xref: /freebsd/sys/kern/subr_witness.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
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/syslog.h>
110 #include <sys/systm.h>
111 
112 #ifdef DDB
113 #include <ddb/ddb.h>
114 #endif
115 
116 #include <machine/stdarg.h>
117 
118 #if !defined(DDB) && !defined(STACK)
119 #error "DDB or STACK options are required for WITNESS"
120 #endif
121 
122 /* Note that these traces do not work with KTR_ALQ. */
123 #if 0
124 #define	KTR_WITNESS	KTR_SUBSYS
125 #else
126 #define	KTR_WITNESS	0
127 #endif
128 
129 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
130 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
131 #define	LI_NORELEASE	0x00020000	/* Lock not allowed to be released. */
132 
133 /* Define this to check for blessed mutexes */
134 #undef BLESSING
135 
136 #ifndef WITNESS_COUNT
137 #define	WITNESS_COUNT 		1536
138 #endif
139 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
140 #define	WITNESS_PENDLIST	(1024 + MAXCPU)
141 
142 /* Allocate 256 KB of stack data space */
143 #define	WITNESS_LO_DATA_COUNT	2048
144 
145 /* Prime, gives load factor of ~2 at full load */
146 #define	WITNESS_LO_HASH_SIZE	1021
147 
148 /*
149  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
150  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
151  * probably be safe for the most part, but it's still a SWAG.
152  */
153 #define	LOCK_NCHILDREN	5
154 #define	LOCK_CHILDCOUNT	2048
155 
156 #define	MAX_W_NAME	64
157 
158 #define	FULLGRAPH_SBUF_SIZE	512
159 
160 /*
161  * These flags go in the witness relationship matrix and describe the
162  * relationship between any two struct witness objects.
163  */
164 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
165 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
166 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
167 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
168 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
169 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
170 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
171 #define	WITNESS_RELATED_MASK						\
172 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
173 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
174 					  * observed. */
175 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
176 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
177 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
178 
179 /* Descendant to ancestor flags */
180 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
181 
182 /* Ancestor to descendant flags */
183 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
184 
185 #define	WITNESS_INDEX_ASSERT(i)						\
186 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
187 
188 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189 
190 /*
191  * Lock instances.  A lock instance is the data associated with a lock while
192  * it is held by witness.  For example, a lock instance will hold the
193  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
194  * are held in a per-cpu list while sleep locks are held in per-thread list.
195  */
196 struct lock_instance {
197 	struct lock_object	*li_lock;
198 	const char		*li_file;
199 	int			li_line;
200 	u_int			li_flags;
201 };
202 
203 /*
204  * A simple list type used to build the list of locks held by a thread
205  * or CPU.  We can't simply embed the list in struct lock_object since a
206  * lock may be held by more than one thread if it is a shared lock.  Locks
207  * are added to the head of the list, so we fill up each list entry from
208  * "the back" logically.  To ease some of the arithmetic, we actually fill
209  * in each list entry the normal way (children[0] then children[1], etc.) but
210  * when we traverse the list we read children[count-1] as the first entry
211  * down to children[0] as the final entry.
212  */
213 struct lock_list_entry {
214 	struct lock_list_entry	*ll_next;
215 	struct lock_instance	ll_children[LOCK_NCHILDREN];
216 	u_int			ll_count;
217 };
218 
219 /*
220  * The main witness structure. One of these per named lock type in the system
221  * (for example, "vnode interlock").
222  */
223 struct witness {
224 	char  			w_name[MAX_W_NAME];
225 	uint32_t 		w_index;  /* Index in the relationship matrix */
226 	struct lock_class	*w_class;
227 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
228 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
229 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
230 	const char		*w_file; /* File where last acquired */
231 	uint32_t 		w_line; /* Line where last acquired */
232 	uint32_t 		w_refcount;
233 	uint16_t 		w_num_ancestors; /* direct/indirect
234 						  * ancestor count */
235 	uint16_t 		w_num_descendants; /* direct/indirect
236 						    * descendant count */
237 	int16_t 		w_ddb_level;
238 	unsigned		w_displayed:1;
239 	unsigned		w_reversed:1;
240 };
241 
242 STAILQ_HEAD(witness_list, witness);
243 
244 /*
245  * The witness hash table. Keys are witness names (const char *), elements are
246  * witness objects (struct witness *).
247  */
248 struct witness_hash {
249 	struct witness	*wh_array[WITNESS_HASH_SIZE];
250 	uint32_t	wh_size;
251 	uint32_t	wh_count;
252 };
253 
254 /*
255  * Key type for the lock order data hash table.
256  */
257 struct witness_lock_order_key {
258 	uint16_t	from;
259 	uint16_t	to;
260 };
261 
262 struct witness_lock_order_data {
263 	struct stack			wlod_stack;
264 	struct witness_lock_order_key	wlod_key;
265 	struct witness_lock_order_data	*wlod_next;
266 };
267 
268 /*
269  * The witness lock order data hash table. Keys are witness index tuples
270  * (struct witness_lock_order_key), elements are lock order data objects
271  * (struct witness_lock_order_data).
272  */
273 struct witness_lock_order_hash {
274 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
275 	u_int	wloh_size;
276 	u_int	wloh_count;
277 };
278 
279 #ifdef BLESSING
280 struct witness_blessed {
281 	const char	*b_lock1;
282 	const char	*b_lock2;
283 };
284 #endif
285 
286 struct witness_pendhelp {
287 	const char		*wh_type;
288 	struct lock_object	*wh_lock;
289 };
290 
291 struct witness_order_list_entry {
292 	const char		*w_name;
293 	struct lock_class	*w_class;
294 };
295 
296 /*
297  * Returns 0 if one of the locks is a spin lock and the other is not.
298  * Returns 1 otherwise.
299  */
300 static __inline int
301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
302 {
303 
304 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
305 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
306 }
307 
308 static __inline int
309 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
310     const struct witness_lock_order_key *b)
311 {
312 
313 	return (a->from == b->from && a->to == b->to);
314 }
315 
316 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
317 		    const char *fname);
318 static void	adopt(struct witness *parent, struct witness *child);
319 #ifdef BLESSING
320 static int	blessed(struct witness *, struct witness *);
321 #endif
322 static void	depart(struct witness *w);
323 static struct witness	*enroll(const char *description,
324 			    struct lock_class *lock_class);
325 static struct lock_instance	*find_instance(struct lock_list_entry *list,
326 				    const struct lock_object *lock);
327 static int	isitmychild(struct witness *parent, struct witness *child);
328 static int	isitmydescendant(struct witness *parent, struct witness *child);
329 static void	itismychild(struct witness *parent, struct witness *child);
330 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
331 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
332 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
333 static int	sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
334 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
335 #ifdef DDB
336 static void	witness_ddb_compute_levels(void);
337 static void	witness_ddb_display(int(*)(const char *fmt, ...));
338 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
339 		    struct witness *, int indent);
340 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
341 		    struct witness_list *list);
342 static void	witness_ddb_level_descendants(struct witness *parent, int l);
343 static void	witness_ddb_list(struct thread *td);
344 #endif
345 static void	witness_debugger(int cond, const char *msg);
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 int	witness_output(const char *fmt, ...) __printflike(1, 2);
365 static int	witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
366 static void	witness_setflag(struct lock_object *lock, int flag, int set);
367 
368 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
369     "Witness Locking");
370 
371 /*
372  * If set to 0, lock order checking is disabled.  If set to -1,
373  * witness is completely disabled.  Otherwise witness performs full
374  * lock order checking for all locks.  At runtime, lock order checking
375  * may be toggled.  However, witness cannot be reenabled once it is
376  * completely disabled.
377  */
378 static int witness_watch = 1;
379 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
380     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
381 
382 #ifdef KDB
383 /*
384  * When KDB is enabled and witness_kdb is 1, it will cause the system
385  * to drop into kdebug() when:
386  *	- a lock hierarchy violation occurs
387  *	- locks are held when going to sleep.
388  */
389 #ifdef WITNESS_KDB
390 int	witness_kdb = 1;
391 #else
392 int	witness_kdb = 0;
393 #endif
394 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
395 #endif /* KDB */
396 
397 #if defined(DDB) || defined(KDB)
398 /*
399  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
400  * to print a stack trace:
401  *	- a lock hierarchy violation occurs
402  *	- locks are held when going to sleep.
403  */
404 int	witness_trace = 1;
405 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
406 #endif /* DDB || KDB */
407 
408 #ifdef WITNESS_SKIPSPIN
409 int	witness_skipspin = 1;
410 #else
411 int	witness_skipspin = 0;
412 #endif
413 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
414 
415 int badstack_sbuf_size;
416 
417 int witness_count = WITNESS_COUNT;
418 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
419     &witness_count, 0, "");
420 
421 /*
422  * Output channel for witness messages.  By default we print to the console.
423  */
424 enum witness_channel {
425 	WITNESS_CONSOLE,
426 	WITNESS_LOG,
427 	WITNESS_NONE,
428 };
429 
430 static enum witness_channel witness_channel = WITNESS_CONSOLE;
431 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
432     CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
433     "Output channel for warnings");
434 
435 /*
436  * Call this to print out the relations between locks.
437  */
438 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
439     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
440 
441 /*
442  * Call this to print out the witness faulty stacks.
443  */
444 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
445     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
446 
447 static struct mtx w_mtx;
448 
449 /* w_list */
450 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
451 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
452 
453 /* w_typelist */
454 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
455 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
456 
457 /* lock list */
458 static struct lock_list_entry *w_lock_list_free = NULL;
459 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
460 static u_int pending_cnt;
461 
462 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
463 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
464 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
465 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
466     "");
467 
468 static struct witness *w_data;
469 static uint8_t **w_rmatrix;
470 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
471 static struct witness_hash w_hash;	/* The witness hash table. */
472 
473 /* The lock order data hash */
474 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
475 static struct witness_lock_order_data *w_lofree = NULL;
476 static struct witness_lock_order_hash w_lohash;
477 static int w_max_used_index = 0;
478 static unsigned int w_generation = 0;
479 static const char w_notrunning[] = "Witness not running\n";
480 static const char w_stillcold[] = "Witness is still cold\n";
481 
482 
483 static struct witness_order_list_entry order_lists[] = {
484 	/*
485 	 * sx locks
486 	 */
487 	{ "proctree", &lock_class_sx },
488 	{ "allproc", &lock_class_sx },
489 	{ "allprison", &lock_class_sx },
490 	{ NULL, NULL },
491 	/*
492 	 * Various mutexes
493 	 */
494 	{ "Giant", &lock_class_mtx_sleep },
495 	{ "pipe mutex", &lock_class_mtx_sleep },
496 	{ "sigio lock", &lock_class_mtx_sleep },
497 	{ "process group", &lock_class_mtx_sleep },
498 	{ "process lock", &lock_class_mtx_sleep },
499 	{ "session", &lock_class_mtx_sleep },
500 	{ "uidinfo hash", &lock_class_rw },
501 #ifdef	HWPMC_HOOKS
502 	{ "pmc-sleep", &lock_class_mtx_sleep },
503 #endif
504 	{ "time lock", &lock_class_mtx_sleep },
505 	{ NULL, NULL },
506 	/*
507 	 * umtx
508 	 */
509 	{ "umtx lock", &lock_class_mtx_sleep },
510 	{ NULL, NULL },
511 	/*
512 	 * Sockets
513 	 */
514 	{ "accept", &lock_class_mtx_sleep },
515 	{ "so_snd", &lock_class_mtx_sleep },
516 	{ "so_rcv", &lock_class_mtx_sleep },
517 	{ "sellck", &lock_class_mtx_sleep },
518 	{ NULL, NULL },
519 	/*
520 	 * Routing
521 	 */
522 	{ "so_rcv", &lock_class_mtx_sleep },
523 	{ "radix node head", &lock_class_rw },
524 	{ "rtentry", &lock_class_mtx_sleep },
525 	{ "ifaddr", &lock_class_mtx_sleep },
526 	{ NULL, NULL },
527 	/*
528 	 * IPv4 multicast:
529 	 * protocol locks before interface locks, after UDP locks.
