subr_witness.c revision 122849
1153570Sjhb/*-
2153570Sjhb * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3153570Sjhb *
4153570Sjhb * Redistribution and use in source and binary forms, with or without
5153570Sjhb * modification, are permitted provided that the following conditions
6153570Sjhb * are met:
7153570Sjhb * 1. Redistributions of source code must retain the above copyright
8153570Sjhb *    notice, this list of conditions and the following disclaimer.
9153570Sjhb * 2. Redistributions in binary form must reproduce the above copyright
10153570Sjhb *    notice, this list of conditions and the following disclaimer in the
11153570Sjhb *    documentation and/or other materials provided with the distribution.
12153570Sjhb * 3. Berkeley Software Design Inc's name may not be used to endorse or
13153570Sjhb *    promote products derived from this software without specific prior
14153570Sjhb *    written permission.
15153570Sjhb *
16153570Sjhb * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17153570Sjhb * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18153570Sjhb * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19153570Sjhb * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20153570Sjhb * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21153570Sjhb * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22153570Sjhb * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23153570Sjhb * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24153570Sjhb * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25153570Sjhb * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26153570Sjhb * SUCH DAMAGE.
27153570Sjhb *
28153570Sjhb *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29153570Sjhb *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30153570Sjhb */
31153570Sjhb
32153570Sjhb/*
33153570Sjhb * Implementation of the `witness' lock verifier.  Originally implemented for
34153570Sjhb * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
35153570Sjhb * classes in FreeBSD.
36153570Sjhb */
37153570Sjhb
38153570Sjhb/*
39153570Sjhb *	Main Entry: witness
40153570Sjhb *	Pronunciation: 'wit-n&s
41153570Sjhb *	Function: noun
42153570Sjhb *	Etymology: Middle English witnesse, from Old English witnes knowledge,
43153570Sjhb *	    testimony, witness, from 2wit
44153570Sjhb *	Date: before 12th century
45153570Sjhb *	1 : attestation of a fact or event : TESTIMONY
46153570Sjhb *	2 : one that gives evidence; specifically : one who testifies in
47153570Sjhb *	    a cause or before a judicial tribunal
48153570Sjhb *	3 : one asked to be present at a transaction so as to be able to
49153570Sjhb *	    testify to its having taken place
50153570Sjhb *	4 : one who has personal knowledge of something
51153570Sjhb *	5 a : something serving as evidence or proof : SIGN
52153570Sjhb *	  b : public affirmation by word or example of usually
53153570Sjhb *	      religious faith or conviction <the heroic witness to divine
54153570Sjhb *	      life -- Pilot>
55153570Sjhb *	6 capitalized : a member of the Jehovah's Witnesses
56153570Sjhb */
57153570Sjhb
58153570Sjhb/*
59153570Sjhb * Special rules concerning Giant and lock orders:
60153570Sjhb *
61153570Sjhb * 1) Giant must be acquired before any other mutexes.  Stated another way,
62153570Sjhb *    no other mutex may be held when Giant is acquired.
63153570Sjhb *
64153570Sjhb * 2) Giant must be released when blocking on a sleepable lock.
65153570Sjhb *
66153570Sjhb * This rule is less obvious, but is a result of Giant providing the same
67153570Sjhb * semantics as spl().  Basically, when a thread sleeps, it must release
68153570Sjhb * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
69153570Sjhb * 2).
70153570Sjhb *
71153570Sjhb * 3) Giant may be acquired before or after sleepable locks.
72153570Sjhb *
73153570Sjhb * This rule is also not quite as obvious.  Giant may be acquired after
74153570Sjhb * a sleepable lock because it is a non-sleepable lock and non-sleepable
75153570Sjhb * locks may always be acquired while holding a sleepable lock.  The second
76153570Sjhb * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
77153570Sjhb * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
78153570Sjhb * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
79153570Sjhb * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
80153570Sjhb * execute.  Thus, acquiring Giant both before and after a sleepable lock
81219902Sjhb * will not result in a lock order reversal.
82153570Sjhb */
83153570Sjhb
84153570Sjhb#include <sys/cdefs.h>
85153570Sjhb__FBSDID("$FreeBSD: head/sys/kern/subr_witness.c 122849 2003-11-17 08:58:16Z peter $");
86153570Sjhb
87153570Sjhb#include "opt_ddb.h"
88153570Sjhb#include "opt_witness.h"
89153570Sjhb
90153570Sjhb#include <sys/param.h>
91153570Sjhb#include <sys/bus.h>
92153570Sjhb#include <sys/kernel.h>
93153570Sjhb#include <sys/ktr.h>
94153570Sjhb#include <sys/lock.h>
95153570Sjhb#include <sys/malloc.h>
96153570Sjhb#include <sys/mutex.h>
97153570Sjhb#include <sys/proc.h>
98153570Sjhb#include <sys/sysctl.h>
99153570Sjhb#include <sys/systm.h>
100153570Sjhb
101153570Sjhb#include <ddb/ddb.h>
102153570Sjhb
103153570Sjhb#include <machine/stdarg.h>
104153570Sjhb
105153570Sjhb/* Define this to check for blessed mutexes */
106153570Sjhb#undef BLESSING
107153570Sjhb
108153570Sjhb#define WITNESS_COUNT 200
109153570Sjhb#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
110153570Sjhb/*
111153570Sjhb * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
112153570Sjhb * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
113153570Sjhb * probably be safe for the most part, but it's still a SWAG.
