xref: /freebsd/sys/kern/kern_rmlock.c (revision c6ec7d31830ab1c80edae95ad5e4b9dba10c47ac) 
1 /*-
2  * Copyright (c) 2007 Stephan Uphoff <ups@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the author nor the names of any co-contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 /*
31  * Machine independent bits of reader/writer lock implementation.
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_ddb.h"
38 #include "opt_kdtrace.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 
43 #include <sys/kernel.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 
55 #ifdef DDB
56 #include <ddb/ddb.h>
57 #endif
58 
59 #define RMPF_ONQUEUE	1
60 #define RMPF_SIGNAL	2
61 
62 /*
63  * To support usage of rmlock in CVs and msleep yet another list for the
64  * priority tracker would be needed.  Using this lock for cv and msleep also
65  * does not seem very useful
66  */
67 
68 static void	assert_rm(const struct lock_object *lock, int what);
69 static void	lock_rm(struct lock_object *lock, int how);
70 #ifdef KDTRACE_HOOKS
71 static int	owner_rm(const struct lock_object *lock, struct thread **owner);
72 #endif
73 static int	unlock_rm(struct lock_object *lock);
74 
75 struct lock_class lock_class_rm = {
76 	.lc_name = "rm",
77 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
78 	.lc_assert = assert_rm,
79 #if 0
80 #ifdef DDB
81 	.lc_ddb_show = db_show_rwlock,
82 #endif
83 #endif
84 	.lc_lock = lock_rm,
85 	.lc_unlock = unlock_rm,
86 #ifdef KDTRACE_HOOKS
87 	.lc_owner = owner_rm,
88 #endif
89 };
90 
91 static void
92 assert_rm(const struct lock_object *lock, int what)
93 {
94 
95 	panic("assert_rm called");
96 }
97 
98 static void
99 lock_rm(struct lock_object *lock, int how)
100 {
101 
102 	panic("lock_rm called");
103 }
104 
105 static int
106 unlock_rm(struct lock_object *lock)
107 {
108 
109 	panic("unlock_rm called");
110 }
111 
112 #ifdef KDTRACE_HOOKS
113 static int
114 owner_rm(const struct lock_object *lock, struct thread **owner)
115 {
116 
117 	panic("owner_rm called");
118 }
119 #endif
120 
121 static struct mtx rm_spinlock;
122 
123 MTX_SYSINIT(rm_spinlock, &rm_spinlock, "rm_spinlock", MTX_SPIN);
124 
125 /*
126  * Add or remove tracker from per-cpu list.
127  *
128  * The per-cpu list can be traversed at any time in forward direction from an
129  * interrupt on the *local* cpu.
130  */
131 static void inline
132 rm_tracker_add(struct pcpu *pc, struct rm_priotracker *tracker)
133 {
134 	struct rm_queue *next;
135 
136 	/* Initialize all tracker pointers */
137 	tracker->rmp_cpuQueue.rmq_prev = &pc->pc_rm_queue;
138 	next = pc->pc_rm_queue.rmq_next;
139 	tracker->rmp_cpuQueue.rmq_next = next;
140 
141 	/* rmq_prev is not used during froward traversal. */
142 	next->rmq_prev = &tracker->rmp_cpuQueue;
143 
144 	/* Update pointer to first element. */
145 	pc->pc_rm_queue.rmq_next = &tracker->rmp_cpuQueue;
146 }
147 
148 static void inline
149 rm_tracker_remove(struct pcpu *pc, struct rm_priotracker *tracker)
150 {
151 	struct rm_queue *next, *prev;
152 
153 	next = tracker->rmp_cpuQueue.rmq_next;
154 	prev = tracker->rmp_cpuQueue.rmq_prev;
155 
156 	/* Not used during forward traversal. */
157 	next->rmq_prev = prev;
158 
159 	/* Remove from list. */
160 	prev->rmq_next = next;
161 }
162 
163 static void
