xref: /netbsd-src/sys/kern/subr_lockdebug.c (revision ed75d7a867996c84cfa88e3b8906816277e957f7)
1 /*	$NetBSD: subr_lockdebug.c,v 1.74 2020/01/21 20:31:57 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 2006, 2007, 2008, 2020 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Basic lock debugging code shared among lock primitives.
34  */
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: subr_lockdebug.c,v 1.74 2020/01/21 20:31:57 ad Exp $");
38 
39 #ifdef _KERNEL_OPT
40 #include "opt_ddb.h"
41 #endif
42 
43 #include <sys/param.h>
44 #include <sys/proc.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
47 #include <sys/kmem.h>
48 #include <sys/lockdebug.h>
49 #include <sys/sleepq.h>
50 #include <sys/cpu.h>
51 #include <sys/atomic.h>
52 #include <sys/lock.h>
53 #include <sys/rbtree.h>
54 #include <sys/ksyms.h>
55 
56 #include <machine/lock.h>
57 
58 unsigned int		ld_panic;
59 
60 #ifdef LOCKDEBUG
61 
62 #ifdef __ia64__
63 #define	LD_BATCH_SHIFT	16
64 #else
65 #define	LD_BATCH_SHIFT	9
66 #endif
67 #define	LD_BATCH	(1 << LD_BATCH_SHIFT)
68 #define	LD_BATCH_MASK	(LD_BATCH - 1)
69 #define	LD_MAX_LOCKS	1048576
70 #define	LD_SLOP		16
71 
72 #define	LD_LOCKED	0x01
73 #define	LD_SLEEPER	0x02
74 
75 #define	LD_WRITE_LOCK	0x80000000
76 
77 typedef struct lockdebug {
78 	struct rb_node	ld_rb_node;
79 	__cpu_simple_lock_t ld_spinlock;
80 	_TAILQ_ENTRY(struct lockdebug, volatile) ld_chain;
81 	_TAILQ_ENTRY(struct lockdebug, volatile) ld_achain;
82 	volatile void	*ld_lock;
83 	lockops_t	*ld_lockops;
84 	struct lwp	*ld_lwp;
85 	uintptr_t	ld_locked;
86 	uintptr_t	ld_unlocked;
87 	uintptr_t	ld_initaddr;
88 	uint16_t	ld_shares;
89 	uint16_t	ld_cpu;
90 	uint8_t		ld_flags;
91 	uint8_t		ld_shwant;	/* advisory */
92 	uint8_t		ld_exwant;	/* advisory */
93 	uint8_t		ld_unused;
94 } volatile lockdebug_t;
95 
96 typedef _TAILQ_HEAD(lockdebuglist, struct lockdebug, volatile) lockdebuglist_t;
97 
98 __cpu_simple_lock_t	ld_mod_lk;
99 lockdebuglist_t		ld_free = TAILQ_HEAD_INITIALIZER(ld_free);
100 lockdebuglist_t		ld_all = TAILQ_HEAD_INITIALIZER(ld_all);
101 int			ld_nfree;
102 int			ld_freeptr;
103 int			ld_recurse;
104 bool			ld_nomore;
105 lockdebug_t		ld_prime[LD_BATCH];
106 
107 static void	lockdebug_abort1(const char *, size_t, lockdebug_t *, int,
108     const char *, bool);
109 static int	lockdebug_more(int);
110 static void	lockdebug_init(void);
111 static void	lockdebug_dump(lwp_t *, lockdebug_t *,
112     void (*)(const char *, ...)
113     __printflike(1, 2));
114 
115 static signed int
116 ld_rbto_compare_nodes(void *ctx, const void *n1, const void *n2)
117 {
118 	const lockdebug_t *ld1 = n1;
119 	const lockdebug_t *ld2 = n2;
120 	const uintptr_t a = (uintptr_t)ld1->ld_lock;
121 	const uintptr_t b = (uintptr_t)ld2->ld_lock;
122 
123 	if (a < b)
124 		return -1;
125 	if (a > b)
126 		return 1;
127 	return 0;
128 }
129 
130 static signed int
131 ld_rbto_compare_key(void *ctx, const void *n, const void *key)
132 {
133 	const lockdebug_t *ld = n;
134 	const uintptr_t a = (uintptr_t)ld->ld_lock;
135 	const uintptr_t b = (uintptr_t)key;
136 
137 	if (a < b)
138 		return -1;
139 	if (a > b)
140 		return 1;
141 	return 0;
142 }
143 
144 static rb_tree_t ld_rb_tree;
145 
146 static const rb_tree_ops_t ld_rb_tree_ops = {
147 	.rbto_compare_nodes = ld_rbto_compare_nodes,
148 	.rbto_compare_key = ld_rbto_compare_key,
149 	.rbto_node_offset = offsetof(lockdebug_t, ld_rb_node),
150 	.rbto_context = NULL
151 };
152 
153 static inline lockdebug_t *
154 lockdebug_lookup1(const volatile void *lock)
155 {
156 	lockdebug_t *ld;
157 	struct cpu_info *ci;
158 
159 	ci = curcpu();
160 	__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
161 	ld = rb_tree_find_node(&ld_rb_tree, (void *)(intptr_t)lock);
162 	__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
163 	if (ld == NULL) {
164 		return NULL;
165 	}
166 	__cpu_simple_lock(&ld->ld_spinlock);
167 
168 	return ld;
169 }
170 
171 static void
172 lockdebug_lock_cpus(void)
173 {
174 	CPU_INFO_ITERATOR cii;
175 	struct cpu_info *ci;
176 
177 	for (CPU_INFO_FOREACH(cii, ci)) {
178 		__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
179 	}
180 }
181 
182 static void
183 lockdebug_unlock_cpus(void)
184 {
185 	CPU_INFO_ITERATOR cii;
186 	struct cpu_info *ci;
187 
188 	for (CPU_INFO_FOREACH(cii, ci)) {
189 		__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
190 	}
191 }
192 
193 /*
194  * lockdebug_lookup:
195  *
196  *	Find a lockdebug structure by a pointer to a lock and return it locked.
