xref: /freebsd/sys/kern/subr_witness.c (revision 7660b554bc59a07be0431c17e0e33815818baa69)
1 /*-
2  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
13  *    promote products derived from this software without specific prior
14  *    written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30  */
31 
32 /*
33  * Implementation of the `witness' lock verifier.  Originally implemented for
34  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
35  * classes in FreeBSD.
36  */
37 
38 /*
39  *	Main Entry: witness
40  *	Pronunciation: 'wit-n&s
41  *	Function: noun
42  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
43  *	    testimony, witness, from 2wit
44  *	Date: before 12th century
45  *	1 : attestation of a fact or event : TESTIMONY
46  *	2 : one that gives evidence; specifically : one who testifies in
47  *	    a cause or before a judicial tribunal
48  *	3 : one asked to be present at a transaction so as to be able to
49  *	    testify to its having taken place
50  *	4 : one who has personal knowledge of something
51  *	5 a : something serving as evidence or proof : SIGN
52  *	  b : public affirmation by word or example of usually
53  *	      religious faith or conviction <the heroic witness to divine
54  *	      life -- Pilot>
55  *	6 capitalized : a member of the Jehovah's Witnesses
56  */
57 
58 /*
59  * Special rules concerning Giant and lock orders:
60  *
61  * 1) Giant must be acquired before any other mutexes.  Stated another way,
62  *    no other mutex may be held when Giant is acquired.
63  *
64  * 2) Giant must be released when blocking on a sleepable lock.
65  *
66  * This rule is less obvious, but is a result of Giant providing the same
67  * semantics as spl().  Basically, when a thread sleeps, it must release
68  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
69  * 2).
70  *
71  * 3) Giant may be acquired before or after sleepable locks.
72  *
73  * This rule is also not quite as obvious.  Giant may be acquired after
74  * a sleepable lock because it is a non-sleepable lock and non-sleepable
75  * locks may always be acquired while holding a sleepable lock.  The second
76  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
77  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
78  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
79  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
80  * execute.  Thus, acquiring Giant both before and after a sleepable lock
81  * will not result in a lock order reversal.
82  */
83 
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86 
87 #include "opt_ddb.h"
88 #include "opt_witness.h"
89 
90 #include <sys/param.h>
91 #include <sys/bus.h>
92 #include <sys/kernel.h>
93 #include <sys/ktr.h>
94 #include <sys/lock.h>
95 #include <sys/malloc.h>
96 #include <sys/mutex.h>
97 #include <sys/proc.h>
98 #include <sys/sysctl.h>
99 #include <sys/systm.h>
100 
101 #include <ddb/ddb.h>
102 
103 #include <machine/stdarg.h>
104 
105 /* Define this to check for blessed mutexes */
106 #undef BLESSING
107 
108 #define WITNESS_COUNT 200
109 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
110 /*
111  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
112  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
113  * probably be safe for the most part, but it's still a SWAG.
114  */
115 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
116 
117 #define	WITNESS_NCHILDREN 6
118 
119 struct witness_child_list_entry;
120 
121 struct witness {
122 	const	char *w_name;
123 	struct	lock_class *w_class;
124 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
125 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
126 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
127 	const	char *w_file;
128 	int	w_line;
129 	u_int	w_level;
130 	u_int	w_refcount;
131 	u_char	w_Giant_squawked:1;
132 	u_char	w_other_squawked:1;
133 	u_char	w_same_squawked:1;
134 	u_char	w_displayed:1;
135 };
136 
137 struct witness_child_list_entry {
138 	struct	witness_child_list_entry *wcl_next;
139 	struct	witness *wcl_children[WITNESS_NCHILDREN];
140 	u_int	wcl_count;
141 };
142 
143 STAILQ_HEAD(witness_list, witness);
144 
145 #ifdef BLESSING
146 struct witness_blessed {
147 	const	char *b_lock1;
148 	const	char *b_lock2;
149 };
150 #endif
151 
152 struct witness_order_list_entry {
153 	const	char *w_name;
154 	struct	lock_class *w_class;
155 };
156 
157 #ifdef BLESSING
158 static int	blessed(struct witness *, struct witness *);
159 #endif
160 static int	depart(struct witness *w);
161 static struct	witness *enroll(const char *description,
162 				struct lock_class *lock_class);
163 static int	insertchild(struct witness *parent, struct witness *child);
164 static int	isitmychild(struct witness *parent, struct witness *child);
165 static int	isitmydescendant(struct witness *parent, struct witness *child);
166 static int	itismychild(struct witness *parent, struct witness *child);
167 static int	rebalancetree(struct witness_list *list);
168 static void	removechild(struct witness *parent, struct witness *child);
169 static int	reparentchildren(struct witness *newparent,
170 		    struct witness *oldparent);
171 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
172 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
173 					   struct witness *, int indent);
174 static const char *fixup_filename(const char *file);
175 static void	witness_leveldescendents(struct witness *parent, int level);
176 static void	witness_levelall(void);
177 static struct	witness *witness_get(void);
178 static void	witness_free(struct witness *m);
179 static struct	witness_child_list_entry *witness_child_get(void);
180 static void	witness_child_free(struct witness_child_list_entry *wcl);
181 static struct	lock_list_entry *witness_lock_list_get(void);
182 static void	witness_lock_list_free(struct lock_list_entry *lle);
183 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
184 					     struct lock_object *lock);
185 static void	witness_list_lock(struct lock_instance *instance);
186 #ifdef DDB
187 static void	witness_list(struct thread *td);
188 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
189 				     struct witness_list *list);
190 static void	witness_display(void(*)(const char *fmt, ...));
191 #endif
192 
193 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
194 
195 /*
196  * If set to 0, witness is disabled.  If set to 1, witness performs full lock
197  * order checking for all locks.  If set to 2 or higher, then witness skips
198  * the full lock order check if the lock being acquired is at a higher level
199  * (i.e. farther down in the tree) than the current lock.  This last mode is
200  * somewhat experimental and not considered fully safe.  At runtime, this
201  * value may be set to 0 to turn off witness.  witness is not allowed be
202  * turned on once it is turned off, however.
203  */
204 static int witness_watch = 1;
205 TUNABLE_INT("debug.witness_watch", &witness_watch);
206 SYSCTL_PROC(_debug, OID_AUTO, witness_watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
207     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
208 
209 #ifdef DDB
210 /*
211  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
212  * drop into kdebug() when:
213  *	- a lock heirarchy violation occurs
214  *	- locks are held when going to sleep.
215  */
216 #ifdef WITNESS_DDB
217 int	witness_ddb = 1;
218 #else
219 int	witness_ddb = 0;
220 #endif
221 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
222 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
223 
224 /*
225  * When DDB is enabled and witness_trace is set to 1, it will cause the system
226  * to print a stack trace:
227  *	- a lock heirarchy violation occurs
228  *	- locks are held when going to sleep.