164 rm_cleanIPI(void *arg)
165 {
166 	struct pcpu *pc;
167 	struct rmlock *rm = arg;
168 	struct rm_priotracker *tracker;
169 	struct rm_queue *queue;
170 	pc = pcpu_find(curcpu);
171 
172 	for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue;
173 	    queue = queue->rmq_next) {
174 		tracker = (struct rm_priotracker *)queue;
175 		if (tracker->rmp_rmlock == rm && tracker->rmp_flags == 0) {
176 			tracker->rmp_flags = RMPF_ONQUEUE;
177 			mtx_lock_spin(&rm_spinlock);
178 			LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker,
179 			    rmp_qentry);
180 			mtx_unlock_spin(&rm_spinlock);
181 		}
182 	}
183 }
184 
185 CTASSERT((RM_SLEEPABLE & LO_CLASSFLAGS) == RM_SLEEPABLE);
186 
187 void
188 rm_init_flags(struct rmlock *rm, const char *name, int opts)
189 {
190 	int liflags;
191 
192 	liflags = 0;
193 	if (!(opts & RM_NOWITNESS))
194 		liflags |= LO_WITNESS;
195 	if (opts & RM_RECURSE)
196 		liflags |= LO_RECURSABLE;
197 	rm->rm_writecpus = all_cpus;
198 	LIST_INIT(&rm->rm_activeReaders);
199 	if (opts & RM_SLEEPABLE) {
200 		liflags |= RM_SLEEPABLE;
201 		sx_init_flags(&rm->rm_lock_sx, "rmlock_sx", SX_RECURSE);
202 	} else
203 		mtx_init(&rm->rm_lock_mtx, name, "rmlock_mtx", MTX_NOWITNESS);
204 	lock_init(&rm->lock_object, &lock_class_rm, name, NULL, liflags);
205 }
206 
207 void
208 rm_init(struct rmlock *rm, const char *name)
209 {
210 
211 	rm_init_flags(rm, name, 0);
212 }
213 
214 void
215 rm_destroy(struct rmlock *rm)
216 {
217 
218 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
219 		sx_destroy(&rm->rm_lock_sx);
220 	else
221 		mtx_destroy(&rm->rm_lock_mtx);
222 	lock_destroy(&rm->lock_object);
223 }
224 
225 int
226 rm_wowned(const struct rmlock *rm)
227 {
228 
229 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
230 		return (sx_xlocked(&rm->rm_lock_sx));
231 	else
232 		return (mtx_owned(&rm->rm_lock_mtx));
233 }
234 
235 void
236 rm_sysinit(void *arg)
237 {
238 	struct rm_args *args = arg;
239 
240 	rm_init(args->ra_rm, args->ra_desc);
241 }
242 
243 void
244 rm_sysinit_flags(void *arg)
245 {
246 	struct rm_args_flags *args = arg;
247 
248 	rm_init_flags(args->ra_rm, args->ra_desc, args->ra_opts);
249 }
250 
251 static int
252 _rm_rlock_hard(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
253 {
254 	struct pcpu *pc;
255 	struct rm_queue *queue;
256 	struct rm_priotracker *atracker;
257 
258 	critical_enter();
259 	pc = pcpu_find(curcpu);
260 
261 	/* Check if we just need to do a proper critical_exit. */
262 	if (!CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)) {
263 		critical_exit();
264 		return (1);
265 	}
266 
267 	/* Remove our tracker from the per-cpu list. */
268 	rm_tracker_remove(pc, tracker);
269 
270 	/* Check to see if the IPI granted us the lock after all. */
271 	if (tracker->rmp_flags) {
272 		/* Just add back tracker - we hold the lock. */
273 		rm_tracker_add(pc, tracker);
274 		critical_exit();
275 		return (1);
276 	}
277 
278 	/*
279 	 * We allow readers to aquire a lock even if a writer is blocked if
280 	 * the lock is recursive and the reader already holds the lock.
281 	 */
282 	if ((rm->lock_object.lo_flags & LO_RECURSABLE) != 0) {
283 		/*
284 		 * Just grant the lock if this thread already has a tracker
285 		 * for this lock on the per-cpu queue.