114153570Sjhb */
115153570Sjhb#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
116153570Sjhb
117153570Sjhb#define	WITNESS_NCHILDREN 6
118153570Sjhb
119153570Sjhbstruct witness_child_list_entry;
120153570Sjhb
121153570Sjhbstruct witness {
122153570Sjhb	const	char *w_name;
123153570Sjhb	struct	lock_class *w_class;
124153570Sjhb	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
125153570Sjhb	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
126153570Sjhb	struct	witness_child_list_entry *w_children;	/* Great evilness... */
127153570Sjhb	const	char *w_file;
128153570Sjhb	int	w_line;
129153570Sjhb	u_int	w_level;
130153570Sjhb	u_int	w_refcount;
131153570Sjhb	u_char	w_Giant_squawked:1;
132153570Sjhb	u_char	w_other_squawked:1;
133153570Sjhb	u_char	w_same_squawked:1;
134153570Sjhb	u_char	w_displayed:1;
135153570Sjhb};
136153570Sjhb
137153570Sjhbstruct witness_child_list_entry {
138153570Sjhb	struct	witness_child_list_entry *wcl_next;
139153570Sjhb	struct	witness *wcl_children[WITNESS_NCHILDREN];
140153570Sjhb	u_int	wcl_count;
141153570Sjhb};
142153570Sjhb
143153570SjhbSTAILQ_HEAD(witness_list, witness);
144153570Sjhb
145153570Sjhb#ifdef BLESSING
146153570Sjhbstruct witness_blessed {
147153570Sjhb	const	char *b_lock1;
148153570Sjhb	const	char *b_lock2;
149153570Sjhb};
150153570Sjhb#endif
151153570Sjhb
152153570Sjhbstruct witness_order_list_entry {
153153570Sjhb	const	char *w_name;
154153570Sjhb	struct	lock_class *w_class;
155153570Sjhb};
156153570Sjhb
157153570Sjhb#ifdef BLESSING
158153570Sjhbstatic int	blessed(struct witness *, struct witness *);
159153570Sjhb#endif
160153570Sjhbstatic int	depart(struct witness *w);
161153570Sjhbstatic struct	witness *enroll(const char *description,
162153570Sjhb				struct lock_class *lock_class);
163153570Sjhbstatic int	insertchild(struct witness *parent, struct witness *child);
164153570Sjhbstatic int	isitmychild(struct witness *parent, struct witness *child);
165153570Sjhbstatic int	isitmydescendant(struct witness *parent, struct witness *child);
166153570Sjhbstatic int	itismychild(struct witness *parent, struct witness *child);
167153570Sjhbstatic int	rebalancetree(struct witness_list *list);
168153570Sjhbstatic void	removechild(struct witness *parent, struct witness *child);
169153570Sjhbstatic int	reparentchildren(struct witness *newparent,
170153570Sjhb		    struct witness *oldparent);
171153570Sjhbstatic int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
172153570Sjhbstatic void	witness_displaydescendants(void(*)(const char *fmt, ...),
173153570Sjhb					   struct witness *, int indent);
174153570Sjhbstatic const char *fixup_filename(const char *file);
175153570Sjhbstatic void	witness_leveldescendents(struct witness *parent, int level);
176153570Sjhbstatic void	witness_levelall(void);
177153570Sjhbstatic struct	witness *witness_get(void);
178153570Sjhbstatic void	witness_free(struct witness *m);
179153570Sjhbstatic struct	witness_child_list_entry *witness_child_get(void);
180153570Sjhbstatic void	witness_child_free(struct witness_child_list_entry *wcl);
181153570Sjhbstatic struct	lock_list_entry *witness_lock_list_get(void);
182153570Sjhbstatic void	witness_lock_list_free(struct lock_list_entry *lle);
183153570Sjhbstatic struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
184153570Sjhb					     struct lock_object *lock);
185153570Sjhbstatic void	witness_list_lock(struct lock_instance *instance);
186153570Sjhb#ifdef DDB
187153570Sjhbstatic void	witness_list(struct thread *td);
188153570Sjhbstatic void	witness_display_list(void(*prnt)(const char *fmt, ...),
189153570Sjhb				     struct witness_list *list);
190153570Sjhbstatic void	witness_display(void(*)(const char *fmt, ...));
191153570Sjhb#endif
192153570Sjhb
193153570SjhbMALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
194153570Sjhb
195153570Sjhb/*
196153570Sjhb * If set to 0, witness is disabled.  If set to 1, witness performs full lock
197153570Sjhb * order checking for all locks.  If set to 2 or higher, then witness skips
198153570Sjhb * the full lock order check if the lock being acquired is at a higher level
199153570Sjhb * (i.e. farther down in the tree) than the current lock.  This last mode is
200153570Sjhb * somewhat experimental and not considered fully safe.  At runtime, this
201153570Sjhb * value may be set to 0 to turn off witness.  witness is not allowed be
202153570Sjhb * turned on once it is turned off, however.
203153570Sjhb */
204153570Sjhbstatic int witness_watch = 1;
205153570SjhbTUNABLE_INT("debug.witness_watch", &witness_watch);
206153570SjhbSYSCTL_PROC(_debug, OID_AUTO, witness_watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
207153570Sjhb    sysctl_debug_witness_watch, "I", "witness is watching lock operations");
208153570Sjhb
209153570Sjhb#ifdef DDB
210153570Sjhb/*
211153570Sjhb * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
212153570Sjhb * drop into kdebug() when:
213153570Sjhb *	- a lock heirarchy violation occurs
214153570Sjhb *	- locks are held when going to sleep.
215153570Sjhb */
216153570Sjhb#ifdef WITNESS_DDB
217153570Sjhbint	witness_ddb = 1;
218153570Sjhb#else
219153570Sjhbint	witness_ddb = 0;
220153570Sjhb#endif
221153570SjhbTUNABLE_INT("debug.witness_ddb", &witness_ddb);
222153570SjhbSYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
223153570Sjhb
224153570Sjhb/*
225153570Sjhb * When DDB is enabled and witness_trace is set to 1, it will cause the system
226153570Sjhb * to print a stack trace:
227153570Sjhb *	- a lock heirarchy violation occurs
228153570Sjhb *	- locks are held when going to sleep.
229153570Sjhb */
230153570Sjhbint	witness_trace = 1;
231153570SjhbTUNABLE_INT("debug.witness_trace", &witness_trace);
232153570SjhbSYSCTL_INT(_debug, OID_AUTO, witness_trace, CTLFLAG_RW, &witness_trace, 0, "");
233153570Sjhb#endif /* DDB */
234153570Sjhb
235153570Sjhb#ifdef WITNESS_SKIPSPIN
236153570Sjhbint	witness_skipspin = 1;
237153570Sjhb#else
238153570Sjhbint	witness_skipspin = 0;
239153570Sjhb#endif
240153570SjhbTUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
241153570SjhbSYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0,
242153570Sjhb    "");
243153570Sjhb
244153570Sjhbstatic struct mtx w_mtx;
245153570Sjhbstatic struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
246153570Sjhbstatic struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
247153570Sjhbstatic struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
248153570Sjhbstatic struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
249153570Sjhbstatic struct witness_child_list_entry *w_child_free = NULL;
250static struct lock_list_entry *w_lock_list_free = NULL;
251
252static struct witness w_data[WITNESS_COUNT];
253static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
254static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
255
256static 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#endif
276	{ "sio", &lock_class_mtx_spin },
277#ifdef __i386__
278	{ "cy", &lock_class_mtx_spin },
279#endif
280	{ "sabtty", &lock_class_mtx_spin },
281	{ "zstty", &lock_class_mtx_spin },
282	{ "ng_node", &lock_class_mtx_spin },
283	{ "ng_worklist", &lock_class_mtx_spin },
284	{ "taskqueue_fast", &lock_class_mtx_spin },
285	{ "intr table", &lock_class_mtx_spin },
286	{ "ithread table lock", &lock_class_mtx_spin },
287	{ "sched lock", &lock_class_mtx_spin },
288	{ "turnstile chain", &lock_class_mtx_spin },
289	{ "td_contested", &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(__amd64__)
300	{ "tlb", &lock_class_mtx_spin },
301	{ "lazypmap", &lock_class_mtx_spin },
302#endif
303#ifdef __sparc64__
304	{ "ipi", &lock_class_mtx_spin },
305#endif
306#endif
307	{ "clk", &lock_class_mtx_spin },
308	{ "mutex profiling lock", &lock_class_mtx_spin },
309	{ "kse zombie lock", &lock_class_mtx_spin },
310	{ "ALD Queue", &lock_class_mtx_spin },
311#ifdef __ia64__
312	{ "MCA spin lock", &lock_class_mtx_spin },
313#endif
314#if defined(__i386__) || defined(__amd64__)
315	{ "pcicfg", &lock_class_mtx_spin },
316#endif
317	{ NULL, NULL },
318	{ NULL, NULL }
319};
320
321#ifdef BLESSING
322/*
323 * Pairs of locks which have been blessed
324 * Don't complain about order problems with blessed locks
325 */
326static struct witness_blessed blessed_list[] = {
327};
328static int blessed_count =
329	sizeof(blessed_list) / sizeof(struct witness_blessed);
330#endif
331
332/*
333 * List of all locks in the system.
