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: stable/11/sys/kern/kern_lock.c 315375 2017-03-16 06:14:33Z mjg $");
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>
59PMC_SOFT_DECLARE( , , lock, failed);
60#endif
61
62CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
63    (LK_ADAPTIVE | LK_NOSHARE));
64CTASSERT(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
137static void	assert_lockmgr(const struct lock_object *lock, int how);
138#ifdef DDB
139static void	db_show_lockmgr(const struct lock_object *lock);
140#endif
141static void	lock_lockmgr(struct lock_object *lock, uintptr_t how);
142#ifdef KDTRACE_HOOKS
143static int	owner_lockmgr(const struct lock_object *lock,
144		    struct thread **owner);
145#endif
146static uintptr_t unlock_lockmgr(struct lock_object *lock);
147
148struct 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
163static u_int alk_retries = 10;
164static u_int alk_loops = 10000;
165static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL,
166    "lockmgr debugging");
167SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
168SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
169#endif
170
171static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp,
172    int flags);
173static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t x);
174
175static void
176lockmgr_note_shared_acquire(struct lock *lk, int contested,
177    uint64_t waittime, const char *file, int line, int flags)
178{
179
180	lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
181	    file, line);
182	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line);
183	WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line);
184	TD_LOCKS_INC(curthread);
185	TD_SLOCKS_INC(curthread);
186	STACK_SAVE(lk);
187}
188
189static void
190lockmgr_note_shared_release(struct lock *lk, const char *file, int line)
191{
192
193	lock_profile_release_lock(&lk->lock_object);
194	WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
195	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
196	TD_LOCKS_DEC(curthread);
197	TD_SLOCKS_DEC(curthread);
198}
199
200static void
201lockmgr_note_exclusive_acquire(struct lock *lk, int contested,
202    uint64_t waittime, const char *file, int line, int flags)
203{
204
205	lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
206	    file, line);
207	LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
208	WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file,
209	    line);
210	TD_LOCKS_INC(curthread);
211	STACK_SAVE(lk);
212}
213
214static void
215lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line)
216{
217
218	lock_profile_release_lock(&lk->lock_object);
219	LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file,
220	    line);
221	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
222	TD_LOCKS_DEC(curthread);
223}
224
225static void
226lockmgr_note_exclusive_upgrade(struct lock *lk, const char *file, int line,
227    int flags)
228{
229
230	LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
231	    line);
232	WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
233	    LK_TRYWIT(flags), file, line);
234	TD_SLOCKS_DEC(curthread);
235}
236
237static __inline struct thread *
238lockmgr_xholder(const struct lock *lk)
239{
240	uintptr_t x;
241
242	x = lk->lk_lock;
243	return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
244}
245
246/*
247 * It assumes sleepq_lock held and returns with this one unheld.
248 * It also assumes the generic interlock is sane and previously checked.
249 * If LK_INTERLOCK is specified the interlock is not reacquired after the
250 * sleep.
251 */
252static __inline int
253sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
254    const char *wmesg, int pri, int timo, int queue)
255{
256	GIANT_DECLARE;
257	struct lock_class *class;
258	int catch, error;
259
260	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
261	catch = pri & PCATCH;
262	pri &= PRIMASK;
263	error = 0;
264
265	LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
266	    (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
267
268	if (flags & LK_INTERLOCK)
269		class->lc_unlock(ilk);
270	if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
271		lk->lk_exslpfail++;
272	GIANT_SAVE();
273	sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
274	    SLEEPQ_INTERRUPTIBLE : 0), queue);
275	if ((flags & LK_TIMELOCK) && timo)
276		sleepq_set_timeout(&lk->lock_object, timo);
277
278	/*
279	 * Decisional switch for real sleeping.
280	 */
281	if ((flags & LK_TIMELOCK) && timo && catch)
282		error = sleepq_timedwait_sig(&lk->lock_object, pri);
283	else if ((flags & LK_TIMELOCK) && timo)
284		error = sleepq_timedwait(&lk->lock_object, pri);
285	else if (catch)
286		error = sleepq_wait_sig(&lk->lock_object, pri);
287	else
288		sleepq_wait(&lk->lock_object, pri);
289	GIANT_RESTORE();
290	if ((flags & LK_SLEEPFAIL) && error == 0)
291		error = ENOLCK;
292
293	return (error);
294}
295
296static __inline int
297wakeupshlk(struct lock *lk, const char *file, int line)
298{
299	uintptr_t v, x;
300	u_int realexslp;
301	int queue, wakeup_swapper;
302
303	wakeup_swapper = 0;
304	for (;;) {
305		x = lk->lk_lock;
306		if (lockmgr_sunlock_try(lk, x))
307			break;
308
309		/*
310		 * We should have a sharer with waiters, so enter the hard
311		 * path in order to handle wakeups correctly.
312		 */
313		sleepq_lock(&lk->lock_object);
314		x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
315		v = LK_UNLOCKED;
316
317		/*
318		 * If the lock has exclusive waiters, give them preference in
319		 * order to avoid deadlock with shared runners up.
320		 * If interruptible sleeps left the exclusive queue empty
321		 * avoid a starvation for the threads sleeping on the shared
322		 * queue by giving them precedence and cleaning up the
323		 * exclusive waiters bit anyway.
324		 * Please note that lk_exslpfail count may be lying about
325		 * the real number of waiters with the LK_SLEEPFAIL flag on
326		 * because they may be used in conjunction with interruptible
327		 * sleeps so lk_exslpfail might be considered an 'upper limit'
328		 * bound, including the edge cases.
329		 */
330		realexslp = sleepq_sleepcnt(&lk->lock_object,
331		    SQ_EXCLUSIVE_QUEUE);
332		if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
333			if (lk->lk_exslpfail < realexslp) {
334				lk->lk_exslpfail = 0;
335				queue = SQ_EXCLUSIVE_QUEUE;
336				v |= (x & LK_SHARED_WAITERS);
337			} else {
338				lk->lk_exslpfail = 0;
339				LOCK_LOG2(lk,
340				    "%s: %p has only LK_SLEEPFAIL sleepers",
341				    __func__, lk);
342				LOCK_LOG2(lk,
343			    "%s: %p waking up threads on the exclusive queue",
344				    __func__, lk);
345				wakeup_swapper =
346				    sleepq_broadcast(&lk->lock_object,
347				    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
348				queue = SQ_SHARED_QUEUE;
349			}
350
351		} else {
352
353			/*
354			 * Exclusive waiters sleeping with LK_SLEEPFAIL on
355			 * and using interruptible sleeps/timeout may have
356			 * left spourious lk_exslpfail counts on, so clean
357			 * it up anyway.