197  */
198 static inline lockdebug_t *
199 lockdebug_lookup(const char *func, size_t line, const volatile void *lock,
200     uintptr_t where)
201 {
202 	lockdebug_t *ld;
203 
204 	ld = lockdebug_lookup1(lock);
205 	if (__predict_false(ld == NULL)) {
206 		panic("%s,%zu: uninitialized lock (lock=%p, from=%08"
207 		    PRIxPTR ")", func, line, lock, where);
208 	}
209 	return ld;
210 }
211 
212 /*
213  * lockdebug_init:
214  *
215  *	Initialize the lockdebug system.  Allocate an initial pool of
216  *	lockdebug structures before the VM system is up and running.
217  */
218 static void
219 lockdebug_init(void)
220 {
221 	lockdebug_t *ld;
222 	int i;
223 
224 	TAILQ_INIT(&curcpu()->ci_data.cpu_ld_locks);
225 	TAILQ_INIT(&curlwp->l_ld_locks);
226 	__cpu_simple_lock_init(&curcpu()->ci_data.cpu_ld_lock);
227 	__cpu_simple_lock_init(&ld_mod_lk);
228 
229 	rb_tree_init(&ld_rb_tree, &ld_rb_tree_ops);
230 
231 	ld = ld_prime;
232 	for (i = 1, ld++; i < LD_BATCH; i++, ld++) {
233 		__cpu_simple_lock_init(&ld->ld_spinlock);
234 		TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
235 		TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
236 	}
237 	ld_freeptr = 1;
238 	ld_nfree = LD_BATCH - 1;
239 }
240 
241 /*
242  * lockdebug_alloc:
243  *
244  *	A lock is being initialized, so allocate an associated debug
245  *	structure.
246  */
247 bool
248 lockdebug_alloc(const char *func, size_t line, volatile void *lock,
249     lockops_t *lo, uintptr_t initaddr)
250 {
251 	struct cpu_info *ci;
252 	lockdebug_t *ld;
253 	int s;
254 
255 	if (__predict_false(lo == NULL || panicstr != NULL || ld_panic))
256 		return false;
257 	if (__predict_false(ld_freeptr == 0))
258 		lockdebug_init();
259 
260 	s = splhigh();
261 	__cpu_simple_lock(&ld_mod_lk);
262 	if (__predict_false((ld = lockdebug_lookup1(lock)) != NULL)) {
263 		__cpu_simple_unlock(&ld_mod_lk);
264 		lockdebug_abort1(func, line, ld, s, "already initialized",
265 		    true);
266 		return false;
267 	}
268 
269 	/*
270 	 * Pinch a new debug structure.  We may recurse because we call
271 	 * kmem_alloc(), which may need to initialize new locks somewhere
272 	 * down the path.  If not recursing, we try to maintain at least
273 	 * LD_SLOP structures free, which should hopefully be enough to
274 	 * satisfy kmem_alloc().  If we can't provide a structure, not to
275 	 * worry: we'll just mark the lock as not having an ID.
