xref: /freebsd/sys/kern/kern_lock.c (revision d0ba1baed3f6e4936a0c1b89c25f6c59168ef6de)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice(s), this list of conditions and the following disclaimer as
12  *    the first lines of this file unmodified other than the possible
13  *    addition of one or more copyright notices.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice(s), this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
28  * DAMAGE.
29  */
30 
31 #include "opt_ddb.h"
32 #include "opt_hwpmc_hooks.h"
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include <sys/param.h>
38 #include <sys/kdb.h>
39 #include <sys/ktr.h>
40 #include <sys/lock.h>
41 #include <sys/lock_profile.h>
42 #include <sys/lockmgr.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/sleepqueue.h>
46 #ifdef DEBUG_LOCKS
47 #include <sys/stack.h>
48 #endif
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 
52 #include <machine/cpu.h>
53 
54 #ifdef DDB
55 #include <ddb/ddb.h>
56 #endif
57 
58 #ifdef HWPMC_HOOKS
59 #include <sys/pmckern.h>
60 PMC_SOFT_DECLARE( , , lock, failed);
61 #endif
62 
63 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
64     (LK_ADAPTIVE | LK_NOSHARE));
65 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
66     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
67 
68 #define	SQ_EXCLUSIVE_QUEUE	0
69 #define	SQ_SHARED_QUEUE		1
70 
71 #ifndef INVARIANTS
72 #define	_lockmgr_assert(lk, what, file, line)
73 #endif
74 
75 #define	TD_SLOCKS_INC(td)	((td)->td_lk_slocks++)
76 #define	TD_SLOCKS_DEC(td)	((td)->td_lk_slocks--)
77 
78 #ifndef DEBUG_LOCKS
79 #define	STACK_PRINT(lk)
80 #define	STACK_SAVE(lk)
81 #define	STACK_ZERO(lk)
82 #else
83 #define	STACK_PRINT(lk)	stack_print_ddb(&(lk)->lk_stack)
84 #define	STACK_SAVE(lk)	stack_save(&(lk)->lk_stack)
85 #define	STACK_ZERO(lk)	stack_zero(&(lk)->lk_stack)
86 #endif
87 
88 #define	LOCK_LOG2(lk, string, arg1, arg2)				\
89 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
90 		CTR2(KTR_LOCK, (string), (arg1), (arg2))
91 #define	LOCK_LOG3(lk, string, arg1, arg2, arg3)				\
92 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
93 		CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
94 
95 #define	GIANT_DECLARE							\
96 	int _i = 0;							\
97 	WITNESS_SAVE_DECL(Giant)
98 #define	GIANT_RESTORE() do {						\
99 	if (_i > 0) {							\
100 		while (_i--)						\
101 			mtx_lock(&Giant);				\
102 		WITNESS_RESTORE(&Giant.lock_object, Giant);		\
103 	}								\
104 } while (0)
105 #define	GIANT_SAVE() do {						\
106 	if (mtx_owned(&Giant)) {					\
107 		WITNESS_SAVE(&Giant.lock_object, Giant);		\
108 		while (mtx_owned(&Giant)) {				\
109 			_i++;						\
110 			mtx_unlock(&Giant);				\
111 		}							\
112 	}								\
113 } while (0)
114 
115 #define	LK_CAN_SHARE(x, flags)						\
116 	(((x) & LK_SHARE) &&						\
117 	(((x) & (LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 0 ||	\
118 	(curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||	\
119 	(curthread->td_pflags & TDP_DEADLKTREAT)))
120 #define	LK_TRYOP(x)							\
121 	((x) & LK_NOWAIT)
122 
123 #define	LK_CAN_WITNESS(x)						\
124 	(((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
125 #define	LK_TRYWIT(x)							\
126 	(LK_TRYOP(x) ? LOP_TRYLOCK : 0)
127 
128 #define	LK_CAN_ADAPT(lk, f)						\
129 	(((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&		\
130 	((f) & LK_SLEEPFAIL) == 0)
131 
132 #define	lockmgr_disowned(lk)						\
133 	(((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
134 
135 #define	lockmgr_xlocked_v(v)						\
136 	(((v) & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
137 
138 #define	lockmgr_xlocked(lk) lockmgr_xlocked_v((lk)->lk_lock)
139 
140 static void	assert_lockmgr(const struct lock_object *lock, int how);
141 #ifdef DDB
142 static void	db_show_lockmgr(const struct lock_object *lock);
143 #endif
144 static void	lock_lockmgr(struct lock_object *lock, uintptr_t how);
145 #ifdef KDTRACE_HOOKS
146 static int	owner_lockmgr(const struct lock_object *lock,
147 		    struct thread **owner);
148 #endif
149 static uintptr_t unlock_lockmgr(struct lock_object *lock);
150 
151 struct lock_class lock_class_lockmgr = {
152 	.lc_name = "lockmgr",
153 	.lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
154 	.lc_assert = assert_lockmgr,
155 #ifdef DDB
156 	.lc_ddb_show = db_show_lockmgr,
157 #endif
158 	.lc_lock = lock_lockmgr,
159 	.lc_unlock = unlock_lockmgr,
160 #ifdef KDTRACE_HOOKS
161 	.lc_owner = owner_lockmgr,
162 #endif
163 };
164 
165 struct lockmgr_wait {
166 	const char *iwmesg;
167 	int ipri;
168 	int itimo;
169 };
170 
171 static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp,
172     int flags);
173 static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp);
174 
175 static void
176 lockmgr_exit(u_int flags, struct lock_object *ilk, int wakeup_swapper)
177 {
178 	struct lock_class *class;
179 
180 	if (flags & LK_INTERLOCK) {
181 		class = LOCK_CLASS(ilk);
182 		class->lc_unlock(ilk);
183 	}
184 
185 	if (__predict_false(wakeup_swapper))
186 		kick_proc0();
187 }
188 
189 static void
190 lockmgr_note_shared_acquire(struct lock *lk, int contested,
191     uint64_t waittime, const char *file, int line, int flags)
192 {
193 
194 	lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
195 	    file, line);
196 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line);
197 	WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line);
198 	TD_LOCKS_INC(curthread);
199 	TD_SLOCKS_INC(curthread);
200 	STACK_SAVE(lk);
201 }
202 
203 static void
204 lockmgr_note_shared_release(struct lock *lk, const char *file, int line)
205 {
206 
207 	lock_profile_release_lock(&lk->lock_object);
208 	WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
209 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
210 	TD_LOCKS_DEC(curthread);
211 	TD_SLOCKS_DEC(curthread);
212 }
213 
214 static void
215 lockmgr_note_exclusive_acquire(struct lock *lk, int contested,
216     uint64_t waittime, const char *file, int line, int flags)
217 {
218 
219 	lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
220 	    file, line);
221 	LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
222 	WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file,
223 	    line);
224 	TD_LOCKS_INC(curthread);
225 	STACK_SAVE(lk);
226 }
227 
228 static void
229 lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line)
230 {
231 
232 	lock_profile_release_lock(&lk->lock_object);
233 	LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file,
234 	    line);
235 	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
236 	TD_LOCKS_DEC(curthread);
237 }
238 
239 static __inline struct thread *
240 lockmgr_xholder(const struct lock *lk)
241 {
242 	uintptr_t x;
243 
244 	x = lk->lk_lock;
245 	return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
246 }
247 
248 /*
249  * It assumes sleepq_lock held and returns with this one unheld.
