xref: /netbsd-src/lib/libpthread/pthread.c (revision 6a9b3088d8d2341ca1454531d365c15fe9c1c589)
1 /*	$NetBSD: pthread.c,v 1.117 2010/11/14 22:25:23 tron Exp $	*/
2 
3 /*-
4  * Copyright (c) 2001, 2002, 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 and 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.c,v 1.117 2010/11/14 22:25:23 tron Exp $");
34 
35 #define	__EXPOSE_STACK	1
36 
37 #include <sys/param.h>
38 #include <sys/mman.h>
39 #include <sys/sysctl.h>
40 #include <sys/lwpctl.h>
41 
42 #include <err.h>
43 #include <errno.h>
44 #include <lwp.h>
45 #include <signal.h>
46 #include <stdio.h>
47 #include <stdlib.h>
48 #include <string.h>
49 #include <syslog.h>
50 #include <ucontext.h>
51 #include <unistd.h>
52 #include <sched.h>
53 
54 #include "pthread.h"
55 #include "pthread_int.h"
56 
57 pthread_rwlock_t pthread__alltree_lock = PTHREAD_RWLOCK_INITIALIZER;
58 RB_HEAD(__pthread__alltree, __pthread_st) pthread__alltree;
59 
60 #ifndef lint
61 static int	pthread__cmp(struct __pthread_st *, struct __pthread_st *);
62 RB_PROTOTYPE_STATIC(__pthread__alltree, __pthread_st, pt_alltree, pthread__cmp)
63 #endif
64 
65 static void	pthread__create_tramp(void *);
66 static void	pthread__initthread(pthread_t);
67 static void	pthread__scrubthread(pthread_t, char *, int);
68 static int	pthread__stackid_setup(void *, size_t, pthread_t *);
69 static int	pthread__stackalloc(pthread_t *);
70 static void	pthread__initmain(pthread_t *);
71 static void	pthread__fork_callback(void);
72 static void	pthread__reap(pthread_t);
73 static void	pthread__child_callback(void);
74 static void	pthread__start(void);
75 
76 void	pthread__init(void);
77 
78 int pthread__started;
79 pthread_mutex_t pthread__deadqueue_lock = PTHREAD_MUTEX_INITIALIZER;
80 pthread_queue_t pthread__deadqueue;
81 pthread_queue_t pthread__allqueue;
82 
83 static pthread_attr_t pthread_default_attr;
84 static lwpctl_t pthread__dummy_lwpctl = { .lc_curcpu = LWPCTL_CPU_NONE };
85 static pthread_t pthread__first;
86 
87 enum {
88 	DIAGASSERT_ABORT =	1<<0,
89 	DIAGASSERT_STDERR =	1<<1,
90 	DIAGASSERT_SYSLOG =	1<<2
91 };
92 
93 static int pthread__diagassert;
94 
95 int pthread__concurrency;
96 int pthread__nspins;
97 int pthread__unpark_max = PTHREAD__UNPARK_MAX;
98 
99 /*
100  * We have to initialize the pthread_stack* variables here because
101  * mutexes are used before pthread_init() and thus pthread__initmain()
102  * are called.  Since mutexes only save the stack pointer and not a
103  * pointer to the thread data, it is safe to change the mapping from
104  * stack pointer to thread data afterwards.
105  */
106 #define	_STACKSIZE_LG 18
107 int	pthread__stacksize_lg = _STACKSIZE_LG;
108 size_t	pthread__stacksize = 1 << _STACKSIZE_LG;
109 vaddr_t	pthread__stackmask = (1 << _STACKSIZE_LG) - 1;
110 vaddr_t pthread__threadmask = (vaddr_t)~((1 << _STACKSIZE_LG) - 1);
111 #undef	_STACKSIZE_LG
112 
113 int _sys___sigprocmask14(int, const sigset_t *, sigset_t *);
114 
115 __strong_alias(__libc_thr_self,pthread_self)
116 __strong_alias(__libc_thr_create,pthread_create)
117 __strong_alias(__libc_thr_exit,pthread_exit)
118 __strong_alias(__libc_thr_errno,pthread__errno)
119 __strong_alias(__libc_thr_setcancelstate,pthread_setcancelstate)
120 __strong_alias(__libc_thr_equal,pthread_equal)
121 __strong_alias(__libc_thr_init,pthread__init)
122 
123 /*
124  * Static library kludge.  Place a reference to a symbol any library
125  * file which does not already have a reference here.
126  */
127 extern int pthread__cancel_stub_binder;
128 
129 void *pthread__static_lib_binder[] = {
130 	&pthread__cancel_stub_binder,
131 	pthread_cond_init,
132 	pthread_mutex_init,
133 	pthread_rwlock_init,
134 	pthread_barrier_init,
135 	pthread_key_create,
136 	pthread_setspecific,
137 };
138 
139 #define	NHASHLOCK	64
140 
141 static union hashlock {
142 	pthread_mutex_t	mutex;
143 	char		pad[64];
144 } hashlocks[NHASHLOCK] __aligned(64);
145 
146 /*
147  * This needs to be started by the library loading code, before main()
148  * gets to run, for various things that use the state of the initial thread
149  * to work properly (thread-specific data is an application-visible example;
150  * spinlock counts for mutexes is an internal example).
