xref: /netbsd-src/lib/libpthread/pthread_mutex.c (revision 2de962bd804263c16657f586aa00f1704045df8e)
1 /*	$NetBSD: pthread_mutex.c,v 1.48 2008/04/28 20:23:01 martin Exp $	*/
2 
3 /*-
4  * Copyright (c) 2001, 2003, 2006, 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Nathan J. Williams, by Jason R. Thorpe, and by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __RCSID("$NetBSD: pthread_mutex.c,v 1.48 2008/04/28 20:23:01 martin Exp $");
34 
35 #include <sys/types.h>
36 #include <sys/lwpctl.h>
37 
38 #include <errno.h>
39 #include <limits.h>
40 #include <stdlib.h>
41 #include <string.h>
42 #include <stdio.h>
43 
44 #include "pthread.h"
45 #include "pthread_int.h"
46 
47 #define	pt_nextwaiter			pt_sleep.ptqe_next
48 
49 #define	MUTEX_WAITERS_BIT		((uintptr_t)0x01)
50 #define	MUTEX_RECURSIVE_BIT		((uintptr_t)0x02)
51 #define	MUTEX_DEFERRED_BIT		((uintptr_t)0x04)
52 #define	MUTEX_THREAD			((uintptr_t)-16L)
53 
54 #define	MUTEX_HAS_WAITERS(x)		((uintptr_t)(x) & MUTEX_WAITERS_BIT)
55 #define	MUTEX_RECURSIVE(x)		((uintptr_t)(x) & MUTEX_RECURSIVE_BIT)
56 #define	MUTEX_OWNER(x)			((uintptr_t)(x) & MUTEX_THREAD)
57 
58 #if __GNUC_PREREQ__(3, 0)
59 #define	NOINLINE		__attribute ((noinline))
60 #else
61 #define	NOINLINE		/* nothing */
62 #endif
63 
64 static void	pthread__mutex_wakeup(pthread_t, pthread_mutex_t *);
65 static int	pthread__mutex_lock_slow(pthread_mutex_t *);
66 static int	pthread__mutex_unlock_slow(pthread_mutex_t *);
67 static void	pthread__mutex_pause(void);
68 
69 int		_pthread_mutex_held_np(pthread_mutex_t *);
70 pthread_t	_pthread_mutex_owner_np(pthread_mutex_t *);
71 
72 __weak_alias(pthread_mutex_held_np,_pthread_mutex_held_np)
73 __weak_alias(pthread_mutex_owner_np,_pthread_mutex_owner_np)
74 
75 __strong_alias(__libc_mutex_init,pthread_mutex_init)
76 __strong_alias(__libc_mutex_lock,pthread_mutex_lock)
77 __strong_alias(__libc_mutex_trylock,pthread_mutex_trylock)
78 __strong_alias(__libc_mutex_unlock,pthread_mutex_unlock)
79 __strong_alias(__libc_mutex_destroy,pthread_mutex_destroy)
80 
81 __strong_alias(__libc_mutexattr_init,pthread_mutexattr_init)
82 __strong_alias(__libc_mutexattr_destroy,pthread_mutexattr_destroy)
83 __strong_alias(__libc_mutexattr_settype,pthread_mutexattr_settype)
84 
85 __strong_alias(__libc_thr_once,pthread_once)
86 
87 int
88 pthread_mutex_init(pthread_mutex_t *ptm, const pthread_mutexattr_t *attr)
89 {
90 	intptr_t type;
91 
92 	if (attr == NULL)
93 		type = PTHREAD_MUTEX_NORMAL;
94 	else
95 		type = (intptr_t)attr->ptma_private;
96 
97 	switch (type) {
98 	case PTHREAD_MUTEX_ERRORCHECK:
99 		ptm->ptm_errorcheck = 1;
100 		ptm->ptm_owner = NULL;
101 		break;
102 	case PTHREAD_MUTEX_RECURSIVE:
103 		ptm->ptm_errorcheck = 0;
104 		ptm->ptm_owner = (void *)MUTEX_RECURSIVE_BIT;
105 		break;
106 	default:
107 		ptm->ptm_errorcheck = 0;
108 		ptm->ptm_owner = NULL;
109 		break;
110 	}
111 
112 	ptm->ptm_magic = _PT_MUTEX_MAGIC;
113 	ptm->ptm_waiters = NULL;
114 	ptm->ptm_recursed = 0;
115 
116 	return 0;
117 }
118 
119 
120 int
121 pthread_mutex_destroy(pthread_mutex_t *ptm)
122 {
123 
124 	pthread__error(EINVAL, "Invalid mutex",
125 	    ptm->ptm_magic == _PT_MUTEX_MAGIC);
126 	pthread__error(EBUSY, "Destroying locked mutex",
