xref: /freebsd/sys/kern/kern_rmlock.c (revision 9a14aa017b21c292740c00ee098195cd46642730) 
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 __inline void compiler_memory_barrier(void) {
69 	__asm __volatile("":::"memory");
70 }
71 
72 static void	assert_rm(const struct lock_object *lock, int what);
73 static void	lock_rm(struct lock_object *lock, int how);
74 #ifdef KDTRACE_HOOKS
75 static int	owner_rm(const struct lock_object *lock, struct thread **owner);
76 #endif
77 static int	unlock_rm(struct lock_object *lock);
78 
79 struct lock_class lock_class_rm = {
80 	.lc_name = "rm",
81 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
82 	.lc_assert = assert_rm,
83 #if 0
84 #ifdef DDB
85 	.lc_ddb_show = db_show_rwlock,
86 #endif
87 #endif
88 	.lc_lock = lock_rm,
89 	.lc_unlock = unlock_rm,
90 #ifdef KDTRACE_HOOKS
91 	.lc_owner = owner_rm,
92 #endif
93 };
94 
95 static void
96 assert_rm(const struct lock_object *lock, int what)
97 {
98 
99 	panic("assert_rm called");
100 }
101 
102 static void
103 lock_rm(struct lock_object *lock, int how)
104 {
105 
106 	panic("lock_rm called");
107 }
108 
109 static int
110 unlock_rm(struct lock_object *lock)
111 {
112 
113 	panic("unlock_rm called");
114 }
115 
116 #ifdef KDTRACE_HOOKS
117 static int
118 owner_rm(const struct lock_object *lock, struct thread **owner)
119 {
120 
121 	panic("owner_rm called");
122 }
123 #endif
124 
125 static struct mtx rm_spinlock;
126 
127 MTX_SYSINIT(rm_spinlock, &rm_spinlock, "rm_spinlock", MTX_SPIN);
128 
129 /*
130  * Add or remove tracker from per-cpu list.
131  *
132  * The per-cpu list can be traversed at any time in forward direction from an
133  * interrupt on the *local* cpu.
134  */
135 static void inline
136 rm_tracker_add(struct pcpu *pc, struct rm_priotracker *tracker)
137 {
138 	struct rm_queue *next;
139 
140 	/* Initialize all tracker pointers */
141 	tracker->rmp_cpuQueue.rmq_prev = &pc->pc_rm_queue;
142 	next = pc->pc_rm_queue.rmq_next;
143 	tracker->rmp_cpuQueue.rmq_next = next;
144 
145 	/* rmq_prev is not used during froward traversal. */
146 	next->rmq_prev = &tracker->rmp_cpuQueue;
147 
148 	/* Update pointer to first element. */
149 	pc->pc_rm_queue.rmq_next = &tracker->rmp_cpuQueue;
150 }
151 
152 static void inline
153 rm_tracker_remove(struct pcpu *pc, struct rm_priotracker *tracker)
154 {
155 	struct rm_queue *next, *prev;
156 
157 	next = tracker->rmp_cpuQueue.rmq_next;
158 	prev = tracker->rmp_cpuQueue.rmq_prev;
159 
160 	/* Not used during forward traversal. */
161 	next->rmq_prev = prev;
162 
163 	/* Remove from list. */
164 	prev->rmq_next = next;
165 }
166 
167 static void
168 rm_cleanIPI(void *arg)
169 {
170 	struct pcpu *pc;
171 	struct rmlock *rm = arg;
172 	struct rm_priotracker *tracker;
173 	struct rm_queue *queue;
174 	pc = pcpu_find(curcpu);
175 
176 	for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue;
177 	    queue = queue->rmq_next) {
178 		tracker = (struct rm_priotracker *)queue;
