xref: /freebsd/sys/kern/kern_rmlock.c (revision a3cf0ef5a295c885c895fabfd56470c0d1db322d) 
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/systm.h>
52 #include <sys/turnstile.h>
53 #include <sys/lock_profile.h>
54 #include <machine/cpu.h>
55 
56 #ifdef DDB
57 #include <ddb/ddb.h>
58 #endif
59 
60 #define RMPF_ONQUEUE	1
61 #define RMPF_SIGNAL	2
62 
63 /*
64  * To support usage of rmlock in CVs and msleep yet another list for the
65  * priority tracker would be needed.  Using this lock for cv and msleep also
66  * does not seem very useful
67  */
68 
69 static __inline void compiler_memory_barrier(void) {
70 	__asm __volatile("":::"memory");
71 }
72 
73 static void	assert_rm(struct lock_object *lock, int what);
74 static void	lock_rm(struct lock_object *lock, int how);
75 #ifdef KDTRACE_HOOKS
76 static int	owner_rm(struct lock_object *lock, struct thread **owner);
77 #endif
78 static int	unlock_rm(struct lock_object *lock);
79 
80 struct lock_class lock_class_rm = {
81 	.lc_name = "rm",
82 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
83 	.lc_assert = assert_rm,
84 #if 0
85 #ifdef DDB
86 	.lc_ddb_show = db_show_rwlock,
87 #endif
88 #endif
89 	.lc_lock = lock_rm,
90 	.lc_unlock = unlock_rm,
91 #ifdef KDTRACE_HOOKS
92 	.lc_owner = owner_rm,
93 #endif
94 };
95 
96 static void
97 assert_rm(struct lock_object *lock, int what)
98 {
99 
100 	panic("assert_rm called");
101 }
102 
103 static void
104 lock_rm(struct lock_object *lock, int how)
105 {
106 
107 	panic("lock_rm called");
108 }
109 
110 static int
111 unlock_rm(struct lock_object *lock)
112 {
113 
114 	panic("unlock_rm called");
115 }
116 
117 #ifdef KDTRACE_HOOKS
118 static int
119 owner_rm(struct lock_object *lock, struct thread **owner)
120 {
121 
122 	panic("owner_rm called");
123 }
124 #endif
125 
126 static struct mtx rm_spinlock;
127 
128 MTX_SYSINIT(rm_spinlock, &rm_spinlock, "rm_spinlock", MTX_SPIN);
129 
130 /*
131  * Add or remove tracker from per-cpu list.
132  *
133  * The per-cpu list can be traversed at any time in forward direction from an
134  * interrupt on the *local* cpu.
135  */
136 static void inline
137 rm_tracker_add(struct pcpu *pc, struct rm_priotracker *tracker)
138 {
139 	struct rm_queue *next;
140 
141 	/* Initialize all tracker pointers */
142 	tracker->rmp_cpuQueue.rmq_prev = &pc->pc_rm_queue;
143 	next = pc->pc_rm_queue.rmq_next;
144 	tracker->rmp_cpuQueue.rmq_next = next;
145 
146 	/* rmq_prev is not used during froward traversal. */
147 	next->rmq_prev = &tracker->rmp_cpuQueue;
148 
149 	/* Update pointer to first element. */
150 	pc->pc_rm_queue.rmq_next = &tracker->rmp_cpuQueue;
151 }
152 
153 static void inline
154 rm_tracker_remove(struct pcpu *pc, struct rm_priotracker *tracker)
155 {
156 	struct rm_queue *next, *prev;
157 
158 	next = tracker->rmp_cpuQueue.rmq_next;
159 	prev = tracker->rmp_cpuQueue.rmq_prev;
160 
161 	/* Not used during forward traversal. */
162 	next->rmq_prev = prev;
163 
164 	/* Remove from list. */
165 	prev->rmq_next = next;
166 }
167 
168 static void
169 rm_cleanIPI(void *arg)
170 {
171 	struct pcpu *pc;
172 	struct rmlock *rm = arg;
173 	struct rm_priotracker *tracker;
