * protocol locks before interface locks, after UDP locks.
518 	 */
519 	{ "udpinp", &lock_class_rw },
520 	{ "in_multi_mtx", &lock_class_mtx_sleep },
521 	{ "igmp_mtx", &lock_class_mtx_sleep },
522 	{ "if_addr_lock", &lock_class_rw },
523 	{ NULL, NULL },
524 	/*
525 	 * IPv6 multicast:
526 	 * protocol locks before interface locks, after UDP locks.
527 	 */
528 	{ "udpinp", &lock_class_rw },
529 	{ "in6_multi_mtx", &lock_class_mtx_sleep },
530 	{ "mld_mtx", &lock_class_mtx_sleep },
531 	{ "if_addr_lock", &lock_class_rw },
532 	{ NULL, NULL },
533 	/*
534 	 * UNIX Domain Sockets
535 	 */
536 	{ "unp_link_rwlock", &lock_class_rw },
537 	{ "unp_list_lock", &lock_class_mtx_sleep },
538 	{ "unp", &lock_class_mtx_sleep },
539 	{ "so_snd", &lock_class_mtx_sleep },
540 	{ NULL, NULL },
541 	/*
542 	 * UDP/IP
543 	 */
544 	{ "udp", &lock_class_rw },
545 	{ "udpinp", &lock_class_rw },
546 	{ "so_snd", &lock_class_mtx_sleep },
547 	{ NULL, NULL },
548 	/*
549 	 * TCP/IP
550 	 */
551 	{ "tcp", &lock_class_rw },
552 	{ "tcpinp", &lock_class_rw },
553 	{ "so_snd", &lock_class_mtx_sleep },
554 	{ NULL, NULL },
555 	/*
556 	 * BPF
557 	 */
558 	{ "bpf global lock", &lock_class_mtx_sleep },
559 	{ "bpf interface lock", &lock_class_rw },
560 	{ "bpf cdev lock", &lock_class_mtx_sleep },
561 	{ NULL, NULL },
562 	/*
563 	 * NFS server
564 	 */
565 	{ "nfsd_mtx", &lock_class_mtx_sleep },
566 	{ "so_snd", &lock_class_mtx_sleep },
567 	{ NULL, NULL },
568 
569 	/*
570 	 * IEEE 802.11
571 	 */
572 	{ "802.11 com lock", &lock_class_mtx_sleep},
573 	{ NULL, NULL },
574 	/*
575 	 * Network drivers
576 	 */
577 	{ "network driver", &lock_class_mtx_sleep},
578 	{ NULL, NULL },
579 
580 	/*
581 	 * Netgraph
582 	 */
583 	{ "ng_node", &lock_class_mtx_sleep },
584 	{ "ng_worklist", &lock_class_mtx_sleep },
585 	{ NULL, NULL },
586 	/*
587 	 * CDEV
588 	 */
589 	{ "vm map (system)", &lock_class_mtx_sleep },
590 	{ "vm page queue", &lock_class_mtx_sleep },
591 	{ "vnode interlock", &lock_class_mtx_sleep },
592 	{ "cdev", &lock_class_mtx_sleep },
593 	{ NULL, NULL },
594 	/*
595 	 * VM
596 	 */
597 	{ "vm map (user)", &lock_class_sx },
598 	{ "vm object", &lock_class_rw },
599 	{ "vm page", &lock_class_mtx_sleep },
600 	{ "vm page queue", &lock_class_mtx_sleep },
601 	{ "pmap pv global", &lock_class_rw },
602 	{ "pmap", &lock_class_mtx_sleep },
603 	{ "pmap pv list", &lock_class_rw },
604 	{ "vm page free queue", &lock_class_mtx_sleep },
605 	{ NULL, NULL },
606 	/*
607 	 * kqueue/VFS interaction
608 	 */
609 	{ "kqueue", &lock_class_mtx_sleep },
610 	{ "struct mount mtx", &lock_class_mtx_sleep },
611 	{ "vnode interlock", &lock_class_mtx_sleep },
612 	{ NULL, NULL },
613 	/*
614 	 * ZFS locking
615 	 */
616 	{ "dn->dn_mtx", &lock_class_sx },
617 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
618 	{ "db->db_mtx", &lock_class_sx },
619 	{ NULL, NULL },
620 	/*
621 	 * spin locks
622 	 */
623 #ifdef SMP
624 	{ "ap boot", &lock_class_mtx_spin },
625 #endif
626 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
627 	{ "sio", &lock_class_mtx_spin },
628 	{ "scrlock", &lock_class_mtx_spin },
629 #ifdef __i386__
630 	{ "cy", &lock_class_mtx_spin },
631 #endif
632 #ifdef __sparc64__
633 	{ "pcib_mtx", &lock_class_mtx_spin },
634 	{ "rtc_mtx", &lock_class_mtx_spin },
635 #endif
636 	{ "scc_hwmtx", &lock_class_mtx_spin },
637 	{ "uart_hwmtx", &lock_class_mtx_spin },
638 	{ "fast_taskqueue", &lock_class_mtx_spin },
639 	{ "intr table", &lock_class_mtx_spin },
640 #ifdef	HWPMC_HOOKS
641 	{ "pmc-per-proc", &lock_class_mtx_spin },
642 #endif
643 	{ "process slock", &lock_class_mtx_spin },
644 	{ "sleepq chain", &lock_class_mtx_spin },
645 	{ "rm_spinlock", &lock_class_mtx_spin },
646 	{ "turnstile chain", &lock_class_mtx_spin },
647 	{ "turnstile lock", &lock_class_mtx_spin },
648 	{ "sched lock", &lock_class_mtx_spin },
649 	{ "td_contested", &lock_class_mtx_spin },
650 	{ "callout", &lock_class_mtx_spin },
651 	{ "entropy harvest mutex", &lock_class_mtx_spin },
652 	{ "syscons video lock", &lock_class_mtx_spin },
653 #ifdef SMP
654 	{ "smp rendezvous", &lock_class_mtx_spin },
655 #endif
656 #ifdef __powerpc__
657 	{ "tlb0", &lock_class_mtx_spin },
658 #endif
659 	/*
660 	 * leaf locks
661 	 */
662 	{ "intrcnt", &lock_class_mtx_spin },
663 	{ "icu", &lock_class_mtx_spin },
664 #if defined(SMP) && defined(__sparc64__)
665 	{ "ipi", &lock_class_mtx_spin },
666 #endif
667 #ifdef __i386__
668 	{ "allpmaps", &lock_class_mtx_spin },
669 	{ "descriptor tables", &lock_class_mtx_spin },
670 #endif
671 	{ "clk", &lock_class_mtx_spin },
672 	{ "cpuset", &lock_class_mtx_spin },
673 	{ "mprof lock", &lock_class_mtx_spin },
674 	{ "zombie lock", &lock_class_mtx_spin },
675 	{ "ALD Queue", &lock_class_mtx_spin },
676 #if defined(__i386__) || defined(__amd64__)
677 	{ "pcicfg", &lock_class_mtx_spin },
678 	{ "NDIS thread lock", &lock_class_mtx_spin },
679 #endif
680 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
681 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
682 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
683 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
684 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
685 #ifdef	HWPMC_HOOKS
686 	{ "pmc-leaf", &lock_class_mtx_spin },
687 #endif
688 	{ "blocked lock", &lock_class_mtx_spin },
689 	{ NULL, NULL },
690 	{ NULL, NULL }
691 };
692 
693 #ifdef BLESSING
694 /*
695  * Pairs of locks which have been blessed
696  * Don't complain about order problems with blessed locks
697  */
698 static struct witness_blessed blessed_list[] = {
699 };
700 static int blessed_count =
701 	sizeof(blessed_list) / sizeof(struct witness_blessed);
702 #endif
703 
704 /*
705  * This global is set to 0 once it becomes safe to use the witness code.
706  */
707 static int witness_cold = 1;
708 
709 /*
710  * This global is set to 1 once the static lock orders have been enrolled
711  * so that a warning can be issued for any spin locks enrolled later.
712  */
713 static int witness_spin_warn = 0;
714 
715 /* Trim useless garbage from filenames. */
716 static const char *
717 fixup_filename(const char *file)
718 {
719 
720 	if (file == NULL)
721 		return (NULL);
722 	while (strncmp(file, "../", 3) == 0)
723 		file += 3;
724 	return (file);
725 }
726 
727 /*
728  * The WITNESS-enabled diagnostic code.  Note that the witness code does
729  * assume that the early boot is single-threaded at least until after this
730  * routine is completed.
731  */
732 static void
733 witness_initialize(void *dummy __unused)
734 {
735 	struct lock_object *lock;
736 	struct witness_order_list_entry *order;
737 	struct witness *w, *w1;
738 	int i;
739 
740 	w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
741 	    M_WAITOK | M_ZERO);
742 
743 	w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
744 	    M_WITNESS, M_WAITOK | M_ZERO);
745 
746 	for (i = 0; i < witness_count + 1; i++) {
747 		w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
748 		    (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
749 	}
750 	badstack_sbuf_size = witness_count * 256;
751 
752 	/*
753 	 * We have to release Giant before initializing its witness
754 	 * structure so that WITNESS doesn't get confused.
