xref: /freebsd/sys/kern/kern_lock.c (revision 907b59d76938e654f0d040a888e8dfca3de1e222)
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_hwpmc_hooks.h"
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/kdb.h>
38 #include <sys/ktr.h>
39 #include <sys/lock.h>
40 #include <sys/lock_profile.h>
41 #include <sys/lockmgr.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/sleepqueue.h>
45 #ifdef DEBUG_LOCKS
46 #include <sys/stack.h>
47 #endif
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 
51 #include <machine/cpu.h>
52 
53 #ifdef DDB
54 #include <ddb/ddb.h>
55 #endif
56 
57 #ifdef HWPMC_HOOKS
58 #include <sys/pmckern.h>
59 PMC_SOFT_DECLARE( , , lock, failed);
60 #endif
61 
62 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
63     (LK_ADAPTIVE | LK_NOSHARE));
64 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
65     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
66 
67 #define	SQ_EXCLUSIVE_QUEUE	0
68 #define	SQ_SHARED_QUEUE		1
69 
70 #ifndef INVARIANTS
71 #define	_lockmgr_assert(lk, what, file, line)
72 #endif
73 
74 #define	TD_SLOCKS_INC(td)	((td)->td_lk_slocks++)
75 #define	TD_SLOCKS_DEC(td)	((td)->td_lk_slocks--)
76 
77 #ifndef DEBUG_LOCKS
78 #define	STACK_PRINT(lk)
79 #define	STACK_SAVE(lk)
80 #define	STACK_ZERO(lk)
81 #else
82 #define	STACK_PRINT(lk)	stack_print_ddb(&(lk)->lk_stack)
83 #define	STACK_SAVE(lk)	stack_save(&(lk)->lk_stack)
84 #define	STACK_ZERO(lk)	stack_zero(&(lk)->lk_stack)
85 #endif
86 
87 #define	LOCK_LOG2(lk, string, arg1, arg2)				\
88 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
89 		CTR2(KTR_LOCK, (string), (arg1), (arg2))
90 #define	LOCK_LOG3(lk, string, arg1, arg2, arg3)				\
91 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
92 		CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
93 
94 #define	GIANT_DECLARE							\
95 	int _i = 0;							\
96 	WITNESS_SAVE_DECL(Giant)
97 #define	GIANT_RESTORE() do {						\
98 	if (_i > 0) {							\
99 		while (_i--)						\
100 			mtx_lock(&Giant);				\
101 		WITNESS_RESTORE(&Giant.lock_object, Giant);		\
102 	}								\
103 } while (0)
104 #define	GIANT_SAVE() do {						\
105 	if (mtx_owned(&Giant)) {					\
106 		WITNESS_SAVE(&Giant.lock_object, Giant);		\
107 		while (mtx_owned(&Giant)) {				\
108 			_i++;						\
109 			mtx_unlock(&Giant);				\
110 		}							\
111 	}								\
112 } while (0)
113 
114 #define	LK_CAN_SHARE(x, flags)						\
115 	(((x) & LK_SHARE) &&						\
116 	(((x) & (LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 0 ||	\
117 	(curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||	\
118 	(curthread->td_pflags & TDP_DEADLKTREAT)))
119 #define	LK_TRYOP(x)							\
120 	((x) & LK_NOWAIT)
121 
122 #define	LK_CAN_WITNESS(x)						\
123 	(((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
124 #define	LK_TRYWIT(x)							\
125 	(LK_TRYOP(x) ? LOP_TRYLOCK : 0)
126 
127 #define	LK_CAN_ADAPT(lk, f)						\
128 	(((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&		\
129 	((f) & LK_SLEEPFAIL) == 0)
130 
131 #define	lockmgr_disowned(lk)						\
132 	(((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
133 
134 #define	lockmgr_xlocked(lk)						\
135 	(((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
136 
137 static void	assert_lockmgr(const struct lock_object *lock, int how);
138 #ifdef DDB
139 static void	db_show_lockmgr(const struct lock_object *lock);
140 #endif
141 static void	lock_lockmgr(struct lock_object *lock, uintptr_t how);
142 #ifdef KDTRACE_HOOKS
143 static int	owner_lockmgr(const struct lock_object *lock,
144 		    struct thread **owner);
145 #endif
146 static uintptr_t unlock_lockmgr(struct lock_object *lock);
147 
148 struct lock_class lock_class_lockmgr = {
149 	.lc_name = "lockmgr",
150 	.lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
151 	.lc_assert = assert_lockmgr,
152 #ifdef DDB
153 	.lc_ddb_show = db_show_lockmgr,
154 #endif
155 	.lc_lock = lock_lockmgr,
156 	.lc_unlock = unlock_lockmgr,
157 #ifdef KDTRACE_HOOKS
158 	.lc_owner = owner_lockmgr,
159 #endif
160 };
161 
162 #ifdef ADAPTIVE_LOCKMGRS
163 static u_int alk_retries = 10;
164 static u_int alk_loops = 10000;
165 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL,
166     "lockmgr debugging");
167 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
168 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
169 #endif
170 
171 static __inline struct thread *
172 lockmgr_xholder(const struct lock *lk)
173 {
174 	uintptr_t x;
175 
176 	x = lk->lk_lock;
177 	return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
178 }
179 
180 /*
181  * It assumes sleepq_lock held and returns with this one unheld.
182  * It also assumes the generic interlock is sane and previously checked.
183  * If LK_INTERLOCK is specified the interlock is not reacquired after the
184  * sleep.
185  */
186 static __inline int
187 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
188     const char *wmesg, int pri, int timo, int queue)
189 {
190 	GIANT_DECLARE;
191 	struct lock_class *class;
192 	int catch, error;
193 
194 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
195 	catch = pri & PCATCH;
196 	pri &= PRIMASK;
197 	error = 0;
198 
199 	LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
200 	    (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
201 
202 	if (flags & LK_INTERLOCK)
203 		class->lc_unlock(ilk);
204 	if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
205 		lk->lk_exslpfail++;
206 	GIANT_SAVE();
207 	sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
208 	    SLEEPQ_INTERRUPTIBLE : 0), queue);
209 	if ((flags & LK_TIMELOCK) && timo)
210 		sleepq_set_timeout(&lk->lock_object, timo);
211 
212 	/*
213 	 * Decisional switch for real sleeping.