174 	struct rm_queue *queue;
175 	pc = pcpu_find(curcpu);
176 
177 	for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue;
178 	    queue = queue->rmq_next) {
179 		tracker = (struct rm_priotracker *)queue;
180 		if (tracker->rmp_rmlock == rm && tracker->rmp_flags == 0) {
181 			tracker->rmp_flags = RMPF_ONQUEUE;
182 			mtx_lock_spin(&rm_spinlock);
183 			LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker,
184 			    rmp_qentry);
185 			mtx_unlock_spin(&rm_spinlock);
186 		}
187 	}
188 }
189 
190 CTASSERT((RM_SLEEPABLE & LO_CLASSFLAGS) == RM_SLEEPABLE);
191 
192 void
193 rm_init_flags(struct rmlock *rm, const char *name, int opts)
194 {
195 	int liflags;
196 
197 	liflags = 0;
198 	if (!(opts & RM_NOWITNESS))
199 		liflags |= LO_WITNESS;
200 	if (opts & RM_RECURSE)
201 		liflags |= LO_RECURSABLE;
202 	rm->rm_writecpus = all_cpus;
203 	LIST_INIT(&rm->rm_activeReaders);
204 	if (opts & RM_SLEEPABLE) {
205 		liflags |= RM_SLEEPABLE;
206 		sx_init_flags(&rm->rm_lock_sx, "rmlock_sx", SX_RECURSE);
207 	} else
208 		mtx_init(&rm->rm_lock_mtx, name, "rmlock_mtx", MTX_NOWITNESS);
209 	lock_init(&rm->lock_object, &lock_class_rm, name, NULL, liflags);
210 }
211 
212 void
213 rm_init(struct rmlock *rm, const char *name)
214 {
215 
216 	rm_init_flags(rm, name, 0);
217 }
218 
219 void
220 rm_destroy(struct rmlock *rm)
221 {
222 
223 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
224 		sx_destroy(&rm->rm_lock_sx);
225 	else
226 		mtx_destroy(&rm->rm_lock_mtx);
227 	lock_destroy(&rm->lock_object);
228 }
229 
230 int
231 rm_wowned(struct rmlock *rm)
232 {
233 
234 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
235 		return (sx_xlocked(&rm->rm_lock_sx));
236 	else
237 		return (mtx_owned(&rm->rm_lock_mtx));
238 }
239 
240 void
241 rm_sysinit(void *arg)
242 {
243 	struct rm_args *args = arg;
244 
245 	rm_init(args->ra_rm, args->ra_desc);
246 }
247 
248 void
249 rm_sysinit_flags(void *arg)
250 {
251 	struct rm_args_flags *args = arg;
252 
253 	rm_init_flags(args->ra_rm, args->ra_desc, args->ra_opts);
254 }
255 
256 static int
257 _rm_rlock_hard(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
258 {
259 	struct pcpu *pc;
260 	struct rm_queue *queue;
261 	struct rm_priotracker *atracker;
262 
263 	critical_enter();
264 	pc = pcpu_find(curcpu);
265 
266 	/* Check if we just need to do a proper critical_exit. */
267 	if (!(pc->pc_cpumask & rm->rm_writecpus)) {
268 		critical_exit();
269 		return (1);
270 	}
271 
272 	/* Remove our tracker from the per-cpu list. */
273 	rm_tracker_remove(pc, tracker);
274 
275 	/* Check to see if the IPI granted us the lock after all. */
276 	if (tracker->rmp_flags) {
277 		/* Just add back tracker - we hold the lock. */
278 		rm_tracker_add(pc, tracker);
279 		critical_exit();
280 		return (1);
281 	}
282 
283 	/*
284 	 * We allow readers to aquire a lock even if a writer is blocked if
285 	 * the lock is recursive and the reader already holds the lock.
286 	 */
287 	if ((rm->lock_object.lo_flags & LO_RECURSABLE) != 0) {
288 		/*
289 		 * Just grant the lock if this thread already has a tracker
290 		 * for this lock on the per-cpu queue.