229  */
230 int	witness_trace = 1;
231 TUNABLE_INT("debug.witness_trace", &witness_trace);
232 SYSCTL_INT(_debug, OID_AUTO, witness_trace, CTLFLAG_RW, &witness_trace, 0, "");
233 #endif /* DDB */
234 
235 #ifdef WITNESS_SKIPSPIN
236 int	witness_skipspin = 1;
237 #else
238 int	witness_skipspin = 0;
239 #endif
240 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
241 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
242     "");
243 
244 static struct mtx w_mtx;
245 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
246 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
247 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
248 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
249 static struct witness_child_list_entry *w_child_free = NULL;
250 static struct lock_list_entry *w_lock_list_free = NULL;
251 
252 static struct witness w_data[WITNESS_COUNT];
253 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
254 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
255 
256 static struct witness_order_list_entry order_lists[] = {
257 	{ "proctree", &lock_class_sx },
258 	{ "allproc", &lock_class_sx },
259 	{ "Giant", &lock_class_mtx_sleep },
260 	{ "filedesc structure", &lock_class_mtx_sleep },
261 	{ "pipe mutex", &lock_class_mtx_sleep },
262 	{ "sigio lock", &lock_class_mtx_sleep },
263 	{ "process group", &lock_class_mtx_sleep },
264 	{ "process lock", &lock_class_mtx_sleep },
265 	{ "session", &lock_class_mtx_sleep },
266 	{ "uidinfo hash", &lock_class_mtx_sleep },
267 	{ "uidinfo struct", &lock_class_mtx_sleep },
268 	{ "allprison", &lock_class_mtx_sleep },
269 	{ NULL, NULL },
270 	/*
271 	 * spin locks
272 	 */
273 #ifdef SMP
274 	{ "ap boot", &lock_class_mtx_spin },
275 #ifdef __i386__
276 	{ "com", &lock_class_mtx_spin },
277 #endif
278 #endif
279 	{ "sio", &lock_class_mtx_spin },
280 #ifdef __i386__
281 	{ "cy", &lock_class_mtx_spin },
282 #endif
283 	{ "sabtty", &lock_class_mtx_spin },
284 	{ "zstty", &lock_class_mtx_spin },
285 	{ "ng_node", &lock_class_mtx_spin },
286 	{ "ng_worklist", &lock_class_mtx_spin },
287 	{ "taskqueue_fast", &lock_class_mtx_spin },
288 	{ "ithread table lock", &lock_class_mtx_spin },
289 	{ "sched lock", &lock_class_mtx_spin },
290 	{ "callout", &lock_class_mtx_spin },
291 	/*
292 	 * leaf locks
293 	 */
294 	{ "allpmaps", &lock_class_mtx_spin },
295 	{ "vm page queue free mutex", &lock_class_mtx_spin },
296 	{ "icu", &lock_class_mtx_spin },
297 #ifdef SMP
298 	{ "smp rendezvous", &lock_class_mtx_spin },
299 #if defined(__i386__) && defined(APIC_IO)
300 	{ "tlb", &lock_class_mtx_spin },
301 #endif
302 #ifdef __i386__
303 	{ "lazypmap", &lock_class_mtx_spin },
304 #endif
305 #ifdef __sparc64__
306 	{ "ipi", &lock_class_mtx_spin },
307 #endif
308 #endif
309 	{ "clk", &lock_class_mtx_spin },
310 	{ "mutex profiling lock", &lock_class_mtx_spin },
311 	{ "kse zombie lock", &lock_class_mtx_spin },
312 	{ "ALD Queue", &lock_class_mtx_spin },
313 #ifdef __ia64__
314 	{ "MCA spin lock", &lock_class_mtx_spin },
315 #endif
316 #if defined(__i386__) || defined(__amd64__)
317 	{ "pcicfg", &lock_class_mtx_spin },
318 #endif
319 	{ NULL, NULL },
320 	{ NULL, NULL }
321 };
322 
323 #ifdef BLESSING
324 /*
325  * Pairs of locks which have been blessed
326  * Don't complain about order problems with blessed locks
327  */
328 static struct witness_blessed blessed_list[] = {
329 };
330 static int blessed_count =
331 	sizeof(blessed_list) / sizeof(struct witness_blessed);
332 #endif
333 
334 /*
335  * List of all locks in the system.
336  */
337 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
338 
339 static struct mtx all_mtx = {
340 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
341 	  "All locks list",		/* mtx_object.lo_name */
342 	  "All locks list",		/* mtx_object.lo_type */
343 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
344 	  { NULL, NULL },		/* mtx_object.lo_list */
345 	  NULL },			/* mtx_object.lo_witness */
346 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
347 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
348 	{ NULL, NULL }			/* mtx_contested */
349 };
350 
351 /*
352  * This global is set to 0 once it becomes safe to use the witness code.
353  */
354 static int witness_cold = 1;
355 
356 /*
357  * Global variables for book keeping.
358  */
359 static int lock_cur_cnt;
360 static int lock_max_cnt;
361 
362 /*
363  * The WITNESS-enabled diagnostic code.
364  */
365 static void
366 witness_initialize(void *dummy __unused)
367 {
368 	struct lock_object *lock;
369 	struct witness_order_list_entry *order;
370 	struct witness *w, *w1;
371 	int i;
372 
373 	/*
374 	 * We have to release Giant before initializing its witness
375 	 * structure so that WITNESS doesn't get confused.
376 	 */
377 	mtx_unlock(&Giant);
378 	mtx_assert(&Giant, MA_NOTOWNED);
379 
380 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
381 	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
382 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
383 	    MTX_NOWITNESS);
384 	for (i = 0; i < WITNESS_COUNT; i++)
385 		witness_free(&w_data[i]);
386 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
387 		witness_child_free(&w_childdata[i]);
388 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
389 		witness_lock_list_free(&w_locklistdata[i]);
390 
391 	/* First add in all the specified order lists. */
392 	for (order = order_lists; order->w_name != NULL; order++) {
393 		w = enroll(order->w_name, order->w_class);
394 		if (w == NULL)
395 			continue;
396 		w->w_file = "order list";
397 		for (order++; order->w_name != NULL; order++) {
398 			w1 = enroll(order->w_name, order->w_class);
399 			if (w1 == NULL)
400 				continue;
401 			w1->w_file = "order list";
402 			if (!itismychild(w, w1))
403 				panic("Not enough memory for static orders!");
404 			w = w1;
405 		}
406 	}
407 
408 	/* Iterate through all locks and add them to witness. */
409 	mtx_lock(&all_mtx);
410 	TAILQ_FOREACH(lock, &all_locks, lo_list) {
411 		if (lock->lo_flags & LO_WITNESS)
412 			lock->lo_witness = enroll(lock->lo_type,
413 			    lock->lo_class);
414 		else
415 			lock->lo_witness = NULL;
416 	}
417 	mtx_unlock(&all_mtx);
418 
419 	/* Mark the witness code as being ready for use. */
420 	atomic_store_rel_int(&witness_cold, 0);
421 
422 	mtx_lock(&Giant);
423 }
424 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
425 
426 static int
427 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
428 {
429 	int error, value;
430 
431 	value = witness_watch;
432 	error = sysctl_handle_int(oidp, &value, 0, req);
433 	if (error != 0 || req->newptr == NULL)
434 		return (error);
435 	error = suser(req->td);
436 	if (error != 0)
437 		return (error);
438 	if (value == witness_watch)
439 		return (0);
440 	if (value != 0)
441 		return (EINVAL);
