xref: /freebsd/sys/kern/subr_witness.c (revision 97549c34ecaf74580941fdc9c5bd1050e4b1f6ce)
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 	 * VFS namecache
627 	 */
628 	{ "ncvn", &lock_class_mtx_sleep },
629 	{ "ncbuc", &lock_class_rw },
630 	{ "vnode interlock", &lock_class_mtx_sleep },
631 	{ "ncneg", &lock_class_mtx_sleep },
632 	{ NULL, NULL },
633 	/*
634 	 * ZFS locking
635 	 */
636 	{ "dn->dn_mtx", &lock_class_sx },
637 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
638 	{ "db->db_mtx", &lock_class_sx },
639 	{ NULL, NULL },
640 	/*
641 	 * spin locks
642 	 */
643 #ifdef SMP
644 	{ "ap boot", &lock_class_mtx_spin },
645 #endif
646 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
647 	{ "sio", &lock_class_mtx_spin },
648 #ifdef __i386__
649 	{ "cy", &lock_class_mtx_spin },
650 #endif
651 #ifdef __sparc64__
652 	{ "pcib_mtx", &lock_class_mtx_spin },
653 	{ "rtc_mtx", &lock_class_mtx_spin },
654 #endif
655 	{ "scc_hwmtx", &lock_class_mtx_spin },
656 	{ "uart_hwmtx", &lock_class_mtx_spin },
657 	{ "fast_taskqueue", &lock_class_mtx_spin },
658 	{ "intr table", &lock_class_mtx_spin },
659 #ifdef	HWPMC_HOOKS
660 	{ "pmc-per-proc", &lock_class_mtx_spin },
661 #endif
662 	{ "process slock", &lock_class_mtx_spin },
663 	{ "syscons video lock", &lock_class_mtx_spin },
664 	{ "sleepq chain", &lock_class_mtx_spin },
665 	{ "rm_spinlock", &lock_class_mtx_spin },
666 	{ "turnstile chain", &lock_class_mtx_spin },
667 	{ "turnstile lock", &lock_class_mtx_spin },
668 	{ "sched lock", &lock_class_mtx_spin },
669 	{ "td_contested", &lock_class_mtx_spin },
670 	{ "callout", &lock_class_mtx_spin },
671 	{ "entropy harvest mutex", &lock_class_mtx_spin },
672 #ifdef SMP
673 	{ "smp rendezvous", &lock_class_mtx_spin },
674 #endif
675 #ifdef __powerpc__
676 	{ "tlb0", &lock_class_mtx_spin },
677 #endif
678 	/*
679 	 * leaf locks
680 	 */
681 	{ "intrcnt", &lock_class_mtx_spin },
682 	{ "icu", &lock_class_mtx_spin },
683 #if defined(SMP) && defined(__sparc64__)
684 	{ "ipi", &lock_class_mtx_spin },
685 #endif
686 #ifdef __i386__
687 	{ "allpmaps", &lock_class_mtx_spin },
688 	{ "descriptor tables", &lock_class_mtx_spin },
689 #endif
690 	{ "clk", &lock_class_mtx_spin },
691 	{ "cpuset", &lock_class_mtx_spin },
692 	{ "mprof lock", &lock_class_mtx_spin },
693 	{ "zombie lock", &lock_class_mtx_spin },
694 	{ "ALD Queue", &lock_class_mtx_spin },
695 #if defined(__i386__) || defined(__amd64__)
696 	{ "pcicfg", &lock_class_mtx_spin },
697 	{ "NDIS thread lock", &lock_class_mtx_spin },
698 #endif
699 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
700 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
701 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
702 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
703 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
704 #ifdef	HWPMC_HOOKS
705 	{ "pmc-leaf", &lock_class_mtx_spin },
706 #endif
707 	{ "blocked lock", &lock_class_mtx_spin },
708 	{ NULL, NULL },
709 	{ NULL, NULL }
710 };
711 
712 #ifdef BLESSING
713 /*
714  * Pairs of locks which have been blessed
715  * Don't complain about order problems with blessed locks
716  */
717 static struct witness_blessed blessed_list[] = {
718 };
719 #endif
720 
721 /*
722  * This global is set to 0 once it becomes safe to use the witness code.
723  */
724 static int witness_cold = 1;
725 
726 /*
727  * This global is set to 1 once the static lock orders have been enrolled
728  * so that a warning can be issued for any spin locks enrolled later.
729  */
730 static int witness_spin_warn = 0;
731 
732 /* Trim useless garbage from filenames. */
733 static const char *
734 fixup_filename(const char *file)
735 {
736 
737 	if (file == NULL)
738 		return (NULL);
739 	while (strncmp(file, "../", 3) == 0)
740 		file += 3;
741 	return (file);
742 }
743 
744 /*
745  * The WITNESS-enabled diagnostic code.  Note that the witness code does
746  * assume that the early boot is single-threaded at least until after this
747  * routine is completed.
748  */
749 static void
750 witness_initialize(void *dummy __unused)
751 {
752 	struct lock_object *lock;
753 	struct witness_order_list_entry *order;
754 	struct witness *w, *w1;
755 	int i;
756 
757 	w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
758 	    M_WAITOK | M_ZERO);
759 
760 	w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
761 	    M_WITNESS, M_WAITOK | M_ZERO);
762 
763 	for (i = 0; i < witness_count + 1; i++) {
764 		w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
765 		    (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
766 	}
767 	badstack_sbuf_size = witness_count * 256;
768 
769 	/*
770 	 * We have to release Giant before initializing its witness
771 	 * structure so that WITNESS doesn't get confused.
772 	 */
773 	mtx_unlock(&Giant);
774 	mtx_assert(&Giant, MA_NOTOWNED);
775 
776 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
777 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
778 	    MTX_NOWITNESS | MTX_NOPROFILE);
779 	for (i = witness_count - 1; i >= 0; i--) {
780 		w = &w_data[i];
781 		memset(w, 0, sizeof(*w));
782 		w_data[i].w_index = i;	/* Witness index never changes. */
783 		witness_free(w);
784 	}
785 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
786 	    ("%s: Invalid list of free witness objects", __func__));
787 
788 	/* Witness with index 0 is not used to aid in debugging. */
789 	STAILQ_REMOVE_HEAD(&w_free, w_list);
790 	w_free_cnt--;
791 
792 	for (i = 0; i < witness_count; i++) {
793 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
794 		    (witness_count + 1));
795 	}
796 
797 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
798 		witness_lock_list_free(&w_locklistdata[i]);
799 	witness_init_hash_tables();
800 
801 	/* First add in all the specified order lists. */
802 	for (order = order_lists; order->w_name != NULL; order++) {
803 		w = enroll(order->w_name, order->w_class);
804 		if (w == NULL)
805 			continue;
806 		w->w_file = "order list";
807 		for (order++; order->w_name != NULL; order++) {
808 			w1 = enroll(order->w_name, order->w_class);
809 			if (w1 == NULL)
810 				continue;
811 			w1->w_file = "order list";
812 			itismychild(w, w1);
813 			w = w1;
814 		}
815 	}
816 	witness_spin_warn = 1;
817 
818 	/* Iterate through all locks and add them to witness. */
819 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
820 		lock = pending_locks[i].wh_lock;
821 		KASSERT(lock->lo_flags & LO_WITNESS,
822 		    ("%s: lock %s is on pending list but not LO_WITNESS",
823 		    __func__, lock->lo_name));
824 		lock->lo_witness = enroll(pending_locks[i].wh_type,
825 		    LOCK_CLASS(lock));
826 	}
827 
828 	/* Mark the witness code as being ready for use. */
829 	witness_cold = 0;
830 
831 	mtx_lock(&Giant);
832 }
833 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
834     NULL);
835 
836 void
837 witness_init(struct lock_object *lock, const char *type)
838 {
839 	struct lock_class *class;
840 
841 	/* Various sanity checks. */
842 	class = LOCK_CLASS(lock);
843 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
844 	    (class->lc_flags & LC_RECURSABLE) == 0)
845 		kassert_panic("%s: lock (%s) %s can not be recursable",
846 		    __func__, class->lc_name, lock->lo_name);
847 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
848 	    (class->lc_flags & LC_SLEEPABLE) == 0)
849 		kassert_panic("%s: lock (%s) %s can not be sleepable",
850 		    __func__, class->lc_name, lock->lo_name);
851 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
852 	    (class->lc_flags & LC_UPGRADABLE) == 0)
853 		kassert_panic("%s: lock (%s) %s can not be upgradable",
854 		    __func__, class->lc_name, lock->lo_name);
855 
856 	/*
857 	 * If we shouldn't watch this lock, then just clear lo_witness.
858 	 * Otherwise, if witness_cold is set, then it is too early to
859 	 * enroll this lock, so defer it to witness_initialize() by adding
860 	 * it to the pending_locks list.  If it is not too early, then enroll
861 	 * the lock now.