291 		 */
292 		for (queue = pc->pc_rm_queue.rmq_next;
293 		    queue != &pc->pc_rm_queue; queue = queue->rmq_next) {
294 			atracker = (struct rm_priotracker *)queue;
295 			if ((atracker->rmp_rmlock == rm) &&
296 			    (atracker->rmp_thread == tracker->rmp_thread)) {
297 				mtx_lock_spin(&rm_spinlock);
298 				LIST_INSERT_HEAD(&rm->rm_activeReaders,
299 				    tracker, rmp_qentry);
300 				tracker->rmp_flags = RMPF_ONQUEUE;
301 				mtx_unlock_spin(&rm_spinlock);
302 				rm_tracker_add(pc, tracker);
303 				critical_exit();
304 				return (1);
305 			}
306 		}
307 	}
308 
309 	sched_unpin();
310 	critical_exit();
311 
312 	if (trylock) {
313 		if (rm->lock_object.lo_flags & RM_SLEEPABLE) {
314 			if (!sx_try_xlock(&rm->rm_lock_sx))
315 				return (0);
316 		} else {
317 			if (!mtx_trylock(&rm->rm_lock_mtx))
318 				return (0);
319 		}
320 	} else {
321 		if (rm->lock_object.lo_flags & RM_SLEEPABLE)
322 			sx_xlock(&rm->rm_lock_sx);
323 		else
324 			mtx_lock(&rm->rm_lock_mtx);
325 	}
326 
327 	critical_enter();
328 	pc = pcpu_find(curcpu);
329 	rm->rm_writecpus &= ~pc->pc_cpumask;
330 	rm_tracker_add(pc, tracker);
331 	sched_pin();
332 	critical_exit();
333 
334 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
335 		sx_xunlock(&rm->rm_lock_sx);
336 	else
337 		mtx_unlock(&rm->rm_lock_mtx);
338 
339 	return (1);
340 }
341 
342 int
343 _rm_rlock(struct rmlock *rm, struct rm_priotracker *tracker, int trylock)
344 {
345 	struct thread *td = curthread;
346 	struct pcpu *pc;
347 
348 	tracker->rmp_flags  = 0;
349 	tracker->rmp_thread = td;
350 	tracker->rmp_rmlock = rm;
351 
352 	td->td_critnest++;	/* critical_enter(); */
353 
354 	compiler_memory_barrier();
355 
356 	pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */
357 
358 	rm_tracker_add(pc, tracker);
359 
360 	sched_pin();
361 
362 	compiler_memory_barrier();
363 
364 	td->td_critnest--;
365 
366 	/*
367 	 * Fast path to combine two common conditions into a single
368 	 * conditional jump.
369 	 */
370 	if (0 == (td->td_owepreempt | (rm->rm_writecpus & pc->pc_cpumask)))
371 		return (1);
372 
373 	/* We do not have a read token and need to acquire one. */
374 	return _rm_rlock_hard(rm, tracker, trylock);
375 }
376 
377 static void
378 _rm_unlock_hard(struct thread *td,struct rm_priotracker *tracker)
379 {
380 
381 	if (td->td_owepreempt) {
382 		td->td_critnest++;
383 		critical_exit();
384 	}
385 
386 	if (!tracker->rmp_flags)
387 		return;
388 
389 	mtx_lock_spin(&rm_spinlock);
390 	LIST_REMOVE(tracker, rmp_qentry);
391 
392 	if (tracker->rmp_flags & RMPF_SIGNAL) {
393 		struct rmlock *rm;
394 		struct turnstile *ts;
395 
396 		rm = tracker->rmp_rmlock;
397 
398 		turnstile_chain_lock(&rm->lock_object);
399 		mtx_unlock_spin(&rm_spinlock);
400 
401 		ts = turnstile_lookup(&rm->lock_object);
402 
403 		turnstile_signal(ts, TS_EXCLUSIVE_QUEUE);
404 		turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
405 		turnstile_chain_unlock(&rm->lock_object);
406 	} else
407 		mtx_unlock_spin(&rm_spinlock);
408 }
409 
410 void
411 _rm_runlock(struct rmlock *rm, struct rm_priotracker *tracker)
412 {
413 	struct pcpu *pc;
414 	struct thread *td = tracker->rmp_thread;
415 
416 	td->td_critnest++;	/* critical_enter(); */
417 	pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */
418 	rm_tracker_remove(pc, tracker);
419 	td->td_critnest--;
420 	sched_unpin();
421 
422 	if (0 == (td->td_owepreempt | tracker->rmp_flags))
423 		return;
424 
425 	_rm_unlock_hard(td, tracker);
426 }
427 
428 void
429 _rm_wlock(struct rmlock *rm)
430 {
431 	struct rm_priotracker *prio;
432 	struct turnstile *ts;
433 	cpumask_t readcpus;
434 
435 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
436 		sx_xlock(&rm->rm_lock_sx);
437 	else
438 		mtx_lock(&rm->rm_lock_mtx);
439 
440 	if (rm->rm_writecpus != all_cpus) {
441 		/* Get all read tokens back */
442 
443 		readcpus = all_cpus & (all_cpus & ~rm->rm_writecpus);
444 		rm->rm_writecpus = all_cpus;
445 
446 		/*
447 		 * Assumes rm->rm_writecpus update is visible on other CPUs
448 		 * before rm_cleanIPI is called.
