xref: /freebsd/sys/kern/kern_lock.c (revision fdafd315)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice(s), this list of conditions and the following disclaimer as
12  *    the first lines of this file unmodified other than the possible
13  *    addition of one or more copyright notices.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice(s), this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
28  * DAMAGE.
29  */
30 
31 #include "opt_ddb.h"
32 #include "opt_hwpmc_hooks.h"
33 
34 #include <sys/param.h>
35 #include <sys/kdb.h>
36 #include <sys/ktr.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/lock_profile.h>
40 #include <sys/lockmgr.h>
41 #include <sys/lockstat.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 /*
63  * Hack. There should be prio_t or similar so that this is not necessary.
64  */
65 _Static_assert((PRILASTFLAG * 2) - 1 <= USHRT_MAX,
66     "prio flags wont fit in u_short pri in struct lock");
67 
68 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
69     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
70 
71 #define	SQ_EXCLUSIVE_QUEUE	0
72 #define	SQ_SHARED_QUEUE		1
73 
74 #ifndef INVARIANTS
75 #define	_lockmgr_assert(lk, what, file, line)
76 #endif
77 
78 #define	TD_SLOCKS_INC(td)	((td)->td_lk_slocks++)
79 #define	TD_SLOCKS_DEC(td)	((td)->td_lk_slocks--)
80 
81 #ifndef DEBUG_LOCKS
82 #define	STACK_PRINT(lk)
83 #define	STACK_SAVE(lk)
84 #define	STACK_ZERO(lk)
85 #else
86 #define	STACK_PRINT(lk)	stack_print_ddb(&(lk)->lk_stack)
87 #define	STACK_SAVE(lk)	stack_save(&(lk)->lk_stack)
88 #define	STACK_ZERO(lk)	stack_zero(&(lk)->lk_stack)
89 #endif
90 
91 #define	LOCK_LOG2(lk, string, arg1, arg2)				\
92 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
93 		CTR2(KTR_LOCK, (string), (arg1), (arg2))
94 #define	LOCK_LOG3(lk, string, arg1, arg2, arg3)				\
95 	if (LOCK_LOG_TEST(&(lk)->lock_object, 0))			\
96 		CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
97 
98 #define	GIANT_DECLARE							\
99 	int _i = 0;							\
100 	WITNESS_SAVE_DECL(Giant)
101 #define	GIANT_RESTORE() do {						\
102 	if (__predict_false(_i > 0)) {					\
103 		while (_i--)						\
104 			mtx_lock(&Giant);				\
105 		WITNESS_RESTORE(&Giant.lock_object, Giant);		\
106 	}								\
107 } while (0)
108 #define	GIANT_SAVE() do {						\
109 	if (__predict_false(mtx_owned(&Giant))) {			\
110 		WITNESS_SAVE(&Giant.lock_object, Giant);		\
111 		while (mtx_owned(&Giant)) {				\
112 			_i++;						\
113 			mtx_unlock(&Giant);				\
114 		}							\
115 	}								\
116 } while (0)
117 
118 static bool __always_inline
LK_CAN_SHARE(uintptr_t x,int flags,bool fp)119 LK_CAN_SHARE(uintptr_t x, int flags, bool fp)
120 {
121 
122 	if ((x & (LK_SHARE | LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) ==
123 	    LK_SHARE)
124 		return (true);
125 	if (fp || (!(x & LK_SHARE)))
126 		return (false);
127 	if ((curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||
128 	    (curthread->td_pflags & TDP_DEADLKTREAT))
129 		return (true);
130 	return (false);
131 }
132 
133 #define	LK_TRYOP(x)							\
134 	((x) & LK_NOWAIT)
135 
136 #define	LK_CAN_WITNESS(x)						\
137 	(((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
138 #define	LK_TRYWIT(x)							\
139 	(LK_TRYOP(x) ? LOP_TRYLOCK : 0)
140 
141 #define	lockmgr_disowned(lk)						\
142 	(((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
143 
144 #define	lockmgr_xlocked_v(v)						\
145 	(((v) & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
146 
147 #define	lockmgr_xlocked(lk) lockmgr_xlocked_v(lockmgr_read_value(lk))
148 
149 static void	assert_lockmgr(const struct lock_object *lock, int how);
150 #ifdef DDB
151 static void	db_show_lockmgr(const struct lock_object *lock);
152 #endif
153 static void	lock_lockmgr(struct lock_object *lock, uintptr_t how);
154 #ifdef KDTRACE_HOOKS
155 static int	owner_lockmgr(const struct lock_object *lock,
156 		    struct thread **owner);
157 #endif
158 static uintptr_t unlock_lockmgr(struct lock_object *lock);
159 
160 struct lock_class lock_class_lockmgr = {
161 	.lc_name = "lockmgr",
162 	.lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
163 	.lc_assert = assert_lockmgr,
164 #ifdef DDB
165 	.lc_ddb_show = db_show_lockmgr,
166 #endif
167 	.lc_lock = lock_lockmgr,
168 	.lc_unlock = unlock_lockmgr,
169 #ifdef KDTRACE_HOOKS
170 	.lc_owner = owner_lockmgr,
171 #endif
172 };
173 
174 static __read_mostly bool lk_adaptive = true;
175 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging");
176 SYSCTL_BOOL(_debug_lockmgr, OID_AUTO, adaptive_spinning, CTLFLAG_RW, &lk_adaptive,
177     0, "");
178 #define lockmgr_delay  locks_delay
179 
180 struct lockmgr_wait {
181 	const char *iwmesg;
182 	int ipri;
183 	int itimo;
184 };
185 
186 static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp,
187     int flags, bool fp);
188 static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp);
189 
190 static void
lockmgr_exit(u_int flags,struct lock_object * ilk,int wakeup_swapper)191 lockmgr_exit(u_int flags, struct lock_object *ilk, int wakeup_swapper)
192 {
193 	struct lock_class *class;
194 
195 	if (flags & LK_INTERLOCK) {
196 		class = LOCK_CLASS(ilk);
197 		class->lc_unlock(ilk);
198 	}
199 
200 	if (__predict_false(wakeup_swapper))
201 		kick_proc0();
202 }
203 
204 static void
lockmgr_note_shared_acquire(struct lock * lk,int contested,uint64_t waittime,const char * file,int line,int flags)205 lockmgr_note_shared_acquire(struct lock *lk, int contested,
206     uint64_t waittime, const char *file, int line, int flags)
207 {
208 
209 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested,
210 	    waittime, file, line, LOCKSTAT_READER);
211 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line);
212 	WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line);
213 	TD_LOCKS_INC(curthread);
214 	TD_SLOCKS_INC(curthread);
215 	STACK_SAVE(lk);
216 }
217 
218 static void
lockmgr_note_shared_release(struct lock * lk,const char * file,int line)219 lockmgr_note_shared_release(struct lock *lk, const char *file, int line)
220 {
221 
222 	WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
223 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
224 	TD_LOCKS_DEC(curthread);
225 	TD_SLOCKS_DEC(curthread);
226 }
227 
228 static void
lockmgr_note_exclusive_acquire(struct lock * lk,int contested,uint64_t waittime,const char * file,int line,int flags)229 lockmgr_note_exclusive_acquire(struct lock *lk, int contested,
230     uint64_t waittime, const char *file, int line, int flags)
231 {
232 
233 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested,
234 	    waittime, file, line, LOCKSTAT_WRITER);
235 	LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
236 	WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file,
237 	    line);
238 	TD_LOCKS_INC(curthread);
239 	STACK_SAVE(lk);
240 }
241 
242 static void
lockmgr_note_exclusive_release(struct lock * lk,const char * file,int line)243 lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line)
244 {
245 
246 	if (LK_HOLDER(lockmgr_read_value(lk)) != LK_KERNPROC) {
247 		WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
248 		TD_LOCKS_DEC(curthread);
249 	}
250 	LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file,
251 	    line);
252 }
253 
254 static __inline struct thread *
lockmgr_xholder(const struct lock * lk)255 lockmgr_xholder(const struct lock *lk)
256 {
257 	uintptr_t x;
258 
259 	x = lockmgr_read_value(lk);
260 	return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
261 }
262 
263 /*
264  * It assumes sleepq_lock held and returns with this one unheld.