276 	 */
277 	ci = curcpu();
278 	ci->ci_lkdebug_recurse++;
279 	if (TAILQ_EMPTY(&ld_free)) {
280 		if (ci->ci_lkdebug_recurse > 1 || ld_nomore) {
281 			ci->ci_lkdebug_recurse--;
282 			__cpu_simple_unlock(&ld_mod_lk);
283 			splx(s);
284 			return false;
285 		}
286 		s = lockdebug_more(s);
287 	} else if (ci->ci_lkdebug_recurse == 1 && ld_nfree < LD_SLOP) {
288 		s = lockdebug_more(s);
289 	}
290 	if (__predict_false((ld = TAILQ_FIRST(&ld_free)) == NULL)) {
291 		__cpu_simple_unlock(&ld_mod_lk);
292 		splx(s);
293 		return false;
294 	}
295 	TAILQ_REMOVE(&ld_free, ld, ld_chain);
296 	ld_nfree--;
297 	ci->ci_lkdebug_recurse--;
298 
299 	if (__predict_false(ld->ld_lock != NULL)) {
300 		panic("%s,%zu: corrupt table ld %p", func, line, ld);
301 	}
302 
303 	/* Initialise the structure. */
304 	ld->ld_lock = lock;
305 	ld->ld_lockops = lo;
306 	ld->ld_locked = 0;
307 	ld->ld_unlocked = 0;
308 	ld->ld_lwp = NULL;
309 	ld->ld_initaddr = initaddr;
310 	ld->ld_flags = (lo->lo_type == LOCKOPS_SLEEP ? LD_SLEEPER : 0);
311 	lockdebug_lock_cpus();
312 	(void)rb_tree_insert_node(&ld_rb_tree, __UNVOLATILE(ld));
313 	lockdebug_unlock_cpus();
314 	__cpu_simple_unlock(&ld_mod_lk);
315 
316 	splx(s);
317 	return true;
318 }
319 
320 /*
321  * lockdebug_free:
322  *
323  *	A lock is being destroyed, so release debugging resources.
324  */
325 void
326 lockdebug_free(const char *func, size_t line, volatile void *lock)
327 {
328 	lockdebug_t *ld;
329 	int s;
330 
331 	if (__predict_false(panicstr != NULL || ld_panic))
332 		return;
333 
334 	s = splhigh();
335 	__cpu_simple_lock(&ld_mod_lk);
336 	ld = lockdebug_lookup(func, line, lock,
337 	    (uintptr_t) __builtin_return_address(0));
338 	if (__predict_false(ld == NULL)) {
339 		__cpu_simple_unlock(&ld_mod_lk);
340 		panic("%s,%zu: destroying uninitialized object %p"
341 		    "(ld_lock=%p)", func, line, lock, ld->ld_lock);
342 		return;
343 	}
344 	if (__predict_false((ld->ld_flags & LD_LOCKED) != 0 ||
345 	    ld->ld_shares != 0)) {
346 		__cpu_simple_unlock(&ld_mod_lk);
347 		lockdebug_abort1(func, line, ld, s, "is locked or in use",
348 		    true);
349 		return;
350 	}
351 	lockdebug_lock_cpus();
352 	rb_tree_remove_node(&ld_rb_tree, __UNVOLATILE(ld));
353 	lockdebug_unlock_cpus();
354 	ld->ld_lock = NULL;
355 	TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
356 	ld_nfree++;
357 	__cpu_simple_unlock(&ld->ld_spinlock);
358 	__cpu_simple_unlock(&ld_mod_lk);
359 	splx(s);
360 }
361 
362 /*
363  * lockdebug_more:
364  *
365  *	Allocate a batch of debug structures and add to the free list.
366  *	Must be called with ld_mod_lk held.
367  */
368 static int
369 lockdebug_more(int s)
370 {
371 	lockdebug_t *ld;
372 	void *block;
373 	int i, base, m;
374 
375 	/*
376 	 * Can't call kmem_alloc() if in interrupt context.  XXX We could
377 	 * deadlock, because we don't know which locks the caller holds.
378 	 */
379 	if (cpu_intr_p() || cpu_softintr_p()) {
380 		return s;
381 	}
382 
383 	while (ld_nfree < LD_SLOP) {
384 		__cpu_simple_unlock(&ld_mod_lk);
385 		splx(s);
386 		block = kmem_zalloc(LD_BATCH * sizeof(lockdebug_t), KM_SLEEP);
387 		s = splhigh();
388 		__cpu_simple_lock(&ld_mod_lk);
389 
390 		if (ld_nfree > LD_SLOP) {
391 			/* Somebody beat us to it. */
392 			__cpu_simple_unlock(&ld_mod_lk);
393 			splx(s);
394 			kmem_free(block, LD_BATCH * sizeof(lockdebug_t));
395 			s = splhigh();
396 			__cpu_simple_lock(&ld_mod_lk);
397 			continue;
398 		}
399 
400 		base = ld_freeptr;
401 		ld_nfree += LD_BATCH;
402 		ld = block;
403 		base <<= LD_BATCH_SHIFT;
404 		m = uimin(LD_MAX_LOCKS, base + LD_BATCH);
405 
406 		if (m == LD_MAX_LOCKS)
407 			ld_nomore = true;
408 
409 		for (i = base; i < m; i++, ld++) {
410 			__cpu_simple_lock_init(&ld->ld_spinlock);
411 			TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
412 			TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
413 		}
414 
415 		membar_producer();
416 	}
417 
418 	return s;
419 }
420 
421 /*
422  * lockdebug_wantlock:
423  *
424  *	Process the preamble to a lock acquire.  The "shared"
425  *	parameter controls which ld_{ex,sh}want counter is
426  *	updated; a negative value of shared updates neither.
