xref: /freebsd/sys/kern/kern_rmlock.c (revision 2e3f49888ec8851bafb22011533217487764fdb0)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2007 Stephan Uphoff <ups@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, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the author nor the names of any co-contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * Machine independent bits of reader/writer lock implementation.
34  */
35 
36 #include <sys/cdefs.h>
37 #include "opt_ddb.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 
42 #include <sys/kernel.h>
43 #include <sys/kdb.h>
44 #include <sys/ktr.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/proc.h>
48 #include <sys/rmlock.h>
49 #include <sys/sched.h>
50 #include <sys/smp.h>
51 #include <sys/turnstile.h>
52 #include <sys/lock_profile.h>
53 #include <machine/cpu.h>
54 #include <vm/uma.h>
55 
56 #ifdef DDB
57 #include <ddb/ddb.h>
58 #endif
59 
60 /*
61  * A cookie to mark destroyed rmlocks.  This is stored in the head of
62  * rm_activeReaders.
63  */
64 #define	RM_DESTROYED	((void *)0xdead)
65 
66 #define	rm_destroyed(rm)						\
67 	(LIST_FIRST(&(rm)->rm_activeReaders) == RM_DESTROYED)
68 
69 #define RMPF_ONQUEUE	1
70 #define RMPF_SIGNAL	2
71 
72 #ifndef INVARIANTS
73 #define	_rm_assert(c, what, file, line)
74 #endif
75 
76 static void	assert_rm(const struct lock_object *lock, int what);
77 #ifdef DDB
78 static void	db_show_rm(const struct lock_object *lock);
79 #endif
80 static void	lock_rm(struct lock_object *lock, uintptr_t how);
81 #ifdef KDTRACE_HOOKS
82 static int	owner_rm(const struct lock_object *lock, struct thread **owner);
83 #endif
84 static uintptr_t unlock_rm(struct lock_object *lock);
85 
86 struct lock_class lock_class_rm = {
87 	.lc_name = "rm",
88 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
89 	.lc_assert = assert_rm,
90 #ifdef DDB
91 	.lc_ddb_show = db_show_rm,
92 #endif
93 	.lc_lock = lock_rm,
94 	.lc_unlock = unlock_rm,
95 #ifdef KDTRACE_HOOKS
96 	.lc_owner = owner_rm,
97 #endif
98 };
99 
100 struct lock_class lock_class_rm_sleepable = {
101 	.lc_name = "sleepable rm",
102 	.lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE,
103 	.lc_assert = assert_rm,
104 #ifdef DDB
105 	.lc_ddb_show = db_show_rm,
106 #endif
107 	.lc_lock = lock_rm,
108 	.lc_unlock = unlock_rm,
109 #ifdef KDTRACE_HOOKS
110 	.lc_owner = owner_rm,
111 #endif
112 };
113 
114 static void
115 assert_rm(const struct lock_object *lock, int what)
116 {
117 
118 	rm_assert((const struct rmlock *)lock, what);
119 }
120 
121 static void
122 lock_rm(struct lock_object *lock, uintptr_t how)
123 {
124 	struct rmlock *rm;
125 	struct rm_priotracker *tracker;
126 
127 	rm = (struct rmlock *)lock;
128 	if (how == 0)
129 		rm_wlock(rm);
130 	else {
131 		tracker = (struct rm_priotracker *)how;
132 		rm_rlock(rm, tracker);
133 	}
134 }
135 
136 static uintptr_t
137 unlock_rm(struct lock_object *lock)
138 {
139 	struct thread *td;
140 	struct pcpu *pc;
141 	struct rmlock *rm;
142 	struct rm_queue *queue;
143 	struct rm_priotracker *tracker;
144 	uintptr_t how;
145 
146 	rm = (struct rmlock *)lock;
147 	tracker = NULL;
148 	how = 0;
149 	rm_assert(rm, RA_LOCKED | RA_NOTRECURSED);
150 	if (rm_wowned(rm))
151 		rm_wunlock(rm);
152 	else {
153 		/*
154 		 * Find the right rm_priotracker structure for curthread.
155 		 * The guarantee about its uniqueness is given by the fact
156 		 * we already asserted the lock wasn't recursively acquired.
157 		 */
158 		critical_enter();
159 		td = curthread;
160 		pc = get_pcpu();
161 		for (queue = pc->pc_rm_queue.rmq_next;
162 		    queue != &pc->pc_rm_queue; queue = queue->rmq_next) {
163 			tracker = (struct rm_priotracker *)queue;
164 				if ((tracker->rmp_rmlock == rm) &&
165 				    (tracker->rmp_thread == td)) {
166 					how = (uintptr_t)tracker;
167 					break;
168 				}
169 		}
170 		KASSERT(tracker != NULL,
171 		    ("rm_priotracker is non-NULL when lock held in read mode"));
172 		critical_exit();
173 		rm_runlock(rm, tracker);
174 	}
175 	return (how);
176 }
177 
178 #ifdef KDTRACE_HOOKS
179 static int
180 owner_rm(const struct lock_object *lock, struct thread **owner)
181 {
182 	const struct rmlock *rm;
183 	struct lock_class *lc;
184 
185 	rm = (const struct rmlock *)lock;
186 	lc = LOCK_CLASS(&rm->rm_wlock_object);
187 	return (lc->lc_owner(&rm->rm_wlock_object, owner));
188 }
189 #endif
190 
191 static struct mtx rm_spinlock;
192 
193 MTX_SYSINIT(rm_spinlock, &rm_spinlock, "rm_spinlock", MTX_SPIN);
194 
195 /*
196  * Add or remove tracker from per-cpu list.