265  * It also assumes the generic interlock is sane and previously checked.
266  * If LK_INTERLOCK is specified the interlock is not reacquired after the
267  * sleep.
268  */
269 static __inline int
sleeplk(struct lock * lk,u_int flags,struct lock_object * ilk,const char * wmesg,int pri,int timo,int queue)270 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
271     const char *wmesg, int pri, int timo, int queue)
272 {
273 	GIANT_DECLARE;
274 	struct lock_class *class;
275 	int catch, error;
276 
277 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
278 	catch = pri & PCATCH;
279 	pri &= PRIMASK;
280 	error = 0;
281 
282 	LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
283 	    (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
284 
285 	if (flags & LK_INTERLOCK)
286 		class->lc_unlock(ilk);
287 	if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0) {
288 		if (lk->lk_exslpfail < USHRT_MAX)
289 			lk->lk_exslpfail++;
290 	}
291 	GIANT_SAVE();
292 	sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
293 	    SLEEPQ_INTERRUPTIBLE : 0), queue);
294 	if ((flags & LK_TIMELOCK) && timo)
295 		sleepq_set_timeout(&lk->lock_object, timo);
296 
297 	/*
298 	 * Decisional switch for real sleeping.
299 	 */
300 	if ((flags & LK_TIMELOCK) && timo && catch)
301 		error = sleepq_timedwait_sig(&lk->lock_object, pri);
302 	else if ((flags & LK_TIMELOCK) && timo)
303 		error = sleepq_timedwait(&lk->lock_object, pri);
304 	else if (catch)
305 		error = sleepq_wait_sig(&lk->lock_object, pri);
306 	else
307 		sleepq_wait(&lk->lock_object, pri);
308 	GIANT_RESTORE();
309 	if ((flags & LK_SLEEPFAIL) && error == 0)
310 		error = ENOLCK;
311 
312 	return (error);
313 }
314 
315 static __inline int
wakeupshlk(struct lock * lk,const char * file,int line)316 wakeupshlk(struct lock *lk, const char *file, int line)
317 {
318 	uintptr_t v, x, orig_x;
319 	u_int realexslp;
320 	int queue, wakeup_swapper;
321 
322 	wakeup_swapper = 0;
323 	for (;;) {
324 		x = lockmgr_read_value(lk);
325 		if (lockmgr_sunlock_try(lk, &x))
326 			break;
327 
328 		/*
329 		 * We should have a sharer with waiters, so enter the hard
330 		 * path in order to handle wakeups correctly.
331 		 */
332 		sleepq_lock(&lk->lock_object);
333 		orig_x = lockmgr_read_value(lk);
334 retry_sleepq:
335 		x = orig_x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
336 		v = LK_UNLOCKED;
337 
338 		/*
339 		 * If the lock has exclusive waiters, give them preference in
340 		 * order to avoid deadlock with shared runners up.
341 		 * If interruptible sleeps left the exclusive queue empty
342 		 * avoid a starvation for the threads sleeping on the shared
343 		 * queue by giving them precedence and cleaning up the
344 		 * exclusive waiters bit anyway.
345 		 * Please note that lk_exslpfail count may be lying about
346 		 * the real number of waiters with the LK_SLEEPFAIL flag on
347 		 * because they may be used in conjunction with interruptible
348 		 * sleeps so lk_exslpfail might be considered an 'upper limit'
349 		 * bound, including the edge cases.
350 		 */
351 		realexslp = sleepq_sleepcnt(&lk->lock_object,
352 		    SQ_EXCLUSIVE_QUEUE);
353 		if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
354 			if (lk->lk_exslpfail != USHRT_MAX && lk->lk_exslpfail < realexslp) {
355 				lk->lk_exslpfail = 0;
356 				queue = SQ_EXCLUSIVE_QUEUE;
357 				v |= (x & LK_SHARED_WAITERS);
358 			} else {
359 				lk->lk_exslpfail = 0;
360 				LOCK_LOG2(lk,
361 				    "%s: %p has only LK_SLEEPFAIL sleepers",
362 				    __func__, lk);
363 				LOCK_LOG2(lk,
364 			    "%s: %p waking up threads on the exclusive queue",
365 				    __func__, lk);
366 				wakeup_swapper =
367 				    sleepq_broadcast(&lk->lock_object,
368 				    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
369 				queue = SQ_SHARED_QUEUE;
370 			}
371 		} else {
372 			/*
373 			 * Exclusive waiters sleeping with LK_SLEEPFAIL on
374 			 * and using interruptible sleeps/timeout may have
375 			 * left spourious lk_exslpfail counts on, so clean
376 			 * it up anyway.
377 			 */
378 			lk->lk_exslpfail = 0;
379 			queue = SQ_SHARED_QUEUE;
380 		}
381 
382 		if (lockmgr_sunlock_try(lk, &orig_x)) {
383 			sleepq_release(&lk->lock_object);
384 			break;
385 		}
386 
387 		x |= LK_SHARERS_LOCK(1);
388 		if (!atomic_fcmpset_rel_ptr(&lk->lk_lock, &x, v)) {
389 			orig_x = x;
390 			goto retry_sleepq;
391 		}
392 		LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
393 		    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
394 		    "exclusive");
395 		wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
396 		    0, queue);
397 		sleepq_release(&lk->lock_object);
398 		break;
399 	}
400 
401 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER);
402 	return (wakeup_swapper);
403 }
404 
405 static void
assert_lockmgr(const struct lock_object * lock,int what)406 assert_lockmgr(const struct lock_object *lock, int what)
407 {
408 
409 	panic("lockmgr locks do not support assertions");
410 }
411 
412 static void
lock_lockmgr(struct lock_object * lock,uintptr_t how)413 lock_lockmgr(struct lock_object *lock, uintptr_t how)
414 {
415 
416 	panic("lockmgr locks do not support sleep interlocking");
417 }
418 
419 static uintptr_t
unlock_lockmgr(struct lock_object * lock)420 unlock_lockmgr(struct lock_object *lock)
421 {
422 
423 	panic("lockmgr locks do not support sleep interlocking");
424 }
425 
426 #ifdef KDTRACE_HOOKS
427 static int
owner_lockmgr(const struct lock_object * lock,struct thread ** owner)428 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
429 {
430 
431 	panic("lockmgr locks do not support owner inquiring");
432 }
433 #endif
434 
435 void
lockinit(struct lock * lk,int pri,const char * wmesg,int timo,int flags)436 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
437 {
438 	int iflags;
439 
440 	MPASS((flags & ~LK_INIT_MASK) == 0);
441 	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
442             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
443             &lk->lk_lock));
444 
445 	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
446 	if (flags & LK_CANRECURSE)
447 		iflags |= LO_RECURSABLE;
448 	if ((flags & LK_NODUP) == 0)
449 		iflags |= LO_DUPOK;
450 	if (flags & LK_NOPROFILE)
451 		iflags |= LO_NOPROFILE;
452 	if ((flags & LK_NOWITNESS) == 0)
453 		iflags |= LO_WITNESS;
454 	if (flags & LK_QUIET)
455 		iflags |= LO_QUIET;
456 	if (flags & LK_IS_VNODE)
457 		iflags |= LO_IS_VNODE;
458 	if (flags & LK_NEW)
459 		iflags |= LO_NEW;
460 	iflags |= flags & LK_NOSHARE;
461 
462 	lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
463 	lk->lk_lock = LK_UNLOCKED;
464 	lk->lk_recurse = 0;
465 	lk->lk_exslpfail = 0;
466 	lk->lk_timo = timo;
467 	lk->lk_pri = pri;
468 	STACK_ZERO(lk);
469 }
470 
471 /*
472  * XXX: Gross hacks to manipulate external lock flags after
473  * initialization.  Used for certain vnode and buf locks.
