xref: /freebsd/sys/kern/kern_lock.c (revision aa0a1e58)
1 /*-
2  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice(s), this list of conditions and the following disclaimer as
10  *    the first lines of this file unmodified other than the possible
11  *    addition of one or more copyright notices.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice(s), this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23  * 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 SUCH
26  * DAMAGE.
27  */
28 
29 #include "opt_adaptive_lockmgrs.h"
30 #include "opt_ddb.h"
31 #include "opt_kdtrace.h"
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/ktr.h>
38 #include <sys/lock.h>
39 #include <sys/lock_profile.h>
40 #include <sys/lockmgr.h>
41 #include <sys/mutex.h>
42 #include <sys/proc.h>
43 #include <sys/sleepqueue.h>
44 #ifdef DEBUG_LOCKS
45 #include <sys/stack.h>
46 #endif
47 #include <sys/sysctl.h>
48 #include <sys/systm.h>
49 
50 #include <machine/cpu.h>
51 
52 #ifdef DDB
53 #include <ddb/ddb.h>
54 #endif
55 
56 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
57     (LK_ADAPTIVE | LK_NOSHARE));
58 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
59     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
60 
61 #define	SQ_EXCLUSIVE_QUEUE	0
62 #define	SQ_SHARED_QUEUE		1
63 
64 #ifndef INVARIANTS
65 #define	_lockmgr_assert(lk, what, file, line)
66 #define	TD_LOCKS_INC(td)
67 #define	TD_LOCKS_DEC(td)
68 #else
69 #define	TD_LOCKS_INC(td)	((td)->td_locks++)
70 #define	TD_LOCKS_DEC(td)	((td)->td_locks--)
71 #endif
72 #define	TD_SLOCKS_INC(td)	((td)->td_lk_slocks++)
73 #define	TD_SLOCKS_DEC(td)	((td)->td_lk_slocks--)
74 
75 #ifndef DEBUG_LOCKS
76 #define	STACK_PRINT(lk)
77 #define	STACK_SAVE(lk)
78 #define	STACK_ZERO(lk)
79 #else
80 #define	STACK_PRINT(lk)	stack_print_ddb(&(lk)->lk_stack)
81 #define	STACK_SAVE(lk)	stack_save(&(lk)->lk_stack)
82 #define	STACK_ZERO(lk)	stack_zero(&(lk)->lk_stack)
83 #endif
84 
85 #define	LOCK_LOG2(lk, string, arg1, arg2)				\
86 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
87 		CTR2(KTR_LOCK, (string), (arg1), (arg2))
88 #define	LOCK_LOG3(lk, string, arg1, arg2, arg3)				\
89 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
90 		CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
91 
92 #define	GIANT_DECLARE							\
93 	int _i = 0;							\
94 	WITNESS_SAVE_DECL(Giant)
95 #define	GIANT_RESTORE() do {						\
96 	if (_i > 0) {							\
97 		while (_i--)						\
98 			mtx_lock(&Giant);				\
99 		WITNESS_RESTORE(&Giant.lock_object, Giant);		\
100 	}								\
101 } while (0)
102 #define	GIANT_SAVE() do {						\
103 	if (mtx_owned(&Giant)) {					\
104 		WITNESS_SAVE(&Giant.lock_object, Giant);		\
105 		while (mtx_owned(&Giant)) {				\
106 			_i++;						\
107 			mtx_unlock(&Giant);				\
108 		}							\
109 	}								\
110 } while (0)
111 
112 #define	LK_CAN_SHARE(x)							\
113 	(((x) & LK_SHARE) && (((x) & LK_EXCLUSIVE_WAITERS) == 0 ||	\
114 	((x) & LK_EXCLUSIVE_SPINNERS) == 0 ||				\
115 	curthread->td_lk_slocks || (curthread->td_pflags & TDP_DEADLKTREAT)))
116 #define	LK_TRYOP(x)							\
117 	((x) & LK_NOWAIT)
118 
119 #define	LK_CAN_WITNESS(x)						\
120 	(((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
121 #define	LK_TRYWIT(x)							\
122 	(LK_TRYOP(x) ? LOP_TRYLOCK : 0)
123 
124 #define	LK_CAN_ADAPT(lk, f)						\
125 	(((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&		\
126 	((f) & LK_SLEEPFAIL) == 0)
127 
128 #define	lockmgr_disowned(lk)						\
129 	(((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
130 
131 #define	lockmgr_xlocked(lk)						\
132 	(((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
133 
134 static void	 assert_lockmgr(struct lock_object *lock, int how);
135 #ifdef DDB
136 static void	 db_show_lockmgr(struct lock_object *lock);
137 #endif
138 static void	 lock_lockmgr(struct lock_object *lock, int how);
139 #ifdef KDTRACE_HOOKS
140 static int	 owner_lockmgr(struct lock_object *lock, struct thread **owner);
141 #endif
142 static int	 unlock_lockmgr(struct lock_object *lock);
143 
144 struct lock_class lock_class_lockmgr = {
145 	.lc_name = "lockmgr",
146 	.lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
147 	.lc_assert = assert_lockmgr,
148 #ifdef DDB
149 	.lc_ddb_show = db_show_lockmgr,
150 #endif
151 	.lc_lock = lock_lockmgr,
152 	.lc_unlock = unlock_lockmgr,
153 #ifdef KDTRACE_HOOKS
154 	.lc_owner = owner_lockmgr,
155 #endif
156 };
157 
158 #ifdef ADAPTIVE_LOCKMGRS
159 static u_int alk_retries = 10;
160 static u_int alk_loops = 10000;
161 SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging");
162 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
163 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
164 #endif
165 
166 static __inline struct thread *
167 lockmgr_xholder(struct lock *lk)
168 {
169 	uintptr_t x;
170 
171 	x = lk->lk_lock;
172 	return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
173 }
174 
175 /*
176  * It assumes sleepq_lock held and returns with this one unheld.
177  * It also assumes the generic interlock is sane and previously checked.
178  * If LK_INTERLOCK is specified the interlock is not reacquired after the
179  * sleep.