214 	 */
215 	if ((flags & LK_TIMELOCK) && timo && catch)
216 		error = sleepq_timedwait_sig(&lk->lock_object, pri);
217 	else if ((flags & LK_TIMELOCK) && timo)
218 		error = sleepq_timedwait(&lk->lock_object, pri);
219 	else if (catch)
220 		error = sleepq_wait_sig(&lk->lock_object, pri);
221 	else
222 		sleepq_wait(&lk->lock_object, pri);
223 	GIANT_RESTORE();
224 	if ((flags & LK_SLEEPFAIL) && error == 0)
225 		error = ENOLCK;
226 
227 	return (error);
228 }
229 
230 static __inline int
231 wakeupshlk(struct lock *lk, const char *file, int line)
232 {
233 	uintptr_t v, x;
234 	u_int realexslp;
235 	int queue, wakeup_swapper;
236 
237 	WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
238 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
239 
240 	wakeup_swapper = 0;
241 	for (;;) {
242 		x = lk->lk_lock;
243 
244 		/*
245 		 * If there is more than one shared lock held, just drop one
246 		 * and return.
247 		 */
248 		if (LK_SHARERS(x) > 1) {
249 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
250 			    x - LK_ONE_SHARER))
251 				break;
252 			continue;
253 		}
254 
255 		/*
256 		 * If there are not waiters on the exclusive queue, drop the
257 		 * lock quickly.
258 		 */
259 		if ((x & LK_ALL_WAITERS) == 0) {
260 			MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
261 			    LK_SHARERS_LOCK(1));
262 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
263 				break;
264 			continue;
265 		}
266 
267 		/*
268 		 * We should have a sharer with waiters, so enter the hard
269 		 * path in order to handle wakeups correctly.
270 		 */
271 		sleepq_lock(&lk->lock_object);
272 		x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
273 		v = LK_UNLOCKED;
274 
275 		/*
276 		 * If the lock has exclusive waiters, give them preference in
277 		 * order to avoid deadlock with shared runners up.
278 		 * If interruptible sleeps left the exclusive queue empty
279 		 * avoid a starvation for the threads sleeping on the shared
280 		 * queue by giving them precedence and cleaning up the
281 		 * exclusive waiters bit anyway.
282 		 * Please note that lk_exslpfail count may be lying about
283 		 * the real number of waiters with the LK_SLEEPFAIL flag on
284 		 * because they may be used in conjunction with interruptible
285 		 * sleeps so lk_exslpfail might be considered an 'upper limit'
286 		 * bound, including the edge cases.
287 		 */
288 		realexslp = sleepq_sleepcnt(&lk->lock_object,
289 		    SQ_EXCLUSIVE_QUEUE);
290 		if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
291 			if (lk->lk_exslpfail < realexslp) {
292 				lk->lk_exslpfail = 0;
293 				queue = SQ_EXCLUSIVE_QUEUE;
294 				v |= (x & LK_SHARED_WAITERS);
295 			} else {
296 				lk->lk_exslpfail = 0;
297 				LOCK_LOG2(lk,
298 				    "%s: %p has only LK_SLEEPFAIL sleepers",
299 				    __func__, lk);
300 				LOCK_LOG2(lk,
301 			    "%s: %p waking up threads on the exclusive queue",
302 				    __func__, lk);
303 				wakeup_swapper =
304 				    sleepq_broadcast(&lk->lock_object,
305 				    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
306 				queue = SQ_SHARED_QUEUE;
307 			}
308 
309 		} else {
310 
311 			/*
312 			 * Exclusive waiters sleeping with LK_SLEEPFAIL on
313 			 * and using interruptible sleeps/timeout may have
314 			 * left spourious lk_exslpfail counts on, so clean
315 			 * it up anyway.
316 			 */
317 			lk->lk_exslpfail = 0;
318 			queue = SQ_SHARED_QUEUE;
319 		}
320 
321 		if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
322 		    v)) {
323 			sleepq_release(&lk->lock_object);
324 			continue;
325 		}
326 		LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
327 		    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
328 		    "exclusive");
329 		wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
330 		    0, queue);
331 		sleepq_release(&lk->lock_object);
332 		break;
333 	}
334 
335 	lock_profile_release_lock(&lk->lock_object);
336 	TD_LOCKS_DEC(curthread);
337 	TD_SLOCKS_DEC(curthread);
338 	return (wakeup_swapper);
339 }
340 
341 static void
342 assert_lockmgr(const struct lock_object *lock, int what)
343 {
344 
345 	panic("lockmgr locks do not support assertions");
346 }
347 
348 static void
349 lock_lockmgr(struct lock_object *lock, uintptr_t how)
350 {
351 
352 	panic("lockmgr locks do not support sleep interlocking");
353 }
354 
355 static uintptr_t
356 unlock_lockmgr(struct lock_object *lock)
357 {
358 
359 	panic("lockmgr locks do not support sleep interlocking");
360 }
361 
362 #ifdef KDTRACE_HOOKS
363 static int
364 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
365 {
366 
367 	panic("lockmgr locks do not support owner inquiring");
368 }
369 #endif
370 
371 void
372 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
373 {
374 	int iflags;
375 
376 	MPASS((flags & ~LK_INIT_MASK) == 0);
377 	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
378             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
379             &lk->lk_lock));
380 
381 	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
382 	if (flags & LK_CANRECURSE)
383 		iflags |= LO_RECURSABLE;
384 	if ((flags & LK_NODUP) == 0)
385 		iflags |= LO_DUPOK;
386 	if (flags & LK_NOPROFILE)
387 		iflags |= LO_NOPROFILE;
388 	if ((flags & LK_NOWITNESS) == 0)
389 		iflags |= LO_WITNESS;
390 	if (flags & LK_QUIET)
391 		iflags |= LO_QUIET;
392 	if (flags & LK_IS_VNODE)
393 		iflags |= LO_IS_VNODE;
394 	iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
395 
396 	lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
397 	lk->lk_lock = LK_UNLOCKED;
398 	lk->lk_recurse = 0;
399 	lk->lk_exslpfail = 0;
400 	lk->lk_timo = timo;
401 	lk->lk_pri = pri;
402 	STACK_ZERO(lk);
403 }
404 
405 /*
406  * XXX: Gross hacks to manipulate external lock flags after
407  * initialization.  Used for certain vnode and buf locks.