358			 */
359			lk->lk_exslpfail = 0;
360			queue = SQ_SHARED_QUEUE;
361		}
362
363		if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
364		    v)) {
365			sleepq_release(&lk->lock_object);
366			continue;
367		}
368		LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
369		    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
370		    "exclusive");
371		wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
372		    0, queue);
373		sleepq_release(&lk->lock_object);
374		break;
375	}
376
377	lockmgr_note_shared_release(lk, file, line);
378	return (wakeup_swapper);
379}
380
381static void
382assert_lockmgr(const struct lock_object *lock, int what)
383{
384
385	panic("lockmgr locks do not support assertions");
386}
387
388static void
389lock_lockmgr(struct lock_object *lock, uintptr_t how)
390{
391
392	panic("lockmgr locks do not support sleep interlocking");
393}
394
395static uintptr_t
396unlock_lockmgr(struct lock_object *lock)
397{
398
399	panic("lockmgr locks do not support sleep interlocking");
400}
401
402#ifdef KDTRACE_HOOKS
403static int
404owner_lockmgr(const struct lock_object *lock, struct thread **owner)
405{
406
407	panic("lockmgr locks do not support owner inquiring");
408}
409#endif
410
411void
412lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
413{
414	int iflags;
415
416	MPASS((flags & ~LK_INIT_MASK) == 0);
417	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
418            ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
419            &lk->lk_lock));
420
421	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
422	if (flags & LK_CANRECURSE)
423		iflags |= LO_RECURSABLE;
424	if ((flags & LK_NODUP) == 0)
425		iflags |= LO_DUPOK;
426	if (flags & LK_NOPROFILE)
427		iflags |= LO_NOPROFILE;
428	if ((flags & LK_NOWITNESS) == 0)
429		iflags |= LO_WITNESS;
430	if (flags & LK_QUIET)
431		iflags |= LO_QUIET;
432	if (flags & LK_IS_VNODE)
433		iflags |= LO_IS_VNODE;
434	iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
435
436	lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
437	lk->lk_lock = LK_UNLOCKED;
438	lk->lk_recurse = 0;
439	lk->lk_exslpfail = 0;
440	lk->lk_timo = timo;
441	lk->lk_pri = pri;
442	STACK_ZERO(lk);
443}
444
445/*
446 * XXX: Gross hacks to manipulate external lock flags after
447 * initialization.  Used for certain vnode and buf locks.
448 */
449void
450lockallowshare(struct lock *lk)
451{
452
453	lockmgr_assert(lk, KA_XLOCKED);
454	lk->lock_object.lo_flags &= ~LK_NOSHARE;
455}
456
457void
458lockdisableshare(struct lock *lk)
459{
460
461	lockmgr_assert(lk, KA_XLOCKED);
462	lk->lock_object.lo_flags |= LK_NOSHARE;
463}
464
465void
466lockallowrecurse(struct lock *lk)
467{
468
469	lockmgr_assert(lk, KA_XLOCKED);
470	lk->lock_object.lo_flags |= LO_RECURSABLE;
471}
472
473void
474lockdisablerecurse(struct lock *lk)
475{
476
477	lockmgr_assert(lk, KA_XLOCKED);
478	lk->lock_object.lo_flags &= ~LO_RECURSABLE;
479}
480
481void
482lockdestroy(struct lock *lk)
483{
484
485	KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
486	KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
487	KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
488	lock_destroy(&lk->lock_object);
489}
490
491static bool __always_inline
492lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags)
493{
494
495	/*
496	 * If no other thread has an exclusive lock, or
497	 * no exclusive waiter is present, bump the count of
498	 * sharers.  Since we have to preserve the state of
499	 * waiters, if we fail to acquire the shared lock
500	 * loop back and retry.
501	 */
502	*xp = lk->lk_lock;
503	while (LK_CAN_SHARE(*xp, flags)) {
504		if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
505		    *xp + LK_ONE_SHARER)) {
506			return (true);
507		}
508	}
509	return (false);
510}
511
512static bool __always_inline
513lockmgr_sunlock_try(struct lock *lk, uintptr_t x)
514{
515
516	for (;;) {
517		/*
518		 * If there is more than one shared lock held, just drop one
519		 * and return.
520		 */
521		if (LK_SHARERS(x) > 1) {
522			if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &x,
523			    x - LK_ONE_SHARER))
524				return (true);
525			continue;
526		}
527
528		/*
529		 * If there are not waiters on the exclusive queue, drop the
530		 * lock quickly.