334 */
335TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
336
337static struct mtx all_mtx = {
338	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
339	  "All locks list",		/* mtx_object.lo_name */
340	  "All locks list",		/* mtx_object.lo_type */
341	  LO_INITIALIZED,		/* mtx_object.lo_flags */
342	  { NULL, NULL },		/* mtx_object.lo_list */
343	  NULL },			/* mtx_object.lo_witness */
344	MTX_UNOWNED, 0			/* mtx_lock, mtx_recurse */
345};
346
347/*
348 * This global is set to 0 once it becomes safe to use the witness code.
349 */
350static int witness_cold = 1;
351
352/*
353 * Global variables for book keeping.
354 */
355static int lock_cur_cnt;
356static int lock_max_cnt;
357
358/*
359 * The WITNESS-enabled diagnostic code.
360 */
361static void
362witness_initialize(void *dummy __unused)
363{
364	struct lock_object *lock;
365	struct witness_order_list_entry *order;
366	struct witness *w, *w1;
367	int i;
368
369	/*
370	 * We have to release Giant before initializing its witness
371	 * structure so that WITNESS doesn't get confused.
372	 */
373	mtx_unlock(&Giant);
374	mtx_assert(&Giant, MA_NOTOWNED);
375
376	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
377	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
378	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
379	    MTX_NOWITNESS);
380	for (i = 0; i < WITNESS_COUNT; i++)
381		witness_free(&w_data[i]);
382	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
383		witness_child_free(&w_childdata[i]);
384	for (i = 0; i < LOCK_CHILDCOUNT; i++)
385		witness_lock_list_free(&w_locklistdata[i]);
386
387	/* First add in all the specified order lists. */
388	for (order = order_lists; order->w_name != NULL; order++) {
389		w = enroll(order->w_name, order->w_class);
390		if (w == NULL)
391			continue;
392		w->w_file = "order list";
393		for (order++; order->w_name != NULL; order++) {
394			w1 = enroll(order->w_name, order->w_class);
395			if (w1 == NULL)
396				continue;
397			w1->w_file = "order list";
398			if (!itismychild(w, w1))
399				panic("Not enough memory for static orders!");
400			w = w1;
401		}
402	}
403
404	/* Iterate through all locks and add them to witness. */
405	mtx_lock(&all_mtx);
406	TAILQ_FOREACH(lock, &all_locks, lo_list) {
407		if (lock->lo_flags & LO_WITNESS)
408			lock->lo_witness = enroll(lock->lo_type,
409			    lock->lo_class);
410		else
411			lock->lo_witness = NULL;
412	}
413	mtx_unlock(&all_mtx);
414
415	/* Mark the witness code as being ready for use. */
416	atomic_store_rel_int(&witness_cold, 0);
417
418	mtx_lock(&Giant);
419}
420SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
421
422static int
423sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
424{
425	int error, value;
426
427	value = witness_watch;
428	error = sysctl_handle_int(oidp, &value, 0, req);
429	if (error != 0 || req->newptr == NULL)
430		return (error);
431	error = suser(req->td);
432	if (error != 0)
433		return (error);
434	if (value == witness_watch)
435		return (0);
436	if (value != 0)
437		return (EINVAL);
438	witness_watch = 0;
439	return (0);
440}
441
442void
443witness_init(struct lock_object *lock)
444{
445	struct lock_class *class;
446
447	class = lock->lo_class;
448	if (lock->lo_flags & LO_INITIALIZED)
449		panic("%s: lock (%s) %s is already initialized", __func__,
450		    class->lc_name, lock->lo_name);
451	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
452	    (class->lc_flags & LC_RECURSABLE) == 0)
453		panic("%s: lock (%s) %s can not be recursable", __func__,
454		    class->lc_name, lock->lo_name);
455	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
456	    (class->lc_flags & LC_SLEEPABLE) == 0)
457		panic("%s: lock (%s) %s can not be sleepable", __func__,
458		    class->lc_name, lock->lo_name);
459	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
460	    (class->lc_flags & LC_UPGRADABLE) == 0)
461		panic("%s: lock (%s) %s can not be upgradable", __func__,
462		    class->lc_name, lock->lo_name);
463
464	mtx_lock(&all_mtx);
465	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
466	lock->lo_flags |= LO_INITIALIZED;
467	lock_cur_cnt++;
468	if (lock_cur_cnt > lock_max_cnt)
469		lock_max_cnt = lock_cur_cnt;
470	mtx_unlock(&all_mtx);
471	if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
472	    (lock->lo_flags & LO_WITNESS) != 0)
473		lock->lo_witness = enroll(lock->lo_type, class);
474	else
475		lock->lo_witness = NULL;
476}
477
478void
479witness_destroy(struct lock_object *lock)
480{
481	struct witness *w;
482
483	if (witness_cold)
484		panic("lock (%s) %s destroyed while witness_cold",
485		    lock->lo_class->lc_name, lock->lo_name);
486	if ((lock->lo_flags & LO_INITIALIZED) == 0)
487		panic("%s: lock (%s) %s is not initialized", __func__,
488		    lock->lo_class->lc_name, lock->lo_name);
489
490	/* XXX: need to verify that no one holds the lock */
491	w = lock->lo_witness;
492	if (w != NULL) {
493		mtx_lock_spin(&w_mtx);
494		MPASS(w->w_refcount > 0);
495		w->w_refcount--;
496
497		/*
498		 * Lock is already released if we have an allocation failure
499		 * and depart() fails.
500		 */
501		if (w->w_refcount != 0 || depart(w))
502			mtx_unlock_spin(&w_mtx);
503	}
504
505	mtx_lock(&all_mtx);
506	lock_cur_cnt--;
507	TAILQ_REMOVE(&all_locks, lock, lo_list);
508	lock->lo_flags &= ~LO_INITIALIZED;
509	mtx_unlock(&all_mtx);
510}
511
512#ifdef DDB
513static void
514witness_display_list(void(*prnt)(const char *fmt, ...),
515		     struct witness_list *list)
516{
517	struct witness *w;
518
519	STAILQ_FOREACH(w, list, w_typelist) {
520		if (w->w_file == NULL || w->w_level > 0)
521			continue;
522		/*
523		 * This lock has no anscestors, display its descendants.
524		 */
525		witness_displaydescendants(prnt, w, 0);
526	}
527}
528
529static void
530witness_display(void(*prnt)(const char *fmt, ...))
531{
532	struct witness *w;
533
534	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
535	witness_levelall();
536
537	/* Clear all the displayed flags. */
538	STAILQ_FOREACH(w, &w_all, w_list) {
539		w->w_displayed = 0;
540	}
541
542	/*
543	 * First, handle sleep locks which have been acquired at least
544	 * once.
545	 */
546	prnt("Sleep locks:\n");
547	witness_display_list(prnt, &w_sleep);
548
549	/*
550	 * Now do spin locks which have been acquired at least once.
551	 */
552	prnt("\nSpin locks:\n");
553	witness_display_list(prnt, &w_spin);
554
555	/*
556	 * Finally, any locks which have not been acquired yet.