151  */
152 void
153 pthread__init(void)
154 {
155 	pthread_t first;
156 	char *p;
157 	int i, mib[2];
158 	size_t len;
159 	extern int __isthreaded;
160 
161 	mib[0] = CTL_HW;
162 	mib[1] = HW_NCPU;
163 
164 	len = sizeof(pthread__concurrency);
165 	if (sysctl(mib, 2, &pthread__concurrency, &len, NULL, 0) == -1)
166 		err(1, "sysctl(hw.ncpu");
167 
168 	mib[0] = CTL_KERN;
169 	mib[1] = KERN_OSREV;
170 
171 	/* Initialize locks first; they're needed elsewhere. */
172 	pthread__lockprim_init();
173 	for (i = 0; i < NHASHLOCK; i++) {
174 		pthread_mutex_init(&hashlocks[i].mutex, NULL);
175 	}
176 
177 	/* Fetch parameters. */
178 	i = (int)_lwp_unpark_all(NULL, 0, NULL);
179 	if (i == -1)
180 		err(1, "_lwp_unpark_all");
181 	if (i < pthread__unpark_max)
182 		pthread__unpark_max = i;
183 
184 	/* Basic data structure setup */
185 	pthread_attr_init(&pthread_default_attr);
186 	PTQ_INIT(&pthread__allqueue);
187 	PTQ_INIT(&pthread__deadqueue);
188 	RB_INIT(&pthread__alltree);
189 
190 	/* Create the thread structure corresponding to main() */
191 	pthread__initmain(&first);
192 	pthread__initthread(first);
193 	pthread__scrubthread(first, NULL, 0);
194 
195 	first->pt_lid = _lwp_self();
196 	PTQ_INSERT_HEAD(&pthread__allqueue, first, pt_allq);
197 	RB_INSERT(__pthread__alltree, &pthread__alltree, first);
198 
199 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &first->pt_lwpctl) != 0) {
200 		err(1, "_lwp_ctl");
201 	}
202 
203 	/* Start subsystems */
204 	PTHREAD_MD_INIT
205 
206 	for (p = pthread__getenv("PTHREAD_DIAGASSERT"); p && *p; p++) {
207 		switch (*p) {
208 		case 'a':
209 			pthread__diagassert |= DIAGASSERT_ABORT;
210 			break;
211 		case 'A':
212 			pthread__diagassert &= ~DIAGASSERT_ABORT;
213 			break;
214 		case 'e':
215 			pthread__diagassert |= DIAGASSERT_STDERR;
216 			break;
217 		case 'E':
218 			pthread__diagassert &= ~DIAGASSERT_STDERR;
219 			break;
220 		case 'l':
221 			pthread__diagassert |= DIAGASSERT_SYSLOG;
222 			break;
223 		case 'L':
224 			pthread__diagassert &= ~DIAGASSERT_SYSLOG;
225 			break;
226 		}
227 	}
228 
229 	/* Tell libc that we're here and it should role-play accordingly. */
230 	pthread__first = first;
231 	pthread_atfork(NULL, NULL, pthread__fork_callback);
232 	__isthreaded = 1;
233 }
234 
235 static void
236 pthread__fork_callback(void)
237 {
238 	struct __pthread_st *self;
239 
240 	/* lwpctl state is not copied across fork. */
241 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &pthread__first->pt_lwpctl)) {
242 		err(1, "_lwp_ctl");
243 	}
244 	self = pthread__self();
245 	self->pt_lid = _lwp_self();
246 }
247 
248 static void
249 pthread__child_callback(void)
250 {
251 
252 	/*
253 	 * Clean up data structures that a forked child process might
254 	 * trip over. Note that if threads have been created (causing
255 	 * this handler to be registered) the standards say that the
256 	 * child will trigger undefined behavior if it makes any
257 	 * pthread_* calls (or any other calls that aren't
258 	 * async-signal-safe), so we don't really have to clean up
259 	 * much. Anything that permits some pthread_* calls to work is
260 	 * merely being polite.
261 	 */
262 	pthread__started = 0;
263 }
264 
265 static void
266 pthread__start(void)
267 {
268 
269 	/*
270 	 * Per-process timers are cleared by fork(); despite the
271 	 * various restrictions on fork() and threads, it's legal to
272 	 * fork() before creating any threads.