180  */
181 static __inline int
182 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
183     const char *wmesg, int pri, int timo, int queue)
184 {
185 	GIANT_DECLARE;
186 	struct lock_class *class;
187 	int catch, error;
188 
189 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
190 	catch = pri & PCATCH;
191 	pri &= PRIMASK;
192 	error = 0;
193 
194 	LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
195 	    (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
196 
197 	if (flags & LK_INTERLOCK)
198 		class->lc_unlock(ilk);
199 	if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
200 		lk->lk_exslpfail++;
201 	GIANT_SAVE();
202 	sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
203 	    SLEEPQ_INTERRUPTIBLE : 0), queue);
204 	if ((flags & LK_TIMELOCK) && timo)
205 		sleepq_set_timeout(&lk->lock_object, timo);
206 
207 	/*
208 	 * Decisional switch for real sleeping.
209 	 */
210 	if ((flags & LK_TIMELOCK) && timo && catch)
211 		error = sleepq_timedwait_sig(&lk->lock_object, pri);
212 	else if ((flags & LK_TIMELOCK) && timo)
213 		error = sleepq_timedwait(&lk->lock_object, pri);
214 	else if (catch)
215 		error = sleepq_wait_sig(&lk->lock_object, pri);
216 	else
217 		sleepq_wait(&lk->lock_object, pri);
218 	GIANT_RESTORE();
219 	if ((flags & LK_SLEEPFAIL) && error == 0)
220 		error = ENOLCK;
221 
222 	return (error);
223 }
224 
225 static __inline int
226 wakeupshlk(struct lock *lk, const char *file, int line)
227 {
228 	uintptr_t v, x;
229 	u_int realexslp;
230 	int queue, wakeup_swapper;
231 
232 	TD_LOCKS_DEC(curthread);
233 	TD_SLOCKS_DEC(curthread);
234 	WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
235 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
236 
237 	wakeup_swapper = 0;
238 	for (;;) {
239 		x = lk->lk_lock;
240 
241 		/*
242 		 * If there is more than one shared lock held, just drop one
243 		 * and return.
244 		 */
245 		if (LK_SHARERS(x) > 1) {
246 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
247 			    x - LK_ONE_SHARER))
248 				break;
249 			continue;
250 		}
251 
252 		/*
253 		 * If there are not waiters on the exclusive queue, drop the
254 		 * lock quickly.
255 		 */
256 		if ((x & LK_ALL_WAITERS) == 0) {
257 			MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
258 			    LK_SHARERS_LOCK(1));
259 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
260 				break;
261 			continue;
262 		}
263 
264 		/*
265 		 * We should have a sharer with waiters, so enter the hard
266 		 * path in order to handle wakeups correctly.
267 		 */
268 		sleepq_lock(&lk->lock_object);
269 		x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
270 		v = LK_UNLOCKED;
271 
272 		/*
273 		 * If the lock has exclusive waiters, give them preference in
274 		 * order to avoid deadlock with shared runners up.
275 		 * If interruptible sleeps left the exclusive queue empty
276 		 * avoid a starvation for the threads sleeping on the shared
277 		 * queue by giving them precedence and cleaning up the
278 		 * exclusive waiters bit anyway.
279 		 * Please note that lk_exslpfail count may be lying about
280 		 * the real number of waiters with the LK_SLEEPFAIL flag on
281 		 * because they may be used in conjuction with interruptible
282 		 * sleeps so lk_exslpfail might be considered an 'upper limit'
283 		 * bound, including the edge cases.
284 		 */
285 		realexslp = sleepq_sleepcnt(&lk->lock_object,
286 		    SQ_EXCLUSIVE_QUEUE);
287 		if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
288 			if (lk->lk_exslpfail < realexslp) {
289 				lk->lk_exslpfail = 0;
290 				queue = SQ_EXCLUSIVE_QUEUE;
291 				v |= (x & LK_SHARED_WAITERS);
292 			} else {
293 				lk->lk_exslpfail = 0;
294 				LOCK_LOG2(lk,
295 				    "%s: %p has only LK_SLEEPFAIL sleepers",
296 				    __func__, lk);
297 				LOCK_LOG2(lk,
298 			    "%s: %p waking up threads on the exclusive queue",
299 				    __func__, lk);
300 				wakeup_swapper =
301 				    sleepq_broadcast(&lk->lock_object,
302 				    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
303 				queue = SQ_SHARED_QUEUE;
304 			}
305 
306 		} else {
307 
308 			/*
309 			 * Exclusive waiters sleeping with LK_SLEEPFAIL on
310 			 * and using interruptible sleeps/timeout may have
311 			 * left spourious lk_exslpfail counts on, so clean
312 			 * it up anyway.
313 			 */
314 			lk->lk_exslpfail = 0;
315 			queue = SQ_SHARED_QUEUE;
316 		}
317 
318 		if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
319 		    v)) {
320 			sleepq_release(&lk->lock_object);
321 			continue;
322 		}
323 		LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
324 		    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
325 		    "exclusive");
326 		wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
327 		    0, queue);
328 		sleepq_release(&lk->lock_object);
329 		break;
330 	}
331 
332 	lock_profile_release_lock(&lk->lock_object);
333 	return (wakeup_swapper);
334 }
335 
336 static void
337 assert_lockmgr(struct lock_object *lock, int what)
338 {
339 
340 	panic("lockmgr locks do not support assertions");
341 }
342 
343 static void
344 lock_lockmgr(struct lock_object *lock, int how)
345 {
346 
347 	panic("lockmgr locks do not support sleep interlocking");
348 }
349 
350 static int
351 unlock_lockmgr(struct lock_object *lock)
352 {
353 
354 	panic("lockmgr locks do not support sleep interlocking");
355 }
356 
357 #ifdef KDTRACE_HOOKS
358 static int
359 owner_lockmgr(struct lock_object *lock, struct thread **owner)
360 {
361 
362 	panic("lockmgr locks do not support owner inquiring");
363 }
364 #endif
365 
366 void
367 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
368 {
369 	int iflags;
370 
371 	MPASS((flags & ~LK_INIT_MASK) == 0);
372 	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
373             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
374             &lk->lk_lock));
375 
376 	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
377 	if (flags & LK_CANRECURSE)
378 		iflags |= LO_RECURSABLE;
379 	if ((flags & LK_NODUP) == 0)
380 		iflags |= LO_DUPOK;
381 	if (flags & LK_NOPROFILE)
382 		iflags |= LO_NOPROFILE;
383 	if ((flags & LK_NOWITNESS) == 0)
384 		iflags |= LO_WITNESS;
385 	if (flags & LK_QUIET)
386 		iflags |= LO_QUIET;
387 	iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
388 
389 	lk->lk_lock = LK_UNLOCKED;
390 	lk->lk_recurse = 0;
391 	lk->lk_exslpfail = 0;
392 	lk->lk_timo = timo;
393 	lk->lk_pri = pri;
394 	lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
395 	STACK_ZERO(lk);
396 }
397 
398 /*
399  * XXX: Gross hacks to manipulate external lock flags after
400  * initialization.  Used for certain vnode and buf locks.