197  *
198  * The per-cpu list can be traversed at any time in forward direction from an
199  * interrupt on the *local* cpu.
200  */
201 static void inline
202 rm_tracker_add(struct pcpu *pc, struct rm_priotracker *tracker)
203 {
204 	struct rm_queue *next;
205 
206 	/* Initialize all tracker pointers */
207 	tracker->rmp_cpuQueue.rmq_prev = &pc->pc_rm_queue;
208 	next = pc->pc_rm_queue.rmq_next;
209 	tracker->rmp_cpuQueue.rmq_next = next;
210 
211 	/* rmq_prev is not used during froward traversal. */
212 	next->rmq_prev = &tracker->rmp_cpuQueue;
213 
214 	/* Update pointer to first element. */
215 	pc->pc_rm_queue.rmq_next = &tracker->rmp_cpuQueue;
216 }
217 
218 /*
219  * Return a count of the number of trackers the thread 'td' already
220  * has on this CPU for the lock 'rm'.
221  */
222 static int
223 rm_trackers_present(const struct pcpu *pc, const struct rmlock *rm,
224     const struct thread *td)
225 {
226 	struct rm_queue *queue;
227 	struct rm_priotracker *tracker;
228 	int count;
229 
230 	count = 0;
231 	for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue;
232 	    queue = queue->rmq_next) {
233 		tracker = (struct rm_priotracker *)queue;
234 		if ((tracker->rmp_rmlock == rm) && (tracker->rmp_thread == td))
235 			count++;
236 	}
237 	return (count);
238 }
239 
240 static void inline
241 rm_tracker_remove(struct pcpu *pc, struct rm_priotracker *tracker)
242 {
243 	struct rm_queue *next, *prev;
244 
245 	next = tracker->rmp_cpuQueue.rmq_next;
246 	prev = tracker->rmp_cpuQueue.rmq_prev;
247 
248 	/* Not used during forward traversal. */
249 	next->rmq_prev = prev;
250 
251 	/* Remove from list. */
252 	prev->rmq_next = next;
253 }
254 
255 static void
256 rm_cleanIPI(void *arg)
257 {
258 	struct pcpu *pc;
259 	struct rmlock *rm = arg;
260 	struct rm_priotracker *tracker;
261 	struct rm_queue *queue;
262 	pc = get_pcpu();
263 
264 	for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue;
265 	    queue = queue->rmq_next) {
266 		tracker = (struct rm_priotracker *)queue;
267 		if (tracker->rmp_rmlock == rm && tracker->rmp_flags == 0) {
268 			tracker->rmp_flags = RMPF_ONQUEUE;
269 			mtx_lock_spin(&rm_spinlock);
270 			LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker,
271 			    rmp_qentry);
272 			mtx_unlock_spin(&rm_spinlock);
273 		}
274 	}
275 }
276 
277 void
278 rm_init_flags(struct rmlock *rm, const char *name, int opts)
279 {
280 	struct lock_class *lc;
281 	int liflags, xflags;
282 
283 	liflags = 0;
284 	if (!(opts & RM_NOWITNESS))
285 		liflags |= LO_WITNESS;
286 	if (opts & RM_RECURSE)
287 		liflags |= LO_RECURSABLE;
288 	if (opts & RM_NEW)
289 		liflags |= LO_NEW;
290 	if (opts & RM_DUPOK)
291 		liflags |= LO_DUPOK;
292 	rm->rm_writecpus = all_cpus;
293 	LIST_INIT(&rm->rm_activeReaders);
294 	if (opts & RM_SLEEPABLE) {
295 		liflags |= LO_SLEEPABLE;
296 		lc = &lock_class_rm_sleepable;
297 		xflags = (opts & RM_NEW ? SX_NEW : 0);
298 		sx_init_flags(&rm->rm_lock_sx, "rmlock_sx",
299 		    xflags | SX_NOWITNESS);
300 	} else {
301 		lc = &lock_class_rm;
302 		xflags = (opts & RM_NEW ? MTX_NEW : 0);
303 		mtx_init(&rm->rm_lock_mtx, name, "rmlock_mtx",
304 		    xflags | MTX_NOWITNESS);
305 	}
306 	lock_init(&rm->lock_object, lc, name, NULL, liflags);
307 }
308 
309 void
310 rm_init(struct rmlock *rm, const char *name)
311 {
312 
313 	rm_init_flags(rm, name, 0);
314 }
315 
316 void
317 rm_destroy(struct rmlock *rm)
318 {
319 
320 	rm_assert(rm, RA_UNLOCKED);
321 	LIST_FIRST(&rm->rm_activeReaders) = RM_DESTROYED;
322 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
323 		sx_destroy(&rm->rm_lock_sx);
324 	else
325 		mtx_destroy(&rm->rm_lock_mtx);
326 	lock_destroy(&rm->lock_object);
327 }
328 
329 int
330 rm_wowned(const struct rmlock *rm)
331 {
332 
333 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
334 		return (sx_xlocked(&rm->rm_lock_sx));
335 	else
336 		return (mtx_owned(&rm->rm_lock_mtx));
337 }
338 
339 void
340 rm_sysinit(void *arg)
341 {
342 	struct rm_args *args;
343 
344 	args = arg;
345 	rm_init_flags(args->ra_rm, args->ra_desc, args->ra_flags);
346 }
347 
348 static __noinline int
349 _rm_rlock_hard(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
350 {
351 	struct pcpu *pc;
352 
353 	critical_enter();
354 	pc = get_pcpu();
355 
356 	/* Check if we just need to do a proper critical_exit. */
357 	if (!CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)) {
358 		critical_exit();
359 		return (1);
360 	}
361 
362 	/* Remove our tracker from the per-cpu list. */
363 	rm_tracker_remove(pc, tracker);
364 
365 	/*
366 	 * Check to see if the IPI granted us the lock after all.  The load of
367 	 * rmp_flags must happen after the tracker is removed from the list.