286 		 */
287 		for (queue = pc->pc_rm_queue.rmq_next;
288 		    queue != &pc->pc_rm_queue; queue = queue->rmq_next) {
289 			atracker = (struct rm_priotracker *)queue;
290 			if ((atracker->rmp_rmlock == rm) &&
291 			    (atracker->rmp_thread == tracker->rmp_thread)) {
292 				mtx_lock_spin(&rm_spinlock);
293 				LIST_INSERT_HEAD(&rm->rm_activeReaders,
294 				    tracker, rmp_qentry);
295 				tracker->rmp_flags = RMPF_ONQUEUE;
296 				mtx_unlock_spin(&rm_spinlock);
297 				rm_tracker_add(pc, tracker);
298 				critical_exit();
299 				return (1);
300 			}
301 		}
302 	}
303 
304 	sched_unpin();
305 	critical_exit();
306 
307 	if (trylock) {
308 		if (rm->lock_object.lo_flags & RM_SLEEPABLE) {
309 			if (!sx_try_xlock(&rm->rm_lock_sx))
310 				return (0);
311 		} else {
312 			if (!mtx_trylock(&rm->rm_lock_mtx))
313 				return (0);
314 		}
315 	} else {
316 		if (rm->lock_object.lo_flags & RM_SLEEPABLE)
317 			sx_xlock(&rm->rm_lock_sx);
318 		else
319 			mtx_lock(&rm->rm_lock_mtx);
320 	}
321 
322 	critical_enter();
323 	pc = pcpu_find(curcpu);
324 	CPU_CLR(pc->pc_cpuid, &rm->rm_writecpus);
325 	rm_tracker_add(pc, tracker);
326 	sched_pin();
327 	critical_exit();
328 
329 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
330 		sx_xunlock(&rm->rm_lock_sx);
331 	else
332 		mtx_unlock(&rm->rm_lock_mtx);
333 
334 	return (1);
335 }
336 
337 int
338 _rm_rlock(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
339 {
340 	struct thread *td = curthread;
341 	struct pcpu *pc;
342 
343 	if (SCHEDULER_STOPPED())
344 		return (1);
345 
346 	tracker->rmp_flags  = 0;
347 	tracker->rmp_thread = td;
348 	tracker->rmp_rmlock = rm;
349 
350 	td->td_critnest++;	/* critical_enter(); */
351 
352 	__compiler_membar();
353 
354 	pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */
355 
356 	rm_tracker_add(pc, tracker);
357 
358 	sched_pin();
359 
360 	__compiler_membar();
361 
362 	td->td_critnest--;
363 
364 	/*
365 	 * Fast path to combine two common conditions into a single
366 	 * conditional jump.
367 	 */
368 	if (0 == (td->td_owepreempt |
369 	    CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)))
370 		return (1);
371 
372 	/* We do not have a read token and need to acquire one. */
373 	return _rm_rlock_hard(rm, tracker, trylock);
374 }
375 
376 static void
377 _rm_unlock_hard(struct thread *td,struct rm_priotracker *tracker)
378 {
379 
380 	if (td->td_owepreempt) {
381 		td->td_critnest++;
382 		critical_exit();
383 	}
384 
385 	if (!tracker->rmp_flags)
386 		return;
387 
388 	mtx_lock_spin(&rm_spinlock);
389 	LIST_REMOVE(tracker, rmp_qentry);
390 
391 	if (tracker->rmp_flags & RMPF_SIGNAL) {
392 		struct rmlock *rm;
393 		struct turnstile *ts;
394 
395 		rm = tracker->rmp_rmlock;
396 
397 		turnstile_chain_lock(&rm->lock_object);
398 		mtx_unlock_spin(&rm_spinlock);
399 
400 		ts = turnstile_lookup(&rm->lock_object);
401 
402 		turnstile_signal(ts, TS_EXCLUSIVE_QUEUE);
403 		turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
404 		turnstile_chain_unlock(&rm->lock_object);
405 	} else
406 		mtx_unlock_spin(&rm_spinlock);
407 }
408 
409 void
410 _rm_runlock(struct rmlock *rm, struct rm_priotracker *tracker)
411 {
412 	struct pcpu *pc;
413 	struct thread *td = tracker->rmp_thread;
414 
415 	if (SCHEDULER_STOPPED())
416 		return;
417 
418 	td->td_critnest++;	/* critical_enter(); */
419 	pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */
420 	rm_tracker_remove(pc, tracker);
421 	td->td_critnest--;
422 	sched_unpin();
423 
424 	if (0 == (td->td_owepreempt | tracker->rmp_flags))
425 		return;
426 
427 	_rm_unlock_hard(td, tracker);
428 }
429 
430 void
431 _rm_wlock(struct rmlock *rm)
432 {
433 	struct rm_priotracker *prio;
434 	struct turnstile *ts;
435 	cpuset_t readcpus;
436 
437 	if (SCHEDULER_STOPPED())
438 		return;
439 
440 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
441 		sx_xlock(&rm->rm_lock_sx);
442 	else
443 		mtx_lock(&rm->rm_lock_mtx);
444 
445 	if (CPU_CMP(&rm->rm_writecpus, &all_cpus)) {
446 		/* Get all read tokens back */
447 		readcpus = all_cpus;
448 		CPU_NAND(&readcpus, &rm->rm_writecpus);
449 		rm->rm_writecpus = all_cpus;
450 
451 		/*
452 		 * Assumes rm->rm_writecpus update is visible on other CPUs
453 		 * before rm_cleanIPI is called.