250  * It also assumes the generic interlock is sane and previously checked.
251  * If LK_INTERLOCK is specified the interlock is not reacquired after the
252  * sleep.
253  */
254 static __inline int
255 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
256     const char *wmesg, int pri, int timo, int queue)
257 {
258 	GIANT_DECLARE;
259 	struct lock_class *class;
260 	int catch, error;
261 
262 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
263 	catch = pri & PCATCH;
264 	pri &= PRIMASK;
265 	error = 0;
266 
267 	LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
268 	    (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
269 
270 	if (flags & LK_INTERLOCK)
271 		class->lc_unlock(ilk);
272 	if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
273 		lk->lk_exslpfail++;
274 	GIANT_SAVE();
275 	sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
276 	    SLEEPQ_INTERRUPTIBLE : 0), queue);
277 	if ((flags & LK_TIMELOCK) && timo)
278 		sleepq_set_timeout(&lk->lock_object, timo);
279 
280 	/*
281 	 * Decisional switch for real sleeping.
282 	 */
283 	if ((flags & LK_TIMELOCK) && timo && catch)
284 		error = sleepq_timedwait_sig(&lk->lock_object, pri);
285 	else if ((flags & LK_TIMELOCK) && timo)
286 		error = sleepq_timedwait(&lk->lock_object, pri);
287 	else if (catch)
288 		error = sleepq_wait_sig(&lk->lock_object, pri);
289 	else
290 		sleepq_wait(&lk->lock_object, pri);
291 	GIANT_RESTORE();
292 	if ((flags & LK_SLEEPFAIL) && error == 0)
293 		error = ENOLCK;
294 
295 	return (error);
296 }
297 
298 static __inline int
299 wakeupshlk(struct lock *lk, const char *file, int line)
300 {
301 	uintptr_t v, x, orig_x;
302 	u_int realexslp;
303 	int queue, wakeup_swapper;
304 
305 	wakeup_swapper = 0;
306 	for (;;) {
307 		x = lk->lk_lock;
308 		if (lockmgr_sunlock_try(lk, &x))
309 			break;
310 
311 		/*
312 		 * We should have a sharer with waiters, so enter the hard
313 		 * path in order to handle wakeups correctly.
314 		 */
315 		sleepq_lock(&lk->lock_object);
316 		orig_x = lk->lk_lock;
317 retry_sleepq:
318 		x = orig_x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
319 		v = LK_UNLOCKED;
320 
321 		/*
322 		 * If the lock has exclusive waiters, give them preference in
323 		 * order to avoid deadlock with shared runners up.
324 		 * If interruptible sleeps left the exclusive queue empty
325 		 * avoid a starvation for the threads sleeping on the shared
326 		 * queue by giving them precedence and cleaning up the
327 		 * exclusive waiters bit anyway.
328 		 * Please note that lk_exslpfail count may be lying about
329 		 * the real number of waiters with the LK_SLEEPFAIL flag on
330 		 * because they may be used in conjunction with interruptible
331 		 * sleeps so lk_exslpfail might be considered an 'upper limit'
332 		 * bound, including the edge cases.
333 		 */
334 		realexslp = sleepq_sleepcnt(&lk->lock_object,
335 		    SQ_EXCLUSIVE_QUEUE);
336 		if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
337 			if (lk->lk_exslpfail < realexslp) {
338 				lk->lk_exslpfail = 0;
339 				queue = SQ_EXCLUSIVE_QUEUE;
340 				v |= (x & LK_SHARED_WAITERS);
341 			} else {
342 				lk->lk_exslpfail = 0;
343 				LOCK_LOG2(lk,
344 				    "%s: %p has only LK_SLEEPFAIL sleepers",
345 				    __func__, lk);
346 				LOCK_LOG2(lk,
347 			    "%s: %p waking up threads on the exclusive queue",
348 				    __func__, lk);
349 				wakeup_swapper =
350 				    sleepq_broadcast(&lk->lock_object,
351 				    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
352 				queue = SQ_SHARED_QUEUE;
353 			}
354 
355 		} else {
356 
357 			/*
358 			 * Exclusive waiters sleeping with LK_SLEEPFAIL on
359 			 * and using interruptible sleeps/timeout may have
360 			 * left spourious lk_exslpfail counts on, so clean
361 			 * it up anyway.
362 			 */
363 			lk->lk_exslpfail = 0;
364 			queue = SQ_SHARED_QUEUE;
365 		}
366 
367 		if (lockmgr_sunlock_try(lk, &orig_x)) {
368 			sleepq_release(&lk->lock_object);
369 			break;
370 		}
371 
372 		x |= LK_SHARERS_LOCK(1);
373 		if (!atomic_fcmpset_rel_ptr(&lk->lk_lock, &x, v)) {
374 			orig_x = x;
375 			goto retry_sleepq;
376 		}
377 		LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
378 		    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
379 		    "exclusive");
380 		wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
381 		    0, queue);
382 		sleepq_release(&lk->lock_object);
383 		break;
384 	}
385 
386 	lockmgr_note_shared_release(lk, file, line);
387 	return (wakeup_swapper);
388 }
389 
390 static void
391 assert_lockmgr(const struct lock_object *lock, int what)
392 {
393 
394 	panic("lockmgr locks do not support assertions");
395 }
396 
397 static void
398 lock_lockmgr(struct lock_object *lock, uintptr_t how)
399 {
400 
401 	panic("lockmgr locks do not support sleep interlocking");
402 }
403 
404 static uintptr_t
405 unlock_lockmgr(struct lock_object *lock)
406 {
407 
408 	panic("lockmgr locks do not support sleep interlocking");
409 }
410 
411 #ifdef KDTRACE_HOOKS
412 static int
413 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
414 {
415 
416 	panic("lockmgr locks do not support owner inquiring");
417 }
418 #endif
419 
420 void
421 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
422 {
423 	int iflags;
424 
425 	MPASS((flags & ~LK_INIT_MASK) == 0);
426 	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
427             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
428             &lk->lk_lock));
429 
430 	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
431 	if (flags & LK_CANRECURSE)
432 		iflags |= LO_RECURSABLE;
433 	if ((flags & LK_NODUP) == 0)
434 		iflags |= LO_DUPOK;
435 	if (flags & LK_NOPROFILE)
436 		iflags |= LO_NOPROFILE;
437 	if ((flags & LK_NOWITNESS) == 0)
438 		iflags |= LO_WITNESS;
439 	if (flags & LK_QUIET)
440 		iflags |= LO_QUIET;
441 	if (flags & LK_IS_VNODE)
442 		iflags |= LO_IS_VNODE;
443 	iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
444 
445 	lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
446 	lk->lk_lock = LK_UNLOCKED;
447 	lk->lk_recurse = 0;
448 	lk->lk_exslpfail = 0;
449 	lk->lk_timo = timo;
450 	lk->lk_pri = pri;
451 	STACK_ZERO(lk);
452 }
453 
454 /*
455  * XXX: Gross hacks to manipulate external lock flags after
456  * initialization.  Used for certain vnode and buf locks.