408  */
409 void
410 lockallowshare(struct lock *lk)
411 {
412 
413 	lockmgr_assert(lk, KA_XLOCKED);
414 	lk->lock_object.lo_flags &= ~LK_NOSHARE;
415 }
416 
417 void
418 lockdisableshare(struct lock *lk)
419 {
420 
421 	lockmgr_assert(lk, KA_XLOCKED);
422 	lk->lock_object.lo_flags |= LK_NOSHARE;
423 }
424 
425 void
426 lockallowrecurse(struct lock *lk)
427 {
428 
429 	lockmgr_assert(lk, KA_XLOCKED);
430 	lk->lock_object.lo_flags |= LO_RECURSABLE;
431 }
432 
433 void
434 lockdisablerecurse(struct lock *lk)
435 {
436 
437 	lockmgr_assert(lk, KA_XLOCKED);
438 	lk->lock_object.lo_flags &= ~LO_RECURSABLE;
439 }
440 
441 void
442 lockdestroy(struct lock *lk)
443 {
444 
445 	KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
446 	KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
447 	KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
448 	lock_destroy(&lk->lock_object);
449 }
450 
451 int
452 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
453     const char *wmesg, int pri, int timo, const char *file, int line)
454 {
455 	GIANT_DECLARE;
456 	struct lock_class *class;
457 	const char *iwmesg;
458 	uintptr_t tid, v, x;
459 	u_int op, realexslp;
460 	int error, ipri, itimo, queue, wakeup_swapper;
461 #ifdef LOCK_PROFILING
462 	uint64_t waittime = 0;
463 	int contested = 0;
464 #endif
465 #ifdef ADAPTIVE_LOCKMGRS
466 	volatile struct thread *owner;
467 	u_int i, spintries = 0;
468 #endif
469 
470 	error = 0;
471 	tid = (uintptr_t)curthread;
472 	op = (flags & LK_TYPE_MASK);
473 	iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
474 	ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
475 	itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
476 
477 	MPASS((flags & ~LK_TOTAL_MASK) == 0);
478 	KASSERT((op & (op - 1)) == 0,
479 	    ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
480 	KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
481 	    (op != LK_DOWNGRADE && op != LK_RELEASE),
482 	    ("%s: Invalid flags in regard of the operation desired @ %s:%d",
483 	    __func__, file, line));
484 	KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
485 	    ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
486 	    __func__, file, line));
487 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
488 	    ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
489 	    lk->lock_object.lo_name, file, line));
490 
491 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
492 	if (panicstr != NULL) {
493 		if (flags & LK_INTERLOCK)
494 			class->lc_unlock(ilk);
495 		return (0);
496 	}
497 
498 	if (lk->lock_object.lo_flags & LK_NOSHARE) {
499 		switch (op) {
500 		case LK_SHARED:
501 			op = LK_EXCLUSIVE;
502 			break;
503 		case LK_UPGRADE:
504 		case LK_TRYUPGRADE:
505 		case LK_DOWNGRADE:
506 			_lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
507 			    file, line);
508 			if (flags & LK_INTERLOCK)
509 				class->lc_unlock(ilk);
510 			return (0);
511 		}
512 	}
513 
514 	wakeup_swapper = 0;
515 	switch (op) {
516 	case LK_SHARED:
517 		if (LK_CAN_WITNESS(flags))
518 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
519 			    file, line, flags & LK_INTERLOCK ? ilk : NULL);
520 		for (;;) {
521 			x = lk->lk_lock;
522 
523 			/*
524 			 * If no other thread has an exclusive lock, or
525 			 * no exclusive waiter is present, bump the count of
526 			 * sharers.  Since we have to preserve the state of
527 			 * waiters, if we fail to acquire the shared lock
528 			 * loop back and retry.
529 			 */
530 			if (LK_CAN_SHARE(x, flags)) {
531 				if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
532 				    x + LK_ONE_SHARER))
533 					break;
534 				continue;
535 			}
536 #ifdef HWPMC_HOOKS
537 			PMC_SOFT_CALL( , , lock, failed);
538 #endif
539 			lock_profile_obtain_lock_failed(&lk->lock_object,
540 			    &contested, &waittime);
541 
542 			/*
543 			 * If the lock is already held by curthread in
544 			 * exclusive way avoid a deadlock.
545 			 */
546 			if (LK_HOLDER(x) == tid) {
547 				LOCK_LOG2(lk,
548 				    "%s: %p already held in exclusive mode",
549 				    __func__, lk);
550 				error = EDEADLK;
551 				break;
552 			}
553 
554 			/*
555 			 * If the lock is expected to not sleep just give up
556 			 * and return.
557 			 */
558 			if (LK_TRYOP(flags)) {
559 				LOCK_LOG2(lk, "%s: %p fails the try operation",
560 				    __func__, lk);
561 				error = EBUSY;
562 				break;
563 			}
564 
565 #ifdef ADAPTIVE_LOCKMGRS
566 			/*
567 			 * If the owner is running on another CPU, spin until
568 			 * the owner stops running or the state of the lock
569 			 * changes.  We need a double-state handle here
570 			 * because for a failed acquisition the lock can be
571 			 * either held in exclusive mode or shared mode
572 			 * (for the writer starvation avoidance technique).
573 			 */
574 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
575 			    LK_HOLDER(x) != LK_KERNPROC) {
576 				owner = (struct thread *)LK_HOLDER(x);
577 				if (LOCK_LOG_TEST(&lk->lock_object, 0))
578 					CTR3(KTR_LOCK,
579 					    "%s: spinning on %p held by %p",
580 					    __func__, lk, owner);
581 				KTR_STATE1(KTR_SCHED, "thread",
582 				    sched_tdname(td), "spinning",
583 				    "lockname:\"%s\"", lk->lock_object.lo_name);
584 
585 				/*
586 				 * If we are holding also an interlock drop it
587 				 * in order to avoid a deadlock if the lockmgr
588 				 * owner is adaptively spinning on the
589 				 * interlock itself.
590 				 */
591 				if (flags & LK_INTERLOCK) {
592 					class->lc_unlock(ilk);
593 					flags &= ~LK_INTERLOCK;
594 				}
595 				GIANT_SAVE();
596 				while (LK_HOLDER(lk->lk_lock) ==
597 				    (uintptr_t)owner && TD_IS_RUNNING(owner))
598 					cpu_spinwait();
599 				KTR_STATE0(KTR_SCHED, "thread",
600 				    sched_tdname(td), "running");
601 				GIANT_RESTORE();
602 				continue;
603 			} else if (LK_CAN_ADAPT(lk, flags) &&
604 			    (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
605 			    spintries < alk_retries) {
606 				KTR_STATE1(KTR_SCHED, "thread",
607 				    sched_tdname(td), "spinning",
608 				    "lockname:\"%s\"", lk->lock_object.lo_name);
609 				if (flags & LK_INTERLOCK) {
610 					class->lc_unlock(ilk);
611 					flags &= ~LK_INTERLOCK;
612 				}
613 				GIANT_SAVE();
614 				spintries++;
615 				for (i = 0; i < alk_loops; i++) {
616 					if (LOCK_LOG_TEST(&lk->lock_object, 0))
617 						CTR4(KTR_LOCK,
618 				    "%s: shared spinning on %p with %u and %u",
619 						    __func__, lk, spintries, i);
620 					x = lk->lk_lock;
621 					if ((x & LK_SHARE) == 0 ||
622 					    LK_CAN_SHARE(x, flags) != 0)
623 						break;
624 					cpu_spinwait();
625 				}
626 				KTR_STATE0(KTR_SCHED, "thread",
627 				    sched_tdname(td), "running");
628 				GIANT_RESTORE();
629 				if (i != alk_loops)
630 					continue;
631 			}
632 #endif
633 
634 			/*
635 			 * Acquire the sleepqueue chain lock because we
636 			 * probabilly will need to manipulate waiters flags.