179 		if (tracker->rmp_rmlock == rm && tracker->rmp_flags == 0) {
180 			tracker->rmp_flags = RMPF_ONQUEUE;
181 			mtx_lock_spin(&rm_spinlock);
182 			LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker,
183 			    rmp_qentry);
184 			mtx_unlock_spin(&rm_spinlock);
185 		}
186 	}
187 }
188 
189 CTASSERT((RM_SLEEPABLE & LO_CLASSFLAGS) == RM_SLEEPABLE);
190 
191 void
192 rm_init_flags(struct rmlock *rm, const char *name, int opts)
193 {
194 	int liflags;
195 
196 	liflags = 0;
197 	if (!(opts & RM_NOWITNESS))
198 		liflags |= LO_WITNESS;
199 	if (opts & RM_RECURSE)
200 		liflags |= LO_RECURSABLE;
201 	rm->rm_writecpus = all_cpus;
202 	LIST_INIT(&rm->rm_activeReaders);
203 	if (opts & RM_SLEEPABLE) {
204 		liflags |= RM_SLEEPABLE;
205 		sx_init_flags(&rm->rm_lock_sx, "rmlock_sx", SX_RECURSE);
206 	} else
207 		mtx_init(&rm->rm_lock_mtx, name, "rmlock_mtx", MTX_NOWITNESS);
208 	lock_init(&rm->lock_object, &lock_class_rm, name, NULL, liflags);
209 }
210 
211 void
212 rm_init(struct rmlock *rm, const char *name)
213 {
214 
215 	rm_init_flags(rm, name, 0);
216 }
217 
218 void
219 rm_destroy(struct rmlock *rm)
220 {
221 
222 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
223 		sx_destroy(&rm->rm_lock_sx);
224 	else
225 		mtx_destroy(&rm->rm_lock_mtx);
226 	lock_destroy(&rm->lock_object);
227 }
228 
229 int
230 rm_wowned(const struct rmlock *rm)
231 {
232 
233 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
234 		return (sx_xlocked(&rm->rm_lock_sx));
235 	else
236 		return (mtx_owned(&rm->rm_lock_mtx));
237 }
238 
239 void
240 rm_sysinit(void *arg)
241 {
242 	struct rm_args *args = arg;
243 
244 	rm_init(args->ra_rm, args->ra_desc);
245 }
246 
247 void
248 rm_sysinit_flags(void *arg)
249 {
250 	struct rm_args_flags *args = arg;
251 
252 	rm_init_flags(args->ra_rm, args->ra_desc, args->ra_opts);
253 }
254 
255 static int
256 _rm_rlock_hard(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
257 {
258 	struct pcpu *pc;
259 	struct rm_queue *queue;
260 	struct rm_priotracker *atracker;
261 
262 	critical_enter();
263 	pc = pcpu_find(curcpu);
264 
265 	/* Check if we just need to do a proper critical_exit. */
266 	if (!CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)) {
267 		critical_exit();
268 		return (1);
269 	}
270 
271 	/* Remove our tracker from the per-cpu list. */
272 	rm_tracker_remove(pc, tracker);
273 
274 	/* Check to see if the IPI granted us the lock after all. */
275 	if (tracker->rmp_flags) {
276 		/* Just add back tracker - we hold the lock. */
277 		rm_tracker_add(pc, tracker);
278 		critical_exit();
279 		return (1);
280 	}
281 
282 	/*
283 	 * We allow readers to aquire a lock even if a writer is blocked if
284 	 * the lock is recursive and the reader already holds the lock.
285 	 */
286 	if ((rm->lock_object.lo_flags & LO_RECURSABLE) != 0) {
287 		/*
288 		 * Just grant the lock if this thread already has a tracker
289 		 * for this lock on the per-cpu queue.