755 	 */
756 	mtx_unlock(&Giant);
757 	mtx_assert(&Giant, MA_NOTOWNED);
758 
759 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
760 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
761 	    MTX_NOWITNESS | MTX_NOPROFILE);
762 	for (i = witness_count - 1; i >= 0; i--) {
763 		w = &w_data[i];
764 		memset(w, 0, sizeof(*w));
765 		w_data[i].w_index = i;	/* Witness index never changes. */
766 		witness_free(w);
767 	}
768 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
769 	    ("%s: Invalid list of free witness objects", __func__));
770 
771 	/* Witness with index 0 is not used to aid in debugging. */
772 	STAILQ_REMOVE_HEAD(&w_free, w_list);
773 	w_free_cnt--;
774 
775 	for (i = 0; i < witness_count; i++) {
776 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
777 		    (witness_count + 1));
778 	}
779 
780 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
781 		witness_lock_list_free(&w_locklistdata[i]);
782 	witness_init_hash_tables();
783 
784 	/* First add in all the specified order lists. */
785 	for (order = order_lists; order->w_name != NULL; order++) {
786 		w = enroll(order->w_name, order->w_class);
787 		if (w == NULL)
788 			continue;
789 		w->w_file = "order list";
790 		for (order++; order->w_name != NULL; order++) {
791 			w1 = enroll(order->w_name, order->w_class);
792 			if (w1 == NULL)
793 				continue;
794 			w1->w_file = "order list";
795 			itismychild(w, w1);
796 			w = w1;
797 		}
798 	}
799 	witness_spin_warn = 1;
800 
801 	/* Iterate through all locks and add them to witness. */
802 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
803 		lock = pending_locks[i].wh_lock;
804 		KASSERT(lock->lo_flags & LO_WITNESS,
805 		    ("%s: lock %s is on pending list but not LO_WITNESS",
806 		    __func__, lock->lo_name));
807 		lock->lo_witness = enroll(pending_locks[i].wh_type,
808 		    LOCK_CLASS(lock));
809 	}
810 
811 	/* Mark the witness code as being ready for use. */
812 	witness_cold = 0;
813 
814 	mtx_lock(&Giant);
815 }
816 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
817     NULL);
818 
819 void
820 witness_init(struct lock_object *lock, const char *type)
821 {
822 	struct lock_class *class;
823 
824 	/* Various sanity checks. */
825 	class = LOCK_CLASS(lock);
826 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
827 	    (class->lc_flags & LC_RECURSABLE) == 0)
828 		kassert_panic("%s: lock (%s) %s can not be recursable",
829 		    __func__, class->lc_name, lock->lo_name);
830 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
831 	    (class->lc_flags & LC_SLEEPABLE) == 0)
832 		kassert_panic("%s: lock (%s) %s can not be sleepable",
833 		    __func__, class->lc_name, lock->lo_name);
834 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
835 	    (class->lc_flags & LC_UPGRADABLE) == 0)
836 		kassert_panic("%s: lock (%s) %s can not be upgradable",
837 		    __func__, class->lc_name, lock->lo_name);
838 
839 	/*
840 	 * If we shouldn't watch this lock, then just clear lo_witness.
841 	 * Otherwise, if witness_cold is set, then it is too early to
842 	 * enroll this lock, so defer it to witness_initialize() by adding
843 	 * it to the pending_locks list.  If it is not too early, then enroll
844 	 * the lock now.
845 	 */
846 	if (witness_watch < 1 || panicstr != NULL ||
847 	    (lock->lo_flags & LO_WITNESS) == 0)
848 		lock->lo_witness = NULL;
849 	else if (witness_cold) {
850 		pending_locks[pending_cnt].wh_lock = lock;
851 		pending_locks[pending_cnt++].wh_type = type;
852 		if (pending_cnt > WITNESS_PENDLIST)
853 			panic("%s: pending locks list is too small, "
854 			    "increase WITNESS_PENDLIST\n",
855 			    __func__);
856 	} else
857 		lock->lo_witness = enroll(type, class);
858 }
859 
860 void
861 witness_destroy(struct lock_object *lock)
862 {
863 	struct lock_class *class;
864 	struct witness *w;
865 
866 	class = LOCK_CLASS(lock);
867 
868 	if (witness_cold)
869 		panic("lock (%s) %s destroyed while witness_cold",
870 		    class->lc_name, lock->lo_name);
871 
872 	/* XXX: need to verify that no one holds the lock */
873 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
874 		return;
875 	w = lock->lo_witness;
876 
877 	mtx_lock_spin(&w_mtx);
878 	MPASS(w->w_refcount > 0);
879 	w->w_refcount--;
880 
881 	if (w->w_refcount == 0)
882 		depart(w);
883 	mtx_unlock_spin(&w_mtx);
884 }
885 
886 #ifdef DDB
887 static void
888 witness_ddb_compute_levels(void)
889 {
890 	struct witness *w;
891 
892 	/*
893 	 * First clear all levels.
894 	 */
895 	STAILQ_FOREACH(w, &w_all, w_list)
896 		w->w_ddb_level = -1;
897 
898 	/*
899 	 * Look for locks with no parents and level all their descendants.
900 	 */
901 	STAILQ_FOREACH(w, &w_all, w_list) {
902 
903 		/* If the witness has ancestors (is not a root), skip it. */
904 		if (w->w_num_ancestors > 0)
905 			continue;
906 		witness_ddb_level_descendants(w, 0);
907 	}
908 }
909 
910 static void
911 witness_ddb_level_descendants(struct witness *w, int l)
912 {
913 	int i;
914 
915 	if (w->w_ddb_level >= l)
916 		return;
917 
918 	w->w_ddb_level = l;
919 	l++;
920 
921 	for (i = 1; i <= w_max_used_index; i++) {
922 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
923 			witness_ddb_level_descendants(&w_data[i], l);
924 	}
925 }
926 
927 static void
928 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
929     struct witness *w, int indent)
930 {
931 	int i;
932 
933  	for (i = 0; i < indent; i++)
934  		prnt(" ");
935 	prnt("%s (type: %s, depth: %d, active refs: %d)",
936 	     w->w_name, w->w_class->lc_name,
937 	     w->w_ddb_level, w->w_refcount);
938  	if (w->w_displayed) {
939  		prnt(" -- (already displayed)\n");
940  		return;
941  	}
942  	w->w_displayed = 1;
943 	if (w->w_file != NULL && w->w_line != 0)
944 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
945 		    w->w_line);
946 	else
947 		prnt(" -- never acquired\n");
948 	indent++;
949 	WITNESS_INDEX_ASSERT(w->w_index);
950 	for (i = 1; i <= w_max_used_index; i++) {
951 		if (db_pager_quit)
952 			return;
953 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
954 			witness_ddb_display_descendants(prnt, &w_data[i],
955 			    indent);
956 	}
957 }
958 
959 static void
960 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
961     struct witness_list *list)
962 {
963 	struct witness *w;
964 
965 	STAILQ_FOREACH(w, list, w_typelist) {
966 		if (w->w_file == NULL || w->w_ddb_level > 0)
967 			continue;
968 
969 		/* This lock has no anscestors - display its descendants. */
970 		witness_ddb_display_descendants(prnt, w, 0);
971 		if (db_pager_quit)
972 			return;
973 	}
974 }
975 
976 static void
977 witness_ddb_display(int(*prnt)(const char *fmt, ...))
978 {
979 	struct witness *w;
980 
981 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
982 	witness_ddb_compute_levels();
983 
984 	/* Clear all the displayed flags. */
985 	STAILQ_FOREACH(w, &w_all, w_list)
986 		w->w_displayed = 0;
987 
988 	/*
989 	 * First, handle sleep locks which have been acquired at least
990 	 * once.
991 	 */
992 	prnt("Sleep locks:\n");
993 	witness_ddb_display_list(prnt, &w_sleep);
994 	if (db_pager_quit)
995 		return;
996 
997 	/*
998 	 * Now do spin locks which have been acquired at least once.
999 	 */
1000 	prnt("\nSpin locks:\n");
1001 	witness_ddb_display_list(prnt, &w_spin);
1002 	if (db_pager_quit)
1003 		return;
1004 
1005 	/*
1006 	 * Finally, any locks which have not been acquired yet.
1007 	 */
1008 	prnt("\nLocks which were never acquired:\n");
1009 	STAILQ_FOREACH(w, &w_all, w_list) {
1010 		if (w->w_file != NULL || w->w_refcount == 0)
1011 			continue;
1012 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1013 		    w->w_class->lc_name, w->w_ddb_level);
1014 		if (db_pager_quit)
1015 			return;
1016 	}
1017 }
1018 #endif /* DDB */
1019 
1020 int
1021 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1022 {
1023 
1024 	if (witness_watch == -1 || panicstr != NULL)
1025 		return (0);
1026 
1027 	/* Require locks that witness knows about. */
1028 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1029 	    lock2->lo_witness == NULL)
1030 		return (EINVAL);
1031 
1032 	mtx_assert(&w_mtx, MA_NOTOWNED);
1033 	mtx_lock_spin(&w_mtx);
1034 
1035 	/*
1036 	 * If we already have either an explicit or implied lock order that
1037 	 * is the other way around, then return an error.
1038 	 */
1039 	if (witness_watch &&
1040 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1041 		mtx_unlock_spin(&w_mtx);
1042 		return (EDOOFUS);
1043 	}
1044 
1045 	/* Try to add the new order. */
1046 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1047 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1048 	itismychild(lock1->lo_witness, lock2->lo_witness);
1049 	mtx_unlock_spin(&w_mtx);
1050 	return (0);
1051 }
1052 
1053 void
1054 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1055     int line, struct lock_object *interlock)
1056 {
1057 	struct lock_list_entry *lock_list, *lle;
1058 	struct lock_instance *lock1, *lock2, *plock;
1059 	struct lock_class *class, *iclass;
1060 	struct witness *w, *w1;
1061 	struct thread *td;
1062 	int i, j;
1063 
1064 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1065 	    panicstr != NULL)
1066 		return;
1067 
1068 	w = lock->lo_witness;
1069 	class = LOCK_CLASS(lock);
1070 	td = curthread;
1071 
1072 	if (class->lc_flags & LC_SLEEPLOCK) {
1073 
1074 		/*
1075 		 * Since spin locks include a critical section, this check
1076 		 * implicitly enforces a lock order of all sleep locks before
1077 		 * all spin locks.