442 	witness_watch = 0;
443 	return (0);
444 }
445 
446 void
447 witness_init(struct lock_object *lock)
448 {
449 	struct lock_class *class;
450 
451 	class = lock->lo_class;
452 	if (lock->lo_flags & LO_INITIALIZED)
453 		panic("%s: lock (%s) %s is already initialized", __func__,
454 		    class->lc_name, lock->lo_name);
455 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
456 	    (class->lc_flags & LC_RECURSABLE) == 0)
457 		panic("%s: lock (%s) %s can not be recursable", __func__,
458 		    class->lc_name, lock->lo_name);
459 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
460 	    (class->lc_flags & LC_SLEEPABLE) == 0)
461 		panic("%s: lock (%s) %s can not be sleepable", __func__,
462 		    class->lc_name, lock->lo_name);
463 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
464 	    (class->lc_flags & LC_UPGRADABLE) == 0)
465 		panic("%s: lock (%s) %s can not be upgradable", __func__,
466 		    class->lc_name, lock->lo_name);
467 
468 	mtx_lock(&all_mtx);
469 	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
470 	lock->lo_flags |= LO_INITIALIZED;
471 	lock_cur_cnt++;
472 	if (lock_cur_cnt > lock_max_cnt)
473 		lock_max_cnt = lock_cur_cnt;
474 	mtx_unlock(&all_mtx);
475 	if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
476 	    (lock->lo_flags & LO_WITNESS) != 0)
477 		lock->lo_witness = enroll(lock->lo_type, class);
478 	else
479 		lock->lo_witness = NULL;
480 }
481 
482 void
483 witness_destroy(struct lock_object *lock)
484 {
485 	struct witness *w;
486 
487 	if (witness_cold)
488 		panic("lock (%s) %s destroyed while witness_cold",
489 		    lock->lo_class->lc_name, lock->lo_name);
490 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
491 		panic("%s: lock (%s) %s is not initialized", __func__,
492 		    lock->lo_class->lc_name, lock->lo_name);
493 
494 	/* XXX: need to verify that no one holds the lock */
495 	w = lock->lo_witness;
496 	if (w != NULL) {
497 		mtx_lock_spin(&w_mtx);
498 		MPASS(w->w_refcount > 0);
499 		w->w_refcount--;
500 
501 		/*
502 		 * Lock is already released if we have an allocation failure
503 		 * and depart() fails.
504 		 */
505 		if (w->w_refcount != 0 || depart(w))
506 			mtx_unlock_spin(&w_mtx);
507 	}
508 
509 	mtx_lock(&all_mtx);
510 	lock_cur_cnt--;
511 	TAILQ_REMOVE(&all_locks, lock, lo_list);
512 	lock->lo_flags &= ~LO_INITIALIZED;
513 	mtx_unlock(&all_mtx);
514 }
515 
516 #ifdef DDB
517 static void
518 witness_display_list(void(*prnt)(const char *fmt, ...),
519 		     struct witness_list *list)
520 {
521 	struct witness *w;
522 
523 	STAILQ_FOREACH(w, list, w_typelist) {
524 		if (w->w_file == NULL || w->w_level > 0)
525 			continue;
526 		/*
527 		 * This lock has no anscestors, display its descendants.
528 		 */
529 		witness_displaydescendants(prnt, w, 0);
530 	}
531 }
532 
533 static void
534 witness_display(void(*prnt)(const char *fmt, ...))
535 {
536 	struct witness *w;
537 
538 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
539 	witness_levelall();
540 
541 	/* Clear all the displayed flags. */
542 	STAILQ_FOREACH(w, &w_all, w_list) {
543 		w->w_displayed = 0;
544 	}
545 
546 	/*
547 	 * First, handle sleep locks which have been acquired at least
548 	 * once.
549 	 */
550 	prnt("Sleep locks:\n");
551 	witness_display_list(prnt, &w_sleep);
552 
553 	/*
554 	 * Now do spin locks which have been acquired at least once.
555 	 */
556 	prnt("\nSpin locks:\n");
557 	witness_display_list(prnt, &w_spin);
558 
559 	/*
560 	 * Finally, any locks which have not been acquired yet.
561 	 */
562 	prnt("\nLocks which were never acquired:\n");
563 	STAILQ_FOREACH(w, &w_all, w_list) {
564 		if (w->w_file != NULL || w->w_refcount == 0)
565 			continue;
566 		prnt("%s\n", w->w_name);
567 	}
568 }
569 #endif /* DDB */
570 
571 /* Trim useless garbage from filenames. */
572 static const char *
573 fixup_filename(const char *file)
574 {
575 
576 	if (file == NULL)
577 		return (NULL);
578 	while (strncmp(file, "../", 3) == 0)
579 		file += 3;
580 	return (file);
581 }
582 
583 void
584 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
585 {
586 	struct lock_list_entry **lock_list, *lle;
587 	struct lock_instance *lock1, *lock2;
588 	struct lock_class *class;
589 	struct witness *w, *w1;
590 	struct thread *td;
591 	int i, j;
592 #ifdef DDB
593 	int go_into_ddb = 0;
594 #endif
595 
596 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
597 	    panicstr != NULL)
598 		return;
599 	w = lock->lo_witness;
600 	class = lock->lo_class;
601 	td = curthread;
602 	file = fixup_filename(file);
603 
604 	if (class->lc_flags & LC_SLEEPLOCK) {
605 		/*
606 		 * Since spin locks include a critical section, this check
607 		 * impliclty enforces a lock order of all sleep locks before
608 		 * all spin locks.
609 		 */
610 		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
611 			panic("blockable sleep lock (%s) %s @ %s:%d",
612 			    class->lc_name, lock->lo_name, file, line);
613 		lock_list = &td->td_sleeplocks;
614 	} else
615 		lock_list = PCPU_PTR(spinlocks);
616 
617 	/*
618 	 * Is this the first lock acquired?  If so, then no order checking
619 	 * is needed.
620 	 */
621 	if (*lock_list == NULL)
622 		goto out;
623 
624 	/*
625 	 * Check to see if we are recursing on a lock we already own.
626 	 */
627 	lock1 = find_instance(*lock_list, lock);
628 	if (lock1 != NULL) {
629 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
630 		    (flags & LOP_EXCLUSIVE) == 0) {
631 			printf("shared lock of (%s) %s @ %s:%d\n",
632 			    class->lc_name, lock->lo_name, file, line);
633 			printf("while exclusively locked from %s:%d\n",
634 			    lock1->li_file, lock1->li_line);
635 			panic("share->excl");
636 		}
637 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
638 		    (flags & LOP_EXCLUSIVE) != 0) {
639 			printf("exclusive lock of (%s) %s @ %s:%d\n",
640 			    class->lc_name, lock->lo_name, file, line);
641 			printf("while share locked from %s:%d\n",
642 			    lock1->li_file, lock1->li_line);
643 			panic("excl->share");
644 		}
645 		lock1->li_flags++;
646 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
647 			printf(
648 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
649 			    class->lc_name, lock->lo_name, file, line);
650 			printf("first acquired @ %s:%d\n", lock1->li_file,
651 			    lock1->li_line);
652 			panic("recurse");
653 		}
654 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
655 		    td->td_proc->p_pid, lock->lo_name,
656 		    lock1->li_flags & LI_RECURSEMASK);
657 		lock1->li_file = file;
658 		lock1->li_line = line;
659 		return;
660 	}
661 
662 	/*
663 	 * Try locks do not block if they fail to acquire the lock, thus
664 	 * there is no danger of deadlocks or of switching while holding a
665 	 * spin lock if we acquire a lock via a try operation.