401  */
402 void
403 lockallowshare(struct lock *lk)
404 {
405 
406 	lockmgr_assert(lk, KA_XLOCKED);
407 	lk->lock_object.lo_flags &= ~LK_NOSHARE;
408 }
409 
410 void
411 lockallowrecurse(struct lock *lk)
412 {
413 
414 	lockmgr_assert(lk, KA_XLOCKED);
415 	lk->lock_object.lo_flags |= LO_RECURSABLE;
416 }
417 
418 void
419 lockdisablerecurse(struct lock *lk)
420 {
421 
422 	lockmgr_assert(lk, KA_XLOCKED);
423 	lk->lock_object.lo_flags &= ~LO_RECURSABLE;
424 }
425 
426 void
427 lockdestroy(struct lock *lk)
428 {
429 
430 	KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
431 	KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
432 	KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
433 	lock_destroy(&lk->lock_object);
434 }
435 
436 int
437 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
438     const char *wmesg, int pri, int timo, const char *file, int line)
439 {
440 	GIANT_DECLARE;
441 	struct lock_class *class;
442 	const char *iwmesg;
443 	uintptr_t tid, v, x;
444 	u_int op, realexslp;
445 	int error, ipri, itimo, queue, wakeup_swapper;
446 #ifdef LOCK_PROFILING
447 	uint64_t waittime = 0;
448 	int contested = 0;
449 #endif
450 #ifdef ADAPTIVE_LOCKMGRS
451 	volatile struct thread *owner;
452 	u_int i, spintries = 0;
453 #endif
454 
455 	error = 0;
456 	tid = (uintptr_t)curthread;
457 	op = (flags & LK_TYPE_MASK);
458 	iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
459 	ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
460 	itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
461 
462 	MPASS((flags & ~LK_TOTAL_MASK) == 0);
463 	KASSERT((op & (op - 1)) == 0,
464 	    ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
465 	KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
466 	    (op != LK_DOWNGRADE && op != LK_RELEASE),
467 	    ("%s: Invalid flags in regard of the operation desired @ %s:%d",
468 	    __func__, file, line));
469 	KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
470 	    ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
471 	    __func__, file, line));
472 
473 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
474 	if (panicstr != NULL) {
475 		if (flags & LK_INTERLOCK)
476 			class->lc_unlock(ilk);
477 		return (0);
478 	}
479 
480 	if (op == LK_SHARED && (lk->lock_object.lo_flags & LK_NOSHARE))
481 		op = LK_EXCLUSIVE;
482 
483 	wakeup_swapper = 0;
484 	switch (op) {
485 	case LK_SHARED:
486 		if (LK_CAN_WITNESS(flags))
487 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
488 			    file, line, ilk);
489 		for (;;) {
490 			x = lk->lk_lock;
491 
492 			/*
493 			 * If no other thread has an exclusive lock, or
494 			 * no exclusive waiter is present, bump the count of
495 			 * sharers.  Since we have to preserve the state of
496 			 * waiters, if we fail to acquire the shared lock
497 			 * loop back and retry.
498 			 */
499 			if (LK_CAN_SHARE(x)) {
500 				if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
501 				    x + LK_ONE_SHARER))
502 					break;
503 				continue;
504 			}
505 			lock_profile_obtain_lock_failed(&lk->lock_object,
506 			    &contested, &waittime);
507 
508 			/*
509 			 * If the lock is already held by curthread in
510 			 * exclusive way avoid a deadlock.
511 			 */
512 			if (LK_HOLDER(x) == tid) {
513 				LOCK_LOG2(lk,
514 				    "%s: %p already held in exclusive mode",
515 				    __func__, lk);
516 				error = EDEADLK;
517 				break;
518 			}
519 
520 			/*
521 			 * If the lock is expected to not sleep just give up
522 			 * and return.
523 			 */
524 			if (LK_TRYOP(flags)) {
525 				LOCK_LOG2(lk, "%s: %p fails the try operation",
526 				    __func__, lk);
527 				error = EBUSY;
528 				break;
529 			}
530 
531 #ifdef ADAPTIVE_LOCKMGRS
532 			/*
533 			 * If the owner is running on another CPU, spin until
534 			 * the owner stops running or the state of the lock
535 			 * changes.  We need a double-state handle here
536 			 * because for a failed acquisition the lock can be
537 			 * either held in exclusive mode or shared mode
538 			 * (for the writer starvation avoidance technique).
539 			 */
540 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
541 			    LK_HOLDER(x) != LK_KERNPROC) {
542 				owner = (struct thread *)LK_HOLDER(x);
543 				if (LOCK_LOG_TEST(&lk->lock_object, 0))
544 					CTR3(KTR_LOCK,
545 					    "%s: spinning on %p held by %p",
546 					    __func__, lk, owner);
547 
548 				/*
549 				 * If we are holding also an interlock drop it
550 				 * in order to avoid a deadlock if the lockmgr
551 				 * owner is adaptively spinning on the
552 				 * interlock itself.