273 	 */
274 	pthread_atfork(NULL, NULL, pthread__child_callback);
275 }
276 
277 
278 /* General-purpose thread data structure sanitization. */
279 /* ARGSUSED */
280 static void
281 pthread__initthread(pthread_t t)
282 {
283 
284 	t->pt_self = t;
285 	t->pt_magic = PT_MAGIC;
286 	t->pt_willpark = 0;
287 	t->pt_unpark = 0;
288 	t->pt_nwaiters = 0;
289 	t->pt_sleepobj = NULL;
290 	t->pt_signalled = 0;
291 	t->pt_havespecific = 0;
292 	t->pt_early = NULL;
293 	t->pt_lwpctl = &pthread__dummy_lwpctl;
294 	t->pt_blocking = 0;
295 	t->pt_droplock = NULL;
296 
297 	memcpy(&t->pt_lockops, pthread__lock_ops, sizeof(t->pt_lockops));
298 	pthread_mutex_init(&t->pt_lock, NULL);
299 	PTQ_INIT(&t->pt_cleanup_stack);
300 	pthread_cond_init(&t->pt_joiners, NULL);
301 	memset(&t->pt_specific, 0, sizeof(t->pt_specific));
302 }
303 
304 static void
305 pthread__scrubthread(pthread_t t, char *name, int flags)
306 {
307 
308 	t->pt_state = PT_STATE_RUNNING;
309 	t->pt_exitval = NULL;
310 	t->pt_flags = flags;
311 	t->pt_cancel = 0;
312 	t->pt_errno = 0;
313 	t->pt_name = name;
314 	t->pt_lid = 0;
315 }
316 
317 
318 int
319 pthread_create(pthread_t *thread, const pthread_attr_t *attr,
320 	    void *(*startfunc)(void *), void *arg)
321 {
322 	pthread_t newthread;
323 	pthread_attr_t nattr;
324 	struct pthread_attr_private *p;
325 	char * volatile name;
326 	unsigned long flag;
327 	int ret;
328 
329 	/*
330 	 * It's okay to check this without a lock because there can
331 	 * only be one thread before it becomes true.
332 	 */
333 	if (pthread__started == 0) {
334 		pthread__start();
335 		pthread__started = 1;
336 	}
337 
338 	if (attr == NULL)
339 		nattr = pthread_default_attr;
340 	else if (attr->pta_magic == PT_ATTR_MAGIC)
341 		nattr = *attr;
342 	else
343 		return EINVAL;
344 
345 	/* Fetch misc. attributes from the attr structure. */
346 	name = NULL;
347 	if ((p = nattr.pta_private) != NULL)
348 		if (p->ptap_name[0] != '\0')
349 			if ((name = strdup(p->ptap_name)) == NULL)
350 				return ENOMEM;
351 
352 	newthread = NULL;
353 
354 	/*
355 	 * Try to reclaim a dead thread.
356 	 */
357 	if (!PTQ_EMPTY(&pthread__deadqueue)) {
358 		pthread_mutex_lock(&pthread__deadqueue_lock);
359 		PTQ_FOREACH(newthread, &pthread__deadqueue, pt_deadq) {
360 			/* Still running? */
361 			if (newthread->pt_lwpctl->lc_curcpu ==
362 			    LWPCTL_CPU_EXITED ||
363 			    (_lwp_kill(newthread->pt_lid, 0) == -1 &&
364 			    errno == ESRCH))
365 				break;
366 		}
367 		if (newthread)
368 			PTQ_REMOVE(&pthread__deadqueue, newthread, pt_deadq);
369 		pthread_mutex_unlock(&pthread__deadqueue_lock);
370 	}
371 
372 	/*
373 	 * If necessary set up a stack, allocate space for a pthread_st,
374 	 * and initialize it.
375 	 */
376 	if (newthread == NULL) {
377 		ret = pthread__stackalloc(&newthread);
378 		if (ret != 0) {
379 			if (name)
380 				free(name);
381 			return ret;
382 		}
383 
384 		/* This is used only when creating the thread. */
385 		_INITCONTEXT_U(&newthread->pt_uc);
386 #ifdef PTHREAD_MACHINE_HAS_ID_REGISTER
387 		pthread__uc_id(&newthread->pt_uc) = newthread;
388 #endif
389 		newthread->pt_uc.uc_stack = newthread->pt_stack;
390 		newthread->pt_uc.uc_link = NULL;
391 
392 		/* Add to list of all threads. */
393 		pthread_rwlock_wrlock(&pthread__alltree_lock);
394 		PTQ_INSERT_TAIL(&pthread__allqueue, newthread, pt_allq);
395 		RB_INSERT(__pthread__alltree, &pthread__alltree, newthread);
396 		pthread_rwlock_unlock(&pthread__alltree_lock);
397 
398 		/* Will be reset by the thread upon exit. */
399 		pthread__initthread(newthread);
400 	}
401 
402 	/*
403 	 * Create the new LWP.