557	 */
558	prnt("\nLocks which were never acquired:\n");
559	STAILQ_FOREACH(w, &w_all, w_list) {
560		if (w->w_file != NULL || w->w_refcount == 0)
561			continue;
562		prnt("%s\n", w->w_name);
563	}
564}
565#endif /* DDB */
566
567/* Trim useless garbage from filenames. */
568static const char *
569fixup_filename(const char *file)
570{
571
572	if (file == NULL)
573		return (NULL);
574	while (strncmp(file, "../", 3) == 0)
575		file += 3;
576	return (file);
577}
578
579void
580witness_lock(struct lock_object *lock, int flags, const char *file, int line)
581{
582	struct lock_list_entry **lock_list, *lle;
583	struct lock_instance *lock1, *lock2;
584	struct lock_class *class;
585	struct witness *w, *w1;
586	struct thread *td;
587	int i, j;
588#ifdef DDB
589	int go_into_ddb = 0;
590#endif
591
592	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
593	    panicstr != NULL)
594		return;
595	w = lock->lo_witness;
596	class = lock->lo_class;
597	td = curthread;
598	file = fixup_filename(file);
599
600	if (class->lc_flags & LC_SLEEPLOCK) {
601		/*
602		 * Since spin locks include a critical section, this check
603		 * impliclty enforces a lock order of all sleep locks before
604		 * all spin locks.
605		 */
606		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
607			panic("blockable sleep lock (%s) %s @ %s:%d",
608			    class->lc_name, lock->lo_name, file, line);
609		lock_list = &td->td_sleeplocks;
610	} else
611		lock_list = PCPU_PTR(spinlocks);
612
613	/*
614	 * Is this the first lock acquired?  If so, then no order checking
615	 * is needed.
616	 */
617	if (*lock_list == NULL)
618		goto out;
619
620	/*
621	 * Check to see if we are recursing on a lock we already own.
622	 */
623	lock1 = find_instance(*lock_list, lock);
624	if (lock1 != NULL) {
625		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
626		    (flags & LOP_EXCLUSIVE) == 0) {
627			printf("shared lock of (%s) %s @ %s:%d\n",
628			    class->lc_name, lock->lo_name, file, line);
629			printf("while exclusively locked from %s:%d\n",
630			    lock1->li_file, lock1->li_line);
631			panic("share->excl");
632		}
633		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
634		    (flags & LOP_EXCLUSIVE) != 0) {
635			printf("exclusive lock of (%s) %s @ %s:%d\n",
636			    class->lc_name, lock->lo_name, file, line);
637			printf("while share locked from %s:%d\n",
638			    lock1->li_file, lock1->li_line);
639			panic("excl->share");
640		}
641		lock1->li_flags++;
642		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
643			printf(
644			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
645			    class->lc_name, lock->lo_name, file, line);
646			printf("first acquired @ %s:%d\n", lock1->li_file,
647			    lock1->li_line);
648			panic("recurse");
649		}
650		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
651		    td->td_proc->p_pid, lock->lo_name,
652		    lock1->li_flags & LI_RECURSEMASK);
653		lock1->li_file = file;
654		lock1->li_line = line;
655		return;
656	}
657
658	/*
659	 * Try locks do not block if they fail to acquire the lock, thus
660	 * there is no danger of deadlocks or of switching while holding a
661	 * spin lock if we acquire a lock via a try operation.
662	 */
663	if (flags & LOP_TRYLOCK)
664		goto out;
665
666	/*
667	 * Check for duplicate locks of the same type.  Note that we only
668	 * have to check for this on the last lock we just acquired.  Any
669	 * other cases will be caught as lock order violations.
670	 */
671	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
672	w1 = lock1->li_lock->lo_witness;
673	if (w1 == w) {
674		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
675			goto out;
676		w->w_same_squawked = 1;
677		printf("acquiring duplicate lock of same type: \"%s\"\n",
678			lock->lo_type);
679		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
680		    lock1->li_file, lock1->li_line);
681		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
682#ifdef DDB
683		go_into_ddb = 1;
684#endif
685		goto out;
686	}
687	MPASS(!mtx_owned(&w_mtx));
688	mtx_lock_spin(&w_mtx);
689	/*
690	 * If we have a known higher number just say ok
691	 */
692	if (witness_watch > 1 && w->w_level > w1->w_level) {
693		mtx_unlock_spin(&w_mtx);
694		goto out;
695	}
696	/*
697	 * If we know that the the lock we are acquiring comes after
698	 * the lock we most recently acquired in the lock order tree,
699	 * then there is no need for any further checks.
700	 */
701	if (isitmydescendant(w1, w)) {
702		mtx_unlock_spin(&w_mtx);
703		goto out;
704	}
705	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
706		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
707
708			MPASS(j < WITNESS_COUNT);
709			lock1 = &lle->ll_children[i];
710			w1 = lock1->li_lock->lo_witness;
711
712			/*
713			 * If this lock doesn't undergo witness checking,
714			 * then skip it.
715			 */
716			if (w1 == NULL) {
717				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
718				    ("lock missing witness structure"));
719				continue;
720			}
721			/*
722			 * If we are locking Giant and this is a sleepable
723			 * lock, then skip it.
724			 */
725			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
726			    lock == &Giant.mtx_object)
727				continue;
728			/*
729			 * If we are locking a sleepable lock and this lock
730			 * is Giant, then skip it.
731			 */
732			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
733			    lock1->li_lock == &Giant.mtx_object)
734				continue;
735			/*
736			 * If we are locking a sleepable lock and this lock
737			 * isn't sleepable, we want to treat it as a lock
738			 * order violation to enfore a general lock order of
739			 * sleepable locks before non-sleepable locks.
740			 */
741			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
742			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
743			    /*
744			     * Check the lock order hierarchy for a reveresal.
745			     */
746			    if (!isitmydescendant(w, w1))
747				continue;
748			/*
749			 * We have a lock order violation, check to see if it
750			 * is allowed or has already been yelled about.
751			 */
752			mtx_unlock_spin(&w_mtx);
753#ifdef BLESSING
754			if (blessed(w, w1))
755				goto out;
756#endif
757			if (lock1->li_lock == &Giant.mtx_object) {
758				if (w1->w_Giant_squawked)
759					goto out;
760				else
761					w1->w_Giant_squawked = 1;
762			} else {
763				if (w1->w_other_squawked)
764					goto out;
765				else
766					w1->w_other_squawked = 1;
767			}
768			/*
769			 * Ok, yell about it.
770			 */
771			printf("lock order reversal\n");
772			/*
773			 * Try to locate an earlier lock with
774			 * witness w in our list.
775			 */
776			do {
777				lock2 = &lle->ll_children[i];
778				MPASS(lock2->li_lock != NULL);
779				if (lock2->li_lock->lo_witness == w)
780					break;
781				i--;
782				if (i == 0 && lle->ll_next != NULL) {
783					lle = lle->ll_next;
784					i = lle->ll_count - 1;
785					MPASS(i >= 0 && i < LOCK_NCHILDREN);
786				}
787			} while (i >= 0);
788			if (i < 0) {
789				printf(" 1st %p %s (%s) @ %s:%d\n",
790				    lock1->li_lock, lock1->li_lock->lo_name,
791				    lock1->li_lock->lo_type, lock1->li_file,
792				    lock1->li_line);
793				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
794				    lock->lo_name, lock->lo_type, file, line);
795			} else {
796				printf(" 1st %p %s (%s) @ %s:%d\n",
797				    lock2->li_lock, lock2->li_lock->lo_name,
798				    lock2->li_lock->lo_type, lock2->li_file,
799				    lock2->li_line);
800				printf(" 2nd %p %s (%s) @ %s:%d\n",
801				    lock1->li_lock, lock1->li_lock->lo_name,
802				    lock1->li_lock->lo_type, lock1->li_file,
803				    lock1->li_line);
804				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
805				    lock->lo_name, lock->lo_type, file, line);
806			}
807#ifdef DDB
808			go_into_ddb = 1;
809#endif
810			goto out;
811		}
812	}
813	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
814	/*
815	 * Don't build a new relationship between a sleepable lock and
816	 * Giant if it is the wrong direction.  The real lock order is that
817	 * sleepable locks come before Giant.