666 	 */
667 	if (flags & LOP_TRYLOCK)
668 		goto out;
669 
670 	/*
671 	 * Check for duplicate locks of the same type.  Note that we only
672 	 * have to check for this on the last lock we just acquired.  Any
673 	 * other cases will be caught as lock order violations.
674 	 */
675 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
676 	w1 = lock1->li_lock->lo_witness;
677 	if (w1 == w) {
678 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
679 			goto out;
680 		w->w_same_squawked = 1;
681 		printf("acquiring duplicate lock of same type: \"%s\"\n",
682 			lock->lo_type);
683 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
684 		    lock1->li_file, lock1->li_line);
685 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
686 #ifdef DDB
687 		go_into_ddb = 1;
688 #endif
689 		goto out;
690 	}
691 	MPASS(!mtx_owned(&w_mtx));
692 	mtx_lock_spin(&w_mtx);
693 	/*
694 	 * If we have a known higher number just say ok
695 	 */
696 	if (witness_watch > 1 && w->w_level > w1->w_level) {
697 		mtx_unlock_spin(&w_mtx);
698 		goto out;
699 	}
700 	/*
701 	 * If we know that the the lock we are acquiring comes after
702 	 * the lock we most recently acquired in the lock order tree,
703 	 * then there is no need for any further checks.
704 	 */
705 	if (isitmydescendant(w1, w)) {
706 		mtx_unlock_spin(&w_mtx);
707 		goto out;
708 	}
709 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
710 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
711 
712 			MPASS(j < WITNESS_COUNT);
713 			lock1 = &lle->ll_children[i];
714 			w1 = lock1->li_lock->lo_witness;
715 
716 			/*
717 			 * If this lock doesn't undergo witness checking,
718 			 * then skip it.
719 			 */
720 			if (w1 == NULL) {
721 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
722 				    ("lock missing witness structure"));
723 				continue;
724 			}
725 			/*
726 			 * If we are locking Giant and this is a sleepable
727 			 * lock, then skip it.
728 			 */
729 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
730 			    lock == &Giant.mtx_object)
731 				continue;
732 			/*
733 			 * If we are locking a sleepable lock and this lock
734 			 * is Giant, then skip it.
735 			 */
736 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
737 			    lock1->li_lock == &Giant.mtx_object)
738 				continue;
739 			/*
740 			 * If we are locking a sleepable lock and this lock
741 			 * isn't sleepable, we want to treat it as a lock
742 			 * order violation to enfore a general lock order of
743 			 * sleepable locks before non-sleepable locks.
744 			 */
745 			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
746 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
747 			    /*
748 			     * Check the lock order hierarchy for a reveresal.
749 			     */
750 			    if (!isitmydescendant(w, w1))
751 				continue;
752 			/*
753 			 * We have a lock order violation, check to see if it
754 			 * is allowed or has already been yelled about.
755 			 */
756 			mtx_unlock_spin(&w_mtx);
757 #ifdef BLESSING
758 			if (blessed(w, w1))
759 				goto out;
760 #endif
761 			if (lock1->li_lock == &Giant.mtx_object) {
762 				if (w1->w_Giant_squawked)
763 					goto out;
764 				else
765 					w1->w_Giant_squawked = 1;
766 			} else {
767 				if (w1->w_other_squawked)
768 					goto out;
769 				else
770 					w1->w_other_squawked = 1;
771 			}
772 			/*
773 			 * Ok, yell about it.
774 			 */
775 			printf("lock order reversal\n");
776 			/*
777 			 * Try to locate an earlier lock with
778 			 * witness w in our list.
779 			 */
780 			do {
781 				lock2 = &lle->ll_children[i];
782 				MPASS(lock2->li_lock != NULL);
783 				if (lock2->li_lock->lo_witness == w)
784 					break;
785 				i--;
786 				if (i == 0 && lle->ll_next != NULL) {
787 					lle = lle->ll_next;
788 					i = lle->ll_count - 1;
789 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
790 				}
791 			} while (i >= 0);
792 			if (i < 0) {
793 				printf(" 1st %p %s (%s) @ %s:%d\n",
794 				    lock1->li_lock, lock1->li_lock->lo_name,
795 				    lock1->li_lock->lo_type, lock1->li_file,
796 				    lock1->li_line);
797 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
798 				    lock->lo_name, lock->lo_type, file, line);
799 			} else {
800 				printf(" 1st %p %s (%s) @ %s:%d\n",
801 				    lock2->li_lock, lock2->li_lock->lo_name,
802 				    lock2->li_lock->lo_type, lock2->li_file,
803 				    lock2->li_line);
804 				printf(" 2nd %p %s (%s) @ %s:%d\n",
805 				    lock1->li_lock, lock1->li_lock->lo_name,
806 				    lock1->li_lock->lo_type, lock1->li_file,
807 				    lock1->li_line);
808 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
809 				    lock->lo_name, lock->lo_type, file, line);
810 			}
811 #ifdef DDB
812 			go_into_ddb = 1;
813 #endif
814 			goto out;
815 		}
816 	}
817 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
818 	/*
819 	 * Don't build a new relationship between a sleepable lock and
820 	 * Giant if it is the wrong direction.  The real lock order is that
821 	 * sleepable locks come before Giant.