404 	 */
405 	pthread__scrubthread(newthread, name, nattr.pta_flags);
406 	newthread->pt_func = startfunc;
407 	newthread->pt_arg = arg;
408 	_lwp_makecontext(&newthread->pt_uc, pthread__create_tramp,
409 	    newthread, newthread, newthread->pt_stack.ss_sp,
410 	    newthread->pt_stack.ss_size);
411 
412 	flag = LWP_DETACHED;
413 	if ((newthread->pt_flags & PT_FLAG_SUSPENDED) != 0 ||
414 	    (nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0)
415 		flag |= LWP_SUSPENDED;
416 	ret = _lwp_create(&newthread->pt_uc, flag, &newthread->pt_lid);
417 	if (ret != 0) {
418 		pthread_mutex_lock(&newthread->pt_lock);
419 		/* Will unlock and free name. */
420 		pthread__reap(newthread);
421 		return ret;
422 	}
423 
424 	if ((nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0) {
425 		if (p != NULL) {
426 			(void)pthread_setschedparam(newthread, p->ptap_policy,
427 			    &p->ptap_sp);
428 		}
429 		if ((newthread->pt_flags & PT_FLAG_SUSPENDED) == 0) {
430 			(void)_lwp_continue(newthread->pt_lid);
431 		}
432 	}
433 
434 	*thread = newthread;
435 
436 	return 0;
437 }
438 
439 
440 static void
441 pthread__create_tramp(void *cookie)
442 {
443 	pthread_t self;
444 	void *retval;
445 
446 	self = cookie;
447 
448 	/*
449 	 * Throw away some stack in a feeble attempt to reduce cache
450 	 * thrash.  May help for SMT processors.  XXX We should not
451 	 * be allocating stacks on fixed 2MB boundaries.  Needs a
452 	 * thread register or decent thread local storage.
453 	 *
454 	 * Note that we may race with the kernel in _lwp_create(),
455 	 * and so pt_lid can be unset at this point, but we don't
456 	 * care.
457 	 */
458 	(void)alloca(((unsigned)self->pt_lid & 7) << 8);
459 
460 	if (self->pt_name != NULL) {
461 		pthread_mutex_lock(&self->pt_lock);
462 		if (self->pt_name != NULL)
463 			(void)_lwp_setname(0, self->pt_name);
464 		pthread_mutex_unlock(&self->pt_lock);
465 	}
466 
467 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &self->pt_lwpctl)) {
468 		err(1, "_lwp_ctl");
469 	}
470 
471 	retval = (*self->pt_func)(self->pt_arg);
472 
473 	pthread_exit(retval);
474 
475 	/*NOTREACHED*/
476 	pthread__abort();
477 }
478 
479 int
480 pthread_suspend_np(pthread_t thread)
481 {
482 	pthread_t self;
483 
484 	self = pthread__self();
485 	if (self == thread) {
486 		return EDEADLK;
487 	}
488 	if (pthread__find(thread) != 0)
489 		return ESRCH;
490 	if (_lwp_suspend(thread->pt_lid) == 0)
491 		return 0;
492 	return errno;
493 }
494 
495 int
496 pthread_resume_np(pthread_t thread)
497 {
498 
499 	if (pthread__find(thread) != 0)
500 		return ESRCH;
501 	if (_lwp_continue(thread->pt_lid) == 0)
502 		return 0;
503 	return errno;
504 }
505 
506 void
507 pthread_exit(void *retval)
508 {
509 	pthread_t self;
510 	struct pt_clean_t *cleanup;
511 	char *name;
512 
513 	self = pthread__self();
514 
515 	/* Disable cancellability. */
516 	pthread_mutex_lock(&self->pt_lock);
517 	self->pt_flags |= PT_FLAG_CS_DISABLED;
518 	self->pt_cancel = 0;
519 
520 	/* Call any cancellation cleanup handlers */
521 	if (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
522 		pthread_mutex_unlock(&self->pt_lock);
523 		while (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
524 			cleanup = PTQ_FIRST(&self->pt_cleanup_stack);
525 			PTQ_REMOVE(&self->pt_cleanup_stack, cleanup, ptc_next);
526 			(*cleanup->ptc_cleanup)(cleanup->ptc_arg);
527 		}
528 		pthread_mutex_lock(&self->pt_lock);
529 	}
530 
531 	/* Perform cleanup of thread-specific data */
532 	pthread__destroy_tsd(self);
533 
534 	/* Signal our exit. */
535 	self->pt_exitval = retval;
536 	if (self->pt_flags & PT_FLAG_DETACHED) {
537 		self->pt_state = PT_STATE_DEAD;
538 		name = self->pt_name;
539 		self->pt_name = NULL;
540 		pthread_mutex_unlock(&self->pt_lock);
541 		if (name != NULL)
542 			free(name);
543 		pthread_mutex_lock(&pthread__deadqueue_lock);
544 		PTQ_INSERT_TAIL(&pthread__deadqueue, self, pt_deadq);
545 		pthread_mutex_unlock(&pthread__deadqueue_lock);
546 		_lwp_exit();
547 	} else {
548 		self->pt_state = PT_STATE_ZOMBIE;
549 		pthread_cond_broadcast(&self->pt_joiners);
550 		pthread_mutex_unlock(&self->pt_lock);