531		 */
532		if ((x & LK_ALL_WAITERS) == 0) {
533			MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
534			    LK_SHARERS_LOCK(1));
535			if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &x,
536			    LK_UNLOCKED))
537				return (true);
538			continue;
539		}
540		break;
541	}
542	return (false);
543}
544
545int
546lockmgr_lock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk,
547    const char *file, int line)
548{
549	struct lock_class *class;
550	uintptr_t x, v, tid;
551	u_int op;
552	bool locked;
553
554	op = flags & LK_TYPE_MASK;
555	locked = false;
556	switch (op) {
557	case LK_SHARED:
558		if (LK_CAN_WITNESS(flags))
559			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
560			    file, line, flags & LK_INTERLOCK ? ilk : NULL);
561		if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
562			break;
563		if (lockmgr_slock_try(lk, &x, flags)) {
564			lockmgr_note_shared_acquire(lk, 0, 0,
565			    file, line, flags);
566			locked = true;
567		}
568		break;
569	case LK_EXCLUSIVE:
570		if (LK_CAN_WITNESS(flags))
571			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
572			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
573			    ilk : NULL);
574		tid = (uintptr_t)curthread;
575		if (lk->lk_lock == LK_UNLOCKED &&
576		    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
577			lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
578			    flags);
579			locked = true;
580		}
581		break;
582	case LK_UPGRADE:
583	case LK_TRYUPGRADE:
584		_lockmgr_assert(lk, KA_SLOCKED, file, line);
585		tid = (uintptr_t)curthread;
586		v = lk->lk_lock;
587		x = v & LK_ALL_WAITERS;
588		v &= LK_EXCLUSIVE_SPINNERS;
589		if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
590		    tid | x)) {
591			lockmgr_note_exclusive_upgrade(lk, file, line, flags);
592			locked = true;
593		}
594		break;
595	default:
596		break;
597	}
598	if (__predict_true(locked)) {
599		if (__predict_false(flags & LK_INTERLOCK)) {
600			class = LOCK_CLASS(ilk);
601			class->lc_unlock(ilk);
602		}
603		return (0);
604	} else {
605		return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT,
606		    LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line));
607	}
608}
609
610int
611lockmgr_unlock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk)
612{
613	struct lock_class *class;
614	uintptr_t x, tid;
615	bool unlocked;
616	const char *file;
617	int line;
618
619	file = __FILE__;
620	line = __LINE__;
621
622	_lockmgr_assert(lk, KA_LOCKED, file, line);
623	unlocked = false;
624	x = lk->lk_lock;
625	if (__predict_true(x & LK_SHARE) != 0) {
626		if (lockmgr_sunlock_try(lk, x)) {
627			lockmgr_note_shared_release(lk, file, line);
628			unlocked = true;
629		}
630	} else {
631		tid = (uintptr_t)curthread;
632		if (!lockmgr_recursed(lk) &&
633		    atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) {
634			lockmgr_note_exclusive_release(lk, file, line);
635			unlocked = true;
636		}
637	}
638	if (__predict_true(unlocked)) {
639		if (__predict_false(flags & LK_INTERLOCK)) {
640			class = LOCK_CLASS(ilk);
641			class->lc_unlock(ilk);
642		}
643		return (0);
644	} else {
645		return (__lockmgr_args(lk, flags | LK_RELEASE, ilk, LK_WMESG_DEFAULT,
646		    LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, LOCK_FILE, LOCK_LINE));
647	}
648}
649
650int
651__lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
652    const char *wmesg, int pri, int timo, const char *file, int line)
653{
654	GIANT_DECLARE;
655	struct lock_class *class;
656	const char *iwmesg;
657	uintptr_t tid, v, x;
658	u_int op, realexslp;
659	int error, ipri, itimo, queue, wakeup_swapper;
660#ifdef LOCK_PROFILING
661	uint64_t waittime = 0;
662	int contested = 0;
663#endif
664#ifdef ADAPTIVE_LOCKMGRS
665	volatile struct thread *owner;
666	u_int i, spintries = 0;
667#endif
668
669	error = 0;
670	tid = (uintptr_t)curthread;
671	op = (flags & LK_TYPE_MASK);
672	iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
673	ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
674	itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
675
676	MPASS((flags & ~LK_TOTAL_MASK) == 0);
677	KASSERT((op & (op - 1)) == 0,
678	    ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
679	KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
680	    (op != LK_DOWNGRADE && op != LK_RELEASE),
681	    ("%s: Invalid flags in regard of the operation desired @ %s:%d",
682	    __func__, file, line));
683	KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
684	    ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
685	    __func__, file, line));
686	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
687	    ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
688	    lk->lock_object.lo_name, file, line));
689
690	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
691	if (panicstr != NULL) {
692		if (flags & LK_INTERLOCK)
693			class->lc_unlock(ilk);
694		return (0);
695	}
696
697	if (lk->lock_object.lo_flags & LK_NOSHARE) {
698		switch (op) {
699		case LK_SHARED:
700			op = LK_EXCLUSIVE;
701			break;
702		case LK_UPGRADE:
703		case LK_TRYUPGRADE:
704		case LK_DOWNGRADE:
705			_lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
706			    file, line);
707			if (flags & LK_INTERLOCK)
708				class->lc_unlock(ilk);
709			return (0);
710		}
711	}
712
713	wakeup_swapper = 0;
714	switch (op) {
715	case LK_SHARED:
716		if (LK_CAN_WITNESS(flags))
717			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
718			    file, line, flags & LK_INTERLOCK ? ilk : NULL);
719		for (;;) {
720			if (lockmgr_slock_try(lk, &x, flags))
721				break;
722#ifdef HWPMC_HOOKS
723			PMC_SOFT_CALL( , , lock, failed);
724#endif
725			lock_profile_obtain_lock_failed(&lk->lock_object,
726			    &contested, &waittime);
727
728			/*
729			 * If the lock is already held by curthread in
730			 * exclusive way avoid a deadlock.
731			 */
732			if (LK_HOLDER(x) == tid) {
733				LOCK_LOG2(lk,
734				    "%s: %p already held in exclusive mode",
735				    __func__, lk);
736				error = EDEADLK;
737				break;
738			}
739
740			/*
741			 * If the lock is expected to not sleep just give up
742			 * and return.
743			 */
744			if (LK_TRYOP(flags)) {
745				LOCK_LOG2(lk, "%s: %p fails the try operation",
746				    __func__, lk);
747				error = EBUSY;
748				break;
749			}
750
751#ifdef ADAPTIVE_LOCKMGRS
752			/*
753			 * If the owner is running on another CPU, spin until
754			 * the owner stops running or the state of the lock
755			 * changes.  We need a double-state handle here
756			 * because for a failed acquisition the lock can be
757			 * either held in exclusive mode or shared mode
758			 * (for the writer starvation avoidance technique).
759			 */
760			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
761			    LK_HOLDER(x) != LK_KERNPROC) {
762				owner = (struct thread *)LK_HOLDER(x);
763				if (LOCK_LOG_TEST(&lk->lock_object, 0))
764					CTR3(KTR_LOCK,
765					    "%s: spinning on %p held by %p",
766					    __func__, lk, owner);
767				KTR_STATE1(KTR_SCHED, "thread",
768				    sched_tdname(td), "spinning",
769				    "lockname:\"%s\"", lk->lock_object.lo_name);
770
771				/*
772				 * If we are holding also an interlock drop it
773				 * in order to avoid a deadlock if the lockmgr
774				 * owner is adaptively spinning on the
775				 * interlock itself.