637 			 */
638 			sleepq_lock(&lk->lock_object);
639 			x = lk->lk_lock;
640 
641 			/*
642 			 * if the lock can be acquired in shared mode, try
643 			 * again.
644 			 */
645 			if (LK_CAN_SHARE(x, flags)) {
646 				sleepq_release(&lk->lock_object);
647 				continue;
648 			}
649 
650 #ifdef ADAPTIVE_LOCKMGRS
651 			/*
652 			 * The current lock owner might have started executing
653 			 * on another CPU (or the lock could have changed
654 			 * owner) while we were waiting on the turnstile
655 			 * chain lock.  If so, drop the turnstile lock and try
656 			 * again.
657 			 */
658 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
659 			    LK_HOLDER(x) != LK_KERNPROC) {
660 				owner = (struct thread *)LK_HOLDER(x);
661 				if (TD_IS_RUNNING(owner)) {
662 					sleepq_release(&lk->lock_object);
663 					continue;
664 				}
665 			}
666 #endif
667 
668 			/*
669 			 * Try to set the LK_SHARED_WAITERS flag.  If we fail,
670 			 * loop back and retry.
671 			 */
672 			if ((x & LK_SHARED_WAITERS) == 0) {
673 				if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
674 				    x | LK_SHARED_WAITERS)) {
675 					sleepq_release(&lk->lock_object);
676 					continue;
677 				}
678 				LOCK_LOG2(lk, "%s: %p set shared waiters flag",
679 				    __func__, lk);
680 			}
681 
682 			/*
683 			 * As far as we have been unable to acquire the
684 			 * shared lock and the shared waiters flag is set,
685 			 * we will sleep.
686 			 */
687 			error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
688 			    SQ_SHARED_QUEUE);
689 			flags &= ~LK_INTERLOCK;
690 			if (error) {
691 				LOCK_LOG3(lk,
692 				    "%s: interrupted sleep for %p with %d",
693 				    __func__, lk, error);
694 				break;
695 			}
696 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
697 			    __func__, lk);
698 		}
699 		if (error == 0) {
700 			lock_profile_obtain_lock_success(&lk->lock_object,
701 			    contested, waittime, file, line);
702 			LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
703 			    line);
704 			WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
705 			    line);
706 			TD_LOCKS_INC(curthread);
707 			TD_SLOCKS_INC(curthread);
708 			STACK_SAVE(lk);
709 		}
710 		break;
711 	case LK_UPGRADE:
712 	case LK_TRYUPGRADE:
713 		_lockmgr_assert(lk, KA_SLOCKED, file, line);
714 		v = lk->lk_lock;
715 		x = v & LK_ALL_WAITERS;
716 		v &= LK_EXCLUSIVE_SPINNERS;
717 
718 		/*
719 		 * Try to switch from one shared lock to an exclusive one.
720 		 * We need to preserve waiters flags during the operation.
721 		 */
722 		if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
723 		    tid | x)) {
724 			LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
725 			    line);
726 			WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
727 			    LK_TRYWIT(flags), file, line);
728 			TD_SLOCKS_DEC(curthread);
729 			break;
730 		}
731 
732 		/*
733 		 * In LK_TRYUPGRADE mode, do not drop the lock,
734 		 * returning EBUSY instead.
735 		 */
736 		if (op == LK_TRYUPGRADE) {
737 			LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
738 			    __func__, lk);
739 			error = EBUSY;
740 			break;
741 		}
742 
743 		/*
744 		 * We have been unable to succeed in upgrading, so just
745 		 * give up the shared lock.
746 		 */
747 		wakeup_swapper |= wakeupshlk(lk, file, line);
748 
749 		/* FALLTHROUGH */
750 	case LK_EXCLUSIVE:
751 		if (LK_CAN_WITNESS(flags))
752 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
753 			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
754 			    ilk : NULL);
755 
756 		/*
757 		 * If curthread already holds the lock and this one is
758 		 * allowed to recurse, simply recurse on it.
759 		 */
760 		if (lockmgr_xlocked(lk)) {
761 			if ((flags & LK_CANRECURSE) == 0 &&
762 			    (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
763 
764 				/*
765 				 * If the lock is expected to not panic just
766 				 * give up and return.
767 				 */
768 				if (LK_TRYOP(flags)) {
769 					LOCK_LOG2(lk,
770 					    "%s: %p fails the try operation",
771 					    __func__, lk);
772 					error = EBUSY;
773 					break;
774 				}
775 				if (flags & LK_INTERLOCK)
776 					class->lc_unlock(ilk);
777 		panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
778 				    __func__, iwmesg, file, line);
779 			}
780 			lk->lk_recurse++;
781 			LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
782 			LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
783 			    lk->lk_recurse, file, line);
784 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
785 			    LK_TRYWIT(flags), file, line);
786 			TD_LOCKS_INC(curthread);
787 			break;
788 		}
789 
790 		for (;;) {
791 			if (lk->lk_lock == LK_UNLOCKED &&
792 			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
793 				break;
794 #ifdef HWPMC_HOOKS
795 			PMC_SOFT_CALL( , , lock, failed);
796 #endif
797 			lock_profile_obtain_lock_failed(&lk->lock_object,
798 			    &contested, &waittime);
799 
800 			/*
801 			 * If the lock is expected to not sleep just give up
802 			 * and return.
803 			 */
804 			if (LK_TRYOP(flags)) {
805 				LOCK_LOG2(lk, "%s: %p fails the try operation",
806 				    __func__, lk);
807 				error = EBUSY;
808 				break;
809 			}
810 
811 #ifdef ADAPTIVE_LOCKMGRS
812 			/*
813 			 * If the owner is running on another CPU, spin until
814 			 * the owner stops running or the state of the lock
815 			 * changes.