551 		/* Note: name will be freed by the joiner. */
552 		_lwp_exit();
553 	}
554 
555 	/*NOTREACHED*/
556 	pthread__abort();
557 	exit(1);
558 }
559 
560 
561 int
562 pthread_join(pthread_t thread, void **valptr)
563 {
564 	pthread_t self;
565 	int error;
566 
567 	self = pthread__self();
568 
569 	if (pthread__find(thread) != 0)
570 		return ESRCH;
571 
572 	if (thread->pt_magic != PT_MAGIC)
573 		return EINVAL;
574 
575 	if (thread == self)
576 		return EDEADLK;
577 
578 	self->pt_droplock = &thread->pt_lock;
579 	pthread_mutex_lock(&thread->pt_lock);
580 	for (;;) {
581 		if (thread->pt_state == PT_STATE_ZOMBIE)
582 			break;
583 		if (thread->pt_state == PT_STATE_DEAD) {
584 			pthread_mutex_unlock(&thread->pt_lock);
585 			self->pt_droplock = NULL;
586 			return ESRCH;
587 		}
588 		if ((thread->pt_flags & PT_FLAG_DETACHED) != 0) {
589 			pthread_mutex_unlock(&thread->pt_lock);
590 			self->pt_droplock = NULL;
591 			return EINVAL;
592 		}
593 		error = pthread_cond_wait(&thread->pt_joiners,
594 		    &thread->pt_lock);
595 		if (error != 0) {
596 			pthread__errorfunc(__FILE__, __LINE__,
597 			    __func__, "unexpected return from cond_wait()");
598 		}
599 
600 	}
601 	pthread__testcancel(self);
602 	if (valptr != NULL)
603 		*valptr = thread->pt_exitval;
604 	/* pthread__reap() will drop the lock. */
605 	pthread__reap(thread);
606 	self->pt_droplock = NULL;
607 
608 	return 0;
609 }
610 
611 static void
612 pthread__reap(pthread_t thread)
613 {
614 	char *name;
615 
616 	name = thread->pt_name;
617 	thread->pt_name = NULL;
618 	thread->pt_state = PT_STATE_DEAD;
619 	pthread_mutex_unlock(&thread->pt_lock);
620 
621 	pthread_mutex_lock(&pthread__deadqueue_lock);
622 	PTQ_INSERT_HEAD(&pthread__deadqueue, thread, pt_deadq);
623 	pthread_mutex_unlock(&pthread__deadqueue_lock);
624 
625 	if (name != NULL)
626 		free(name);
627 }
628 
629 int
630 pthread_equal(pthread_t t1, pthread_t t2)
631 {
632 
633 	/* Nothing special here. */
634 	return (t1 == t2);
635 }
636 
637 
638 int
639 pthread_detach(pthread_t thread)
640 {
641 
642 	if (pthread__find(thread) != 0)
643 		return ESRCH;
644 
645 	if (thread->pt_magic != PT_MAGIC)
646 		return EINVAL;
647 
648 	pthread_mutex_lock(&thread->pt_lock);
649 	thread->pt_flags |= PT_FLAG_DETACHED;
650 	if (thread->pt_state == PT_STATE_ZOMBIE) {
651 		/* pthread__reap() will drop the lock. */
652 		pthread__reap(thread);
653 	} else {
654 		/*
655 		 * Not valid for threads to be waiting in
656 		 * pthread_join() (there are intractable
657 		 * sync issues from the application
658 		 * perspective), but give those threads
659 		 * a chance anyway.
660 		 */
661 		pthread_cond_broadcast(&thread->pt_joiners);
662 		pthread_mutex_unlock(&thread->pt_lock);
663 	}
664 
665 	return 0;
666 }
667 
668 
669 int
670 pthread_getname_np(pthread_t thread, char *name, size_t len)
671 {
672 
673 	if (pthread__find(thread) != 0)
674 		return ESRCH;
675 
676 	if (thread->pt_magic != PT_MAGIC)
677 		return EINVAL;
678 
679 	pthread_mutex_lock(&thread->pt_lock);
680 	if (thread->pt_name == NULL)
681 		name[0] = '\0';
682 	else
683 		strlcpy(name, thread->pt_name, len);
684 	pthread_mutex_unlock(&thread->pt_lock);
685 
686 	return 0;
687 }
688 
689 
690 int
691 pthread_setname_np(pthread_t thread, const char *name, void *arg)
692 {
693 	char *oldname, *cp, newname[PTHREAD_MAX_NAMELEN_NP];
694 	int namelen;
695 
696 	if (pthread__find(thread) != 0)
697 		return ESRCH;
698 
699 	if (thread->pt_magic != PT_MAGIC)
700 		return EINVAL;
701 
702 	namelen = snprintf(newname, sizeof(newname), name, arg);
703 	if (namelen >= PTHREAD_MAX_NAMELEN_NP)
704 		return EINVAL;
705 
706 	cp = strdup(newname);
707 	if (cp == NULL)
708 		return ENOMEM;
709 
710 	pthread_mutex_lock(&thread->pt_lock);
711 	oldname = thread->pt_name;
712 	thread->pt_name = cp;
713 	(void)_lwp_setname(thread->pt_lid, cp);
714 	pthread_mutex_unlock(&thread->pt_lock);
715 
716 	if (oldname != NULL)
717 		free(oldname);
718 
719 	return 0;
720 }
721 
722 
723 
724 /*
725  * XXX There should be a way for applications to use the efficent
726  *  inline version, but there are opacity/namespace issues.