862 	 */
863 	if (witness_watch < 1 || panicstr != NULL ||
864 	    (lock->lo_flags & LO_WITNESS) == 0)
865 		lock->lo_witness = NULL;
866 	else if (witness_cold) {
867 		pending_locks[pending_cnt].wh_lock = lock;
868 		pending_locks[pending_cnt++].wh_type = type;
869 		if (pending_cnt > WITNESS_PENDLIST)
870 			panic("%s: pending locks list is too small, "
871 			    "increase WITNESS_PENDLIST\n",
872 			    __func__);
873 	} else
874 		lock->lo_witness = enroll(type, class);
875 }
876 
877 void
878 witness_destroy(struct lock_object *lock)
879 {
880 	struct lock_class *class;
881 	struct witness *w;
882 
883 	class = LOCK_CLASS(lock);
884 
885 	if (witness_cold)
886 		panic("lock (%s) %s destroyed while witness_cold",
887 		    class->lc_name, lock->lo_name);
888 
889 	/* XXX: need to verify that no one holds the lock */
890 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
891 		return;
892 	w = lock->lo_witness;
893 
894 	mtx_lock_spin(&w_mtx);
895 	MPASS(w->w_refcount > 0);
896 	w->w_refcount--;
897 
898 	if (w->w_refcount == 0)
899 		depart(w);
900 	mtx_unlock_spin(&w_mtx);
901 }
902 
903 #ifdef DDB
904 static void
905 witness_ddb_compute_levels(void)
906 {
907 	struct witness *w;
908 
909 	/*
910 	 * First clear all levels.
911 	 */
912 	STAILQ_FOREACH(w, &w_all, w_list)
913 		w->w_ddb_level = -1;
914 
915 	/*
916 	 * Look for locks with no parents and level all their descendants.
917 	 */
918 	STAILQ_FOREACH(w, &w_all, w_list) {
919 
920 		/* If the witness has ancestors (is not a root), skip it. */
921 		if (w->w_num_ancestors > 0)
922 			continue;
923 		witness_ddb_level_descendants(w, 0);
924 	}
925 }
926 
927 static void
928 witness_ddb_level_descendants(struct witness *w, int l)
929 {
930 	int i;
931 
932 	if (w->w_ddb_level >= l)
933 		return;
934 
935 	w->w_ddb_level = l;
936 	l++;
937 
938 	for (i = 1; i <= w_max_used_index; i++) {
939 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
940 			witness_ddb_level_descendants(&w_data[i], l);
941 	}
942 }
943 
944 static void
945 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
946     struct witness *w, int indent)
947 {
948 	int i;
949 
950  	for (i = 0; i < indent; i++)
951  		prnt(" ");
952 	prnt("%s (type: %s, depth: %d, active refs: %d)",
953 	     w->w_name, w->w_class->lc_name,
954 	     w->w_ddb_level, w->w_refcount);
955  	if (w->w_displayed) {
956  		prnt(" -- (already displayed)\n");
957  		return;
958  	}
959  	w->w_displayed = 1;
960 	if (w->w_file != NULL && w->w_line != 0)
961 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
962 		    w->w_line);
963 	else
964 		prnt(" -- never acquired\n");
965 	indent++;
966 	WITNESS_INDEX_ASSERT(w->w_index);
967 	for (i = 1; i <= w_max_used_index; i++) {
968 		if (db_pager_quit)
969 			return;
970 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
971 			witness_ddb_display_descendants(prnt, &w_data[i],
972 			    indent);
973 	}
974 }
975 
976 static void
977 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
978     struct witness_list *list)
979 {
980 	struct witness *w;
981 
982 	STAILQ_FOREACH(w, list, w_typelist) {
983 		if (w->w_file == NULL || w->w_ddb_level > 0)
984 			continue;
985 
986 		/* This lock has no anscestors - display its descendants. */
987 		witness_ddb_display_descendants(prnt, w, 0);
988 		if (db_pager_quit)
989 			return;
990 	}
991 }
992 
993 static void
994 witness_ddb_display(int(*prnt)(const char *fmt, ...))
995 {
996 	struct witness *w;
997 
998 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
999 	witness_ddb_compute_levels();
1000 
1001 	/* Clear all the displayed flags. */
1002 	STAILQ_FOREACH(w, &w_all, w_list)
1003 		w->w_displayed = 0;
1004 
1005 	/*
1006 	 * First, handle sleep locks which have been acquired at least
1007 	 * once.
1008 	 */
1009 	prnt("Sleep locks:\n");
1010 	witness_ddb_display_list(prnt, &w_sleep);
1011 	if (db_pager_quit)
1012 		return;
1013 
1014 	/*
1015 	 * Now do spin locks which have been acquired at least once.
1016 	 */
1017 	prnt("\nSpin locks:\n");
1018 	witness_ddb_display_list(prnt, &w_spin);
1019 	if (db_pager_quit)
1020 		return;
1021 
1022 	/*
1023 	 * Finally, any locks which have not been acquired yet.
1024 	 */
1025 	prnt("\nLocks which were never acquired:\n");
1026 	STAILQ_FOREACH(w, &w_all, w_list) {
1027 		if (w->w_file != NULL || w->w_refcount == 0)
1028 			continue;
1029 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1030 		    w->w_class->lc_name, w->w_ddb_level);
1031 		if (db_pager_quit)
1032 			return;
1033 	}
1034 }
1035 #endif /* DDB */
1036 
1037 int
1038 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1039 {
1040 
1041 	if (witness_watch == -1 || panicstr != NULL)
1042 		return (0);
1043 
1044 	/* Require locks that witness knows about. */
1045 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1046 	    lock2->lo_witness == NULL)
1047 		return (EINVAL);
1048 
1049 	mtx_assert(&w_mtx, MA_NOTOWNED);
1050 	mtx_lock_spin(&w_mtx);
1051 
1052 	/*
1053 	 * If we already have either an explicit or implied lock order that
1054 	 * is the other way around, then return an error.
1055 	 */
1056 	if (witness_watch &&
1057 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1058 		mtx_unlock_spin(&w_mtx);
1059 		return (EDOOFUS);
1060 	}
1061 
1062 	/* Try to add the new order. */
1063 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1064 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1065 	itismychild(lock1->lo_witness, lock2->lo_witness);
1066 	mtx_unlock_spin(&w_mtx);
1067 	return (0);
1068 }
1069 
1070 void
1071 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1072     int line, struct lock_object *interlock)
1073 {
1074 	struct lock_list_entry *lock_list, *lle;
1075 	struct lock_instance *lock1, *lock2, *plock;
1076 	struct lock_class *class, *iclass;
1077 	struct witness *w, *w1;
1078 	struct thread *td;
1079 	int i, j;
1080 
1081 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1082 	    panicstr != NULL)
1083 		return;
1084 
1085 	w = lock->lo_witness;
1086 	class = LOCK_CLASS(lock);
1087 	td = curthread;
1088 
1089 	if (class->lc_flags & LC_SLEEPLOCK) {
1090 
1091 		/*
1092 		 * Since spin locks include a critical section, this check
1093 		 * implicitly enforces a lock order of all sleep locks before
1094 		 * all spin locks.
1095 		 */
1096 		if (td->td_critnest != 0 && !kdb_active)
1097 			kassert_panic("acquiring blockable sleep lock with "
1098 			    "spinlock or critical section held (%s) %s @ %s:%d",
1099 			    class->lc_name, lock->lo_name,
1100 			    fixup_filename(file), line);
1101 
1102 		/*
1103 		 * If this is the first lock acquired then just return as
1104 		 * no order checking is needed.
1105 		 */
1106 		lock_list = td->td_sleeplocks;
1107 		if (lock_list == NULL || lock_list->ll_count == 0)
1108 			return;
1109 	} else {
1110 
1111 		/*
1112 		 * If this is the first lock, just return as no order
1113 		 * checking is needed.  Avoid problems with thread
1114 		 * migration pinning the thread while checking if
1115 		 * spinlocks are held.  If at least one spinlock is held
1116 		 * the thread is in a safe path and it is allowed to
1117 		 * unpin it.
1118 		 */
1119 		sched_pin();
1120 		lock_list = PCPU_GET(spinlocks);
1121 		if (lock_list == NULL || lock_list->ll_count == 0) {
1122 			sched_unpin();
1123 			return;
1124 		}
1125 		sched_unpin();
1126 	}
1127 
1128 	/*
1129 	 * Check to see if we are recursing on a lock we already own.  If
1130 	 * so, make sure that we don't mismatch exclusive and shared lock
1131 	 * acquires.
1132 	 */
1133 	lock1 = find_instance(lock_list, lock);
1134 	if (lock1 != NULL) {
1135 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1136 		    (flags & LOP_EXCLUSIVE) == 0) {
1137 			witness_output("shared lock of (%s) %s @ %s:%d\n",
1138 			    class->lc_name, lock->lo_name,
1139 			    fixup_filename(file), line);
1140 			witness_output("while exclusively locked from %s:%d\n",
1141 			    fixup_filename(lock1->li_file), lock1->li_line);
1142 			kassert_panic("excl->share");
1143 		}
1144 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1145 		    (flags & LOP_EXCLUSIVE) != 0) {
1146 			witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1147 			    class->lc_name, lock->lo_name,
1148 			    fixup_filename(file), line);
1149 			witness_output("while share locked from %s:%d\n",
1150 			    fixup_filename(lock1->li_file), lock1->li_line);
1151 			kassert_panic("share->excl");
1152 		}
1153 		return;
1154 	}
1155 
1156 	/* Warn if the interlock is not locked exactly once. */
1157 	if (interlock != NULL) {
1158 		iclass = LOCK_CLASS(interlock);
1159 		lock1 = find_instance(lock_list, interlock);
1160 		if (lock1 == NULL)
1161 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1162 			    iclass->lc_name, interlock->lo_name,
1163 			    fixup_filename(file), line);
1164 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1165 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1166 			    iclass->lc_name, interlock->lo_name,
1167 			    fixup_filename(file), line);
1168 	}
1169 
1170 	/*
1171 	 * Find the previously acquired lock, but ignore interlocks.
1172 	 */
1173 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1174 	if (interlock != NULL && plock->li_lock == interlock) {
1175 		if (lock_list->ll_count > 1)
1176 			plock =
1177 			    &lock_list->ll_children[lock_list->ll_count - 2];
1178 		else {
1179 			lle = lock_list->ll_next;
1180 
1181 			/*
1182 			 * The interlock is the only lock we hold, so
1183 			 * simply return.