457  */
458 void
459 lockallowshare(struct lock *lk)
460 {
461 
462 	lockmgr_assert(lk, KA_XLOCKED);
463 	lk->lock_object.lo_flags &= ~LK_NOSHARE;
464 }
465 
466 void
467 lockdisableshare(struct lock *lk)
468 {
469 
470 	lockmgr_assert(lk, KA_XLOCKED);
471 	lk->lock_object.lo_flags |= LK_NOSHARE;
472 }
473 
474 void
475 lockallowrecurse(struct lock *lk)
476 {
477 
478 	lockmgr_assert(lk, KA_XLOCKED);
479 	lk->lock_object.lo_flags |= LO_RECURSABLE;
480 }
481 
482 void
483 lockdisablerecurse(struct lock *lk)
484 {
485 
486 	lockmgr_assert(lk, KA_XLOCKED);
487 	lk->lock_object.lo_flags &= ~LO_RECURSABLE;
488 }
489 
490 void
491 lockdestroy(struct lock *lk)
492 {
493 
494 	KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
495 	KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
496 	KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
497 	lock_destroy(&lk->lock_object);
498 }
499 
500 static bool __always_inline
501 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags)
502 {
503 
504 	/*
505 	 * If no other thread has an exclusive lock, or
506 	 * no exclusive waiter is present, bump the count of
507 	 * sharers.  Since we have to preserve the state of
508 	 * waiters, if we fail to acquire the shared lock
509 	 * loop back and retry.
510 	 */
511 	*xp = lk->lk_lock;
512 	while (LK_CAN_SHARE(*xp, flags)) {
513 		if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
514 		    *xp + LK_ONE_SHARER)) {
515 			return (true);
516 		}
517 	}
518 	return (false);
519 }
520 
521 static bool __always_inline
522 lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp)
523 {
524 
525 	for (;;) {
526 		/*
527 		 * If there is more than one shared lock held, just drop one
528 		 * and return.
529 		 */
530 		if (LK_SHARERS(*xp) > 1) {
531 			if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp,
532 			    *xp - LK_ONE_SHARER))
533 				return (true);
534 			continue;
535 		}
536 
537 		/*
538 		 * If there are not waiters on the exclusive queue, drop the
539 		 * lock quickly.
540 		 */
541 		if ((*xp & LK_ALL_WAITERS) == 0) {
542 			MPASS((*xp & ~LK_EXCLUSIVE_SPINNERS) ==
543 			    LK_SHARERS_LOCK(1));
544 			if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp,
545 			    LK_UNLOCKED))
546 				return (true);
547 			continue;
548 		}
549 		break;
550 	}
551 	return (false);
552 }
553 
554 static __noinline int
555 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
556     const char *file, int line, struct lockmgr_wait *lwa)
557 {
558 	uintptr_t tid, x;
559 	int error = 0;
560 	const char *iwmesg;
561 	int ipri, itimo;
562 
563 #ifdef LOCK_PROFILING
564 	uint64_t waittime = 0;
565 	int contested = 0;
566 #endif
567 
568 	if (__predict_false(panicstr != NULL))
569 		goto out;
570 
571 	tid = (uintptr_t)curthread;
572 
573 	if (LK_CAN_WITNESS(flags))
574 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
575 		    file, line, flags & LK_INTERLOCK ? ilk : NULL);
576 	for (;;) {
577 		if (lockmgr_slock_try(lk, &x, flags))
578 			break;
579 #ifdef HWPMC_HOOKS
580 		PMC_SOFT_CALL( , , lock, failed);
581 #endif
582 		lock_profile_obtain_lock_failed(&lk->lock_object,
583 		    &contested, &waittime);
584 
585 		/*
586 		 * If the lock is already held by curthread in
587 		 * exclusive way avoid a deadlock.
588 		 */
589 		if (LK_HOLDER(x) == tid) {
590 			LOCK_LOG2(lk,
591 			    "%s: %p already held in exclusive mode",
592 			    __func__, lk);
593 			error = EDEADLK;
594 			break;
595 		}
596 
597 		/*
598 		 * If the lock is expected to not sleep just give up
599 		 * and return.
600 		 */
601 		if (LK_TRYOP(flags)) {
602 			LOCK_LOG2(lk, "%s: %p fails the try operation",
603 			    __func__, lk);
604 			error = EBUSY;
605 			break;
606 		}
607 
608 		/*
609 		 * Acquire the sleepqueue chain lock because we
610 		 * probabilly will need to manipulate waiters flags.
611 		 */
612 		sleepq_lock(&lk->lock_object);
613 		x = lk->lk_lock;
614 retry_sleepq:
615 
616 		/*
617 		 * if the lock can be acquired in shared mode, try
618 		 * again.
619 		 */
620 		if (LK_CAN_SHARE(x, flags)) {
621 			sleepq_release(&lk->lock_object);
622 			continue;
623 		}
624 
625 		/*
626 		 * Try to set the LK_SHARED_WAITERS flag.  If we fail,
627 		 * loop back and retry.
628 		 */
629 		if ((x & LK_SHARED_WAITERS) == 0) {
630 			if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
631 			    x | LK_SHARED_WAITERS)) {
632 				goto retry_sleepq;
633 			}
634 			LOCK_LOG2(lk, "%s: %p set shared waiters flag",
635 			    __func__, lk);
636 		}
637 
638 		if (lwa == NULL) {
639 			iwmesg = lk->lock_object.lo_name;
640 			ipri = lk->lk_pri;
641 			itimo = lk->lk_timo;
642 		} else {
643 			iwmesg = lwa->iwmesg;
644 			ipri = lwa->ipri;
645 			itimo = lwa->itimo;
646 		}
647 
648 		/*
649 		 * As far as we have been unable to acquire the
650 		 * shared lock and the shared waiters flag is set,
651 		 * we will sleep.