818	 */
819	if (!(lock1->li_lock == &Giant.mtx_object &&
820	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
821		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
822		    lock->lo_type, lock1->li_lock->lo_type);
823		if (!itismychild(lock1->li_lock->lo_witness, w))
824			/* Witness is dead. */
825			return;
826	}
827	mtx_unlock_spin(&w_mtx);
828
829out:
830#ifdef DDB
831	if (go_into_ddb) {
832		if (witness_trace)
833			backtrace();
834		if (witness_ddb)
835			Debugger(__func__);
836	}
837#endif
838	w->w_file = file;
839	w->w_line = line;
840
841	lle = *lock_list;
842	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
843		lle = witness_lock_list_get();
844		if (lle == NULL)
845			return;
846		lle->ll_next = *lock_list;
847		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
848		    td->td_proc->p_pid, lle);
849		*lock_list = lle;
850	}
851	lock1 = &lle->ll_children[lle->ll_count++];
852	lock1->li_lock = lock;
853	lock1->li_line = line;
854	lock1->li_file = file;
855	if ((flags & LOP_EXCLUSIVE) != 0)
856		lock1->li_flags = LI_EXCLUSIVE;
857	else
858		lock1->li_flags = 0;
859	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
860	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
861}
862
863void
864witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
865{
866	struct lock_instance *instance;
867	struct lock_class *class;
868
869	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
870	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
871		return;
872	class = lock->lo_class;
873	file = fixup_filename(file);
874	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
875		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
876		    class->lc_name, lock->lo_name, file, line);
877	if ((flags & LOP_TRYLOCK) == 0)
878		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
879		    lock->lo_name, file, line);
880	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
881		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
882		    class->lc_name, lock->lo_name, file, line);
883	instance = find_instance(curthread->td_sleeplocks, lock);
884	if (instance == NULL)
885		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
886		    class->lc_name, lock->lo_name, file, line);
887	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
888		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
889		    class->lc_name, lock->lo_name, file, line);
890	if ((instance->li_flags & LI_RECURSEMASK) != 0)
891		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
892		    class->lc_name, lock->lo_name,
893		    instance->li_flags & LI_RECURSEMASK, file, line);
894	instance->li_flags |= LI_EXCLUSIVE;
895}
896
897void
898witness_downgrade(struct lock_object *lock, int flags, const char *file,
899    int line)
900{
901	struct lock_instance *instance;
902	struct lock_class *class;
903
904	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
905	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
906		return;
907	class = lock->lo_class;
908	file = fixup_filename(file);
909	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
910		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
911		    class->lc_name, lock->lo_name, file, line);
912	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
913		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
914		    class->lc_name, lock->lo_name, file, line);
915	instance = find_instance(curthread->td_sleeplocks, lock);
916	if (instance == NULL)
917		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
918		    class->lc_name, lock->lo_name, file, line);
919	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
920		panic("downgrade of shared lock (%s) %s @ %s:%d",
921		    class->lc_name, lock->lo_name, file, line);
922	if ((instance->li_flags & LI_RECURSEMASK) != 0)
923		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
924		    class->lc_name, lock->lo_name,
925		    instance->li_flags & LI_RECURSEMASK, file, line);
926	instance->li_flags &= ~LI_EXCLUSIVE;
927}
928
929void
930witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
931{
932	struct lock_list_entry **lock_list, *lle;
933	struct lock_instance *instance;
934	struct lock_class *class;
935	struct thread *td;
936	register_t s;
937	int i, j;
938
939	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
940	    panicstr != NULL)
941		return;
942	td = curthread;
943	class = lock->lo_class;
944	file = fixup_filename(file);
945	if (class->lc_flags & LC_SLEEPLOCK)
946		lock_list = &td->td_sleeplocks;
947	else
948		lock_list = PCPU_PTR(spinlocks);
949	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
950		for (i = 0; i < (*lock_list)->ll_count; i++) {
951			instance = &(*lock_list)->ll_children[i];
952			if (instance->li_lock == lock) {
953				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
954				    (flags & LOP_EXCLUSIVE) == 0) {
955					printf(
956					"shared unlock of (%s) %s @ %s:%d\n",
957					    class->lc_name, lock->lo_name,
958					    file, line);
959					printf(
960					"while exclusively locked from %s:%d\n",
961					    instance->li_file,
962					    instance->li_line);
963					panic("excl->ushare");
964				}
965				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
966				    (flags & LOP_EXCLUSIVE) != 0) {
967					printf(
968					"exclusive unlock of (%s) %s @ %s:%d\n",
969					    class->lc_name, lock->lo_name,
970					    file, line);
971					printf(
972					"while share locked from %s:%d\n",
973					    instance->li_file,
974					    instance->li_line);
975					panic("share->uexcl");
976				}
977				/* If we are recursed, unrecurse. */
978				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
979					CTR4(KTR_WITNESS,
980				    "%s: pid %d unrecursed on %s r=%d", __func__,
981					    td->td_proc->p_pid,
982					    instance->li_lock->lo_name,
983					    instance->li_flags);
984					instance->li_flags--;
985					return;
986				}
987				s = intr_disable();
988				CTR4(KTR_WITNESS,
989				    "%s: pid %d removed %s from lle[%d]", __func__,
990				    td->td_proc->p_pid,
991				    instance->li_lock->lo_name,
992				    (*lock_list)->ll_count - 1);
993				for (j = i; j < (*lock_list)->ll_count - 1; j++)
994					(*lock_list)->ll_children[j] =
995					    (*lock_list)->ll_children[j + 1];
996				(*lock_list)->ll_count--;
997				intr_restore(s);
998				if ((*lock_list)->ll_count == 0) {
999					lle = *lock_list;
1000					*lock_list = lle->ll_next;
1001					CTR3(KTR_WITNESS,
1002					    "%s: pid %d removed lle %p", __func__,
1003					    td->td_proc->p_pid, lle);
1004					witness_lock_list_free(lle);
1005				}
1006				return;
1007			}
1008		}
1009	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1010	    file, line);
1011}
1012
1013/*
1014 * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1015 * exempt Giant and sleepable locks from the checks as well.  If any
1016 * non-exempt locks are held, then a supplied message is printed to the
1017 * console along with a list of the offending locks.  If indicated in the
1018 * flags then a failure results in a panic as well.
1019 */
1020int
1021witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1022{
1023	struct lock_list_entry *lle;
1024	struct lock_instance *lock1;
1025	struct thread *td;
1026	va_list ap;
1027	int i, n;
1028
1029	if (witness_cold || witness_watch == 0 || panicstr != NULL)
1030		return (0);
1031	n = 0;
1032	td = curthread;
1033	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1034		for (i = lle->ll_count - 1; i >= 0; i--) {
1035			lock1 = &lle->ll_children[i];
1036			if (lock1->li_lock == lock)
1037				continue;
1038			if (flags & WARN_GIANTOK &&
1039			    lock1->li_lock == &Giant.mtx_object)
1040				continue;
1041			if (flags & WARN_SLEEPOK &&
1042			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1043				continue;
1044			if (n == 0) {
1045				va_start(ap, fmt);
1046				vprintf(fmt, ap);
1047				va_end(ap);
1048				printf(" with the following");
1049				if (flags & WARN_SLEEPOK)
1050					printf(" non-sleepable");
1051				printf(" locks held:\n");
1052			}
1053			n++;
1054			witness_list_lock(lock1);
1055		}
1056	if (PCPU_GET(spinlocks) != NULL) {
1057		/*
1058		 * Since we already hold a spinlock preemption is
1059		 * already blocked.