776				 */
777				if (flags & LK_INTERLOCK) {
778					class->lc_unlock(ilk);
779					flags &= ~LK_INTERLOCK;
780				}
781				GIANT_SAVE();
782				while (LK_HOLDER(lk->lk_lock) ==
783				    (uintptr_t)owner && TD_IS_RUNNING(owner))
784					cpu_spinwait();
785				KTR_STATE0(KTR_SCHED, "thread",
786				    sched_tdname(td), "running");
787				GIANT_RESTORE();
788				continue;
789			} else if (LK_CAN_ADAPT(lk, flags) &&
790			    (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
791			    spintries < alk_retries) {
792				KTR_STATE1(KTR_SCHED, "thread",
793				    sched_tdname(td), "spinning",
794				    "lockname:\"%s\"", lk->lock_object.lo_name);
795				if (flags & LK_INTERLOCK) {
796					class->lc_unlock(ilk);
797					flags &= ~LK_INTERLOCK;
798				}
799				GIANT_SAVE();
800				spintries++;
801				for (i = 0; i < alk_loops; i++) {
802					if (LOCK_LOG_TEST(&lk->lock_object, 0))
803						CTR4(KTR_LOCK,
804				    "%s: shared spinning on %p with %u and %u",
805						    __func__, lk, spintries, i);
806					x = lk->lk_lock;
807					if ((x & LK_SHARE) == 0 ||
808					    LK_CAN_SHARE(x, flags) != 0)
809						break;
810					cpu_spinwait();
811				}
812				KTR_STATE0(KTR_SCHED, "thread",
813				    sched_tdname(td), "running");
814				GIANT_RESTORE();
815				if (i != alk_loops)
816					continue;
817			}
818#endif
819
820			/*
821			 * Acquire the sleepqueue chain lock because we
822			 * probabilly will need to manipulate waiters flags.
823			 */
824			sleepq_lock(&lk->lock_object);
825			x = lk->lk_lock;
826
827			/*
828			 * if the lock can be acquired in shared mode, try
829			 * again.
830			 */
831			if (LK_CAN_SHARE(x, flags)) {
832				sleepq_release(&lk->lock_object);
833				continue;
834			}
835
836#ifdef ADAPTIVE_LOCKMGRS
837			/*
838			 * The current lock owner might have started executing
839			 * on another CPU (or the lock could have changed
840			 * owner) while we were waiting on the turnstile
841			 * chain lock.  If so, drop the turnstile lock and try
842			 * again.
843			 */
844			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
845			    LK_HOLDER(x) != LK_KERNPROC) {
846				owner = (struct thread *)LK_HOLDER(x);
847				if (TD_IS_RUNNING(owner)) {
848					sleepq_release(&lk->lock_object);
849					continue;
850				}
851			}
852#endif
853
854			/*
855			 * Try to set the LK_SHARED_WAITERS flag.  If we fail,
856			 * loop back and retry.
857			 */
858			if ((x & LK_SHARED_WAITERS) == 0) {
859				if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
860				    x | LK_SHARED_WAITERS)) {
861					sleepq_release(&lk->lock_object);
862					continue;
863				}
864				LOCK_LOG2(lk, "%s: %p set shared waiters flag",
865				    __func__, lk);
866			}
867
868			/*
869			 * As far as we have been unable to acquire the
870			 * shared lock and the shared waiters flag is set,
871			 * we will sleep.
872			 */
873			error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
874			    SQ_SHARED_QUEUE);
875			flags &= ~LK_INTERLOCK;
876			if (error) {
877				LOCK_LOG3(lk,
878				    "%s: interrupted sleep for %p with %d",
879				    __func__, lk, error);
880				break;
881			}
882			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
883			    __func__, lk);
884		}
885		if (error == 0) {
886#ifdef LOCK_PROFILING
887			lockmgr_note_shared_acquire(lk, contested, waittime,
888			    file, line, flags);
889#else
890			lockmgr_note_shared_acquire(lk, 0, 0, file, line,
891			    flags);
892#endif
893		}
894		break;
895	case LK_UPGRADE:
896	case LK_TRYUPGRADE:
897		_lockmgr_assert(lk, KA_SLOCKED, file, line);
898		v = lk->lk_lock;
899		x = v & LK_ALL_WAITERS;
900		v &= LK_EXCLUSIVE_SPINNERS;
901
902		/*
903		 * Try to switch from one shared lock to an exclusive one.
904		 * We need to preserve waiters flags during the operation.
905		 */
906		if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
907		    tid | x)) {
908			LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
909			    line);
910			WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
911			    LK_TRYWIT(flags), file, line);
912			TD_SLOCKS_DEC(curthread);
913			break;
914		}
915
916		/*
917		 * In LK_TRYUPGRADE mode, do not drop the lock,
918		 * returning EBUSY instead.
919		 */
920		if (op == LK_TRYUPGRADE) {
921			LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
922			    __func__, lk);
923			error = EBUSY;
924			break;
925		}
926
927		/*
928		 * We have been unable to succeed in upgrading, so just
929		 * give up the shared lock.
930		 */
931		wakeup_swapper |= wakeupshlk(lk, file, line);
932
933		/* FALLTHROUGH */
934	case LK_EXCLUSIVE:
935		if (LK_CAN_WITNESS(flags))
936			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
937			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
938			    ilk : NULL);
939
940		/*
941		 * If curthread already holds the lock and this one is
942		 * allowed to recurse, simply recurse on it.
943		 */
944		if (lockmgr_xlocked(lk)) {
945			if ((flags & LK_CANRECURSE) == 0 &&
946			    (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
947
948				/*
949				 * If the lock is expected to not panic just
950				 * give up and return.
951				 */
952				if (LK_TRYOP(flags)) {
953					LOCK_LOG2(lk,
954					    "%s: %p fails the try operation",
955					    __func__, lk);
956					error = EBUSY;
957					break;
958				}
959				if (flags & LK_INTERLOCK)
960					class->lc_unlock(ilk);
961		panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
962				    __func__, iwmesg, file, line);
963			}
964			lk->lk_recurse++;
965			LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
966			LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
967			    lk->lk_recurse, file, line);
968			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
969			    LK_TRYWIT(flags), file, line);
970			TD_LOCKS_INC(curthread);
971			break;
972		}
973
974		for (;;) {
975			if (lk->lk_lock == LK_UNLOCKED &&
976			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
977				break;
978#ifdef HWPMC_HOOKS
979			PMC_SOFT_CALL( , , lock, failed);
980#endif
981			lock_profile_obtain_lock_failed(&lk->lock_object,
982			    &contested, &waittime);
983
984			/*
985			 * If the lock is expected to not sleep just give up
986			 * and return.
987			 */
988			if (LK_TRYOP(flags)) {
989				LOCK_LOG2(lk, "%s: %p fails the try operation",
990				    __func__, lk);
991				error = EBUSY;
992				break;
993			}
994
995#ifdef ADAPTIVE_LOCKMGRS
996			/*
997			 * If the owner is running on another CPU, spin until
998			 * the owner stops running or the state of the lock
999			 * changes.