1078 		 */
1079 		if (td->td_critnest != 0 && !kdb_active)
1080 			kassert_panic("acquiring blockable sleep lock with "
1081 			    "spinlock or critical section held (%s) %s @ %s:%d",
1082 			    class->lc_name, lock->lo_name,
1083 			    fixup_filename(file), line);
1084 
1085 		/*
1086 		 * If this is the first lock acquired then just return as
1087 		 * no order checking is needed.
1088 		 */
1089 		lock_list = td->td_sleeplocks;
1090 		if (lock_list == NULL || lock_list->ll_count == 0)
1091 			return;
1092 	} else {
1093 
1094 		/*
1095 		 * If this is the first lock, just return as no order
1096 		 * checking is needed.  Avoid problems with thread
1097 		 * migration pinning the thread while checking if
1098 		 * spinlocks are held.  If at least one spinlock is held
1099 		 * the thread is in a safe path and it is allowed to
1100 		 * unpin it.
1101 		 */
1102 		sched_pin();
1103 		lock_list = PCPU_GET(spinlocks);
1104 		if (lock_list == NULL || lock_list->ll_count == 0) {
1105 			sched_unpin();
1106 			return;
1107 		}
1108 		sched_unpin();
1109 	}
1110 
1111 	/*
1112 	 * Check to see if we are recursing on a lock we already own.  If
1113 	 * so, make sure that we don't mismatch exclusive and shared lock
1114 	 * acquires.
1115 	 */
1116 	lock1 = find_instance(lock_list, lock);
1117 	if (lock1 != NULL) {
1118 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1119 		    (flags & LOP_EXCLUSIVE) == 0) {
1120 			printf("shared lock of (%s) %s @ %s:%d\n",
1121 			    class->lc_name, lock->lo_name,
1122 			    fixup_filename(file), line);
1123 			printf("while exclusively locked from %s:%d\n",
1124 			    fixup_filename(lock1->li_file), lock1->li_line);
1125 			kassert_panic("excl->share");
1126 		}
1127 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1128 		    (flags & LOP_EXCLUSIVE) != 0) {
1129 			printf("exclusive lock of (%s) %s @ %s:%d\n",
1130 			    class->lc_name, lock->lo_name,
1131 			    fixup_filename(file), line);
1132 			printf("while share locked from %s:%d\n",
1133 			    fixup_filename(lock1->li_file), lock1->li_line);
1134 			kassert_panic("share->excl");
1135 		}
1136 		return;
1137 	}
1138 
1139 	/* Warn if the interlock is not locked exactly once. */
1140 	if (interlock != NULL) {
1141 		iclass = LOCK_CLASS(interlock);
1142 		lock1 = find_instance(lock_list, interlock);
1143 		if (lock1 == NULL)
1144 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1145 			    iclass->lc_name, interlock->lo_name,
1146 			    fixup_filename(file), line);
1147 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1148 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1149 			    iclass->lc_name, interlock->lo_name,
1150 			    fixup_filename(file), line);
1151 	}
1152 
1153 	/*
1154 	 * Find the previously acquired lock, but ignore interlocks.
1155 	 */
1156 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1157 	if (interlock != NULL && plock->li_lock == interlock) {
1158 		if (lock_list->ll_count > 1)
1159 			plock =
1160 			    &lock_list->ll_children[lock_list->ll_count - 2];
1161 		else {
1162 			lle = lock_list->ll_next;
1163 
1164 			/*
1165 			 * The interlock is the only lock we hold, so
1166 			 * simply return.
1167 			 */
1168 			if (lle == NULL)
1169 				return;
1170 			plock = &lle->ll_children[lle->ll_count - 1];
1171 		}
1172 	}
1173 
1174 	/*
1175 	 * Try to perform most checks without a lock.  If this succeeds we
1176 	 * can skip acquiring the lock and return success.  Otherwise we redo
1177 	 * the check with the lock held to handle races with concurrent updates.
1178 	 */
1179 	w1 = plock->li_lock->lo_witness;
1180 	if (witness_lock_order_check(w1, w))
1181 		return;
1182 
1183 	mtx_lock_spin(&w_mtx);
1184 	if (witness_lock_order_check(w1, w)) {
1185 		mtx_unlock_spin(&w_mtx);
1186 		return;
1187 	}
1188 	witness_lock_order_add(w1, w);
1189 
1190 	/*
1191 	 * Check for duplicate locks of the same type.  Note that we only
1192 	 * have to check for this on the last lock we just acquired.  Any
1193 	 * other cases will be caught as lock order violations.
1194 	 */
1195 	if (w1 == w) {
1196 		i = w->w_index;
1197 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1198 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1199 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1200 			w->w_reversed = 1;
1201 			mtx_unlock_spin(&w_mtx);
1202 			printf(
1203 			    "acquiring duplicate lock of same type: \"%s\"\n",
1204 			    w->w_name);
1205 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1206 			    fixup_filename(plock->li_file), plock->li_line);
1207 			printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1208 			    fixup_filename(file), line);
1209 			witness_debugger(1);
1210 		} else
1211 			mtx_unlock_spin(&w_mtx);
1212 		return;
1213 	}
1214 	mtx_assert(&w_mtx, MA_OWNED);
1215 
1216 	/*
1217 	 * If we know that the lock we are acquiring comes after
1218 	 * the lock we most recently acquired in the lock order tree,
1219 	 * then there is no need for any further checks.
1220 	 */
1221 	if (isitmychild(w1, w))
1222 		goto out;
1223 
1224 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1225 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1226 
1227 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1228 			lock1 = &lle->ll_children[i];
1229 
1230 			/*
1231 			 * Ignore the interlock.
1232 			 */
1233 			if (interlock == lock1->li_lock)
1234 				continue;
1235 
1236 			/*
1237 			 * If this lock doesn't undergo witness checking,
1238 			 * then skip it.
1239 			 */
1240 			w1 = lock1->li_lock->lo_witness;
1241 			if (w1 == NULL) {
1242 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1243 				    ("lock missing witness structure"));
1244 				continue;
1245 			}
1246 
1247 			/*
1248 			 * If we are locking Giant and this is a sleepable
1249 			 * lock, then skip it.
1250 			 */
1251 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1252 			    lock == &Giant.lock_object)
1253 				continue;
1254 
1255 			/*
1256 			 * If we are locking a sleepable lock and this lock
1257 			 * is Giant, then skip it.
1258 			 */
1259 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1260 			    lock1->li_lock == &Giant.lock_object)
1261 				continue;
1262 
1263 			/*
1264 			 * If we are locking a sleepable lock and this lock
1265 			 * isn't sleepable, we want to treat it as a lock
1266 			 * order violation to enfore a general lock order of
1267 			 * sleepable locks before non-sleepable locks.
1268 			 */
1269 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1270 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1271 				goto reversal;
1272 
1273 			/*
1274 			 * If we are locking Giant and this is a non-sleepable
1275 			 * lock, then treat it as a reversal.
1276 			 */
1277 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1278 			    lock == &Giant.lock_object)
1279 				goto reversal;
1280 
1281 			/*
1282 			 * Check the lock order hierarchy for a reveresal.
1283 			 */
1284 			if (!isitmydescendant(w, w1))
1285 				continue;
1286 		reversal:
1287 
1288 			/*
1289 			 * We have a lock order violation, check to see if it
1290 			 * is allowed or has already been yelled about.
1291 			 */
1292 #ifdef BLESSING
1293 
1294 			/*
1295 			 * If the lock order is blessed, just bail.  We don't
1296 			 * look for other lock order violations though, which
1297 			 * may be a bug.
1298 			 */
1299 			if (blessed(w, w1))
1300 				goto out;
1301 #endif
1302 
1303 			/* Bail if this violation is known */
1304 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1305 				goto out;
1306 
1307 			/* Record this as a violation */
1308 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1309 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1310 			w->w_reversed = w1->w_reversed = 1;
1311 			witness_increment_graph_generation();
1312 			mtx_unlock_spin(&w_mtx);
1313 
1314 #ifdef WITNESS_NO_VNODE
1315 			/*
1316 			 * There are known LORs between VNODE locks. They are
1317 			 * not an indication of a bug. VNODE locks are flagged
1318 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1319 			 * is between 2 VNODE locks.
1320 			 */
1321 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1322 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1323 				return;
1324 #endif
1325 
1326 			/*
1327 			 * Ok, yell about it.
1328 			 */
1329 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1330 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1331 				printf(
1332 		"lock order reversal: (sleepable after non-sleepable)\n");
1333 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1334 			    && lock == &Giant.lock_object)
1335 				printf(
1336 		"lock order reversal: (Giant after non-sleepable)\n");
1337 			else
1338 				printf("lock order reversal:\n");
1339 
1340 			/*
1341 			 * Try to locate an earlier lock with
1342 			 * witness w in our list.