474  */
475 void
lockallowshare(struct lock * lk)476 lockallowshare(struct lock *lk)
477 {
478 
479 	lockmgr_assert(lk, KA_XLOCKED);
480 	lk->lock_object.lo_flags &= ~LK_NOSHARE;
481 }
482 
483 void
lockdisableshare(struct lock * lk)484 lockdisableshare(struct lock *lk)
485 {
486 
487 	lockmgr_assert(lk, KA_XLOCKED);
488 	lk->lock_object.lo_flags |= LK_NOSHARE;
489 }
490 
491 void
lockallowrecurse(struct lock * lk)492 lockallowrecurse(struct lock *lk)
493 {
494 
495 	lockmgr_assert(lk, KA_XLOCKED);
496 	lk->lock_object.lo_flags |= LO_RECURSABLE;
497 }
498 
499 void
lockdisablerecurse(struct lock * lk)500 lockdisablerecurse(struct lock *lk)
501 {
502 
503 	lockmgr_assert(lk, KA_XLOCKED);
504 	lk->lock_object.lo_flags &= ~LO_RECURSABLE;
505 }
506 
507 void
lockdestroy(struct lock * lk)508 lockdestroy(struct lock *lk)
509 {
510 
511 	KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
512 	KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
513 	KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
514 	lock_destroy(&lk->lock_object);
515 }
516 
517 static bool __always_inline
lockmgr_slock_try(struct lock * lk,uintptr_t * xp,int flags,bool fp)518 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp)
519 {
520 
521 	/*
522 	 * If no other thread has an exclusive lock, or
523 	 * no exclusive waiter is present, bump the count of
524 	 * sharers.  Since we have to preserve the state of
525 	 * waiters, if we fail to acquire the shared lock
526 	 * loop back and retry.
527 	 */
528 	while (LK_CAN_SHARE(*xp, flags, fp)) {
529 		if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
530 		    *xp + LK_ONE_SHARER)) {
531 			return (true);
532 		}
533 	}
534 	return (false);
535 }
536 
537 static bool __always_inline
lockmgr_sunlock_try(struct lock * lk,uintptr_t * xp)538 lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp)
539 {
540 
541 	for (;;) {
542 		if (LK_SHARERS(*xp) > 1 || !(*xp & LK_ALL_WAITERS)) {
543 			if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp,
544 			    *xp - LK_ONE_SHARER))
545 				return (true);
546 			continue;
547 		}
548 		break;
549 	}
550 	return (false);
551 }
552 
553 static bool
lockmgr_slock_adaptive(struct lock_delay_arg * lda,struct lock * lk,uintptr_t * xp,int flags)554 lockmgr_slock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp,
555     int flags)
556 {
557 	struct thread *owner;
558 	uintptr_t x;
559 
560 	x = *xp;
561 	MPASS(x != LK_UNLOCKED);
562 	owner = (struct thread *)LK_HOLDER(x);
563 	for (;;) {
564 		MPASS(owner != curthread);
565 		if (owner == (struct thread *)LK_KERNPROC)
566 			return (false);
567 		if ((x & LK_SHARE) && LK_SHARERS(x) > 0)
568 			return (false);
569 		if (owner == NULL)
570 			return (false);
571 		if (!TD_IS_RUNNING(owner))
572 			return (false);
573 		if ((x & LK_ALL_WAITERS) != 0)
574 			return (false);
575 		lock_delay(lda);
576 		x = lockmgr_read_value(lk);
577 		if (LK_CAN_SHARE(x, flags, false)) {
578 			*xp = x;
579 			return (true);
580 		}
581 		owner = (struct thread *)LK_HOLDER(x);
582 	}
583 }
584 
585 static __noinline int
lockmgr_slock_hard(struct lock * lk,u_int flags,struct lock_object * ilk,const char * file,int line,struct lockmgr_wait * lwa)586 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
587     const char *file, int line, struct lockmgr_wait *lwa)
588 {
589 	uintptr_t tid, x;
590 	int error = 0;
591 	const char *iwmesg;
592 	int ipri, itimo;
593 
594 #ifdef KDTRACE_HOOKS
595 	uint64_t sleep_time = 0;
596 #endif
597 #ifdef LOCK_PROFILING
598 	uint64_t waittime = 0;
599 	int contested = 0;
600 #endif
601 	struct lock_delay_arg lda;
602 
603 	if (SCHEDULER_STOPPED())
604 		goto out;
605 
606 	tid = (uintptr_t)curthread;
607 
608 	if (LK_CAN_WITNESS(flags))
609 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
610 		    file, line, flags & LK_INTERLOCK ? ilk : NULL);
611 	x = lockmgr_read_value(lk);
612 	lock_delay_arg_init(&lda, &lockmgr_delay);
613 	if (!lk_adaptive)
614 		flags &= ~LK_ADAPTIVE;
615 	/*
616 	 * The lock may already be locked exclusive by curthread,
617 	 * avoid deadlock.
618 	 */
619 	if (LK_HOLDER(x) == tid) {
620 		LOCK_LOG2(lk,
621 		    "%s: %p already held in exclusive mode",
622 		    __func__, lk);
623 		error = EDEADLK;
624 		goto out;
625 	}
626 
627 	for (;;) {
628 		if (lockmgr_slock_try(lk, &x, flags, false))
629 			break;
630 
631 		lock_profile_obtain_lock_failed(&lk->lock_object, false,
632 		    &contested, &waittime);
633 
634 		if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) {
635 			if (lockmgr_slock_adaptive(&lda, lk, &x, flags))
636 				continue;
637 		}
638 
639 #ifdef HWPMC_HOOKS
640 		PMC_SOFT_CALL( , , lock, failed);
641 #endif
642 
643 		/*
644 		 * If the lock is expected to not sleep just give up
645 		 * and return.
646 		 */
647 		if (LK_TRYOP(flags)) {
648 			LOCK_LOG2(lk, "%s: %p fails the try operation",
649 			    __func__, lk);
650 			error = EBUSY;
651 			break;
652 		}
653 
654 		/*
655 		 * Acquire the sleepqueue chain lock because we
656 		 * probabilly will need to manipulate waiters flags.
657 		 */
658 		sleepq_lock(&lk->lock_object);
659 		x = lockmgr_read_value(lk);
660 retry_sleepq:
661 
662 		/*
663 		 * if the lock can be acquired in shared mode, try
664 		 * again.
665 		 */
666 		if (LK_CAN_SHARE(x, flags, false)) {
667 			sleepq_release(&lk->lock_object);
668 			continue;
669 		}
670 
671 		/*
672 		 * Try to set the LK_SHARED_WAITERS flag.  If we fail,
673 		 * loop back and retry.
674 		 */
675 		if ((x & LK_SHARED_WAITERS) == 0) {
676 			if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
677 			    x | LK_SHARED_WAITERS)) {
678 				goto retry_sleepq;
679 			}
680 			LOCK_LOG2(lk, "%s: %p set shared waiters flag",
681 			    __func__, lk);
682 		}
683 
684 		if (lwa == NULL) {
685 			iwmesg = lk->lock_object.lo_name;
686 			ipri = lk->lk_pri;
687 			itimo = lk->lk_timo;
688 		} else {
689 			iwmesg = lwa->iwmesg;
690 			ipri = lwa->ipri;
691 			itimo = lwa->itimo;
692 		}
693 
694 		/*
695 		 * As far as we have been unable to acquire the
696 		 * shared lock and the shared waiters flag is set,
697 		 * we will sleep.