553 				 */
554 				if (flags & LK_INTERLOCK) {
555 					class->lc_unlock(ilk);
556 					flags &= ~LK_INTERLOCK;
557 				}
558 				GIANT_SAVE();
559 				while (LK_HOLDER(lk->lk_lock) ==
560 				    (uintptr_t)owner && TD_IS_RUNNING(owner))
561 					cpu_spinwait();
562 				GIANT_RESTORE();
563 				continue;
564 			} else if (LK_CAN_ADAPT(lk, flags) &&
565 			    (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
566 			    spintries < alk_retries) {
567 				if (flags & LK_INTERLOCK) {
568 					class->lc_unlock(ilk);
569 					flags &= ~LK_INTERLOCK;
570 				}
571 				GIANT_SAVE();
572 				spintries++;
573 				for (i = 0; i < alk_loops; i++) {
574 					if (LOCK_LOG_TEST(&lk->lock_object, 0))
575 						CTR4(KTR_LOCK,
576 				    "%s: shared spinning on %p with %u and %u",
577 						    __func__, lk, spintries, i);
578 					x = lk->lk_lock;
579 					if ((x & LK_SHARE) == 0 ||
580 					    LK_CAN_SHARE(x) != 0)
581 						break;
582 					cpu_spinwait();
583 				}
584 				GIANT_RESTORE();
585 				if (i != alk_loops)
586 					continue;
587 			}
588 #endif
589 
590 			/*
591 			 * Acquire the sleepqueue chain lock because we
592 			 * probabilly will need to manipulate waiters flags.
593 			 */
594 			sleepq_lock(&lk->lock_object);
595 			x = lk->lk_lock;
596 
597 			/*
598 			 * if the lock can be acquired in shared mode, try
599 			 * again.
600 			 */
601 			if (LK_CAN_SHARE(x)) {
602 				sleepq_release(&lk->lock_object);
603 				continue;
604 			}
605 
606 #ifdef ADAPTIVE_LOCKMGRS
607 			/*
608 			 * The current lock owner might have started executing
609 			 * on another CPU (or the lock could have changed
610 			 * owner) while we were waiting on the turnstile
611 			 * chain lock.  If so, drop the turnstile lock and try
612 			 * again.
613 			 */
614 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
615 			    LK_HOLDER(x) != LK_KERNPROC) {
616 				owner = (struct thread *)LK_HOLDER(x);
617 				if (TD_IS_RUNNING(owner)) {
618 					sleepq_release(&lk->lock_object);
619 					continue;
620 				}
621 			}
622 #endif
623 
624 			/*
625 			 * Try to set the LK_SHARED_WAITERS flag.  If we fail,
626 			 * loop back and retry.
627 			 */
628 			if ((x & LK_SHARED_WAITERS) == 0) {
629 				if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
630 				    x | LK_SHARED_WAITERS)) {
631 					sleepq_release(&lk->lock_object);
632 					continue;
633 				}
634 				LOCK_LOG2(lk, "%s: %p set shared waiters flag",
635 				    __func__, lk);
636 			}
637 
638 			/*
639 			 * As far as we have been unable to acquire the
640 			 * shared lock and the shared waiters flag is set,
641 			 * we will sleep.
642 			 */
643 			error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
644 			    SQ_SHARED_QUEUE);
645 			flags &= ~LK_INTERLOCK;
646 			if (error) {
647 				LOCK_LOG3(lk,
648 				    "%s: interrupted sleep for %p with %d",
649 				    __func__, lk, error);
650 				break;
651 			}
652 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
653 			    __func__, lk);
654 		}
655 		if (error == 0) {
656 			lock_profile_obtain_lock_success(&lk->lock_object,
657 			    contested, waittime, file, line);
658 			LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
659 			    line);
660 			WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
661 			    line);
662 			TD_LOCKS_INC(curthread);
663 			TD_SLOCKS_INC(curthread);
664 			STACK_SAVE(lk);
665 		}
666 		break;
667 	case LK_UPGRADE:
668 		_lockmgr_assert(lk, KA_SLOCKED, file, line);
669 		v = lk->lk_lock;
670 		x = v & LK_ALL_WAITERS;
671 		v &= LK_EXCLUSIVE_SPINNERS;
672 
673 		/*
674 		 * Try to switch from one shared lock to an exclusive one.
675 		 * We need to preserve waiters flags during the operation.
676 		 */
677 		if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
678 		    tid | x)) {
679 			LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
680 			    line);
681 			WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
682 			    LK_TRYWIT(flags), file, line);
683 			TD_SLOCKS_DEC(curthread);
684 			break;
685 		}
686 
687 		/*
688 		 * We have been unable to succeed in upgrading, so just
689 		 * give up the shared lock.
690 		 */
691 		wakeup_swapper |= wakeupshlk(lk, file, line);
692 
693 		/* FALLTHROUGH */
694 	case LK_EXCLUSIVE:
695 		if (LK_CAN_WITNESS(flags))
696 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
697 			    LOP_EXCLUSIVE, file, line, ilk);
698 
699 		/*
700 		 * If curthread already holds the lock and this one is
701 		 * allowed to recurse, simply recurse on it.
702 		 */
703 		if (lockmgr_xlocked(lk)) {
704 			if ((flags & LK_CANRECURSE) == 0 &&
705 			    (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
706 
707 				/*
708 				 * If the lock is expected to not panic just
709 				 * give up and return.
710 				 */
711 				if (LK_TRYOP(flags)) {
712 					LOCK_LOG2(lk,
713 					    "%s: %p fails the try operation",
714 					    __func__, lk);
715 					error = EBUSY;
716 					break;
717 				}
718 				if (flags & LK_INTERLOCK)
719 					class->lc_unlock(ilk);
720 		panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
721 				    __func__, iwmesg, file, line);
722 			}
723 			lk->lk_recurse++;
724 			LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
725 			LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
726 			    lk->lk_recurse, file, line);
727 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
728 			    LK_TRYWIT(flags), file, line);
729 			TD_LOCKS_INC(curthread);
730 			break;
731 		}
732 
733 		while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
734 		    tid)) {
735 			lock_profile_obtain_lock_failed(&lk->lock_object,
736 			    &contested, &waittime);
737 
738 			/*
739 			 * If the lock is expected to not sleep just give up
740 			 * and return.
741 			 */
742 			if (LK_TRYOP(flags)) {
743 				LOCK_LOG2(lk, "%s: %p fails the try operation",
744 				    __func__, lk);
745 				error = EBUSY;
746 				break;
747 			}
748 
749 #ifdef ADAPTIVE_LOCKMGRS
750 			/*
751 			 * If the owner is running on another CPU, spin until
752 			 * the owner stops running or the state of the lock
753 			 * changes.