652 		 */
653 		error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
654 		    SQ_SHARED_QUEUE);
655 		flags &= ~LK_INTERLOCK;
656 		if (error) {
657 			LOCK_LOG3(lk,
658 			    "%s: interrupted sleep for %p with %d",
659 			    __func__, lk, error);
660 			break;
661 		}
662 		LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
663 		    __func__, lk);
664 	}
665 	if (error == 0) {
666 #ifdef LOCK_PROFILING
667 		lockmgr_note_shared_acquire(lk, contested, waittime,
668 		    file, line, flags);
669 #else
670 		lockmgr_note_shared_acquire(lk, 0, 0, file, line,
671 		    flags);
672 #endif
673 	}
674 
675 out:
676 	lockmgr_exit(flags, ilk, 0);
677 	return (error);
678 }
679 
680 static __noinline int
681 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
682     const char *file, int line, struct lockmgr_wait *lwa)
683 {
684 	struct lock_class *class;
685 	uintptr_t tid, x, v;
686 	int error = 0;
687 	const char *iwmesg;
688 	int ipri, itimo;
689 
690 #ifdef LOCK_PROFILING
691 	uint64_t waittime = 0;
692 	int contested = 0;
693 #endif
694 
695 	if (__predict_false(panicstr != NULL))
696 		goto out;
697 
698 	tid = (uintptr_t)curthread;
699 
700 	if (LK_CAN_WITNESS(flags))
701 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
702 		    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
703 		    ilk : NULL);
704 
705 	/*
706 	 * If curthread already holds the lock and this one is
707 	 * allowed to recurse, simply recurse on it.
708 	 */
709 	if (lockmgr_xlocked(lk)) {
710 		if ((flags & LK_CANRECURSE) == 0 &&
711 		    (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
712 			/*
713 			 * If the lock is expected to not panic just
714 			 * give up and return.
715 			 */
716 			if (LK_TRYOP(flags)) {
717 				LOCK_LOG2(lk,
718 				    "%s: %p fails the try operation",
719 				    __func__, lk);
720 				error = EBUSY;
721 				goto out;
722 			}
723 			if (flags & LK_INTERLOCK) {
724 				class = LOCK_CLASS(ilk);
725 				class->lc_unlock(ilk);
726 			}
727 			panic("%s: recursing on non recursive lockmgr %p "
728 			    "@ %s:%d\n", __func__, lk, file, line);
729 		}
730 		lk->lk_recurse++;
731 		LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
732 		LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
733 		    lk->lk_recurse, file, line);
734 		WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
735 		    LK_TRYWIT(flags), file, line);
736 		TD_LOCKS_INC(curthread);
737 		goto out;
738 	}
739 
740 	for (;;) {
741 		if (lk->lk_lock == LK_UNLOCKED &&
742 		    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
743 			break;
744 #ifdef HWPMC_HOOKS
745 		PMC_SOFT_CALL( , , lock, failed);
746 #endif
747 		lock_profile_obtain_lock_failed(&lk->lock_object,
748 		    &contested, &waittime);
749 
750 		/*
751 		 * If the lock is expected to not sleep just give up
752 		 * and return.
753 		 */
754 		if (LK_TRYOP(flags)) {
755 			LOCK_LOG2(lk, "%s: %p fails the try operation",
756 			    __func__, lk);
757 			error = EBUSY;
758 			break;
759 		}
760 
761 		/*
762 		 * Acquire the sleepqueue chain lock because we
763 		 * probabilly will need to manipulate waiters flags.
764 		 */
765 		sleepq_lock(&lk->lock_object);
766 		x = lk->lk_lock;
767 retry_sleepq:
768 
769 		/*
770 		 * if the lock has been released while we spun on
771 		 * the sleepqueue chain lock just try again.
772 		 */
773 		if (x == LK_UNLOCKED) {
774 			sleepq_release(&lk->lock_object);
775 			continue;
776 		}
777 
778 		/*
779 		 * The lock can be in the state where there is a
780 		 * pending queue of waiters, but still no owner.
781 		 * This happens when the lock is contested and an
782 		 * owner is going to claim the lock.
783 		 * If curthread is the one successfully acquiring it
784 		 * claim lock ownership and return, preserving waiters
785 		 * flags.
786 		 */
787 		v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
788 		if ((x & ~v) == LK_UNLOCKED) {
789 			v &= ~LK_EXCLUSIVE_SPINNERS;
790 			if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
791 			    tid | v)) {
792 				sleepq_release(&lk->lock_object);
793 				LOCK_LOG2(lk,
794 				    "%s: %p claimed by a new writer",
795 				    __func__, lk);
796 				break;
797 			}
798 			goto retry_sleepq;
799 		}
800 
801 		/*
802 		 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
803 		 * fail, loop back and retry.
804 		 */
805 		if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
806 			if (!atomic_fcmpset_ptr(&lk->lk_lock, &x,
807 			    x | LK_EXCLUSIVE_WAITERS)) {
808 				goto retry_sleepq;
809 			}
810 			LOCK_LOG2(lk, "%s: %p set excl waiters flag",
811 			    __func__, lk);
812 		}
813 
814 		if (lwa == NULL) {
815 			iwmesg = lk->lock_object.lo_name;
816 			ipri = lk->lk_pri;
817 			itimo = lk->lk_timo;
818 		} else {
819 			iwmesg = lwa->iwmesg;
820 			ipri = lwa->ipri;
821 			itimo = lwa->itimo;
822 		}
823 
824 		/*
825 		 * As far as we have been unable to acquire the
826 		 * exclusive lock and the exclusive waiters flag
827 		 * is set, we will sleep.
828 		 */
829 		error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
830 		    SQ_EXCLUSIVE_QUEUE);
831 		flags &= ~LK_INTERLOCK;
832 		if (error) {
833 			LOCK_LOG3(lk,
834 			    "%s: interrupted sleep for %p with %d",
835 			    __func__, lk, error);
836 			break;
837 		}
838 		LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
839 		    __func__, lk);
840 	}
841 	if (error == 0) {
842 #ifdef LOCK_PROFILING
843 		lockmgr_note_exclusive_acquire(lk, contested, waittime,
844 		    file, line, flags);
845 #else
846 		lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
847 		    flags);
848 #endif
849 	}
850 
851 out:
852 	lockmgr_exit(flags, ilk, 0);
853 	return (error);
854 }
855 
856 static __noinline int
857 lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk,
858     const char *file, int line, struct lockmgr_wait *lwa)
859 {
860 	uintptr_t tid, x, v;
861 	int error = 0;
862 	int wakeup_swapper = 0;
863 	int op;
864 
865 	if (__predict_false(panicstr != NULL))
866 		goto out;
867 
868 	tid = (uintptr_t)curthread;
869 
870 	_lockmgr_assert(lk, KA_SLOCKED, file, line);
871 	v = lk->lk_lock;
872 	x = v & LK_ALL_WAITERS;
873 	v &= LK_EXCLUSIVE_SPINNERS;
874 
875 	/*
876 	 * Try to switch from one shared lock to an exclusive one.
877 	 * We need to preserve waiters flags during the operation.
878 	 */
879 	if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
880 	    tid | x)) {
881 		LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
882 		    line);
883 		WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
884 		    LK_TRYWIT(flags), file, line);
885 		TD_SLOCKS_DEC(curthread);
886 		goto out;
887 	}
888 
889 	op = flags & LK_TYPE_MASK;
890 
891 	/*
892 	 * In LK_TRYUPGRADE mode, do not drop the lock,
893 	 * returning EBUSY instead.