698 		 */
699 #ifdef KDTRACE_HOOKS
700 		sleep_time -= lockstat_nsecs(&lk->lock_object);
701 #endif
702 		error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
703 		    SQ_SHARED_QUEUE);
704 #ifdef KDTRACE_HOOKS
705 		sleep_time += lockstat_nsecs(&lk->lock_object);
706 #endif
707 		flags &= ~LK_INTERLOCK;
708 		if (error) {
709 			LOCK_LOG3(lk,
710 			    "%s: interrupted sleep for %p with %d",
711 			    __func__, lk, error);
712 			break;
713 		}
714 		LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
715 		    __func__, lk);
716 		x = lockmgr_read_value(lk);
717 	}
718 	if (error == 0) {
719 #ifdef KDTRACE_HOOKS
720 		if (sleep_time != 0)
721 			LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time,
722 			    LOCKSTAT_READER, (x & LK_SHARE) == 0,
723 			    (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x));
724 #endif
725 #ifdef LOCK_PROFILING
726 		lockmgr_note_shared_acquire(lk, contested, waittime,
727 		    file, line, flags);
728 #else
729 		lockmgr_note_shared_acquire(lk, 0, 0, file, line,
730 		    flags);
731 #endif
732 	}
733 
734 out:
735 	lockmgr_exit(flags, ilk, 0);
736 	return (error);
737 }
738 
739 static bool
lockmgr_xlock_adaptive(struct lock_delay_arg * lda,struct lock * lk,uintptr_t * xp)740 lockmgr_xlock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp)
741 {
742 	struct thread *owner;
743 	uintptr_t x;
744 
745 	x = *xp;
746 	MPASS(x != LK_UNLOCKED);
747 	owner = (struct thread *)LK_HOLDER(x);
748 	for (;;) {
749 		MPASS(owner != curthread);
750 		if (owner == NULL)
751 			return (false);
752 		if ((x & LK_SHARE) && LK_SHARERS(x) > 0)
753 			return (false);
754 		if (owner == (struct thread *)LK_KERNPROC)
755 			return (false);
756 		if (!TD_IS_RUNNING(owner))
757 			return (false);
758 		if ((x & LK_ALL_WAITERS) != 0)
759 			return (false);
760 		lock_delay(lda);
761 		x = lockmgr_read_value(lk);
762 		if (x == LK_UNLOCKED) {
763 			*xp = x;
764 			return (true);
765 		}
766 		owner = (struct thread *)LK_HOLDER(x);
767 	}
768 }
769 
770 static __noinline int
lockmgr_xlock_hard(struct lock * lk,u_int flags,struct lock_object * ilk,const char * file,int line,struct lockmgr_wait * lwa)771 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
772     const char *file, int line, struct lockmgr_wait *lwa)
773 {
774 	struct lock_class *class;
775 	uintptr_t tid, x, v;
776 	int error = 0;
777 	const char *iwmesg;
778 	int ipri, itimo;
779 
780 #ifdef KDTRACE_HOOKS
781 	uint64_t sleep_time = 0;
782 #endif
783 #ifdef LOCK_PROFILING
784 	uint64_t waittime = 0;
785 	int contested = 0;
786 #endif
787 	struct lock_delay_arg lda;
788 
789 	if (SCHEDULER_STOPPED())
790 		goto out;
791 
792 	tid = (uintptr_t)curthread;
793 
794 	if (LK_CAN_WITNESS(flags))
795 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
796 		    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
797 		    ilk : NULL);
798 
799 	/*
800 	 * If curthread already holds the lock and this one is
801 	 * allowed to recurse, simply recurse on it.
802 	 */
803 	if (lockmgr_xlocked(lk)) {
804 		if ((flags & LK_CANRECURSE) == 0 &&
805 		    (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
806 			/*
807 			 * If the lock is expected to not panic just
808 			 * give up and return.
809 			 */
810 			if (LK_TRYOP(flags)) {
811 				LOCK_LOG2(lk,
812 				    "%s: %p fails the try operation",
813 				    __func__, lk);
814 				error = EBUSY;
815 				goto out;
816 			}
817 			if (flags & LK_INTERLOCK) {
818 				class = LOCK_CLASS(ilk);
819 				class->lc_unlock(ilk);
820 			}
821 			STACK_PRINT(lk);
822 			panic("%s: recursing on non recursive lockmgr %p "
823 			    "@ %s:%d\n", __func__, lk, file, line);
824 		}
825 		atomic_set_ptr(&lk->lk_lock, LK_WRITER_RECURSED);
826 		lk->lk_recurse++;
827 		LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
828 		LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
829 		    lk->lk_recurse, file, line);
830 		WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
831 		    LK_TRYWIT(flags), file, line);
832 		TD_LOCKS_INC(curthread);
833 		goto out;
834 	}
835 
836 	x = LK_UNLOCKED;
837 	lock_delay_arg_init(&lda, &lockmgr_delay);
838 	if (!lk_adaptive)
839 		flags &= ~LK_ADAPTIVE;
840 	for (;;) {
841 		if (x == LK_UNLOCKED) {
842 			if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, tid))
843 				break;
844 			continue;
845 		}
846 
847 		lock_profile_obtain_lock_failed(&lk->lock_object, false,
848 		    &contested, &waittime);
849 
850 		if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) {
851 			if (lockmgr_xlock_adaptive(&lda, lk, &x))
852 				continue;
853 		}
854 #ifdef HWPMC_HOOKS
855 		PMC_SOFT_CALL( , , lock, failed);
856 #endif
857 
858 		/*
859 		 * If the lock is expected to not sleep just give up
860 		 * and return.
861 		 */
862 		if (LK_TRYOP(flags)) {
863 			LOCK_LOG2(lk, "%s: %p fails the try operation",
864 			    __func__, lk);
865 			error = EBUSY;
866 			break;
867 		}
868 
869 		/*
870 		 * Acquire the sleepqueue chain lock because we
871 		 * probabilly will need to manipulate waiters flags.
872 		 */
873 		sleepq_lock(&lk->lock_object);
874 		x = lockmgr_read_value(lk);
875 retry_sleepq:
876 
877 		/*
878 		 * if the lock has been released while we spun on
879 		 * the sleepqueue chain lock just try again.
880 		 */
881 		if (x == LK_UNLOCKED) {
882 			sleepq_release(&lk->lock_object);
883 			continue;
884 		}
885 
886 		/*
887 		 * The lock can be in the state where there is a
888 		 * pending queue of waiters, but still no owner.
889 		 * This happens when the lock is contested and an
890 		 * owner is going to claim the lock.
891 		 * If curthread is the one successfully acquiring it
892 		 * claim lock ownership and return, preserving waiters
893 		 * flags.
894 		 */
895 		v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
896 		if ((x & ~v) == LK_UNLOCKED) {
897 			v &= ~LK_EXCLUSIVE_SPINNERS;
898 			if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
899 			    tid | v)) {
900 				sleepq_release(&lk->lock_object);
901 				LOCK_LOG2(lk,
902 				    "%s: %p claimed by a new writer",
903 				    __func__, lk);
904 				break;
905 			}
906 			goto retry_sleepq;
907 		}
908 
909 		/*
910 		 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
911 		 * fail, loop back and retry.
912 		 */
913 		if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
914 			if (!atomic_fcmpset_ptr(&lk->lk_lock, &x,
915 			    x | LK_EXCLUSIVE_WAITERS)) {
916 				goto retry_sleepq;
917 			}
918 			LOCK_LOG2(lk, "%s: %p set excl waiters flag",
919 			    __func__, lk);
920 		}
921 
922 		if (lwa == NULL) {
923 			iwmesg = lk->lock_object.lo_name;
924 			ipri = lk->lk_pri;
925 			itimo = lk->lk_timo;
926 		} else {
927 			iwmesg = lwa->iwmesg;
928 			ipri = lwa->ipri;
929 			itimo = lwa->itimo;
930 		}
931 
932 		/*
933 		 * As far as we have been unable to acquire the
934 		 * exclusive lock and the exclusive waiters flag
935 		 * is set, we will sleep.