368 	 */
369 	atomic_interrupt_fence();
370 	if (tracker->rmp_flags) {
371 		/* Just add back tracker - we hold the lock. */
372 		rm_tracker_add(pc, tracker);
373 		critical_exit();
374 		return (1);
375 	}
376 
377 	/*
378 	 * We allow readers to acquire a lock even if a writer is blocked if
379 	 * the lock is recursive and the reader already holds the lock.
380 	 */
381 	if ((rm->lock_object.lo_flags & LO_RECURSABLE) != 0) {
382 		/*
383 		 * Just grant the lock if this thread already has a tracker
384 		 * for this lock on the per-cpu queue.
385 		 */
386 		if (rm_trackers_present(pc, rm, curthread) != 0) {
387 			mtx_lock_spin(&rm_spinlock);
388 			LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker,
389 			    rmp_qentry);
390 			tracker->rmp_flags = RMPF_ONQUEUE;
391 			mtx_unlock_spin(&rm_spinlock);
392 			rm_tracker_add(pc, tracker);
393 			critical_exit();
394 			return (1);
395 		}
396 	}
397 
398 	sched_unpin();
399 	critical_exit();
400 
401 	if (trylock) {
402 		if (rm->lock_object.lo_flags & LO_SLEEPABLE) {
403 			if (!sx_try_xlock(&rm->rm_lock_sx))
404 				return (0);
405 		} else {
406 			if (!mtx_trylock(&rm->rm_lock_mtx))
407 				return (0);
408 		}
409 	} else {
410 		if (rm->lock_object.lo_flags & LO_SLEEPABLE) {
411 			THREAD_SLEEPING_OK();
412 			sx_xlock(&rm->rm_lock_sx);
413 			THREAD_NO_SLEEPING();
414 		} else
415 			mtx_lock(&rm->rm_lock_mtx);
416 	}
417 
418 	critical_enter();
419 	pc = get_pcpu();
420 	CPU_CLR(pc->pc_cpuid, &rm->rm_writecpus);
421 	rm_tracker_add(pc, tracker);
422 	sched_pin();
423 	critical_exit();
424 
425 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
426 		sx_xunlock(&rm->rm_lock_sx);
427 	else
428 		mtx_unlock(&rm->rm_lock_mtx);
429 
430 	return (1);
431 }
432 
433 int
434 _rm_rlock(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
435 {
436 	struct thread *td = curthread;
437 	struct pcpu *pc;
438 
439 	if (SCHEDULER_STOPPED())
440 		return (1);
441 
442 	tracker->rmp_flags  = 0;
443 	tracker->rmp_thread = td;
444 	tracker->rmp_rmlock = rm;
445 
446 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
447 		THREAD_NO_SLEEPING();
448 
449 	td->td_critnest++;	/* critical_enter(); */
450 	atomic_interrupt_fence();
451 
452 	pc = cpuid_to_pcpu[td->td_oncpu];
453 	rm_tracker_add(pc, tracker);
454 	sched_pin();
455 
456 	atomic_interrupt_fence();
457 	td->td_critnest--;
458 
459 	/*
460 	 * Fast path to combine two common conditions into a single
461 	 * conditional jump.