894 	 */
895 	if (op == LK_TRYUPGRADE) {
896 		LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
897 		    __func__, lk);
898 		error = EBUSY;
899 		goto out;
900 	}
901 
902 	/*
903 	 * We have been unable to succeed in upgrading, so just
904 	 * give up the shared lock.
905 	 */
906 	wakeup_swapper |= wakeupshlk(lk, file, line);
907 	error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa);
908 	flags &= ~LK_INTERLOCK;
909 out:
910 	lockmgr_exit(flags, ilk, wakeup_swapper);
911 	return (error);
912 }
913 
914 int
915 lockmgr_lock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk,
916     const char *file, int line)
917 {
918 	struct lock_class *class;
919 	uintptr_t x, tid;
920 	u_int op;
921 	bool locked;
922 
923 	if (__predict_false(panicstr != NULL))
924 		return (0);
925 
926 	op = flags & LK_TYPE_MASK;
927 	locked = false;
928 	switch (op) {
929 	case LK_SHARED:
930 		if (LK_CAN_WITNESS(flags))
931 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
932 			    file, line, flags & LK_INTERLOCK ? ilk : NULL);
933 		if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
934 			break;
935 		if (lockmgr_slock_try(lk, &x, flags)) {
936 			lockmgr_note_shared_acquire(lk, 0, 0,
937 			    file, line, flags);
938 			locked = true;
939 		} else {
940 			return (lockmgr_slock_hard(lk, flags, ilk, file, line,
941 			    NULL));
942 		}
943 		break;
944 	case LK_EXCLUSIVE:
945 		if (LK_CAN_WITNESS(flags))
946 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
947 			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
948 			    ilk : NULL);
949 		tid = (uintptr_t)curthread;
950 		if (lk->lk_lock == LK_UNLOCKED &&
951 		    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
952 			lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
953 			    flags);
954 			locked = true;
955 		} else {
956 			return (lockmgr_xlock_hard(lk, flags, ilk, file, line,
957 			    NULL));
958 		}
959 		break;
960 	case LK_UPGRADE:
961 	case LK_TRYUPGRADE:
962 		return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL));
963 	default:
964 		break;
965 	}
966 	if (__predict_true(locked)) {
967 		if (__predict_false(flags & LK_INTERLOCK)) {
968 			class = LOCK_CLASS(ilk);
969 			class->lc_unlock(ilk);
970 		}
971 		return (0);
972 	} else {
973 		return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT,
974 		    LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line));
975 	}
976 }
977 
978 static __noinline int
979 lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
980     const char *file, int line)
981 
982 {
983 	int wakeup_swapper = 0;
984 
985 	if (__predict_false(panicstr != NULL))
986 		goto out;
987 
988 	wakeup_swapper = wakeupshlk(lk, file, line);
989 
990 out:
991 	lockmgr_exit(flags, ilk, wakeup_swapper);
992 	return (0);
993 }
994 
995 static __noinline int
996 lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
997     const char *file, int line)
998 {
999 	uintptr_t tid, v;
1000 	int wakeup_swapper = 0;
1001 	u_int realexslp;
1002 	int queue;
1003 
1004 	if (__predict_false(panicstr != NULL))
1005 		goto out;
1006 
1007 	tid = (uintptr_t)curthread;
1008 
1009 	/*
1010 	 * As first option, treact the lock as if it has not
1011 	 * any waiter.
1012 	 * Fix-up the tid var if the lock has been disowned.
1013 	 */
1014 	if (LK_HOLDER(x) == LK_KERNPROC)
1015 		tid = LK_KERNPROC;
1016 	else {
1017 		WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1018 		TD_LOCKS_DEC(curthread);
1019 	}
1020 	LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
1021 
1022 	/*
1023 	 * The lock is held in exclusive mode.
1024 	 * If the lock is recursed also, then unrecurse it.
1025 	 */
1026 	if (lockmgr_xlocked_v(x) && lockmgr_recursed(lk)) {
1027 		LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk);
1028 		lk->lk_recurse--;
1029 		goto out;
1030 	}
1031 	if (tid != LK_KERNPROC)
1032 		lock_profile_release_lock(&lk->lock_object);
1033 
1034 	if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED))
1035 		goto out;
1036 
1037 	sleepq_lock(&lk->lock_object);
1038 	x = lk->lk_lock;
1039 	v = LK_UNLOCKED;
1040 
1041 	/*
1042 	 * If the lock has exclusive waiters, give them
1043 	 * preference in order to avoid deadlock with
1044 	 * shared runners up.
1045 	 * If interruptible sleeps left the exclusive queue
1046 	 * empty avoid a starvation for the threads sleeping
1047 	 * on the shared queue by giving them precedence
1048 	 * and cleaning up the exclusive waiters bit anyway.
1049 	 * Please note that lk_exslpfail count may be lying
1050 	 * about the real number of waiters with the
1051 	 * LK_SLEEPFAIL flag on because they may be used in
1052 	 * conjunction with interruptible sleeps so
1053 	 * lk_exslpfail might be considered an 'upper limit'
1054 	 * bound, including the edge cases.
1055 	 */
1056 	MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1057 	realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE);
1058 	if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1059 		if (lk->lk_exslpfail < realexslp) {
1060 			lk->lk_exslpfail = 0;
1061 			queue = SQ_EXCLUSIVE_QUEUE;
1062 			v |= (x & LK_SHARED_WAITERS);
1063 		} else {
1064 			lk->lk_exslpfail = 0;
1065 			LOCK_LOG2(lk,
1066 			    "%s: %p has only LK_SLEEPFAIL sleepers",
1067 			    __func__, lk);
1068 			LOCK_LOG2(lk,
1069 			    "%s: %p waking up threads on the exclusive queue",
1070 			    __func__, lk);
1071 			wakeup_swapper = sleepq_broadcast(&lk->lock_object,
1072 			    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1073 			queue = SQ_SHARED_QUEUE;
1074 		}
1075 	} else {
1076 
1077 		/*
1078 		 * Exclusive waiters sleeping with LK_SLEEPFAIL
1079 		 * on and using interruptible sleeps/timeout
1080 		 * may have left spourious lk_exslpfail counts
1081 		 * on, so clean it up anyway.