936 		 */
937 #ifdef KDTRACE_HOOKS
938 		sleep_time -= lockstat_nsecs(&lk->lock_object);
939 #endif
940 		error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
941 		    SQ_EXCLUSIVE_QUEUE);
942 #ifdef KDTRACE_HOOKS
943 		sleep_time += lockstat_nsecs(&lk->lock_object);
944 #endif
945 		flags &= ~LK_INTERLOCK;
946 		if (error) {
947 			LOCK_LOG3(lk,
948 			    "%s: interrupted sleep for %p with %d",
949 			    __func__, lk, error);
950 			break;
951 		}
952 		LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
953 		    __func__, lk);
954 		x = lockmgr_read_value(lk);
955 	}
956 	if (error == 0) {
957 #ifdef KDTRACE_HOOKS
958 		if (sleep_time != 0)
959 			LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time,
960 			    LOCKSTAT_WRITER, (x & LK_SHARE) == 0,
961 			    (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x));
962 #endif
963 #ifdef LOCK_PROFILING
964 		lockmgr_note_exclusive_acquire(lk, contested, waittime,
965 		    file, line, flags);
966 #else
967 		lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
968 		    flags);
969 #endif
970 	}
971 
972 out:
973 	lockmgr_exit(flags, ilk, 0);
974 	return (error);
975 }
976 
977 static __noinline int
lockmgr_upgrade(struct lock * lk,u_int flags,struct lock_object * ilk,const char * file,int line,struct lockmgr_wait * lwa)978 lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk,
979     const char *file, int line, struct lockmgr_wait *lwa)
980 {
981 	uintptr_t tid, v, setv;
982 	int error = 0;
983 	int op;
984 
985 	if (SCHEDULER_STOPPED())
986 		goto out;
987 
988 	tid = (uintptr_t)curthread;
989 
990 	_lockmgr_assert(lk, KA_SLOCKED, file, line);
991 
992 	op = flags & LK_TYPE_MASK;
993 	v = lockmgr_read_value(lk);
994 	for (;;) {
995 		if (LK_SHARERS(v) > 1) {
996 			if (op == LK_TRYUPGRADE) {
997 				LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
998 				    __func__, lk);
999 				error = EBUSY;
1000 				goto out;
1001 			}
1002 			if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &v,
1003 			    v - LK_ONE_SHARER)) {
1004 				lockmgr_note_shared_release(lk, file, line);
1005 				goto out_xlock;
1006 			}
1007 			continue;
1008 		}
1009 		MPASS((v & ~LK_ALL_WAITERS) == LK_SHARERS_LOCK(1));
1010 
1011 		setv = tid;
1012 		setv |= (v & LK_ALL_WAITERS);
1013 
1014 		/*
1015 		 * Try to switch from one shared lock to an exclusive one.
1016 		 * We need to preserve waiters flags during the operation.
1017 		 */
1018 		if (atomic_fcmpset_ptr(&lk->lk_lock, &v, setv)) {
1019 			LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
1020 			    line);
1021 			WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
1022 			    LK_TRYWIT(flags), file, line);
1023 			LOCKSTAT_RECORD0(lockmgr__upgrade, lk);
1024 			TD_SLOCKS_DEC(curthread);
1025 			goto out;
1026 		}
1027 	}
1028 
1029 out_xlock:
1030 	error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa);
1031 	flags &= ~LK_INTERLOCK;
1032 out:
1033 	lockmgr_exit(flags, ilk, 0);
1034 	return (error);
1035 }
1036 
1037 int
lockmgr_lock_flags(struct lock * lk,u_int flags,struct lock_object * ilk,const char * file,int line)1038 lockmgr_lock_flags(struct lock *lk, u_int flags, struct lock_object *ilk,
1039     const char *file, int line)
1040 {
1041 	struct lock_class *class;
1042 	uintptr_t x, tid;
1043 	u_int op;
1044 	bool locked;
1045 
1046 	if (SCHEDULER_STOPPED())
1047 		return (0);
1048 
1049 	op = flags & LK_TYPE_MASK;
1050 	locked = false;
1051 	switch (op) {
1052 	case LK_SHARED:
1053 		if (LK_CAN_WITNESS(flags))
1054 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
1055 			    file, line, flags & LK_INTERLOCK ? ilk : NULL);
1056 		if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
1057 			break;
1058 		x = lockmgr_read_value(lk);
1059 		if (lockmgr_slock_try(lk, &x, flags, true)) {
1060 			lockmgr_note_shared_acquire(lk, 0, 0,
1061 			    file, line, flags);
1062 			locked = true;
1063 		} else {
1064 			return (lockmgr_slock_hard(lk, flags, ilk, file, line,
1065 			    NULL));
1066 		}
1067 		break;
1068 	case LK_EXCLUSIVE:
1069 		if (LK_CAN_WITNESS(flags))
1070 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1071 			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1072 			    ilk : NULL);
1073 		tid = (uintptr_t)curthread;
1074 		if (lockmgr_read_value(lk) == LK_UNLOCKED &&
1075 		    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1076 			lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
1077 			    flags);
1078 			locked = true;
1079 		} else {
1080 			return (lockmgr_xlock_hard(lk, flags, ilk, file, line,
1081 			    NULL));
1082 		}
1083 		break;
1084 	case LK_UPGRADE:
1085 	case LK_TRYUPGRADE:
1086 		return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL));
1087 	default:
1088 		break;
1089 	}
1090 	if (__predict_true(locked)) {
1091 		if (__predict_false(flags & LK_INTERLOCK)) {
1092 			class = LOCK_CLASS(ilk);
1093 			class->lc_unlock(ilk);
1094 		}
1095 		return (0);
1096 	} else {
1097 		return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT,
1098 		    LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line));
1099 	}
1100 }
1101 
1102 static __noinline int
lockmgr_sunlock_hard(struct lock * lk,uintptr_t x,u_int flags,struct lock_object * ilk,const char * file,int line)1103 lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
1104     const char *file, int line)
1105 
1106 {
1107 	int wakeup_swapper = 0;
1108 
1109 	if (SCHEDULER_STOPPED())
1110 		goto out;
1111 
1112 	wakeup_swapper = wakeupshlk(lk, file, line);
1113 
1114 out:
1115 	lockmgr_exit(flags, ilk, wakeup_swapper);
1116 	return (0);
1117 }
1118 
1119 static __noinline int
lockmgr_xunlock_hard(struct lock * lk,uintptr_t x,u_int flags,struct lock_object * ilk,const char * file,int line)1120 lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
1121     const char *file, int line)
1122 {
1123 	uintptr_t tid, v;
1124 	int wakeup_swapper = 0;
1125 	u_int realexslp;
1126 	int queue;
1127 
1128 	if (SCHEDULER_STOPPED())
1129 		goto out;
1130 
1131 	tid = (uintptr_t)curthread;
1132 
1133 	/*
1134 	 * As first option, treact the lock as if it has not
1135 	 * any waiter.
1136 	 * Fix-up the tid var if the lock has been disowned.
1137 	 */
1138 	if (LK_HOLDER(x) == LK_KERNPROC)
1139 		tid = LK_KERNPROC;
1140 
1141 	/*
1142 	 * The lock is held in exclusive mode.
1143 	 * If the lock is recursed also, then unrecurse it.
1144 	 */
1145 	if (lockmgr_recursed_v(x)) {
1146 		LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk);
1147 		lk->lk_recurse--;
1148 		if (lk->lk_recurse == 0)
1149 			atomic_clear_ptr(&lk->lk_lock, LK_WRITER_RECURSED);
1150 		goto out;
1151 	}
1152 	if (tid != LK_KERNPROC)
1153 		LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk,
1154 		    LOCKSTAT_WRITER);
1155 
1156 	if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED))
1157 		goto out;
1158 
1159 	sleepq_lock(&lk->lock_object);
1160 	x = lockmgr_read_value(lk);
1161 	v = LK_UNLOCKED;
1162 
1163 	/*
1164 	 * If the lock has exclusive waiters, give them
1165 	 * preference in order to avoid deadlock with
1166 	 * shared runners up.
1167 	 * If interruptible sleeps left the exclusive queue
1168 	 * empty avoid a starvation for the threads sleeping
1169 	 * on the shared queue by giving them precedence
1170 	 * and cleaning up the exclusive waiters bit anyway.
1171 	 * Please note that lk_exslpfail count may be lying
1172 	 * about the real number of waiters with the
1173 	 * LK_SLEEPFAIL flag on because they may be used in
1174 	 * conjunction with interruptible sleeps so
1175 	 * lk_exslpfail might be considered an 'upper limit'
1176 	 * bound, including the edge cases.
1177 	 */
1178 	MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1179 	realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE);
1180 	if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1181 		if (lk->lk_exslpfail != USHRT_MAX && lk->lk_exslpfail < realexslp) {
1182 			lk->lk_exslpfail = 0;
1183 			queue = SQ_EXCLUSIVE_QUEUE;
1184 			v |= (x & LK_SHARED_WAITERS);
1185 		} else {
1186 			lk->lk_exslpfail = 0;
1187 			LOCK_LOG2(lk,
1188 			    "%s: %p has only LK_SLEEPFAIL sleepers",
1189 			    __func__, lk);
1190 			LOCK_LOG2(lk,
1191 			    "%s: %p waking up threads on the exclusive queue",
1192 			    __func__, lk);
1193 			wakeup_swapper = sleepq_broadcast(&lk->lock_object,
1194 			    SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1195 			queue = SQ_SHARED_QUEUE;
1196 		}
1197 	} else {
1198 		/*
1199 		 * Exclusive waiters sleeping with LK_SLEEPFAIL
1200 		 * on and using interruptible sleeps/timeout
1201 		 * may have left spourious lk_exslpfail counts
1202 		 * on, so clean it up anyway.