1000			 */
1001			x = lk->lk_lock;
1002			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
1003			    LK_HOLDER(x) != LK_KERNPROC) {
1004				owner = (struct thread *)LK_HOLDER(x);
1005				if (LOCK_LOG_TEST(&lk->lock_object, 0))
1006					CTR3(KTR_LOCK,
1007					    "%s: spinning on %p held by %p",
1008					    __func__, lk, owner);
1009				KTR_STATE1(KTR_SCHED, "thread",
1010				    sched_tdname(td), "spinning",
1011				    "lockname:\"%s\"", lk->lock_object.lo_name);
1012
1013				/*
1014				 * If we are holding also an interlock drop it
1015				 * in order to avoid a deadlock if the lockmgr
1016				 * owner is adaptively spinning on the
1017				 * interlock itself.
1018				 */
1019				if (flags & LK_INTERLOCK) {
1020					class->lc_unlock(ilk);
1021					flags &= ~LK_INTERLOCK;
1022				}
1023				GIANT_SAVE();
1024				while (LK_HOLDER(lk->lk_lock) ==
1025				    (uintptr_t)owner && TD_IS_RUNNING(owner))
1026					cpu_spinwait();
1027				KTR_STATE0(KTR_SCHED, "thread",
1028				    sched_tdname(td), "running");
1029				GIANT_RESTORE();
1030				continue;
1031			} else if (LK_CAN_ADAPT(lk, flags) &&
1032			    (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
1033			    spintries < alk_retries) {
1034				if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
1035				    !atomic_cmpset_ptr(&lk->lk_lock, x,
1036				    x | LK_EXCLUSIVE_SPINNERS))
1037					continue;
1038				KTR_STATE1(KTR_SCHED, "thread",
1039				    sched_tdname(td), "spinning",
1040				    "lockname:\"%s\"", lk->lock_object.lo_name);
1041				if (flags & LK_INTERLOCK) {
1042					class->lc_unlock(ilk);
1043					flags &= ~LK_INTERLOCK;
1044				}
1045				GIANT_SAVE();
1046				spintries++;
1047				for (i = 0; i < alk_loops; i++) {
1048					if (LOCK_LOG_TEST(&lk->lock_object, 0))
1049						CTR4(KTR_LOCK,
1050				    "%s: shared spinning on %p with %u and %u",
1051						    __func__, lk, spintries, i);
1052					if ((lk->lk_lock &
1053					    LK_EXCLUSIVE_SPINNERS) == 0)
1054						break;
1055					cpu_spinwait();
1056				}
1057				KTR_STATE0(KTR_SCHED, "thread",
1058				    sched_tdname(td), "running");
1059				GIANT_RESTORE();
1060				if (i != alk_loops)
1061					continue;
1062			}
1063#endif
1064
1065			/*
1066			 * Acquire the sleepqueue chain lock because we
1067			 * probabilly will need to manipulate waiters flags.
1068			 */
1069			sleepq_lock(&lk->lock_object);
1070			x = lk->lk_lock;
1071
1072			/*
1073			 * if the lock has been released while we spun on
1074			 * the sleepqueue chain lock just try again.
1075			 */
1076			if (x == LK_UNLOCKED) {
1077				sleepq_release(&lk->lock_object);
1078				continue;
1079			}
1080
1081#ifdef ADAPTIVE_LOCKMGRS
1082			/*
1083			 * The current lock owner might have started executing
1084			 * on another CPU (or the lock could have changed
1085			 * owner) while we were waiting on the turnstile
1086			 * chain lock.  If so, drop the turnstile lock and try
1087			 * again.
1088			 */
1089			if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
1090			    LK_HOLDER(x) != LK_KERNPROC) {
1091				owner = (struct thread *)LK_HOLDER(x);
1092				if (TD_IS_RUNNING(owner)) {
1093					sleepq_release(&lk->lock_object);
1094					continue;
1095				}
1096			}
1097#endif
1098
1099			/*
1100			 * The lock can be in the state where there is a
1101			 * pending queue of waiters, but still no owner.
1102			 * This happens when the lock is contested and an
1103			 * owner is going to claim the lock.
1104			 * If curthread is the one successfully acquiring it
1105			 * claim lock ownership and return, preserving waiters
1106			 * flags.
1107			 */
1108			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1109			if ((x & ~v) == LK_UNLOCKED) {
1110				v &= ~LK_EXCLUSIVE_SPINNERS;
1111				if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
1112				    tid | v)) {
1113					sleepq_release(&lk->lock_object);
1114					LOCK_LOG2(lk,
1115					    "%s: %p claimed by a new writer",
1116					    __func__, lk);
1117					break;
1118				}
1119				sleepq_release(&lk->lock_object);
1120				continue;
1121			}
1122
1123			/*
1124			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1125			 * fail, loop back and retry.
1126			 */
1127			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1128				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1129				    x | LK_EXCLUSIVE_WAITERS)) {
1130					sleepq_release(&lk->lock_object);
1131					continue;
1132				}
1133				LOCK_LOG2(lk, "%s: %p set excl waiters flag",
1134				    __func__, lk);
1135			}
1136
1137			/*
1138			 * As far as we have been unable to acquire the
1139			 * exclusive lock and the exclusive waiters flag
1140			 * is set, we will sleep.
1141			 */
1142			error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
1143			    SQ_EXCLUSIVE_QUEUE);
1144			flags &= ~LK_INTERLOCK;
1145			if (error) {
1146				LOCK_LOG3(lk,
1147				    "%s: interrupted sleep for %p with %d",
1148				    __func__, lk, error);
1149				break;
1150			}
1151			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1152			    __func__, lk);
1153		}
1154		if (error == 0) {
1155#ifdef LOCK_PROFILING
1156			lockmgr_note_exclusive_acquire(lk, contested, waittime,
1157			    file, line, flags);
1158#else
1159			lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
1160			    flags);
1161#endif
1162		}
1163		break;
1164	case LK_DOWNGRADE:
1165		_lockmgr_assert(lk, KA_XLOCKED, file, line);
1166		LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
1167		WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
1168
1169		/*
1170		 * Panic if the lock is recursed.