1082 		 */
1083 		lk->lk_exslpfail = 0;
1084 		queue = SQ_SHARED_QUEUE;
1085 	}
1086 
1087 	LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
1088 	    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1089 	    "exclusive");
1090 	atomic_store_rel_ptr(&lk->lk_lock, v);
1091 	wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue);
1092 	sleepq_release(&lk->lock_object);
1093 
1094 out:
1095 	lockmgr_exit(flags, ilk, wakeup_swapper);
1096 	return (0);
1097 }
1098 
1099 int
1100 lockmgr_unlock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk)
1101 {
1102 	struct lock_class *class;
1103 	uintptr_t x, tid;
1104 	const char *file;
1105 	int line;
1106 
1107 	if (__predict_false(panicstr != NULL))
1108 		return (0);
1109 
1110 	file = __FILE__;
1111 	line = __LINE__;
1112 
1113 	_lockmgr_assert(lk, KA_LOCKED, file, line);
1114 	x = lk->lk_lock;
1115 	if (__predict_true(x & LK_SHARE) != 0) {
1116 		if (lockmgr_sunlock_try(lk, &x)) {
1117 			lockmgr_note_shared_release(lk, file, line);
1118 		} else {
1119 			return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1120 		}
1121 	} else {
1122 		tid = (uintptr_t)curthread;
1123 		if (!lockmgr_recursed(lk) &&
1124 		    atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) {
1125 			lockmgr_note_exclusive_release(lk, file, line);
1126 		} else {
1127 			return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1128 		}
1129 	}
1130 	if (__predict_false(flags & LK_INTERLOCK)) {
1131 		class = LOCK_CLASS(ilk);
1132 		class->lc_unlock(ilk);
1133 	}
1134 	return (0);
1135 }
1136 
1137 int
1138 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
1139     const char *wmesg, int pri, int timo, const char *file, int line)
1140 {
1141 	GIANT_DECLARE;
1142 	struct lockmgr_wait lwa;
1143 	struct lock_class *class;
1144 	const char *iwmesg;
1145 	uintptr_t tid, v, x;
1146 	u_int op, realexslp;
1147 	int error, ipri, itimo, queue, wakeup_swapper;
1148 #ifdef LOCK_PROFILING
1149 	uint64_t waittime = 0;
1150 	int contested = 0;
1151 #endif
1152 
1153 	if (panicstr != NULL)
1154 		return (0);
1155 
1156 	error = 0;
1157 	tid = (uintptr_t)curthread;
1158 	op = (flags & LK_TYPE_MASK);
1159 	iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
1160 	ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
1161 	itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
1162 
1163 	lwa.iwmesg = iwmesg;
1164 	lwa.ipri = ipri;
1165 	lwa.itimo = itimo;
1166 
1167 	MPASS((flags & ~LK_TOTAL_MASK) == 0);
1168 	KASSERT((op & (op - 1)) == 0,
1169 	    ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
1170 	KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
1171 	    (op != LK_DOWNGRADE && op != LK_RELEASE),
1172 	    ("%s: Invalid flags in regard of the operation desired @ %s:%d",
1173 	    __func__, file, line));
1174 	KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
1175 	    ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
1176 	    __func__, file, line));
1177 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
1178 	    ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
1179 	    lk->lock_object.lo_name, file, line));
1180 
1181 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
1182 
1183 	if (lk->lock_object.lo_flags & LK_NOSHARE) {
1184 		switch (op) {
1185 		case LK_SHARED:
1186 			op = LK_EXCLUSIVE;
1187 			break;
1188 		case LK_UPGRADE:
1189 		case LK_TRYUPGRADE:
1190 		case LK_DOWNGRADE:
1191 			_lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
1192 			    file, line);
1193 			if (flags & LK_INTERLOCK)
1194 				class->lc_unlock(ilk);
1195 			return (0);
1196 		}
1197 	}
1198 
1199 	wakeup_swapper = 0;
1200 	switch (op) {
1201 	case LK_SHARED:
1202 		return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa));
1203 		break;
1204 	case LK_UPGRADE:
1205 	case LK_TRYUPGRADE:
1206 		return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa));
1207 		break;
1208 	case LK_EXCLUSIVE:
1209 		return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa));
1210 		break;
1211 	case LK_DOWNGRADE:
1212 		_lockmgr_assert(lk, KA_XLOCKED, file, line);
1213 		LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
1214 		WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
1215 
1216 		/*
1217 		 * Panic if the lock is recursed.
1218 		 */
1219 		if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1220 			if (flags & LK_INTERLOCK)
1221 				class->lc_unlock(ilk);
1222 			panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
1223 			    __func__, iwmesg, file, line);
1224 		}
1225 		TD_SLOCKS_INC(curthread);
1226 
1227 		/*
1228 		 * In order to preserve waiters flags, just spin.
1229 		 */
1230 		for (;;) {
1231 			x = lk->lk_lock;
1232 			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1233 			x &= LK_ALL_WAITERS;
1234 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1235 			    LK_SHARERS_LOCK(1) | x))
1236 				break;
1237 			cpu_spinwait();
1238 		}
1239 		break;
1240 	case LK_RELEASE:
1241 		_lockmgr_assert(lk, KA_LOCKED, file, line);
1242 		x = lk->lk_lock;
1243 
1244 		if (__predict_true(x & LK_SHARE) != 0) {
1245 			return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1246 		} else {
1247 			return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1248 		}
1249 		break;
1250 	case LK_DRAIN:
1251 		if (LK_CAN_WITNESS(flags))
1252 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1253 			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1254 			    ilk : NULL);
1255 
1256 		/*
1257 		 * Trying to drain a lock we already own will result in a
1258 		 * deadlock.
1259 		 */
1260 		if (lockmgr_xlocked(lk)) {
1261 			if (flags & LK_INTERLOCK)
1262 				class->lc_unlock(ilk);
1263 			panic("%s: draining %s with the lock held @ %s:%d\n",
1264 			    __func__, iwmesg, file, line);
1265 		}
1266 
1267 		for (;;) {
1268 			if (lk->lk_lock == LK_UNLOCKED &&
1269 			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1270 				break;
1271 
1272 #ifdef HWPMC_HOOKS
1273 			PMC_SOFT_CALL( , , lock, failed);
1274 #endif
1275 			lock_profile_obtain_lock_failed(&lk->lock_object,
1276 			    &contested, &waittime);
1277 
1278 			/*
1279 			 * If the lock is expected to not sleep just give up
1280 			 * and return.
1281 			 */
1282 			if (LK_TRYOP(flags)) {
1283 				LOCK_LOG2(lk, "%s: %p fails the try operation",
1284 				    __func__, lk);
1285 				error = EBUSY;
1286 				break;
1287 			}
1288 
1289 			/*
1290 			 * Acquire the sleepqueue chain lock because we
1291 			 * probabilly will need to manipulate waiters flags.
1292 			 */
1293 			sleepq_lock(&lk->lock_object);
1294 			x = lk->lk_lock;
1295 
1296 			/*
1297 			 * if the lock has been released while we spun on
1298 			 * the sleepqueue chain lock just try again.
1299 			 */
1300 			if (x == LK_UNLOCKED) {
1301 				sleepq_release(&lk->lock_object);
1302 				continue;
1303 			}
1304 
1305 			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1306 			if ((x & ~v) == LK_UNLOCKED) {
1307 				v = (x & ~LK_EXCLUSIVE_SPINNERS);
1308 
1309 				/*
1310 				 * If interruptible sleeps left the exclusive
1311 				 * queue empty avoid a starvation for the
1312 				 * threads sleeping on the shared queue by
1313 				 * giving them precedence and cleaning up the
1314 				 * exclusive waiters bit anyway.