1203 		 */
1204 		lk->lk_exslpfail = 0;
1205 		queue = SQ_SHARED_QUEUE;
1206 	}
1207 
1208 	LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
1209 	    __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1210 	    "exclusive");
1211 	atomic_store_rel_ptr(&lk->lk_lock, v);
1212 	wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue);
1213 	sleepq_release(&lk->lock_object);
1214 
1215 out:
1216 	lockmgr_exit(flags, ilk, wakeup_swapper);
1217 	return (0);
1218 }
1219 
1220 /*
1221  * Lightweight entry points for common operations.
1222  *
1223  * Functionality is similar to sx locks, in that none of the additional lockmgr
1224  * features are supported. To be clear, these are NOT supported:
1225  * 1. shared locking disablement
1226  * 2. returning with an error after sleep
1227  * 3. unlocking the interlock
1228  *
1229  * If in doubt, use lockmgr_lock_flags.
1230  */
1231 int
lockmgr_slock(struct lock * lk,u_int flags,const char * file,int line)1232 lockmgr_slock(struct lock *lk, u_int flags, const char *file, int line)
1233 {
1234 	uintptr_t x;
1235 
1236 	MPASS((flags & LK_TYPE_MASK) == LK_SHARED);
1237 	MPASS((flags & LK_INTERLOCK) == 0);
1238 	MPASS((lk->lock_object.lo_flags & LK_NOSHARE) == 0);
1239 
1240 	if (LK_CAN_WITNESS(flags))
1241 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
1242 		    file, line, NULL);
1243 	x = lockmgr_read_value(lk);
1244 	if (__predict_true(lockmgr_slock_try(lk, &x, flags, true))) {
1245 		lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags);
1246 		return (0);
1247 	}
1248 
1249 	return (lockmgr_slock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL));
1250 }
1251 
1252 int
lockmgr_xlock(struct lock * lk,u_int flags,const char * file,int line)1253 lockmgr_xlock(struct lock *lk, u_int flags, const char *file, int line)
1254 {
1255 	uintptr_t tid;
1256 
1257 	MPASS((flags & LK_TYPE_MASK) == LK_EXCLUSIVE);
1258 	MPASS((flags & LK_INTERLOCK) == 0);
1259 
1260 	if (LK_CAN_WITNESS(flags))
1261 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1262 		    LOP_EXCLUSIVE, file, line, NULL);
1263 	tid = (uintptr_t)curthread;
1264 	if (atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1265 		lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
1266 		    flags);
1267 		return (0);
1268 	}
1269 
1270 	return (lockmgr_xlock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL));
1271 }
1272 
1273 int
lockmgr_unlock(struct lock * lk)1274 lockmgr_unlock(struct lock *lk)
1275 {
1276 	uintptr_t x, tid;
1277 	const char *file;
1278 	int line;
1279 
1280 	file = __FILE__;
1281 	line = __LINE__;
1282 
1283 	_lockmgr_assert(lk, KA_LOCKED, file, line);
1284 	x = lockmgr_read_value(lk);
1285 	if (__predict_true(x & LK_SHARE) != 0) {
1286 		lockmgr_note_shared_release(lk, file, line);
1287 		if (lockmgr_sunlock_try(lk, &x)) {
1288 			LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER);
1289 		} else {
1290 			return (lockmgr_sunlock_hard(lk, x, LK_RELEASE, NULL, file, line));
1291 		}
1292 	} else {
1293 		tid = (uintptr_t)curthread;
1294 		lockmgr_note_exclusive_release(lk, file, line);
1295 		if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) {
1296 			LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk,LOCKSTAT_WRITER);
1297 		} else {
1298 			return (lockmgr_xunlock_hard(lk, x, LK_RELEASE, NULL, file, line));
1299 		}
1300 	}
1301 	return (0);
1302 }
1303 
1304 int
__lockmgr_args(struct lock * lk,u_int flags,struct lock_object * ilk,const char * wmesg,int pri,int timo,const char * file,int line)1305 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
1306     const char *wmesg, int pri, int timo, const char *file, int line)
1307 {
1308 	GIANT_DECLARE;
1309 	struct lockmgr_wait lwa;
1310 	struct lock_class *class;
1311 	const char *iwmesg;
1312 	uintptr_t tid, v, x;
1313 	u_int op, realexslp;
1314 	int error, ipri, itimo, queue, wakeup_swapper;
1315 #ifdef LOCK_PROFILING
1316 	uint64_t waittime = 0;
1317 	int contested = 0;
1318 #endif
1319 
1320 	if (SCHEDULER_STOPPED())
1321 		return (0);
1322 
1323 	error = 0;
1324 	tid = (uintptr_t)curthread;
1325 	op = (flags & LK_TYPE_MASK);
1326 	iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
1327 	ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
1328 	itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
1329 
1330 	lwa.iwmesg = iwmesg;
1331 	lwa.ipri = ipri;
1332 	lwa.itimo = itimo;
1333 
1334 	MPASS((flags & ~LK_TOTAL_MASK) == 0);
1335 	KASSERT((op & (op - 1)) == 0,
1336 	    ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
1337 	KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
1338 	    (op != LK_DOWNGRADE && op != LK_RELEASE),
1339 	    ("%s: Invalid flags in regard of the operation desired @ %s:%d",
1340 	    __func__, file, line));
1341 	KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
1342 	    ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
1343 	    __func__, file, line));
1344 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
1345 	    ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
1346 	    lk->lock_object.lo_name, file, line));
1347 
1348 	class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
1349 
1350 	if (lk->lock_object.lo_flags & LK_NOSHARE) {
1351 		switch (op) {
1352 		case LK_SHARED:
1353 			op = LK_EXCLUSIVE;
1354 			break;
1355 		case LK_UPGRADE:
1356 		case LK_TRYUPGRADE:
1357 		case LK_DOWNGRADE:
1358 			_lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
1359 			    file, line);
1360 			if (flags & LK_INTERLOCK)
1361 				class->lc_unlock(ilk);
1362 			return (0);
1363 		}
1364 	}
1365 
1366 	wakeup_swapper = 0;
1367 	switch (op) {
1368 	case LK_SHARED:
1369 		return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa));
1370 		break;
1371 	case LK_UPGRADE:
1372 	case LK_TRYUPGRADE:
1373 		return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa));
1374 		break;
1375 	case LK_EXCLUSIVE:
1376 		return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa));
1377 		break;
1378 	case LK_DOWNGRADE:
1379 		_lockmgr_assert(lk, KA_XLOCKED, file, line);
1380 		WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
1381 
1382 		/*
1383 		 * Panic if the lock is recursed.
1384 		 */
1385 		if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1386 			if (flags & LK_INTERLOCK)
1387 				class->lc_unlock(ilk);
1388 			panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
1389 			    __func__, iwmesg, file, line);
1390 		}
1391 		TD_SLOCKS_INC(curthread);
1392 
1393 		/*
1394 		 * In order to preserve waiters flags, just spin.
1395 		 */
1396 		for (;;) {
1397 			x = lockmgr_read_value(lk);
1398 			MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1399 			x &= LK_ALL_WAITERS;
1400 			if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1401 			    LK_SHARERS_LOCK(1) | x))
1402 				break;
1403 			cpu_spinwait();
1404 		}
1405 		LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
1406 		LOCKSTAT_RECORD0(lockmgr__downgrade, lk);
1407 		break;
1408 	case LK_RELEASE:
1409 		_lockmgr_assert(lk, KA_LOCKED, file, line);
1410 		x = lockmgr_read_value(lk);
1411 
1412 		if (__predict_true(x & LK_SHARE) != 0) {
1413 			lockmgr_note_shared_release(lk, file, line);
1414 			return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1415 		} else {
1416 			lockmgr_note_exclusive_release(lk, file, line);
1417 			return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1418 		}
1419 		break;
1420 	case LK_DRAIN:
1421 		if (LK_CAN_WITNESS(flags))
1422 			WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1423 			    LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1424 			    ilk : NULL);
1425 
1426 		/*
1427 		 * Trying to drain a lock we already own will result in a
1428 		 * deadlock.