1343 			 */
1344 			do {
1345 				lock2 = &lle->ll_children[i];
1346 				MPASS(lock2->li_lock != NULL);
1347 				if (lock2->li_lock->lo_witness == w)
1348 					break;
1349 				if (i == 0 && lle->ll_next != NULL) {
1350 					lle = lle->ll_next;
1351 					i = lle->ll_count - 1;
1352 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1353 				} else
1354 					i--;
1355 			} while (i >= 0);
1356 			if (i < 0) {
1357 				printf(" 1st %p %s (%s) @ %s:%d\n",
1358 				    lock1->li_lock, lock1->li_lock->lo_name,
1359 				    w1->w_name, fixup_filename(lock1->li_file),
1360 				    lock1->li_line);
1361 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1362 				    lock->lo_name, w->w_name,
1363 				    fixup_filename(file), line);
1364 			} else {
1365 				printf(" 1st %p %s (%s) @ %s:%d\n",
1366 				    lock2->li_lock, lock2->li_lock->lo_name,
1367 				    lock2->li_lock->lo_witness->w_name,
1368 				    fixup_filename(lock2->li_file),
1369 				    lock2->li_line);
1370 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1371 				    lock1->li_lock, lock1->li_lock->lo_name,
1372 				    w1->w_name, fixup_filename(lock1->li_file),
1373 				    lock1->li_line);
1374 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1375 				    lock->lo_name, w->w_name,
1376 				    fixup_filename(file), line);
1377 			}
1378 			witness_debugger(1);
1379 			return;
1380 		}
1381 	}
1382 
1383 	/*
1384 	 * If requested, build a new lock order.  However, don't build a new
1385 	 * relationship between a sleepable lock and Giant if it is in the
1386 	 * wrong direction.  The correct lock order is that sleepable locks
1387 	 * always come before Giant.
1388 	 */
1389 	if (flags & LOP_NEWORDER &&
1390 	    !(plock->li_lock == &Giant.lock_object &&
1391 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1392 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1393 		    w->w_name, plock->li_lock->lo_witness->w_name);
1394 		itismychild(plock->li_lock->lo_witness, w);
1395 	}
1396 out:
1397 	mtx_unlock_spin(&w_mtx);
1398 }
1399 
1400 void
1401 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1402 {
1403 	struct lock_list_entry **lock_list, *lle;
1404 	struct lock_instance *instance;
1405 	struct witness *w;
1406 	struct thread *td;
1407 
1408 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1409 	    panicstr != NULL)
1410 		return;
1411 	w = lock->lo_witness;
1412 	td = curthread;
1413 
1414 	/* Determine lock list for this lock. */
1415 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1416 		lock_list = &td->td_sleeplocks;
1417 	else
1418 		lock_list = PCPU_PTR(spinlocks);
1419 
1420 	/* Check to see if we are recursing on a lock we already own. */
1421 	instance = find_instance(*lock_list, lock);
1422 	if (instance != NULL) {
1423 		instance->li_flags++;
1424 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1425 		    td->td_proc->p_pid, lock->lo_name,
1426 		    instance->li_flags & LI_RECURSEMASK);
1427 		instance->li_file = file;
1428 		instance->li_line = line;
1429 		return;
1430 	}
1431 
1432 	/* Update per-witness last file and line acquire. */
1433 	w->w_file = file;
1434 	w->w_line = line;
1435 
1436 	/* Find the next open lock instance in the list and fill it. */
1437 	lle = *lock_list;
1438 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1439 		lle = witness_lock_list_get();
1440 		if (lle == NULL)
1441 			return;
1442 		lle->ll_next = *lock_list;
1443 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1444 		    td->td_proc->p_pid, lle);
1445 		*lock_list = lle;
1446 	}
1447 	instance = &lle->ll_children[lle->ll_count++];
1448 	instance->li_lock = lock;
1449 	instance->li_line = line;
1450 	instance->li_file = file;
1451 	if ((flags & LOP_EXCLUSIVE) != 0)
1452 		instance->li_flags = LI_EXCLUSIVE;
1453 	else
1454 		instance->li_flags = 0;
1455 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1456 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1457 }
1458 
1459 void
1460 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1461 {
1462 	struct lock_instance *instance;
1463 	struct lock_class *class;
1464 
1465 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1466 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1467 		return;
1468 	class = LOCK_CLASS(lock);
1469 	if (witness_watch) {
1470 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1471 			kassert_panic(
1472 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1473 			    class->lc_name, lock->lo_name,
1474 			    fixup_filename(file), line);
1475 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1476 			kassert_panic(
1477 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1478 			    class->lc_name, lock->lo_name,
1479 			    fixup_filename(file), line);
1480 	}
1481 	instance = find_instance(curthread->td_sleeplocks, lock);
1482 	if (instance == NULL) {
1483 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1484 		    class->lc_name, lock->lo_name,
1485 		    fixup_filename(file), line);
1486 		return;
1487 	}
1488 	if (witness_watch) {
1489 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1490 			kassert_panic(
1491 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1492 			    class->lc_name, lock->lo_name,
1493 			    fixup_filename(file), line);
1494 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1495 			kassert_panic(
1496 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1497 			    class->lc_name, lock->lo_name,
1498 			    instance->li_flags & LI_RECURSEMASK,
1499 			    fixup_filename(file), line);
1500 	}
1501 	instance->li_flags |= LI_EXCLUSIVE;
1502 }
1503 
1504 void
1505 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1506     int line)
1507 {
1508 	struct lock_instance *instance;
1509 	struct lock_class *class;
1510 
1511 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1512 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1513 		return;
1514 	class = LOCK_CLASS(lock);
1515 	if (witness_watch) {
1516 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1517 			kassert_panic(
1518 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1519 			    class->lc_name, lock->lo_name,
1520 			    fixup_filename(file), line);
1521 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1522 			kassert_panic(
1523 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1524 			    class->lc_name, lock->lo_name,
1525 			    fixup_filename(file), line);
1526 	}
1527 	instance = find_instance(curthread->td_sleeplocks, lock);
1528 	if (instance == NULL) {
1529 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1530 		    class->lc_name, lock->lo_name,
1531 		    fixup_filename(file), line);
1532 		return;
1533 	}
1534 	if (witness_watch) {
1535 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1536 			kassert_panic(
1537 			    "downgrade of shared lock (%s) %s @ %s:%d",
1538 			    class->lc_name, lock->lo_name,
1539 			    fixup_filename(file), line);
1540 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1541 			kassert_panic(
1542 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1543 			    class->lc_name, lock->lo_name,
1544 			    instance->li_flags & LI_RECURSEMASK,
1545 			    fixup_filename(file), line);
1546 	}
1547 	instance->li_flags &= ~LI_EXCLUSIVE;
1548 }
1549 
1550 void
1551 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1552 {
1553 	struct lock_list_entry **lock_list, *lle;
1554 	struct lock_instance *instance;
1555 	struct lock_class *class;
1556 	struct thread *td;
1557 	register_t s;
1558 	int i, j;
1559 
1560 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1561 		return;
1562 	td = curthread;
1563 	class = LOCK_CLASS(lock);
1564 
1565 	/* Find lock instance associated with this lock. */
1566 	if (class->lc_flags & LC_SLEEPLOCK)
1567 		lock_list = &td->td_sleeplocks;
1568 	else
1569 		lock_list = PCPU_PTR(spinlocks);
1570 	lle = *lock_list;
1571 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1572 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1573 			instance = &(*lock_list)->ll_children[i];
1574 			if (instance->li_lock == lock)
1575 				goto found;
1576 		}
1577 
1578 	/*
1579 	 * When disabling WITNESS through witness_watch we could end up in
1580 	 * having registered locks in the td_sleeplocks queue.
1581 	 * We have to make sure we flush these queues, so just search for
1582 	 * eventual register locks and remove them.
1583 	 */
1584 	if (witness_watch > 0) {
1585 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1586 		    lock->lo_name, fixup_filename(file), line);
1587 		return;
1588 	} else {
1589 		return;
1590 	}
1591 found:
1592 
1593 	/* First, check for shared/exclusive mismatches. */
1594 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1595 	    (flags & LOP_EXCLUSIVE) == 0) {
1596 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1597 		    lock->lo_name, fixup_filename(file), line);
1598 		printf("while exclusively locked from %s:%d\n",
1599 		    fixup_filename(instance->li_file), instance->li_line);
1600 		kassert_panic("excl->ushare");
1601 	}
1602 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1603 	    (flags & LOP_EXCLUSIVE) != 0) {
1604 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1605 		    lock->lo_name, fixup_filename(file), line);
1606 		printf("while share locked from %s:%d\n",
1607 		    fixup_filename(instance->li_file),
1608 		    instance->li_line);
1609 		kassert_panic("share->uexcl");
1610 	}
1611 	/* If we are recursed, unrecurse. */
1612 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1613 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1614 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1615 		    instance->li_flags);
1616 		instance->li_flags--;
1617 		return;
1618 	}
1619 	/* The lock is now being dropped, check for NORELEASE flag */
1620 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1621 		printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1622 		    lock->lo_name, fixup_filename(file), line);
1623 		kassert_panic("lock marked norelease");
1624 	}
1625 
1626 	/* Otherwise, remove this item from the list. */
1627 	s = intr_disable();
1628 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1629 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1630 	    (*lock_list)->ll_count - 1);
1631 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1632 		(*lock_list)->ll_children[j] =
1633 		    (*lock_list)->ll_children[j + 1];
1634 	(*lock_list)->ll_count--;
1635 	intr_restore(s);
1636 
1637 	/*
1638 	 * In order to reduce contention on w_mtx, we want to keep always an
1639 	 * head object into lists so that frequent allocation from the
1640 	 * free witness pool (and subsequent locking) is avoided.
1641 	 * In order to maintain the current code simple, when the head
1642 	 * object is totally unloaded it means also that we do not have
1643 	 * further objects in the list, so the list ownership needs to be
1644 	 * hand over to another object if the current head needs to be freed.