727  */
728 pthread_t
729 pthread_self(void)
730 {
731 
732 	return pthread__self();
733 }
734 
735 
736 int
737 pthread_cancel(pthread_t thread)
738 {
739 
740 	if (pthread__find(thread) != 0)
741 		return ESRCH;
742 	pthread_mutex_lock(&thread->pt_lock);
743 	thread->pt_flags |= PT_FLAG_CS_PENDING;
744 	if ((thread->pt_flags & PT_FLAG_CS_DISABLED) == 0) {
745 		thread->pt_cancel = 1;
746 		pthread_mutex_unlock(&thread->pt_lock);
747 		_lwp_wakeup(thread->pt_lid);
748 	} else
749 		pthread_mutex_unlock(&thread->pt_lock);
750 
751 	return 0;
752 }
753 
754 
755 int
756 pthread_setcancelstate(int state, int *oldstate)
757 {
758 	pthread_t self;
759 	int retval;
760 
761 	self = pthread__self();
762 	retval = 0;
763 
764 	pthread_mutex_lock(&self->pt_lock);
765 
766 	if (oldstate != NULL) {
767 		if (self->pt_flags & PT_FLAG_CS_DISABLED)
768 			*oldstate = PTHREAD_CANCEL_DISABLE;
769 		else
770 			*oldstate = PTHREAD_CANCEL_ENABLE;
771 	}
772 
773 	if (state == PTHREAD_CANCEL_DISABLE) {
774 		self->pt_flags |= PT_FLAG_CS_DISABLED;
775 		if (self->pt_cancel) {
776 			self->pt_flags |= PT_FLAG_CS_PENDING;
777 			self->pt_cancel = 0;
778 		}
779 	} else if (state == PTHREAD_CANCEL_ENABLE) {
780 		self->pt_flags &= ~PT_FLAG_CS_DISABLED;
781 		/*
782 		 * If a cancellation was requested while cancellation
783 		 * was disabled, note that fact for future
784 		 * cancellation tests.
785 		 */
786 		if (self->pt_flags & PT_FLAG_CS_PENDING) {
787 			self->pt_cancel = 1;
788 			/* This is not a deferred cancellation point. */
789 			if (self->pt_flags & PT_FLAG_CS_ASYNC) {
790 				pthread_mutex_unlock(&self->pt_lock);
791 				pthread__cancelled();
792 			}
793 		}
794 	} else
795 		retval = EINVAL;
796 
797 	pthread_mutex_unlock(&self->pt_lock);
798 
799 	return retval;
800 }
801 
802 
803 int
804 pthread_setcanceltype(int type, int *oldtype)
805 {
806 	pthread_t self;
807 	int retval;
808 
809 	self = pthread__self();
810 	retval = 0;
811 
812 	pthread_mutex_lock(&self->pt_lock);
813 
814 	if (oldtype != NULL) {
815 		if (self->pt_flags & PT_FLAG_CS_ASYNC)
816 			*oldtype = PTHREAD_CANCEL_ASYNCHRONOUS;
817 		else
818 			*oldtype = PTHREAD_CANCEL_DEFERRED;
819 	}
820 
821 	if (type == PTHREAD_CANCEL_ASYNCHRONOUS) {
822 		self->pt_flags |= PT_FLAG_CS_ASYNC;
823 		if (self->pt_cancel) {
824 			pthread_mutex_unlock(&self->pt_lock);
825 			pthread__cancelled();
826 		}
827 	} else if (type == PTHREAD_CANCEL_DEFERRED)
828 		self->pt_flags &= ~PT_FLAG_CS_ASYNC;
829 	else
830 		retval = EINVAL;
831 
832 	pthread_mutex_unlock(&self->pt_lock);
833 
834 	return retval;
835 }
836 
837 
838 void
839 pthread_testcancel(void)
840 {
841 	pthread_t self;
842 
843 	self = pthread__self();
844 	if (self->pt_cancel)
845 		pthread__cancelled();
846 }
847 
848 
849 /*
850  * POSIX requires that certain functions return an error rather than
851  * invoking undefined behavior even when handed completely bogus
852  * pthread_t values, e.g. stack garbage or (pthread_t)666. This
853  * utility routine searches the list of threads for the pthread_t
854  * value without dereferencing it.