822 	 */
823 	if (!(lock1->li_lock == &Giant.mtx_object &&
824 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
825 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
826 		    lock->lo_type, lock1->li_lock->lo_type);
827 		if (!itismychild(lock1->li_lock->lo_witness, w))
828 			/* Witness is dead. */
829 			return;
830 	}
831 	mtx_unlock_spin(&w_mtx);
832 
833 out:
834 #ifdef DDB
835 	if (go_into_ddb) {
836 		if (witness_trace)
837 			backtrace();
838 		if (witness_ddb)
839 			Debugger(__func__);
840 	}
841 #endif
842 	w->w_file = file;
843 	w->w_line = line;
844 
845 	lle = *lock_list;
846 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
847 		lle = witness_lock_list_get();
848 		if (lle == NULL)
849 			return;
850 		lle->ll_next = *lock_list;
851 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
852 		    td->td_proc->p_pid, lle);
853 		*lock_list = lle;
854 	}
855 	lock1 = &lle->ll_children[lle->ll_count++];
856 	lock1->li_lock = lock;
857 	lock1->li_line = line;
858 	lock1->li_file = file;
859 	if ((flags & LOP_EXCLUSIVE) != 0)
860 		lock1->li_flags = LI_EXCLUSIVE;
861 	else
862 		lock1->li_flags = 0;
863 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
864 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
865 }
866 
867 void
868 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
869 {
870 	struct lock_instance *instance;
871 	struct lock_class *class;
872 
873 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
874 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
875 		return;
876 	class = lock->lo_class;
877 	file = fixup_filename(file);
878 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
879 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
880 		    class->lc_name, lock->lo_name, file, line);
881 	if ((flags & LOP_TRYLOCK) == 0)
882 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
883 		    lock->lo_name, file, line);
884 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
885 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
886 		    class->lc_name, lock->lo_name, file, line);
887 	instance = find_instance(curthread->td_sleeplocks, lock);
888 	if (instance == NULL)
889 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
890 		    class->lc_name, lock->lo_name, file, line);
891 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
892 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
893 		    class->lc_name, lock->lo_name, file, line);
894 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
895 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
896 		    class->lc_name, lock->lo_name,
897 		    instance->li_flags & LI_RECURSEMASK, file, line);
898 	instance->li_flags |= LI_EXCLUSIVE;
899 }
900 
901 void
902 witness_downgrade(struct lock_object *lock, int flags, const char *file,
903     int line)
904 {
905 	struct lock_instance *instance;
906 	struct lock_class *class;
907 
908 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
909 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
910 		return;
911 	class = lock->lo_class;
912 	file = fixup_filename(file);
913 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
914 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
915 		    class->lc_name, lock->lo_name, file, line);
916 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
917 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
918 		    class->lc_name, lock->lo_name, file, line);
919 	instance = find_instance(curthread->td_sleeplocks, lock);
920 	if (instance == NULL)
921 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
922 		    class->lc_name, lock->lo_name, file, line);
923 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
924 		panic("downgrade of shared lock (%s) %s @ %s:%d",
925 		    class->lc_name, lock->lo_name, file, line);
926 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
927 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
928 		    class->lc_name, lock->lo_name,
929 		    instance->li_flags & LI_RECURSEMASK, file, line);
930 	instance->li_flags &= ~LI_EXCLUSIVE;
931 }
932 
933 void
934 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
935 {
936 	struct lock_list_entry **lock_list, *lle;
937 	struct lock_instance *instance;
938 	struct lock_class *class;
939 	struct thread *td;
940 	register_t s;
941 	int i, j;
942 
943 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
944 	    panicstr != NULL)
945 		return;
946 	td = curthread;
947 	class = lock->lo_class;
948 	file = fixup_filename(file);
949 	if (class->lc_flags & LC_SLEEPLOCK)
950 		lock_list = &td->td_sleeplocks;
951 	else
952 		lock_list = PCPU_PTR(spinlocks);
953 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
954 		for (i = 0; i < (*lock_list)->ll_count; i++) {
955 			instance = &(*lock_list)->ll_children[i];
956 			if (instance->li_lock == lock) {
957 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
958 				    (flags & LOP_EXCLUSIVE) == 0) {
959 					printf(
960 					"shared unlock of (%s) %s @ %s:%d\n",
961 					    class->lc_name, lock->lo_name,
962 					    file, line);
963 					printf(
964 					"while exclusively locked from %s:%d\n",
965 					    instance->li_file,
966 					    instance->li_line);
967 					panic("excl->ushare");
968 				}
969 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
970 				    (flags & LOP_EXCLUSIVE) != 0) {
971 					printf(
972 					"exclusive unlock of (%s) %s @ %s:%d\n",
973 					    class->lc_name, lock->lo_name,
974 					    file, line);
975 					printf(
976 					"while share locked from %s:%d\n",
977 					    instance->li_file,
978 					    instance->li_line);
979 					panic("share->uexcl");
980 				}
981 				/* If we are recursed, unrecurse. */
982 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
983 					CTR4(KTR_WITNESS,
984 				    "%s: pid %d unrecursed on %s r=%d", __func__,
985 					    td->td_proc->p_pid,
986 					    instance->li_lock->lo_name,
987 					    instance->li_flags);
988 					instance->li_flags--;
989 					return;
990 				}
991 				s = intr_disable();
992 				CTR4(KTR_WITNESS,
993 				    "%s: pid %d removed %s from lle[%d]", __func__,
994 				    td->td_proc->p_pid,
995 				    instance->li_lock->lo_name,
996 				    (*lock_list)->ll_count - 1);
997 				for (j = i; j < (*lock_list)->ll_count - 1; j++)
998 					(*lock_list)->ll_children[j] =
999 					    (*lock_list)->ll_children[j + 1];
1000 				(*lock_list)->ll_count--;
1001 				intr_restore(s);
1002 				if ((*lock_list)->ll_count == 0) {
1003 					lle = *lock_list;
1004 					*lock_list = lle->ll_next;
1005 					CTR3(KTR_WITNESS,
1006 					    "%s: pid %d removed lle %p", __func__,
1007 					    td->td_proc->p_pid, lle);
1008 					witness_lock_list_free(lle);
1009 				}
1010 				return;
1011 			}
1012 		}
1013 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1014 	    file, line);
1015 }
1016 
1017 /*
1018  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1019  * exempt Giant and sleepable locks from the checks as well.  If any
1020  * non-exempt locks are held, then a supplied message is printed to the
1021  * console along with a list of the offending locks.  If indicated in the
1022  * flags then a failure results in a panic as well.
1023  */
1024 int
1025 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1026 {
1027 	struct lock_list_entry *lle;
1028 	struct lock_instance *lock1;
1029 	struct thread *td;
1030 	va_list ap;
1031 	int i, n;
1032 
1033 	if (witness_cold || witness_watch == 0 || panicstr != NULL)
1034 		return (0);
1035 	n = 0;
1036 	td = curthread;
1037 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1038 		for (i = lle->ll_count - 1; i >= 0; i--) {
1039 			lock1 = &lle->ll_children[i];
1040 			if (lock1->li_lock == lock)
1041 				continue;
1042 			if (flags & WARN_GIANTOK &&
1043 			    lock1->li_lock == &Giant.mtx_object)
1044 				continue;
1045 			if (flags & WARN_SLEEPOK &&
1046 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1047 				continue;
1048 			if (n == 0) {
1049 				va_start(ap, fmt);
1050 				vprintf(fmt, ap);
1051 				va_end(ap);
1052 				printf(" with the following");
1053 				if (flags & WARN_SLEEPOK)
1054 					printf(" non-sleepable");
1055 				printf(" locks held:\n");
1056 			}
1057 			n++;
1058 			witness_list_lock(lock1);
1059 		}
1060 	if (PCPU_GET(spinlocks) != NULL) {
1061 		/*
1062 		 * Since we already hold a spinlock preemption is
1063 		 * already blocked.