754 			 */
755 			x = lk->lk_lock;
756 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
757 			    LK_HOLDER(x) != LK_KERNPROC) {
758 				owner = (struct thread *)LK_HOLDER(x);
759 				if (LOCK_LOG_TEST(&lk->lock_object, 0))
760 					CTR3(KTR_LOCK,
761 					    "%s: spinning on %p held by %p",
762 					    __func__, lk, owner);
763 
764 				/*
765 				 * If we are holding also an interlock drop it
766 				 * in order to avoid a deadlock if the lockmgr
767 				 * owner is adaptively spinning on the
768 				 * interlock itself.
769 				 */
770 				if (flags & LK_INTERLOCK) {
771 					class->lc_unlock(ilk);
772 					flags &= ~LK_INTERLOCK;
773 				}
774 				GIANT_SAVE();
775 				while (LK_HOLDER(lk->lk_lock) ==
776 				    (uintptr_t)owner && TD_IS_RUNNING(owner))
777 					cpu_spinwait();
778 				GIANT_RESTORE();
779 				continue;
780 			} else if (LK_CAN_ADAPT(lk, flags) &&
781 			    (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
782 			    spintries < alk_retries) {
783 				if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
784 				    !atomic_cmpset_ptr(&lk->lk_lock, x,
785 				    x | LK_EXCLUSIVE_SPINNERS))
786 					continue;
787 				if (flags & LK_INTERLOCK) {
788 					class->lc_unlock(ilk);
789 					flags &= ~LK_INTERLOCK;
790 				}
791 				GIANT_SAVE();
792 				spintries++;
793 				for (i = 0; i < alk_loops; i++) {
794 					if (LOCK_LOG_TEST(&lk->lock_object, 0))
795 						CTR4(KTR_LOCK,
796 				    "%s: shared spinning on %p with %u and %u",
797 						    __func__, lk, spintries, i);
798 					if ((lk->lk_lock &
799 					    LK_EXCLUSIVE_SPINNERS) == 0)
800 						break;
801 					cpu_spinwait();
802 				}
803 				GIANT_RESTORE();
804 				if (i != alk_loops)
805 					continue;
806 			}
807 #endif
808 
809 			/*
810 			 * Acquire the sleepqueue chain lock because we
811 			 * probabilly will need to manipulate waiters flags.
812 			 */
813 			sleepq_lock(&lk->lock_object);
814 			x = lk->lk_lock;
815 
816 			/*
817 			 * if the lock has been released while we spun on
818 			 * the sleepqueue chain lock just try again.
819 			 */
820 			if (x == LK_UNLOCKED) {
821 				sleepq_release(&lk->lock_object);
822 				continue;
823 			}
824 
825 #ifdef ADAPTIVE_LOCKMGRS
826 			/*
827 			 * The current lock owner might have started executing
828 			 * on another CPU (or the lock could have changed
829 			 * owner) while we were waiting on the turnstile
830 			 * chain lock.  If so, drop the turnstile lock and try
831 			 * again.
832 			 */
833 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
834 			    LK_HOLDER(x) != LK_KERNPROC) {
835 				owner = (struct thread *)LK_HOLDER(x);
836 				if (TD_IS_RUNNING(owner)) {
837 					sleepq_release(&lk->lock_object);
838 					continue;
839 				}
840 			}
841 #endif
842 
843 			/*
844 			 * The lock can be in the state where there is a
845 			 * pending queue of waiters, but still no owner.
846 			 * This happens when the lock is contested and an
847 			 * owner is going to claim the lock.
848 			 * If curthread is the one successfully acquiring it
849 			 * claim lock ownership and return, preserving waiters
850 			 * flags.
851 			 */
852 			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
853 			if ((x & ~v) == LK_UNLOCKED) {
854 				v &= ~LK_EXCLUSIVE_SPINNERS;
855 				if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
856 				    tid | v)) {
857 					sleepq_release(&lk->lock_object);
858 					LOCK_LOG2(lk,
859 					    "%s: %p claimed by a new writer",
860 					    __func__, lk);
861 					break;
862 				}
863 				sleepq_release(&lk->lock_object);
864 				continue;
865 			}
866 
867 			/*
868 			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
869 			 * fail, loop back and retry.
870 			 */
871 			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
872 				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
873 				    x | LK_EXCLUSIVE_WAITERS)) {
874 					sleepq_release(&lk->lock_object);
875 					continue;
876 				}
877 				LOCK_LOG2(lk, "%s: %p set excl waiters flag",
878 				    __func__, lk);
879 			}
880 
881 			/*
882 			 * As far as we have been unable to acquire the
883 			 * exclusive lock and the exclusive waiters flag
884 			 * is set, we will sleep.
885 			 */
886 			error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
887 			    SQ_EXCLUSIVE_QUEUE);
888 			flags &= ~LK_INTERLOCK;
889 			if (error) {
890 				LOCK_LOG3(lk,
891 				    "%s: interrupted sleep for %p with %d",
892 				    __func__, lk, error);
893 				break;
894 			}
895 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
896 			    __func__, lk);
897 		}
898 		if (error == 0) {
899 			lock_profile_obtain_lock_success(&lk->lock_object,
900 			    contested, waittime, file, line);
901 			LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
902 			    lk->lk_recurse, file, line);
903 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
904 			    LK_TRYWIT(flags), file, line);
905 			TD_LOCKS_INC(curthread);
906 			STACK_SAVE(lk);
907 		}
908 		break;
909 	case LK_DOWNGRADE:
910 		_lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line);
911 		LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
912 		WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
913 		TD_SLOCKS_INC(curthread);
914 
915 		/*
916 		 * In order to preserve waiters flags, just spin.
917 		 */
918 		for (;;) {
919 			x = lk->lk_lock;
920 			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
921 			x &= LK_ALL_WAITERS;
922 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
923 			    LK_SHARERS_LOCK(1) | x))
924 				break;
925 			cpu_spinwait();
926 		}
927 		break;
928 	case LK_RELEASE:
929 		_lockmgr_assert(lk, KA_LOCKED, file, line);
930 		x = lk->lk_lock;
931 
932 		if ((x & LK_SHARE) == 0) {
933 
934 			/*
935 			 * As first option, treact the lock as if it has not
936 			 * any waiter.
937 			 * Fix-up the tid var if the lock has been disowned.