127 	    MUTEX_OWNER(ptm->ptm_owner) == 0);
128 
129 	ptm->ptm_magic = _PT_MUTEX_DEAD;
130 	return 0;
131 }
132 
133 int
134 pthread_mutex_lock(pthread_mutex_t *ptm)
135 {
136 	pthread_t self;
137 	void *val;
138 
139 	self = pthread__self();
140 	val = atomic_cas_ptr(&ptm->ptm_owner, NULL, self);
141 	if (__predict_true(val == NULL)) {
142 #ifndef PTHREAD__ATOMIC_IS_MEMBAR
143 		membar_enter();
144 #endif
145 		return 0;
146 	}
147 	return pthread__mutex_lock_slow(ptm);
148 }
149 
150 /* We want function call overhead. */
151 NOINLINE static void
152 pthread__mutex_pause(void)
153 {
154 
155 	pthread__smt_pause();
156 }
157 
158 /*
159  * Spin while the holder is running.  'lwpctl' gives us the true
160  * status of the thread.  pt_blocking is set by libpthread in order
161  * to cut out system call and kernel spinlock overhead on remote CPUs
162  * (could represent many thousands of clock cycles).  pt_blocking also
163  * makes this thread yield if the target is calling sched_yield().
164  */
165 NOINLINE static void *
166 pthread__mutex_spin(pthread_mutex_t *ptm, pthread_t owner)
167 {
168 	pthread_t thread;
169 	unsigned int count, i;
170 
171 	for (count = 2;; owner = ptm->ptm_owner) {
172 		thread = (pthread_t)MUTEX_OWNER(owner);
173 		if (thread == NULL)
174 			break;
175 		if (thread->pt_lwpctl->lc_curcpu == LWPCTL_CPU_NONE ||
176 		    thread->pt_blocking)
177 			break;
178 		if (count < 128)
179 			count += count;
180 		for (i = count; i != 0; i--)
181 			pthread__mutex_pause();
182 	}
183 
184 	return owner;
185 }
186 
187 NOINLINE static int
188 pthread__mutex_lock_slow(pthread_mutex_t *ptm)
189 {
190 	void *waiters, *new, *owner, *next;
191 	pthread_t self;
192 
193 	pthread__error(EINVAL, "Invalid mutex",
194 	    ptm->ptm_magic == _PT_MUTEX_MAGIC);
195 
196 	owner = ptm->ptm_owner;
197 	self = pthread__self();
198 
199 	/* Recursive or errorcheck? */
200 	if (MUTEX_OWNER(owner) == (uintptr_t)self) {
201 		if (MUTEX_RECURSIVE(owner)) {
202 			if (ptm->ptm_recursed == INT_MAX)
203 				return EAGAIN;
204 			ptm->ptm_recursed++;
205 			return 0;
206 		}
207 		if (ptm->ptm_errorcheck)
208 			return EDEADLK;
209 	}
210 
211 	for (;; owner = ptm->ptm_owner) {
212 		/* Spin while the owner is running. */
213 		owner = pthread__mutex_spin(ptm, owner);
214 
215 		/* If it has become free, try to acquire it again. */
216 		if (MUTEX_OWNER(owner) == 0) {
217 			do {
218 				new = (void *)
219 				    ((uintptr_t)self | (uintptr_t)owner);
220 				next = atomic_cas_ptr(&ptm->ptm_owner, owner,
221 				    new);
222 				if (next == owner) {
223 #ifndef PTHREAD__ATOMIC_IS_MEMBAR
224 					membar_enter();
225 #endif
226 					return 0;
227 				}
228 				owner = next;
229 			} while (MUTEX_OWNER(owner) == 0);
230 			/*
231 			 * We have lost the race to acquire the mutex.
232 			 * The new owner could be running on another
233 			 * CPU, in which case we should spin and avoid
234 			 * the overhead of blocking.
235 			 */
236 			continue;
237 		}
238 
239 		/*
240 		 * Nope, still held.  Add thread to the list of waiters.
241 		 * Issue a memory barrier to ensure sleeponq/nextwaiter
242 		 * are visible before we enter the waiters list.