530 	 */
531 	{ "udpinp", &lock_class_rw },
532 	{ "in_multi_mtx", &lock_class_mtx_sleep },
533 	{ "igmp_mtx", &lock_class_mtx_sleep },
534 	{ "if_addr_lock", &lock_class_rw },
535 	{ NULL, NULL },
536 	/*
537 	 * IPv6 multicast:
538 	 * protocol locks before interface locks, after UDP locks.
539 	 */
540 	{ "udpinp", &lock_class_rw },
541 	{ "in6_multi_mtx", &lock_class_mtx_sleep },
542 	{ "mld_mtx", &lock_class_mtx_sleep },
543 	{ "if_addr_lock", &lock_class_rw },
544 	{ NULL, NULL },
545 	/*
546 	 * UNIX Domain Sockets
547 	 */
548 	{ "unp_link_rwlock", &lock_class_rw },
549 	{ "unp_list_lock", &lock_class_mtx_sleep },
550 	{ "unp", &lock_class_mtx_sleep },
551 	{ "so_snd", &lock_class_mtx_sleep },
552 	{ NULL, NULL },
553 	/*
554 	 * UDP/IP
555 	 */
556 	{ "udp", &lock_class_rw },
557 	{ "udpinp", &lock_class_rw },
558 	{ "so_snd", &lock_class_mtx_sleep },
559 	{ NULL, NULL },
560 	/*
561 	 * TCP/IP
562 	 */
563 	{ "tcp", &lock_class_rw },
564 	{ "tcpinp", &lock_class_rw },
565 	{ "so_snd", &lock_class_mtx_sleep },
566 	{ NULL, NULL },
567 	/*
568 	 * BPF
569 	 */
570 	{ "bpf global lock", &lock_class_mtx_sleep },
571 	{ "bpf interface lock", &lock_class_rw },
572 	{ "bpf cdev lock", &lock_class_mtx_sleep },
573 	{ NULL, NULL },
574 	/*
575 	 * NFS server
576 	 */
577 	{ "nfsd_mtx", &lock_class_mtx_sleep },
578 	{ "so_snd", &lock_class_mtx_sleep },
579 	{ NULL, NULL },
580 
581 	/*
582 	 * IEEE 802.11
583 	 */
584 	{ "802.11 com lock", &lock_class_mtx_sleep},
585 	{ NULL, NULL },
586 	/*
587 	 * Network drivers
588 	 */
589 	{ "network driver", &lock_class_mtx_sleep},
590 	{ NULL, NULL },
591 
592 	/*
593 	 * Netgraph
594 	 */
595 	{ "ng_node", &lock_class_mtx_sleep },
596 	{ "ng_worklist", &lock_class_mtx_sleep },
597 	{ NULL, NULL },
598 	/*
599 	 * CDEV
600 	 */
601 	{ "vm map (system)", &lock_class_mtx_sleep },
602 	{ "vm page queue", &lock_class_mtx_sleep },
603 	{ "vnode interlock", &lock_class_mtx_sleep },
604 	{ "cdev", &lock_class_mtx_sleep },
605 	{ NULL, NULL },
606 	/*
607 	 * VM
608 	 */
609 	{ "vm map (user)", &lock_class_sx },
610 	{ "vm object", &lock_class_rw },
611 	{ "vm page", &lock_class_mtx_sleep },
612 	{ "vm page queue", &lock_class_mtx_sleep },
613 	{ "pmap pv global", &lock_class_rw },
614 	{ "pmap", &lock_class_mtx_sleep },
615 	{ "pmap pv list", &lock_class_rw },
616 	{ "vm page free queue", &lock_class_mtx_sleep },
617 	{ NULL, NULL },
618 	/*
619 	 * kqueue/VFS interaction
620 	 */
621 	{ "kqueue", &lock_class_mtx_sleep },
622 	{ "struct mount mtx", &lock_class_mtx_sleep },
623 	{ "vnode interlock", &lock_class_mtx_sleep },
624 	{ NULL, NULL },
625 	/*
626 	 * ZFS locking
627 	 */
628 	{ "dn->dn_mtx", &lock_class_sx },
629 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
630 	{ "db->db_mtx", &lock_class_sx },
631 	{ NULL, NULL },
632 	/*
633 	 * spin locks
634 	 */
635 #ifdef SMP
636 	{ "ap boot", &lock_class_mtx_spin },
637 #endif
638 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
639 	{ "sio", &lock_class_mtx_spin },
640 #ifdef __i386__
641 	{ "cy", &lock_class_mtx_spin },
642 #endif
643 #ifdef __sparc64__
644 	{ "pcib_mtx", &lock_class_mtx_spin },
645 	{ "rtc_mtx", &lock_class_mtx_spin },
646 #endif
647 	{ "scc_hwmtx", &lock_class_mtx_spin },
648 	{ "uart_hwmtx", &lock_class_mtx_spin },
649 	{ "fast_taskqueue", &lock_class_mtx_spin },
650 	{ "intr table", &lock_class_mtx_spin },
651 #ifdef	HWPMC_HOOKS
652 	{ "pmc-per-proc", &lock_class_mtx_spin },
653 #endif
654 	{ "process slock", &lock_class_mtx_spin },
655 	{ "syscons video lock", &lock_class_mtx_spin },
656 	{ "sleepq chain", &lock_class_mtx_spin },
657 	{ "rm_spinlock", &lock_class_mtx_spin },
658 	{ "turnstile chain", &lock_class_mtx_spin },
659 	{ "turnstile lock", &lock_class_mtx_spin },
660 	{ "sched lock", &lock_class_mtx_spin },
661 	{ "td_contested", &lock_class_mtx_spin },
662 	{ "callout", &lock_class_mtx_spin },
663 	{ "entropy harvest mutex", &lock_class_mtx_spin },
664 #ifdef SMP
665 	{ "smp rendezvous", &lock_class_mtx_spin },
666 #endif
667 #ifdef __powerpc__
668 	{ "tlb0", &lock_class_mtx_spin },
669 #endif
670 	/*
671 	 * leaf locks
672 	 */
673 	{ "intrcnt", &lock_class_mtx_spin },
674 	{ "icu", &lock_class_mtx_spin },
675 #if defined(SMP) && defined(__sparc64__)
676 	{ "ipi", &lock_class_mtx_spin },
677 #endif
678 #ifdef __i386__
679 	{ "allpmaps", &lock_class_mtx_spin },
680 	{ "descriptor tables", &lock_class_mtx_spin },
681 #endif
682 	{ "clk", &lock_class_mtx_spin },
683 	{ "cpuset", &lock_class_mtx_spin },
684 	{ "mprof lock", &lock_class_mtx_spin },
685 	{ "zombie lock", &lock_class_mtx_spin },
686 	{ "ALD Queue", &lock_class_mtx_spin },
687 #if defined(__i386__) || defined(__amd64__)
688 	{ "pcicfg", &lock_class_mtx_spin },
689 	{ "NDIS thread lock", &lock_class_mtx_spin },
690 #endif
691 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
692 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
693 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
694 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
695 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
696 #ifdef	HWPMC_HOOKS
697 	{ "pmc-leaf", &lock_class_mtx_spin },
698 #endif
699 	{ "blocked lock", &lock_class_mtx_spin },
700 	{ NULL, NULL },
701 	{ NULL, NULL }
702 };
703 
704 #ifdef BLESSING
705 /*
706  * Pairs of locks which have been blessed
707  * Don't complain about order problems with blessed locks
708  */
709 static struct witness_blessed blessed_list[] = {
710 };
711 #endif
712 
713 /*
714  * This global is set to 0 once it becomes safe to use the witness code.
715  */
716 static int witness_cold = 1;
717 
718 /*
719  * This global is set to 1 once the static lock orders have been enrolled
720  * so that a warning can be issued for any spin locks enrolled later.
721  */
722 static int witness_spin_warn = 0;
723 
724 /* Trim useless garbage from filenames. */
725 static const char *
726 fixup_filename(const char *file)
727 {
728 
729 	if (file == NULL)
730 		return (NULL);
731 	while (strncmp(file, "../", 3) == 0)
732 		file += 3;
733 	return (file);
734 }
735 
736 /*
737  * The WITNESS-enabled diagnostic code.  Note that the witness code does
738  * assume that the early boot is single-threaded at least until after this
739  * routine is completed.
740  */
741 static void
742 witness_initialize(void *dummy __unused)
743 {
744 	struct lock_object *lock;
745 	struct witness_order_list_entry *order;
746 	struct witness *w, *w1;
747 	int i;
748 
749 	w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
750 	    M_WAITOK | M_ZERO);
751 
752 	w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
753 	    M_WITNESS, M_WAITOK | M_ZERO);
754 
755 	for (i = 0; i < witness_count + 1; i++) {
756 		w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
757 		    (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
758 	}
759 	badstack_sbuf_size = witness_count * 256;
760 
761 	/*
762 	 * We have to release Giant before initializing its witness
763 	 * structure so that WITNESS doesn't get confused.
764 	 */
765 	mtx_unlock(&Giant);
766 	mtx_assert(&Giant, MA_NOTOWNED);
767 
768 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
769 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
770 	    MTX_NOWITNESS | MTX_NOPROFILE);
771 	for (i = witness_count - 1; i >= 0; i--) {
772 		w = &w_data[i];
773 		memset(w, 0, sizeof(*w));
774 		w_data[i].w_index = i;	/* Witness index never changes. */
775 		witness_free(w);
776 	}
777 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
778 	    ("%s: Invalid list of free witness objects", __func__));
779 
780 	/* Witness with index 0 is not used to aid in debugging. */
781 	STAILQ_REMOVE_HEAD(&w_free, w_list);
782 	w_free_cnt--;
783 
784 	for (i = 0; i < witness_count; i++) {
785 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
786 		    (witness_count + 1));
787 	}
788 
789 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
790 		witness_lock_list_free(&w_locklistdata[i]);
791 	witness_init_hash_tables();
792 
793 	/* First add in all the specified order lists. */
794 	for (order = order_lists; order->w_name != NULL; order++) {
795 		w = enroll(order->w_name, order->w_class);
796 		if (w == NULL)
797 			continue;
798 		w->w_file = "order list";
799 		for (order++; order->w_name != NULL; order++) {
800 			w1 = enroll(order->w_name, order->w_class);
801 			if (w1 == NULL)
802 				continue;
803 			w1->w_file = "order list";
804 			itismychild(w, w1);
805 			w = w1;
806 		}
807 	}
808 	witness_spin_warn = 1;
809 
810 	/* Iterate through all locks and add them to witness. */
811 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
812 		lock = pending_locks[i].wh_lock;
813 		KASSERT(lock->lo_flags & LO_WITNESS,
814 		    ("%s: lock %s is on pending list but not LO_WITNESS",
815 		    __func__, lock->lo_name));
816 		lock->lo_witness = enroll(pending_locks[i].wh_type,
817 		    LOCK_CLASS(lock));
818 	}
819 
820 	/* Mark the witness code as being ready for use. */
821 	witness_cold = 0;
822 
823 	mtx_lock(&Giant);
824 }
825 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
826     NULL);
827 
828 void
829 witness_init(struct lock_object *lock, const char *type)
830 {
831 	struct lock_class *class;
832 
833 	/* Various sanity checks. */
834 	class = LOCK_CLASS(lock);
835 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
836 	    (class->lc_flags & LC_RECURSABLE) == 0)
837 		kassert_panic("%s: lock (%s) %s can not be recursable",
838 		    __func__, class->lc_name, lock->lo_name);
839 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
840 	    (class->lc_flags & LC_SLEEPABLE) == 0)
841 		kassert_panic("%s: lock (%s) %s can not be sleepable",
842 		    __func__, class->lc_name, lock->lo_name);
843 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
844 	    (class->lc_flags & LC_UPGRADABLE) == 0)
845 		kassert_panic("%s: lock (%s) %s can not be upgradable",
846 		    __func__, class->lc_name, lock->lo_name);
847 
848 	/*
849 	 * If we shouldn't watch this lock, then just clear lo_witness.