1315 				 * Please note that lk_exslpfail count may be
1316 				 * lying about the real number of waiters with
1317 				 * the LK_SLEEPFAIL flag on because they may
1318 				 * be used in conjunction with interruptible
1319 				 * sleeps so lk_exslpfail might be considered
1320 				 * an 'upper limit' bound, including the edge
1321 				 * cases.
1322 				 */
1323 				if (v & LK_EXCLUSIVE_WAITERS) {
1324 					queue = SQ_EXCLUSIVE_QUEUE;
1325 					v &= ~LK_EXCLUSIVE_WAITERS;
1326 				} else {
1327 
1328 					/*
1329 					 * Exclusive waiters sleeping with
1330 					 * LK_SLEEPFAIL on and using
1331 					 * interruptible sleeps/timeout may
1332 					 * have left spourious lk_exslpfail
1333 					 * counts on, so clean it up anyway.
1334 					 */
1335 					MPASS(v & LK_SHARED_WAITERS);
1336 					lk->lk_exslpfail = 0;
1337 					queue = SQ_SHARED_QUEUE;
1338 					v &= ~LK_SHARED_WAITERS;
1339 				}
1340 				if (queue == SQ_EXCLUSIVE_QUEUE) {
1341 					realexslp =
1342 					    sleepq_sleepcnt(&lk->lock_object,
1343 					    SQ_EXCLUSIVE_QUEUE);
1344 					if (lk->lk_exslpfail >= realexslp) {
1345 						lk->lk_exslpfail = 0;
1346 						queue = SQ_SHARED_QUEUE;
1347 						v &= ~LK_SHARED_WAITERS;
1348 						if (realexslp != 0) {
1349 							LOCK_LOG2(lk,
1350 					"%s: %p has only LK_SLEEPFAIL sleepers",
1351 							    __func__, lk);
1352 							LOCK_LOG2(lk,
1353 			"%s: %p waking up threads on the exclusive queue",
1354 							    __func__, lk);
1355 							wakeup_swapper =
1356 							    sleepq_broadcast(
1357 							    &lk->lock_object,
1358 							    SLEEPQ_LK, 0,
1359 							    SQ_EXCLUSIVE_QUEUE);
1360 						}
1361 					} else
1362 						lk->lk_exslpfail = 0;
1363 				}
1364 				if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1365 					sleepq_release(&lk->lock_object);
1366 					continue;
1367 				}
1368 				LOCK_LOG3(lk,
1369 				"%s: %p waking up all threads on the %s queue",
1370 				    __func__, lk, queue == SQ_SHARED_QUEUE ?
1371 				    "shared" : "exclusive");
1372 				wakeup_swapper |= sleepq_broadcast(
1373 				    &lk->lock_object, SLEEPQ_LK, 0, queue);
1374 
1375 				/*
1376 				 * If shared waiters have been woken up we need
1377 				 * to wait for one of them to acquire the lock
1378 				 * before to set the exclusive waiters in
1379 				 * order to avoid a deadlock.
1380 				 */
1381 				if (queue == SQ_SHARED_QUEUE) {
1382 					for (v = lk->lk_lock;
1383 					    (v & LK_SHARE) && !LK_SHARERS(v);
1384 					    v = lk->lk_lock)
1385 						cpu_spinwait();
1386 				}
1387 			}
1388 
1389 			/*
1390 			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1391 			 * fail, loop back and retry.
1392 			 */
1393 			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1394 				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1395 				    x | LK_EXCLUSIVE_WAITERS)) {
1396 					sleepq_release(&lk->lock_object);
1397 					continue;
1398 				}
1399 				LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1400 				    __func__, lk);
1401 			}
1402 
1403 			/*
1404 			 * As far as we have been unable to acquire the
1405 			 * exclusive lock and the exclusive waiters flag
1406 			 * is set, we will sleep.
1407 			 */
1408 			if (flags & LK_INTERLOCK) {
1409 				class->lc_unlock(ilk);
1410 				flags &= ~LK_INTERLOCK;
1411 			}
1412 			GIANT_SAVE();
1413 			sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1414 			    SQ_EXCLUSIVE_QUEUE);
1415 			sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1416 			GIANT_RESTORE();
1417 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1418 			    __func__, lk);
1419 		}
1420 
1421 		if (error == 0) {
1422 			lock_profile_obtain_lock_success(&lk->lock_object,
1423 			    contested, waittime, file, line);
1424 			LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1425 			    lk->lk_recurse, file, line);
1426 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1427 			    LK_TRYWIT(flags), file, line);
1428 			TD_LOCKS_INC(curthread);
1429 			STACK_SAVE(lk);
1430 		}
1431 		break;
1432 	default:
1433 		if (flags & LK_INTERLOCK)
1434 			class->lc_unlock(ilk);
1435 		panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1436 	}
1437 
1438 	if (flags & LK_INTERLOCK)
1439 		class->lc_unlock(ilk);
1440 	if (wakeup_swapper)
1441 		kick_proc0();
1442 
1443 	return (error);
1444 }
1445 
1446 void
1447 _lockmgr_disown(struct lock *lk, const char *file, int line)
1448 {
1449 	uintptr_t tid, x;
1450 
1451 	if (SCHEDULER_STOPPED())
1452 		return;
1453 
1454 	tid = (uintptr_t)curthread;
1455 	_lockmgr_assert(lk, KA_XLOCKED, file, line);
1456 
1457 	/*
1458 	 * Panic if the lock is recursed.
1459 	 */
1460 	if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1461 		panic("%s: disown a recursed lockmgr @ %s:%d\n",
1462 		    __func__,  file, line);
1463 
1464 	/*
1465 	 * If the owner is already LK_KERNPROC just skip the whole operation.
1466 	 */
1467 	if (LK_HOLDER(lk->lk_lock) != tid)
1468 		return;
1469 	lock_profile_release_lock(&lk->lock_object);
1470 	LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1471 	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1472 	TD_LOCKS_DEC(curthread);
1473 	STACK_SAVE(lk);
1474 
1475 	/*
1476 	 * In order to preserve waiters flags, just spin.