462 	 */
463 	if (__predict_true(0 == (td->td_owepreempt |
464 	    CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus))))
465 		return (1);
466 
467 	/* We do not have a read token and need to acquire one. */
468 	return _rm_rlock_hard(rm, tracker, trylock);
469 }
470 
471 static __noinline void
472 _rm_unlock_hard(struct thread *td,struct rm_priotracker *tracker)
473 {
474 
475 	if (td->td_owepreempt) {
476 		td->td_critnest++;
477 		critical_exit();
478 	}
479 
480 	if (!tracker->rmp_flags)
481 		return;
482 
483 	mtx_lock_spin(&rm_spinlock);
484 	LIST_REMOVE(tracker, rmp_qentry);
485 
486 	if (tracker->rmp_flags & RMPF_SIGNAL) {
487 		struct rmlock *rm;
488 		struct turnstile *ts;
489 
490 		rm = tracker->rmp_rmlock;
491 
492 		turnstile_chain_lock(&rm->lock_object);
493 		mtx_unlock_spin(&rm_spinlock);
494 
495 		ts = turnstile_lookup(&rm->lock_object);
496 
497 		turnstile_signal(ts, TS_EXCLUSIVE_QUEUE);
498 		turnstile_unpend(ts);
499 		turnstile_chain_unlock(&rm->lock_object);
500 	} else
501 		mtx_unlock_spin(&rm_spinlock);
502 }
503 
504 void
505 _rm_runlock(struct rmlock *rm, struct rm_priotracker *tracker)
506 {
507 	struct pcpu *pc;
508 	struct thread *td = tracker->rmp_thread;
509 
510 	if (SCHEDULER_STOPPED())
511 		return;
512 
513 	td->td_critnest++;	/* critical_enter(); */
514 	atomic_interrupt_fence();
515 
516 	pc = cpuid_to_pcpu[td->td_oncpu];
517 	rm_tracker_remove(pc, tracker);
518 
519 	atomic_interrupt_fence();
520 	td->td_critnest--;
521 	sched_unpin();
522 
523 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
524 		THREAD_SLEEPING_OK();
525 
526 	if (__predict_true(0 == (td->td_owepreempt | tracker->rmp_flags)))
527 		return;
528 
529 	_rm_unlock_hard(td, tracker);
530 }
531 
532 void
533 _rm_wlock(struct rmlock *rm)
534 {
535 	struct rm_priotracker *prio;
536 	struct turnstile *ts;
537 	cpuset_t readcpus;
538 
539 	if (SCHEDULER_STOPPED())
540 		return;
541 
542 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
543 		sx_xlock(&rm->rm_lock_sx);
544 	else
545 		mtx_lock(&rm->rm_lock_mtx);
546 
547 	if (CPU_CMP(&rm->rm_writecpus, &all_cpus)) {
548 		/* Get all read tokens back */
549 		readcpus = all_cpus;
550 		CPU_ANDNOT(&readcpus, &readcpus, &rm->rm_writecpus);
551 		rm->rm_writecpus = all_cpus;
552 
553 		/*
554 		 * Assumes rm->rm_writecpus update is visible on other CPUs
555 		 * before rm_cleanIPI is called.
556 		 */
557 #ifdef SMP
558 		smp_rendezvous_cpus(readcpus,
559 		    smp_no_rendezvous_barrier,
560 		    rm_cleanIPI,
561 		    smp_no_rendezvous_barrier,
562 		    rm);
563 
564 #else
565 		rm_cleanIPI(rm);
566 #endif
567 
568 		mtx_lock_spin(&rm_spinlock);
569 		while ((prio = LIST_FIRST(&rm->rm_activeReaders)) != NULL) {
570 			ts = turnstile_trywait(&rm->lock_object);
571 			prio->rmp_flags = RMPF_ONQUEUE | RMPF_SIGNAL;
572 			mtx_unlock_spin(&rm_spinlock);
573 			turnstile_wait(ts, prio->rmp_thread,
574 			    TS_EXCLUSIVE_QUEUE);
575 			mtx_lock_spin(&rm_spinlock);
576 		}
577 		mtx_unlock_spin(&rm_spinlock);
578 	}
579 }
580 
581 void
582 _rm_wunlock(struct rmlock *rm)
583 {
584 
585 	if (rm->lock_object.lo_flags & LO_SLEEPABLE)
586 		sx_xunlock(&rm->rm_lock_sx);
587 	else
588 		mtx_unlock(&rm->rm_lock_mtx);
589 }
590 
591 #if LOCK_DEBUG > 0
592 
593 void
594 _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
595 {
596 
597 	if (SCHEDULER_STOPPED())
598 		return;
599 
600 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
601 	    ("rm_wlock() by idle thread %p on rmlock %s @ %s:%d",
602 	    curthread, rm->lock_object.lo_name, file, line));
603 	KASSERT(!rm_destroyed(rm),
604 	    ("rm_wlock() of destroyed rmlock @ %s:%d", file, line));
605 	_rm_assert(rm, RA_UNLOCKED, file, line);
606 
607 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE,
608 	    file, line, NULL);
609 
610 	_rm_wlock(rm);
611 
612 	LOCK_LOG_LOCK("RMWLOCK", &rm->lock_object, 0, 0, file, line);
613 	WITNESS_LOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
614 	TD_LOCKS_INC(curthread);
615 }
616 
617 void
618 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
619 {
620 
621 	if (SCHEDULER_STOPPED())
622 		return;
623 
624 	KASSERT(!rm_destroyed(rm),
625 	    ("rm_wunlock() of destroyed rmlock @ %s:%d", file, line));
626 	_rm_assert(rm, RA_WLOCKED, file, line);