850 	 * Otherwise, if witness_cold is set, then it is too early to
851 	 * enroll this lock, so defer it to witness_initialize() by adding
852 	 * it to the pending_locks list.  If it is not too early, then enroll
853 	 * the lock now.
854 	 */
855 	if (witness_watch < 1 || panicstr != NULL ||
856 	    (lock->lo_flags & LO_WITNESS) == 0)
857 		lock->lo_witness = NULL;
858 	else if (witness_cold) {
859 		pending_locks[pending_cnt].wh_lock = lock;
860 		pending_locks[pending_cnt++].wh_type = type;
861 		if (pending_cnt > WITNESS_PENDLIST)
862 			panic("%s: pending locks list is too small, "
863 			    "increase WITNESS_PENDLIST\n",
864 			    __func__);
865 	} else
866 		lock->lo_witness = enroll(type, class);
867 }
868 
869 void
870 witness_destroy(struct lock_object *lock)
871 {
872 	struct lock_class *class;
873 	struct witness *w;
874 
875 	class = LOCK_CLASS(lock);
876 
877 	if (witness_cold)
878 		panic("lock (%s) %s destroyed while witness_cold",
879 		    class->lc_name, lock->lo_name);
880 
881 	/* XXX: need to verify that no one holds the lock */
882 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
883 		return;
884 	w = lock->lo_witness;
885 
886 	mtx_lock_spin(&w_mtx);
887 	MPASS(w->w_refcount > 0);
888 	w->w_refcount--;
889 
890 	if (w->w_refcount == 0)
891 		depart(w);
892 	mtx_unlock_spin(&w_mtx);
893 }
894 
895 #ifdef DDB
896 static void
897 witness_ddb_compute_levels(void)
898 {
899 	struct witness *w;
900 
901 	/*
902 	 * First clear all levels.
903 	 */
904 	STAILQ_FOREACH(w, &w_all, w_list)
905 		w->w_ddb_level = -1;
906 
907 	/*
908 	 * Look for locks with no parents and level all their descendants.
909 	 */
910 	STAILQ_FOREACH(w, &w_all, w_list) {
911 
912 		/* If the witness has ancestors (is not a root), skip it. */
913 		if (w->w_num_ancestors > 0)
914 			continue;
915 		witness_ddb_level_descendants(w, 0);
916 	}
917 }
918 
919 static void
920 witness_ddb_level_descendants(struct witness *w, int l)
921 {
922 	int i;
923 
924 	if (w->w_ddb_level >= l)
925 		return;
926 
927 	w->w_ddb_level = l;
928 	l++;
929 
930 	for (i = 1; i <= w_max_used_index; i++) {
931 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
932 			witness_ddb_level_descendants(&w_data[i], l);
933 	}
934 }
935 
936 static void
937 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
938     struct witness *w, int indent)
939 {
940 	int i;
941 
942  	for (i = 0; i < indent; i++)
943  		prnt(" ");
944 	prnt("%s (type: %s, depth: %d, active refs: %d)",
945 	     w->w_name, w->w_class->lc_name,
946 	     w->w_ddb_level, w->w_refcount);
947  	if (w->w_displayed) {
948  		prnt(" -- (already displayed)\n");
949  		return;
950  	}
951  	w->w_displayed = 1;
952 	if (w->w_file != NULL && w->w_line != 0)
953 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
954 		    w->w_line);
955 	else
956 		prnt(" -- never acquired\n");
957 	indent++;
958 	WITNESS_INDEX_ASSERT(w->w_index);
959 	for (i = 1; i <= w_max_used_index; i++) {
960 		if (db_pager_quit)
961 			return;
962 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
963 			witness_ddb_display_descendants(prnt, &w_data[i],
964 			    indent);
965 	}
966 }
967 
968 static void
969 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
970     struct witness_list *list)
971 {
972 	struct witness *w;
973 
974 	STAILQ_FOREACH(w, list, w_typelist) {
975 		if (w->w_file == NULL || w->w_ddb_level > 0)
976 			continue;
977 
978 		/* This lock has no anscestors - display its descendants. */
979 		witness_ddb_display_descendants(prnt, w, 0);
980 		if (db_pager_quit)
981 			return;
982 	}
983 }
984 
985 static void
986 witness_ddb_display(int(*prnt)(const char *fmt, ...))
987 {
988 	struct witness *w;
989 
990 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
991 	witness_ddb_compute_levels();
992 
993 	/* Clear all the displayed flags. */
994 	STAILQ_FOREACH(w, &w_all, w_list)
995 		w->w_displayed = 0;
996 
997 	/*
998 	 * First, handle sleep locks which have been acquired at least
999 	 * once.
1000 	 */
1001 	prnt("Sleep locks:\n");
1002 	witness_ddb_display_list(prnt, &w_sleep);
1003 	if (db_pager_quit)
1004 		return;
1005 
1006 	/*
1007 	 * Now do spin locks which have been acquired at least once.
1008 	 */
1009 	prnt("\nSpin locks:\n");
1010 	witness_ddb_display_list(prnt, &w_spin);
1011 	if (db_pager_quit)
1012 		return;
1013 
1014 	/*
1015 	 * Finally, any locks which have not been acquired yet.
1016 	 */
1017 	prnt("\nLocks which were never acquired:\n");
1018 	STAILQ_FOREACH(w, &w_all, w_list) {
1019 		if (w->w_file != NULL || w->w_refcount == 0)
1020 			continue;
1021 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1022 		    w->w_class->lc_name, w->w_ddb_level);
1023 		if (db_pager_quit)
1024 			return;
1025 	}
1026 }
1027 #endif /* DDB */
1028 
1029 int
1030 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1031 {
1032 
1033 	if (witness_watch == -1 || panicstr != NULL)
1034 		return (0);
1035 
1036 	/* Require locks that witness knows about. */
1037 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1038 	    lock2->lo_witness == NULL)
1039 		return (EINVAL);
1040 
1041 	mtx_assert(&w_mtx, MA_NOTOWNED);
1042 	mtx_lock_spin(&w_mtx);
1043 
1044 	/*
1045 	 * If we already have either an explicit or implied lock order that
1046 	 * is the other way around, then return an error.
1047 	 */
1048 	if (witness_watch &&
1049 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1050 		mtx_unlock_spin(&w_mtx);
1051 		return (EDOOFUS);
1052 	}
1053 
1054 	/* Try to add the new order. */
1055 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1056 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1057 	itismychild(lock1->lo_witness, lock2->lo_witness);
1058 	mtx_unlock_spin(&w_mtx);
1059 	return (0);
1060 }
1061 
1062 void
1063 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1064     int line, struct lock_object *interlock)
1065 {
1066 	struct lock_list_entry *lock_list, *lle;
1067 	struct lock_instance *lock1, *lock2, *plock;
1068 	struct lock_class *class, *iclass;
1069 	struct witness *w, *w1;
1070 	struct thread *td;
1071 	int i, j;
1072 
1073 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1074 	    panicstr != NULL)
1075 		return;
1076 
1077 	w = lock->lo_witness;
1078 	class = LOCK_CLASS(lock);
1079 	td = curthread;
1080 
1081 	if (class->lc_flags & LC_SLEEPLOCK) {
1082 
1083 		/*
1084 		 * Since spin locks include a critical section, this check
1085 		 * implicitly enforces a lock order of all sleep locks before
1086 		 * all spin locks.
1087 		 */
1088 		if (td->td_critnest != 0 && !kdb_active)
1089 			kassert_panic("acquiring blockable sleep lock with "
1090 			    "spinlock or critical section held (%s) %s @ %s:%d",
1091 			    class->lc_name, lock->lo_name,
1092 			    fixup_filename(file), line);
1093 
1094 		/*
1095 		 * If this is the first lock acquired then just return as
1096 		 * no order checking is needed.
1097 		 */
1098 		lock_list = td->td_sleeplocks;
1099 		if (lock_list == NULL || lock_list->ll_count == 0)
1100 			return;
1101 	} else {
1102 
1103 		/*
1104 		 * If this is the first lock, just return as no order
1105 		 * checking is needed.  Avoid problems with thread
1106 		 * migration pinning the thread while checking if
1107 		 * spinlocks are held.  If at least one spinlock is held
1108 		 * the thread is in a safe path and it is allowed to
1109 		 * unpin it.
1110 		 */
1111 		sched_pin();
1112 		lock_list = PCPU_GET(spinlocks);
1113 		if (lock_list == NULL || lock_list->ll_count == 0) {
1114 			sched_unpin();
1115 			return;
1116 		}
1117 		sched_unpin();
1118 	}
1119 
1120 	/*
1121 	 * Check to see if we are recursing on a lock we already own.  If
1122 	 * so, make sure that we don't mismatch exclusive and shared lock
1123 	 * acquires.
1124 	 */
1125 	lock1 = find_instance(lock_list, lock);
1126 	if (lock1 != NULL) {
1127 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1128 		    (flags & LOP_EXCLUSIVE) == 0) {
1129 			witness_output("shared lock of (%s) %s @ %s:%d\n",
1130 			    class->lc_name, lock->lo_name,
1131 			    fixup_filename(file), line);
1132 			witness_output("while exclusively locked from %s:%d\n",
1133 			    fixup_filename(lock1->li_file), lock1->li_line);
1134 			kassert_panic("excl->share");
1135 		}
1136 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1137 		    (flags & LOP_EXCLUSIVE) != 0) {
1138 			witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1139 			    class->lc_name, lock->lo_name,
1140 			    fixup_filename(file), line);
1141 			witness_output("while share locked from %s:%d\n",
1142 			    fixup_filename(lock1->li_file), lock1->li_line);
1143 			kassert_panic("share->excl");
1144 		}
1145 		return;
1146 	}
1147 
1148 	/* Warn if the interlock is not locked exactly once. */
1149 	if (interlock != NULL) {
1150 		iclass = LOCK_CLASS(interlock);
1151 		lock1 = find_instance(lock_list, interlock);
1152 		if (lock1 == NULL)
1153 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1154 			    iclass->lc_name, interlock->lo_name,
1155 			    fixup_filename(file), line);
1156 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1157 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1158 			    iclass->lc_name, interlock->lo_name,
1159 			    fixup_filename(file), line);
1160 	}
1161 
1162 	/*
1163 	 * Find the previously acquired lock, but ignore interlocks.