816 			 */
817 			x = lk->lk_lock;
818 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
819 			    LK_HOLDER(x) != LK_KERNPROC) {
820 				owner = (struct thread *)LK_HOLDER(x);
821 				if (LOCK_LOG_TEST(&lk->lock_object, 0))
822 					CTR3(KTR_LOCK,
823 					    "%s: spinning on %p held by %p",
824 					    __func__, lk, owner);
825 				KTR_STATE1(KTR_SCHED, "thread",
826 				    sched_tdname(td), "spinning",
827 				    "lockname:\"%s\"", lk->lock_object.lo_name);
828 
829 				/*
830 				 * If we are holding also an interlock drop it
831 				 * in order to avoid a deadlock if the lockmgr
832 				 * owner is adaptively spinning on the
833 				 * interlock itself.
834 				 */
835 				if (flags & LK_INTERLOCK) {
836 					class->lc_unlock(ilk);
837 					flags &= ~LK_INTERLOCK;
838 				}
839 				GIANT_SAVE();
840 				while (LK_HOLDER(lk->lk_lock) ==
841 				    (uintptr_t)owner && TD_IS_RUNNING(owner))
842 					cpu_spinwait();
843 				KTR_STATE0(KTR_SCHED, "thread",
844 				    sched_tdname(td), "running");
845 				GIANT_RESTORE();
846 				continue;
847 			} else if (LK_CAN_ADAPT(lk, flags) &&
848 			    (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
849 			    spintries < alk_retries) {
850 				if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
851 				    !atomic_cmpset_ptr(&lk->lk_lock, x,
852 				    x | LK_EXCLUSIVE_SPINNERS))
853 					continue;
854 				KTR_STATE1(KTR_SCHED, "thread",
855 				    sched_tdname(td), "spinning",
856 				    "lockname:\"%s\"", lk->lock_object.lo_name);
857 				if (flags & LK_INTERLOCK) {
858 					class->lc_unlock(ilk);
859 					flags &= ~LK_INTERLOCK;
860 				}
861 				GIANT_SAVE();
862 				spintries++;
863 				for (i = 0; i < alk_loops; i++) {
864 					if (LOCK_LOG_TEST(&lk->lock_object, 0))
865 						CTR4(KTR_LOCK,
866 				    "%s: shared spinning on %p with %u and %u",
867 						    __func__, lk, spintries, i);
868 					if ((lk->lk_lock &
869 					    LK_EXCLUSIVE_SPINNERS) == 0)
870 						break;
871 					cpu_spinwait();
872 				}
873 				KTR_STATE0(KTR_SCHED, "thread",
874 				    sched_tdname(td), "running");
875 				GIANT_RESTORE();
876 				if (i != alk_loops)
877 					continue;
878 			}
879 #endif
880 
881 			/*
882 			 * Acquire the sleepqueue chain lock because we
883 			 * probabilly will need to manipulate waiters flags.
884 			 */
885 			sleepq_lock(&lk->lock_object);
886 			x = lk->lk_lock;
887 
888 			/*
889 			 * if the lock has been released while we spun on
890 			 * the sleepqueue chain lock just try again.
891 			 */
892 			if (x == LK_UNLOCKED) {
893 				sleepq_release(&lk->lock_object);
894 				continue;
895 			}
896 
897 #ifdef ADAPTIVE_LOCKMGRS
898 			/*
899 			 * The current lock owner might have started executing
900 			 * on another CPU (or the lock could have changed
901 			 * owner) while we were waiting on the turnstile
902 			 * chain lock.  If so, drop the turnstile lock and try
903 			 * again.
904 			 */
905 			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
906 			    LK_HOLDER(x) != LK_KERNPROC) {
907 				owner = (struct thread *)LK_HOLDER(x);
908 				if (TD_IS_RUNNING(owner)) {
909 					sleepq_release(&lk->lock_object);
910 					continue;
911 				}
912 			}
913 #endif
914 
915 			/*
916 			 * The lock can be in the state where there is a
917 			 * pending queue of waiters, but still no owner.
918 			 * This happens when the lock is contested and an
919 			 * owner is going to claim the lock.
920 			 * If curthread is the one successfully acquiring it
921 			 * claim lock ownership and return, preserving waiters
922 			 * flags.
923 			 */
924 			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
925 			if ((x & ~v) == LK_UNLOCKED) {
926 				v &= ~LK_EXCLUSIVE_SPINNERS;
927 				if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
928 				    tid | v)) {
929 					sleepq_release(&lk->lock_object);
930 					LOCK_LOG2(lk,
931 					    "%s: %p claimed by a new writer",
932 					    __func__, lk);
933 					break;
934 				}
935 				sleepq_release(&lk->lock_object);
936 				continue;
937 			}
938 
939 			/*
940 			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
941 			 * fail, loop back and retry.
942 			 */
943 			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
944 				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
945 				    x | LK_EXCLUSIVE_WAITERS)) {
946 					sleepq_release(&lk->lock_object);
947 					continue;
948 				}
949 				LOCK_LOG2(lk, "%s: %p set excl waiters flag",
950 				    __func__, lk);
951 			}
952 
953 			/*
954 			 * As far as we have been unable to acquire the
955 			 * exclusive lock and the exclusive waiters flag
956 			 * is set, we will sleep.
957 			 */
958 			error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
959 			    SQ_EXCLUSIVE_QUEUE);
960 			flags &= ~LK_INTERLOCK;
961 			if (error) {
962 				LOCK_LOG3(lk,
963 				    "%s: interrupted sleep for %p with %d",
964 				    __func__, lk, error);
965 				break;
966 			}
967 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
968 			    __func__, lk);
969 		}
970 		if (error == 0) {
971 			lock_profile_obtain_lock_success(&lk->lock_object,
972 			    contested, waittime, file, line);
973 			LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
974 			    lk->lk_recurse, file, line);
975 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
976 			    LK_TRYWIT(flags), file, line);
977 			TD_LOCKS_INC(curthread);
978 			STACK_SAVE(lk);
979 		}
980 		break;
981 	case LK_DOWNGRADE:
982 		_lockmgr_assert(lk, KA_XLOCKED, file, line);
983 		LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
984 		WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
985 
986 		/*
987 		 * Panic if the lock is recursed.
988 		 */
989 		if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
990 			if (flags & LK_INTERLOCK)
991 				class->lc_unlock(ilk);
992 			panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
993 			    __func__, iwmesg, file, line);
994 		}
995 		TD_SLOCKS_INC(curthread);
996 
997 		/*
998 		 * In order to preserve waiters flags, just spin.
999 		 */
1000 		for (;;) {
1001 			x = lk->lk_lock;
1002 			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1003 			x &= LK_ALL_WAITERS;
1004 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1005 			    LK_SHARERS_LOCK(1) | x))
1006 				break;
1007 			cpu_spinwait();
1008 		}
1009 		break;
1010 	case LK_RELEASE:
1011 		_lockmgr_assert(lk, KA_LOCKED, file, line);
1012 		x = lk->lk_lock;
1013 
1014 		if ((x & LK_SHARE) == 0) {
1015 
1016 			/*
1017 			 * As first option, treact the lock as if it has not
1018 			 * any waiter.