1184 			 */
1185 			if (lle == NULL)
1186 				return;
1187 			plock = &lle->ll_children[lle->ll_count - 1];
1188 		}
1189 	}
1190 
1191 	/*
1192 	 * Try to perform most checks without a lock.  If this succeeds we
1193 	 * can skip acquiring the lock and return success.  Otherwise we redo
1194 	 * the check with the lock held to handle races with concurrent updates.
1195 	 */
1196 	w1 = plock->li_lock->lo_witness;
1197 	if (witness_lock_order_check(w1, w))
1198 		return;
1199 
1200 	mtx_lock_spin(&w_mtx);
1201 	if (witness_lock_order_check(w1, w)) {
1202 		mtx_unlock_spin(&w_mtx);
1203 		return;
1204 	}
1205 	witness_lock_order_add(w1, w);
1206 
1207 	/*
1208 	 * Check for duplicate locks of the same type.  Note that we only
1209 	 * have to check for this on the last lock we just acquired.  Any
1210 	 * other cases will be caught as lock order violations.
1211 	 */
1212 	if (w1 == w) {
1213 		i = w->w_index;
1214 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1215 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1216 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1217 			w->w_reversed = 1;
1218 			mtx_unlock_spin(&w_mtx);
1219 			witness_output(
1220 			    "acquiring duplicate lock of same type: \"%s\"\n",
1221 			    w->w_name);
1222 			witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1223 			    fixup_filename(plock->li_file), plock->li_line);
1224 			witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1225 			    fixup_filename(file), line);
1226 			witness_debugger(1, __func__);
1227 		} else
1228 			mtx_unlock_spin(&w_mtx);
1229 		return;
1230 	}
1231 	mtx_assert(&w_mtx, MA_OWNED);
1232 
1233 	/*
1234 	 * If we know that the lock we are acquiring comes after
1235 	 * the lock we most recently acquired in the lock order tree,
1236 	 * then there is no need for any further checks.
1237 	 */
1238 	if (isitmychild(w1, w))
1239 		goto out;
1240 
1241 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1242 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1243 
1244 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1245 			lock1 = &lle->ll_children[i];
1246 
1247 			/*
1248 			 * Ignore the interlock.
1249 			 */
1250 			if (interlock == lock1->li_lock)
1251 				continue;
1252 
1253 			/*
1254 			 * If this lock doesn't undergo witness checking,
1255 			 * then skip it.
1256 			 */
1257 			w1 = lock1->li_lock->lo_witness;
1258 			if (w1 == NULL) {
1259 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1260 				    ("lock missing witness structure"));
1261 				continue;
1262 			}
1263 
1264 			/*
1265 			 * If we are locking Giant and this is a sleepable
1266 			 * lock, then skip it.
1267 			 */
1268 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1269 			    lock == &Giant.lock_object)
1270 				continue;
1271 
1272 			/*
1273 			 * If we are locking a sleepable lock and this lock
1274 			 * is Giant, then skip it.
1275 			 */
1276 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1277 			    lock1->li_lock == &Giant.lock_object)
1278 				continue;
1279 
1280 			/*
1281 			 * If we are locking a sleepable lock and this lock
1282 			 * isn't sleepable, we want to treat it as a lock
1283 			 * order violation to enfore a general lock order of
1284 			 * sleepable locks before non-sleepable locks.
1285 			 */
1286 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1287 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1288 				goto reversal;
1289 
1290 			/*
1291 			 * If we are locking Giant and this is a non-sleepable
1292 			 * lock, then treat it as a reversal.
1293 			 */
1294 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1295 			    lock == &Giant.lock_object)
1296 				goto reversal;
1297 
1298 			/*
1299 			 * Check the lock order hierarchy for a reveresal.
1300 			 */
1301 			if (!isitmydescendant(w, w1))
1302 				continue;
1303 		reversal:
1304 
1305 			/*
1306 			 * We have a lock order violation, check to see if it
1307 			 * is allowed or has already been yelled about.
1308 			 */
1309 #ifdef BLESSING
1310 
1311 			/*
1312 			 * If the lock order is blessed, just bail.  We don't
1313 			 * look for other lock order violations though, which
1314 			 * may be a bug.
1315 			 */
1316 			if (blessed(w, w1))
1317 				goto out;
1318 #endif
1319 
1320 			/* Bail if this violation is known */
1321 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1322 				goto out;
1323 
1324 			/* Record this as a violation */
1325 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1326 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1327 			w->w_reversed = w1->w_reversed = 1;
1328 			witness_increment_graph_generation();
1329 			mtx_unlock_spin(&w_mtx);
1330 
1331 #ifdef WITNESS_NO_VNODE
1332 			/*
1333 			 * There are known LORs between VNODE locks. They are
1334 			 * not an indication of a bug. VNODE locks are flagged
1335 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1336 			 * is between 2 VNODE locks.
1337 			 */
1338 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1339 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1340 				return;
1341 #endif
1342 
1343 			/*
1344 			 * Ok, yell about it.
1345 			 */
1346 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1347 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1348 				witness_output(
1349 		"lock order reversal: (sleepable after non-sleepable)\n");
1350 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1351 			    && lock == &Giant.lock_object)
1352 				witness_output(
1353 		"lock order reversal: (Giant after non-sleepable)\n");
1354 			else
1355 				witness_output("lock order reversal:\n");
1356 
1357 			/*
1358 			 * Try to locate an earlier lock with
1359 			 * witness w in our list.
1360 			 */
1361 			do {
1362 				lock2 = &lle->ll_children[i];
1363 				MPASS(lock2->li_lock != NULL);
1364 				if (lock2->li_lock->lo_witness == w)
1365 					break;
1366 				if (i == 0 && lle->ll_next != NULL) {
1367 					lle = lle->ll_next;
1368 					i = lle->ll_count - 1;
1369 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1370 				} else
1371 					i--;
1372 			} while (i >= 0);
1373 			if (i < 0) {
1374 				witness_output(" 1st %p %s (%s) @ %s:%d\n",
1375 				    lock1->li_lock, lock1->li_lock->lo_name,
1376 				    w1->w_name, fixup_filename(lock1->li_file),
1377 				    lock1->li_line);
1378 				witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1379 				    lock->lo_name, w->w_name,
1380 				    fixup_filename(file), line);
1381 			} else {
1382 				witness_output(" 1st %p %s (%s) @ %s:%d\n",
1383 				    lock2->li_lock, lock2->li_lock->lo_name,
1384 				    lock2->li_lock->lo_witness->w_name,
1385 				    fixup_filename(lock2->li_file),
1386 				    lock2->li_line);
1387 				witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1388 				    lock1->li_lock, lock1->li_lock->lo_name,
1389 				    w1->w_name, fixup_filename(lock1->li_file),
1390 				    lock1->li_line);
1391 				witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1392 				    lock->lo_name, w->w_name,
1393 				    fixup_filename(file), line);
1394 			}
1395 			witness_debugger(1, __func__);
1396 			return;
1397 		}
1398 	}
1399 
1400 	/*
1401 	 * If requested, build a new lock order.  However, don't build a new
1402 	 * relationship between a sleepable lock and Giant if it is in the
1403 	 * wrong direction.  The correct lock order is that sleepable locks
1404 	 * always come before Giant.