1164 	 */
1165 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1166 	if (interlock != NULL && plock->li_lock == interlock) {
1167 		if (lock_list->ll_count > 1)
1168 			plock =
1169 			    &lock_list->ll_children[lock_list->ll_count - 2];
1170 		else {
1171 			lle = lock_list->ll_next;
1172 
1173 			/*
1174 			 * The interlock is the only lock we hold, so
1175 			 * simply return.
1176 			 */
1177 			if (lle == NULL)
1178 				return;
1179 			plock = &lle->ll_children[lle->ll_count - 1];
1180 		}
1181 	}
1182 
1183 	/*
1184 	 * Try to perform most checks without a lock.  If this succeeds we
1185 	 * can skip acquiring the lock and return success.  Otherwise we redo
1186 	 * the check with the lock held to handle races with concurrent updates.
1187 	 */
1188 	w1 = plock->li_lock->lo_witness;
1189 	if (witness_lock_order_check(w1, w))
1190 		return;
1191 
1192 	mtx_lock_spin(&w_mtx);
1193 	if (witness_lock_order_check(w1, w)) {
1194 		mtx_unlock_spin(&w_mtx);
1195 		return;
1196 	}
1197 	witness_lock_order_add(w1, w);
1198 
1199 	/*
1200 	 * Check for duplicate locks of the same type.  Note that we only
1201 	 * have to check for this on the last lock we just acquired.  Any
1202 	 * other cases will be caught as lock order violations.
1203 	 */
1204 	if (w1 == w) {
1205 		i = w->w_index;
1206 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1207 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1208 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1209 			w->w_reversed = 1;
1210 			mtx_unlock_spin(&w_mtx);
1211 			witness_output(
1212 			    "acquiring duplicate lock of same type: \"%s\"\n",
1213 			    w->w_name);
1214 			witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1215 			    fixup_filename(plock->li_file), plock->li_line);
1216 			witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1217 			    fixup_filename(file), line);
1218 			witness_debugger(1, __func__);
1219 		} else
1220 			mtx_unlock_spin(&w_mtx);
1221 		return;
1222 	}
1223 	mtx_assert(&w_mtx, MA_OWNED);
1224 
1225 	/*
1226 	 * If we know that the lock we are acquiring comes after
1227 	 * the lock we most recently acquired in the lock order tree,
1228 	 * then there is no need for any further checks.
1229 	 */
1230 	if (isitmychild(w1, w))
1231 		goto out;
1232 
1233 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1234 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1235 
1236 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1237 			lock1 = &lle->ll_children[i];
1238 
1239 			/*
1240 			 * Ignore the interlock.
1241 			 */
1242 			if (interlock == lock1->li_lock)
1243 				continue;
1244 
1245 			/*
1246 			 * If this lock doesn't undergo witness checking,
1247 			 * then skip it.
1248 			 */
1249 			w1 = lock1->li_lock->lo_witness;
1250 			if (w1 == NULL) {
1251 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1252 				    ("lock missing witness structure"));
1253 				continue;
1254 			}
1255 
1256 			/*
1257 			 * If we are locking Giant and this is a sleepable
1258 			 * lock, then skip it.
1259 			 */
1260 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1261 			    lock == &Giant.lock_object)
1262 				continue;
1263 
1264 			/*
1265 			 * If we are locking a sleepable lock and this lock
1266 			 * is Giant, then skip it.
1267 			 */
1268 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1269 			    lock1->li_lock == &Giant.lock_object)
1270 				continue;
1271 
1272 			/*
1273 			 * If we are locking a sleepable lock and this lock
1274 			 * isn't sleepable, we want to treat it as a lock
1275 			 * order violation to enfore a general lock order of
1276 			 * sleepable locks before non-sleepable locks.
1277 			 */
1278 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1279 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1280 				goto reversal;
1281 
1282 			/*
1283 			 * If we are locking Giant and this is a non-sleepable
1284 			 * lock, then treat it as a reversal.
1285 			 */
1286 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1287 			    lock == &Giant.lock_object)
1288 				goto reversal;
1289 
1290 			/*
1291 			 * Check the lock order hierarchy for a reveresal.
1292 			 */
1293 			if (!isitmydescendant(w, w1))
1294 				continue;
1295 		reversal:
1296 
1297 			/*
1298 			 * We have a lock order violation, check to see if it
1299 			 * is allowed or has already been yelled about.
1300 			 */
1301 #ifdef BLESSING
1302 
1303 			/*
1304 			 * If the lock order is blessed, just bail.  We don't
1305 			 * look for other lock order violations though, which
1306 			 * may be a bug.
1307 			 */
1308 			if (blessed(w, w1))
1309 				goto out;
1310 #endif
1311 
1312 			/* Bail if this violation is known */
1313 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1314 				goto out;
1315 
1316 			/* Record this as a violation */
1317 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1318 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1319 			w->w_reversed = w1->w_reversed = 1;
1320 			witness_increment_graph_generation();
1321 			mtx_unlock_spin(&w_mtx);
1322 
1323 #ifdef WITNESS_NO_VNODE
1324 			/*
1325 			 * There are known LORs between VNODE locks. They are
1326 			 * not an indication of a bug. VNODE locks are flagged
1327 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1328 			 * is between 2 VNODE locks.
1329 			 */
1330 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1331 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1332 				return;
1333 #endif
1334 
1335 			/*
1336 			 * Ok, yell about it.
1337 			 */
1338 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1339 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1340 				witness_output(
1341 		"lock order reversal: (sleepable after non-sleepable)\n");
1342 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1343 			    && lock == &Giant.lock_object)
1344 				witness_output(
1345 		"lock order reversal: (Giant after non-sleepable)\n");
1346 			else
1347 				witness_output("lock order reversal:\n");
1348 
1349 			/*
1350 			 * Try to locate an earlier lock with
1351 			 * witness w in our list.
1352 			 */
1353 			do {
1354 				lock2 = &lle->ll_children[i];
1355 				MPASS(lock2->li_lock != NULL);
1356 				if (lock2->li_lock->lo_witness == w)
1357 					break;
1358 				if (i == 0 && lle->ll_next != NULL) {
1359 					lle = lle->ll_next;
1360 					i = lle->ll_count - 1;
1361 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1362 				} else
1363 					i--;
1364 			} while (i >= 0);
1365 			if (i < 0) {
1366 				witness_output(" 1st %p %s (%s) @ %s:%d\n",
1367 				    lock1->li_lock, lock1->li_lock->lo_name,
1368 				    w1->w_name, fixup_filename(lock1->li_file),
1369 				    lock1->li_line);
1370 				witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1371 				    lock->lo_name, w->w_name,
1372 				    fixup_filename(file), line);
1373 			} else {
1374 				witness_output(" 1st %p %s (%s) @ %s:%d\n",
1375 				    lock2->li_lock, lock2->li_lock->lo_name,
1376 				    lock2->li_lock->lo_witness->w_name,
1377 				    fixup_filename(lock2->li_file),
1378 				    lock2->li_line);
1379 				witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1380 				    lock1->li_lock, lock1->li_lock->lo_name,
1381 				    w1->w_name, fixup_filename(lock1->li_file),
1382 				    lock1->li_line);
1383 				witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1384 				    lock->lo_name, w->w_name,
1385 				    fixup_filename(file), line);
1386 			}
1387 			witness_debugger(1, __func__);
1388 			return;
1389 		}
1390 	}
1391 
1392 	/*
1393 	 * If requested, build a new lock order.  However, don't build a new
1394 	 * relationship between a sleepable lock and Giant if it is in the
1395 	 * wrong direction.  The correct lock order is that sleepable locks
1396 	 * always come before Giant.
1397 	 */
1398 	if (flags & LOP_NEWORDER &&
1399 	    !(plock->li_lock == &Giant.lock_object &&
1400 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1401 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1402 		    w->w_name, plock->li_lock->lo_witness->w_name);
1403 		itismychild(plock->li_lock->lo_witness, w);
1404 	}
1405 out:
1406 	mtx_unlock_spin(&w_mtx);
1407 }
1408 
1409 void
1410 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1411 {
1412 	struct lock_list_entry **lock_list, *lle;
1413 	struct lock_instance *instance;
1414 	struct witness *w;
1415 	struct thread *td;
1416 
1417 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1418 	    panicstr != NULL)
1419 		return;
1420 	w = lock->lo_witness;
1421 	td = curthread;
1422 
1423 	/* Determine lock list for this lock. */
1424 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1425 		lock_list = &td->td_sleeplocks;
1426 	else
1427 		lock_list = PCPU_PTR(spinlocks);
1428 
1429 	/* Check to see if we are recursing on a lock we already own. */
1430 	instance = find_instance(*lock_list, lock);
1431 	if (instance != NULL) {
1432 		instance->li_flags++;
1433 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1434 		    td->td_proc->p_pid, lock->lo_name,
1435 		    instance->li_flags & LI_RECURSEMASK);
1436 		instance->li_file = file;
1437 		instance->li_line = line;
1438 		return;
1439 	}
1440 
1441 	/* Update per-witness last file and line acquire. */
1442 	w->w_file = file;
1443 	w->w_line = line;
1444 
1445 	/* Find the next open lock instance in the list and fill it. */
1446 	lle = *lock_list;
1447 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1448 		lle = witness_lock_list_get();
1449 		if (lle == NULL)
1450 			return;
1451 		lle->ll_next = *lock_list;
1452 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1453 		    td->td_proc->p_pid, lle);
1454 		*lock_list = lle;
1455 	}
1456 	instance = &lle->ll_children[lle->ll_count++];
1457 	instance->li_lock = lock;
1458 	instance->li_line = line;
1459 	instance->li_file = file;
1460 	if ((flags & LOP_EXCLUSIVE) != 0)
1461 		instance->li_flags = LI_EXCLUSIVE;
1462 	else
1463 		instance->li_flags = 0;
1464 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1465 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1466 }
1467 
1468 void
1469 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1470 {
1471 	struct lock_instance *instance;
1472 	struct lock_class *class;
1473 
1474 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1475 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1476 		return;
1477 	class = LOCK_CLASS(lock);
1478 	if (witness_watch) {
1479 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1480 			kassert_panic(
1481 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1482 			    class->lc_name, lock->lo_name,
1483 			    fixup_filename(file), line);
1484 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1485 			kassert_panic(
1486 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1487 			    class->lc_name, lock->lo_name,
1488 			    fixup_filename(file), line);
1489 	}
1490 	instance = find_instance(curthread->td_sleeplocks, lock);
1491 	if (instance == NULL) {
1492 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1493 		    class->lc_name, lock->lo_name,
1494 		    fixup_filename(file), line);
1495 		return;
1496 	}
1497 	if (witness_watch) {
1498 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1499 			kassert_panic(
1500 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1501 			    class->lc_name, lock->lo_name,
1502 			    fixup_filename(file), line);
1503 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1504 			kassert_panic(
1505 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1506 			    class->lc_name, lock->lo_name,
1507 			    instance->li_flags & LI_RECURSEMASK,
1508 			    fixup_filename(file), line);
1509 	}
1510 	instance->li_flags |= LI_EXCLUSIVE;
1511 }
1512 
1513 void
1514 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1515     int line)
1516 {
1517 	struct lock_instance *instance;
1518 	struct lock_class *class;
1519 
1520 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1521 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1522 		return;
1523 	class = LOCK_CLASS(lock);
1524 	if (witness_watch) {
1525 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1526 			kassert_panic(
1527 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1528 			    class->lc_name, lock->lo_name,
1529 			    fixup_filename(file), line);
1530 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1531 			kassert_panic(
1532 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1533 			    class->lc_name, lock->lo_name,
1534 			    fixup_filename(file), line);
1535 	}
1536 	instance = find_instance(curthread->td_sleeplocks, lock);
1537 	if (instance == NULL) {
1538 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1539 		    class->lc_name, lock->lo_name,
1540 		    fixup_filename(file), line);
1541 		return;
1542 	}
1543 	if (witness_watch) {
1544 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1545 			kassert_panic(
1546 			    "downgrade of shared lock (%s) %s @ %s:%d",
1547 			    class->lc_name, lock->lo_name,
1548 			    fixup_filename(file), line);
1549 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1550 			kassert_panic(
1551 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1552 			    class->lc_name, lock->lo_name,
1553 			    instance->li_flags & LI_RECURSEMASK,
1554 			    fixup_filename(file), line);
1555 	}
1556 	instance->li_flags &= ~LI_EXCLUSIVE;
1557 }
1558 
1559 void
1560 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1561 {
1562 	struct lock_list_entry **lock_list, *lle;
1563 	struct lock_instance *instance;
1564 	struct lock_class *class;
1565 	struct thread *td;
1566 	register_t s;
1567 	int i, j;
1568 
1569 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1570 		return;
1571 	td = curthread;
1572 	class = LOCK_CLASS(lock);
1573 
1574 	/* Find lock instance associated with this lock. */
1575 	if (class->lc_flags & LC_SLEEPLOCK)
1576 		lock_list = &td->td_sleeplocks;
1577 	else
1578 		lock_list = PCPU_PTR(spinlocks);
1579 	lle = *lock_list;
1580 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1581 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1582 			instance = &(*lock_list)->ll_children[i];
1583 			if (instance->li_lock == lock)
1584 				goto found;
1585 		}
1586 
1587 	/*
1588 	 * When disabling WITNESS through witness_watch we could end up in
1589 	 * having registered locks in the td_sleeplocks queue.