1429 		 */
1430 		if (lockmgr_xlocked(lk)) {
1431 			if (flags & LK_INTERLOCK)
1432 				class->lc_unlock(ilk);
1433 			panic("%s: draining %s with the lock held @ %s:%d\n",
1434 			    __func__, iwmesg, file, line);
1435 		}
1436 
1437 		for (;;) {
1438 			if (lk->lk_lock == LK_UNLOCKED &&
1439 			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1440 				break;
1441 
1442 #ifdef HWPMC_HOOKS
1443 			PMC_SOFT_CALL( , , lock, failed);
1444 #endif
1445 			lock_profile_obtain_lock_failed(&lk->lock_object, false,
1446 			    &contested, &waittime);
1447 
1448 			/*
1449 			 * If the lock is expected to not sleep just give up
1450 			 * and return.
1451 			 */
1452 			if (LK_TRYOP(flags)) {
1453 				LOCK_LOG2(lk, "%s: %p fails the try operation",
1454 				    __func__, lk);
1455 				error = EBUSY;
1456 				break;
1457 			}
1458 
1459 			/*
1460 			 * Acquire the sleepqueue chain lock because we
1461 			 * probabilly will need to manipulate waiters flags.
1462 			 */
1463 			sleepq_lock(&lk->lock_object);
1464 			x = lockmgr_read_value(lk);
1465 
1466 			/*
1467 			 * if the lock has been released while we spun on
1468 			 * the sleepqueue chain lock just try again.
1469 			 */
1470 			if (x == LK_UNLOCKED) {
1471 				sleepq_release(&lk->lock_object);
1472 				continue;
1473 			}
1474 
1475 			v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1476 			if ((x & ~v) == LK_UNLOCKED) {
1477 				v = (x & ~LK_EXCLUSIVE_SPINNERS);
1478 
1479 				/*
1480 				 * If interruptible sleeps left the exclusive
1481 				 * queue empty avoid a starvation for the
1482 				 * threads sleeping on the shared queue by
1483 				 * giving them precedence and cleaning up the
1484 				 * exclusive waiters bit anyway.
1485 				 * Please note that lk_exslpfail count may be
1486 				 * lying about the real number of waiters with
1487 				 * the LK_SLEEPFAIL flag on because they may
1488 				 * be used in conjunction with interruptible
1489 				 * sleeps so lk_exslpfail might be considered
1490 				 * an 'upper limit' bound, including the edge
1491 				 * cases.
1492 				 */
1493 				if (v & LK_EXCLUSIVE_WAITERS) {
1494 					queue = SQ_EXCLUSIVE_QUEUE;
1495 					v &= ~LK_EXCLUSIVE_WAITERS;
1496 				} else {
1497 					/*
1498 					 * Exclusive waiters sleeping with
1499 					 * LK_SLEEPFAIL on and using
1500 					 * interruptible sleeps/timeout may
1501 					 * have left spourious lk_exslpfail
1502 					 * counts on, so clean it up anyway.
1503 					 */
1504 					MPASS(v & LK_SHARED_WAITERS);
1505 					lk->lk_exslpfail = 0;
1506 					queue = SQ_SHARED_QUEUE;
1507 					v &= ~LK_SHARED_WAITERS;
1508 				}
1509 				if (queue == SQ_EXCLUSIVE_QUEUE) {
1510 					realexslp =
1511 					    sleepq_sleepcnt(&lk->lock_object,
1512 					    SQ_EXCLUSIVE_QUEUE);
1513 					if (lk->lk_exslpfail >= realexslp) {
1514 						lk->lk_exslpfail = 0;
1515 						queue = SQ_SHARED_QUEUE;
1516 						v &= ~LK_SHARED_WAITERS;
1517 						if (realexslp != 0) {
1518 							LOCK_LOG2(lk,
1519 					"%s: %p has only LK_SLEEPFAIL sleepers",
1520 							    __func__, lk);
1521 							LOCK_LOG2(lk,
1522 			"%s: %p waking up threads on the exclusive queue",
1523 							    __func__, lk);
1524 							wakeup_swapper =
1525 							    sleepq_broadcast(
1526 							    &lk->lock_object,
1527 							    SLEEPQ_LK, 0,
1528 							    SQ_EXCLUSIVE_QUEUE);
1529 						}
1530 					} else
1531 						lk->lk_exslpfail = 0;
1532 				}
1533 				if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1534 					sleepq_release(&lk->lock_object);
1535 					continue;
1536 				}
1537 				LOCK_LOG3(lk,
1538 				"%s: %p waking up all threads on the %s queue",
1539 				    __func__, lk, queue == SQ_SHARED_QUEUE ?
1540 				    "shared" : "exclusive");
1541 				wakeup_swapper |= sleepq_broadcast(
1542 				    &lk->lock_object, SLEEPQ_LK, 0, queue);
1543 
1544 				/*
1545 				 * If shared waiters have been woken up we need
1546 				 * to wait for one of them to acquire the lock
1547 				 * before to set the exclusive waiters in
1548 				 * order to avoid a deadlock.
1549 				 */
1550 				if (queue == SQ_SHARED_QUEUE) {
1551 					for (v = lk->lk_lock;
1552 					    (v & LK_SHARE) && !LK_SHARERS(v);
1553 					    v = lk->lk_lock)
1554 						cpu_spinwait();
1555 				}
1556 			}
1557 
1558 			/*
1559 			 * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1560 			 * fail, loop back and retry.
1561 			 */
1562 			if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1563 				if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1564 				    x | LK_EXCLUSIVE_WAITERS)) {
1565 					sleepq_release(&lk->lock_object);
1566 					continue;
1567 				}
1568 				LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1569 				    __func__, lk);
1570 			}
1571 
1572 			/*
1573 			 * As far as we have been unable to acquire the
1574 			 * exclusive lock and the exclusive waiters flag
1575 			 * is set, we will sleep.
1576 			 */
1577 			if (flags & LK_INTERLOCK) {
1578 				class->lc_unlock(ilk);
1579 				flags &= ~LK_INTERLOCK;
1580 			}
1581 			GIANT_SAVE();
1582 			sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1583 			    SQ_EXCLUSIVE_QUEUE);
1584 			sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1585 			GIANT_RESTORE();
1586 			LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1587 			    __func__, lk);
1588 		}
1589 
1590 		if (error == 0) {
1591 			lock_profile_obtain_lock_success(&lk->lock_object,
1592 			    false, contested, waittime, file, line);
1593 			LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1594 			    lk->lk_recurse, file, line);
1595 			WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1596 			    LK_TRYWIT(flags), file, line);
1597 			TD_LOCKS_INC(curthread);
1598 			STACK_SAVE(lk);
1599 		}
1600 		break;
1601 	default:
1602 		if (flags & LK_INTERLOCK)
1603 			class->lc_unlock(ilk);
1604 		panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1605 	}
1606 
1607 	if (flags & LK_INTERLOCK)
1608 		class->lc_unlock(ilk);
1609 	if (wakeup_swapper)
1610 		kick_proc0();
1611 
1612 	return (error);
1613 }
1614 
1615 void
_lockmgr_disown(struct lock * lk,const char * file,int line)1616 _lockmgr_disown(struct lock *lk, const char *file, int line)
1617 {
1618 	uintptr_t tid, x;
1619 
1620 	if (SCHEDULER_STOPPED())
1621 		return;
1622 
1623 	tid = (uintptr_t)curthread;
1624 	_lockmgr_assert(lk, KA_XLOCKED, file, line);
1625 
1626 	/*
1627 	 * Panic if the lock is recursed.
1628 	 */
1629 	if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1630 		panic("%s: disown a recursed lockmgr @ %s:%d\n",
1631 		    __func__,  file, line);
1632 
1633 	/*
1634 	 * If the owner is already LK_KERNPROC just skip the whole operation.
1635 	 */
1636 	if (LK_HOLDER(lk->lk_lock) != tid)
1637 		return;
1638 	lock_profile_release_lock(&lk->lock_object, false);
1639 	LOCKSTAT_RECORD1(lockmgr__disown, lk, LOCKSTAT_WRITER);
1640 	LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1641 	WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1642 	TD_LOCKS_DEC(curthread);
1643 	STACK_SAVE(lk);
1644 
1645 	/*
1646 	 * In order to preserve waiters flags, just spin.