1019 			 * Fix-up the tid var if the lock has been disowned.
1020 			 */
1021 			if (LK_HOLDER(x) == LK_KERNPROC)
1022 				tid = LK_KERNPROC;
1023 			else {
1024 				WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
1025 				    file, line);
1026 				TD_LOCKS_DEC(curthread);
1027 			}
1028 			LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
1029 			    lk->lk_recurse, file, line);
1030 
1031 			/*
1032 			 * The lock is held in exclusive mode.
1033 			 * If the lock is recursed also, then unrecurse it.
1034 			 */
1035 			if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1036 				LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1037 				    lk);
1038 				lk->lk_recurse--;
1039 				break;
1040 			}
1041 			if (tid != LK_KERNPROC)
1042 				lock_profile_release_lock(&lk->lock_object);
1043 
1044 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1045 			    LK_UNLOCKED))
1046 				break;
1047 
1048 			sleepq_lock(&lk->lock_object);
1049 			x = lk->lk_lock;
1050 			v = LK_UNLOCKED;
1051 
1052 			/*
1053 		 	 * If the lock has exclusive waiters, give them
1054 			 * preference in order to avoid deadlock with
1055 			 * shared runners up.
1056 			 * If interruptible sleeps left the exclusive queue
1057 			 * empty avoid a starvation for the threads sleeping
1058 			 * on the shared queue by giving them precedence
1059 			 * and cleaning up the exclusive waiters bit anyway.
1060 			 * Please note that lk_exslpfail count may be lying
1061 			 * about the real number of waiters with the
1062 			 * LK_SLEEPFAIL flag on because they may be used in
1063 			 * conjunction with interruptible sleeps so
1064 			 * lk_exslpfail might be considered an 'upper limit'
1065 			 * bound, including the edge cases.
1066 			 */
1067 			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1068 			realexslp = sleepq_sleepcnt(&lk->lock_object,
1069 			    SQ_EXCLUSIVE_QUEUE);
1070 			if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1071 				if (lk->lk_exslpfail < realexslp) {
1072 					lk->lk_exslpfail = 0;
1073 					queue = SQ_EXCLUSIVE_QUEUE;
1074 					v |= (x & LK_SHARED_WAITERS);
1075 				} else {
1076 					lk->lk_exslpfail = 0;
1077 					LOCK_LOG2(lk,
1078 					"%s: %p has only LK_SLEEPFAIL sleepers",
1079 					    __func__, lk);
1080 					LOCK_LOG2(lk,
1081 			"%s: %p waking up threads on the exclusive queue",
1082 					    __func__, lk);
1083 					wakeup_swapper =
1084 					    sleepq_broadcast(&lk->lock_object,
1085 					    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1086 					queue = SQ_SHARED_QUEUE;
1087 				}
1088 			} else {
1089 
1090 				/*
1091 				 * Exclusive waiters sleeping with LK_SLEEPFAIL
1092 				 * on and using interruptible sleeps/timeout
1093 				 * may have left spourious lk_exslpfail counts
1094 				 * on, so clean it up anyway.
1095 				 */
1096 				lk->lk_exslpfail = 0;
1097 				queue = SQ_SHARED_QUEUE;
1098 			}
1099 
1100 			LOCK_LOG3(lk,
1101 			    "%s: %p waking up threads on the %s queue",
1102 			    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1103 			    "exclusive");
1104 			atomic_store_rel_ptr(&lk->lk_lock, v);
1105 			wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1106 			    SLEEPQ_LK, 0, queue);
1107 			sleepq_release(&lk->lock_object);
1108 			break;
1109 		} else
1110 			wakeup_swapper = wakeupshlk(lk, file, line);
1111 		break;
1112 	case LK_DRAIN:
1113 		if (LK_CAN_WITNESS(flags))
1114 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1115 			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1116 			    ilk : NULL);
1117 
1118 		/*
1119 		 * Trying to drain a lock we already own will result in a
1120 		 * deadlock.
1121 		 */
1122 		if (lockmgr_xlocked(lk)) {
1123 			if (flags & LK_INTERLOCK)
1124 				class->lc_unlock(ilk);
1125 			panic("%s: draining %s with the lock held @ %s:%d\n",
1126 			    __func__, iwmesg, file, line);
1127 		}
1128 
1129 		for (;;) {
1130 			if (lk->lk_lock == LK_UNLOCKED &&
1131 			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1132 				break;
1133 
1134 #ifdef HWPMC_HOOKS
1135 			PMC_SOFT_CALL( , , lock, failed);
1136 #endif
1137 			lock_profile_obtain_lock_failed(&lk->lock_object,
1138 			    &contested, &waittime);
1139 
1140 			/*
1141 			 * If the lock is expected to not sleep just give up
1142 			 * and return.
1143 			 */
1144 			if (LK_TRYOP(flags)) {
1145 				LOCK_LOG2(lk, "%s: %p fails the try operation",
1146 				    __func__, lk);
1147 				error = EBUSY;
1148 				break;
1149 			}
1150 
1151 			/*
1152 			 * Acquire the sleepqueue chain lock because we
1153 			 * probabilly will need to manipulate waiters flags.
1154 			 */
1155 			sleepq_lock(&lk->lock_object);
1156 			x = lk->lk_lock;
1157 
1158 			/*
1159 			 * if the lock has been released while we spun on
1160 			 * the sleepqueue chain lock just try again.
1161 			 */
1162 			if (x == LK_UNLOCKED) {
1163 				sleepq_release(&lk->lock_object);
1164 				continue;
1165 			}
1166 
1167 			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1168 			if ((x & ~v) == LK_UNLOCKED) {
1169 				v = (x & ~LK_EXCLUSIVE_SPINNERS);
1170 
1171 				/*
1172 				 * If interruptible sleeps left the exclusive
1173 				 * queue empty avoid a starvation for the
1174 				 * threads sleeping on the shared queue by
1175 				 * giving them precedence and cleaning up the
1176 				 * exclusive waiters bit anyway.
1177 				 * Please note that lk_exslpfail count may be
1178 				 * lying about the real number of waiters with
1179 				 * the LK_SLEEPFAIL flag on because they may
1180 				 * be used in conjunction with interruptible
1181 				 * sleeps so lk_exslpfail might be considered
1182 				 * an 'upper limit' bound, including the edge
1183 				 * cases.