290 		 */
291 		for (queue = pc->pc_rm_queue.rmq_next;
292 		    queue != &pc->pc_rm_queue; queue = queue->rmq_next) {
293 			atracker = (struct rm_priotracker *)queue;
294 			if ((atracker->rmp_rmlock == rm) &&
295 			    (atracker->rmp_thread == tracker->rmp_thread)) {
296 				mtx_lock_spin(&rm_spinlock);
297 				LIST_INSERT_HEAD(&rm->rm_activeReaders,
298 				    tracker, rmp_qentry);
299 				tracker->rmp_flags = RMPF_ONQUEUE;
300 				mtx_unlock_spin(&rm_spinlock);
301 				rm_tracker_add(pc, tracker);
302 				critical_exit();
303 				return (1);
304 			}
305 		}
306 	}
307 
308 	sched_unpin();
309 	critical_exit();
310 
311 	if (trylock) {
312 		if (rm->lock_object.lo_flags & RM_SLEEPABLE) {
313 			if (!sx_try_xlock(&rm->rm_lock_sx))
314 				return (0);
315 		} else {
316 			if (!mtx_trylock(&rm->rm_lock_mtx))
317 				return (0);
318 		}
319 	} else {
320 		if (rm->lock_object.lo_flags & RM_SLEEPABLE)
321 			sx_xlock(&rm->rm_lock_sx);
322 		else
323 			mtx_lock(&rm->rm_lock_mtx);
324 	}
325 
326 	critical_enter();
327 	pc = pcpu_find(curcpu);
328 	CPU_CLR(pc->pc_cpuid, &rm->rm_writecpus);
329 	rm_tracker_add(pc, tracker);
330 	sched_pin();
331 	critical_exit();
332 
333 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
334 		sx_xunlock(&rm->rm_lock_sx);
335 	else
336 		mtx_unlock(&rm->rm_lock_mtx);
337 
338 	return (1);
339 }
340 
341 int
342 _rm_rlock(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
343 {
344 	struct thread *td = curthread;
345 	struct pcpu *pc;
346 
347 	if (SCHEDULER_STOPPED())
348 		return (1);
349 
350 	tracker->rmp_flags  = 0;
351 	tracker->rmp_thread = td;
352 	tracker->rmp_rmlock = rm;
353 
354 	td->td_critnest++;	/* critical_enter(); */
355 
356 	compiler_memory_barrier();
357 
358 	pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */
359 
360 	rm_tracker_add(pc, tracker);
361 
362 	sched_pin();
363 
364 	compiler_memory_barrier();
365 
366 	td->td_critnest--;
367 
368 	/*
369 	 * Fast path to combine two common conditions into a single
370 	 * conditional jump.
371 	 */
372 	if (0 == (td->td_owepreempt |
373 	    CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)))
374 		return (1);
375 
376 	/* We do not have a read token and need to acquire one. */
377 	return _rm_rlock_hard(rm, tracker, trylock);
378 }
379 
380 static void
381 _rm_unlock_hard(struct thread *td,struct rm_priotracker *tracker)
382 {
383 
384 	if (td->td_owepreempt) {
385 		td->td_critnest++;
386 		critical_exit();
387 	}
388 
389 	if (!tracker->rmp_flags)
390 		return;
391 
392 	mtx_lock_spin(&rm_spinlock);
393 	LIST_REMOVE(tracker, rmp_qentry);
394 
395 	if (tracker->rmp_flags & RMPF_SIGNAL) {
396 		struct rmlock *rm;
397 		struct turnstile *ts;
398 
399 		rm = tracker->rmp_rmlock;
400 
401 		turnstile_chain_lock(&rm->lock_object);
402 		mtx_unlock_spin(&rm_spinlock);
403 
404 		ts = turnstile_lookup(&rm->lock_object);
405 
406 		turnstile_signal(ts, TS_EXCLUSIVE_QUEUE);
407 		turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
408 		turnstile_chain_unlock(&rm->lock_object);
409 	} else
410 		mtx_unlock_spin(&rm_spinlock);
411 }
412 
413 void
414 _rm_runlock(struct rmlock *rm, struct rm_priotracker *tracker)
415 {
416 	struct pcpu *pc;
417 	struct thread *td = tracker->rmp_thread;
418 
419 	if (SCHEDULER_STOPPED())
420 		return;
421 
422 	td->td_critnest++;	/* critical_enter(); */
423 	pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */
424 	rm_tracker_remove(pc, tracker);
425 	td->td_critnest--;
426 	sched_unpin();
427 
428 	if (0 == (td->td_owepreempt | tracker->rmp_flags))
429 		return;
430 
431 	_rm_unlock_hard(td, tracker);
432 }
433 
434 void
435 _rm_wlock(struct rmlock *rm)
436 {
437 	struct rm_priotracker *prio;
438 	struct turnstile *ts;
439 	cpuset_t readcpus;
440 
441 	if (SCHEDULER_STOPPED())
442 		return;
443 
444 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
445 		sx_xlock(&rm->rm_lock_sx);
446 	else
447 		mtx_lock(&rm->rm_lock_mtx);
448 
449 	if (CPU_CMP(&rm->rm_writecpus, &all_cpus)) {
450 		/* Get all read tokens back */
451 		readcpus = all_cpus;
452 		CPU_NAND(&readcpus, &rm->rm_writecpus);
453 		rm->rm_writecpus = all_cpus;
454 
455 		/*
456 		 * Assumes rm->rm_writecpus update is visible on other CPUs
457 		 * before rm_cleanIPI is called.