427  */
428 void
429 lockdebug_wantlock(const char *func, size_t line,
430     const volatile void *lock, uintptr_t where, int shared)
431 {
432 	struct lwp *l = curlwp;
433 	lockdebug_t *ld;
434 	bool recurse;
435 	int s;
436 
437 	(void)shared;
438 	recurse = false;
439 
440 	if (__predict_false(panicstr != NULL || ld_panic))
441 		return;
442 
443 	s = splhigh();
444 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
445 		splx(s);
446 		return;
447 	}
448 	if ((ld->ld_flags & LD_LOCKED) != 0 || ld->ld_shares != 0) {
449 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
450 			if (ld->ld_lwp == l)
451 				recurse = true;
452 		} else if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
453 			recurse = true;
454 	}
455 	if (cpu_intr_p()) {
456 		if (__predict_false((ld->ld_flags & LD_SLEEPER) != 0)) {
457 			lockdebug_abort1(func, line, ld, s,
458 			    "acquiring sleep lock from interrupt context",
459 			    true);
460 			return;
461 		}
462 	}
463 	if (shared > 0)
464 		ld->ld_shwant++;
465 	else if (shared == 0)
466 		ld->ld_exwant++;
467 	if (__predict_false(recurse)) {
468 		lockdebug_abort1(func, line, ld, s, "locking against myself",
469 		    true);
470 		return;
471 	}
472 	if (l->l_ld_wanted == NULL) {
473 		l->l_ld_wanted = ld;
474 	}
475 	__cpu_simple_unlock(&ld->ld_spinlock);
476 	splx(s);
477 }
478 
479 /*
480  * lockdebug_locked:
481  *
482  *	Process a lock acquire operation.
483  */
484 void
485 lockdebug_locked(const char *func, size_t line,
486     volatile void *lock, void *cvlock, uintptr_t where, int shared)
487 {
488 	struct lwp *l = curlwp;
489 	lockdebug_t *ld;
490 	int s;
491 
492 	if (__predict_false(panicstr != NULL || ld_panic))
493 		return;
494 
495 	s = splhigh();
496 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
497 		splx(s);
498 		return;
499 	}
500 	if (cvlock) {
501 		KASSERT(ld->ld_lockops->lo_type == LOCKOPS_CV);
502 		if (lock == (void *)&lbolt) {
503 			/* nothing */
504 		} else if (ld->ld_shares++ == 0) {
505 			ld->ld_locked = (uintptr_t)cvlock;
506 		} else if (__predict_false(cvlock != (void *)ld->ld_locked)) {
507 			lockdebug_abort1(func, line, ld, s,
508 			    "multiple locks used with condition variable",
509 			    true);
510 			return;
511 		}
512 	} else if (shared) {
513 		l->l_shlocks++;
514 		ld->ld_locked = where;
515 		ld->ld_shares++;
516 		ld->ld_shwant--;
517 	} else {
518 		if (__predict_false((ld->ld_flags & LD_LOCKED) != 0)) {
519 			lockdebug_abort1(func, line, ld, s, "already locked",
520 			    true);
521 			return;
522 		}
523 		ld->ld_flags |= LD_LOCKED;
524 		ld->ld_locked = where;
525 		ld->ld_exwant--;
526 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
527 			TAILQ_INSERT_TAIL(&l->l_ld_locks, ld, ld_chain);
528 		} else {
529 			TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_ld_locks,
530 			    ld, ld_chain);
531 		}
532 	}
533 	ld->ld_cpu = (uint16_t)cpu_index(curcpu());
534 	ld->ld_lwp = l;
535 	__cpu_simple_unlock(&ld->ld_spinlock);
536 	if (l->l_ld_wanted == ld) {
537 		l->l_ld_wanted = NULL;
538 	}
539 	splx(s);
540 }
541 
542 /*
543  * lockdebug_unlocked:
544  *
545  *	Process a lock release operation.