627 	WITNESS_UNLOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
628 	LOCK_LOG_LOCK("RMWUNLOCK", &rm->lock_object, 0, 0, file, line);
629 	_rm_wunlock(rm);
630 	TD_LOCKS_DEC(curthread);
631 }
632 
633 int
634 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
635     int trylock, const char *file, int line)
636 {
637 
638 	if (SCHEDULER_STOPPED())
639 		return (1);
640 
641 #ifdef INVARIANTS
642 	if (!(rm->lock_object.lo_flags & LO_RECURSABLE) && !trylock) {
643 		critical_enter();
644 		KASSERT(rm_trackers_present(get_pcpu(), rm,
645 		    curthread) == 0,
646 		    ("rm_rlock: recursed on non-recursive rmlock %s @ %s:%d\n",
647 		    rm->lock_object.lo_name, file, line));
648 		critical_exit();
649 	}
650 #endif
651 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
652 	    ("rm_rlock() by idle thread %p on rmlock %s @ %s:%d",
653 	    curthread, rm->lock_object.lo_name, file, line));
654 	KASSERT(!rm_destroyed(rm),
655 	    ("rm_rlock() of destroyed rmlock @ %s:%d", file, line));
656 	if (!trylock) {
657 		KASSERT(!rm_wowned(rm),
658 		    ("rm_rlock: wlock already held for %s @ %s:%d",
659 		    rm->lock_object.lo_name, file, line));
660 		WITNESS_CHECKORDER(&rm->lock_object,
661 		    LOP_NEWORDER | LOP_NOSLEEP, file, line, NULL);
662 	}
663 
664 	if (_rm_rlock(rm, tracker, trylock)) {
665 		if (trylock)
666 			LOCK_LOG_TRY("RMRLOCK", &rm->lock_object, 0, 1, file,
667 			    line);
668 		else
669 			LOCK_LOG_LOCK("RMRLOCK", &rm->lock_object, 0, 0, file,
670 			    line);
671 		WITNESS_LOCK(&rm->lock_object, LOP_NOSLEEP, file, line);
672 		TD_LOCKS_INC(curthread);
673 		return (1);
674 	} else if (trylock)
675 		LOCK_LOG_TRY("RMRLOCK", &rm->lock_object, 0, 0, file, line);
676 
677 	return (0);
678 }
679 
680 void
681 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
682     const char *file, int line)
683 {
684 
685 	if (SCHEDULER_STOPPED())
686 		return;
687 
688 	KASSERT(!rm_destroyed(rm),
689 	    ("rm_runlock() of destroyed rmlock @ %s:%d", file, line));
690 	_rm_assert(rm, RA_RLOCKED, file, line);
691 	WITNESS_UNLOCK(&rm->lock_object, 0, file, line);
692 	LOCK_LOG_LOCK("RMRUNLOCK", &rm->lock_object, 0, 0, file, line);
693 	_rm_runlock(rm, tracker);
694 	TD_LOCKS_DEC(curthread);
695 }
696 
697 #else
698 
699 /*
700  * Just strip out file and line arguments if no lock debugging is enabled in
701  * the kernel - we are called from a kernel module.
702  */
703 void
704 _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
705 {
706 
707 	_rm_wlock(rm);
708 }
709 
710 void
711 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
712 {
713 
714 	_rm_wunlock(rm);
715 }
716 
717 int
718 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
719     int trylock, const char *file, int line)
720 {
721 
722 	return _rm_rlock(rm, tracker, trylock);
723 }
724 
725 void
726 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
727     const char *file, int line)
728 {
729 
730 	_rm_runlock(rm, tracker);
731 }
732 
733 #endif
734 
735 #ifdef INVARIANT_SUPPORT
736 #ifndef INVARIANTS
737 #undef _rm_assert
738 #endif
739 
740 /*
741  * Note that this does not need to use witness_assert() for read lock
742  * assertions since an exact count of read locks held by this thread
743  * is computable.
744  */
745 void
746 _rm_assert(const struct rmlock *rm, int what, const char *file, int line)
747 {
748 	int count;
749 
750 	if (SCHEDULER_STOPPED())
751 		return;
752 	switch (what) {
753 	case RA_LOCKED:
754 	case RA_LOCKED | RA_RECURSED:
755 	case RA_LOCKED | RA_NOTRECURSED:
756 	case RA_RLOCKED:
757 	case RA_RLOCKED | RA_RECURSED:
758 	case RA_RLOCKED | RA_NOTRECURSED:
759 		/*
760 		 * Handle the write-locked case.  Unlike other
761 		 * primitives, writers can never recurse.
762 		 */
763 		if (rm_wowned(rm)) {
764 			if (what & RA_RLOCKED)
765 				panic("Lock %s exclusively locked @ %s:%d\n",
766 				    rm->lock_object.lo_name, file, line);
767 			if (what & RA_RECURSED)
768 				panic("Lock %s not recursed @ %s:%d\n",
769 				    rm->lock_object.lo_name, file, line);
770 			break;
771 		}
772 
773 		critical_enter();
774 		count = rm_trackers_present(get_pcpu(), rm, curthread);
775 		critical_exit();
776 
777 		if (count == 0)
778 			panic("Lock %s not %slocked @ %s:%d\n",
779 			    rm->lock_object.lo_name, (what & RA_RLOCKED) ?