243 		 */
244 		self->pt_sleeponq = 1;
245 		for (waiters = ptm->ptm_waiters;; waiters = next) {
246 			self->pt_nextwaiter = waiters;
247 			membar_producer();
248 			next = atomic_cas_ptr(&ptm->ptm_waiters, waiters, self);
249 			if (next == waiters)
250 			    	break;
251 		}
252 
253 		/*
254 		 * Set the waiters bit and block.
255 		 *
256 		 * Note that the mutex can become unlocked before we set
257 		 * the waiters bit.  If that happens it's not safe to sleep
258 		 * as we may never be awoken: we must remove the current
259 		 * thread from the waiters list and try again.
260 		 *
261 		 * Because we are doing this atomically, we can't remove
262 		 * one waiter: we must remove all waiters and awken them,
263 		 * then sleep in _lwp_park() until we have been awoken.
264 		 *
265 		 * Issue a memory barrier to ensure that we are reading
266 		 * the value of ptm_owner/pt_sleeponq after we have entered
267 		 * the waiters list (the CAS itself must be atomic).
268 		 */
269 		membar_consumer();
270 		for (owner = ptm->ptm_owner;; owner = next) {
271 			if (MUTEX_HAS_WAITERS(owner))
272 				break;
273 			if (MUTEX_OWNER(owner) == 0) {
274 				pthread__mutex_wakeup(self, ptm);
275 				break;
276 			}
277 			new = (void *)((uintptr_t)owner | MUTEX_WAITERS_BIT);
278 			next = atomic_cas_ptr(&ptm->ptm_owner, owner, new);
279 			if (next == owner) {
280 				/*
281 				 * pthread_mutex_unlock() can do a
282 				 * non-interlocked CAS.  We cannot
283 				 * know if our attempt to set the
284 				 * waiters bit has succeeded while
285 				 * the holding thread is running.
286 				 * There are many assumptions; see
287 				 * sys/kern/kern_mutex.c for details.
288 				 * In short, we must spin if we see
289 				 * that the holder is running again.
290 				 */
291 				membar_sync();
292 				next = pthread__mutex_spin(ptm, owner);
293 			}
294 		}
295 
296 		/*
297 		 * We may have been awoken by the current thread above,
298 		 * or will be awoken by the current holder of the mutex.
299 		 * The key requirement is that we must not proceed until
300 		 * told that we are no longer waiting (via pt_sleeponq
301 		 * being set to zero).  Otherwise it is unsafe to re-enter
302 		 * the thread onto the waiters list.
303 		 */
304 		while (self->pt_sleeponq) {
305 			self->pt_blocking++;
306 			(void)_lwp_park(NULL, 0,
307 			    __UNVOLATILE(&ptm->ptm_waiters), NULL);
308 			self->pt_blocking--;
309 			membar_sync();
310 		}
311 	}
312 }
313 
314 int
315 pthread_mutex_trylock(pthread_mutex_t *ptm)
316 {
317 	pthread_t self;
318 	void *val, *new, *next;
319 
320 	self = pthread__self();
321 	val = atomic_cas_ptr(&ptm->ptm_owner, NULL, self);
322 	if (__predict_true(val == NULL)) {
323 #ifndef PTHREAD__ATOMIC_IS_MEMBAR
324 		membar_enter();
325 #endif
326 		return 0;
327 	}
328 
329 	if (MUTEX_RECURSIVE(val)) {
330 		if (MUTEX_OWNER(val) == 0) {
331 			new = (void *)((uintptr_t)self | (uintptr_t)val);
332 			next = atomic_cas_ptr(&ptm->ptm_owner, val, new);
333 			if (__predict_true(next == val)) {
334 #ifndef PTHREAD__ATOMIC_IS_MEMBAR
335 				membar_enter();
336 #endif
337 				return 0;
338 			}
339 		}
340 		if (MUTEX_OWNER(val) == (uintptr_t)self) {
341 			if (ptm->ptm_recursed == INT_MAX)
342 				return EAGAIN;
343 			ptm->ptm_recursed++;
344 			return 0;
345 		}
346 	}
347 
348 	return EBUSY;
349 }
350 
351 int
352 pthread_mutex_unlock(pthread_mutex_t *ptm)
353 {
354 	pthread_t self;
355 	void *value;
356 
357 	/*
358 	 * Note this may be a non-interlocked CAS.  See lock_slow()
359 	 * above and sys/kern/kern_mutex.c for details.