546  */
547 void
548 lockdebug_unlocked(const char *func, size_t line,
549     volatile void *lock, uintptr_t where, int shared)
550 {
551 	struct lwp *l = curlwp;
552 	lockdebug_t *ld;
553 	int s;
554 
555 	if (__predict_false(panicstr != NULL || ld_panic))
556 		return;
557 
558 	s = splhigh();
559 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
560 		splx(s);
561 		return;
562 	}
563 	if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
564 		if (lock == (void *)&lbolt) {
565 			/* nothing */
566 		} else {
567 			ld->ld_shares--;
568 		}
569 	} else if (shared) {
570 		if (__predict_false(l->l_shlocks == 0)) {
571 			lockdebug_abort1(func, line, ld, s,
572 			    "no shared locks held by LWP", true);
573 			return;
574 		}
575 		if (__predict_false(ld->ld_shares == 0)) {
576 			lockdebug_abort1(func, line, ld, s,
577 			    "no shared holds on this lock", true);
578 			return;
579 		}
580 		l->l_shlocks--;
581 		ld->ld_shares--;
582 		if (ld->ld_lwp == l) {
583 			ld->ld_unlocked = where;
584 			ld->ld_lwp = NULL;
585 		}
586 		if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
587 			ld->ld_cpu = (uint16_t)-1;
588 	} else {
589 		if (__predict_false((ld->ld_flags & LD_LOCKED) == 0)) {
590 			lockdebug_abort1(func, line, ld, s, "not locked", true);
591 			return;
592 		}
593 
594 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
595 			if (__predict_false(ld->ld_lwp != curlwp)) {
596 				lockdebug_abort1(func, line, ld, s,
597 				    "not held by current LWP", true);
598 				return;
599 			}
600 			TAILQ_REMOVE(&l->l_ld_locks, ld, ld_chain);
601 		} else {
602 			uint16_t idx = (uint16_t)cpu_index(curcpu());
603 			if (__predict_false(ld->ld_cpu != idx)) {
604 				lockdebug_abort1(func, line, ld, s,
605 				    "not held by current CPU", true);
606 				return;
607 			}
608 			TAILQ_REMOVE(&curcpu()->ci_data.cpu_ld_locks, ld,
609 			    ld_chain);
610 		}
611 		ld->ld_flags &= ~LD_LOCKED;
612 		ld->ld_unlocked = where;
613 		ld->ld_lwp = NULL;
614 	}
615 	__cpu_simple_unlock(&ld->ld_spinlock);
616 	splx(s);
617 }
618 
619 /*
620  * lockdebug_wakeup:
621  *
622  *	Process a wakeup on a condition variable.
623  */
624 void
625 lockdebug_wakeup(const char *func, size_t line, volatile void *lock,
626     uintptr_t where)
627 {
628 	lockdebug_t *ld;
629 	int s;
630 
631 	if (__predict_false(panicstr != NULL || ld_panic || lock == (void *)&lbolt))
632 		return;
633 
634 	s = splhigh();
635 	/* Find the CV... */
636 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
637 		splx(s);
638 		return;
639 	}
640 	/*
641 	 * If it has any waiters, ensure that they are using the
642 	 * same interlock.
643 	 */
644 	if (__predict_false(ld->ld_shares != 0 &&
645 	    !mutex_owned((kmutex_t *)ld->ld_locked))) {
646 		lockdebug_abort1(func, line, ld, s, "interlocking mutex not "
647 		    "held during wakeup", true);
648 		return;
649 	}
650 	__cpu_simple_unlock(&ld->ld_spinlock);
651 	splx(s);
652 }
653 
654 /*
655  * lockdebug_barrier:
656  *
657  *	Panic if we hold more than one specified lock, and optionally, if we
658  *	hold any sleep locks.
659  */
660 void
661 lockdebug_barrier(const char *func, size_t line, volatile void *onelock,
662     int slplocks)
663 {
664 	struct lwp *l = curlwp;
665 	lockdebug_t *ld;
666 	int s;
667 
668 	if (__predict_false(panicstr != NULL || ld_panic))
669 		return;
670 
671 	s = splhigh();
672 	if ((l->l_pflag & LP_INTR) == 0) {
673 		TAILQ_FOREACH(ld, &curcpu()->ci_data.cpu_ld_locks, ld_chain) {
674 			if (ld->ld_lock == onelock) {
675 				continue;
676 			}
677 			__cpu_simple_lock(&ld->ld_spinlock);
678 			lockdebug_abort1(func, line, ld, s,
679 			    "spin lock held", true);
680 			return;
681 		}
682 	}
683 	if (slplocks) {
684 		splx(s);
685 		return;
686 	}
687 	ld = TAILQ_FIRST(&l->l_ld_locks);
688 	if (__predict_false(ld != NULL && ld->ld_lock != onelock)) {
689 		__cpu_simple_lock(&ld->ld_spinlock);
690 		lockdebug_abort1(func, line, ld, s, "sleep lock held", true);
691 		return;
692 	}
693 	splx(s);
694 	if (l->l_shlocks != 0) {
695 		TAILQ_FOREACH(ld, &ld_all, ld_achain) {
696 			if (ld->ld_lock == onelock) {
697 				continue;
698 			}
699 			if (ld->ld_lockops->lo_type == LOCKOPS_CV)
700 				continue;
701 			if (ld->ld_lwp == l)
702 				lockdebug_dump(l, ld, printf);
703 		}
704 		panic("%s,%zu: holding %d shared locks", func, line,
705 		    l->l_shlocks);
706 	}
707 }
708 
709 /*
710  * lockdebug_mem_check:
711  *
712  *	Check for in-use locks within a memory region that is
713  *	being freed.