1064 		 */
1065 		if (n == 0) {
1066 			va_start(ap, fmt);
1067 			vprintf(fmt, ap);
1068 			va_end(ap);
1069 			printf(" with the following");
1070 			if (flags & WARN_SLEEPOK)
1071 				printf(" non-sleepable");
1072 			printf(" locks held:\n");
1073 		}
1074 		n += witness_list_locks(PCPU_PTR(spinlocks));
1075 	}
1076 	if (flags & WARN_PANIC && n)
1077 		panic("witness_warn");
1078 #ifdef DDB
1079 	else if (witness_ddb && n)
1080 		Debugger(__func__);
1081 #endif
1082 	return (n);
1083 }
1084 
1085 const char *
1086 witness_file(struct lock_object *lock)
1087 {
1088 	struct witness *w;
1089 
1090 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1091 		return ("?");
1092 	w = lock->lo_witness;
1093 	return (w->w_file);
1094 }
1095 
1096 int
1097 witness_line(struct lock_object *lock)
1098 {
1099 	struct witness *w;
1100 
1101 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1102 		return (0);
1103 	w = lock->lo_witness;
1104 	return (w->w_line);
1105 }
1106 
1107 static struct witness *
1108 enroll(const char *description, struct lock_class *lock_class)
1109 {
1110 	struct witness *w;
1111 
1112 	if (!witness_watch || witness_watch == 0 || panicstr != NULL)
1113 		return (NULL);
1114 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1115 		return (NULL);
1116 	mtx_lock_spin(&w_mtx);
1117 	STAILQ_FOREACH(w, &w_all, w_list) {
1118 		if (w->w_name == description || (w->w_refcount > 0 &&
1119 		    strcmp(description, w->w_name) == 0)) {
1120 			w->w_refcount++;
1121 			mtx_unlock_spin(&w_mtx);
1122 			if (lock_class != w->w_class)
1123 				panic(
1124 				"lock (%s) %s does not match earlier (%s) lock",
1125 				    description, lock_class->lc_name,
1126 				    w->w_class->lc_name);
1127 			return (w);
1128 		}
1129 	}
1130 	/*
1131 	 * This isn't quite right, as witness_cold is still 0 while we
1132 	 * enroll all the locks initialized before witness_initialize().
1133 	 */
1134 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1135 		mtx_unlock_spin(&w_mtx);
1136 		panic("spin lock %s not in order list", description);
1137 	}
1138 	if ((w = witness_get()) == NULL)
1139 		return (NULL);
1140 	w->w_name = description;
1141 	w->w_class = lock_class;
1142 	w->w_refcount = 1;
1143 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1144 	if (lock_class->lc_flags & LC_SPINLOCK)
1145 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1146 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
1147 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1148 	else {
1149 		mtx_unlock_spin(&w_mtx);
1150 		panic("lock class %s is not sleep or spin",
1151 		    lock_class->lc_name);
1152 	}
1153 	mtx_unlock_spin(&w_mtx);
1154 	return (w);
1155 }
1156 
1157 /* Don't let the door bang you on the way out... */
1158 static int
1159 depart(struct witness *w)
1160 {
1161 	struct witness_child_list_entry *wcl, *nwcl;
1162 	struct witness_list *list;
1163 	struct witness *parent;
1164 
1165 	MPASS(w->w_refcount == 0);
1166 	if (w->w_class->lc_flags & LC_SLEEPLOCK)
1167 		list = &w_sleep;
1168 	else
1169 		list = &w_spin;
1170 	/*
1171 	 * First, we run through the entire tree looking for any
1172 	 * witnesses that the outgoing witness is a child of.  For
1173 	 * each parent that we find, we reparent all the direct
1174 	 * children of the outgoing witness to its parent.
1175 	 */
1176 	STAILQ_FOREACH(parent, list, w_typelist) {
1177 		if (!isitmychild(parent, w))
1178 			continue;
1179 		removechild(parent, w);
1180 		if (!reparentchildren(parent, w))
1181 			return (0);
1182 	}
1183 
1184 	/*
1185 	 * Now we go through and free up the child list of the
1186 	 * outgoing witness.
1187 	 */
1188 	for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1189 		nwcl = wcl->wcl_next;
1190 		witness_child_free(wcl);
1191 	}
1192 
1193 	/*
1194 	 * Detach from various lists and free.
1195 	 */
1196 	STAILQ_REMOVE(list, w, witness, w_typelist);
1197 	STAILQ_REMOVE(&w_all, w, witness, w_list);
1198 	witness_free(w);
1199 
1200 	/* Finally, fixup the tree. */
1201 	return (rebalancetree(list));
1202 }
1203 
1204 /*
1205  * Prune an entire lock order tree.  We look for cases where a lock
1206  * is now both a descendant and a direct child of a given lock.  In
1207  * that case, we want to remove the direct child link from the tree.
1208  *
1209  * Returns false if insertchild() fails.
1210  */
1211 static int
1212 rebalancetree(struct witness_list *list)
1213 {
1214 	struct witness *child, *parent;
1215 
1216 	STAILQ_FOREACH(child, list, w_typelist) {
1217 		STAILQ_FOREACH(parent, list, w_typelist) {
1218 			if (!isitmychild(parent, child))
1219 				continue;
1220 			removechild(parent, child);
1221 			if (isitmydescendant(parent, child))
1222 				continue;
1223 			if (!insertchild(parent, child))
1224 				return (0);
1225 		}
1226 	}
1227 	witness_levelall();
1228 	return (1);
1229 }
1230 
1231 /*
1232  * Add "child" as a direct child of "parent".  Returns false if
1233  * we fail due to out of memory.
1234  */
1235 static int
1236 insertchild(struct witness *parent, struct witness *child)
1237 {
1238 	struct witness_child_list_entry **wcl;
1239 
1240 	MPASS(child != NULL && parent != NULL);
1241 
1242 	/*
1243 	 * Insert "child" after "parent"
1244 	 */
1245 	wcl = &parent->w_children;
1246 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1247 		wcl = &(*wcl)->wcl_next;
1248 	if (*wcl == NULL) {
1249 		*wcl = witness_child_get();
1250 		if (*wcl == NULL)
1251 			return (0);
1252 	}
1253 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1254 
1255 	return (1);
1256 }
1257 
1258 /*
1259  * Make all the direct descendants of oldparent be direct descendants
1260  * of newparent.
1261  */
1262 static int
1263 reparentchildren(struct witness *newparent, struct witness *oldparent)
1264 {
1265 	struct witness_child_list_entry *wcl;
1266 	int i;
1267 
1268 	/* Avoid making a witness a child of itself. */
1269 	MPASS(!isitmychild(oldparent, newparent));
1270 
1271 	for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1272 		for (i = 0; i < wcl->wcl_count; i++)
1273 			if (!insertchild(newparent, wcl->wcl_children[i]))
1274 				return (0);
1275 	return (1);
1276 }
1277 
1278 static int
1279 itismychild(struct witness *parent, struct witness *child)
1280 {
1281 	struct witness_list *list;
1282 
1283 	MPASS(child != NULL && parent != NULL);
1284 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1285 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1286 		panic(
1287 		"%s: parent (%s) and child (%s) are not the same lock type",
1288 		    __func__, parent->w_class->lc_name,
1289 		    child->w_class->lc_name);
1290 
1291 	if (!insertchild(parent, child))
1292 		return (0);
1293 
1294 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1295 		list = &w_sleep;
1296 	else
1297 		list = &w_spin;
1298 	return (rebalancetree(list));
1299 }
1300 
1301 static void
1302 removechild(struct witness *parent, struct witness *child)
1303 {
1304 	struct witness_child_list_entry **wcl, *wcl1;
1305 	int i;
1306 
1307 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1308 		for (i = 0; i < (*wcl)->wcl_count; i++)
1309 			if ((*wcl)->wcl_children[i] == child)
1310 				goto found;
1311 	return;
1312 found:
1313 	(*wcl)->wcl_count--;
1314 	if ((*wcl)->wcl_count > i)
1315 		(*wcl)->wcl_children[i] =
1316 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1317 	MPASS((*wcl)->wcl_children[i] != NULL);
1318 	if ((*wcl)->wcl_count != 0)
1319 		return;
1320 	wcl1 = *wcl;
1321 	*wcl = wcl1->wcl_next;
1322 	witness_child_free(wcl1);
1323 }
1324 
1325 static int
1326 isitmychild(struct witness *parent, struct witness *child)
1327 {
1328 	struct witness_child_list_entry *wcl;
1329 	int i;
1330 
1331 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1332 		for (i = 0; i < wcl->wcl_count; i++) {
1333 			if (wcl->wcl_children[i] == child)
1334 				return (1);
1335 		}
1336 	}
1337 	return (0);
1338 }
1339 
1340 static int
1341 isitmydescendant(struct witness *parent, struct witness *child)
1342 {
1343 	struct witness_child_list_entry *wcl;
1344 	int i, j;
1345 
1346 	if (isitmychild(parent, child))
1347 		return (1);
1348 	j = 0;
1349 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1350 		MPASS(j < 1000);
1351 		for (i = 0; i < wcl->wcl_count; i++) {
1352 			if (isitmydescendant(wcl->wcl_children[i], child))
1353 				return (1);
1354 		}
1355 		j++;
1356 	}
1357 	return (0);
1358 }
1359 
1360 static void
1361 witness_levelall (void)
1362 {
1363 	struct witness_list *list;
1364 	struct witness *w, *w1;
1365 
1366 	/*
1367 	 * First clear all levels.