938 			 */
939 			if (LK_HOLDER(x) == LK_KERNPROC)
940 				tid = LK_KERNPROC;
941 			else {
942 				WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
943 				    file, line);
944 				TD_LOCKS_DEC(curthread);
945 			}
946 			LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
947 			    lk->lk_recurse, file, line);
948 
949 			/*
950 			 * The lock is held in exclusive mode.
951 			 * If the lock is recursed also, then unrecurse it.
952 			 */
953 			if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
954 				LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
955 				    lk);
956 				lk->lk_recurse--;
957 				break;
958 			}
959 			if (tid != LK_KERNPROC)
960 				lock_profile_release_lock(&lk->lock_object);
961 
962 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
963 			    LK_UNLOCKED))
964 				break;
965 
966 			sleepq_lock(&lk->lock_object);
967 			x = lk->lk_lock;
968 			v = LK_UNLOCKED;
969 
970 			/*
971 		 	 * If the lock has exclusive waiters, give them
972 			 * preference in order to avoid deadlock with
973 			 * shared runners up.
974 			 * If interruptible sleeps left the exclusive queue
975 			 * empty avoid a starvation for the threads sleeping
976 			 * on the shared queue by giving them precedence
977 			 * and cleaning up the exclusive waiters bit anyway.
978 			 * Please note that lk_exslpfail count may be lying
979 			 * about the real number of waiters with the
980 			 * LK_SLEEPFAIL flag on because they may be used in
981 			 * conjuction with interruptible sleeps so
982 			 * lk_exslpfail might be considered an 'upper limit'
983 			 * bound, including the edge cases.
984 			 */
985 			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
986 			realexslp = sleepq_sleepcnt(&lk->lock_object,
987 			    SQ_EXCLUSIVE_QUEUE);
988 			if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
989 				if (lk->lk_exslpfail < realexslp) {
990 					lk->lk_exslpfail = 0;
991 					queue = SQ_EXCLUSIVE_QUEUE;
992 					v |= (x & LK_SHARED_WAITERS);
993 				} else {
994 					lk->lk_exslpfail = 0;
995 					LOCK_LOG2(lk,
996 					"%s: %p has only LK_SLEEPFAIL sleepers",
997 					    __func__, lk);
998 					LOCK_LOG2(lk,
999 			"%s: %p waking up threads on the exclusive queue",
1000 					    __func__, lk);
1001 					wakeup_swapper =
1002 					    sleepq_broadcast(&lk->lock_object,
1003 					    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1004 					queue = SQ_SHARED_QUEUE;
1005 				}
1006 			} else {
1007 
1008 				/*
1009 				 * Exclusive waiters sleeping with LK_SLEEPFAIL
1010 				 * on and using interruptible sleeps/timeout
1011 				 * may have left spourious lk_exslpfail counts
1012 				 * on, so clean it up anyway.
1013 				 */
1014 				lk->lk_exslpfail = 0;
1015 				queue = SQ_SHARED_QUEUE;
1016 			}
1017 
1018 			LOCK_LOG3(lk,
1019 			    "%s: %p waking up threads on the %s queue",
1020 			    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1021 			    "exclusive");
1022 			atomic_store_rel_ptr(&lk->lk_lock, v);
1023 			wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1024 			    SLEEPQ_LK, 0, queue);
1025 			sleepq_release(&lk->lock_object);
1026 			break;
1027 		} else
1028 			wakeup_swapper = wakeupshlk(lk, file, line);
1029 		break;
1030 	case LK_DRAIN:
1031 		if (LK_CAN_WITNESS(flags))
1032 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1033 			    LOP_EXCLUSIVE, file, line, ilk);
1034 
1035 		/*
1036 		 * Trying to drain a lock we already own will result in a
1037 		 * deadlock.
1038 		 */
1039 		if (lockmgr_xlocked(lk)) {
1040 			if (flags & LK_INTERLOCK)
1041 				class->lc_unlock(ilk);
1042 			panic("%s: draining %s with the lock held @ %s:%d\n",
1043 			    __func__, iwmesg, file, line);
1044 		}
1045 
1046 		while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1047 			lock_profile_obtain_lock_failed(&lk->lock_object,
1048 			    &contested, &waittime);
1049 
1050 			/*
1051 			 * If the lock is expected to not sleep just give up
1052 			 * and return.
1053 			 */
1054 			if (LK_TRYOP(flags)) {
1055 				LOCK_LOG2(lk, "%s: %p fails the try operation",
1056 				    __func__, lk);
1057 				error = EBUSY;
1058 				break;
1059 			}
1060 
1061 			/*
1062 			 * Acquire the sleepqueue chain lock because we
1063 			 * probabilly will need to manipulate waiters flags.
1064 			 */
1065 			sleepq_lock(&lk->lock_object);
1066 			x = lk->lk_lock;
1067 
1068 			/*
1069 			 * if the lock has been released while we spun on
1070 			 * the sleepqueue chain lock just try again.
1071 			 */
1072 			if (x == LK_UNLOCKED) {
1073 				sleepq_release(&lk->lock_object);
1074 				continue;
1075 			}
1076 
1077 			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1078 			if ((x & ~v) == LK_UNLOCKED) {
1079 				v = (x & ~LK_EXCLUSIVE_SPINNERS);
1080 
1081 				/*
1082 				 * If interruptible sleeps left the exclusive
1083 				 * queue empty avoid a starvation for the
1084 				 * threads sleeping on the shared queue by
1085 				 * giving them precedence and cleaning up the
1086 				 * exclusive waiters bit anyway.
1087 				 * Please note that lk_exslpfail count may be
1088 				 * lying about the real number of waiters with
1089 				 * the LK_SLEEPFAIL flag on because they may
1090 				 * be used in conjuction with interruptible
1091 				 * sleeps so lk_exslpfail might be considered
1092 				 * an 'upper limit' bound, including the edge
1093 				 * cases.
1094 				 */
1095 				if (v & LK_EXCLUSIVE_WAITERS) {
1096 					queue = SQ_EXCLUSIVE_QUEUE;
1097 					v &= ~LK_EXCLUSIVE_WAITERS;
1098 				} else {
1099 
1100 					/*
1101 					 * Exclusive waiters sleeping with
1102 					 * LK_SLEEPFAIL on and using
1103 					 * interruptible sleeps/timeout may
1104 					 * have left spourious lk_exslpfail
1105 					 * counts on, so clean it up anyway.