1590 	 * We have to make sure we flush these queues, so just search for
1591 	 * eventual register locks and remove them.
1592 	 */
1593 	if (witness_watch > 0) {
1594 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1595 		    lock->lo_name, fixup_filename(file), line);
1596 		return;
1597 	} else {
1598 		return;
1599 	}
1600 found:
1601 
1602 	/* First, check for shared/exclusive mismatches. */
1603 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1604 	    (flags & LOP_EXCLUSIVE) == 0) {
1605 		witness_output("shared unlock of (%s) %s @ %s:%d\n",
1606 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1607 		witness_output("while exclusively locked from %s:%d\n",
1608 		    fixup_filename(instance->li_file), instance->li_line);
1609 		kassert_panic("excl->ushare");
1610 	}
1611 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1612 	    (flags & LOP_EXCLUSIVE) != 0) {
1613 		witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1614 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1615 		witness_output("while share locked from %s:%d\n",
1616 		    fixup_filename(instance->li_file),
1617 		    instance->li_line);
1618 		kassert_panic("share->uexcl");
1619 	}
1620 	/* If we are recursed, unrecurse. */
1621 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1622 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1623 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1624 		    instance->li_flags);
1625 		instance->li_flags--;
1626 		return;
1627 	}
1628 	/* The lock is now being dropped, check for NORELEASE flag */
1629 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1630 		witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1631 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1632 		kassert_panic("lock marked norelease");
1633 	}
1634 
1635 	/* Otherwise, remove this item from the list. */
1636 	s = intr_disable();
1637 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1638 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1639 	    (*lock_list)->ll_count - 1);
1640 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1641 		(*lock_list)->ll_children[j] =
1642 		    (*lock_list)->ll_children[j + 1];
1643 	(*lock_list)->ll_count--;
1644 	intr_restore(s);
1645 
1646 	/*
1647 	 * In order to reduce contention on w_mtx, we want to keep always an
1648 	 * head object into lists so that frequent allocation from the
1649 	 * free witness pool (and subsequent locking) is avoided.
1650 	 * In order to maintain the current code simple, when the head
1651 	 * object is totally unloaded it means also that we do not have
1652 	 * further objects in the list, so the list ownership needs to be
1653 	 * hand over to another object if the current head needs to be freed.
1654 	 */
1655 	if ((*lock_list)->ll_count == 0) {
1656 		if (*lock_list == lle) {
1657 			if (lle->ll_next == NULL)
1658 				return;
1659 		} else
1660 			lle = *lock_list;
1661 		*lock_list = lle->ll_next;
1662 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1663 		    td->td_proc->p_pid, lle);
1664 		witness_lock_list_free(lle);
1665 	}
1666 }
1667 
1668 void
1669 witness_thread_exit(struct thread *td)
1670 {
1671 	struct lock_list_entry *lle;
1672 	int i, n;
1673 
1674 	lle = td->td_sleeplocks;
1675 	if (lle == NULL || panicstr != NULL)
1676 		return;
1677 	if (lle->ll_count != 0) {
1678 		for (n = 0; lle != NULL; lle = lle->ll_next)
1679 			for (i = lle->ll_count - 1; i >= 0; i--) {
1680 				if (n == 0)
1681 					witness_output(
1682 		    "Thread %p exiting with the following locks held:\n", td);
1683 				n++;
1684 				witness_list_lock(&lle->ll_children[i],
1685 				    witness_output);
1686 
1687 			}
1688 		kassert_panic(
1689 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1690 	}
1691 	witness_lock_list_free(lle);
1692 }
1693 
1694 /*
1695  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1696  * exempt Giant and sleepable locks from the checks as well.  If any
1697  * non-exempt locks are held, then a supplied message is printed to the
1698  * output channel along with a list of the offending locks.  If indicated in the
1699  * flags then a failure results in a panic as well.
1700  */
1701 int
1702 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1703 {
1704 	struct lock_list_entry *lock_list, *lle;
1705 	struct lock_instance *lock1;
1706 	struct thread *td;
1707 	va_list ap;
1708 	int i, n;
1709 
1710 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1711 		return (0);
1712 	n = 0;
1713 	td = curthread;
1714 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1715 		for (i = lle->ll_count - 1; i >= 0; i--) {
1716 			lock1 = &lle->ll_children[i];
1717 			if (lock1->li_lock == lock)
1718 				continue;
1719 			if (flags & WARN_GIANTOK &&
1720 			    lock1->li_lock == &Giant.lock_object)
1721 				continue;
1722 			if (flags & WARN_SLEEPOK &&
1723 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1724 				continue;
1725 			if (n == 0) {
1726 				va_start(ap, fmt);
1727 				witness_voutput(fmt, ap);
1728 				va_end(ap);
1729 				witness_output(
1730 				    " with the following %slocks held:\n",
1731 				    (flags & WARN_SLEEPOK) != 0 ?
1732 				    "non-sleepable " : "");
1733 			}
1734 			n++;
1735 			witness_list_lock(lock1, witness_output);
1736 		}
1737 
1738 	/*
1739 	 * Pin the thread in order to avoid problems with thread migration.
1740 	 * Once that all verifies are passed about spinlocks ownership,
1741 	 * the thread is in a safe path and it can be unpinned.
1742 	 */
1743 	sched_pin();
1744 	lock_list = PCPU_GET(spinlocks);
1745 	if (lock_list != NULL && lock_list->ll_count != 0) {
1746 		sched_unpin();
1747 
1748 		/*
1749 		 * We should only have one spinlock and as long as
1750 		 * the flags cannot match for this locks class,
1751 		 * check if the first spinlock is the one curthread
1752 		 * should hold.
1753 		 */
1754 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1755 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1756 		    lock1->li_lock == lock && n == 0)
1757 			return (0);
1758 
1759 		va_start(ap, fmt);
1760 		witness_voutput(fmt, ap);
1761 		va_end(ap);
1762 		witness_output(" with the following %slocks held:\n",
1763 		    (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1764 		n += witness_list_locks(&lock_list, witness_output);
1765 	} else
1766 		sched_unpin();
1767 	if (flags & WARN_PANIC && n)
1768 		kassert_panic("%s", __func__);
1769 	else
1770 		witness_debugger(n, __func__);
1771 	return (n);
1772 }
1773 
1774 const char *
1775 witness_file(struct lock_object *lock)
1776 {
1777 	struct witness *w;
1778 
1779 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1780 		return ("?");
1781 	w = lock->lo_witness;
1782 	return (w->w_file);
1783 }
1784 
1785 int
1786 witness_line(struct lock_object *lock)
1787 {
1788 	struct witness *w;
1789 
1790 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1791 		return (0);
1792 	w = lock->lo_witness;
1793 	return (w->w_line);
1794 }
1795 
1796 static struct witness *
1797 enroll(const char *description, struct lock_class *lock_class)
1798 {
1799 	struct witness *w;
1800 	struct witness_list *typelist;
1801 
1802 	MPASS(description != NULL);
1803 
1804 	if (witness_watch == -1 || panicstr != NULL)
1805 		return (NULL);
1806 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1807 		if (witness_skipspin)
1808 			return (NULL);
1809 		else
1810 			typelist = &w_spin;
1811 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1812 		typelist = &w_sleep;
1813 	} else {
1814 		kassert_panic("lock class %s is not sleep or spin",
1815 		    lock_class->lc_name);
1816 		return (NULL);
1817 	}
1818 
1819 	mtx_lock_spin(&w_mtx);
1820 	w = witness_hash_get(description);
1821 	if (w)
1822 		goto found;
1823 	if ((w = witness_get()) == NULL)
1824 		return (NULL);
1825 	MPASS(strlen(description) < MAX_W_NAME);
1826 	strcpy(w->w_name, description);
1827 	w->w_class = lock_class;
1828 	w->w_refcount = 1;
1829 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1830 	if (lock_class->lc_flags & LC_SPINLOCK) {
1831 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1832 		w_spin_cnt++;
1833 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1834 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1835 		w_sleep_cnt++;
1836 	}
1837 
1838 	/* Insert new witness into the hash */
1839 	witness_hash_put(w);
1840 	witness_increment_graph_generation();
1841 	mtx_unlock_spin(&w_mtx);
1842 	return (w);
1843 found:
1844 	w->w_refcount++;
1845 	mtx_unlock_spin(&w_mtx);
1846 	if (lock_class != w->w_class)
1847 		kassert_panic(
1848 			"lock (%s) %s does not match earlier (%s) lock",
1849 			description, lock_class->lc_name,
1850 			w->w_class->lc_name);
1851 	return (w);
1852 }
1853 
1854 static void
1855 depart(struct witness *w)
1856 {
1857 	struct witness_list *list;
1858 
1859 	MPASS(w->w_refcount == 0);
1860 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1861 		list = &w_sleep;
1862 		w_sleep_cnt--;
1863 	} else {
1864 		list = &w_spin;
1865 		w_spin_cnt--;
1866 	}
1867 	/*
1868 	 * Set file to NULL as it may point into a loadable module.
1869 	 */
1870 	w->w_file = NULL;
1871 	w->w_line = 0;
1872 	witness_increment_graph_generation();
1873 }
1874 
1875 
1876 static void
1877 adopt(struct witness *parent, struct witness *child)
1878 {
1879 	int pi, ci, i, j;
1880 
1881 	if (witness_cold == 0)
1882 		mtx_assert(&w_mtx, MA_OWNED);
1883 
1884 	/* If the relationship is already known, there's no work to be done. */
1885 	if (isitmychild(parent, child))
1886 		return;
1887 
1888 	/* When the structure of the graph changes, bump up the generation. */
1889 	witness_increment_graph_generation();
1890 
1891 	/*
1892 	 * The hard part ... create the direct relationship, then propagate all
1893 	 * indirect relationships.