1184 				 */
1185 				if (v & LK_EXCLUSIVE_WAITERS) {
1186 					queue = SQ_EXCLUSIVE_QUEUE;
1187 					v &= ~LK_EXCLUSIVE_WAITERS;
1188 				} else {
1189 
1190 					/*
1191 					 * Exclusive waiters sleeping with
1192 					 * LK_SLEEPFAIL on and using
1193 					 * interruptible sleeps/timeout may
1194 					 * have left spourious lk_exslpfail
1195 					 * counts on, so clean it up anyway.
1196 					 */
1197 					MPASS(v & LK_SHARED_WAITERS);
1198 					lk->lk_exslpfail = 0;
1199 					queue = SQ_SHARED_QUEUE;
1200 					v &= ~LK_SHARED_WAITERS;
1201 				}
1202 				if (queue == SQ_EXCLUSIVE_QUEUE) {
1203 					realexslp =
1204 					    sleepq_sleepcnt(&lk->lock_object,
1205 					    SQ_EXCLUSIVE_QUEUE);
1206 					if (lk->lk_exslpfail >= realexslp) {
1207 						lk->lk_exslpfail = 0;
1208 						queue = SQ_SHARED_QUEUE;
1209 						v &= ~LK_SHARED_WAITERS;
1210 						if (realexslp != 0) {
1211 							LOCK_LOG2(lk,
1212 					"%s: %p has only LK_SLEEPFAIL sleepers",
1213 							    __func__, lk);
1214 							LOCK_LOG2(lk,
1215 			"%s: %p waking up threads on the exclusive queue",
1216 							    __func__, lk);
1217 							wakeup_swapper =
1218 							    sleepq_broadcast(
1219 							    &lk->lock_object,
1220 							    SLEEPQ_LK, 0,
1221 							    SQ_EXCLUSIVE_QUEUE);
1222 						}
1223 					} else
1224 						lk->lk_exslpfail = 0;
1225 				}
1226 				if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1227 					sleepq_release(&lk->lock_object);
1228 					continue;
1229 				}
1230 				LOCK_LOG3(lk,
1231 				"%s: %p waking up all threads on the %s queue",
1232 				    __func__, lk, queue == SQ_SHARED_QUEUE ?
1233 				    "shared" : "exclusive");
1234 				wakeup_swapper |= sleepq_broadcast(
1235 				    &lk->lock_object, SLEEPQ_LK, 0, queue);
1236 
1237 				/*
1238 				 * If shared waiters have been woken up we need
1239 				 * to wait for one of them to acquire the lock
1240 				 * before to set the exclusive waiters in
1241 				 * order to avoid a deadlock.
1242 				 */
1243 				if (queue == SQ_SHARED_QUEUE) {
1244 					for (v = lk->lk_lock;
1245 					    (v & LK_SHARE) && !LK_SHARERS(v);
1246 					    v = lk->lk_lock)
1247 						cpu_spinwait();
1248 				}
1249 			}
1250 
1251 			/*
1252 			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1253 			 * fail, loop back and retry.
1254 			 */
1255 			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1256 				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1257 				    x | LK_EXCLUSIVE_WAITERS)) {
1258 					sleepq_release(&lk->lock_object);
1259 					continue;
1260 				}
1261 				LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1262 				    __func__, lk);
1263 			}
1264 
1265 			/*
1266 			 * As far as we have been unable to acquire the
1267 			 * exclusive lock and the exclusive waiters flag
1268 			 * is set, we will sleep.
1269 			 */
1270 			if (flags & LK_INTERLOCK) {
1271 				class->lc_unlock(ilk);
1272 				flags &= ~LK_INTERLOCK;
1273 			}
1274 			GIANT_SAVE();
1275 			sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1276 			    SQ_EXCLUSIVE_QUEUE);
1277 			sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1278 			GIANT_RESTORE();
1279 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1280 			    __func__, lk);
1281 		}
1282 
1283 		if (error == 0) {
1284 			lock_profile_obtain_lock_success(&lk->lock_object,
1285 			    contested, waittime, file, line);
1286 			LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1287 			    lk->lk_recurse, file, line);
1288 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1289 			    LK_TRYWIT(flags), file, line);
1290 			TD_LOCKS_INC(curthread);
1291 			STACK_SAVE(lk);
1292 		}
1293 		break;
1294 	default:
1295 		if (flags & LK_INTERLOCK)
1296 			class->lc_unlock(ilk);
1297 		panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1298 	}
1299 
1300 	if (flags & LK_INTERLOCK)
1301 		class->lc_unlock(ilk);
1302 	if (wakeup_swapper)
1303 		kick_proc0();
1304 
1305 	return (error);
1306 }
1307 
1308 void
1309 _lockmgr_disown(struct lock *lk, const char *file, int line)
1310 {
1311 	uintptr_t tid, x;
1312 
1313 	if (SCHEDULER_STOPPED())
1314 		return;
1315 
1316 	tid = (uintptr_t)curthread;
1317 	_lockmgr_assert(lk, KA_XLOCKED, file, line);
1318 
1319 	/*
1320 	 * Panic if the lock is recursed.
1321 	 */
1322 	if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1323 		panic("%s: disown a recursed lockmgr @ %s:%d\n",
1324 		    __func__,  file, line);
1325 
1326 	/*
1327 	 * If the owner is already LK_KERNPROC just skip the whole operation.
1328 	 */
1329 	if (LK_HOLDER(lk->lk_lock) != tid)
1330 		return;
1331 	lock_profile_release_lock(&lk->lock_object);
1332 	LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1333 	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1334 	TD_LOCKS_DEC(curthread);
1335 	STACK_SAVE(lk);
1336 
1337 	/*
1338 	 * In order to preserve waiters flags, just spin.