1645 	 */
1646 	if ((*lock_list)->ll_count == 0) {
1647 		if (*lock_list == lle) {
1648 			if (lle->ll_next == NULL)
1649 				return;
1650 		} else
1651 			lle = *lock_list;
1652 		*lock_list = lle->ll_next;
1653 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1654 		    td->td_proc->p_pid, lle);
1655 		witness_lock_list_free(lle);
1656 	}
1657 }
1658 
1659 void
1660 witness_thread_exit(struct thread *td)
1661 {
1662 	struct lock_list_entry *lle;
1663 	int i, n;
1664 
1665 	lle = td->td_sleeplocks;
1666 	if (lle == NULL || panicstr != NULL)
1667 		return;
1668 	if (lle->ll_count != 0) {
1669 		for (n = 0; lle != NULL; lle = lle->ll_next)
1670 			for (i = lle->ll_count - 1; i >= 0; i--) {
1671 				if (n == 0)
1672 		printf("Thread %p exiting with the following locks held:\n",
1673 					    td);
1674 				n++;
1675 				witness_list_lock(&lle->ll_children[i], printf);
1676 
1677 			}
1678 		kassert_panic(
1679 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1680 	}
1681 	witness_lock_list_free(lle);
1682 }
1683 
1684 /*
1685  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1686  * exempt Giant and sleepable locks from the checks as well.  If any
1687  * non-exempt locks are held, then a supplied message is printed to the
1688  * console along with a list of the offending locks.  If indicated in the
1689  * flags then a failure results in a panic as well.
1690  */
1691 int
1692 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1693 {
1694 	struct lock_list_entry *lock_list, *lle;
1695 	struct lock_instance *lock1;
1696 	struct thread *td;
1697 	va_list ap;
1698 	int i, n;
1699 
1700 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1701 		return (0);
1702 	n = 0;
1703 	td = curthread;
1704 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1705 		for (i = lle->ll_count - 1; i >= 0; i--) {
1706 			lock1 = &lle->ll_children[i];
1707 			if (lock1->li_lock == lock)
1708 				continue;
1709 			if (flags & WARN_GIANTOK &&
1710 			    lock1->li_lock == &Giant.lock_object)
1711 				continue;
1712 			if (flags & WARN_SLEEPOK &&
1713 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1714 				continue;
1715 			if (n == 0) {
1716 				va_start(ap, fmt);
1717 				vprintf(fmt, ap);
1718 				va_end(ap);
1719 				printf(" with the following");
1720 				if (flags & WARN_SLEEPOK)
1721 					printf(" non-sleepable");
1722 				printf(" locks held:\n");
1723 			}
1724 			n++;
1725 			witness_list_lock(lock1, printf);
1726 		}
1727 
1728 	/*
1729 	 * Pin the thread in order to avoid problems with thread migration.
1730 	 * Once that all verifies are passed about spinlocks ownership,
1731 	 * the thread is in a safe path and it can be unpinned.
1732 	 */
1733 	sched_pin();
1734 	lock_list = PCPU_GET(spinlocks);
1735 	if (lock_list != NULL && lock_list->ll_count != 0) {
1736 		sched_unpin();
1737 
1738 		/*
1739 		 * We should only have one spinlock and as long as
1740 		 * the flags cannot match for this locks class,
1741 		 * check if the first spinlock is the one curthread
1742 		 * should hold.
1743 		 */
1744 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1745 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1746 		    lock1->li_lock == lock && n == 0)
1747 			return (0);
1748 
1749 		va_start(ap, fmt);
1750 		vprintf(fmt, ap);
1751 		va_end(ap);
1752 		printf(" with the following");
1753 		if (flags & WARN_SLEEPOK)
1754 			printf(" non-sleepable");
1755 		printf(" locks held:\n");
1756 		n += witness_list_locks(&lock_list, printf);
1757 	} else
1758 		sched_unpin();
1759 	if (flags & WARN_PANIC && n)
1760 		kassert_panic("%s", __func__);
1761 	else
1762 		witness_debugger(n);
1763 	return (n);
1764 }
1765 
1766 const char *
1767 witness_file(struct lock_object *lock)
1768 {
1769 	struct witness *w;
1770 
1771 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1772 		return ("?");
1773 	w = lock->lo_witness;
1774 	return (w->w_file);
1775 }
1776 
1777 int
1778 witness_line(struct lock_object *lock)
1779 {
1780 	struct witness *w;
1781 
1782 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1783 		return (0);
1784 	w = lock->lo_witness;
1785 	return (w->w_line);
1786 }
1787 
1788 static struct witness *
1789 enroll(const char *description, struct lock_class *lock_class)
1790 {
1791 	struct witness *w;
1792 	struct witness_list *typelist;
1793 
1794 	MPASS(description != NULL);
1795 
1796 	if (witness_watch == -1 || panicstr != NULL)
1797 		return (NULL);
1798 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1799 		if (witness_skipspin)
1800 			return (NULL);
1801 		else
1802 			typelist = &w_spin;
1803 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1804 		typelist = &w_sleep;
1805 	} else {
1806 		kassert_panic("lock class %s is not sleep or spin",
1807 		    lock_class->lc_name);
1808 		return (NULL);
1809 	}
1810 
1811 	mtx_lock_spin(&w_mtx);
1812 	w = witness_hash_get(description);
1813 	if (w)
1814 		goto found;
1815 	if ((w = witness_get()) == NULL)
1816 		return (NULL);
1817 	MPASS(strlen(description) < MAX_W_NAME);
1818 	strcpy(w->w_name, description);
1819 	w->w_class = lock_class;
1820 	w->w_refcount = 1;
1821 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1822 	if (lock_class->lc_flags & LC_SPINLOCK) {
1823 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1824 		w_spin_cnt++;
1825 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1826 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1827 		w_sleep_cnt++;
1828 	}
1829 
1830 	/* Insert new witness into the hash */
1831 	witness_hash_put(w);
1832 	witness_increment_graph_generation();
1833 	mtx_unlock_spin(&w_mtx);
1834 	return (w);
1835 found:
1836 	w->w_refcount++;
1837 	mtx_unlock_spin(&w_mtx);
1838 	if (lock_class != w->w_class)
1839 		kassert_panic(
1840 			"lock (%s) %s does not match earlier (%s) lock",
1841 			description, lock_class->lc_name,
1842 			w->w_class->lc_name);
1843 	return (w);
1844 }
1845 
1846 static void
1847 depart(struct witness *w)
1848 {
1849 	struct witness_list *list;
1850 
1851 	MPASS(w->w_refcount == 0);
1852 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1853 		list = &w_sleep;
1854 		w_sleep_cnt--;
1855 	} else {
1856 		list = &w_spin;
1857 		w_spin_cnt--;
1858 	}
1859 	/*
1860 	 * Set file to NULL as it may point into a loadable module.
1861 	 */
1862 	w->w_file = NULL;
1863 	w->w_line = 0;
1864 	witness_increment_graph_generation();
1865 }
1866 
1867 
1868 static void
1869 adopt(struct witness *parent, struct witness *child)
1870 {
1871 	int pi, ci, i, j;
1872 
1873 	if (witness_cold == 0)
1874 		mtx_assert(&w_mtx, MA_OWNED);
1875 
1876 	/* If the relationship is already known, there's no work to be done. */
1877 	if (isitmychild(parent, child))
1878 		return;
1879 
1880 	/* When the structure of the graph changes, bump up the generation. */
1881 	witness_increment_graph_generation();
1882 
1883 	/*
1884 	 * The hard part ... create the direct relationship, then propagate all
1885 	 * indirect relationships.
1886 	 */
1887 	pi = parent->w_index;
1888 	ci = child->w_index;
1889 	WITNESS_INDEX_ASSERT(pi);
1890 	WITNESS_INDEX_ASSERT(ci);
1891 	MPASS(pi != ci);
1892 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1893 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1894 
1895 	/*
1896 	 * If parent was not already an ancestor of child,
1897 	 * then we increment the descendant and ancestor counters.
1898 	 */
1899 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1900 		parent->w_num_descendants++;
1901 		child->w_num_ancestors++;
1902 	}
1903 
1904 	/*
1905 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1906 	 * an ancestor of 'pi' during this loop.
1907 	 */
1908 	for (i = 1; i <= w_max_used_index; i++) {
1909 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1910 		    (i != pi))
1911 			continue;
1912 
1913 		/* Find each descendant of 'i' and mark it as a descendant. */
1914 		for (j = 1; j <= w_max_used_index; j++) {
1915 
1916 			/*
1917 			 * Skip children that are already marked as
1918 			 * descendants of 'i'.
1919 			 */
1920 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1921 				continue;
1922 
1923 			/*
1924 			 * We are only interested in descendants of 'ci'. Note
1925 			 * that 'ci' itself is counted as a descendant of 'ci'.
1926 			 */
1927 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1928 			    (j != ci))
1929 				continue;
1930 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1931 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1932 			w_data[i].w_num_descendants++;
1933 			w_data[j].w_num_ancestors++;
1934 
1935 			/*
1936 			 * Make sure we aren't marking a node as both an
1937 			 * ancestor and descendant. We should have caught
1938 			 * this as a lock order reversal earlier.
1939 			 */
1940 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1941 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1942 				printf("witness rmatrix paradox! [%d][%d]=%d "
1943 				    "both ancestor and descendant\n",
1944 				    i, j, w_rmatrix[i][j]);
1945 				kdb_backtrace();
1946 				printf("Witness disabled.\n");
1947 				witness_watch = -1;
1948 			}
1949 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1950 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1951 				printf("witness rmatrix paradox! [%d][%d]=%d "
1952 				    "both ancestor and descendant\n",
1953 				    j, i, w_rmatrix[j][i]);
1954 				kdb_backtrace();
1955 				printf("Witness disabled.\n");
1956 				witness_watch = -1;
1957 			}
1958 		}
1959 	}
1960 }
1961 
1962 static void
1963 itismychild(struct witness *parent, struct witness *child)
1964 {
1965 	int unlocked;
1966 
1967 	MPASS(child != NULL && parent != NULL);
1968 	if (witness_cold == 0)
1969 		mtx_assert(&w_mtx, MA_OWNED);
1970 
1971 	if (!witness_lock_type_equal(parent, child)) {
1972 		if (witness_cold == 0) {
1973 			unlocked = 1;
1974 			mtx_unlock_spin(&w_mtx);
1975 		} else {
1976 			unlocked = 0;
1977 		}
1978 		kassert_panic(
1979 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1980 		    "the same lock type", __func__, parent->w_name,
1981 		    parent->w_class->lc_name, child->w_name,
1982 		    child->w_class->lc_name);
1983 		if (unlocked)
1984 			mtx_lock_spin(&w_mtx);
1985 	}
1986 	adopt(parent, child);
1987 }
1988 
1989 /*
1990  * Generic code for the isitmy*() functions. The rmask parameter is the
1991  * expected relationship of w1 to w2.