714  */
715 void
716 lockdebug_mem_check(const char *func, size_t line, void *base, size_t sz)
717 {
718 	lockdebug_t *ld;
719 	struct cpu_info *ci;
720 	int s;
721 
722 	if (__predict_false(panicstr != NULL || ld_panic))
723 		return;
724 
725 	s = splhigh();
726 	ci = curcpu();
727 	__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
728 	ld = (lockdebug_t *)rb_tree_find_node_geq(&ld_rb_tree, base);
729 	if (ld != NULL) {
730 		const uintptr_t lock = (uintptr_t)ld->ld_lock;
731 
732 		if (__predict_false((uintptr_t)base > lock))
733 			panic("%s,%zu: corrupt tree ld=%p, base=%p, sz=%zu",
734 			    func, line, ld, base, sz);
735 		if (lock >= (uintptr_t)base + sz)
736 			ld = NULL;
737 	}
738 	__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
739 	if (__predict_false(ld != NULL)) {
740 		__cpu_simple_lock(&ld->ld_spinlock);
741 		lockdebug_abort1(func, line, ld, s,
742 		    "allocation contains active lock", !cold);
743 		return;
744 	}
745 	splx(s);
746 }
747 
748 /*
749  * lockdebug_dump:
750  *
751  *	Dump information about a lock on panic, or for DDB.
752  */
753 static void
754 lockdebug_dump(lwp_t *l, lockdebug_t *ld, void (*pr)(const char *, ...)
755     __printflike(1, 2))
756 {
757 	int sleeper = (ld->ld_flags & LD_SLEEPER);
758 
759 	(*pr)(
760 	    "lock address : %#018lx type     : %18s\n"
761 	    "initialized  : %#018lx",
762 	    (long)ld->ld_lock, (sleeper ? "sleep/adaptive" : "spin"),
763 	    (long)ld->ld_initaddr);
764 
765 	if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
766 		(*pr)(" interlock: %#018lx\n", (long)ld->ld_locked);
767 	} else {
768 		(*pr)("\n"
769 		    "shared holds : %18u exclusive: %18u\n"
770 		    "shares wanted: %18u exclusive: %18u\n"
771 		    "relevant cpu : %18u last held: %18u\n"
772 		    "relevant lwp : %#018lx last held: %#018lx\n"
773 		    "last locked%c : %#018lx unlocked%c: %#018lx\n",
774 		    (unsigned)ld->ld_shares, ((ld->ld_flags & LD_LOCKED) != 0),
775 		    (unsigned)ld->ld_shwant, (unsigned)ld->ld_exwant,
776 		    (unsigned)cpu_index(l->l_cpu), (unsigned)ld->ld_cpu,
777 		    (long)l, (long)ld->ld_lwp,
778 		    ((ld->ld_flags & LD_LOCKED) ? '*' : ' '),
779 		    (long)ld->ld_locked,
780 		    ((ld->ld_flags & LD_LOCKED) ? ' ' : '*'),
781 		    (long)ld->ld_unlocked);
782 	}
783 
784 	if (ld->ld_lockops->lo_dump != NULL)
785 		(*ld->ld_lockops->lo_dump)(ld->ld_lock, pr);
786 
787 	if (sleeper) {
788 		turnstile_print(ld->ld_lock, pr);
789 	}
790 }
791 
792 /*
793  * lockdebug_abort1:
794  *
795  *	An error has been trapped - dump lock info and panic.
796  */
797 static void
798 lockdebug_abort1(const char *func, size_t line, lockdebug_t *ld, int s,
799 		 const char *msg, bool dopanic)
800 {
801 
802 	/*
803 	 * Don't make the situation worse if the system is already going
804 	 * down in flames.  Once a panic is triggered, lockdebug state
805 	 * becomes stale and cannot be trusted.
806 	 */
807 	if (atomic_inc_uint_nv(&ld_panic) != 1) {
808 		__cpu_simple_unlock(&ld->ld_spinlock);
809 		splx(s);
810 		return;
811 	}
812 
813 	printf_nolog("%s error: %s,%zu: %s\n\n", ld->ld_lockops->lo_name,
814 	    func, line, msg);
815 	lockdebug_dump(curlwp, ld, printf_nolog);
816 	__cpu_simple_unlock(&ld->ld_spinlock);
817 	splx(s);
818 	printf_nolog("\n");
819 	if (dopanic)
820 		panic("LOCKDEBUG: %s error: %s,%zu: %s",
821 		    ld->ld_lockops->lo_name, func, line, msg);
822 }
823 
824 #endif	/* LOCKDEBUG */
825 
826 /*
827  * lockdebug_lock_print:
828  *
829  *	Handle the DDB 'show lock' command.