360 	 */
361 #ifndef PTHREAD__ATOMIC_IS_MEMBAR
362 	membar_exit();
363 #endif
364 	self = pthread__self();
365 	value = atomic_cas_ptr_ni(&ptm->ptm_owner, self, NULL);
366 	if (__predict_true(value == self))
367 		return 0;
368 	return pthread__mutex_unlock_slow(ptm);
369 }
370 
371 NOINLINE static int
372 pthread__mutex_unlock_slow(pthread_mutex_t *ptm)
373 {
374 	pthread_t self, owner, new;
375 	int weown, error, deferred;
376 
377 	pthread__error(EINVAL, "Invalid mutex",
378 	    ptm->ptm_magic == _PT_MUTEX_MAGIC);
379 
380 	self = pthread__self();
381 	owner = ptm->ptm_owner;
382 	weown = (MUTEX_OWNER(owner) == (uintptr_t)self);
383 	deferred = (int)((uintptr_t)owner & MUTEX_DEFERRED_BIT);
384 	error = 0;
385 
386 	if (ptm->ptm_errorcheck) {
387 		if (!weown) {
388 			error = EPERM;
389 			new = owner;
390 		} else {
391 			new = NULL;
392 		}
393 	} else if (MUTEX_RECURSIVE(owner)) {
394 		if (!weown) {
395 			error = EPERM;
396 			new = owner;
397 		} else if (ptm->ptm_recursed) {
398 			ptm->ptm_recursed--;
399 			new = owner;
400 		} else {
401 			new = (pthread_t)MUTEX_RECURSIVE_BIT;
402 		}
403 	} else {
404 		pthread__error(EPERM,
405 		    "Unlocking unlocked mutex", (owner != NULL));
406 		pthread__error(EPERM,
407 		    "Unlocking mutex owned by another thread", weown);
408 		new = NULL;
409 	}
410 
411 	/*
412 	 * Release the mutex.  If there appear to be waiters, then
413 	 * wake them up.
414 	 */
415 	if (new != owner) {
416 		owner = atomic_swap_ptr(&ptm->ptm_owner, new);
417 		if (MUTEX_HAS_WAITERS(owner) != 0) {
418 			pthread__mutex_wakeup(self, ptm);
419 			return 0;
420 		}
421 	}
422 
423 	/*
424 	 * There were no waiters, but we may have deferred waking
425 	 * other threads until mutex unlock - we must wake them now.
426 	 */
427 	if (!deferred)
428 		return error;
429 
430 	if (self->pt_nwaiters == 1) {
431 		/*
432 		 * If the calling thread is about to block, defer
433 		 * unparking the target until _lwp_park() is called.
434 		 */
435 		if (self->pt_willpark && self->pt_unpark == 0) {
436 			self->pt_unpark = self->pt_waiters[0];
437 			self->pt_unparkhint =
438 			    __UNVOLATILE(&ptm->ptm_waiters);
439 		} else {
440 			(void)_lwp_unpark(self->pt_waiters[0],
441 			    __UNVOLATILE(&ptm->ptm_waiters));
442 		}
443 	} else {
444 		(void)_lwp_unpark_all(self->pt_waiters, self->pt_nwaiters,
445 		    __UNVOLATILE(&ptm->ptm_waiters));
446 	}
447 	self->pt_nwaiters = 0;
448 
449 	return error;
450 }
451 
452 static void
453 pthread__mutex_wakeup(pthread_t self, pthread_mutex_t *ptm)
454 {
455 	pthread_t thread, next;
456 	ssize_t n, rv;
457 
458 	/*
459 	 * Take ownership of the current set of waiters.  No
460 	 * need for a memory barrier following this, all loads
461 	 * are dependent upon 'thread'.
462 	 */
463 	thread = atomic_swap_ptr(&ptm->ptm_waiters, NULL);
464 
465 	for (;;) {
466 		/*
467 		 * Pull waiters from the queue and add to our list.
468 		 * Use a memory barrier to ensure that we safely
469 		 * read the value of pt_nextwaiter before 'thread'
470 		 * sees pt_sleeponq being cleared.
471 		 */
472 		for (n = self->pt_nwaiters, self->pt_nwaiters = 0;
473 		    n < pthread__unpark_max && thread != NULL;
474 		    thread = next) {
475 		    	next = thread->pt_nextwaiter;
476 		    	if (thread != self) {
477 				self->pt_waiters[n++] = thread->pt_lid;
478 				membar_sync();
479 			}
480 			thread->pt_sleeponq = 0;
481 			/* No longer safe to touch 'thread' */
482 		}
483 
484 		switch (n) {
485 		case 0:
486 			return;
487 		case 1:
488 			/*
489 			 * If the calling thread is about to block,
490 			 * defer unparking the target until _lwp_park()
491 			 * is called.