454 		 */
455 #ifdef SMP
456 		smp_rendezvous_cpus(readcpus,
457 		    smp_no_rendevous_barrier,
458 		    rm_cleanIPI,
459 		    smp_no_rendevous_barrier,
460 		    rm);
461 
462 #else
463 		rm_cleanIPI(rm);
464 #endif
465 
466 		mtx_lock_spin(&rm_spinlock);
467 		while ((prio = LIST_FIRST(&rm->rm_activeReaders)) != NULL) {
468 			ts = turnstile_trywait(&rm->lock_object);
469 			prio->rmp_flags = RMPF_ONQUEUE | RMPF_SIGNAL;
470 			mtx_unlock_spin(&rm_spinlock);
471 			turnstile_wait(ts, prio->rmp_thread,
472 			    TS_EXCLUSIVE_QUEUE);
473 			mtx_lock_spin(&rm_spinlock);
474 		}
475 		mtx_unlock_spin(&rm_spinlock);
476 	}
477 }
478 
479 void
480 _rm_wunlock(struct rmlock *rm)
481 {
482 
483 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
484 		sx_xunlock(&rm->rm_lock_sx);
485 	else
486 		mtx_unlock(&rm->rm_lock_mtx);
487 }
488 
489 #ifdef LOCK_DEBUG
490 
491 void _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
492 {
493 
494 	if (SCHEDULER_STOPPED())
495 		return;
496 
497 	KASSERT(!TD_IS_IDLETHREAD(curthread),
498 	    ("rm_wlock() by idle thread %p on rmlock %s @ %s:%d",
499 	    curthread, rm->lock_object.lo_name, file, line));
500 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE,
501 	    file, line, NULL);
502 
503 	_rm_wlock(rm);
504 
505 	LOCK_LOG_LOCK("RMWLOCK", &rm->lock_object, 0, 0, file, line);
506 
507 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
508 		WITNESS_LOCK(&rm->rm_lock_sx.lock_object, LOP_EXCLUSIVE,
509 		    file, line);
510 	else
511 		WITNESS_LOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
512 
513 	curthread->td_locks++;
514 
515 }
516 
517 void
518 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
519 {
520 
521 	if (SCHEDULER_STOPPED())
522 		return;
523 
524 	curthread->td_locks--;
525 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
526 		WITNESS_UNLOCK(&rm->rm_lock_sx.lock_object, LOP_EXCLUSIVE,
527 		    file, line);
528 	else
529 		WITNESS_UNLOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
530 	LOCK_LOG_LOCK("RMWUNLOCK", &rm->lock_object, 0, 0, file, line);
531 	_rm_wunlock(rm);
532 }
533 
534 int
535 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
536     int trylock, const char *file, int line)
537 {
538 
539 	if (SCHEDULER_STOPPED())
540 		return (1);
541 
542 	KASSERT(!TD_IS_IDLETHREAD(curthread),
543 	    ("rm_rlock() by idle thread %p on rmlock %s @ %s:%d",
544 	    curthread, rm->lock_object.lo_name, file, line));
545 	if (!trylock && (rm->lock_object.lo_flags & RM_SLEEPABLE))
546 		WITNESS_CHECKORDER(&rm->rm_lock_sx.lock_object, LOP_NEWORDER,
547 		    file, line, NULL);
548 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER, file, line, NULL);
549 
550 	if (_rm_rlock(rm, tracker, trylock)) {
551 		LOCK_LOG_LOCK("RMRLOCK", &rm->lock_object, 0, 0, file, line);
552 
553 		WITNESS_LOCK(&rm->lock_object, 0, file, line);
554 
555 		curthread->td_locks++;
556 
557 		return (1);
558 	}
559 
560 	return (0);
561 }
562 
563 void
564 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
565     const char *file, int line)
566 {
567 
568 	if (SCHEDULER_STOPPED())
569 		return;
570 
571 	curthread->td_locks--;
572 	WITNESS_UNLOCK(&rm->lock_object, 0, file, line);
573 	LOCK_LOG_LOCK("RMRUNLOCK", &rm->lock_object, 0, 0, file, line);
574 	_rm_runlock(rm, tracker);
575 }
576 
577 #else
578 
579 /*
580  * Just strip out file and line arguments if no lock debugging is enabled in
581  * the kernel - we are called from a kernel module.
582  */
583 void
584 _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
585 {
586 
587 	_rm_wlock(rm);
588 }
589 
590 void
591 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
592 {
593 
594 	_rm_wunlock(rm);
595 }
596 
597 int
598 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
599     int trylock, const char *file, int line)
600 {
601 
602 	return _rm_rlock(rm, tracker, trylock);
603 }
604 
605 void
606 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
607     const char *file, int line)
608 {
609 
610 	_rm_runlock(rm, tracker);
611 }
612 
613 #endif
614