830  */
831 #ifdef DDB
832 #include <machine/db_machdep.h>
833 #include <ddb/db_interface.h>
834 
835 void
836 lockdebug_lock_print(void *addr,
837     void (*pr)(const char *, ...) __printflike(1, 2))
838 {
839 #ifdef LOCKDEBUG
840 	lockdebug_t *ld;
841 
842 	TAILQ_FOREACH(ld, &ld_all, ld_achain) {
843 		if (ld->ld_lock == NULL)
844 			continue;
845 		if (addr == NULL || ld->ld_lock == addr) {
846 			lockdebug_dump(curlwp, ld, pr);
847 			if (addr != NULL)
848 				return;
849 		}
850 	}
851 	if (addr != NULL) {
852 		(*pr)("Sorry, no record of a lock with address %p found.\n",
853 		    addr);
854 	}
855 #else
856 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
857 #endif	/* LOCKDEBUG */
858 }
859 
860 #ifdef LOCKDEBUG
861 static void
862 lockdebug_show_one(lwp_t *l, lockdebug_t *ld, int i,
863     void (*pr)(const char *, ...) __printflike(1, 2))
864 {
865 	const char *sym;
866 
867 	ksyms_getname(NULL, &sym, (vaddr_t)ld->ld_initaddr,
868 	    KSYMS_CLOSEST|KSYMS_PROC|KSYMS_ANY);
869 	(*pr)("* Lock %d (initialized at %s)\n", i++, sym);
870 	lockdebug_dump(l, ld, pr);
871 }
872 
873 static void
874 lockdebug_show_trace(const void *ptr,
875     void (*pr)(const char *, ...) __printflike(1, 2))
876 {
877     db_stack_trace_print((db_expr_t)(intptr_t)ptr, true, 32, "a", pr);
878 }
879 
880 static void
881 lockdebug_show_all_locks_lwp(void (*pr)(const char *, ...) __printflike(1, 2),
882     bool show_trace)
883 {
884 	struct proc *p;
885 
886 	LIST_FOREACH(p, &allproc, p_list) {
887 		struct lwp *l;
888 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
889 			lockdebug_t *ld;
890 			int i = 0;
891 			if (TAILQ_EMPTY(&l->l_ld_locks) &&
892 			    l->l_ld_wanted == NULL) {
893 			    	continue;
894 			}
895 			(*pr)("\n****** LWP %d.%d (%s) @ %p, l_stat=%d\n",
896 			    p->p_pid, l->l_lid,
897 			    l->l_name ? l->l_name : p->p_comm, l, l->l_stat);
898 			if (!TAILQ_EMPTY(&l->l_ld_locks)) {
899 				(*pr)("\n*** Locks held: \n");
900 				TAILQ_FOREACH(ld, &l->l_ld_locks, ld_chain) {
901 					(*pr)("\n");
902 					lockdebug_show_one(l, ld, i++, pr);
903 				}
904 			} else {
905 				(*pr)("\n*** Locks held: none\n");
906 			}
907 
908 			if (l->l_ld_wanted != NULL) {
909 				(*pr)("\n*** Locks wanted: \n\n");
910 				lockdebug_show_one(l, l->l_ld_wanted, 0, pr);
911 			} else {
912 				(*pr)("\n*** Locks wanted: none\n");
913 			}
914 			if (show_trace) {
915 				(*pr)("\n*** Traceback: \n\n");
916 				lockdebug_show_trace(l, pr);
917 				(*pr)("\n");
918 			}
919 		}
920 	}
921 }
922 
923 static void
924 lockdebug_show_all_locks_cpu(void (*pr)(const char *, ...) __printflike(1, 2),
925     bool show_trace)
926 {
927 	lockdebug_t *ld;
928 	CPU_INFO_ITERATOR cii;
929 	struct cpu_info *ci;
930 
931 	for (CPU_INFO_FOREACH(cii, ci)) {
932 		int i = 0;
933 		if (TAILQ_EMPTY(&ci->ci_data.cpu_ld_locks))
934 			continue;
935 		(*pr)("\n******* Locks held on %s:\n", cpu_name(ci));
936 		TAILQ_FOREACH(ld, &ci->ci_data.cpu_ld_locks, ld_chain) {
937 			(*pr)("\n");
938 #ifdef MULTIPROCESSOR
939 			lockdebug_show_one(ci->ci_curlwp, ld, i++, pr);
940 			if (show_trace)
941 				lockdebug_show_trace(ci->ci_curlwp, pr);
942 #else
943 			lockdebug_show_one(curlwp, ld, i++, pr);
944 			if (show_trace)
945 				lockdebug_show_trace(curlwp, pr);
946 #endif
947 		}
948 	}
949 }
950 #endif	/* LOCKDEBUG */
951 
952 void
953 lockdebug_show_all_locks(void (*pr)(const char *, ...) __printflike(1, 2),