1405 	 */
1406 	if (flags & LOP_NEWORDER &&
1407 	    !(plock->li_lock == &Giant.lock_object &&
1408 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1409 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1410 		    w->w_name, plock->li_lock->lo_witness->w_name);
1411 		itismychild(plock->li_lock->lo_witness, w);
1412 	}
1413 out:
1414 	mtx_unlock_spin(&w_mtx);
1415 }
1416 
1417 void
1418 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1419 {
1420 	struct lock_list_entry **lock_list, *lle;
1421 	struct lock_instance *instance;
1422 	struct witness *w;
1423 	struct thread *td;
1424 
1425 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1426 	    panicstr != NULL)
1427 		return;
1428 	w = lock->lo_witness;
1429 	td = curthread;
1430 
1431 	/* Determine lock list for this lock. */
1432 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1433 		lock_list = &td->td_sleeplocks;
1434 	else
1435 		lock_list = PCPU_PTR(spinlocks);
1436 
1437 	/* Check to see if we are recursing on a lock we already own. */
1438 	instance = find_instance(*lock_list, lock);
1439 	if (instance != NULL) {
1440 		instance->li_flags++;
1441 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1442 		    td->td_proc->p_pid, lock->lo_name,
1443 		    instance->li_flags & LI_RECURSEMASK);
1444 		instance->li_file = file;
1445 		instance->li_line = line;
1446 		return;
1447 	}
1448 
1449 	/* Update per-witness last file and line acquire. */
1450 	w->w_file = file;
1451 	w->w_line = line;
1452 
1453 	/* Find the next open lock instance in the list and fill it. */
1454 	lle = *lock_list;
1455 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1456 		lle = witness_lock_list_get();
1457 		if (lle == NULL)
1458 			return;
1459 		lle->ll_next = *lock_list;
1460 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1461 		    td->td_proc->p_pid, lle);
1462 		*lock_list = lle;
1463 	}
1464 	instance = &lle->ll_children[lle->ll_count++];
1465 	instance->li_lock = lock;
1466 	instance->li_line = line;
1467 	instance->li_file = file;
1468 	if ((flags & LOP_EXCLUSIVE) != 0)
1469 		instance->li_flags = LI_EXCLUSIVE;
1470 	else
1471 		instance->li_flags = 0;
1472 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1473 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1474 }
1475 
1476 void
1477 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1478 {
1479 	struct lock_instance *instance;
1480 	struct lock_class *class;
1481 
1482 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1483 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1484 		return;
1485 	class = LOCK_CLASS(lock);
1486 	if (witness_watch) {
1487 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1488 			kassert_panic(
1489 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1490 			    class->lc_name, lock->lo_name,
1491 			    fixup_filename(file), line);
1492 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1493 			kassert_panic(
1494 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1495 			    class->lc_name, lock->lo_name,
1496 			    fixup_filename(file), line);
1497 	}
1498 	instance = find_instance(curthread->td_sleeplocks, lock);
1499 	if (instance == NULL) {
1500 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1501 		    class->lc_name, lock->lo_name,
1502 		    fixup_filename(file), line);
1503 		return;
1504 	}
1505 	if (witness_watch) {
1506 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1507 			kassert_panic(
1508 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1509 			    class->lc_name, lock->lo_name,
1510 			    fixup_filename(file), line);
1511 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1512 			kassert_panic(
1513 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1514 			    class->lc_name, lock->lo_name,
1515 			    instance->li_flags & LI_RECURSEMASK,
1516 			    fixup_filename(file), line);
1517 	}
1518 	instance->li_flags |= LI_EXCLUSIVE;
1519 }
1520 
1521 void
1522 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1523     int line)
1524 {
1525 	struct lock_instance *instance;
1526 	struct lock_class *class;
1527 
1528 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1529 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1530 		return;
1531 	class = LOCK_CLASS(lock);
1532 	if (witness_watch) {
1533 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1534 			kassert_panic(
1535 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1536 			    class->lc_name, lock->lo_name,
1537 			    fixup_filename(file), line);
1538 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1539 			kassert_panic(
1540 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1541 			    class->lc_name, lock->lo_name,
1542 			    fixup_filename(file), line);
1543 	}
1544 	instance = find_instance(curthread->td_sleeplocks, lock);
1545 	if (instance == NULL) {
1546 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1547 		    class->lc_name, lock->lo_name,
1548 		    fixup_filename(file), line);
1549 		return;
1550 	}
1551 	if (witness_watch) {
1552 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1553 			kassert_panic(
1554 			    "downgrade of shared lock (%s) %s @ %s:%d",
1555 			    class->lc_name, lock->lo_name,
1556 			    fixup_filename(file), line);
1557 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1558 			kassert_panic(
1559 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1560 			    class->lc_name, lock->lo_name,
1561 			    instance->li_flags & LI_RECURSEMASK,
1562 			    fixup_filename(file), line);
1563 	}
1564 	instance->li_flags &= ~LI_EXCLUSIVE;
1565 }
1566 
1567 void
1568 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1569 {
1570 	struct lock_list_entry **lock_list, *lle;
1571 	struct lock_instance *instance;
1572 	struct lock_class *class;
1573 	struct thread *td;
1574 	register_t s;
1575 	int i, j;
1576 
1577 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1578 		return;
1579 	td = curthread;
1580 	class = LOCK_CLASS(lock);
1581 
1582 	/* Find lock instance associated with this lock. */
1583 	if (class->lc_flags & LC_SLEEPLOCK)
1584 		lock_list = &td->td_sleeplocks;
1585 	else
1586 		lock_list = PCPU_PTR(spinlocks);
1587 	lle = *lock_list;
1588 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1589 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1590 			instance = &(*lock_list)->ll_children[i];
1591 			if (instance->li_lock == lock)
1592 				goto found;
1593 		}
1594 
1595 	/*
1596 	 * When disabling WITNESS through witness_watch we could end up in
1597 	 * having registered locks in the td_sleeplocks queue.
1598 	 * We have to make sure we flush these queues, so just search for
1599 	 * eventual register locks and remove them.
1600 	 */
1601 	if (witness_watch > 0) {
1602 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1603 		    lock->lo_name, fixup_filename(file), line);
1604 		return;
1605 	} else {
1606 		return;
1607 	}
1608 found:
1609 
1610 	/* First, check for shared/exclusive mismatches. */
1611 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1612 	    (flags & LOP_EXCLUSIVE) == 0) {
1613 		witness_output("shared unlock of (%s) %s @ %s:%d\n",
1614 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1615 		witness_output("while exclusively locked from %s:%d\n",
1616 		    fixup_filename(instance->li_file), instance->li_line);
1617 		kassert_panic("excl->ushare");
1618 	}
1619 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1620 	    (flags & LOP_EXCLUSIVE) != 0) {
1621 		witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1622 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1623 		witness_output("while share locked from %s:%d\n",
1624 		    fixup_filename(instance->li_file),
1625 		    instance->li_line);
1626 		kassert_panic("share->uexcl");
1627 	}
1628 	/* If we are recursed, unrecurse. */
1629 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1630 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1631 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1632 		    instance->li_flags);
1633 		instance->li_flags--;
1634 		return;
1635 	}
1636 	/* The lock is now being dropped, check for NORELEASE flag */
1637 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1638 		witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1639 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1640 		kassert_panic("lock marked norelease");
1641 	}
1642 
1643 	/* Otherwise, remove this item from the list. */
1644 	s = intr_disable();
1645 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1646 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1647 	    (*lock_list)->ll_count - 1);
1648 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1649 		(*lock_list)->ll_children[j] =
1650 		    (*lock_list)->ll_children[j + 1];
1651 	(*lock_list)->ll_count--;
1652 	intr_restore(s);
1653 
1654 	/*
1655 	 * In order to reduce contention on w_mtx, we want to keep always an
1656 	 * head object into lists so that frequent allocation from the
1657 	 * free witness pool (and subsequent locking) is avoided.
1658 	 * In order to maintain the current code simple, when the head
1659 	 * object is totally unloaded it means also that we do not have
1660 	 * further objects in the list, so the list ownership needs to be
1661 	 * hand over to another object if the current head needs to be freed.
1662 	 */
1663 	if ((*lock_list)->ll_count == 0) {
1664 		if (*lock_list == lle) {
1665 			if (lle->ll_next == NULL)
1666 				return;
1667 		} else
1668 			lle = *lock_list;
1669 		*lock_list = lle->ll_next;
1670 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1671 		    td->td_proc->p_pid, lle);
1672 		witness_lock_list_free(lle);
1673 	}
1674 }
1675 
1676 void
1677 witness_thread_exit(struct thread *td)
1678 {
1679 	struct lock_list_entry *lle;
1680 	int i, n;
1681 
1682 	lle = td->td_sleeplocks;
1683 	if (lle == NULL || panicstr != NULL)
1684 		return;
1685 	if (lle->ll_count != 0) {
1686 		for (n = 0; lle != NULL; lle = lle->ll_next)
1687 			for (i = lle->ll_count - 1; i >= 0; i--) {
1688 				if (n == 0)
1689 					witness_output(
1690 		    "Thread %p exiting with the following locks held:\n", td);
1691 				n++;
1692 				witness_list_lock(&lle->ll_children[i],
1693 				    witness_output);
1694 
1695 			}
1696 		kassert_panic(
1697 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1698 	}
1699 	witness_lock_list_free(lle);
1700 }
1701 
1702 /*
1703  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1704  * exempt Giant and sleepable locks from the checks as well.  If any
1705  * non-exempt locks are held, then a supplied message is printed to the
1706  * output channel along with a list of the offending locks.  If indicated in the
1707  * flags then a failure results in a panic as well.
1708  */
1709 int
1710 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1711 {
1712 	struct lock_list_entry *lock_list, *lle;
1713 	struct lock_instance *lock1;
1714 	struct thread *td;
1715 	va_list ap;
1716 	int i, n;
1717 
1718 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1719 		return (0);
1720 	n = 0;
1721 	td = curthread;
1722 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1723 		for (i = lle->ll_count - 1; i >= 0; i--) {
1724 			lock1 = &lle->ll_children[i];
1725 			if (lock1->li_lock == lock)
1726 				continue;
1727 			if (flags & WARN_GIANTOK &&
1728 			    lock1->li_lock == &Giant.lock_object)
1729 				continue;
1730 			if (flags & WARN_SLEEPOK &&
1731 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1732 				continue;
1733 			if (n == 0) {
1734 				va_start(ap, fmt);
1735 				witness_voutput(fmt, ap);
1736 				va_end(ap);
1737 				witness_output(
1738 				    " with the following %slocks held:\n",
1739 				    (flags & WARN_SLEEPOK) != 0 ?
1740 				    "non-sleepable " : "");
1741 			}
1742 			n++;
1743 			witness_list_lock(lock1, witness_output);
1744 		}
1745 
1746 	/*
1747 	 * Pin the thread in order to avoid problems with thread migration.
1748 	 * Once that all verifies are passed about spinlocks ownership,
1749 	 * the thread is in a safe path and it can be unpinned.
1750 	 */
1751 	sched_pin();
1752 	lock_list = PCPU_GET(spinlocks);
1753 	if (lock_list != NULL && lock_list->ll_count != 0) {
1754 		sched_unpin();
1755 
1756 		/*
1757 		 * We should only have one spinlock and as long as
1758 		 * the flags cannot match for this locks class,
1759 		 * check if the first spinlock is the one curthread
1760 		 * should hold.