1339 	 */
1340 	for (;;) {
1341 		x = lk->lk_lock;
1342 		MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1343 		x &= LK_ALL_WAITERS;
1344 		if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1345 		    LK_KERNPROC | x))
1346 			return;
1347 		cpu_spinwait();
1348 	}
1349 }
1350 
1351 void
1352 lockmgr_printinfo(const struct lock *lk)
1353 {
1354 	struct thread *td;
1355 	uintptr_t x;
1356 
1357 	if (lk->lk_lock == LK_UNLOCKED)
1358 		printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1359 	else if (lk->lk_lock & LK_SHARE)
1360 		printf("lock type %s: SHARED (count %ju)\n",
1361 		    lk->lock_object.lo_name,
1362 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1363 	else {
1364 		td = lockmgr_xholder(lk);
1365 		if (td == (struct thread *)LK_KERNPROC)
1366 			printf("lock type %s: EXCL by KERNPROC\n",
1367 			    lk->lock_object.lo_name);
1368 		else
1369 			printf("lock type %s: EXCL by thread %p "
1370 			    "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1371 			    td, td->td_proc->p_pid, td->td_proc->p_comm,
1372 			    td->td_tid);
1373 	}
1374 
1375 	x = lk->lk_lock;
1376 	if (x & LK_EXCLUSIVE_WAITERS)
1377 		printf(" with exclusive waiters pending\n");
1378 	if (x & LK_SHARED_WAITERS)
1379 		printf(" with shared waiters pending\n");
1380 	if (x & LK_EXCLUSIVE_SPINNERS)
1381 		printf(" with exclusive spinners pending\n");
1382 
1383 	STACK_PRINT(lk);
1384 }
1385 
1386 int
1387 lockstatus(const struct lock *lk)
1388 {
1389 	uintptr_t v, x;
1390 	int ret;
1391 
1392 	ret = LK_SHARED;
1393 	x = lk->lk_lock;
1394 	v = LK_HOLDER(x);
1395 
1396 	if ((x & LK_SHARE) == 0) {
1397 		if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1398 			ret = LK_EXCLUSIVE;
1399 		else
1400 			ret = LK_EXCLOTHER;
1401 	} else if (x == LK_UNLOCKED)
1402 		ret = 0;
1403 
1404 	return (ret);
1405 }
1406 
1407 #ifdef INVARIANT_SUPPORT
1408 
1409 FEATURE(invariant_support,
1410     "Support for modules compiled with INVARIANTS option");
1411 
1412 #ifndef INVARIANTS
1413 #undef	_lockmgr_assert
1414 #endif
1415 
1416 void
1417 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1418 {
1419 	int slocked = 0;
1420 
1421 	if (panicstr != NULL)
1422 		return;
1423 	switch (what) {
1424 	case KA_SLOCKED:
1425 	case KA_SLOCKED | KA_NOTRECURSED:
1426 	case KA_SLOCKED | KA_RECURSED:
1427 		slocked = 1;
1428 	case KA_LOCKED:
1429 	case KA_LOCKED | KA_NOTRECURSED:
1430 	case KA_LOCKED | KA_RECURSED:
1431 #ifdef WITNESS
1432 
1433 		/*
1434 		 * We cannot trust WITNESS if the lock is held in exclusive
1435 		 * mode and a call to lockmgr_disown() happened.
1436 		 * Workaround this skipping the check if the lock is held in
1437 		 * exclusive mode even for the KA_LOCKED case.
1438 		 */
1439 		if (slocked || (lk->lk_lock & LK_SHARE)) {
1440 			witness_assert(&lk->lock_object, what, file, line);
1441 			break;
1442 		}
1443 #endif
1444 		if (lk->lk_lock == LK_UNLOCKED ||
1445 		    ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1446 		    (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1447 			panic("Lock %s not %slocked @ %s:%d\n",
1448 			    lk->lock_object.lo_name, slocked ? "share" : "",
1449 			    file, line);
1450 
1451 		if ((lk->lk_lock & LK_SHARE) == 0) {
1452 			if (lockmgr_recursed(lk)) {
1453 				if (what & KA_NOTRECURSED)
1454 					panic("Lock %s recursed @ %s:%d\n",
1455 					    lk->lock_object.lo_name, file,
1456 					    line);
1457 			} else if (what & KA_RECURSED)
1458 				panic("Lock %s not recursed @ %s:%d\n",
1459 				    lk->lock_object.lo_name, file, line);
1460 		}
1461 		break;
1462 	case KA_XLOCKED:
1463 	case KA_XLOCKED | KA_NOTRECURSED:
1464 	case KA_XLOCKED | KA_RECURSED:
1465 		if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1466 			panic("Lock %s not exclusively locked @ %s:%d\n",
1467 			    lk->lock_object.lo_name, file, line);
1468 		if (lockmgr_recursed(lk)) {
1469 			if (what & KA_NOTRECURSED)
1470 				panic("Lock %s recursed @ %s:%d\n",
1471 				    lk->lock_object.lo_name, file, line);
1472 		} else if (what & KA_RECURSED)
1473 			panic("Lock %s not recursed @ %s:%d\n",
1474 			    lk->lock_object.lo_name, file, line);
1475 		break;
1476 	case KA_UNLOCKED:
1477 		if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1478 			panic("Lock %s exclusively locked @ %s:%d\n",
1479 			    lk->lock_object.lo_name, file, line);
1480 		break;
1481 	default:
1482 		panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1483 		    line);
1484 	}
1485 }
1486 #endif
1487 
1488 #ifdef DDB
1489 int
1490 lockmgr_chain(struct thread *td, struct thread **ownerp)
1491 {
1492 	struct lock *lk;
1493 
1494 	lk = td->td_wchan;
1495 
1496 	if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1497 		return (0);
1498 	db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1499 	if (lk->lk_lock & LK_SHARE)
1500 		db_printf("SHARED (count %ju)\n",
1501 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1502 	else
1503 		db_printf("EXCL\n");
1504 	*ownerp = lockmgr_xholder(lk);
1505 
1506 	return (1);
1507 }
1508 
1509 static void
1510 db_show_lockmgr(const struct lock_object *lock)
1511 {
1512 	struct thread *td;
1513 	const struct lock *lk;
1514 
1515 	lk = (const struct lock *)lock;
1516 
1517 	db_printf(" state: ");
1518 	if (lk->lk_lock == LK_UNLOCKED)
1519 		db_printf("UNLOCKED\n");
1520 	else if (lk->lk_lock & LK_SHARE)
1521 		db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1522 	else {
1523 		td = lockmgr_xholder(lk);
1524 		if (td == (struct thread *)LK_KERNPROC)
1525 			db_printf("XLOCK: LK_KERNPROC\n");
1526 		else
1527 			db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1528 			    td->td_tid, td->td_proc->p_pid,
1529 			    td->td_proc->p_comm);
1530 		if (lockmgr_recursed(lk))
1531 			db_printf(" recursed: %d\n", lk->lk_recurse);
1532 	}
1533 	db_printf(" waiters: ");
1534 	switch (lk->lk_lock & LK_ALL_WAITERS) {
1535 	case LK_SHARED_WAITERS:
1536 		db_printf("shared\n");
1537 		break;
1538 	case LK_EXCLUSIVE_WAITERS:
1539 		db_printf("exclusive\n");
1540 		break;
1541 	case LK_ALL_WAITERS:
1542 		db_printf("shared and exclusive\n");
1543 		break;
1544 	default:
1545 		db_printf("none\n");
1546 	}
1547 	db_printf(" spinners: ");
1548 	if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1549 		db_printf("exclusive\n");
1550 	else
1551 		db_printf("none\n");
1552 }
1553 #endif
1554