458 		 */
459 #ifdef SMP
460 		smp_rendezvous_cpus(readcpus,
461 		    smp_no_rendevous_barrier,
462 		    rm_cleanIPI,
463 		    smp_no_rendevous_barrier,
464 		    rm);
465 
466 #else
467 		rm_cleanIPI(rm);
468 #endif
469 
470 		mtx_lock_spin(&rm_spinlock);
471 		while ((prio = LIST_FIRST(&rm->rm_activeReaders)) != NULL) {
472 			ts = turnstile_trywait(&rm->lock_object);
473 			prio->rmp_flags = RMPF_ONQUEUE | RMPF_SIGNAL;
474 			mtx_unlock_spin(&rm_spinlock);
475 			turnstile_wait(ts, prio->rmp_thread,
476 			    TS_EXCLUSIVE_QUEUE);
477 			mtx_lock_spin(&rm_spinlock);
478 		}
479 		mtx_unlock_spin(&rm_spinlock);
480 	}
481 }
482 
483 void
484 _rm_wunlock(struct rmlock *rm)
485 {
486 
487 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
488 		sx_xunlock(&rm->rm_lock_sx);
489 	else
490 		mtx_unlock(&rm->rm_lock_mtx);
491 }
492 
493 #ifdef LOCK_DEBUG
494 
495 void _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
496 {
497 
498 	if (SCHEDULER_STOPPED())
499 		return;
500 
501 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE,
502 	    file, line, NULL);
503 
504 	_rm_wlock(rm);
505 
506 	LOCK_LOG_LOCK("RMWLOCK", &rm->lock_object, 0, 0, file, line);
507 
508 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
509 		WITNESS_LOCK(&rm->rm_lock_sx.lock_object, LOP_EXCLUSIVE,
510 		    file, line);
511 	else
512 		WITNESS_LOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
513 
514 	curthread->td_locks++;
515 
516 }
517 
518 void
519 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
520 {
521 
522 	if (SCHEDULER_STOPPED())
523 		return;
524 
525 	curthread->td_locks--;
526 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
527 		WITNESS_UNLOCK(&rm->rm_lock_sx.lock_object, LOP_EXCLUSIVE,
528 		    file, line);
529 	else
530 		WITNESS_UNLOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
531 	LOCK_LOG_LOCK("RMWUNLOCK", &rm->lock_object, 0, 0, file, line);
532 	_rm_wunlock(rm);
533 }
534 
535 int
536 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
537     int trylock, const char *file, int line)
538 {
539 
540 	if (SCHEDULER_STOPPED())
541 		return (1);
542 
543 	if (!trylock && (rm->lock_object.lo_flags & RM_SLEEPABLE))
544 		WITNESS_CHECKORDER(&rm->rm_lock_sx.lock_object, LOP_NEWORDER,
545 		    file, line, NULL);
546 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER, file, line, NULL);
547 
548 	if (_rm_rlock(rm, tracker, trylock)) {
549 		LOCK_LOG_LOCK("RMRLOCK", &rm->lock_object, 0, 0, file, line);
550 
551 		WITNESS_LOCK(&rm->lock_object, 0, file, line);
552 
553 		curthread->td_locks++;
554 
555 		return (1);
556 	}
557 
558 	return (0);
559 }
560 
561 void
562 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
563     const char *file, int line)
564 {
565 
566 	if (SCHEDULER_STOPPED())
567 		return;
568 
569 	curthread->td_locks--;
570 	WITNESS_UNLOCK(&rm->lock_object, 0, file, line);
571 	LOCK_LOG_LOCK("RMRUNLOCK", &rm->lock_object, 0, 0, file, line);
572 	_rm_runlock(rm, tracker);
573 }
574 
575 #else
576 
577 /*
578  * Just strip out file and line arguments if no lock debugging is enabled in
579  * the kernel - we are called from a kernel module.
580  */
581 void
582 _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
583 {
584 
585 	_rm_wlock(rm);
586 }
587 
588 void
589 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
590 {
591 
592 	_rm_wunlock(rm);
593 }
594 
595 int
596 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
597     int trylock, const char *file, int line)
598 {
599 
600 	return _rm_rlock(rm, tracker, trylock);
601 }
602 
603 void
604 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
605     const char *file, int line)
606 {
607 
608 	_rm_runlock(rm, tracker);
609 }
610 
611 #endif
612