780 			    "read " : "", file, line);
781 		if (count > 1) {
782 			if (what & RA_NOTRECURSED)
783 				panic("Lock %s recursed @ %s:%d\n",
784 				    rm->lock_object.lo_name, file, line);
785 		} else if (what & RA_RECURSED)
786 			panic("Lock %s not recursed @ %s:%d\n",
787 			    rm->lock_object.lo_name, file, line);
788 		break;
789 	case RA_WLOCKED:
790 		if (!rm_wowned(rm))
791 			panic("Lock %s not exclusively locked @ %s:%d\n",
792 			    rm->lock_object.lo_name, file, line);
793 		break;
794 	case RA_UNLOCKED:
795 		if (rm_wowned(rm))
796 			panic("Lock %s exclusively locked @ %s:%d\n",
797 			    rm->lock_object.lo_name, file, line);
798 
799 		critical_enter();
800 		count = rm_trackers_present(get_pcpu(), rm, curthread);
801 		critical_exit();
802 
803 		if (count != 0)
804 			panic("Lock %s read locked @ %s:%d\n",
805 			    rm->lock_object.lo_name, file, line);
806 		break;
807 	default:
808 		panic("Unknown rm lock assertion: %d @ %s:%d", what, file,
809 		    line);
810 	}
811 }
812 #endif /* INVARIANT_SUPPORT */
813 
814 #ifdef DDB
815 static void
816 print_tracker(struct rm_priotracker *tr)
817 {
818 	struct thread *td;
819 
820 	td = tr->rmp_thread;
821 	db_printf("   thread %p (tid %d, pid %d, \"%s\") {", td, td->td_tid,
822 	    td->td_proc->p_pid, td->td_name);
823 	if (tr->rmp_flags & RMPF_ONQUEUE) {
824 		db_printf("ONQUEUE");
825 		if (tr->rmp_flags & RMPF_SIGNAL)
826 			db_printf(",SIGNAL");
827 	} else
828 		db_printf("0");
829 	db_printf("}\n");
830 }
831 
832 static void
833 db_show_rm(const struct lock_object *lock)
834 {
835 	struct rm_priotracker *tr;
836 	struct rm_queue *queue;
837 	const struct rmlock *rm;
838 	struct lock_class *lc;
839 	struct pcpu *pc;
840 
841 	rm = (const struct rmlock *)lock;
842 	db_printf(" writecpus: ");
843 	ddb_display_cpuset(__DEQUALIFY(const cpuset_t *, &rm->rm_writecpus));
844 	db_printf("\n");
845 	db_printf(" per-CPU readers:\n");
846 	STAILQ_FOREACH(pc, &cpuhead, pc_allcpu)
847 		for (queue = pc->pc_rm_queue.rmq_next;
848 		    queue != &pc->pc_rm_queue; queue = queue->rmq_next) {
849 			tr = (struct rm_priotracker *)queue;
850 			if (tr->rmp_rmlock == rm)
851 				print_tracker(tr);
852 		}
853 	db_printf(" active readers:\n");
854 	LIST_FOREACH(tr, &rm->rm_activeReaders, rmp_qentry)
855 		print_tracker(tr);
856 	lc = LOCK_CLASS(&rm->rm_wlock_object);
857 	db_printf("Backing write-lock (%s):\n", lc->lc_name);
858 	lc->lc_ddb_show(&rm->rm_wlock_object);
859 }
860 #endif
861 
862 /*
863  * Read-mostly sleepable locks.
864  *
865  * These primitives allow both readers and writers to sleep. However, neither
866  * readers nor writers are tracked and subsequently there is no priority
867  * propagation.
868  *
869  * They are intended to be only used when write-locking is almost never needed
870  * (e.g., they can guard against unloading a kernel module) while read-locking
871  * happens all the time.
872  *
873  * Concurrent writers take turns taking the lock while going off cpu. If this is
874  * of concern for your usecase, this is not the right primitive.
875  *
876  * Neither rms_rlock nor rms_runlock use thread fences. Instead interrupt
877  * fences are inserted to ensure ordering with the code executed in the IPI
878  * handler.
879  *
880  * No attempt is made to track which CPUs read locked at least once,
881  * consequently write locking sends IPIs to all of them. This will become a
882  * problem at some point. The easiest way to lessen it is to provide a bitmap.
883  */
884 
885 #define	RMS_NOOWNER	((void *)0x1)
886 #define	RMS_TRANSIENT	((void *)0x2)
887 #define	RMS_FLAGMASK	0xf
888 
889 struct rmslock_pcpu {
890 	int influx;
891 	int readers;
892 };
893 
894 _Static_assert(sizeof(struct rmslock_pcpu) == 8, "bad size");
895 
896 /*
897  * Internal routines
898  */
899 static struct rmslock_pcpu *
900 rms_int_pcpu(struct rmslock *rms)
901 {
902 
903 	CRITICAL_ASSERT(curthread);
904 	return (zpcpu_get(rms->pcpu));
905 }
906 
907 static struct rmslock_pcpu *
908 rms_int_remote_pcpu(struct rmslock *rms, int cpu)
909 {
910 
911 	return (zpcpu_get_cpu(rms->pcpu, cpu));
912 }
913 
914 static void
915 rms_int_influx_enter(struct rmslock *rms, struct rmslock_pcpu *pcpu)