449 		 */
450 #ifdef SMP
451 		smp_rendezvous_cpus(readcpus,
452 		    smp_no_rendevous_barrier,
453 		    rm_cleanIPI,
454 		    smp_no_rendevous_barrier,
455 		    rm);
456 
457 #else
458 		rm_cleanIPI(rm);
459 #endif
460 
461 		mtx_lock_spin(&rm_spinlock);
462 		while ((prio = LIST_FIRST(&rm->rm_activeReaders)) != NULL) {
463 			ts = turnstile_trywait(&rm->lock_object);
464 			prio->rmp_flags = RMPF_ONQUEUE | RMPF_SIGNAL;
465 			mtx_unlock_spin(&rm_spinlock);
466 			turnstile_wait(ts, prio->rmp_thread,
467 			    TS_EXCLUSIVE_QUEUE);
468 			mtx_lock_spin(&rm_spinlock);
469 		}
470 		mtx_unlock_spin(&rm_spinlock);
471 	}
472 }
473 
474 void
475 _rm_wunlock(struct rmlock *rm)
476 {
477 
478 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
479 		sx_xunlock(&rm->rm_lock_sx);
480 	else
481 		mtx_unlock(&rm->rm_lock_mtx);
482 }
483 
484 #ifdef LOCK_DEBUG
485 
486 void _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
487 {
488 
489 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE,
490 	    file, line, NULL);
491 
492 	_rm_wlock(rm);
493 
494 	LOCK_LOG_LOCK("RMWLOCK", &rm->lock_object, 0, 0, file, line);
495 
496 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
497 		WITNESS_LOCK(&rm->rm_lock_sx.lock_object, LOP_EXCLUSIVE,
498 		    file, line);
499 	else
500 		WITNESS_LOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
501 
502 	curthread->td_locks++;
503 
504 }
505 
506 void
507 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
508 {
509 
510 	curthread->td_locks--;
511 	if (rm->lock_object.lo_flags & RM_SLEEPABLE)
512 		WITNESS_UNLOCK(&rm->rm_lock_sx.lock_object, LOP_EXCLUSIVE,
513 		    file, line);
514 	else
515 		WITNESS_UNLOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line);
516 	LOCK_LOG_LOCK("RMWUNLOCK", &rm->lock_object, 0, 0, file, line);
517 	_rm_wunlock(rm);
518 }
519 
520 int
521 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
522     int trylock, const char *file, int line)
523 {
524 	if (!trylock && (rm->lock_object.lo_flags & RM_SLEEPABLE))
525 		WITNESS_CHECKORDER(&rm->rm_lock_sx.lock_object, LOP_NEWORDER,
526 		    file, line, NULL);
527 	WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER, file, line, NULL);
528 
529 	if (_rm_rlock(rm, tracker, trylock)) {
530 		LOCK_LOG_LOCK("RMRLOCK", &rm->lock_object, 0, 0, file, line);
531 
532 		WITNESS_LOCK(&rm->lock_object, 0, file, line);
533 
534 		curthread->td_locks++;
535 
536 		return (1);
537 	}
538 
539 	return (0);
540 }
541 
542 void
543 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
544     const char *file, int line)
545 {
546 
547 	curthread->td_locks--;
548 	WITNESS_UNLOCK(&rm->lock_object, 0, file, line);
549 	LOCK_LOG_LOCK("RMRUNLOCK", &rm->lock_object, 0, 0, file, line);
550 	_rm_runlock(rm, tracker);
551 }
552 
553 #else
554 
555 /*
556  * Just strip out file and line arguments if no lock debugging is enabled in
557  * the kernel - we are called from a kernel module.
558  */
559 void
560 _rm_wlock_debug(struct rmlock *rm, const char *file, int line)
561 {
562 
563 	_rm_wlock(rm);
564 }
565 
566 void
567 _rm_wunlock_debug(struct rmlock *rm, const char *file, int line)
568 {
569 
570 	_rm_wunlock(rm);
571 }
572 
573 int
574 _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
575     int trylock, const char *file, int line)
576 {
577 
578 	return _rm_rlock(rm, tracker, trylock);
579 }
580 
581 void
582 _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker,
583     const char *file, int line)
584 {
585 
586 	_rm_runlock(rm, tracker);
587 }
588 
589 #endif
590