1894 	 */
1895 	pi = parent->w_index;
1896 	ci = child->w_index;
1897 	WITNESS_INDEX_ASSERT(pi);
1898 	WITNESS_INDEX_ASSERT(ci);
1899 	MPASS(pi != ci);
1900 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1901 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1902 
1903 	/*
1904 	 * If parent was not already an ancestor of child,
1905 	 * then we increment the descendant and ancestor counters.
1906 	 */
1907 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1908 		parent->w_num_descendants++;
1909 		child->w_num_ancestors++;
1910 	}
1911 
1912 	/*
1913 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1914 	 * an ancestor of 'pi' during this loop.
1915 	 */
1916 	for (i = 1; i <= w_max_used_index; i++) {
1917 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1918 		    (i != pi))
1919 			continue;
1920 
1921 		/* Find each descendant of 'i' and mark it as a descendant. */
1922 		for (j = 1; j <= w_max_used_index; j++) {
1923 
1924 			/*
1925 			 * Skip children that are already marked as
1926 			 * descendants of 'i'.
1927 			 */
1928 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1929 				continue;
1930 
1931 			/*
1932 			 * We are only interested in descendants of 'ci'. Note
1933 			 * that 'ci' itself is counted as a descendant of 'ci'.
1934 			 */
1935 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1936 			    (j != ci))
1937 				continue;
1938 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1939 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1940 			w_data[i].w_num_descendants++;
1941 			w_data[j].w_num_ancestors++;
1942 
1943 			/*
1944 			 * Make sure we aren't marking a node as both an
1945 			 * ancestor and descendant. We should have caught
1946 			 * this as a lock order reversal earlier.
1947 			 */
1948 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1949 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1950 				printf("witness rmatrix paradox! [%d][%d]=%d "
1951 				    "both ancestor and descendant\n",
1952 				    i, j, w_rmatrix[i][j]);
1953 				kdb_backtrace();
1954 				printf("Witness disabled.\n");
1955 				witness_watch = -1;
1956 			}
1957 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1958 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1959 				printf("witness rmatrix paradox! [%d][%d]=%d "
1960 				    "both ancestor and descendant\n",
1961 				    j, i, w_rmatrix[j][i]);
1962 				kdb_backtrace();
1963 				printf("Witness disabled.\n");
1964 				witness_watch = -1;
1965 			}
1966 		}
1967 	}
1968 }
1969 
1970 static void
1971 itismychild(struct witness *parent, struct witness *child)
1972 {
1973 	int unlocked;
1974 
1975 	MPASS(child != NULL && parent != NULL);
1976 	if (witness_cold == 0)
1977 		mtx_assert(&w_mtx, MA_OWNED);
1978 
1979 	if (!witness_lock_type_equal(parent, child)) {
1980 		if (witness_cold == 0) {
1981 			unlocked = 1;
1982 			mtx_unlock_spin(&w_mtx);
1983 		} else {
1984 			unlocked = 0;
1985 		}
1986 		kassert_panic(
1987 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1988 		    "the same lock type", __func__, parent->w_name,
1989 		    parent->w_class->lc_name, child->w_name,
1990 		    child->w_class->lc_name);
1991 		if (unlocked)
1992 			mtx_lock_spin(&w_mtx);
1993 	}
1994 	adopt(parent, child);
1995 }
1996 
1997 /*
1998  * Generic code for the isitmy*() functions. The rmask parameter is the
1999  * expected relationship of w1 to w2.
2000  */
2001 static int
2002 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2003 {
2004 	unsigned char r1, r2;
2005 	int i1, i2;
2006 
2007 	i1 = w1->w_index;
2008 	i2 = w2->w_index;
2009 	WITNESS_INDEX_ASSERT(i1);
2010 	WITNESS_INDEX_ASSERT(i2);
2011 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2012 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2013 
2014 	/* The flags on one better be the inverse of the flags on the other */
2015 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2016 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2017 		/* Don't squawk if we're potentially racing with an update. */
2018 		if (!mtx_owned(&w_mtx))
2019 			return (0);
2020 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2021 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2022 		    "w_rmatrix[%d][%d] == %hhx\n",
2023 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2024 		    i2, i1, r2);
2025 		kdb_backtrace();
2026 		printf("Witness disabled.\n");
2027 		witness_watch = -1;
2028 	}
2029 	return (r1 & rmask);
2030 }
2031 
2032 /*
2033  * Checks if @child is a direct child of @parent.
2034  */
2035 static int
2036 isitmychild(struct witness *parent, struct witness *child)
2037 {
2038 
2039 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2040 }
2041 
2042 /*
2043  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2044  */
2045 static int
2046 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2047 {
2048 
2049 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2050 	    __func__));
2051 }
2052 
2053 #ifdef BLESSING
2054 static int
2055 blessed(struct witness *w1, struct witness *w2)
2056 {
2057 	int i;
2058 	struct witness_blessed *b;
2059 
2060 	for (i = 0; i < nitems(blessed_list); i++) {
2061 		b = &blessed_list[i];
2062 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2063 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2064 				return (1);
2065 			continue;
2066 		}
2067 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2068 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2069 				return (1);
2070 	}
2071 	return (0);
2072 }
2073 #endif
2074 
2075 static struct witness *
2076 witness_get(void)
2077 {
2078 	struct witness *w;
2079 	int index;
2080 
2081 	if (witness_cold == 0)
2082 		mtx_assert(&w_mtx, MA_OWNED);
2083 
2084 	if (witness_watch == -1) {
2085 		mtx_unlock_spin(&w_mtx);
2086 		return (NULL);
2087 	}
2088 	if (STAILQ_EMPTY(&w_free)) {
2089 		witness_watch = -1;
2090 		mtx_unlock_spin(&w_mtx);
2091 		printf("WITNESS: unable to allocate a new witness object\n");
2092 		return (NULL);
2093 	}
2094 	w = STAILQ_FIRST(&w_free);
2095 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2096 	w_free_cnt--;
2097 	index = w->w_index;
2098 	MPASS(index > 0 && index == w_max_used_index+1 &&
2099 	    index < witness_count);
2100 	bzero(w, sizeof(*w));
2101 	w->w_index = index;
2102 	if (index > w_max_used_index)
2103 		w_max_used_index = index;
2104 	return (w);
2105 }
2106 
2107 static void
2108 witness_free(struct witness *w)
2109 {
2110 
2111 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2112 	w_free_cnt++;
2113 }
2114 
2115 static struct lock_list_entry *
2116 witness_lock_list_get(void)
2117 {
2118 	struct lock_list_entry *lle;
2119 
2120 	if (witness_watch == -1)
2121 		return (NULL);
2122 	mtx_lock_spin(&w_mtx);
2123 	lle = w_lock_list_free;
2124 	if (lle == NULL) {
2125 		witness_watch = -1;
2126 		mtx_unlock_spin(&w_mtx);
2127 		printf("%s: witness exhausted\n", __func__);
2128 		return (NULL);
2129 	}
2130 	w_lock_list_free = lle->ll_next;
2131 	mtx_unlock_spin(&w_mtx);
2132 	bzero(lle, sizeof(*lle));
2133 	return (lle);
2134 }
2135 
2136 static void
2137 witness_lock_list_free(struct lock_list_entry *lle)
2138 {
2139 
2140 	mtx_lock_spin(&w_mtx);
2141 	lle->ll_next = w_lock_list_free;
2142 	w_lock_list_free = lle;
2143 	mtx_unlock_spin(&w_mtx);
2144 }
2145 
2146 static struct lock_instance *
2147 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2148 {
2149 	struct lock_list_entry *lle;
2150 	struct lock_instance *instance;
2151 	int i;
2152 
2153 	for (lle = list; lle != NULL; lle = lle->ll_next)
2154 		for (i = lle->ll_count - 1; i >= 0; i--) {
2155 			instance = &lle->ll_children[i];
2156 			if (instance->li_lock == lock)
2157 				return (instance);
2158 		}
2159 	return (NULL);
2160 }
2161 
2162 static void
2163 witness_list_lock(struct lock_instance *instance,
2164     int (*prnt)(const char *fmt, ...))
2165 {
2166 	struct lock_object *lock;
2167 
2168 	lock = instance->li_lock;
2169 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2170 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2171 	if (lock->lo_witness->w_name != lock->lo_name)
2172 		prnt(" (%s)", lock->lo_witness->w_name);
2173 	prnt(" r = %d (%p) locked @ %s:%d\n",
2174 	    instance->li_flags & LI_RECURSEMASK, lock,
2175 	    fixup_filename(instance->li_file), instance->li_line);
2176 }
2177 
2178 static int
2179 witness_output(const char *fmt, ...)
2180 {
2181 	va_list ap;
2182 	int ret;
2183 
2184 	va_start(ap, fmt);
2185 	ret = witness_voutput(fmt, ap);
2186 	va_end(ap);
2187 	return (ret);
2188 }
2189 
2190 static int
2191 witness_voutput(const char *fmt, va_list ap)
2192 {
2193 	int ret;
2194 
2195 	ret = 0;
2196 	switch (witness_channel) {
2197 	case WITNESS_CONSOLE:
2198 		ret = vprintf(fmt, ap);
2199 		break;
2200 	case WITNESS_LOG:
2201 		vlog(LOG_NOTICE, fmt, ap);
2202 		break;
2203 	case WITNESS_NONE:
2204 		break;
2205 	}
2206 	return (ret);
2207 }
2208 
2209 #ifdef DDB
2210 static int
2211 witness_thread_has_locks(struct thread *td)
2212 {
2213 
2214 	if (td->td_sleeplocks == NULL)
2215 		return (0);
2216 	return (td->td_sleeplocks->ll_count != 0);
2217 }
2218 
2219 static int
2220 witness_proc_has_locks(struct proc *p)
2221 {
2222 	struct thread *td;
2223 
2224 	FOREACH_THREAD_IN_PROC(p, td) {
2225 		if (witness_thread_has_locks(td))
2226 			return (1);
2227 	}
2228 	return (0);
2229 }
2230 #endif
2231 
2232 int
2233 witness_list_locks(struct lock_list_entry **lock_list,
2234     int (*prnt)(const char *fmt, ...))
2235 {
2236 	struct lock_list_entry *lle;
2237 	int i, nheld;
2238 
2239 	nheld = 0;
2240 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2241 		for (i = lle->ll_count - 1; i >= 0; i--) {
2242 			witness_list_lock(&lle->ll_children[i], prnt);
2243 			nheld++;
2244 		}
2245 	return (nheld);
2246 }
2247 
2248 /*
2249  * This is a bit risky at best.  We call this function when we have timed
2250  * out acquiring a spin lock, and we assume that the other CPU is stuck
2251  * with this lock held.  So, we go groveling around in the other CPU's
2252  * per-cpu data to try to find the lock instance for this spin lock to
2253  * see when it was last acquired.
2254  */
2255 void
2256 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2257     int (*prnt)(const char *fmt, ...))
2258 {
2259 	struct lock_instance *instance;
2260 	struct pcpu *pc;
2261 
2262 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2263 		return;
2264 	pc = pcpu_find(owner->td_oncpu);
2265 	instance = find_instance(pc->pc_spinlocks, lock);
2266 	if (instance != NULL)
2267 		witness_list_lock(instance, prnt);
2268 }
2269 
2270 void
2271 witness_save(struct lock_object *lock, const char **filep, int *linep)
2272 {
2273 	struct lock_list_entry *lock_list;
2274 	struct lock_instance *instance;
2275 	struct lock_class *class;
2276 
2277 	/*
2278 	 * This function is used independently in locking code to deal with
2279 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2280 	 * is gone.