916 {
917 
918 	CRITICAL_ASSERT(curthread);
919 	MPASS(pcpu->influx == 0);
920 	pcpu->influx = 1;
921 }
922 
923 static void
924 rms_int_influx_exit(struct rmslock *rms, struct rmslock_pcpu *pcpu)
925 {
926 
927 	CRITICAL_ASSERT(curthread);
928 	MPASS(pcpu->influx == 1);
929 	pcpu->influx = 0;
930 }
931 
932 #ifdef INVARIANTS
933 static void
934 rms_int_debug_readers_inc(struct rmslock *rms)
935 {
936 	int old;
937 	old = atomic_fetchadd_int(&rms->debug_readers, 1);
938 	KASSERT(old >= 0, ("%s: bad readers count %d\n", __func__, old));
939 }
940 
941 static void
942 rms_int_debug_readers_dec(struct rmslock *rms)
943 {
944 	int old;
945 
946 	old = atomic_fetchadd_int(&rms->debug_readers, -1);
947 	KASSERT(old > 0, ("%s: bad readers count %d\n", __func__, old));
948 }
949 #else
950 static void
951 rms_int_debug_readers_inc(struct rmslock *rms)
952 {
953 }
954 
955 static void
956 rms_int_debug_readers_dec(struct rmslock *rms)
957 {
958 }
959 #endif
960 
961 static void
962 rms_int_readers_inc(struct rmslock *rms, struct rmslock_pcpu *pcpu)
963 {
964 
965 	CRITICAL_ASSERT(curthread);
966 	rms_int_debug_readers_inc(rms);
967 	pcpu->readers++;
968 }
969 
970 static void
971 rms_int_readers_dec(struct rmslock *rms, struct rmslock_pcpu *pcpu)
972 {
973 
974 	CRITICAL_ASSERT(curthread);
975 	rms_int_debug_readers_dec(rms);
976 	pcpu->readers--;
977 }
978 
979 /*
980  * Public API
981  */
982 void
983 rms_init(struct rmslock *rms, const char *name)
984 {
985 
986 	rms->owner = RMS_NOOWNER;
987 	rms->writers = 0;
988 	rms->readers = 0;
989 	rms->debug_readers = 0;
990 	mtx_init(&rms->mtx, name, NULL, MTX_DEF | MTX_NEW);
991 	rms->pcpu = uma_zalloc_pcpu(pcpu_zone_8, M_WAITOK | M_ZERO);
992 }
993 
994 void
995 rms_destroy(struct rmslock *rms)
996 {
997 
998 	MPASS(rms->writers == 0);
999 	MPASS(rms->readers == 0);
1000 	mtx_destroy(&rms->mtx);
1001 	uma_zfree_pcpu(pcpu_zone_8, rms->pcpu);
1002 }
1003 
1004 static void __noinline
1005 rms_rlock_fallback(struct rmslock *rms)
1006 {
1007 
1008 	rms_int_influx_exit(rms, rms_int_pcpu(rms));
1009 	critical_exit();
1010 
1011 	mtx_lock(&rms->mtx);
1012 	while (rms->writers > 0)
1013 		msleep(&rms->readers, &rms->mtx, PUSER - 1, mtx_name(&rms->mtx), 0);
1014 	critical_enter();
1015 	rms_int_readers_inc(rms, rms_int_pcpu(rms));
1016 	mtx_unlock(&rms->mtx);
1017 	critical_exit();
1018 	TD_LOCKS_INC(curthread);
1019 }
1020 
1021 void
1022 rms_rlock(struct rmslock *rms)
1023 {
1024 	struct rmslock_pcpu *pcpu;
1025 
1026 	rms_assert_rlock_ok(rms);
1027 	MPASS(atomic_load_ptr(&rms->owner) != curthread);
1028 
1029 	critical_enter();
1030 	pcpu = rms_int_pcpu(rms);
1031 	rms_int_influx_enter(rms, pcpu);
1032 	atomic_interrupt_fence();
1033 	if (__predict_false(rms->writers > 0)) {
1034 		rms_rlock_fallback(rms);
1035 		return;
1036 	}
1037 	atomic_interrupt_fence();
1038 	rms_int_readers_inc(rms, pcpu);
1039 	atomic_interrupt_fence();
1040 	rms_int_influx_exit(rms, pcpu);
1041 	critical_exit();
1042 	TD_LOCKS_INC(curthread);
1043 }
1044 
1045 int
1046 rms_try_rlock(struct rmslock *rms)
1047 {
1048 	struct rmslock_pcpu *pcpu;
1049 
1050 	MPASS(atomic_load_ptr(&rms->owner) != curthread);
1051 
1052 	critical_enter();
1053 	pcpu = rms_int_pcpu(rms);
1054 	rms_int_influx_enter(rms, pcpu);
1055 	atomic_interrupt_fence();
1056 	if (__predict_false(rms->writers > 0)) {
1057 		rms_int_influx_exit(rms, pcpu);
1058 		critical_exit();
1059 		return (0);
1060 	}
1061 	atomic_interrupt_fence();
1062 	rms_int_readers_inc(rms, pcpu);
1063 	atomic_interrupt_fence();
1064 	rms_int_influx_exit(rms, pcpu);
1065 	critical_exit();
1066 	TD_LOCKS_INC(curthread);
1067 	return (1);
1068 }
1069 
1070 static void __noinline
1071 rms_runlock_fallback(struct rmslock *rms)
1072 {
1073 
1074 	rms_int_influx_exit(rms, rms_int_pcpu(rms));
1075 	critical_exit();
1076 
1077 	mtx_lock(&rms->mtx);
1078 	MPASS(rms->writers > 0);
1079 	MPASS(rms->readers > 0);
1080 	MPASS(rms->debug_readers == rms->readers);
1081 	rms_int_debug_readers_dec(rms);
1082 	rms->readers--;
1083 	if (rms->readers == 0)
1084 		wakeup_one(&rms->writers);
1085 	mtx_unlock(&rms->mtx);
1086 	TD_LOCKS_DEC(curthread);