1477 	 */
1478 	for (;;) {
1479 		x = lk->lk_lock;
1480 		MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1481 		x &= LK_ALL_WAITERS;
1482 		if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1483 		    LK_KERNPROC | x))
1484 			return;
1485 		cpu_spinwait();
1486 	}
1487 }
1488 
1489 void
1490 lockmgr_printinfo(const struct lock *lk)
1491 {
1492 	struct thread *td;
1493 	uintptr_t x;
1494 
1495 	if (lk->lk_lock == LK_UNLOCKED)
1496 		printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1497 	else if (lk->lk_lock & LK_SHARE)
1498 		printf("lock type %s: SHARED (count %ju)\n",
1499 		    lk->lock_object.lo_name,
1500 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1501 	else {
1502 		td = lockmgr_xholder(lk);
1503 		if (td == (struct thread *)LK_KERNPROC)
1504 			printf("lock type %s: EXCL by KERNPROC\n",
1505 			    lk->lock_object.lo_name);
1506 		else
1507 			printf("lock type %s: EXCL by thread %p "
1508 			    "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1509 			    td, td->td_proc->p_pid, td->td_proc->p_comm,
1510 			    td->td_tid);
1511 	}
1512 
1513 	x = lk->lk_lock;
1514 	if (x & LK_EXCLUSIVE_WAITERS)
1515 		printf(" with exclusive waiters pending\n");
1516 	if (x & LK_SHARED_WAITERS)
1517 		printf(" with shared waiters pending\n");
1518 	if (x & LK_EXCLUSIVE_SPINNERS)
1519 		printf(" with exclusive spinners pending\n");
1520 
1521 	STACK_PRINT(lk);
1522 }
1523 
1524 int
1525 lockstatus(const struct lock *lk)
1526 {
1527 	uintptr_t v, x;
1528 	int ret;
1529 
1530 	ret = LK_SHARED;
1531 	x = lk->lk_lock;
1532 	v = LK_HOLDER(x);
1533 
1534 	if ((x & LK_SHARE) == 0) {
1535 		if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1536 			ret = LK_EXCLUSIVE;
1537 		else
1538 			ret = LK_EXCLOTHER;
1539 	} else if (x == LK_UNLOCKED)
1540 		ret = 0;
1541 
1542 	return (ret);
1543 }
1544 
1545 #ifdef INVARIANT_SUPPORT
1546 
1547 FEATURE(invariant_support,
1548     "Support for modules compiled with INVARIANTS option");
1549 
1550 #ifndef INVARIANTS
1551 #undef	_lockmgr_assert
1552 #endif
1553 
1554 void
1555 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1556 {
1557 	int slocked = 0;
1558 
1559 	if (panicstr != NULL)
1560 		return;
1561 	switch (what) {
1562 	case KA_SLOCKED:
1563 	case KA_SLOCKED | KA_NOTRECURSED:
1564 	case KA_SLOCKED | KA_RECURSED:
1565 		slocked = 1;
1566 	case KA_LOCKED:
1567 	case KA_LOCKED | KA_NOTRECURSED:
1568 	case KA_LOCKED | KA_RECURSED:
1569 #ifdef WITNESS
1570 
1571 		/*
1572 		 * We cannot trust WITNESS if the lock is held in exclusive
1573 		 * mode and a call to lockmgr_disown() happened.
1574 		 * Workaround this skipping the check if the lock is held in
1575 		 * exclusive mode even for the KA_LOCKED case.
1576 		 */
1577 		if (slocked || (lk->lk_lock & LK_SHARE)) {
1578 			witness_assert(&lk->lock_object, what, file, line);
1579 			break;
1580 		}
1581 #endif
1582 		if (lk->lk_lock == LK_UNLOCKED ||
1583 		    ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1584 		    (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1585 			panic("Lock %s not %slocked @ %s:%d\n",
1586 			    lk->lock_object.lo_name, slocked ? "share" : "",
1587 			    file, line);
1588 
1589 		if ((lk->lk_lock & LK_SHARE) == 0) {
1590 			if (lockmgr_recursed(lk)) {
1591 				if (what & KA_NOTRECURSED)
1592 					panic("Lock %s recursed @ %s:%d\n",
1593 					    lk->lock_object.lo_name, file,
1594 					    line);
1595 			} else if (what & KA_RECURSED)
1596 				panic("Lock %s not recursed @ %s:%d\n",
1597 				    lk->lock_object.lo_name, file, line);
1598 		}
1599 		break;
1600 	case KA_XLOCKED:
1601 	case KA_XLOCKED | KA_NOTRECURSED:
1602 	case KA_XLOCKED | KA_RECURSED:
1603 		if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1604 			panic("Lock %s not exclusively locked @ %s:%d\n",
1605 			    lk->lock_object.lo_name, file, line);
1606 		if (lockmgr_recursed(lk)) {
1607 			if (what & KA_NOTRECURSED)
1608 				panic("Lock %s recursed @ %s:%d\n",
1609 				    lk->lock_object.lo_name, file, line);
1610 		} else if (what & KA_RECURSED)
1611 			panic("Lock %s not recursed @ %s:%d\n",
1612 			    lk->lock_object.lo_name, file, line);
1613 		break;
1614 	case KA_UNLOCKED:
1615 		if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1616 			panic("Lock %s exclusively locked @ %s:%d\n",
1617 			    lk->lock_object.lo_name, file, line);
1618 		break;
1619 	default:
1620 		panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1621 		    line);
1622 	}
1623 }
1624 #endif
1625 
1626 #ifdef DDB
1627 int
1628 lockmgr_chain(struct thread *td, struct thread **ownerp)
1629 {
1630 	struct lock *lk;
1631 
1632 	lk = td->td_wchan;
1633 
1634 	if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1635 		return (0);
1636 	db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1637 	if (lk->lk_lock & LK_SHARE)
1638 		db_printf("SHARED (count %ju)\n",
1639 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1640 	else
1641 		db_printf("EXCL\n");
1642 	*ownerp = lockmgr_xholder(lk);
1643 
1644 	return (1);
1645 }
1646 
1647 static void
1648 db_show_lockmgr(const struct lock_object *lock)
1649 {
1650 	struct thread *td;
1651 	const struct lock *lk;
1652 
1653 	lk = (const struct lock *)lock;
1654 
1655 	db_printf(" state: ");
1656 	if (lk->lk_lock == LK_UNLOCKED)
1657 		db_printf("UNLOCKED\n");
1658 	else if (lk->lk_lock & LK_SHARE)
1659 		db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1660 	else {
1661 		td = lockmgr_xholder(lk);
1662 		if (td == (struct thread *)LK_KERNPROC)
1663 			db_printf("XLOCK: LK_KERNPROC\n");
1664 		else
1665 			db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1666 			    td->td_tid, td->td_proc->p_pid,
1667 			    td->td_proc->p_comm);
1668 		if (lockmgr_recursed(lk))
1669 			db_printf(" recursed: %d\n", lk->lk_recurse);
1670 	}
1671 	db_printf(" waiters: ");
1672 	switch (lk->lk_lock & LK_ALL_WAITERS) {
1673 	case LK_SHARED_WAITERS:
1674 		db_printf("shared\n");
1675 		break;
1676 	case LK_EXCLUSIVE_WAITERS:
1677 		db_printf("exclusive\n");
1678 		break;
1679 	case LK_ALL_WAITERS:
1680 		db_printf("shared and exclusive\n");
1681 		break;
1682 	default:
1683 		db_printf("none\n");
1684 	}
1685 	db_printf(" spinners: ");
1686 	if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1687 		db_printf("exclusive\n");
1688 	else
1689 		db_printf("none\n");
1690 }
1691 #endif
1692