1060		 */
1061		if (n == 0) {
1062			va_start(ap, fmt);
1063			vprintf(fmt, ap);
1064			va_end(ap);
1065			printf(" with the following");
1066			if (flags & WARN_SLEEPOK)
1067				printf(" non-sleepable");
1068			printf(" locks held:\n");
1069		}
1070		n += witness_list_locks(PCPU_PTR(spinlocks));
1071	}
1072	if (flags & WARN_PANIC && n)
1073		panic("witness_warn");
1074#ifdef DDB
1075	else if (witness_ddb && n)
1076		Debugger(__func__);
1077#endif
1078	return (n);
1079}
1080
1081const char *
1082witness_file(struct lock_object *lock)
1083{
1084	struct witness *w;
1085
1086	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1087		return ("?");
1088	w = lock->lo_witness;
1089	return (w->w_file);
1090}
1091
1092int
1093witness_line(struct lock_object *lock)
1094{
1095	struct witness *w;
1096
1097	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1098		return (0);
1099	w = lock->lo_witness;
1100	return (w->w_line);
1101}
1102
1103static struct witness *
1104enroll(const char *description, struct lock_class *lock_class)
1105{
1106	struct witness *w;
1107
1108	if (!witness_watch || witness_watch == 0 || panicstr != NULL)
1109		return (NULL);
1110	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1111		return (NULL);
1112	mtx_lock_spin(&w_mtx);
1113	STAILQ_FOREACH(w, &w_all, w_list) {
1114		if (w->w_name == description || (w->w_refcount > 0 &&
1115		    strcmp(description, w->w_name) == 0)) {
1116			w->w_refcount++;
1117			mtx_unlock_spin(&w_mtx);
1118			if (lock_class != w->w_class)
1119				panic(
1120				"lock (%s) %s does not match earlier (%s) lock",
1121				    description, lock_class->lc_name,
1122				    w->w_class->lc_name);
1123			return (w);
1124		}
1125	}
1126	/*
1127	 * This isn't quite right, as witness_cold is still 0 while we
1128	 * enroll all the locks initialized before witness_initialize().
1129	 */
1130	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1131		mtx_unlock_spin(&w_mtx);
1132		panic("spin lock %s not in order list", description);
1133	}
1134	if ((w = witness_get()) == NULL)
1135		return (NULL);
1136	w->w_name = description;
1137	w->w_class = lock_class;
1138	w->w_refcount = 1;
1139	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1140	if (lock_class->lc_flags & LC_SPINLOCK)
1141		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1142	else if (lock_class->lc_flags & LC_SLEEPLOCK)
1143		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1144	else {
1145		mtx_unlock_spin(&w_mtx);
1146		panic("lock class %s is not sleep or spin",
1147		    lock_class->lc_name);
1148	}
1149	mtx_unlock_spin(&w_mtx);
1150	return (w);
1151}
1152
1153/* Don't let the door bang you on the way out... */
1154static int
1155depart(struct witness *w)
1156{
1157	struct witness_child_list_entry *wcl, *nwcl;
1158	struct witness_list *list;
1159	struct witness *parent;
1160
1161	MPASS(w->w_refcount == 0);
1162	if (w->w_class->lc_flags & LC_SLEEPLOCK)
1163		list = &w_sleep;
1164	else
1165		list = &w_spin;
1166	/*
1167	 * First, we run through the entire tree looking for any
1168	 * witnesses that the outgoing witness is a child of.  For
1169	 * each parent that we find, we reparent all the direct
1170	 * children of the outgoing witness to its parent.
1171	 */
1172	STAILQ_FOREACH(parent, list, w_typelist) {
1173		if (!isitmychild(parent, w))
1174			continue;
1175		removechild(parent, w);
1176		if (!reparentchildren(parent, w))
1177			return (0);
1178	}
1179
1180	/*
1181	 * Now we go through and free up the child list of the
1182	 * outgoing witness.
1183	 */
1184	for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1185		nwcl = wcl->wcl_next;
1186		witness_child_free(wcl);
1187	}
1188
1189	/*
1190	 * Detach from various lists and free.
1191	 */
1192	STAILQ_REMOVE(list, w, witness, w_typelist);
1193	STAILQ_REMOVE(&w_all, w, witness, w_list);
1194	witness_free(w);
1195
1196	/* Finally, fixup the tree. */
1197	return (rebalancetree(list));
1198}
1199
1200/*
1201 * Prune an entire lock order tree.  We look for cases where a lock
1202 * is now both a descendant and a direct child of a given lock.  In
1203 * that case, we want to remove the direct child link from the tree.
1204 *
1205 * Returns false if insertchild() fails.
1206 */
1207static int
1208rebalancetree(struct witness_list *list)
1209{
1210	struct witness *child, *parent;
1211
1212	STAILQ_FOREACH(child, list, w_typelist) {
1213		STAILQ_FOREACH(parent, list, w_typelist) {
1214			if (!isitmychild(parent, child))
1215				continue;
1216			removechild(parent, child);
1217			if (isitmydescendant(parent, child))
1218				continue;
1219			if (!insertchild(parent, child))
1220				return (0);
1221		}
1222	}
1223	witness_levelall();
1224	return (1);
1225}
1226
1227/*
1228 * Add "child" as a direct child of "parent".  Returns false if
1229 * we fail due to out of memory.
1230 */
1231static int
1232insertchild(struct witness *parent, struct witness *child)
1233{
1234	struct witness_child_list_entry **wcl;
1235
1236	MPASS(child != NULL && parent != NULL);
1237
1238	/*
1239	 * Insert "child" after "parent"
1240	 */
1241	wcl = &parent->w_children;
1242	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1243		wcl = &(*wcl)->wcl_next;
1244	if (*wcl == NULL) {
1245		*wcl = witness_child_get();
1246		if (*wcl == NULL)
1247			return (0);
1248	}
1249	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1250
1251	return (1);
1252}
1253
1254/*
1255 * Make all the direct descendants of oldparent be direct descendants
1256 * of newparent.