2281 	 */
2282 	if (SCHEDULER_STOPPED())
2283 		return;
2284 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2285 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2286 		return;
2287 	class = LOCK_CLASS(lock);
2288 	if (class->lc_flags & LC_SLEEPLOCK)
2289 		lock_list = curthread->td_sleeplocks;
2290 	else {
2291 		if (witness_skipspin)
2292 			return;
2293 		lock_list = PCPU_GET(spinlocks);
2294 	}
2295 	instance = find_instance(lock_list, lock);
2296 	if (instance == NULL) {
2297 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2298 		    class->lc_name, lock->lo_name);
2299 		return;
2300 	}
2301 	*filep = instance->li_file;
2302 	*linep = instance->li_line;
2303 }
2304 
2305 void
2306 witness_restore(struct lock_object *lock, const char *file, int line)
2307 {
2308 	struct lock_list_entry *lock_list;
2309 	struct lock_instance *instance;
2310 	struct lock_class *class;
2311 
2312 	/*
2313 	 * This function is used independently in locking code to deal with
2314 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2315 	 * is gone.
2316 	 */
2317 	if (SCHEDULER_STOPPED())
2318 		return;
2319 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2320 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2321 		return;
2322 	class = LOCK_CLASS(lock);
2323 	if (class->lc_flags & LC_SLEEPLOCK)
2324 		lock_list = curthread->td_sleeplocks;
2325 	else {
2326 		if (witness_skipspin)
2327 			return;
2328 		lock_list = PCPU_GET(spinlocks);
2329 	}
2330 	instance = find_instance(lock_list, lock);
2331 	if (instance == NULL)
2332 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2333 		    class->lc_name, lock->lo_name);
2334 	lock->lo_witness->w_file = file;
2335 	lock->lo_witness->w_line = line;
2336 	if (instance == NULL)
2337 		return;
2338 	instance->li_file = file;
2339 	instance->li_line = line;
2340 }
2341 
2342 void
2343 witness_assert(const struct lock_object *lock, int flags, const char *file,
2344     int line)
2345 {
2346 #ifdef INVARIANT_SUPPORT
2347 	struct lock_instance *instance;
2348 	struct lock_class *class;
2349 
2350 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2351 		return;
2352 	class = LOCK_CLASS(lock);
2353 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2354 		instance = find_instance(curthread->td_sleeplocks, lock);
2355 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2356 		instance = find_instance(PCPU_GET(spinlocks), lock);
2357 	else {
2358 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2359 		    class->lc_name, lock->lo_name);
2360 		return;
2361 	}
2362 	switch (flags) {
2363 	case LA_UNLOCKED:
2364 		if (instance != NULL)
2365 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2366 			    class->lc_name, lock->lo_name,
2367 			    fixup_filename(file), line);
2368 		break;
2369 	case LA_LOCKED:
2370 	case LA_LOCKED | LA_RECURSED:
2371 	case LA_LOCKED | LA_NOTRECURSED:
2372 	case LA_SLOCKED:
2373 	case LA_SLOCKED | LA_RECURSED:
2374 	case LA_SLOCKED | LA_NOTRECURSED:
2375 	case LA_XLOCKED:
2376 	case LA_XLOCKED | LA_RECURSED:
2377 	case LA_XLOCKED | LA_NOTRECURSED:
2378 		if (instance == NULL) {
2379 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2380 			    class->lc_name, lock->lo_name,
2381 			    fixup_filename(file), line);
2382 			break;
2383 		}
2384 		if ((flags & LA_XLOCKED) != 0 &&
2385 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2386 			kassert_panic(
2387 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2388 			    class->lc_name, lock->lo_name,
2389 			    fixup_filename(file), line);
2390 		if ((flags & LA_SLOCKED) != 0 &&
2391 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2392 			kassert_panic(
2393 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2394 			    class->lc_name, lock->lo_name,
2395 			    fixup_filename(file), line);
2396 		if ((flags & LA_RECURSED) != 0 &&
2397 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2398 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2399 			    class->lc_name, lock->lo_name,
2400 			    fixup_filename(file), line);
2401 		if ((flags & LA_NOTRECURSED) != 0 &&
2402 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2403 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2404 			    class->lc_name, lock->lo_name,
2405 			    fixup_filename(file), line);
2406 		break;
2407 	default:
2408 		kassert_panic("Invalid lock assertion at %s:%d.",
2409 		    fixup_filename(file), line);
2410 
2411 	}
2412 #endif	/* INVARIANT_SUPPORT */
2413 }
2414 
2415 static void
2416 witness_setflag(struct lock_object *lock, int flag, int set)
2417 {
2418 	struct lock_list_entry *lock_list;
2419 	struct lock_instance *instance;
2420 	struct lock_class *class;
2421 
2422 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2423 		return;
2424 	class = LOCK_CLASS(lock);
2425 	if (class->lc_flags & LC_SLEEPLOCK)
2426 		lock_list = curthread->td_sleeplocks;
2427 	else {
2428 		if (witness_skipspin)
2429 			return;
2430 		lock_list = PCPU_GET(spinlocks);
2431 	}
2432 	instance = find_instance(lock_list, lock);
2433 	if (instance == NULL) {
2434 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2435 		    class->lc_name, lock->lo_name);
2436 		return;
2437 	}
2438 
2439 	if (set)
2440 		instance->li_flags |= flag;
2441 	else
2442 		instance->li_flags &= ~flag;
2443 }
2444 
2445 void
2446 witness_norelease(struct lock_object *lock)
2447 {
2448 
2449 	witness_setflag(lock, LI_NORELEASE, 1);
2450 }
2451 
2452 void
2453 witness_releaseok(struct lock_object *lock)
2454 {
2455 
2456 	witness_setflag(lock, LI_NORELEASE, 0);
2457 }
2458 
2459 #ifdef DDB
2460 static void
2461 witness_ddb_list(struct thread *td)
2462 {
2463 
2464 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2465 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2466 
2467 	if (witness_watch < 1)
2468 		return;
2469 
2470 	witness_list_locks(&td->td_sleeplocks, db_printf);
2471 
2472 	/*
2473 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2474 	 * if td is currently executing on some other CPU and holds spin locks
2475 	 * as we won't display those locks.  If we had a MI way of getting
2476 	 * the per-cpu data for a given cpu then we could use
2477 	 * td->td_oncpu to get the list of spinlocks for this thread
2478 	 * and "fix" this.
2479 	 *
2480 	 * That still wouldn't really fix this unless we locked the scheduler
2481 	 * lock or stopped the other CPU to make sure it wasn't changing the
2482 	 * list out from under us.  It is probably best to just not try to
2483 	 * handle threads on other CPU's for now.
2484 	 */
2485 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2486 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2487 }
2488 
2489 DB_SHOW_COMMAND(locks, db_witness_list)
2490 {
2491 	struct thread *td;
2492 
2493 	if (have_addr)
2494 		td = db_lookup_thread(addr, true);
2495 	else
2496 		td = kdb_thread;
2497 	witness_ddb_list(td);
2498 }
2499 
2500 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2501 {
2502 	struct thread *td;
2503 	struct proc *p;
2504 
2505 	/*
2506 	 * It would be nice to list only threads and processes that actually
2507 	 * held sleep locks, but that information is currently not exported
2508 	 * by WITNESS.
2509 	 */
2510 	FOREACH_PROC_IN_SYSTEM(p) {
2511 		if (!witness_proc_has_locks(p))
2512 			continue;
2513 		FOREACH_THREAD_IN_PROC(p, td) {
2514 			if (!witness_thread_has_locks(td))
2515 				continue;
2516 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2517 			    p->p_comm, td, td->td_tid);
2518 			witness_ddb_list(td);
2519 			if (db_pager_quit)
2520 				return;
2521 		}
2522 	}
2523 }
2524 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2525 
2526 DB_SHOW_COMMAND(witness, db_witness_display)
2527 {
2528 
2529 	witness_ddb_display(db_printf);
2530 }
2531 #endif
2532 
2533 static int
2534 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2535 {
2536 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2537 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2538 	struct sbuf *sb;
2539 	u_int w_rmatrix1, w_rmatrix2;
2540 	int error, generation, i, j;
2541 
2542 	tmp_data1 = NULL;
2543 	tmp_data2 = NULL;
2544 	tmp_w1 = NULL;
2545 	tmp_w2 = NULL;
2546 	if (witness_watch < 1) {
2547 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2548 		return (error);
2549 	}
2550 	if (witness_cold) {
2551 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2552 		return (error);
2553 	}
2554 	error = 0;
2555 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2556 	if (sb == NULL)
2557 		return (ENOMEM);
2558 
2559 	/* Allocate and init temporary storage space. */
2560 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2561 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2562 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2563 	    M_WAITOK | M_ZERO);
2564 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2565 	    M_WAITOK | M_ZERO);
2566 	stack_zero(&tmp_data1->wlod_stack);
2567 	stack_zero(&tmp_data2->wlod_stack);
2568 
2569 restart:
2570 	mtx_lock_spin(&w_mtx);
2571 	generation = w_generation;
2572 	mtx_unlock_spin(&w_mtx);
2573 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2574 	    w_lohash.wloh_count);
2575 	for (i = 1; i < w_max_used_index; i++) {
2576 		mtx_lock_spin(&w_mtx);
2577 		if (generation != w_generation) {
2578 			mtx_unlock_spin(&w_mtx);
2579 
2580 			/* The graph has changed, try again. */
2581 			req->oldidx = 0;
2582 			sbuf_clear(sb);
2583 			goto restart;
2584 		}
2585 
2586 		w1 = &w_data[i];
2587 		if (w1->w_reversed == 0) {
2588 			mtx_unlock_spin(&w_mtx);
2589 			continue;
2590 		}
2591 
2592 		/* Copy w1 locally so we can release the spin lock. */
2593 		*tmp_w1 = *w1;
2594 		mtx_unlock_spin(&w_mtx);
2595 
2596 		if (tmp_w1->w_reversed == 0)
2597 			continue;
2598 		for (j = 1; j < w_max_used_index; j++) {
2599 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2600 				continue;
2601 
2602 			mtx_lock_spin(&w_mtx);
2603 			if (generation != w_generation) {
2604 				mtx_unlock_spin(&w_mtx);
2605 
2606 				/* The graph has changed, try again. */
2607 				req->oldidx = 0;
2608 				sbuf_clear(sb);
2609 				goto restart;
2610 			}
2611 
2612 			w2 = &w_data[j];
2613 			data1 = witness_lock_order_get(w1, w2);
2614 			data2 = witness_lock_order_get(w2, w1);
2615 
2616 			/*
2617 			 * Copy information locally so we can release the
2618 			 * spin lock.