1171		 */
1172		if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1173			if (flags & LK_INTERLOCK)
1174				class->lc_unlock(ilk);
1175			panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
1176			    __func__, iwmesg, file, line);
1177		}
1178		TD_SLOCKS_INC(curthread);
1179
1180		/*
1181		 * In order to preserve waiters flags, just spin.
1182		 */
1183		for (;;) {
1184			x = lk->lk_lock;
1185			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1186			x &= LK_ALL_WAITERS;
1187			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1188			    LK_SHARERS_LOCK(1) | x))
1189				break;
1190			cpu_spinwait();
1191		}
1192		break;
1193	case LK_RELEASE:
1194		_lockmgr_assert(lk, KA_LOCKED, file, line);
1195		x = lk->lk_lock;
1196
1197		if ((x & LK_SHARE) == 0) {
1198
1199			/*
1200			 * As first option, treact the lock as if it has not
1201			 * any waiter.
1202			 * Fix-up the tid var if the lock has been disowned.
1203			 */
1204			if (LK_HOLDER(x) == LK_KERNPROC)
1205				tid = LK_KERNPROC;
1206			else {
1207				WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
1208				    file, line);
1209				TD_LOCKS_DEC(curthread);
1210			}
1211			LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
1212			    lk->lk_recurse, file, line);
1213
1214			/*
1215			 * The lock is held in exclusive mode.
1216			 * If the lock is recursed also, then unrecurse it.
1217			 */
1218			if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1219				LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1220				    lk);
1221				lk->lk_recurse--;
1222				break;
1223			}
1224			if (tid != LK_KERNPROC)
1225				lock_profile_release_lock(&lk->lock_object);
1226
1227			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1228			    LK_UNLOCKED))
1229				break;
1230
1231			sleepq_lock(&lk->lock_object);
1232			x = lk->lk_lock;
1233			v = LK_UNLOCKED;
1234
1235			/*
1236		 	 * If the lock has exclusive waiters, give them
1237			 * preference in order to avoid deadlock with
1238			 * shared runners up.
1239			 * If interruptible sleeps left the exclusive queue
1240			 * empty avoid a starvation for the threads sleeping
1241			 * on the shared queue by giving them precedence
1242			 * and cleaning up the exclusive waiters bit anyway.
1243			 * Please note that lk_exslpfail count may be lying
1244			 * about the real number of waiters with the
1245			 * LK_SLEEPFAIL flag on because they may be used in
1246			 * conjunction with interruptible sleeps so
1247			 * lk_exslpfail might be considered an 'upper limit'
1248			 * bound, including the edge cases.
1249			 */
1250			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1251			realexslp = sleepq_sleepcnt(&lk->lock_object,
1252			    SQ_EXCLUSIVE_QUEUE);
1253			if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1254				if (lk->lk_exslpfail < realexslp) {
1255					lk->lk_exslpfail = 0;
1256					queue = SQ_EXCLUSIVE_QUEUE;
1257					v |= (x & LK_SHARED_WAITERS);
1258				} else {
1259					lk->lk_exslpfail = 0;
1260					LOCK_LOG2(lk,
1261					"%s: %p has only LK_SLEEPFAIL sleepers",
1262					    __func__, lk);
1263					LOCK_LOG2(lk,
1264			"%s: %p waking up threads on the exclusive queue",
1265					    __func__, lk);
1266					wakeup_swapper =
1267					    sleepq_broadcast(&lk->lock_object,
1268					    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1269					queue = SQ_SHARED_QUEUE;
1270				}
1271			} else {
1272
1273				/*
1274				 * Exclusive waiters sleeping with LK_SLEEPFAIL
1275				 * on and using interruptible sleeps/timeout
1276				 * may have left spourious lk_exslpfail counts
1277				 * on, so clean it up anyway.
1278				 */
1279				lk->lk_exslpfail = 0;
1280				queue = SQ_SHARED_QUEUE;
1281			}
1282
1283			LOCK_LOG3(lk,
1284			    "%s: %p waking up threads on the %s queue",
1285			    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1286			    "exclusive");
1287			atomic_store_rel_ptr(&lk->lk_lock, v);
1288			wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1289			    SLEEPQ_LK, 0, queue);
1290			sleepq_release(&lk->lock_object);
1291			break;
1292		} else
1293			wakeup_swapper = wakeupshlk(lk, file, line);
1294		break;
1295	case LK_DRAIN:
1296		if (LK_CAN_WITNESS(flags))
1297			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1298			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1299			    ilk : NULL);
1300
1301		/*
1302		 * Trying to drain a lock we already own will result in a
1303		 * deadlock.
1304		 */
1305		if (lockmgr_xlocked(lk)) {
1306			if (flags & LK_INTERLOCK)
1307				class->lc_unlock(ilk);
1308			panic("%s: draining %s with the lock held @ %s:%d\n",
1309			    __func__, iwmesg, file, line);
1310		}
1311
1312		for (;;) {
1313			if (lk->lk_lock == LK_UNLOCKED &&
1314			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1315				break;
1316
1317#ifdef HWPMC_HOOKS
1318			PMC_SOFT_CALL( , , lock, failed);
1319#endif
1320			lock_profile_obtain_lock_failed(&lk->lock_object,
1321			    &contested, &waittime);
1322
1323			/*
1324			 * If the lock is expected to not sleep just give up
1325			 * and return.
1326			 */
1327			if (LK_TRYOP(flags)) {
1328				LOCK_LOG2(lk, "%s: %p fails the try operation",
1329				    __func__, lk);
1330				error = EBUSY;
1331				break;
1332			}
1333
1334			/*
1335			 * Acquire the sleepqueue chain lock because we
1336			 * probabilly will need to manipulate waiters flags.
1337			 */
1338			sleepq_lock(&lk->lock_object);
1339			x = lk->lk_lock;
1340
1341			/*
1342			 * if the lock has been released while we spun on
1343			 * the sleepqueue chain lock just try again.
1344			 */
1345			if (x == LK_UNLOCKED) {
1346				sleepq_release(&lk->lock_object);
1347				continue;
1348			}
1349
1350			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1351			if ((x & ~v) == LK_UNLOCKED) {
1352				v = (x & ~LK_EXCLUSIVE_SPINNERS);
1353
1354				/*
1355				 * If interruptible sleeps left the exclusive
1356				 * queue empty avoid a starvation for the
1357				 * threads sleeping on the shared queue by
1358				 * giving them precedence and cleaning up the
1359				 * exclusive waiters bit anyway.