1257 */
1258static int
1259reparentchildren(struct witness *newparent, struct witness *oldparent)
1260{
1261	struct witness_child_list_entry *wcl;
1262	int i;
1263
1264	/* Avoid making a witness a child of itself. */
1265	MPASS(!isitmychild(oldparent, newparent));
1266
1267	for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1268		for (i = 0; i < wcl->wcl_count; i++)
1269			if (!insertchild(newparent, wcl->wcl_children[i]))
1270				return (0);
1271	return (1);
1272}
1273
1274static int
1275itismychild(struct witness *parent, struct witness *child)
1276{
1277	struct witness_list *list;
1278
1279	MPASS(child != NULL && parent != NULL);
1280	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1281	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1282		panic(
1283		"%s: parent (%s) and child (%s) are not the same lock type",
1284		    __func__, parent->w_class->lc_name,
1285		    child->w_class->lc_name);
1286
1287	if (!insertchild(parent, child))
1288		return (0);
1289
1290	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1291		list = &w_sleep;
1292	else
1293		list = &w_spin;
1294	return (rebalancetree(list));
1295}
1296
1297static void
1298removechild(struct witness *parent, struct witness *child)
1299{
1300	struct witness_child_list_entry **wcl, *wcl1;
1301	int i;
1302
1303	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1304		for (i = 0; i < (*wcl)->wcl_count; i++)
1305			if ((*wcl)->wcl_children[i] == child)
1306				goto found;
1307	return;
1308found:
1309	(*wcl)->wcl_count--;
1310	if ((*wcl)->wcl_count > i)
1311		(*wcl)->wcl_children[i] =
1312		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1313	MPASS((*wcl)->wcl_children[i] != NULL);
1314	if ((*wcl)->wcl_count != 0)
1315		return;
1316	wcl1 = *wcl;
1317	*wcl = wcl1->wcl_next;
1318	witness_child_free(wcl1);
1319}
1320
1321static int
1322isitmychild(struct witness *parent, struct witness *child)
1323{
1324	struct witness_child_list_entry *wcl;
1325	int i;
1326
1327	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1328		for (i = 0; i < wcl->wcl_count; i++) {
1329			if (wcl->wcl_children[i] == child)
1330				return (1);
1331		}
1332	}
1333	return (0);
1334}
1335
1336static int
1337isitmydescendant(struct witness *parent, struct witness *child)
1338{
1339	struct witness_child_list_entry *wcl;
1340	int i, j;
1341
1342	if (isitmychild(parent, child))
1343		return (1);
1344	j = 0;
1345	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1346		MPASS(j < 1000);
1347		for (i = 0; i < wcl->wcl_count; i++) {
1348			if (isitmydescendant(wcl->wcl_children[i], child))
1349				return (1);
1350		}
1351		j++;
1352	}
1353	return (0);
1354}
1355
1356static void
1357witness_levelall (void)
1358{
1359	struct witness_list *list;
1360	struct witness *w, *w1;
1361
1362	/*
1363	 * First clear all levels.
1364	 */
1365	STAILQ_FOREACH(w, &w_all, w_list) {
1366		w->w_level = 0;
1367	}
1368
1369	/*
1370	 * Look for locks with no parent and level all their descendants.
1371	 */
1372	STAILQ_FOREACH(w, &w_all, w_list) {
1373		/*
1374		 * This is just an optimization, technically we could get
1375		 * away just walking the all list each time.
1376		 */
1377		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1378			list = &w_sleep;
1379		else
1380			list = &w_spin;
1381		STAILQ_FOREACH(w1, list, w_typelist) {
1382			if (isitmychild(w1, w))
1383				goto skip;
1384		}
1385		witness_leveldescendents(w, 0);
1386	skip:
1387		;	/* silence GCC 3.x */
1388	}
1389}
1390
1391static void
1392witness_leveldescendents(struct witness *parent, int level)
1393{
1394	struct witness_child_list_entry *wcl;
1395	int i;
1396
1397	if (parent->w_level < level)
1398		parent->w_level = level;
1399	level++;
1400	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1401		for (i = 0; i < wcl->wcl_count; i++)
1402			witness_leveldescendents(wcl->wcl_children[i], level);
1403}
1404
1405static void
1406witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1407			   struct witness *parent, int indent)
1408{
1409	struct witness_child_list_entry *wcl;
1410	int i, level;
1411
1412	level = parent->w_level;
1413	prnt("%-2d", level);
1414	for (i = 0; i < indent; i++)
1415		prnt(" ");
1416	if (parent->w_refcount > 0)
1417		prnt("%s", parent->w_name);
1418	else
1419		prnt("(dead)");
1420	if (parent->w_displayed) {
1421		prnt(" -- (already displayed)\n");
1422		return;
1423	}
1424	parent->w_displayed = 1;
1425	if (parent->w_refcount > 0) {
1426		if (parent->w_file != NULL)
1427			prnt(" -- last acquired @ %s:%d", parent->w_file,
1428			    parent->w_line);
1429	}
1430	prnt("\n");
1431	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1432		for (i = 0; i < wcl->wcl_count; i++)
1433			    witness_displaydescendants(prnt,
1434				wcl->wcl_children[i], indent + 1);
1435}
1436
1437#ifdef BLESSING
1438static int
1439blessed(struct witness *w1, struct witness *w2)
1440{
1441	int i;
1442	struct witness_blessed *b;
1443
1444	for (i = 0; i < blessed_count; i++) {
1445		b = &blessed_list[i];
1446		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1447			if (strcmp(w2->w_name, b->b_lock2) == 0)
1448				return (1);
1449			continue;
1450		}
1451		if (strcmp(w1->w_name, b->b_lock2) == 0)
1452			if (strcmp(w2->w_name, b->b_lock1) == 0)
1453				return (1);
1454	}
1455	return (0);
1456}
1457#endif
1458
1459static struct witness *
1460witness_get(void)
1461{
1462	struct witness *w;
1463
1464	if (witness_watch == 0) {
1465		mtx_unlock_spin(&w_mtx);
1466		return (NULL);
1467	}
1468	if (STAILQ_EMPTY(&w_free)) {
1469		witness_watch = 0;
1470		mtx_unlock_spin(&w_mtx);
1471		printf("%s: witness exhausted\n", __func__);
1472		return (NULL);
1473	}
1474	w = STAILQ_FIRST(&w_free);
1475	STAILQ_REMOVE_HEAD(&w_free, w_list);
1476	bzero(w, sizeof(*w));
1477	return (w);
1478}
1479
1480static void
1481witness_free(struct witness *w)
1482{
1483
1484	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1485}
1486
1487static struct witness_child_list_entry *
1488witness_child_get(void)
1489{
1490	struct witness_child_list_entry *wcl;
1491
1492	if (witness_watch == 0) {
1493		mtx_unlock_spin(&w_mtx);
1494		return (NULL);
1495	}
1496	wcl = w_child_free;
1497	if (wcl == NULL) {
1498		witness_watch = 0;
1499		mtx_unlock_spin(&w_mtx);
1500		printf("%s: witness exhausted\n", __func__);
1501		return (NULL);
1502	}
1503	w_child_free = wcl->wcl_next;
1504	bzero(wcl, sizeof(*wcl));
1505	return (wcl);
1506}
1507
1508static void
1509witness_child_free(struct witness_child_list_entry *wcl)
1510{
1511
1512	wcl->wcl_next = w_child_free;
1513	w_child_free = wcl;
1514}
1515
1516static struct lock_list_entry *
1517witness_lock_list_get(void)
1518{
1519	struct lock_list_entry *lle;
1520
1521	if (witness_watch == 0)
1522		return (NULL);
1523	mtx_lock_spin(&w_mtx);
1524	lle = w_lock_list_free;
1525	if (lle == NULL) {
1526		witness_watch = 0;
1527		mtx_unlock_spin(&w_mtx);
1528		printf("%s: witness exhausted\n", __func__);
1529		return (NULL);
1530	}
1531	w_lock_list_free = lle->ll_next;
1532	mtx_unlock_spin(&w_mtx);
1533	bzero(lle, sizeof(*lle));
1534	return (lle);
1535}
1536
1537static void
1538witness_lock_list_free(struct lock_list_entry *lle)
1539{
1540
1541	mtx_lock_spin(&w_mtx);
1542	lle->ll_next = w_lock_list_free;
1543	w_lock_list_free = lle;
1544	mtx_unlock_spin(&w_mtx);
1545}
1546
1547static struct lock_instance *
1548find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1549{
1550	struct lock_list_entry *lle;
1551	struct lock_instance *instance;
1552	int i;
1553
1554	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1555		for (i = lle->ll_count - 1; i >= 0; i--) {
1556			instance = &lle->ll_children[i];
1557			if (instance->li_lock == lock)
1558				return (instance);
1559		}
1560	return (NULL);
1561}
1562
1563static void
1564witness_list_lock(struct lock_instance *instance)
1565{
1566	struct lock_object *lock;
1567
1568	lock = instance->li_lock;
1569	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1570	    "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1571	if (lock->lo_type != lock->lo_name)
1572		printf(" (%s)", lock->lo_type);
1573	printf(" r = %d (%p) locked @ %s:%d\n",
1574	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1575	    instance->li_line);
1576}
1577
1578int
1579witness_list_locks(struct lock_list_entry **lock_list)
1580{
1581	struct lock_list_entry *lle;
1582	int i, nheld;
1583
1584	nheld = 0;
1585	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1586		for (i = lle->ll_count - 1; i >= 0; i--) {
1587			witness_list_lock(&lle->ll_children[i]);
1588			nheld++;
1589		}
1590	return (nheld);
1591}
1592
1593/*
1594 * This is a bit risky at best.  We call this function when we have timed
1595 * out acquiring a spin lock, and we assume that the other CPU is stuck
1596 * with this lock held.  So, we go groveling around in the other CPU's
1597 * per-cpu data to try to find the lock instance for this spin lock to
1598 * see when it was last acquired.