2619 			 */
2620 			*tmp_w2 = *w2;
2621 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2622 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2623 
2624 			if (data1) {
2625 				stack_zero(&tmp_data1->wlod_stack);
2626 				stack_copy(&data1->wlod_stack,
2627 				    &tmp_data1->wlod_stack);
2628 			}
2629 			if (data2 && data2 != data1) {
2630 				stack_zero(&tmp_data2->wlod_stack);
2631 				stack_copy(&data2->wlod_stack,
2632 				    &tmp_data2->wlod_stack);
2633 			}
2634 			mtx_unlock_spin(&w_mtx);
2635 
2636 			sbuf_printf(sb,
2637 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2638 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2639 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2640 			if (data1) {
2641 				sbuf_printf(sb,
2642 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2643 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2644 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2645 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2646 				sbuf_printf(sb, "\n");
2647 			}
2648 			if (data2 && data2 != data1) {
2649 				sbuf_printf(sb,
2650 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2651 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2652 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2653 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2654 				sbuf_printf(sb, "\n");
2655 			}
2656 		}
2657 	}
2658 	mtx_lock_spin(&w_mtx);
2659 	if (generation != w_generation) {
2660 		mtx_unlock_spin(&w_mtx);
2661 
2662 		/*
2663 		 * The graph changed while we were printing stack data,
2664 		 * try again.
2665 		 */
2666 		req->oldidx = 0;
2667 		sbuf_clear(sb);
2668 		goto restart;
2669 	}
2670 	mtx_unlock_spin(&w_mtx);
2671 
2672 	/* Free temporary storage space. */
2673 	free(tmp_data1, M_TEMP);
2674 	free(tmp_data2, M_TEMP);
2675 	free(tmp_w1, M_TEMP);
2676 	free(tmp_w2, M_TEMP);
2677 
2678 	sbuf_finish(sb);
2679 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2680 	sbuf_delete(sb);
2681 
2682 	return (error);
2683 }
2684 
2685 static int
2686 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2687 {
2688 	static const struct {
2689 		enum witness_channel channel;
2690 		const char *name;
2691 	} channels[] = {
2692 		{ WITNESS_CONSOLE, "console" },
2693 		{ WITNESS_LOG, "log" },
2694 		{ WITNESS_NONE, "none" },
2695 	};
2696 	char buf[16];
2697 	u_int i;
2698 	int error;
2699 
2700 	buf[0] = '\0';
2701 	for (i = 0; i < nitems(channels); i++)
2702 		if (witness_channel == channels[i].channel) {
2703 			snprintf(buf, sizeof(buf), "%s", channels[i].name);
2704 			break;
2705 		}
2706 
2707 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2708 	if (error != 0 || req->newptr == NULL)
2709 		return (error);
2710 
2711 	error = EINVAL;
2712 	for (i = 0; i < nitems(channels); i++)
2713 		if (strcmp(channels[i].name, buf) == 0) {
2714 			witness_channel = channels[i].channel;
2715 			error = 0;
2716 			break;
2717 		}
2718 	return (error);
2719 }
2720 
2721 static int
2722 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2723 {
2724 	struct witness *w;
2725 	struct sbuf *sb;
2726 	int error;
2727 
2728 	if (witness_watch < 1) {
2729 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2730 		return (error);
2731 	}
2732 	if (witness_cold) {
2733 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2734 		return (error);
2735 	}
2736 	error = 0;
2737 
2738 	error = sysctl_wire_old_buffer(req, 0);
2739 	if (error != 0)
2740 		return (error);
2741 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2742 	if (sb == NULL)
2743 		return (ENOMEM);
2744 	sbuf_printf(sb, "\n");
2745 
2746 	mtx_lock_spin(&w_mtx);
2747 	STAILQ_FOREACH(w, &w_all, w_list)
2748 		w->w_displayed = 0;
2749 	STAILQ_FOREACH(w, &w_all, w_list)
2750 		witness_add_fullgraph(sb, w);
2751 	mtx_unlock_spin(&w_mtx);
2752 
2753 	/*
2754 	 * Close the sbuf and return to userland.
2755 	 */
2756 	error = sbuf_finish(sb);
2757 	sbuf_delete(sb);
2758 
2759 	return (error);
2760 }
2761 
2762 static int
2763 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2764 {
2765 	int error, value;
2766 
2767 	value = witness_watch;
2768 	error = sysctl_handle_int(oidp, &value, 0, req);
2769 	if (error != 0 || req->newptr == NULL)
2770 		return (error);
2771 	if (value > 1 || value < -1 ||
2772 	    (witness_watch == -1 && value != witness_watch))
2773 		return (EINVAL);
2774 	witness_watch = value;
2775 	return (0);
2776 }
2777 
2778 static void
2779 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2780 {
2781 	int i;
2782 
2783 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2784 		return;
2785 	w->w_displayed = 1;
2786 
2787 	WITNESS_INDEX_ASSERT(w->w_index);
2788 	for (i = 1; i <= w_max_used_index; i++) {
2789 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2790 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2791 			    w_data[i].w_name);
2792 			witness_add_fullgraph(sb, &w_data[i]);
2793 		}
2794 	}
2795 }
2796 
2797 /*
2798  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2799  * interprets the key as a string and reads until the null
2800  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2801  * hash value computed from the key.
2802  */
2803 static uint32_t
2804 witness_hash_djb2(const uint8_t *key, uint32_t size)
2805 {
2806 	unsigned int hash = 5381;
2807 	int i;
2808 
2809 	/* hash = hash * 33 + key[i] */
2810 	if (size)
2811 		for (i = 0; i < size; i++)
2812 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2813 	else
2814 		for (i = 0; key[i] != 0; i++)
2815 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2816 
2817 	return (hash);
2818 }
2819 
2820 
2821 /*
2822  * Initializes the two witness hash tables. Called exactly once from
2823  * witness_initialize().
2824  */
2825 static void
2826 witness_init_hash_tables(void)
2827 {
2828 	int i;
2829 
2830 	MPASS(witness_cold);
2831 
2832 	/* Initialize the hash tables. */
2833 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2834 		w_hash.wh_array[i] = NULL;
2835 
2836 	w_hash.wh_size = WITNESS_HASH_SIZE;
2837 	w_hash.wh_count = 0;
2838 
2839 	/* Initialize the lock order data hash. */
2840 	w_lofree = NULL;
2841 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2842 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2843 		w_lodata[i].wlod_next = w_lofree;
2844 		w_lofree = &w_lodata[i];
2845 	}
2846 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2847 	w_lohash.wloh_count = 0;
2848 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2849 		w_lohash.wloh_array[i] = NULL;
2850 }
2851 
2852 static struct witness *
2853 witness_hash_get(const char *key)
2854 {
2855 	struct witness *w;
2856 	uint32_t hash;
2857 
2858 	MPASS(key != NULL);
2859 	if (witness_cold == 0)
2860 		mtx_assert(&w_mtx, MA_OWNED);
2861 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2862 	w = w_hash.wh_array[hash];
2863 	while (w != NULL) {
2864 		if (strcmp(w->w_name, key) == 0)
2865 			goto out;
2866 		w = w->w_hash_next;
2867 	}
2868 
2869 out:
2870 	return (w);
2871 }
2872 
2873 static void
2874 witness_hash_put(struct witness *w)
2875 {
2876 	uint32_t hash;
2877 
2878 	MPASS(w != NULL);
2879 	MPASS(w->w_name != NULL);
2880 	if (witness_cold == 0)
2881 		mtx_assert(&w_mtx, MA_OWNED);
2882 	KASSERT(witness_hash_get(w->w_name) == NULL,
2883 	    ("%s: trying to add a hash entry that already exists!", __func__));
2884 	KASSERT(w->w_hash_next == NULL,
2885 	    ("%s: w->w_hash_next != NULL", __func__));
2886 
2887 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2888 	w->w_hash_next = w_hash.wh_array[hash];
2889 	w_hash.wh_array[hash] = w;
2890 	w_hash.wh_count++;
2891 }
2892 
2893 
2894 static struct witness_lock_order_data *
2895 witness_lock_order_get(struct witness *parent, struct witness *child)
2896 {
2897 	struct witness_lock_order_data *data = NULL;
2898 	struct witness_lock_order_key key;
2899 	unsigned int hash;
2900 
2901 	MPASS(parent != NULL && child != NULL);
2902 	key.from = parent->w_index;
2903 	key.to = child->w_index;
2904 	WITNESS_INDEX_ASSERT(key.from);
2905 	WITNESS_INDEX_ASSERT(key.to);
2906 	if ((w_rmatrix[parent->w_index][child->w_index]
2907 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2908 		goto out;
2909 
2910 	hash = witness_hash_djb2((const char*)&key,
2911 	    sizeof(key)) % w_lohash.wloh_size;
2912 	data = w_lohash.wloh_array[hash];
2913 	while (data != NULL) {
2914 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2915 			break;
2916 		data = data->wlod_next;
2917 	}
2918 
2919 out:
2920 	return (data);
2921 }
2922 
2923 /*
2924  * Verify that parent and child have a known relationship, are not the same,
2925  * and child is actually a child of parent.  This is done without w_mtx
2926  * to avoid contention in the common case.
2927  */
2928 static int
2929 witness_lock_order_check(struct witness *parent, struct witness *child)
2930 {
2931 
2932 	if (parent != child &&
2933 	    w_rmatrix[parent->w_index][child->w_index]
2934 	    & WITNESS_LOCK_ORDER_KNOWN &&
2935 	    isitmychild(parent, child))
2936 		return (1);
2937 
2938 	return (0);
2939 }
2940 
2941 static int
2942 witness_lock_order_add(struct witness *parent, struct witness *child)
2943 {
2944 	struct witness_lock_order_data *data = NULL;
2945 	struct witness_lock_order_key key;
2946 	unsigned int hash;
2947 
2948 	MPASS(parent != NULL && child != NULL);
2949 	key.from = parent->w_index;
2950 	key.to = child->w_index;
2951 	WITNESS_INDEX_ASSERT(key.from);
2952 	WITNESS_INDEX_ASSERT(key.to);
2953 	if (w_rmatrix[parent->w_index][child->w_index]
2954 	    & WITNESS_LOCK_ORDER_KNOWN)
2955 		return (1);
2956 
2957 	hash = witness_hash_djb2((const char*)&key,
2958 	    sizeof(key)) % w_lohash.wloh_size;
2959 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2960 	data = w_lofree;
2961 	if (data == NULL)
2962 		return (0);
2963 	w_lofree = data->wlod_next;
2964 	data->wlod_next = w_lohash.wloh_array[hash];
2965 	data->wlod_key = key;
2966 	w_lohash.wloh_array[hash] = data;
2967 	w_lohash.wloh_count++;
2968 	stack_zero(&data->wlod_stack);
2969 	stack_save(&data->wlod_stack);
2970 	return (1);
2971 }
2972 
2973 /* Call this whenever the structure of the witness graph changes. */
2974 static void
2975 witness_increment_graph_generation(void)
2976 {
2977 
2978 	if (witness_cold == 0)
2979 		mtx_assert(&w_mtx, MA_OWNED);
2980 	w_generation++;
2981 }
2982 
2983 static int
2984 witness_output_drain(void *arg __unused, const char *data, int len)
2985 {
2986 
2987 	witness_output("%.*s", len, data);
2988 	return (len);
2989 }
2990 
2991 static void
2992 witness_debugger(int cond, const char *msg)
2993 {
2994 	char buf[32];
2995 	struct sbuf sb;
2996 	struct stack st;
2997 
2998 	if (!cond)
2999 		return;
3000 
3001 	if (witness_trace) {
3002 		sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3003 		sbuf_set_drain(&sb, witness_output_drain, NULL);
3004 
3005 		stack_zero(&st);
3006 		stack_save(&st);
3007 		witness_output("stack backtrace:\n");
3008 		stack_sbuf_print_ddb(&sb, &st);
3009 
3010 		sbuf_finish(&sb);
3011 	}
3012 
3013 #ifdef KDB
3014 	if (witness_kdb)
3015 		kdb_enter(KDB_WHY_WITNESS, msg);
3016 #endif
3017 }
3018