954     const char *modif)
955 {
956 #ifdef LOCKDEBUG
957 	bool show_trace = false;
958 	if (modif[0] == 't')
959 		show_trace = true;
960 
961 	(*pr)("[Locks tracked through LWPs]\n");
962 	lockdebug_show_all_locks_lwp(pr, show_trace);
963 	(*pr)("\n");
964 
965 	(*pr)("[Locks tracked through CPUs]\n");
966 	lockdebug_show_all_locks_cpu(pr, show_trace);
967 	(*pr)("\n");
968 #else
969 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
970 #endif	/* LOCKDEBUG */
971 }
972 
973 void
974 lockdebug_show_lockstats(void (*pr)(const char *, ...) __printflike(1, 2))
975 {
976 #ifdef LOCKDEBUG
977 	lockdebug_t *ld;
978 	void *_ld;
979 	uint32_t n_null = 0;
980 	uint32_t n_spin_mutex = 0;
981 	uint32_t n_adaptive_mutex = 0;
982 	uint32_t n_rwlock = 0;
983 	uint32_t n_cv = 0;
984 	uint32_t n_others = 0;
985 
986 	RB_TREE_FOREACH(_ld, &ld_rb_tree) {
987 		ld = _ld;
988 		if (ld->ld_lock == NULL) {
989 			n_null++;
990 			continue;
991 		}
992 		if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
993 			n_cv++;
994 			continue;
995 		}
996 		if (ld->ld_lockops->lo_name[0] == 'M') {
997 			if (ld->ld_lockops->lo_type == LOCKOPS_SLEEP)
998 				n_adaptive_mutex++;
999 			else
1000 				n_spin_mutex++;
1001 			continue;
1002 		}
1003 		if (ld->ld_lockops->lo_name[0] == 'R') {
1004 			n_rwlock++;
1005 			continue;
1006 		}
1007 		n_others++;
1008 	}
1009 	(*pr)(
1010 	    "condvar: %u\n"
1011 	    "spin mutex: %u\n"
1012 	    "adaptive mutex: %u\n"
1013 	    "rwlock: %u\n"
1014 	    "null locks: %u\n"
1015 	    "others: %u\n",
1016 	    n_cv,  n_spin_mutex, n_adaptive_mutex, n_rwlock,
1017 	    n_null, n_others);
1018 #else
1019 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
1020 #endif	/* LOCKDEBUG */
1021 }
1022 #endif	/* DDB */
1023 
1024 /*
1025  * lockdebug_dismiss:
1026  *
1027  *      The system is rebooting, and potentially from an unsafe
1028  *      place so avoid any future aborts.
1029  */
1030 void
1031 lockdebug_dismiss(void)
1032 {
1033 
1034 	atomic_inc_uint_nv(&ld_panic);
1035 }
1036 
1037 /*
1038  * lockdebug_abort:
1039  *
1040  *	An error has been trapped - dump lock info and call panic().
1041  */
1042 void
1043 lockdebug_abort(const char *func, size_t line, const volatile void *lock,
1044     lockops_t *ops, const char *msg)
1045 {
1046 #ifdef LOCKDEBUG
1047 	lockdebug_t *ld;
1048 	int s;
1049 
1050 	s = splhigh();
1051 	if ((ld = lockdebug_lookup(func, line, lock,
1052 			(uintptr_t) __builtin_return_address(0))) != NULL) {
1053 		lockdebug_abort1(func, line, ld, s, msg, true);
1054 		return;
1055 	}
1056 	splx(s);
1057 #endif	/* LOCKDEBUG */
1058 
1059 	/*
1060 	 * Don't make the situation worse if the system is already going
1061 	 * down in flames.  Once a panic is triggered, lockdebug state
1062 	 * becomes stale and cannot be trusted.
1063 	 */
1064 	if (atomic_inc_uint_nv(&ld_panic) > 1)
1065 		return;
1066 
1067 	printf_nolog("%s error: %s,%zu: %s\n\n"
1068 	    "lock address : %#018lx\n"
1069 	    "current cpu  : %18d\n"
1070 	    "current lwp  : %#018lx\n",
1071 	    ops->lo_name, func, line, msg, (long)lock,
1072 	    (int)cpu_index(curcpu()), (long)curlwp);
1073 	(*ops->lo_dump)(lock, printf_nolog);
1074 	printf_nolog("\n");
1075 
1076 	panic("lock error: %s: %s,%zu: %s: lock %p cpu %d lwp %p",
1077 	    ops->lo_name, func, line, msg, lock, cpu_index(curcpu()), curlwp);
1078 }
1079