1599 */
1600void
1601witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1602{
1603	struct lock_instance *instance;
1604	struct pcpu *pc;
1605
1606	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1607		return;
1608	pc = pcpu_find(owner->td_oncpu);
1609	instance = find_instance(pc->pc_spinlocks, lock);
1610	if (instance != NULL)
1611		witness_list_lock(instance);
1612}
1613
1614void
1615witness_save(struct lock_object *lock, const char **filep, int *linep)
1616{
1617	struct lock_instance *instance;
1618
1619	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1620	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1621		return;
1622	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1623		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1624		    lock->lo_class->lc_name, lock->lo_name);
1625	instance = find_instance(curthread->td_sleeplocks, lock);
1626	if (instance == NULL)
1627		panic("%s: lock (%s) %s not locked", __func__,
1628		    lock->lo_class->lc_name, lock->lo_name);
1629	*filep = instance->li_file;
1630	*linep = instance->li_line;
1631}
1632
1633void
1634witness_restore(struct lock_object *lock, const char *file, int line)
1635{
1636	struct lock_instance *instance;
1637
1638	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1639	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1640		return;
1641	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1642		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1643		    lock->lo_class->lc_name, lock->lo_name);
1644	instance = find_instance(curthread->td_sleeplocks, lock);
1645	if (instance == NULL)
1646		panic("%s: lock (%s) %s not locked", __func__,
1647		    lock->lo_class->lc_name, lock->lo_name);
1648	lock->lo_witness->w_file = file;
1649	lock->lo_witness->w_line = line;
1650	instance->li_file = file;
1651	instance->li_line = line;
1652}
1653
1654void
1655witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1656{
1657#ifdef INVARIANT_SUPPORT
1658	struct lock_instance *instance;
1659
1660	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1661		return;
1662	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1663		instance = find_instance(curthread->td_sleeplocks, lock);
1664	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1665		instance = find_instance(PCPU_GET(spinlocks), lock);
1666	else {
1667		panic("Lock (%s) %s is not sleep or spin!",
1668		    lock->lo_class->lc_name, lock->lo_name);
1669	}
1670	file = fixup_filename(file);
1671	switch (flags) {
1672	case LA_UNLOCKED:
1673		if (instance != NULL)
1674			panic("Lock (%s) %s locked @ %s:%d.",
1675			    lock->lo_class->lc_name, lock->lo_name, file, line);
1676		break;
1677	case LA_LOCKED:
1678	case LA_LOCKED | LA_RECURSED:
1679	case LA_LOCKED | LA_NOTRECURSED:
1680	case LA_SLOCKED:
1681	case LA_SLOCKED | LA_RECURSED:
1682	case LA_SLOCKED | LA_NOTRECURSED:
1683	case LA_XLOCKED:
1684	case LA_XLOCKED | LA_RECURSED:
1685	case LA_XLOCKED | LA_NOTRECURSED:
1686		if (instance == NULL) {
1687			panic("Lock (%s) %s not locked @ %s:%d.",
1688			    lock->lo_class->lc_name, lock->lo_name, file, line);
1689			break;
1690		}
1691		if ((flags & LA_XLOCKED) != 0 &&
1692		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1693			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1694			    lock->lo_class->lc_name, lock->lo_name, file, line);
1695		if ((flags & LA_SLOCKED) != 0 &&
1696		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1697			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1698			    lock->lo_class->lc_name, lock->lo_name, file, line);
1699		if ((flags & LA_RECURSED) != 0 &&
1700		    (instance->li_flags & LI_RECURSEMASK) == 0)
1701			panic("Lock (%s) %s not recursed @ %s:%d.",
1702			    lock->lo_class->lc_name, lock->lo_name, file, line);
1703		if ((flags & LA_NOTRECURSED) != 0 &&
1704		    (instance->li_flags & LI_RECURSEMASK) != 0)
1705			panic("Lock (%s) %s recursed @ %s:%d.",
1706			    lock->lo_class->lc_name, lock->lo_name, file, line);
1707		break;
1708	default:
1709		panic("Invalid lock assertion at %s:%d.", file, line);
1710
1711	}
1712#endif	/* INVARIANT_SUPPORT */
1713}
1714
1715#ifdef DDB
1716static void
1717witness_list(struct thread *td)
1718{
1719
1720	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1721	KASSERT(db_active, ("%s: not in the debugger", __func__));
1722
1723	if (witness_watch == 0)
1724		return;
1725
1726	witness_list_locks(&td->td_sleeplocks);
1727
1728	/*
1729	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1730	 * if td is currently executing on some other CPU and holds spin locks
1731	 * as we won't display those locks.  If we had a MI way of getting
1732	 * the per-cpu data for a given cpu then we could use
1733	 * td->td_oncpu to get the list of spinlocks for this thread
1734	 * and "fix" this.
1735	 *
1736	 * That still wouldn't really fix this unless we locked sched_lock
1737	 * or stopped the other CPU to make sure it wasn't changing the list
1738	 * out from under us.  It is probably best to just not try to handle
1739	 * threads on other CPU's for now.
1740	 */
1741	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1742		witness_list_locks(PCPU_PTR(spinlocks));
1743}
1744
1745DB_SHOW_COMMAND(locks, db_witness_list)
1746{
1747	struct thread *td;
1748	pid_t pid;
1749	struct proc *p;
1750
1751	if (have_addr) {
1752		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1753		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1754		    ((addr >> 16) % 16) * 10000;
1755		/* sx_slock(&allproc_lock); */
1756		FOREACH_PROC_IN_SYSTEM(p) {
1757			if (p->p_pid == pid)
1758				break;
1759		}
1760		/* sx_sunlock(&allproc_lock); */
1761		if (p == NULL) {
1762			db_printf("pid %d not found\n", pid);
1763			return;
1764		}
1765		FOREACH_THREAD_IN_PROC(p, td) {
1766			witness_list(td);
1767		}
1768	} else {
1769		td = curthread;
1770		witness_list(td);
1771	}
1772}
1773
1774DB_SHOW_COMMAND(witness, db_witness_display)
1775{
1776
1777	witness_display(db_printf);
1778}
1779#endif
1780