855  */
856 int
857 pthread__find(pthread_t id)
858 {
859 	pthread_t target;
860 
861 	pthread_rwlock_rdlock(&pthread__alltree_lock);
862 	/* LINTED */
863 	target = RB_FIND(__pthread__alltree, &pthread__alltree, id);
864 	pthread_rwlock_unlock(&pthread__alltree_lock);
865 
866 	if (target == NULL || target->pt_state == PT_STATE_DEAD)
867 		return ESRCH;
868 
869 	return 0;
870 }
871 
872 
873 void
874 pthread__testcancel(pthread_t self)
875 {
876 
877 	if (self->pt_cancel)
878 		pthread__cancelled();
879 }
880 
881 
882 void
883 pthread__cancelled(void)
884 {
885 	pthread_mutex_t *droplock;
886 	pthread_t self;
887 
888 	self = pthread__self();
889 	droplock = self->pt_droplock;
890 	self->pt_droplock = NULL;
891 
892 	if (droplock != NULL && pthread_mutex_held_np(droplock))
893 		pthread_mutex_unlock(droplock);
894 
895 	pthread_exit(PTHREAD_CANCELED);
896 }
897 
898 
899 void
900 pthread__cleanup_push(void (*cleanup)(void *), void *arg, void *store)
901 {
902 	pthread_t self;
903 	struct pt_clean_t *entry;
904 
905 	self = pthread__self();
906 	entry = store;
907 	entry->ptc_cleanup = cleanup;
908 	entry->ptc_arg = arg;
909 	PTQ_INSERT_HEAD(&self->pt_cleanup_stack, entry, ptc_next);
910 }
911 
912 
913 void
914 pthread__cleanup_pop(int ex, void *store)
915 {
916 	pthread_t self;
917 	struct pt_clean_t *entry;
918 
919 	self = pthread__self();
920 	entry = store;
921 
922 	PTQ_REMOVE(&self->pt_cleanup_stack, entry, ptc_next);
923 	if (ex)
924 		(*entry->ptc_cleanup)(entry->ptc_arg);
925 }
926 
927 
928 int *
929 pthread__errno(void)
930 {
931 	pthread_t self;
932 
933 	self = pthread__self();
934 
935 	return &(self->pt_errno);
936 }
937 
938 ssize_t	_sys_write(int, const void *, size_t);
939 
940 void
941 pthread__assertfunc(const char *file, int line, const char *function,
942 		    const char *expr)
943 {
944 	char buf[1024];
945 	int len;
946 
947 	/*
948 	 * snprintf should not acquire any locks, or we could
949 	 * end up deadlocked if the assert caller held locks.
950 	 */
951 	len = snprintf(buf, 1024,
952 	    "assertion \"%s\" failed: file \"%s\", line %d%s%s%s\n",
953 	    expr, file, line,
954 	    function ? ", function \"" : "",
955 	    function ? function : "",
956 	    function ? "\"" : "");
957 
958 	_sys_write(STDERR_FILENO, buf, (size_t)len);
959 	(void)kill(getpid(), SIGABRT);
960 
961 	_exit(1);
962 }
963 
964 
965 void
966 pthread__errorfunc(const char *file, int line, const char *function,
967 		   const char *msg)
968 {
969 	char buf[1024];
970 	size_t len;
971 
972 	if (pthread__diagassert == 0)
973 		return;
974 
975 	/*
976 	 * snprintf should not acquire any locks, or we could
977 	 * end up deadlocked if the assert caller held locks.
978 	 */
979 	len = snprintf(buf, 1024,
980 	    "%s: Error detected by libpthread: %s.\n"
981 	    "Detected by file \"%s\", line %d%s%s%s.\n"
982 	    "See pthread(3) for information.\n",
983 	    getprogname(), msg, file, line,
984 	    function ? ", function \"" : "",
985 	    function ? function : "",
986 	    function ? "\"" : "");
987 
988 	if (pthread__diagassert & DIAGASSERT_STDERR)
989 		_sys_write(STDERR_FILENO, buf, len);
990 
991 	if (pthread__diagassert & DIAGASSERT_SYSLOG)
992 		syslog(LOG_DEBUG | LOG_USER, "%s", buf);
993 
994 	if (pthread__diagassert & DIAGASSERT_ABORT) {
995 		(void)kill(getpid(), SIGABRT);
996 		_exit(1);
997 	}
998 }
999 
1000 /*
1001  * Thread park/unpark operations.  The kernel operations are
1002  * modelled after a brief description from "Multithreading in
1003  * the Solaris Operating Environment":
1004  *
1005  * http://www.sun.com/software/whitepapers/solaris9/multithread.pdf
1006  */
1007 
1008 #define	OOPS(msg)			\
1009     pthread__errorfunc(__FILE__, __LINE__, __func__, msg)
1010 
1011 int
1012 pthread__park(pthread_t self, pthread_mutex_t *lock,
1013 	      pthread_queue_t *queue, const struct timespec *abstime,
1014 	      int cancelpt, const void *hint)
1015 {
1016 	int rv, error;
1017 	void *obj;
1018 
1019 	/*
1020 	 * For non-interlocked release of mutexes we need a store
1021 	 * barrier before incrementing pt_blocking away from zero.
1022 	 * This is provided by pthread_mutex_unlock().