1368 	 */
1369 	STAILQ_FOREACH(w, &w_all, w_list) {
1370 		w->w_level = 0;
1371 	}
1372 
1373 	/*
1374 	 * Look for locks with no parent and level all their descendants.
1375 	 */
1376 	STAILQ_FOREACH(w, &w_all, w_list) {
1377 		/*
1378 		 * This is just an optimization, technically we could get
1379 		 * away just walking the all list each time.
1380 		 */
1381 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1382 			list = &w_sleep;
1383 		else
1384 			list = &w_spin;
1385 		STAILQ_FOREACH(w1, list, w_typelist) {
1386 			if (isitmychild(w1, w))
1387 				goto skip;
1388 		}
1389 		witness_leveldescendents(w, 0);
1390 	skip:
1391 		;	/* silence GCC 3.x */
1392 	}
1393 }
1394 
1395 static void
1396 witness_leveldescendents(struct witness *parent, int level)
1397 {
1398 	struct witness_child_list_entry *wcl;
1399 	int i;
1400 
1401 	if (parent->w_level < level)
1402 		parent->w_level = level;
1403 	level++;
1404 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1405 		for (i = 0; i < wcl->wcl_count; i++)
1406 			witness_leveldescendents(wcl->wcl_children[i], level);
1407 }
1408 
1409 static void
1410 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1411 			   struct witness *parent, int indent)
1412 {
1413 	struct witness_child_list_entry *wcl;
1414 	int i, level;
1415 
1416 	level = parent->w_level;
1417 	prnt("%-2d", level);
1418 	for (i = 0; i < indent; i++)
1419 		prnt(" ");
1420 	if (parent->w_refcount > 0)
1421 		prnt("%s", parent->w_name);
1422 	else
1423 		prnt("(dead)");
1424 	if (parent->w_displayed) {
1425 		prnt(" -- (already displayed)\n");
1426 		return;
1427 	}
1428 	parent->w_displayed = 1;
1429 	if (parent->w_refcount > 0) {
1430 		if (parent->w_file != NULL)
1431 			prnt(" -- last acquired @ %s:%d", parent->w_file,
1432 			    parent->w_line);
1433 	}
1434 	prnt("\n");
1435 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1436 		for (i = 0; i < wcl->wcl_count; i++)
1437 			    witness_displaydescendants(prnt,
1438 				wcl->wcl_children[i], indent + 1);
1439 }
1440 
1441 #ifdef BLESSING
1442 static int
1443 blessed(struct witness *w1, struct witness *w2)
1444 {
1445 	int i;
1446 	struct witness_blessed *b;
1447 
1448 	for (i = 0; i < blessed_count; i++) {
1449 		b = &blessed_list[i];
1450 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1451 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1452 				return (1);
1453 			continue;
1454 		}
1455 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1456 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1457 				return (1);
1458 	}
1459 	return (0);
1460 }
1461 #endif
1462 
1463 static struct witness *
1464 witness_get(void)
1465 {
1466 	struct witness *w;
1467 
1468 	if (witness_watch == 0) {
1469 		mtx_unlock_spin(&w_mtx);
1470 		return (NULL);
1471 	}
1472 	if (STAILQ_EMPTY(&w_free)) {
1473 		witness_watch = 0;
1474 		mtx_unlock_spin(&w_mtx);
1475 		printf("%s: witness exhausted\n", __func__);
1476 		return (NULL);
1477 	}
1478 	w = STAILQ_FIRST(&w_free);
1479 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1480 	bzero(w, sizeof(*w));
1481 	return (w);
1482 }
1483 
1484 static void
1485 witness_free(struct witness *w)
1486 {
1487 
1488 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1489 }
1490 
1491 static struct witness_child_list_entry *
1492 witness_child_get(void)
1493 {
1494 	struct witness_child_list_entry *wcl;
1495 
1496 	if (witness_watch == 0) {
1497 		mtx_unlock_spin(&w_mtx);
1498 		return (NULL);
1499 	}
1500 	wcl = w_child_free;
1501 	if (wcl == NULL) {
1502 		witness_watch = 0;
1503 		mtx_unlock_spin(&w_mtx);
1504 		printf("%s: witness exhausted\n", __func__);
1505 		return (NULL);
1506 	}
1507 	w_child_free = wcl->wcl_next;
1508 	bzero(wcl, sizeof(*wcl));
1509 	return (wcl);
1510 }
1511 
1512 static void
1513 witness_child_free(struct witness_child_list_entry *wcl)
1514 {
1515 
1516 	wcl->wcl_next = w_child_free;
1517 	w_child_free = wcl;
1518 }
1519 
1520 static struct lock_list_entry *
1521 witness_lock_list_get(void)
1522 {
1523 	struct lock_list_entry *lle;
1524 
1525 	if (witness_watch == 0)
1526 		return (NULL);
1527 	mtx_lock_spin(&w_mtx);
1528 	lle = w_lock_list_free;
1529 	if (lle == NULL) {
1530 		witness_watch = 0;
1531 		mtx_unlock_spin(&w_mtx);
1532 		printf("%s: witness exhausted\n", __func__);
1533 		return (NULL);
1534 	}
1535 	w_lock_list_free = lle->ll_next;
1536 	mtx_unlock_spin(&w_mtx);
1537 	bzero(lle, sizeof(*lle));
1538 	return (lle);
1539 }
1540 
1541 static void
1542 witness_lock_list_free(struct lock_list_entry *lle)
1543 {
1544 
1545 	mtx_lock_spin(&w_mtx);
1546 	lle->ll_next = w_lock_list_free;
1547 	w_lock_list_free = lle;
1548 	mtx_unlock_spin(&w_mtx);
1549 }
1550 
1551 static struct lock_instance *
1552 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1553 {
1554 	struct lock_list_entry *lle;
1555 	struct lock_instance *instance;
1556 	int i;
1557 
1558 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1559 		for (i = lle->ll_count - 1; i >= 0; i--) {
1560 			instance = &lle->ll_children[i];
1561 			if (instance->li_lock == lock)
1562 				return (instance);
1563 		}
1564 	return (NULL);
1565 }
1566 
1567 static void
1568 witness_list_lock(struct lock_instance *instance)
1569 {
1570 	struct lock_object *lock;
1571 
1572 	lock = instance->li_lock;
1573 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1574 	    "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1575 	if (lock->lo_type != lock->lo_name)
1576 		printf(" (%s)", lock->lo_type);
1577 	printf(" r = %d (%p) locked @ %s:%d\n",
1578 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1579 	    instance->li_line);
1580 }
1581 
1582 int
1583 witness_list_locks(struct lock_list_entry **lock_list)
1584 {
1585 	struct lock_list_entry *lle;
1586 	int i, nheld;
1587 
1588 	nheld = 0;
1589 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1590 		for (i = lle->ll_count - 1; i >= 0; i--) {
1591 			witness_list_lock(&lle->ll_children[i]);
1592 			nheld++;
1593 		}
1594 	return (nheld);
1595 }
1596 
1597 /*
1598  * This is a bit risky at best.  We call this function when we have timed
1599  * out acquiring a spin lock, and we assume that the other CPU is stuck
1600  * with this lock held.  So, we go groveling around in the other CPU's
1601  * per-cpu data to try to find the lock instance for this spin lock to
1602  * see when it was last acquired.