1106 					 */
1107 					MPASS(v & LK_SHARED_WAITERS);
1108 					lk->lk_exslpfail = 0;
1109 					queue = SQ_SHARED_QUEUE;
1110 					v &= ~LK_SHARED_WAITERS;
1111 				}
1112 				if (queue == SQ_EXCLUSIVE_QUEUE) {
1113 					realexslp =
1114 					    sleepq_sleepcnt(&lk->lock_object,
1115 					    SQ_EXCLUSIVE_QUEUE);
1116 					if (lk->lk_exslpfail >= realexslp) {
1117 						lk->lk_exslpfail = 0;
1118 						queue = SQ_SHARED_QUEUE;
1119 						v &= ~LK_SHARED_WAITERS;
1120 						if (realexslp != 0) {
1121 							LOCK_LOG2(lk,
1122 					"%s: %p has only LK_SLEEPFAIL sleepers",
1123 							    __func__, lk);
1124 							LOCK_LOG2(lk,
1125 			"%s: %p waking up threads on the exclusive queue",
1126 							    __func__, lk);
1127 							wakeup_swapper =
1128 							    sleepq_broadcast(
1129 							    &lk->lock_object,
1130 							    SLEEPQ_LK, 0,
1131 							    SQ_EXCLUSIVE_QUEUE);
1132 						}
1133 					} else
1134 						lk->lk_exslpfail = 0;
1135 				}
1136 				if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1137 					sleepq_release(&lk->lock_object);
1138 					continue;
1139 				}
1140 				LOCK_LOG3(lk,
1141 				"%s: %p waking up all threads on the %s queue",
1142 				    __func__, lk, queue == SQ_SHARED_QUEUE ?
1143 				    "shared" : "exclusive");
1144 				wakeup_swapper |= sleepq_broadcast(
1145 				    &lk->lock_object, SLEEPQ_LK, 0, queue);
1146 
1147 				/*
1148 				 * If shared waiters have been woken up we need
1149 				 * to wait for one of them to acquire the lock
1150 				 * before to set the exclusive waiters in
1151 				 * order to avoid a deadlock.
1152 				 */
1153 				if (queue == SQ_SHARED_QUEUE) {
1154 					for (v = lk->lk_lock;
1155 					    (v & LK_SHARE) && !LK_SHARERS(v);
1156 					    v = lk->lk_lock)
1157 						cpu_spinwait();
1158 				}
1159 			}
1160 
1161 			/*
1162 			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1163 			 * fail, loop back and retry.
1164 			 */
1165 			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1166 				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1167 				    x | LK_EXCLUSIVE_WAITERS)) {
1168 					sleepq_release(&lk->lock_object);
1169 					continue;
1170 				}
1171 				LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1172 				    __func__, lk);
1173 			}
1174 
1175 			/*
1176 			 * As far as we have been unable to acquire the
1177 			 * exclusive lock and the exclusive waiters flag
1178 			 * is set, we will sleep.
1179 			 */
1180 			if (flags & LK_INTERLOCK) {
1181 				class->lc_unlock(ilk);
1182 				flags &= ~LK_INTERLOCK;
1183 			}
1184 			GIANT_SAVE();
1185 			sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1186 			    SQ_EXCLUSIVE_QUEUE);
1187 			sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1188 			GIANT_RESTORE();
1189 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1190 			    __func__, lk);
1191 		}
1192 
1193 		if (error == 0) {
1194 			lock_profile_obtain_lock_success(&lk->lock_object,
1195 			    contested, waittime, file, line);
1196 			LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1197 			    lk->lk_recurse, file, line);
1198 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1199 			    LK_TRYWIT(flags), file, line);
1200 			TD_LOCKS_INC(curthread);
1201 			STACK_SAVE(lk);
1202 		}
1203 		break;
1204 	default:
1205 		if (flags & LK_INTERLOCK)
1206 			class->lc_unlock(ilk);
1207 		panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1208 	}
1209 
1210 	if (flags & LK_INTERLOCK)
1211 		class->lc_unlock(ilk);
1212 	if (wakeup_swapper)
1213 		kick_proc0();
1214 
1215 	return (error);
1216 }
1217 
1218 void
1219 _lockmgr_disown(struct lock *lk, const char *file, int line)
1220 {
1221 	uintptr_t tid, x;
1222 
1223 	tid = (uintptr_t)curthread;
1224 	_lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line);
1225 
1226 	/*
1227 	 * If the owner is already LK_KERNPROC just skip the whole operation.
1228 	 */
1229 	if (LK_HOLDER(lk->lk_lock) != tid)
1230 		return;
1231 	lock_profile_release_lock(&lk->lock_object);
1232 	LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1233 	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1234 	TD_LOCKS_DEC(curthread);
1235 	STACK_SAVE(lk);
1236 
1237 	/*
1238 	 * In order to preserve waiters flags, just spin.