1647 	 */
1648 	for (;;) {
1649 		x = lockmgr_read_value(lk);
1650 		MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1651 		x &= LK_ALL_WAITERS;
1652 		if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1653 		    LK_KERNPROC | x))
1654 			return;
1655 		cpu_spinwait();
1656 	}
1657 }
1658 
1659 void
lockmgr_printinfo(const struct lock * lk)1660 lockmgr_printinfo(const struct lock *lk)
1661 {
1662 	struct thread *td;
1663 	uintptr_t x;
1664 
1665 	if (lk->lk_lock == LK_UNLOCKED)
1666 		printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1667 	else if (lk->lk_lock & LK_SHARE)
1668 		printf("lock type %s: SHARED (count %ju)\n",
1669 		    lk->lock_object.lo_name,
1670 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1671 	else {
1672 		td = lockmgr_xholder(lk);
1673 		if (td == (struct thread *)LK_KERNPROC)
1674 			printf("lock type %s: EXCL by KERNPROC\n",
1675 			    lk->lock_object.lo_name);
1676 		else
1677 			printf("lock type %s: EXCL by thread %p "
1678 			    "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1679 			    td, td->td_proc->p_pid, td->td_proc->p_comm,
1680 			    td->td_tid);
1681 	}
1682 
1683 	x = lk->lk_lock;
1684 	if (x & LK_EXCLUSIVE_WAITERS)
1685 		printf(" with exclusive waiters pending\n");
1686 	if (x & LK_SHARED_WAITERS)
1687 		printf(" with shared waiters pending\n");
1688 	if (x & LK_EXCLUSIVE_SPINNERS)
1689 		printf(" with exclusive spinners pending\n");
1690 
1691 	STACK_PRINT(lk);
1692 }
1693 
1694 int
lockstatus(const struct lock * lk)1695 lockstatus(const struct lock *lk)
1696 {
1697 	uintptr_t v, x;
1698 	int ret;
1699 
1700 	ret = LK_SHARED;
1701 	x = lockmgr_read_value(lk);
1702 	v = LK_HOLDER(x);
1703 
1704 	if ((x & LK_SHARE) == 0) {
1705 		if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1706 			ret = LK_EXCLUSIVE;
1707 		else
1708 			ret = LK_EXCLOTHER;
1709 	} else if (x == LK_UNLOCKED)
1710 		ret = 0;
1711 
1712 	return (ret);
1713 }
1714 
1715 #ifdef INVARIANT_SUPPORT
1716 
1717 FEATURE(invariant_support,
1718     "Support for modules compiled with INVARIANTS option");
1719 
1720 #ifndef INVARIANTS
1721 #undef	_lockmgr_assert
1722 #endif
1723 
1724 void
_lockmgr_assert(const struct lock * lk,int what,const char * file,int line)1725 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1726 {
1727 	int slocked = 0;
1728 
1729 	if (SCHEDULER_STOPPED())
1730 		return;
1731 	switch (what) {
1732 	case KA_SLOCKED:
1733 	case KA_SLOCKED | KA_NOTRECURSED:
1734 	case KA_SLOCKED | KA_RECURSED:
1735 		slocked = 1;
1736 	case KA_LOCKED:
1737 	case KA_LOCKED | KA_NOTRECURSED:
1738 	case KA_LOCKED | KA_RECURSED:
1739 #ifdef WITNESS
1740 
1741 		/*
1742 		 * We cannot trust WITNESS if the lock is held in exclusive
1743 		 * mode and a call to lockmgr_disown() happened.
1744 		 * Workaround this skipping the check if the lock is held in
1745 		 * exclusive mode even for the KA_LOCKED case.
1746 		 */
1747 		if (slocked || (lk->lk_lock & LK_SHARE)) {
1748 			witness_assert(&lk->lock_object, what, file, line);
1749 			break;
1750 		}
1751 #endif
1752 		if (lk->lk_lock == LK_UNLOCKED ||
1753 		    ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1754 		    (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1755 			panic("Lock %s not %slocked @ %s:%d\n",
1756 			    lk->lock_object.lo_name, slocked ? "share" : "",
1757 			    file, line);
1758 
1759 		if ((lk->lk_lock & LK_SHARE) == 0) {
1760 			if (lockmgr_recursed(lk)) {
1761 				if (what & KA_NOTRECURSED)
1762 					panic("Lock %s recursed @ %s:%d\n",
1763 					    lk->lock_object.lo_name, file,
1764 					    line);
1765 			} else if (what & KA_RECURSED)
1766 				panic("Lock %s not recursed @ %s:%d\n",
1767 				    lk->lock_object.lo_name, file, line);
1768 		}
1769 		break;
1770 	case KA_XLOCKED:
1771 	case KA_XLOCKED | KA_NOTRECURSED:
1772 	case KA_XLOCKED | KA_RECURSED:
1773 		if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1774 			panic("Lock %s not exclusively locked @ %s:%d\n",
1775 			    lk->lock_object.lo_name, file, line);
1776 		if (lockmgr_recursed(lk)) {
1777 			if (what & KA_NOTRECURSED)
1778 				panic("Lock %s recursed @ %s:%d\n",
1779 				    lk->lock_object.lo_name, file, line);
1780 		} else if (what & KA_RECURSED)
1781 			panic("Lock %s not recursed @ %s:%d\n",
1782 			    lk->lock_object.lo_name, file, line);
1783 		break;
1784 	case KA_UNLOCKED:
1785 		if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1786 			panic("Lock %s exclusively locked @ %s:%d\n",
1787 			    lk->lock_object.lo_name, file, line);
1788 		break;
1789 	default:
1790 		panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1791 		    line);
1792 	}
1793 }
1794 #endif
1795 
1796 #ifdef DDB
1797 int
lockmgr_chain(struct thread * td,struct thread ** ownerp)1798 lockmgr_chain(struct thread *td, struct thread **ownerp)
1799 {
1800 	const struct lock *lk;
1801 
1802 	lk = td->td_wchan;
1803 
1804 	if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1805 		return (0);
1806 	db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1807 	if (lk->lk_lock & LK_SHARE)
1808 		db_printf("SHARED (count %ju)\n",
1809 		    (uintmax_t)LK_SHARERS(lk->lk_lock));
1810 	else
1811 		db_printf("EXCL\n");
1812 	*ownerp = lockmgr_xholder(lk);
1813 
1814 	return (1);
1815 }
1816 
1817 static void
db_show_lockmgr(const struct lock_object * lock)1818 db_show_lockmgr(const struct lock_object *lock)
1819 {
1820 	struct thread *td;
1821 	const struct lock *lk;
1822 
1823 	lk = (const struct lock *)lock;
1824 
1825 	db_printf(" state: ");
1826 	if (lk->lk_lock == LK_UNLOCKED)
1827 		db_printf("UNLOCKED\n");
1828 	else if (lk->lk_lock & LK_SHARE)
1829 		db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1830 	else {
1831 		td = lockmgr_xholder(lk);
1832 		if (td == (struct thread *)LK_KERNPROC)
1833 			db_printf("XLOCK: LK_KERNPROC\n");
1834 		else
1835 			db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1836 			    td->td_tid, td->td_proc->p_pid,
1837 			    td->td_proc->p_comm);
1838 		if (lockmgr_recursed(lk))
1839 			db_printf(" recursed: %d\n", lk->lk_recurse);
1840 	}
1841 	db_printf(" waiters: ");
1842 	switch (lk->lk_lock & LK_ALL_WAITERS) {
1843 	case LK_SHARED_WAITERS:
1844 		db_printf("shared\n");
1845 		break;
1846 	case LK_EXCLUSIVE_WAITERS:
1847 		db_printf("exclusive\n");
1848 		break;
1849 	case LK_ALL_WAITERS:
1850 		db_printf("shared and exclusive\n");
1851 		break;
1852 	default:
1853 		db_printf("none\n");
1854 	}
1855 	db_printf(" spinners: ");
1856 	if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1857 		db_printf("exclusive\n");
1858 	else
1859 		db_printf("none\n");
1860 }
1861 #endif
1862