1360				 * Please note that lk_exslpfail count may be
1361				 * lying about the real number of waiters with
1362				 * the LK_SLEEPFAIL flag on because they may
1363				 * be used in conjunction with interruptible
1364				 * sleeps so lk_exslpfail might be considered
1365				 * an 'upper limit' bound, including the edge
1366				 * cases.
1367				 */
1368				if (v & LK_EXCLUSIVE_WAITERS) {
1369					queue = SQ_EXCLUSIVE_QUEUE;
1370					v &= ~LK_EXCLUSIVE_WAITERS;
1371				} else {
1372
1373					/*
1374					 * Exclusive waiters sleeping with
1375					 * LK_SLEEPFAIL on and using
1376					 * interruptible sleeps/timeout may
1377					 * have left spourious lk_exslpfail
1378					 * counts on, so clean it up anyway.
1379					 */
1380					MPASS(v & LK_SHARED_WAITERS);
1381					lk->lk_exslpfail = 0;
1382					queue = SQ_SHARED_QUEUE;
1383					v &= ~LK_SHARED_WAITERS;
1384				}
1385				if (queue == SQ_EXCLUSIVE_QUEUE) {
1386					realexslp =
1387					    sleepq_sleepcnt(&lk->lock_object,
1388					    SQ_EXCLUSIVE_QUEUE);
1389					if (lk->lk_exslpfail >= realexslp) {
1390						lk->lk_exslpfail = 0;
1391						queue = SQ_SHARED_QUEUE;
1392						v &= ~LK_SHARED_WAITERS;
1393						if (realexslp != 0) {
1394							LOCK_LOG2(lk,
1395					"%s: %p has only LK_SLEEPFAIL sleepers",
1396							    __func__, lk);
1397							LOCK_LOG2(lk,
1398			"%s: %p waking up threads on the exclusive queue",
1399							    __func__, lk);
1400							wakeup_swapper =
1401							    sleepq_broadcast(
1402							    &lk->lock_object,
1403							    SLEEPQ_LK, 0,
1404							    SQ_EXCLUSIVE_QUEUE);
1405						}
1406					} else
1407						lk->lk_exslpfail = 0;
1408				}
1409				if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1410					sleepq_release(&lk->lock_object);
1411					continue;
1412				}
1413				LOCK_LOG3(lk,
1414				"%s: %p waking up all threads on the %s queue",
1415				    __func__, lk, queue == SQ_SHARED_QUEUE ?
1416				    "shared" : "exclusive");
1417				wakeup_swapper |= sleepq_broadcast(
1418				    &lk->lock_object, SLEEPQ_LK, 0, queue);
1419
1420				/*
1421				 * If shared waiters have been woken up we need
1422				 * to wait for one of them to acquire the lock
1423				 * before to set the exclusive waiters in
1424				 * order to avoid a deadlock.
1425				 */
1426				if (queue == SQ_SHARED_QUEUE) {
1427					for (v = lk->lk_lock;
1428					    (v & LK_SHARE) && !LK_SHARERS(v);
1429					    v = lk->lk_lock)
1430						cpu_spinwait();
1431				}
1432			}
1433
1434			/*
1435			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1436			 * fail, loop back and retry.
1437			 */
1438			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1439				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1440				    x | LK_EXCLUSIVE_WAITERS)) {
1441					sleepq_release(&lk->lock_object);
1442					continue;
1443				}
1444				LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1445				    __func__, lk);
1446			}
1447
1448			/*
1449			 * As far as we have been unable to acquire the
1450			 * exclusive lock and the exclusive waiters flag
1451			 * is set, we will sleep.
1452			 */
1453			if (flags & LK_INTERLOCK) {
1454				class->lc_unlock(ilk);
1455				flags &= ~LK_INTERLOCK;
1456			}
1457			GIANT_SAVE();
1458			sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1459			    SQ_EXCLUSIVE_QUEUE);
1460			sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1461			GIANT_RESTORE();
1462			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1463			    __func__, lk);
1464		}
1465
1466		if (error == 0) {
1467			lock_profile_obtain_lock_success(&lk->lock_object,
1468			    contested, waittime, file, line);
1469			LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1470			    lk->lk_recurse, file, line);
1471			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1472			    LK_TRYWIT(flags), file, line);
1473			TD_LOCKS_INC(curthread);
1474			STACK_SAVE(lk);
1475		}
1476		break;
1477	default:
1478		if (flags & LK_INTERLOCK)
1479			class->lc_unlock(ilk);
1480		panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1481	}
1482
1483	if (flags & LK_INTERLOCK)
1484		class->lc_unlock(ilk);
1485	if (wakeup_swapper)
1486		kick_proc0();
1487
1488	return (error);
1489}
1490
1491void
1492_lockmgr_disown(struct lock *lk, const char *file, int line)
1493{
1494	uintptr_t tid, x;
1495
1496	if (SCHEDULER_STOPPED())
1497		return;
1498
1499	tid = (uintptr_t)curthread;
1500	_lockmgr_assert(lk, KA_XLOCKED, file, line);
1501
1502	/*
1503	 * Panic if the lock is recursed.
1504	 */
1505	if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1506		panic("%s: disown a recursed lockmgr @ %s:%d\n",
1507		    __func__,  file, line);
1508
1509	/*
1510	 * If the owner is already LK_KERNPROC just skip the whole operation.
1511	 */
1512	if (LK_HOLDER(lk->lk_lock) != tid)
1513		return;
1514	lock_profile_release_lock(&lk->lock_object);
1515	LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1516	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1517	TD_LOCKS_DEC(curthread);
1518	STACK_SAVE(lk);
1519
1520	/*
1521	 * In order to preserve waiters flags, just spin.