1992  */
1993 static int
1994 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1995 {
1996 	unsigned char r1, r2;
1997 	int i1, i2;
1998 
1999 	i1 = w1->w_index;
2000 	i2 = w2->w_index;
2001 	WITNESS_INDEX_ASSERT(i1);
2002 	WITNESS_INDEX_ASSERT(i2);
2003 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2004 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2005 
2006 	/* The flags on one better be the inverse of the flags on the other */
2007 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2008 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2009 		/* Don't squawk if we're potentially racing with an update. */
2010 		if (!mtx_owned(&w_mtx))
2011 			return (0);
2012 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2013 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2014 		    "w_rmatrix[%d][%d] == %hhx\n",
2015 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2016 		    i2, i1, r2);
2017 		kdb_backtrace();
2018 		printf("Witness disabled.\n");
2019 		witness_watch = -1;
2020 	}
2021 	return (r1 & rmask);
2022 }
2023 
2024 /*
2025  * Checks if @child is a direct child of @parent.
2026  */
2027 static int
2028 isitmychild(struct witness *parent, struct witness *child)
2029 {
2030 
2031 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2032 }
2033 
2034 /*
2035  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2036  */
2037 static int
2038 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2039 {
2040 
2041 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2042 	    __func__));
2043 }
2044 
2045 #ifdef BLESSING
2046 static int
2047 blessed(struct witness *w1, struct witness *w2)
2048 {
2049 	int i;
2050 	struct witness_blessed *b;
2051 
2052 	for (i = 0; i < blessed_count; i++) {
2053 		b = &blessed_list[i];
2054 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2055 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2056 				return (1);
2057 			continue;
2058 		}
2059 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2060 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2061 				return (1);
2062 	}
2063 	return (0);
2064 }
2065 #endif
2066 
2067 static struct witness *
2068 witness_get(void)
2069 {
2070 	struct witness *w;
2071 	int index;
2072 
2073 	if (witness_cold == 0)
2074 		mtx_assert(&w_mtx, MA_OWNED);
2075 
2076 	if (witness_watch == -1) {
2077 		mtx_unlock_spin(&w_mtx);
2078 		return (NULL);
2079 	}
2080 	if (STAILQ_EMPTY(&w_free)) {
2081 		witness_watch = -1;
2082 		mtx_unlock_spin(&w_mtx);
2083 		printf("WITNESS: unable to allocate a new witness object\n");
2084 		return (NULL);
2085 	}
2086 	w = STAILQ_FIRST(&w_free);
2087 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2088 	w_free_cnt--;
2089 	index = w->w_index;
2090 	MPASS(index > 0 && index == w_max_used_index+1 &&
2091 	    index < witness_count);
2092 	bzero(w, sizeof(*w));
2093 	w->w_index = index;
2094 	if (index > w_max_used_index)
2095 		w_max_used_index = index;
2096 	return (w);
2097 }
2098 
2099 static void
2100 witness_free(struct witness *w)
2101 {
2102 
2103 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2104 	w_free_cnt++;
2105 }
2106 
2107 static struct lock_list_entry *
2108 witness_lock_list_get(void)
2109 {
2110 	struct lock_list_entry *lle;
2111 
2112 	if (witness_watch == -1)
2113 		return (NULL);
2114 	mtx_lock_spin(&w_mtx);
2115 	lle = w_lock_list_free;
2116 	if (lle == NULL) {
2117 		witness_watch = -1;
2118 		mtx_unlock_spin(&w_mtx);
2119 		printf("%s: witness exhausted\n", __func__);
2120 		return (NULL);
2121 	}
2122 	w_lock_list_free = lle->ll_next;
2123 	mtx_unlock_spin(&w_mtx);
2124 	bzero(lle, sizeof(*lle));
2125 	return (lle);
2126 }
2127 
2128 static void
2129 witness_lock_list_free(struct lock_list_entry *lle)
2130 {
2131 
2132 	mtx_lock_spin(&w_mtx);
2133 	lle->ll_next = w_lock_list_free;
2134 	w_lock_list_free = lle;
2135 	mtx_unlock_spin(&w_mtx);
2136 }
2137 
2138 static struct lock_instance *
2139 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2140 {
2141 	struct lock_list_entry *lle;
2142 	struct lock_instance *instance;
2143 	int i;
2144 
2145 	for (lle = list; lle != NULL; lle = lle->ll_next)
2146 		for (i = lle->ll_count - 1; i >= 0; i--) {
2147 			instance = &lle->ll_children[i];
2148 			if (instance->li_lock == lock)
2149 				return (instance);
2150 		}
2151 	return (NULL);
2152 }
2153 
2154 static void
2155 witness_list_lock(struct lock_instance *instance,
2156     int (*prnt)(const char *fmt, ...))
2157 {
2158 	struct lock_object *lock;
2159 
2160 	lock = instance->li_lock;
2161 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2162 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2163 	if (lock->lo_witness->w_name != lock->lo_name)
2164 		prnt(" (%s)", lock->lo_witness->w_name);
2165 	prnt(" r = %d (%p) locked @ %s:%d\n",
2166 	    instance->li_flags & LI_RECURSEMASK, lock,
2167 	    fixup_filename(instance->li_file), instance->li_line);
2168 }
2169 
2170 #ifdef DDB
2171 static int
2172 witness_thread_has_locks(struct thread *td)
2173 {
2174 
2175 	if (td->td_sleeplocks == NULL)
2176 		return (0);
2177 	return (td->td_sleeplocks->ll_count != 0);
2178 }
2179 
2180 static int
2181 witness_proc_has_locks(struct proc *p)
2182 {
2183 	struct thread *td;
2184 
2185 	FOREACH_THREAD_IN_PROC(p, td) {
2186 		if (witness_thread_has_locks(td))
2187 			return (1);
2188 	}
2189 	return (0);
2190 }
2191 #endif
2192 
2193 int
2194 witness_list_locks(struct lock_list_entry **lock_list,
2195     int (*prnt)(const char *fmt, ...))
2196 {
2197 	struct lock_list_entry *lle;
2198 	int i, nheld;
2199 
2200 	nheld = 0;
2201 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2202 		for (i = lle->ll_count - 1; i >= 0; i--) {
2203 			witness_list_lock(&lle->ll_children[i], prnt);
2204 			nheld++;
2205 		}
2206 	return (nheld);
2207 }
2208 
2209 /*
2210  * This is a bit risky at best.  We call this function when we have timed
2211  * out acquiring a spin lock, and we assume that the other CPU is stuck
2212  * with this lock held.  So, we go groveling around in the other CPU's
2213  * per-cpu data to try to find the lock instance for this spin lock to
2214  * see when it was last acquired.
2215  */
2216 void
2217 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2218     int (*prnt)(const char *fmt, ...))
2219 {
2220 	struct lock_instance *instance;
2221 	struct pcpu *pc;
2222 
2223 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2224 		return;
2225 	pc = pcpu_find(owner->td_oncpu);
2226 	instance = find_instance(pc->pc_spinlocks, lock);
2227 	if (instance != NULL)
2228 		witness_list_lock(instance, prnt);
2229 }
2230 
2231 void
2232 witness_save(struct lock_object *lock, const char **filep, int *linep)
2233 {
2234 	struct lock_list_entry *lock_list;
2235 	struct lock_instance *instance;
2236 	struct lock_class *class;
2237 
2238 	/*
2239 	 * This function is used independently in locking code to deal with
2240 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2241 	 * is gone.
2242 	 */
2243 	if (SCHEDULER_STOPPED())
2244 		return;
2245 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2246 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2247 		return;
2248 	class = LOCK_CLASS(lock);
2249 	if (class->lc_flags & LC_SLEEPLOCK)
2250 		lock_list = curthread->td_sleeplocks;
2251 	else {
2252 		if (witness_skipspin)
2253 			return;
2254 		lock_list = PCPU_GET(spinlocks);
2255 	}
2256 	instance = find_instance(lock_list, lock);
2257 	if (instance == NULL) {
2258 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2259 		    class->lc_name, lock->lo_name);
2260 		return;
2261 	}
2262 	*filep = instance->li_file;
2263 	*linep = instance->li_line;
2264 }
2265 
2266 void
2267 witness_restore(struct lock_object *lock, const char *file, int line)
2268 {
2269 	struct lock_list_entry *lock_list;
2270 	struct lock_instance *instance;
2271 	struct lock_class *class;
2272 
2273 	/*
2274 	 * This function is used independently in locking code to deal with
2275 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2276 	 * is gone.