1761 		 */
1762 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1763 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1764 		    lock1->li_lock == lock && n == 0)
1765 			return (0);
1766 
1767 		va_start(ap, fmt);
1768 		witness_voutput(fmt, ap);
1769 		va_end(ap);
1770 		witness_output(" with the following %slocks held:\n",
1771 		    (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1772 		n += witness_list_locks(&lock_list, witness_output);
1773 	} else
1774 		sched_unpin();
1775 	if (flags & WARN_PANIC && n)
1776 		kassert_panic("%s", __func__);
1777 	else
1778 		witness_debugger(n, __func__);
1779 	return (n);
1780 }
1781 
1782 const char *
1783 witness_file(struct lock_object *lock)
1784 {
1785 	struct witness *w;
1786 
1787 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1788 		return ("?");
1789 	w = lock->lo_witness;
1790 	return (w->w_file);
1791 }
1792 
1793 int
1794 witness_line(struct lock_object *lock)
1795 {
1796 	struct witness *w;
1797 
1798 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1799 		return (0);
1800 	w = lock->lo_witness;
1801 	return (w->w_line);
1802 }
1803 
1804 static struct witness *
1805 enroll(const char *description, struct lock_class *lock_class)
1806 {
1807 	struct witness *w;
1808 	struct witness_list *typelist;
1809 
1810 	MPASS(description != NULL);
1811 
1812 	if (witness_watch == -1 || panicstr != NULL)
1813 		return (NULL);
1814 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1815 		if (witness_skipspin)
1816 			return (NULL);
1817 		else
1818 			typelist = &w_spin;
1819 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1820 		typelist = &w_sleep;
1821 	} else {
1822 		kassert_panic("lock class %s is not sleep or spin",
1823 		    lock_class->lc_name);
1824 		return (NULL);
1825 	}
1826 
1827 	mtx_lock_spin(&w_mtx);
1828 	w = witness_hash_get(description);
1829 	if (w)
1830 		goto found;
1831 	if ((w = witness_get()) == NULL)
1832 		return (NULL);
1833 	MPASS(strlen(description) < MAX_W_NAME);
1834 	strcpy(w->w_name, description);
1835 	w->w_class = lock_class;
1836 	w->w_refcount = 1;
1837 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1838 	if (lock_class->lc_flags & LC_SPINLOCK) {
1839 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1840 		w_spin_cnt++;
1841 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1842 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1843 		w_sleep_cnt++;
1844 	}
1845 
1846 	/* Insert new witness into the hash */
1847 	witness_hash_put(w);
1848 	witness_increment_graph_generation();
1849 	mtx_unlock_spin(&w_mtx);
1850 	return (w);
1851 found:
1852 	w->w_refcount++;
1853 	mtx_unlock_spin(&w_mtx);
1854 	if (lock_class != w->w_class)
1855 		kassert_panic(
1856 			"lock (%s) %s does not match earlier (%s) lock",
1857 			description, lock_class->lc_name,
1858 			w->w_class->lc_name);
1859 	return (w);
1860 }
1861 
1862 static void
1863 depart(struct witness *w)
1864 {
1865 	struct witness_list *list;
1866 
1867 	MPASS(w->w_refcount == 0);
1868 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1869 		list = &w_sleep;
1870 		w_sleep_cnt--;
1871 	} else {
1872 		list = &w_spin;
1873 		w_spin_cnt--;
1874 	}
1875 	/*
1876 	 * Set file to NULL as it may point into a loadable module.
1877 	 */
1878 	w->w_file = NULL;
1879 	w->w_line = 0;
1880 	witness_increment_graph_generation();
1881 }
1882 
1883 
1884 static void
1885 adopt(struct witness *parent, struct witness *child)
1886 {
1887 	int pi, ci, i, j;
1888 
1889 	if (witness_cold == 0)
1890 		mtx_assert(&w_mtx, MA_OWNED);
1891 
1892 	/* If the relationship is already known, there's no work to be done. */
1893 	if (isitmychild(parent, child))
1894 		return;
1895 
1896 	/* When the structure of the graph changes, bump up the generation. */
1897 	witness_increment_graph_generation();
1898 
1899 	/*
1900 	 * The hard part ... create the direct relationship, then propagate all
1901 	 * indirect relationships.
1902 	 */
1903 	pi = parent->w_index;
1904 	ci = child->w_index;
1905 	WITNESS_INDEX_ASSERT(pi);
1906 	WITNESS_INDEX_ASSERT(ci);
1907 	MPASS(pi != ci);
1908 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1909 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1910 
1911 	/*
1912 	 * If parent was not already an ancestor of child,
1913 	 * then we increment the descendant and ancestor counters.
1914 	 */
1915 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1916 		parent->w_num_descendants++;
1917 		child->w_num_ancestors++;
1918 	}
1919 
1920 	/*
1921 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1922 	 * an ancestor of 'pi' during this loop.
1923 	 */
1924 	for (i = 1; i <= w_max_used_index; i++) {
1925 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1926 		    (i != pi))
1927 			continue;
1928 
1929 		/* Find each descendant of 'i' and mark it as a descendant. */
1930 		for (j = 1; j <= w_max_used_index; j++) {
1931 
1932 			/*
1933 			 * Skip children that are already marked as
1934 			 * descendants of 'i'.
1935 			 */
1936 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1937 				continue;
1938 
1939 			/*
1940 			 * We are only interested in descendants of 'ci'. Note
1941 			 * that 'ci' itself is counted as a descendant of 'ci'.
1942 			 */
1943 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1944 			    (j != ci))
1945 				continue;
1946 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1947 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1948 			w_data[i].w_num_descendants++;
1949 			w_data[j].w_num_ancestors++;
1950 
1951 			/*
1952 			 * Make sure we aren't marking a node as both an
1953 			 * ancestor and descendant. We should have caught
1954 			 * this as a lock order reversal earlier.
1955 			 */
1956 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1957 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1958 				printf("witness rmatrix paradox! [%d][%d]=%d "
1959 				    "both ancestor and descendant\n",
1960 				    i, j, w_rmatrix[i][j]);
1961 				kdb_backtrace();
1962 				printf("Witness disabled.\n");
1963 				witness_watch = -1;
1964 			}
1965 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1966 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1967 				printf("witness rmatrix paradox! [%d][%d]=%d "
1968 				    "both ancestor and descendant\n",
1969 				    j, i, w_rmatrix[j][i]);
1970 				kdb_backtrace();
1971 				printf("Witness disabled.\n");
1972 				witness_watch = -1;
1973 			}
1974 		}
1975 	}
1976 }
1977 
1978 static void
1979 itismychild(struct witness *parent, struct witness *child)
1980 {
1981 	int unlocked;
1982 
1983 	MPASS(child != NULL && parent != NULL);
1984 	if (witness_cold == 0)
1985 		mtx_assert(&w_mtx, MA_OWNED);
1986 
1987 	if (!witness_lock_type_equal(parent, child)) {
1988 		if (witness_cold == 0) {
1989 			unlocked = 1;
1990 			mtx_unlock_spin(&w_mtx);
1991 		} else {
1992 			unlocked = 0;
1993 		}
1994 		kassert_panic(
1995 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1996 		    "the same lock type", __func__, parent->w_name,
1997 		    parent->w_class->lc_name, child->w_name,
1998 		    child->w_class->lc_name);
1999 		if (unlocked)
2000 			mtx_lock_spin(&w_mtx);
2001 	}
2002 	adopt(parent, child);
2003 }
2004 
2005 /*
2006  * Generic code for the isitmy*() functions. The rmask parameter is the
2007  * expected relationship of w1 to w2.
2008  */
2009 static int
2010 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2011 {
2012 	unsigned char r1, r2;
2013 	int i1, i2;
2014 
2015 	i1 = w1->w_index;
2016 	i2 = w2->w_index;
2017 	WITNESS_INDEX_ASSERT(i1);
2018 	WITNESS_INDEX_ASSERT(i2);
2019 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2020 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2021 
2022 	/* The flags on one better be the inverse of the flags on the other */
2023 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2024 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2025 		/* Don't squawk if we're potentially racing with an update. */
2026 		if (!mtx_owned(&w_mtx))
2027 			return (0);
2028 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2029 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2030 		    "w_rmatrix[%d][%d] == %hhx\n",
2031 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2032 		    i2, i1, r2);
2033 		kdb_backtrace();
2034 		printf("Witness disabled.\n");
2035 		witness_watch = -1;
2036 	}
2037 	return (r1 & rmask);
2038 }
2039 
2040 /*
2041  * Checks if @child is a direct child of @parent.