1239 	 */
1240 	for (;;) {
1241 		x = lk->lk_lock;
1242 		MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1243 		x &= LK_ALL_WAITERS;
1244 		if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1245 		    LK_KERNPROC | x))
1246 			return;
1247 		cpu_spinwait();
1248 	}
1249 }
1250 
1251 void
1252 lockmgr_printinfo(struct lock *lk)
1253 {
1254 	struct thread *td;
1255 	uintptr_t x;
1256 
1257 	if (lk->lk_lock == LK_UNLOCKED)
1258 		printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1259 	else if (lk->lk_lock & LK_SHARE)
1260 		printf("lock type %s: SHARED (count %ju)\n",
1261 		    lk->lock_object.lo_name,
1262 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1263 	else {
1264 		td = lockmgr_xholder(lk);
1265 		printf("lock type %s: EXCL by thread %p (pid %d)\n",
1266 		    lk->lock_object.lo_name, td, td->td_proc->p_pid);
1267 	}
1268 
1269 	x = lk->lk_lock;
1270 	if (x & LK_EXCLUSIVE_WAITERS)
1271 		printf(" with exclusive waiters pending\n");
1272 	if (x & LK_SHARED_WAITERS)
1273 		printf(" with shared waiters pending\n");
1274 	if (x & LK_EXCLUSIVE_SPINNERS)
1275 		printf(" with exclusive spinners pending\n");
1276 
1277 	STACK_PRINT(lk);
1278 }
1279 
1280 int
1281 lockstatus(struct lock *lk)
1282 {
1283 	uintptr_t v, x;
1284 	int ret;
1285 
1286 	ret = LK_SHARED;
1287 	x = lk->lk_lock;
1288 	v = LK_HOLDER(x);
1289 
1290 	if ((x & LK_SHARE) == 0) {
1291 		if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1292 			ret = LK_EXCLUSIVE;
1293 		else
1294 			ret = LK_EXCLOTHER;
1295 	} else if (x == LK_UNLOCKED)
1296 		ret = 0;
1297 
1298 	return (ret);
1299 }
1300 
1301 #ifdef INVARIANT_SUPPORT
1302 
1303 FEATURE(invariant_support,
1304     "Support for modules compiled with INVARIANTS option");
1305 
1306 #ifndef INVARIANTS
1307 #undef	_lockmgr_assert
1308 #endif
1309 
1310 void
1311 _lockmgr_assert(struct lock *lk, int what, const char *file, int line)
1312 {
1313 	int slocked = 0;
1314 
1315 	if (panicstr != NULL)
1316 		return;
1317 	switch (what) {
1318 	case KA_SLOCKED:
1319 	case KA_SLOCKED | KA_NOTRECURSED:
1320 	case KA_SLOCKED | KA_RECURSED:
1321 		slocked = 1;
1322 	case KA_LOCKED:
1323 	case KA_LOCKED | KA_NOTRECURSED:
1324 	case KA_LOCKED | KA_RECURSED:
1325 #ifdef WITNESS
1326 
1327 		/*
1328 		 * We cannot trust WITNESS if the lock is held in exclusive
1329 		 * mode and a call to lockmgr_disown() happened.
1330 		 * Workaround this skipping the check if the lock is held in
1331 		 * exclusive mode even for the KA_LOCKED case.
1332 		 */
1333 		if (slocked || (lk->lk_lock & LK_SHARE)) {
1334 			witness_assert(&lk->lock_object, what, file, line);
1335 			break;
1336 		}
1337 #endif
1338 		if (lk->lk_lock == LK_UNLOCKED ||
1339 		    ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1340 		    (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1341 			panic("Lock %s not %slocked @ %s:%d\n",
1342 			    lk->lock_object.lo_name, slocked ? "share" : "",
1343 			    file, line);
1344 
1345 		if ((lk->lk_lock & LK_SHARE) == 0) {
1346 			if (lockmgr_recursed(lk)) {
1347 				if (what & KA_NOTRECURSED)
1348 					panic("Lock %s recursed @ %s:%d\n",
1349 					    lk->lock_object.lo_name, file,
1350 					    line);
1351 			} else if (what & KA_RECURSED)
1352 				panic("Lock %s not recursed @ %s:%d\n",
1353 				    lk->lock_object.lo_name, file, line);
1354 		}
1355 		break;
1356 	case KA_XLOCKED:
1357 	case KA_XLOCKED | KA_NOTRECURSED:
1358 	case KA_XLOCKED | KA_RECURSED:
1359 		if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1360 			panic("Lock %s not exclusively locked @ %s:%d\n",
1361 			    lk->lock_object.lo_name, file, line);
1362 		if (lockmgr_recursed(lk)) {
1363 			if (what & KA_NOTRECURSED)
1364 				panic("Lock %s recursed @ %s:%d\n",
1365 				    lk->lock_object.lo_name, file, line);
1366 		} else if (what & KA_RECURSED)
1367 			panic("Lock %s not recursed @ %s:%d\n",
1368 			    lk->lock_object.lo_name, file, line);
1369 		break;
1370 	case KA_UNLOCKED:
1371 		if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1372 			panic("Lock %s exclusively locked @ %s:%d\n",
1373 			    lk->lock_object.lo_name, file, line);
1374 		break;
1375 	default:
1376 		panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1377 		    line);
1378 	}
1379 }
1380 #endif
1381 
1382 #ifdef DDB
1383 int
1384 lockmgr_chain(struct thread *td, struct thread **ownerp)
1385 {
1386 	struct lock *lk;
1387 
1388 	lk = td->td_wchan;
1389 
1390 	if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1391 		return (0);
1392 	db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1393 	if (lk->lk_lock & LK_SHARE)
1394 		db_printf("SHARED (count %ju)\n",
1395 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1396 	else
1397 		db_printf("EXCL\n");
1398 	*ownerp = lockmgr_xholder(lk);
1399 
1400 	return (1);
1401 }
1402 
1403 static void
1404 db_show_lockmgr(struct lock_object *lock)
1405 {
1406 	struct thread *td;
1407 	struct lock *lk;
1408 
1409 	lk = (struct lock *)lock;
1410 
1411 	db_printf(" state: ");
1412 	if (lk->lk_lock == LK_UNLOCKED)
1413 		db_printf("UNLOCKED\n");
1414 	else if (lk->lk_lock & LK_SHARE)
1415 		db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1416 	else {
1417 		td = lockmgr_xholder(lk);
1418 		if (td == (struct thread *)LK_KERNPROC)
1419 			db_printf("XLOCK: LK_KERNPROC\n");
1420 		else
1421 			db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1422 			    td->td_tid, td->td_proc->p_pid,
1423 			    td->td_proc->p_comm);
1424 		if (lockmgr_recursed(lk))
1425 			db_printf(" recursed: %d\n", lk->lk_recurse);
1426 	}
1427 	db_printf(" waiters: ");
1428 	switch (lk->lk_lock & LK_ALL_WAITERS) {
1429 	case LK_SHARED_WAITERS:
1430 		db_printf("shared\n");
1431 		break;
1432 	case LK_EXCLUSIVE_WAITERS:
1433 		db_printf("exclusive\n");
1434 		break;
1435 	case LK_ALL_WAITERS:
1436 		db_printf("shared and exclusive\n");
1437 		break;
1438 	default:
1439 		db_printf("none\n");
1440 	}
1441 	db_printf(" spinners: ");
1442 	if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1443 		db_printf("exclusive\n");
1444 	else
1445 		db_printf("none\n");
1446 }
1447 #endif
1448