2277 	 */
2278 	if (SCHEDULER_STOPPED())
2279 		return;
2280 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2281 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2282 		return;
2283 	class = LOCK_CLASS(lock);
2284 	if (class->lc_flags & LC_SLEEPLOCK)
2285 		lock_list = curthread->td_sleeplocks;
2286 	else {
2287 		if (witness_skipspin)
2288 			return;
2289 		lock_list = PCPU_GET(spinlocks);
2290 	}
2291 	instance = find_instance(lock_list, lock);
2292 	if (instance == NULL)
2293 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2294 		    class->lc_name, lock->lo_name);
2295 	lock->lo_witness->w_file = file;
2296 	lock->lo_witness->w_line = line;
2297 	if (instance == NULL)
2298 		return;
2299 	instance->li_file = file;
2300 	instance->li_line = line;
2301 }
2302 
2303 void
2304 witness_assert(const struct lock_object *lock, int flags, const char *file,
2305     int line)
2306 {
2307 #ifdef INVARIANT_SUPPORT
2308 	struct lock_instance *instance;
2309 	struct lock_class *class;
2310 
2311 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2312 		return;
2313 	class = LOCK_CLASS(lock);
2314 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2315 		instance = find_instance(curthread->td_sleeplocks, lock);
2316 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2317 		instance = find_instance(PCPU_GET(spinlocks), lock);
2318 	else {
2319 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2320 		    class->lc_name, lock->lo_name);
2321 		return;
2322 	}
2323 	switch (flags) {
2324 	case LA_UNLOCKED:
2325 		if (instance != NULL)
2326 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2327 			    class->lc_name, lock->lo_name,
2328 			    fixup_filename(file), line);
2329 		break;
2330 	case LA_LOCKED:
2331 	case LA_LOCKED | LA_RECURSED:
2332 	case LA_LOCKED | LA_NOTRECURSED:
2333 	case LA_SLOCKED:
2334 	case LA_SLOCKED | LA_RECURSED:
2335 	case LA_SLOCKED | LA_NOTRECURSED:
2336 	case LA_XLOCKED:
2337 	case LA_XLOCKED | LA_RECURSED:
2338 	case LA_XLOCKED | LA_NOTRECURSED:
2339 		if (instance == NULL) {
2340 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2341 			    class->lc_name, lock->lo_name,
2342 			    fixup_filename(file), line);
2343 			break;
2344 		}
2345 		if ((flags & LA_XLOCKED) != 0 &&
2346 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2347 			kassert_panic(
2348 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2349 			    class->lc_name, lock->lo_name,
2350 			    fixup_filename(file), line);
2351 		if ((flags & LA_SLOCKED) != 0 &&
2352 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2353 			kassert_panic(
2354 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2355 			    class->lc_name, lock->lo_name,
2356 			    fixup_filename(file), line);
2357 		if ((flags & LA_RECURSED) != 0 &&
2358 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2359 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2360 			    class->lc_name, lock->lo_name,
2361 			    fixup_filename(file), line);
2362 		if ((flags & LA_NOTRECURSED) != 0 &&
2363 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2364 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2365 			    class->lc_name, lock->lo_name,
2366 			    fixup_filename(file), line);
2367 		break;
2368 	default:
2369 		kassert_panic("Invalid lock assertion at %s:%d.",
2370 		    fixup_filename(file), line);
2371 
2372 	}
2373 #endif	/* INVARIANT_SUPPORT */
2374 }
2375 
2376 static void
2377 witness_setflag(struct lock_object *lock, int flag, int set)
2378 {
2379 	struct lock_list_entry *lock_list;
2380 	struct lock_instance *instance;
2381 	struct lock_class *class;
2382 
2383 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2384 		return;
2385 	class = LOCK_CLASS(lock);
2386 	if (class->lc_flags & LC_SLEEPLOCK)
2387 		lock_list = curthread->td_sleeplocks;
2388 	else {
2389 		if (witness_skipspin)
2390 			return;
2391 		lock_list = PCPU_GET(spinlocks);
2392 	}
2393 	instance = find_instance(lock_list, lock);
2394 	if (instance == NULL) {
2395 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2396 		    class->lc_name, lock->lo_name);
2397 		return;
2398 	}
2399 
2400 	if (set)
2401 		instance->li_flags |= flag;
2402 	else
2403 		instance->li_flags &= ~flag;
2404 }
2405 
2406 void
2407 witness_norelease(struct lock_object *lock)
2408 {
2409 
2410 	witness_setflag(lock, LI_NORELEASE, 1);
2411 }
2412 
2413 void
2414 witness_releaseok(struct lock_object *lock)
2415 {
2416 
2417 	witness_setflag(lock, LI_NORELEASE, 0);
2418 }
2419 
2420 #ifdef DDB
2421 static void
2422 witness_ddb_list(struct thread *td)
2423 {
2424 
2425 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2426 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2427 
2428 	if (witness_watch < 1)
2429 		return;
2430 
2431 	witness_list_locks(&td->td_sleeplocks, db_printf);
2432 
2433 	/*
2434 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2435 	 * if td is currently executing on some other CPU and holds spin locks
2436 	 * as we won't display those locks.  If we had a MI way of getting
2437 	 * the per-cpu data for a given cpu then we could use
2438 	 * td->td_oncpu to get the list of spinlocks for this thread
2439 	 * and "fix" this.
2440 	 *
2441 	 * That still wouldn't really fix this unless we locked the scheduler
2442 	 * lock or stopped the other CPU to make sure it wasn't changing the
2443 	 * list out from under us.  It is probably best to just not try to
2444 	 * handle threads on other CPU's for now.
2445 	 */
2446 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2447 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2448 }
2449 
2450 DB_SHOW_COMMAND(locks, db_witness_list)
2451 {
2452 	struct thread *td;
2453 
2454 	if (have_addr)
2455 		td = db_lookup_thread(addr, true);
2456 	else
2457 		td = kdb_thread;
2458 	witness_ddb_list(td);
2459 }
2460 
2461 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2462 {
2463 	struct thread *td;
2464 	struct proc *p;
2465 
2466 	/*
2467 	 * It would be nice to list only threads and processes that actually
2468 	 * held sleep locks, but that information is currently not exported
2469 	 * by WITNESS.
2470 	 */
2471 	FOREACH_PROC_IN_SYSTEM(p) {
2472 		if (!witness_proc_has_locks(p))
2473 			continue;
2474 		FOREACH_THREAD_IN_PROC(p, td) {
2475 			if (!witness_thread_has_locks(td))
2476 				continue;
2477 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2478 			    p->p_comm, td, td->td_tid);
2479 			witness_ddb_list(td);
2480 			if (db_pager_quit)
2481 				return;
2482 		}
2483 	}
2484 }
2485 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2486 
2487 DB_SHOW_COMMAND(witness, db_witness_display)
2488 {
2489 
2490 	witness_ddb_display(db_printf);
2491 }
2492 #endif
2493 
2494 static int
2495 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2496 {
2497 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2498 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2499 	struct sbuf *sb;
2500 	u_int w_rmatrix1, w_rmatrix2;
2501 	int error, generation, i, j;
2502 
2503 	tmp_data1 = NULL;
2504 	tmp_data2 = NULL;
2505 	tmp_w1 = NULL;
2506 	tmp_w2 = NULL;
2507 	if (witness_watch < 1) {
2508 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2509 		return (error);
2510 	}
2511 	if (witness_cold) {
2512 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2513 		return (error);
2514 	}
2515 	error = 0;
2516 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2517 	if (sb == NULL)
2518 		return (ENOMEM);
2519 
2520 	/* Allocate and init temporary storage space. */
2521 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2522 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2523 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2524 	    M_WAITOK | M_ZERO);
2525 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2526 	    M_WAITOK | M_ZERO);
2527 	stack_zero(&tmp_data1->wlod_stack);
2528 	stack_zero(&tmp_data2->wlod_stack);
2529 
2530 restart:
2531 	mtx_lock_spin(&w_mtx);
2532 	generation = w_generation;
2533 	mtx_unlock_spin(&w_mtx);
2534 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2535 	    w_lohash.wloh_count);
2536 	for (i = 1; i < w_max_used_index; i++) {
2537 		mtx_lock_spin(&w_mtx);
2538 		if (generation != w_generation) {
2539 			mtx_unlock_spin(&w_mtx);
2540 
2541 			/* The graph has changed, try again. */
2542 			req->oldidx = 0;
2543 			sbuf_clear(sb);
2544 			goto restart;
2545 		}
2546 
2547 		w1 = &w_data[i];
2548 		if (w1->w_reversed == 0) {
2549 			mtx_unlock_spin(&w_mtx);
2550 			continue;
2551 		}
2552 
2553 		/* Copy w1 locally so we can release the spin lock. */
2554 		*tmp_w1 = *w1;
2555 		mtx_unlock_spin(&w_mtx);
2556 
2557 		if (tmp_w1->w_reversed == 0)
2558 			continue;
2559 		for (j = 1; j < w_max_used_index; j++) {
2560 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2561 				continue;
2562 
2563 			mtx_lock_spin(&w_mtx);
2564 			if (generation != w_generation) {
2565 				mtx_unlock_spin(&w_mtx);
2566 
2567 				/* The graph has changed, try again. */
2568 				req->oldidx = 0;
2569 				sbuf_clear(sb);
2570 				goto restart;
2571 			}
2572 
2573 			w2 = &w_data[j];
2574 			data1 = witness_lock_order_get(w1, w2);
2575 			data2 = witness_lock_order_get(w2, w1);
2576 
2577 			/*
2578 			 * Copy information locally so we can release the
2579 			 * spin lock.
2580 			 */
2581 			*tmp_w2 = *w2;
2582 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2583 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2584 
2585 			if (data1) {
2586 				stack_zero(&tmp_data1->wlod_stack);
2587 				stack_copy(&data1->wlod_stack,
2588 				    &tmp_data1->wlod_stack);
2589 			}
2590 			if (data2 && data2 != data1) {
2591 				stack_zero(&tmp_data2->wlod_stack);
2592 				stack_copy(&data2->wlod_stack,
2593 				    &tmp_data2->wlod_stack);
2594 			}
2595 			mtx_unlock_spin(&w_mtx);
2596 
2597 			sbuf_printf(sb,
2598 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2599 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2600 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2601 #if 0
2602  			sbuf_printf(sb,
2603 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2604  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2605  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2606 #endif
2607 			if (data1) {
2608 				sbuf_printf(sb,
2609 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2610 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2611 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2612 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2613 				sbuf_printf(sb, "\n");
2614 			}
2615 			if (data2 && data2 != data1) {
2616 				sbuf_printf(sb,
2617 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2618 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2619 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2620 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2621 				sbuf_printf(sb, "\n");
2622 			}
2623 		}
2624 	}
2625 	mtx_lock_spin(&w_mtx);
2626 	if (generation != w_generation) {
2627 		mtx_unlock_spin(&w_mtx);
2628 
2629 		/*
2630 		 * The graph changed while we were printing stack data,
2631 		 * try again.