2042  */
2043 static int
2044 isitmychild(struct witness *parent, struct witness *child)
2045 {
2046 
2047 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2048 }
2049 
2050 /*
2051  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2052  */
2053 static int
2054 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2055 {
2056 
2057 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2058 	    __func__));
2059 }
2060 
2061 #ifdef BLESSING
2062 static int
2063 blessed(struct witness *w1, struct witness *w2)
2064 {
2065 	int i;
2066 	struct witness_blessed *b;
2067 
2068 	for (i = 0; i < nitems(blessed_list); i++) {
2069 		b = &blessed_list[i];
2070 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2071 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2072 				return (1);
2073 			continue;
2074 		}
2075 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2076 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2077 				return (1);
2078 	}
2079 	return (0);
2080 }
2081 #endif
2082 
2083 static struct witness *
2084 witness_get(void)
2085 {
2086 	struct witness *w;
2087 	int index;
2088 
2089 	if (witness_cold == 0)
2090 		mtx_assert(&w_mtx, MA_OWNED);
2091 
2092 	if (witness_watch == -1) {
2093 		mtx_unlock_spin(&w_mtx);
2094 		return (NULL);
2095 	}
2096 	if (STAILQ_EMPTY(&w_free)) {
2097 		witness_watch = -1;
2098 		mtx_unlock_spin(&w_mtx);
2099 		printf("WITNESS: unable to allocate a new witness object\n");
2100 		return (NULL);
2101 	}
2102 	w = STAILQ_FIRST(&w_free);
2103 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2104 	w_free_cnt--;
2105 	index = w->w_index;
2106 	MPASS(index > 0 && index == w_max_used_index+1 &&
2107 	    index < witness_count);
2108 	bzero(w, sizeof(*w));
2109 	w->w_index = index;
2110 	if (index > w_max_used_index)
2111 		w_max_used_index = index;
2112 	return (w);
2113 }
2114 
2115 static void
2116 witness_free(struct witness *w)
2117 {
2118 
2119 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2120 	w_free_cnt++;
2121 }
2122 
2123 static struct lock_list_entry *
2124 witness_lock_list_get(void)
2125 {
2126 	struct lock_list_entry *lle;
2127 
2128 	if (witness_watch == -1)
2129 		return (NULL);
2130 	mtx_lock_spin(&w_mtx);
2131 	lle = w_lock_list_free;
2132 	if (lle == NULL) {
2133 		witness_watch = -1;
2134 		mtx_unlock_spin(&w_mtx);
2135 		printf("%s: witness exhausted\n", __func__);
2136 		return (NULL);
2137 	}
2138 	w_lock_list_free = lle->ll_next;
2139 	mtx_unlock_spin(&w_mtx);
2140 	bzero(lle, sizeof(*lle));
2141 	return (lle);
2142 }
2143 
2144 static void
2145 witness_lock_list_free(struct lock_list_entry *lle)
2146 {
2147 
2148 	mtx_lock_spin(&w_mtx);
2149 	lle->ll_next = w_lock_list_free;
2150 	w_lock_list_free = lle;
2151 	mtx_unlock_spin(&w_mtx);
2152 }
2153 
2154 static struct lock_instance *
2155 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2156 {
2157 	struct lock_list_entry *lle;
2158 	struct lock_instance *instance;
2159 	int i;
2160 
2161 	for (lle = list; lle != NULL; lle = lle->ll_next)
2162 		for (i = lle->ll_count - 1; i >= 0; i--) {
2163 			instance = &lle->ll_children[i];
2164 			if (instance->li_lock == lock)
2165 				return (instance);
2166 		}
2167 	return (NULL);
2168 }
2169 
2170 static void
2171 witness_list_lock(struct lock_instance *instance,
2172     int (*prnt)(const char *fmt, ...))
2173 {
2174 	struct lock_object *lock;
2175 
2176 	lock = instance->li_lock;
2177 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2178 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2179 	if (lock->lo_witness->w_name != lock->lo_name)
2180 		prnt(" (%s)", lock->lo_witness->w_name);
2181 	prnt(" r = %d (%p) locked @ %s:%d\n",
2182 	    instance->li_flags & LI_RECURSEMASK, lock,
2183 	    fixup_filename(instance->li_file), instance->li_line);
2184 }
2185 
2186 static int
2187 witness_output(const char *fmt, ...)
2188 {
2189 	va_list ap;
2190 	int ret;
2191 
2192 	va_start(ap, fmt);
2193 	ret = witness_voutput(fmt, ap);
2194 	va_end(ap);
2195 	return (ret);
2196 }
2197 
2198 static int
2199 witness_voutput(const char *fmt, va_list ap)
2200 {
2201 	int ret;
2202 
2203 	ret = 0;
2204 	switch (witness_channel) {
2205 	case WITNESS_CONSOLE:
2206 		ret = vprintf(fmt, ap);
2207 		break;
2208 	case WITNESS_LOG:
2209 		vlog(LOG_NOTICE, fmt, ap);
2210 		break;
2211 	case WITNESS_NONE:
2212 		break;
2213 	}
2214 	return (ret);
2215 }
2216 
2217 #ifdef DDB
2218 static int
2219 witness_thread_has_locks(struct thread *td)
2220 {
2221 
2222 	if (td->td_sleeplocks == NULL)
2223 		return (0);
2224 	return (td->td_sleeplocks->ll_count != 0);
2225 }
2226 
2227 static int
2228 witness_proc_has_locks(struct proc *p)
2229 {
2230 	struct thread *td;
2231 
2232 	FOREACH_THREAD_IN_PROC(p, td) {
2233 		if (witness_thread_has_locks(td))
2234 			return (1);
2235 	}
2236 	return (0);
2237 }
2238 #endif
2239 
2240 int
2241 witness_list_locks(struct lock_list_entry **lock_list,
2242     int (*prnt)(const char *fmt, ...))
2243 {
2244 	struct lock_list_entry *lle;
2245 	int i, nheld;
2246 
2247 	nheld = 0;
2248 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2249 		for (i = lle->ll_count - 1; i >= 0; i--) {
2250 			witness_list_lock(&lle->ll_children[i], prnt);
2251 			nheld++;
2252 		}
2253 	return (nheld);
2254 }
2255 
2256 /*
2257  * This is a bit risky at best.  We call this function when we have timed
2258  * out acquiring a spin lock, and we assume that the other CPU is stuck
2259  * with this lock held.  So, we go groveling around in the other CPU's
2260  * per-cpu data to try to find the lock instance for this spin lock to
2261  * see when it was last acquired.
2262  */
2263 void
2264 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2265     int (*prnt)(const char *fmt, ...))
2266 {
2267 	struct lock_instance *instance;
2268 	struct pcpu *pc;
2269 
2270 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2271 		return;
2272 	pc = pcpu_find(owner->td_oncpu);
2273 	instance = find_instance(pc->pc_spinlocks, lock);
2274 	if (instance != NULL)
2275 		witness_list_lock(instance, prnt);
2276 }
2277 
2278 void
2279 witness_save(struct lock_object *lock, const char **filep, int *linep)
2280 {
2281 	struct lock_list_entry *lock_list;
2282 	struct lock_instance *instance;
2283 	struct lock_class *class;
2284 
2285 	/*
2286 	 * This function is used independently in locking code to deal with
2287 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2288 	 * is gone.
2289 	 */
2290 	if (SCHEDULER_STOPPED())
2291 		return;
2292 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2293 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2294 		return;
2295 	class = LOCK_CLASS(lock);
2296 	if (class->lc_flags & LC_SLEEPLOCK)
2297 		lock_list = curthread->td_sleeplocks;
2298 	else {
2299 		if (witness_skipspin)
2300 			return;
2301 		lock_list = PCPU_GET(spinlocks);
2302 	}
2303 	instance = find_instance(lock_list, lock);
2304 	if (instance == NULL) {
2305 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2306 		    class->lc_name, lock->lo_name);
2307 		return;
2308 	}
2309 	*filep = instance->li_file;
2310 	*linep = instance->li_line;
2311 }
2312 
2313 void
2314 witness_restore(struct lock_object *lock, const char *file, int line)
2315 {
2316 	struct lock_list_entry *lock_list;
2317 	struct lock_instance *instance;
2318 	struct lock_class *class;
2319 
2320 	/*
2321 	 * This function is used independently in locking code to deal with
2322 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2323 	 * is gone.