492 			 */
493 			if (self->pt_willpark && self->pt_unpark == 0) {
494 				self->pt_unpark = self->pt_waiters[0];
495 				self->pt_unparkhint =
496 				    __UNVOLATILE(&ptm->ptm_waiters);
497 				return;
498 			}
499 			rv = (ssize_t)_lwp_unpark(self->pt_waiters[0],
500 			    __UNVOLATILE(&ptm->ptm_waiters));
501 			if (rv != 0 && errno != EALREADY && errno != EINTR &&
502 			    errno != ESRCH) {
503 				pthread__errorfunc(__FILE__, __LINE__,
504 				    __func__, "_lwp_unpark failed");
505 			}
506 			return;
507 		default:
508 			rv = _lwp_unpark_all(self->pt_waiters, (size_t)n,
509 			    __UNVOLATILE(&ptm->ptm_waiters));
510 			if (rv != 0 && errno != EINTR) {
511 				pthread__errorfunc(__FILE__, __LINE__,
512 				    __func__, "_lwp_unpark_all failed");
513 			}
514 			break;
515 		}
516 	}
517 }
518 int
519 pthread_mutexattr_init(pthread_mutexattr_t *attr)
520 {
521 
522 	attr->ptma_magic = _PT_MUTEXATTR_MAGIC;
523 	attr->ptma_private = (void *)PTHREAD_MUTEX_DEFAULT;
524 	return 0;
525 }
526 
527 int
528 pthread_mutexattr_destroy(pthread_mutexattr_t *attr)
529 {
530 
531 	pthread__error(EINVAL, "Invalid mutex attribute",
532 	    attr->ptma_magic == _PT_MUTEXATTR_MAGIC);
533 
534 	return 0;
535 }
536 
537 
538 int
539 pthread_mutexattr_gettype(const pthread_mutexattr_t *attr, int *typep)
540 {
541 
542 	pthread__error(EINVAL, "Invalid mutex attribute",
543 	    attr->ptma_magic == _PT_MUTEXATTR_MAGIC);
544 
545 	*typep = (int)(intptr_t)attr->ptma_private;
546 	return 0;
547 }
548 
549 
550 int
551 pthread_mutexattr_settype(pthread_mutexattr_t *attr, int type)
552 {
553 
554 	pthread__error(EINVAL, "Invalid mutex attribute",
555 	    attr->ptma_magic == _PT_MUTEXATTR_MAGIC);
556 
557 	switch (type) {
558 	case PTHREAD_MUTEX_NORMAL:
559 	case PTHREAD_MUTEX_ERRORCHECK:
560 	case PTHREAD_MUTEX_RECURSIVE:
561 		attr->ptma_private = (void *)(intptr_t)type;
562 		return 0;
563 	default:
564 		return EINVAL;
565 	}
566 }
567 
568 
569 static void
570 once_cleanup(void *closure)
571 {
572 
573        pthread_mutex_unlock((pthread_mutex_t *)closure);
574 }
575 
576 
577 int
578 pthread_once(pthread_once_t *once_control, void (*routine)(void))
579 {
580 
581 	if (once_control->pto_done == 0) {
582 		pthread_mutex_lock(&once_control->pto_mutex);
583 		pthread_cleanup_push(&once_cleanup, &once_control->pto_mutex);
584 		if (once_control->pto_done == 0) {
585 			routine();
586 			once_control->pto_done = 1;
587 		}
588 		pthread_cleanup_pop(1);
589 	}
590 
591 	return 0;
592 }
593 
594 int
595 pthread__mutex_deferwake(pthread_t thread, pthread_mutex_t *ptm)
596 {
597 
598 	if (MUTEX_OWNER(ptm->ptm_owner) != (uintptr_t)thread)
599 		return 0;
600 	atomic_or_ulong((volatile unsigned long *)
601 	    (uintptr_t)&ptm->ptm_owner,
602 	    (unsigned long)MUTEX_DEFERRED_BIT);
603 	return 1;
604 }
605 
606 int
607 _pthread_mutex_held_np(pthread_mutex_t *ptm)
608 {
609 
610 	return MUTEX_OWNER(ptm->ptm_owner) == (uintptr_t)pthread__self();
611 }
612 
613 pthread_t
614 _pthread_mutex_owner_np(pthread_mutex_t *ptm)
615 {
616 
617 	return (pthread_t)MUTEX_OWNER(ptm->ptm_owner);
618 }
619