2632 		 */
2633 		req->oldidx = 0;
2634 		sbuf_clear(sb);
2635 		goto restart;
2636 	}
2637 	mtx_unlock_spin(&w_mtx);
2638 
2639 	/* Free temporary storage space. */
2640 	free(tmp_data1, M_TEMP);
2641 	free(tmp_data2, M_TEMP);
2642 	free(tmp_w1, M_TEMP);
2643 	free(tmp_w2, M_TEMP);
2644 
2645 	sbuf_finish(sb);
2646 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2647 	sbuf_delete(sb);
2648 
2649 	return (error);
2650 }
2651 
2652 static int
2653 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2654 {
2655 	struct witness *w;
2656 	struct sbuf *sb;
2657 	int error;
2658 
2659 	if (witness_watch < 1) {
2660 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2661 		return (error);
2662 	}
2663 	if (witness_cold) {
2664 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2665 		return (error);
2666 	}
2667 	error = 0;
2668 
2669 	error = sysctl_wire_old_buffer(req, 0);
2670 	if (error != 0)
2671 		return (error);
2672 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2673 	if (sb == NULL)
2674 		return (ENOMEM);
2675 	sbuf_printf(sb, "\n");
2676 
2677 	mtx_lock_spin(&w_mtx);
2678 	STAILQ_FOREACH(w, &w_all, w_list)
2679 		w->w_displayed = 0;
2680 	STAILQ_FOREACH(w, &w_all, w_list)
2681 		witness_add_fullgraph(sb, w);
2682 	mtx_unlock_spin(&w_mtx);
2683 
2684 	/*
2685 	 * Close the sbuf and return to userland.
2686 	 */
2687 	error = sbuf_finish(sb);
2688 	sbuf_delete(sb);
2689 
2690 	return (error);
2691 }
2692 
2693 static int
2694 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2695 {
2696 	int error, value;
2697 
2698 	value = witness_watch;
2699 	error = sysctl_handle_int(oidp, &value, 0, req);
2700 	if (error != 0 || req->newptr == NULL)
2701 		return (error);
2702 	if (value > 1 || value < -1 ||
2703 	    (witness_watch == -1 && value != witness_watch))
2704 		return (EINVAL);
2705 	witness_watch = value;
2706 	return (0);
2707 }
2708 
2709 static void
2710 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2711 {
2712 	int i;
2713 
2714 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2715 		return;
2716 	w->w_displayed = 1;
2717 
2718 	WITNESS_INDEX_ASSERT(w->w_index);
2719 	for (i = 1; i <= w_max_used_index; i++) {
2720 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2721 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2722 			    w_data[i].w_name);
2723 			witness_add_fullgraph(sb, &w_data[i]);
2724 		}
2725 	}
2726 }
2727 
2728 /*
2729  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2730  * interprets the key as a string and reads until the null
2731  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2732  * hash value computed from the key.
2733  */
2734 static uint32_t
2735 witness_hash_djb2(const uint8_t *key, uint32_t size)
2736 {
2737 	unsigned int hash = 5381;
2738 	int i;
2739 
2740 	/* hash = hash * 33 + key[i] */
2741 	if (size)
2742 		for (i = 0; i < size; i++)
2743 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2744 	else
2745 		for (i = 0; key[i] != 0; i++)
2746 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2747 
2748 	return (hash);
2749 }
2750 
2751 
2752 /*
2753  * Initializes the two witness hash tables. Called exactly once from
2754  * witness_initialize().
2755  */
2756 static void
2757 witness_init_hash_tables(void)
2758 {
2759 	int i;
2760 
2761 	MPASS(witness_cold);
2762 
2763 	/* Initialize the hash tables. */
2764 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2765 		w_hash.wh_array[i] = NULL;
2766 
2767 	w_hash.wh_size = WITNESS_HASH_SIZE;
2768 	w_hash.wh_count = 0;
2769 
2770 	/* Initialize the lock order data hash. */
2771 	w_lofree = NULL;
2772 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2773 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2774 		w_lodata[i].wlod_next = w_lofree;
2775 		w_lofree = &w_lodata[i];
2776 	}
2777 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2778 	w_lohash.wloh_count = 0;
2779 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2780 		w_lohash.wloh_array[i] = NULL;
2781 }
2782 
2783 static struct witness *
2784 witness_hash_get(const char *key)
2785 {
2786 	struct witness *w;
2787 	uint32_t hash;
2788 
2789 	MPASS(key != NULL);
2790 	if (witness_cold == 0)
2791 		mtx_assert(&w_mtx, MA_OWNED);
2792 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2793 	w = w_hash.wh_array[hash];
2794 	while (w != NULL) {
2795 		if (strcmp(w->w_name, key) == 0)
2796 			goto out;
2797 		w = w->w_hash_next;
2798 	}
2799 
2800 out:
2801 	return (w);
2802 }
2803 
2804 static void
2805 witness_hash_put(struct witness *w)
2806 {
2807 	uint32_t hash;
2808 
2809 	MPASS(w != NULL);
2810 	MPASS(w->w_name != NULL);
2811 	if (witness_cold == 0)
2812 		mtx_assert(&w_mtx, MA_OWNED);
2813 	KASSERT(witness_hash_get(w->w_name) == NULL,
2814 	    ("%s: trying to add a hash entry that already exists!", __func__));
2815 	KASSERT(w->w_hash_next == NULL,
2816 	    ("%s: w->w_hash_next != NULL", __func__));
2817 
2818 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2819 	w->w_hash_next = w_hash.wh_array[hash];
2820 	w_hash.wh_array[hash] = w;
2821 	w_hash.wh_count++;
2822 }
2823 
2824 
2825 static struct witness_lock_order_data *
2826 witness_lock_order_get(struct witness *parent, struct witness *child)
2827 {
2828 	struct witness_lock_order_data *data = NULL;
2829 	struct witness_lock_order_key key;
2830 	unsigned int hash;
2831 
2832 	MPASS(parent != NULL && child != NULL);
2833 	key.from = parent->w_index;
2834 	key.to = child->w_index;
2835 	WITNESS_INDEX_ASSERT(key.from);
2836 	WITNESS_INDEX_ASSERT(key.to);
2837 	if ((w_rmatrix[parent->w_index][child->w_index]
2838 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2839 		goto out;
2840 
2841 	hash = witness_hash_djb2((const char*)&key,
2842 	    sizeof(key)) % w_lohash.wloh_size;
2843 	data = w_lohash.wloh_array[hash];
2844 	while (data != NULL) {
2845 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2846 			break;
2847 		data = data->wlod_next;
2848 	}
2849 
2850 out:
2851 	return (data);
2852 }
2853 
2854 /*
2855  * Verify that parent and child have a known relationship, are not the same,
2856  * and child is actually a child of parent.  This is done without w_mtx
2857  * to avoid contention in the common case.
2858  */
2859 static int
2860 witness_lock_order_check(struct witness *parent, struct witness *child)
2861 {
2862 
2863 	if (parent != child &&
2864 	    w_rmatrix[parent->w_index][child->w_index]
2865 	    & WITNESS_LOCK_ORDER_KNOWN &&
2866 	    isitmychild(parent, child))
2867 		return (1);
2868 
2869 	return (0);
2870 }
2871 
2872 static int
2873 witness_lock_order_add(struct witness *parent, struct witness *child)
2874 {
2875 	struct witness_lock_order_data *data = NULL;
2876 	struct witness_lock_order_key key;
2877 	unsigned int hash;
2878 
2879 	MPASS(parent != NULL && child != NULL);
2880 	key.from = parent->w_index;
2881 	key.to = child->w_index;
2882 	WITNESS_INDEX_ASSERT(key.from);
2883 	WITNESS_INDEX_ASSERT(key.to);
2884 	if (w_rmatrix[parent->w_index][child->w_index]
2885 	    & WITNESS_LOCK_ORDER_KNOWN)
2886 		return (1);
2887 
2888 	hash = witness_hash_djb2((const char*)&key,
2889 	    sizeof(key)) % w_lohash.wloh_size;
2890 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2891 	data = w_lofree;
2892 	if (data == NULL)
2893 		return (0);
2894 	w_lofree = data->wlod_next;
2895 	data->wlod_next = w_lohash.wloh_array[hash];
2896 	data->wlod_key = key;
2897 	w_lohash.wloh_array[hash] = data;
2898 	w_lohash.wloh_count++;
2899 	stack_zero(&data->wlod_stack);
2900 	stack_save(&data->wlod_stack);
2901 	return (1);
2902 }
2903 
2904 /* Call this whenver the structure of the witness graph changes. */
2905 static void
2906 witness_increment_graph_generation(void)
2907 {
2908 
2909 	if (witness_cold == 0)
2910 		mtx_assert(&w_mtx, MA_OWNED);
2911 	w_generation++;
2912 }
2913 
2914 #ifdef KDB
2915 static void
2916 _witness_debugger(int cond, const char *msg)
2917 {
2918 
2919 	if (witness_trace && cond)
2920 		kdb_backtrace();
2921 	if (witness_kdb && cond)
2922 		kdb_enter(KDB_WHY_WITNESS, msg);
2923 }
2924 #endif
2925