2324 	 */
2325 	if (SCHEDULER_STOPPED())
2326 		return;
2327 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2328 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2329 		return;
2330 	class = LOCK_CLASS(lock);
2331 	if (class->lc_flags & LC_SLEEPLOCK)
2332 		lock_list = curthread->td_sleeplocks;
2333 	else {
2334 		if (witness_skipspin)
2335 			return;
2336 		lock_list = PCPU_GET(spinlocks);
2337 	}
2338 	instance = find_instance(lock_list, lock);
2339 	if (instance == NULL)
2340 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2341 		    class->lc_name, lock->lo_name);
2342 	lock->lo_witness->w_file = file;
2343 	lock->lo_witness->w_line = line;
2344 	if (instance == NULL)
2345 		return;
2346 	instance->li_file = file;
2347 	instance->li_line = line;
2348 }
2349 
2350 void
2351 witness_assert(const struct lock_object *lock, int flags, const char *file,
2352     int line)
2353 {
2354 #ifdef INVARIANT_SUPPORT
2355 	struct lock_instance *instance;
2356 	struct lock_class *class;
2357 
2358 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2359 		return;
2360 	class = LOCK_CLASS(lock);
2361 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2362 		instance = find_instance(curthread->td_sleeplocks, lock);
2363 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2364 		instance = find_instance(PCPU_GET(spinlocks), lock);
2365 	else {
2366 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2367 		    class->lc_name, lock->lo_name);
2368 		return;
2369 	}
2370 	switch (flags) {
2371 	case LA_UNLOCKED:
2372 		if (instance != NULL)
2373 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2374 			    class->lc_name, lock->lo_name,
2375 			    fixup_filename(file), line);
2376 		break;
2377 	case LA_LOCKED:
2378 	case LA_LOCKED | LA_RECURSED:
2379 	case LA_LOCKED | LA_NOTRECURSED:
2380 	case LA_SLOCKED:
2381 	case LA_SLOCKED | LA_RECURSED:
2382 	case LA_SLOCKED | LA_NOTRECURSED:
2383 	case LA_XLOCKED:
2384 	case LA_XLOCKED | LA_RECURSED:
2385 	case LA_XLOCKED | LA_NOTRECURSED:
2386 		if (instance == NULL) {
2387 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2388 			    class->lc_name, lock->lo_name,
2389 			    fixup_filename(file), line);
2390 			break;
2391 		}
2392 		if ((flags & LA_XLOCKED) != 0 &&
2393 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2394 			kassert_panic(
2395 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2396 			    class->lc_name, lock->lo_name,
2397 			    fixup_filename(file), line);
2398 		if ((flags & LA_SLOCKED) != 0 &&
2399 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2400 			kassert_panic(
2401 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2402 			    class->lc_name, lock->lo_name,
2403 			    fixup_filename(file), line);
2404 		if ((flags & LA_RECURSED) != 0 &&
2405 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2406 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2407 			    class->lc_name, lock->lo_name,
2408 			    fixup_filename(file), line);
2409 		if ((flags & LA_NOTRECURSED) != 0 &&
2410 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2411 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2412 			    class->lc_name, lock->lo_name,
2413 			    fixup_filename(file), line);
2414 		break;
2415 	default:
2416 		kassert_panic("Invalid lock assertion at %s:%d.",
2417 		    fixup_filename(file), line);
2418 
2419 	}
2420 #endif	/* INVARIANT_SUPPORT */
2421 }
2422 
2423 static void
2424 witness_setflag(struct lock_object *lock, int flag, int set)
2425 {
2426 	struct lock_list_entry *lock_list;
2427 	struct lock_instance *instance;
2428 	struct lock_class *class;
2429 
2430 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2431 		return;
2432 	class = LOCK_CLASS(lock);
2433 	if (class->lc_flags & LC_SLEEPLOCK)
2434 		lock_list = curthread->td_sleeplocks;
2435 	else {
2436 		if (witness_skipspin)
2437 			return;
2438 		lock_list = PCPU_GET(spinlocks);
2439 	}
2440 	instance = find_instance(lock_list, lock);
2441 	if (instance == NULL) {
2442 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2443 		    class->lc_name, lock->lo_name);
2444 		return;
2445 	}
2446 
2447 	if (set)
2448 		instance->li_flags |= flag;
2449 	else
2450 		instance->li_flags &= ~flag;
2451 }
2452 
2453 void
2454 witness_norelease(struct lock_object *lock)
2455 {
2456 
2457 	witness_setflag(lock, LI_NORELEASE, 1);
2458 }
2459 
2460 void
2461 witness_releaseok(struct lock_object *lock)
2462 {
2463 
2464 	witness_setflag(lock, LI_NORELEASE, 0);
2465 }
2466 
2467 #ifdef DDB
2468 static void
2469 witness_ddb_list(struct thread *td)
2470 {
2471 
2472 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2473 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2474 
2475 	if (witness_watch < 1)
2476 		return;
2477 
2478 	witness_list_locks(&td->td_sleeplocks, db_printf);
2479 
2480 	/*
2481 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2482 	 * if td is currently executing on some other CPU and holds spin locks
2483 	 * as we won't display those locks.  If we had a MI way of getting
2484 	 * the per-cpu data for a given cpu then we could use
2485 	 * td->td_oncpu to get the list of spinlocks for this thread
2486 	 * and "fix" this.
2487 	 *
2488 	 * That still wouldn't really fix this unless we locked the scheduler
2489 	 * lock or stopped the other CPU to make sure it wasn't changing the
2490 	 * list out from under us.  It is probably best to just not try to
2491 	 * handle threads on other CPU's for now.
2492 	 */
2493 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2494 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2495 }
2496 
2497 DB_SHOW_COMMAND(locks, db_witness_list)
2498 {
2499 	struct thread *td;
2500 
2501 	if (have_addr)
2502 		td = db_lookup_thread(addr, true);
2503 	else
2504 		td = kdb_thread;
2505 	witness_ddb_list(td);
2506 }
2507 
2508 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2509 {
2510 	struct thread *td;
2511 	struct proc *p;
2512 
2513 	/*
2514 	 * It would be nice to list only threads and processes that actually
2515 	 * held sleep locks, but that information is currently not exported
2516 	 * by WITNESS.
2517 	 */
2518 	FOREACH_PROC_IN_SYSTEM(p) {
2519 		if (!witness_proc_has_locks(p))
2520 			continue;
2521 		FOREACH_THREAD_IN_PROC(p, td) {
2522 			if (!witness_thread_has_locks(td))
2523 				continue;
2524 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2525 			    p->p_comm, td, td->td_tid);
2526 			witness_ddb_list(td);
2527 			if (db_pager_quit)
2528 				return;
2529 		}
2530 	}
2531 }
2532 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2533 
2534 DB_SHOW_COMMAND(witness, db_witness_display)
2535 {
2536 
2537 	witness_ddb_display(db_printf);
2538 }
2539 #endif
2540 
2541 static int
2542 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2543 {
2544 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2545 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2546 	struct sbuf *sb;
2547 	u_int w_rmatrix1, w_rmatrix2;
2548 	int error, generation, i, j;
2549 
2550 	tmp_data1 = NULL;
2551 	tmp_data2 = NULL;
2552 	tmp_w1 = NULL;
2553 	tmp_w2 = NULL;
2554 	if (witness_watch < 1) {
2555 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2556 		return (error);
2557 	}
2558 	if (witness_cold) {
2559 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2560 		return (error);
2561 	}
2562 	error = 0;
2563 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2564 	if (sb == NULL)
2565 		return (ENOMEM);
2566 
2567 	/* Allocate and init temporary storage space. */
2568 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2569 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2570 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2571 	    M_WAITOK | M_ZERO);
2572 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2573 	    M_WAITOK | M_ZERO);
2574 	stack_zero(&tmp_data1->wlod_stack);
2575 	stack_zero(&tmp_data2->wlod_stack);
2576 
2577 restart:
2578 	mtx_lock_spin(&w_mtx);
2579 	generation = w_generation;
2580 	mtx_unlock_spin(&w_mtx);
2581 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2582 	    w_lohash.wloh_count);
2583 	for (i = 1; i < w_max_used_index; i++) {
2584 		mtx_lock_spin(&w_mtx);
2585 		if (generation != w_generation) {
2586 			mtx_unlock_spin(&w_mtx);
2587 
2588 			/* The graph has changed, try again. */
2589 			req->oldidx = 0;
2590 			sbuf_clear(sb);
2591 			goto restart;
2592 		}
2593 
2594 		w1 = &w_data[i];
2595 		if (w1->w_reversed == 0) {
2596 			mtx_unlock_spin(&w_mtx);
2597 			continue;
2598 		}
2599 
2600 		/* Copy w1 locally so we can release the spin lock. */
2601 		*tmp_w1 = *w1;
2602 		mtx_unlock_spin(&w_mtx);
2603 
2604 		if (tmp_w1->w_reversed == 0)
2605 			continue;
2606 		for (j = 1; j < w_max_used_index; j++) {
2607 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2608 				continue;
2609 
2610 			mtx_lock_spin(&w_mtx);
2611 			if (generation != w_generation) {
2612 				mtx_unlock_spin(&w_mtx);
2613 
2614 				/* The graph has changed, try again. */
2615 				req->oldidx = 0;
2616 				sbuf_clear(sb);
2617 				goto restart;
2618 			}
2619 
2620 			w2 = &w_data[j];
2621 			data1 = witness_lock_order_get(w1, w2);
2622 			data2 = witness_lock_order_get(w2, w1);
2623 
2624 			/*
2625 			 * Copy information locally so we can release the
2626 			 * spin lock.
2627 			 */
2628 			*tmp_w2 = *w2;
2629 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2630 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2631 
2632 			if (data1) {
2633 				stack_zero(&tmp_data1->wlod_stack);
2634 				stack_copy(&data1->wlod_stack,
2635 				    &tmp_data1->wlod_stack);
2636 			}
2637 			if (data2 && data2 != data1) {
2638 				stack_zero(&tmp_data2->wlod_stack);
2639 				stack_copy(&data2->wlod_stack,
2640 				    &tmp_data2->wlod_stack);
2641 			}
2642 			mtx_unlock_spin(&w_mtx);
2643 
2644 			sbuf_printf(sb,
2645 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2646 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2647 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2648 			if (data1) {
2649 				sbuf_printf(sb,
2650 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2651 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2652 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2653 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2654 				sbuf_printf(sb, "\n");
2655 			}
2656 			if (data2 && data2 != data1) {
2657 				sbuf_printf(sb,
2658 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2659 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2660 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2661 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2662 				sbuf_printf(sb, "\n");
2663 			}
2664 		}
2665 	}
2666 	mtx_lock_spin(&w_mtx);
2667 	if (generation != w_generation) {
2668 		mtx_unlock_spin(&w_mtx);
2669 
2670 		/*
2671 		 * The graph changed while we were printing stack data,
2672 		 * try again.