1087 }
1088 
1089 void
1090 rms_runlock(struct rmslock *rms)
1091 {
1092 	struct rmslock_pcpu *pcpu;
1093 
1094 	critical_enter();
1095 	pcpu = rms_int_pcpu(rms);
1096 	rms_int_influx_enter(rms, pcpu);
1097 	atomic_interrupt_fence();
1098 	if (__predict_false(rms->writers > 0)) {
1099 		rms_runlock_fallback(rms);
1100 		return;
1101 	}
1102 	atomic_interrupt_fence();
1103 	rms_int_readers_dec(rms, pcpu);
1104 	atomic_interrupt_fence();
1105 	rms_int_influx_exit(rms, pcpu);
1106 	critical_exit();
1107 	TD_LOCKS_DEC(curthread);
1108 }
1109 
1110 struct rmslock_ipi {
1111 	struct rmslock *rms;
1112 	struct smp_rendezvous_cpus_retry_arg srcra;
1113 };
1114 
1115 static void
1116 rms_action_func(void *arg)
1117 {
1118 	struct rmslock_ipi *rmsipi;
1119 	struct rmslock_pcpu *pcpu;
1120 	struct rmslock *rms;
1121 
1122 	rmsipi = __containerof(arg, struct rmslock_ipi, srcra);
1123 	rms = rmsipi->rms;
1124 	pcpu = rms_int_pcpu(rms);
1125 
1126 	if (pcpu->influx)
1127 		return;
1128 	if (pcpu->readers != 0) {
1129 		atomic_add_int(&rms->readers, pcpu->readers);
1130 		pcpu->readers = 0;
1131 	}
1132 	smp_rendezvous_cpus_done(arg);
1133 }
1134 
1135 static void
1136 rms_wait_func(void *arg, int cpu)
1137 {
1138 	struct rmslock_ipi *rmsipi;
1139 	struct rmslock_pcpu *pcpu;
1140 	struct rmslock *rms;
1141 
1142 	rmsipi = __containerof(arg, struct rmslock_ipi, srcra);
1143 	rms = rmsipi->rms;
1144 	pcpu = rms_int_remote_pcpu(rms, cpu);
1145 
1146 	while (atomic_load_int(&pcpu->influx))
1147 		cpu_spinwait();
1148 }
1149 
1150 #ifdef INVARIANTS
1151 static void
1152 rms_assert_no_pcpu_readers(struct rmslock *rms)
1153 {
1154 	struct rmslock_pcpu *pcpu;
1155 	int cpu;
1156 
1157 	CPU_FOREACH(cpu) {
1158 		pcpu = rms_int_remote_pcpu(rms, cpu);
1159 		if (pcpu->readers != 0) {
1160 			panic("%s: got %d readers on cpu %d\n", __func__,
1161 			    pcpu->readers, cpu);
1162 		}
1163 	}
1164 }
1165 #else
1166 static void
1167 rms_assert_no_pcpu_readers(struct rmslock *rms)
1168 {
1169 }
1170 #endif
1171 
1172 static void
1173 rms_wlock_switch(struct rmslock *rms)
1174 {
1175 	struct rmslock_ipi rmsipi;
1176 
1177 	MPASS(rms->readers == 0);
1178 	MPASS(rms->writers == 1);
1179 
1180 	rmsipi.rms = rms;
1181 
1182 	smp_rendezvous_cpus_retry(all_cpus,
1183 	    smp_no_rendezvous_barrier,
1184 	    rms_action_func,
1185 	    smp_no_rendezvous_barrier,
1186 	    rms_wait_func,
1187 	    &rmsipi.srcra);
1188 }
1189 
1190 void
1191 rms_wlock(struct rmslock *rms)
1192 {
1193 
1194 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
1195 	MPASS(atomic_load_ptr(&rms->owner) != curthread);
1196 
1197 	mtx_lock(&rms->mtx);
1198 	rms->writers++;
1199 	if (rms->writers > 1) {
1200 		msleep(&rms->owner, &rms->mtx, (PUSER - 1),
1201 		    mtx_name(&rms->mtx), 0);
1202 		MPASS(rms->readers == 0);
1203 		KASSERT(rms->owner == RMS_TRANSIENT,
1204 		    ("%s: unexpected owner value %p\n", __func__,
1205 		    rms->owner));
1206 		goto out_grab;
1207 	}
1208 
1209 	KASSERT(rms->owner == RMS_NOOWNER,
1210 	    ("%s: unexpected owner value %p\n", __func__, rms->owner));
1211 
1212 	rms_wlock_switch(rms);
1213 	rms_assert_no_pcpu_readers(rms);
1214 
1215 	if (rms->readers > 0) {
1216 		msleep(&rms->writers, &rms->mtx, (PUSER - 1),
1217 		    mtx_name(&rms->mtx), 0);
1218 	}
1219 
1220 out_grab:
1221 	rms->owner = curthread;
1222 	rms_assert_no_pcpu_readers(rms);
1223 	mtx_unlock(&rms->mtx);
1224 	MPASS(rms->readers == 0);
1225 	TD_LOCKS_INC(curthread);
1226 }
1227 
1228 void
1229 rms_wunlock(struct rmslock *rms)
1230 {
1231 
1232 	mtx_lock(&rms->mtx);
1233 	KASSERT(rms->owner == curthread,
1234 	    ("%s: unexpected owner value %p\n", __func__, rms->owner));
1235 	MPASS(rms->writers >= 1);
1236 	MPASS(rms->readers == 0);
1237 	rms->writers--;
1238 	if (rms->writers > 0) {
1239 		wakeup_one(&rms->owner);
1240 		rms->owner = RMS_TRANSIENT;
1241 	} else {
1242 		wakeup(&rms->readers);
1243 		rms->owner = RMS_NOOWNER;
1244 	}
1245 	mtx_unlock(&rms->mtx);
1246 	TD_LOCKS_DEC(curthread);
1247 }
1248 
1249 void
1250 rms_unlock(struct rmslock *rms)
1251 {
1252 
1253 	if (rms_wowned(rms))
1254 		rms_wunlock(rms);
1255 	else
1256 		rms_runlock(rms);
1257 }
1258