1522	 */
1523	for (;;) {
1524		x = lk->lk_lock;
1525		MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1526		x &= LK_ALL_WAITERS;
1527		if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1528		    LK_KERNPROC | x))
1529			return;
1530		cpu_spinwait();
1531	}
1532}
1533
1534void
1535lockmgr_printinfo(const struct lock *lk)
1536{
1537	struct thread *td;
1538	uintptr_t x;
1539
1540	if (lk->lk_lock == LK_UNLOCKED)
1541		printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1542	else if (lk->lk_lock & LK_SHARE)
1543		printf("lock type %s: SHARED (count %ju)\n",
1544		    lk->lock_object.lo_name,
1545		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1546	else {
1547		td = lockmgr_xholder(lk);
1548		if (td == (struct thread *)LK_KERNPROC)
1549			printf("lock type %s: EXCL by KERNPROC\n",
1550			    lk->lock_object.lo_name);
1551		else
1552			printf("lock type %s: EXCL by thread %p "
1553			    "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1554			    td, td->td_proc->p_pid, td->td_proc->p_comm,
1555			    td->td_tid);
1556	}
1557
1558	x = lk->lk_lock;
1559	if (x & LK_EXCLUSIVE_WAITERS)
1560		printf(" with exclusive waiters pending\n");
1561	if (x & LK_SHARED_WAITERS)
1562		printf(" with shared waiters pending\n");
1563	if (x & LK_EXCLUSIVE_SPINNERS)
1564		printf(" with exclusive spinners pending\n");
1565
1566	STACK_PRINT(lk);
1567}
1568
1569int
1570lockstatus(const struct lock *lk)
1571{
1572	uintptr_t v, x;
1573	int ret;
1574
1575	ret = LK_SHARED;
1576	x = lk->lk_lock;
1577	v = LK_HOLDER(x);
1578
1579	if ((x & LK_SHARE) == 0) {
1580		if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1581			ret = LK_EXCLUSIVE;
1582		else
1583			ret = LK_EXCLOTHER;
1584	} else if (x == LK_UNLOCKED)
1585		ret = 0;
1586
1587	return (ret);
1588}
1589
1590#ifdef INVARIANT_SUPPORT
1591
1592FEATURE(invariant_support,
1593    "Support for modules compiled with INVARIANTS option");
1594
1595#ifndef INVARIANTS
1596#undef	_lockmgr_assert
1597#endif
1598
1599void
1600_lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1601{
1602	int slocked = 0;
1603
1604	if (panicstr != NULL)
1605		return;
1606	switch (what) {
1607	case KA_SLOCKED:
1608	case KA_SLOCKED | KA_NOTRECURSED:
1609	case KA_SLOCKED | KA_RECURSED:
1610		slocked = 1;
1611	case KA_LOCKED:
1612	case KA_LOCKED | KA_NOTRECURSED:
1613	case KA_LOCKED | KA_RECURSED:
1614#ifdef WITNESS
1615
1616		/*
1617		 * We cannot trust WITNESS if the lock is held in exclusive
1618		 * mode and a call to lockmgr_disown() happened.
1619		 * Workaround this skipping the check if the lock is held in
1620		 * exclusive mode even for the KA_LOCKED case.
1621		 */
1622		if (slocked || (lk->lk_lock & LK_SHARE)) {
1623			witness_assert(&lk->lock_object, what, file, line);
1624			break;
1625		}
1626#endif
1627		if (lk->lk_lock == LK_UNLOCKED ||
1628		    ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1629		    (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1630			panic("Lock %s not %slocked @ %s:%d\n",
1631			    lk->lock_object.lo_name, slocked ? "share" : "",
1632			    file, line);
1633
1634		if ((lk->lk_lock & LK_SHARE) == 0) {
1635			if (lockmgr_recursed(lk)) {
1636				if (what & KA_NOTRECURSED)
1637					panic("Lock %s recursed @ %s:%d\n",
1638					    lk->lock_object.lo_name, file,
1639					    line);
1640			} else if (what & KA_RECURSED)
1641				panic("Lock %s not recursed @ %s:%d\n",
1642				    lk->lock_object.lo_name, file, line);
1643		}
1644		break;
1645	case KA_XLOCKED:
1646	case KA_XLOCKED | KA_NOTRECURSED:
1647	case KA_XLOCKED | KA_RECURSED:
1648		if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1649			panic("Lock %s not exclusively locked @ %s:%d\n",
1650			    lk->lock_object.lo_name, file, line);
1651		if (lockmgr_recursed(lk)) {
1652			if (what & KA_NOTRECURSED)
1653				panic("Lock %s recursed @ %s:%d\n",
1654				    lk->lock_object.lo_name, file, line);
1655		} else if (what & KA_RECURSED)
1656			panic("Lock %s not recursed @ %s:%d\n",
1657			    lk->lock_object.lo_name, file, line);
1658		break;
1659	case KA_UNLOCKED:
1660		if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1661			panic("Lock %s exclusively locked @ %s:%d\n",
1662			    lk->lock_object.lo_name, file, line);
1663		break;
1664	default:
1665		panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1666		    line);
1667	}
1668}
1669#endif
1670
1671#ifdef DDB
1672int
1673lockmgr_chain(struct thread *td, struct thread **ownerp)
1674{
1675	struct lock *lk;
1676
1677	lk = td->td_wchan;
1678
1679	if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1680		return (0);
1681	db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1682	if (lk->lk_lock & LK_SHARE)
1683		db_printf("SHARED (count %ju)\n",
1684		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1685	else
1686		db_printf("EXCL\n");
1687	*ownerp = lockmgr_xholder(lk);
1688
1689	return (1);
1690}
1691
1692static void
1693db_show_lockmgr(const struct lock_object *lock)
1694{
1695	struct thread *td;
1696	const struct lock *lk;
1697
1698	lk = (const struct lock *)lock;
1699
1700	db_printf(" state: ");
1701	if (lk->lk_lock == LK_UNLOCKED)
1702		db_printf("UNLOCKED\n");
1703	else if (lk->lk_lock & LK_SHARE)
1704		db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1705	else {
1706		td = lockmgr_xholder(lk);
1707		if (td == (struct thread *)LK_KERNPROC)
1708			db_printf("XLOCK: LK_KERNPROC\n");
1709		else
1710			db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1711			    td->td_tid, td->td_proc->p_pid,
1712			    td->td_proc->p_comm);
1713		if (lockmgr_recursed(lk))
1714			db_printf(" recursed: %d\n", lk->lk_recurse);
1715	}
1716	db_printf(" waiters: ");
1717	switch (lk->lk_lock & LK_ALL_WAITERS) {
1718	case LK_SHARED_WAITERS:
1719		db_printf("shared\n");
1720		break;
1721	case LK_EXCLUSIVE_WAITERS:
1722		db_printf("exclusive\n");
1723		break;
1724	case LK_ALL_WAITERS:
1725		db_printf("shared and exclusive\n");
1726		break;
1727	default:
1728		db_printf("none\n");
1729	}
1730	db_printf(" spinners: ");
1731	if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1732		db_printf("exclusive\n");
1733	else
1734		db_printf("none\n");
1735}
1736#endif
1737