1603  */
1604 void
1605 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1606 {
1607 	struct lock_instance *instance;
1608 	struct pcpu *pc;
1609 
1610 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1611 		return;
1612 	pc = pcpu_find(owner->td_oncpu);
1613 	instance = find_instance(pc->pc_spinlocks, lock);
1614 	if (instance != NULL)
1615 		witness_list_lock(instance);
1616 }
1617 
1618 void
1619 witness_save(struct lock_object *lock, const char **filep, int *linep)
1620 {
1621 	struct lock_instance *instance;
1622 
1623 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1624 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1625 		return;
1626 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1627 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1628 		    lock->lo_class->lc_name, lock->lo_name);
1629 	instance = find_instance(curthread->td_sleeplocks, lock);
1630 	if (instance == NULL)
1631 		panic("%s: lock (%s) %s not locked", __func__,
1632 		    lock->lo_class->lc_name, lock->lo_name);
1633 	*filep = instance->li_file;
1634 	*linep = instance->li_line;
1635 }
1636 
1637 void
1638 witness_restore(struct lock_object *lock, const char *file, int line)
1639 {
1640 	struct lock_instance *instance;
1641 
1642 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1643 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1644 		return;
1645 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1646 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1647 		    lock->lo_class->lc_name, lock->lo_name);
1648 	instance = find_instance(curthread->td_sleeplocks, lock);
1649 	if (instance == NULL)
1650 		panic("%s: lock (%s) %s not locked", __func__,
1651 		    lock->lo_class->lc_name, lock->lo_name);
1652 	lock->lo_witness->w_file = file;
1653 	lock->lo_witness->w_line = line;
1654 	instance->li_file = file;
1655 	instance->li_line = line;
1656 }
1657 
1658 void
1659 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1660 {
1661 #ifdef INVARIANT_SUPPORT
1662 	struct lock_instance *instance;
1663 
1664 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1665 		return;
1666 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1667 		instance = find_instance(curthread->td_sleeplocks, lock);
1668 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1669 		instance = find_instance(PCPU_GET(spinlocks), lock);
1670 	else {
1671 		panic("Lock (%s) %s is not sleep or spin!",
1672 		    lock->lo_class->lc_name, lock->lo_name);
1673 	}
1674 	file = fixup_filename(file);
1675 	switch (flags) {
1676 	case LA_UNLOCKED:
1677 		if (instance != NULL)
1678 			panic("Lock (%s) %s locked @ %s:%d.",
1679 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1680 		break;
1681 	case LA_LOCKED:
1682 	case LA_LOCKED | LA_RECURSED:
1683 	case LA_LOCKED | LA_NOTRECURSED:
1684 	case LA_SLOCKED:
1685 	case LA_SLOCKED | LA_RECURSED:
1686 	case LA_SLOCKED | LA_NOTRECURSED:
1687 	case LA_XLOCKED:
1688 	case LA_XLOCKED | LA_RECURSED:
1689 	case LA_XLOCKED | LA_NOTRECURSED:
1690 		if (instance == NULL) {
1691 			panic("Lock (%s) %s not locked @ %s:%d.",
1692 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1693 			break;
1694 		}
1695 		if ((flags & LA_XLOCKED) != 0 &&
1696 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1697 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1698 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1699 		if ((flags & LA_SLOCKED) != 0 &&
1700 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1701 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1702 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1703 		if ((flags & LA_RECURSED) != 0 &&
1704 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1705 			panic("Lock (%s) %s not recursed @ %s:%d.",
1706 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1707 		if ((flags & LA_NOTRECURSED) != 0 &&
1708 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1709 			panic("Lock (%s) %s recursed @ %s:%d.",
1710 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1711 		break;
1712 	default:
1713 		panic("Invalid lock assertion at %s:%d.", file, line);
1714 
1715 	}
1716 #endif	/* INVARIANT_SUPPORT */
1717 }
1718 
1719 #ifdef DDB
1720 static void
1721 witness_list(struct thread *td)
1722 {
1723 
1724 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1725 	KASSERT(db_active, ("%s: not in the debugger", __func__));
1726 
1727 	if (witness_watch == 0)
1728 		return;
1729 
1730 	witness_list_locks(&td->td_sleeplocks);
1731 
1732 	/*
1733 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1734 	 * if td is currently executing on some other CPU and holds spin locks
1735 	 * as we won't display those locks.  If we had a MI way of getting
1736 	 * the per-cpu data for a given cpu then we could use
1737 	 * td->td_oncpu to get the list of spinlocks for this thread
1738 	 * and "fix" this.
1739 	 *
1740 	 * That still wouldn't really fix this unless we locked sched_lock
1741 	 * or stopped the other CPU to make sure it wasn't changing the list
1742 	 * out from under us.  It is probably best to just not try to handle
1743 	 * threads on other CPU's for now.
1744 	 */
1745 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1746 		witness_list_locks(PCPU_PTR(spinlocks));
1747 }
1748 
1749 DB_SHOW_COMMAND(locks, db_witness_list)
1750 {
1751 	struct thread *td;
1752 	pid_t pid;
1753 	struct proc *p;
1754 
1755 	if (have_addr) {
1756 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1757 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1758 		    ((addr >> 16) % 16) * 10000;
1759 		/* sx_slock(&allproc_lock); */
1760 		FOREACH_PROC_IN_SYSTEM(p) {
1761 			if (p->p_pid == pid)
1762 				break;
1763 		}
1764 		/* sx_sunlock(&allproc_lock); */
1765 		if (p == NULL) {
1766 			db_printf("pid %d not found\n", pid);
1767 			return;
1768 		}
1769 		FOREACH_THREAD_IN_PROC(p, td) {
1770 			witness_list(td);
1771 		}
1772 	} else {
1773 		td = curthread;
1774 		witness_list(td);
1775 	}
1776 }
1777 
1778 DB_SHOW_COMMAND(witness, db_witness_display)
1779 {
1780 
1781 	witness_display(db_printf);
1782 }
1783 #endif
1784