2673 		 */
2674 		req->oldidx = 0;
2675 		sbuf_clear(sb);
2676 		goto restart;
2677 	}
2678 	mtx_unlock_spin(&w_mtx);
2679 
2680 	/* Free temporary storage space. */
2681 	free(tmp_data1, M_TEMP);
2682 	free(tmp_data2, M_TEMP);
2683 	free(tmp_w1, M_TEMP);
2684 	free(tmp_w2, M_TEMP);
2685 
2686 	sbuf_finish(sb);
2687 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2688 	sbuf_delete(sb);
2689 
2690 	return (error);
2691 }
2692 
2693 static int
2694 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2695 {
2696 	static const struct {
2697 		enum witness_channel channel;
2698 		const char *name;
2699 	} channels[] = {
2700 		{ WITNESS_CONSOLE, "console" },
2701 		{ WITNESS_LOG, "log" },
2702 		{ WITNESS_NONE, "none" },
2703 	};
2704 	char buf[16];
2705 	u_int i;
2706 	int error;
2707 
2708 	buf[0] = '\0';
2709 	for (i = 0; i < nitems(channels); i++)
2710 		if (witness_channel == channels[i].channel) {
2711 			snprintf(buf, sizeof(buf), "%s", channels[i].name);
2712 			break;
2713 		}
2714 
2715 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2716 	if (error != 0 || req->newptr == NULL)
2717 		return (error);
2718 
2719 	error = EINVAL;
2720 	for (i = 0; i < nitems(channels); i++)
2721 		if (strcmp(channels[i].name, buf) == 0) {
2722 			witness_channel = channels[i].channel;
2723 			error = 0;
2724 			break;
2725 		}
2726 	return (error);
2727 }
2728 
2729 static int
2730 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2731 {
2732 	struct witness *w;
2733 	struct sbuf *sb;
2734 	int error;
2735 
2736 	if (witness_watch < 1) {
2737 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2738 		return (error);
2739 	}
2740 	if (witness_cold) {
2741 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2742 		return (error);
2743 	}
2744 	error = 0;
2745 
2746 	error = sysctl_wire_old_buffer(req, 0);
2747 	if (error != 0)
2748 		return (error);
2749 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2750 	if (sb == NULL)
2751 		return (ENOMEM);
2752 	sbuf_printf(sb, "\n");
2753 
2754 	mtx_lock_spin(&w_mtx);
2755 	STAILQ_FOREACH(w, &w_all, w_list)
2756 		w->w_displayed = 0;
2757 	STAILQ_FOREACH(w, &w_all, w_list)
2758 		witness_add_fullgraph(sb, w);
2759 	mtx_unlock_spin(&w_mtx);
2760 
2761 	/*
2762 	 * Close the sbuf and return to userland.
2763 	 */
2764 	error = sbuf_finish(sb);
2765 	sbuf_delete(sb);
2766 
2767 	return (error);
2768 }
2769 
2770 static int
2771 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2772 {
2773 	int error, value;
2774 
2775 	value = witness_watch;
2776 	error = sysctl_handle_int(oidp, &value, 0, req);
2777 	if (error != 0 || req->newptr == NULL)
2778 		return (error);
2779 	if (value > 1 || value < -1 ||
2780 	    (witness_watch == -1 && value != witness_watch))
2781 		return (EINVAL);
2782 	witness_watch = value;
2783 	return (0);
2784 }
2785 
2786 static void
2787 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2788 {
2789 	int i;
2790 
2791 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2792 		return;
2793 	w->w_displayed = 1;
2794 
2795 	WITNESS_INDEX_ASSERT(w->w_index);
2796 	for (i = 1; i <= w_max_used_index; i++) {
2797 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2798 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2799 			    w_data[i].w_name);
2800 			witness_add_fullgraph(sb, &w_data[i]);
2801 		}
2802 	}
2803 }
2804 
2805 /*
2806  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2807  * interprets the key as a string and reads until the null
2808  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2809  * hash value computed from the key.
2810  */
2811 static uint32_t
2812 witness_hash_djb2(const uint8_t *key, uint32_t size)
2813 {
2814 	unsigned int hash = 5381;
2815 	int i;
2816 
2817 	/* hash = hash * 33 + key[i] */
2818 	if (size)
2819 		for (i = 0; i < size; i++)
2820 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2821 	else
2822 		for (i = 0; key[i] != 0; i++)
2823 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2824 
2825 	return (hash);
2826 }
2827 
2828 
2829 /*
2830  * Initializes the two witness hash tables. Called exactly once from
2831  * witness_initialize().
2832  */
2833 static void
2834 witness_init_hash_tables(void)
2835 {
2836 	int i;
2837 
2838 	MPASS(witness_cold);
2839 
2840 	/* Initialize the hash tables. */
2841 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2842 		w_hash.wh_array[i] = NULL;
2843 
2844 	w_hash.wh_size = WITNESS_HASH_SIZE;
2845 	w_hash.wh_count = 0;
2846 
2847 	/* Initialize the lock order data hash. */
2848 	w_lofree = NULL;
2849 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2850 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2851 		w_lodata[i].wlod_next = w_lofree;
2852 		w_lofree = &w_lodata[i];
2853 	}
2854 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2855 	w_lohash.wloh_count = 0;
2856 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2857 		w_lohash.wloh_array[i] = NULL;
2858 }
2859 
2860 static struct witness *
2861 witness_hash_get(const char *key)
2862 {
2863 	struct witness *w;
2864 	uint32_t hash;
2865 
2866 	MPASS(key != NULL);
2867 	if (witness_cold == 0)
2868 		mtx_assert(&w_mtx, MA_OWNED);
2869 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2870 	w = w_hash.wh_array[hash];
2871 	while (w != NULL) {
2872 		if (strcmp(w->w_name, key) == 0)
2873 			goto out;
2874 		w = w->w_hash_next;
2875 	}
2876 
2877 out:
2878 	return (w);
2879 }
2880 
2881 static void
2882 witness_hash_put(struct witness *w)
2883 {
2884 	uint32_t hash;
2885 
2886 	MPASS(w != NULL);
2887 	MPASS(w->w_name != NULL);
2888 	if (witness_cold == 0)
2889 		mtx_assert(&w_mtx, MA_OWNED);
2890 	KASSERT(witness_hash_get(w->w_name) == NULL,
2891 	    ("%s: trying to add a hash entry that already exists!", __func__));
2892 	KASSERT(w->w_hash_next == NULL,
2893 	    ("%s: w->w_hash_next != NULL", __func__));
2894 
2895 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2896 	w->w_hash_next = w_hash.wh_array[hash];
2897 	w_hash.wh_array[hash] = w;
2898 	w_hash.wh_count++;
2899 }
2900 
2901 
2902 static struct witness_lock_order_data *
2903 witness_lock_order_get(struct witness *parent, struct witness *child)
2904 {
2905 	struct witness_lock_order_data *data = NULL;
2906 	struct witness_lock_order_key key;
2907 	unsigned int hash;
2908 
2909 	MPASS(parent != NULL && child != NULL);
2910 	key.from = parent->w_index;
2911 	key.to = child->w_index;
2912 	WITNESS_INDEX_ASSERT(key.from);
2913 	WITNESS_INDEX_ASSERT(key.to);
2914 	if ((w_rmatrix[parent->w_index][child->w_index]
2915 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2916 		goto out;
2917 
2918 	hash = witness_hash_djb2((const char*)&key,
2919 	    sizeof(key)) % w_lohash.wloh_size;
2920 	data = w_lohash.wloh_array[hash];
2921 	while (data != NULL) {
2922 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2923 			break;
2924 		data = data->wlod_next;
2925 	}
2926 
2927 out:
2928 	return (data);
2929 }
2930 
2931 /*
2932  * Verify that parent and child have a known relationship, are not the same,
2933  * and child is actually a child of parent.  This is done without w_mtx
2934  * to avoid contention in the common case.
2935  */
2936 static int
2937 witness_lock_order_check(struct witness *parent, struct witness *child)
2938 {
2939 
2940 	if (parent != child &&
2941 	    w_rmatrix[parent->w_index][child->w_index]
2942 	    & WITNESS_LOCK_ORDER_KNOWN &&
2943 	    isitmychild(parent, child))
2944 		return (1);
2945 
2946 	return (0);
2947 }
2948 
2949 static int
2950 witness_lock_order_add(struct witness *parent, struct witness *child)
2951 {
2952 	struct witness_lock_order_data *data = NULL;
2953 	struct witness_lock_order_key key;
2954 	unsigned int hash;
2955 
2956 	MPASS(parent != NULL && child != NULL);
2957 	key.from = parent->w_index;
2958 	key.to = child->w_index;
2959 	WITNESS_INDEX_ASSERT(key.from);
2960 	WITNESS_INDEX_ASSERT(key.to);
2961 	if (w_rmatrix[parent->w_index][child->w_index]
2962 	    & WITNESS_LOCK_ORDER_KNOWN)
2963 		return (1);
2964 
2965 	hash = witness_hash_djb2((const char*)&key,
2966 	    sizeof(key)) % w_lohash.wloh_size;
2967 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2968 	data = w_lofree;
2969 	if (data == NULL)
2970 		return (0);
2971 	w_lofree = data->wlod_next;
2972 	data->wlod_next = w_lohash.wloh_array[hash];
2973 	data->wlod_key = key;
2974 	w_lohash.wloh_array[hash] = data;
2975 	w_lohash.wloh_count++;
2976 	stack_zero(&data->wlod_stack);
2977 	stack_save(&data->wlod_stack);
2978 	return (1);
2979 }
2980 
2981 /* Call this whenever the structure of the witness graph changes. */
2982 static void
2983 witness_increment_graph_generation(void)
2984 {
2985 
2986 	if (witness_cold == 0)
2987 		mtx_assert(&w_mtx, MA_OWNED);
2988 	w_generation++;
2989 }
2990 
2991 static int
2992 witness_output_drain(void *arg __unused, const char *data, int len)
2993 {
2994 
2995 	witness_output("%.*s", len, data);
2996 	return (len);
2997 }
2998 
2999 static void
3000 witness_debugger(int cond, const char *msg)
3001 {
3002 	char buf[32];
3003 	struct sbuf sb;
3004 	struct stack st;
3005 
3006 	if (!cond)
3007 		return;
3008 
3009 	if (witness_trace) {
3010 		sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3011 		sbuf_set_drain(&sb, witness_output_drain, NULL);
3012 
3013 		stack_zero(&st);
3014 		stack_save(&st);
3015 		witness_output("stack backtrace:\n");
3016 		stack_sbuf_print_ddb(&sb, &st);
3017 
3018 		sbuf_finish(&sb);
3019 	}
3020 
3021 #ifdef KDB
3022 	if (witness_kdb)
3023 		kdb_enter(KDB_WHY_WITNESS, msg);
3024 #endif
3025 }
3026