1023 	 */
1024 	self->pt_willpark = 1;
1025 	pthread_mutex_unlock(lock);
1026 	self->pt_willpark = 0;
1027 	self->pt_blocking++;
1028 
1029 	/*
1030 	 * Wait until we are awoken by a pending unpark operation,
1031 	 * a signal, an unpark posted after we have gone asleep,
1032 	 * or an expired timeout.
1033 	 *
1034 	 * It is fine to test the value of pt_sleepobj without
1035 	 * holding any locks, because:
1036 	 *
1037 	 * o Only the blocking thread (this thread) ever sets them
1038 	 *   to a non-NULL value.
1039 	 *
1040 	 * o Other threads may set them NULL, but if they do so they
1041 	 *   must also make this thread return from _lwp_park.
1042 	 *
1043 	 * o _lwp_park, _lwp_unpark and _lwp_unpark_all are system
1044 	 *   calls and all make use of spinlocks in the kernel.  So
1045 	 *   these system calls act as full memory barriers, and will
1046 	 *   ensure that the calling CPU's store buffers are drained.
1047 	 *   In combination with the spinlock release before unpark,
1048 	 *   this means that modification of pt_sleepobj/onq by another
1049 	 *   thread will become globally visible before that thread
1050 	 *   schedules an unpark operation on this thread.
1051 	 *
1052 	 * Note: the test in the while() statement dodges the park op if
1053 	 * we have already been awoken, unless there is another thread to
1054 	 * awaken.  This saves a syscall - if we were already awakened,
1055 	 * the next call to _lwp_park() would need to return early in order
1056 	 * to eat the previous wakeup.
1057 	 */
1058 	rv = 0;
1059 	do {
1060 		/*
1061 		 * If we deferred unparking a thread, arrange to
1062 		 * have _lwp_park() restart it before blocking.
1063 		 */
1064 		error = _lwp_park(abstime, self->pt_unpark, hint, hint);
1065 		self->pt_unpark = 0;
1066 		if (error != 0) {
1067 			switch (rv = errno) {
1068 			case EINTR:
1069 			case EALREADY:
1070 				rv = 0;
1071 				break;
1072 			case ETIMEDOUT:
1073 				break;
1074 			default:
1075 				OOPS("_lwp_park failed");
1076 				break;
1077 			}
1078 		}
1079 		/* Check for cancellation. */
1080 		if (cancelpt && self->pt_cancel)
1081 			rv = EINTR;
1082 	} while (self->pt_sleepobj != NULL && rv == 0);
1083 
1084 	/*
1085 	 * If we have been awoken early but are still on the queue,
1086 	 * then remove ourself.  Again, it's safe to do the test
1087 	 * without holding any locks.
1088 	 */
1089 	if (__predict_false(self->pt_sleepobj != NULL)) {
1090 		pthread_mutex_lock(lock);
1091 		if ((obj = self->pt_sleepobj) != NULL) {
1092 			PTQ_REMOVE(queue, self, pt_sleep);
1093 			self->pt_sleepobj = NULL;
1094 			if (obj != NULL && self->pt_early != NULL)
1095 				(*self->pt_early)(obj);
1096 		}
1097 		pthread_mutex_unlock(lock);
1098 	}
1099 	self->pt_early = NULL;
1100 	self->pt_blocking--;
1101 	membar_sync();
1102 
1103 	return rv;
1104 }
1105 
1106 void
1107 pthread__unpark(pthread_queue_t *queue, pthread_t self,
1108 		pthread_mutex_t *interlock)
1109 {
1110 	pthread_t target;
1111 	u_int max;
1112 	size_t nwaiters;
1113 
1114 	max = pthread__unpark_max;
1115 	nwaiters = self->pt_nwaiters;
1116 	target = PTQ_FIRST(queue);
1117 	if (nwaiters == max) {
1118 		/* Overflow. */
1119 		(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1120 		    __UNVOLATILE(&interlock->ptm_waiters));
1121 		nwaiters = 0;
1122 	}
1123 	target->pt_sleepobj = NULL;
1124 	self->pt_waiters[nwaiters++] = target->pt_lid;
1125 	PTQ_REMOVE(queue, target, pt_sleep);
1126 	self->pt_nwaiters = nwaiters;
1127 	pthread__mutex_deferwake(self, interlock);
1128 }
1129 
1130 void
1131 pthread__unpark_all(pthread_queue_t *queue, pthread_t self,
1132 		    pthread_mutex_t *interlock)
1133 {
1134 	pthread_t target;
1135 	u_int max;
1136 	size_t nwaiters;
1137 
1138 	max = pthread__unpark_max;
1139 	nwaiters = self->pt_nwaiters;
1140 	PTQ_FOREACH(target, queue, pt_sleep) {
1141 		if (nwaiters == max) {
1142 			/* Overflow. */
1143 			(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1144 			    __UNVOLATILE(&interlock->ptm_waiters));
1145 			nwaiters = 0;
1146 		}
1147 		target->pt_sleepobj = NULL;
1148 		self->pt_waiters[nwaiters++] = target->pt_lid;
1149 	}
1150 	self->pt_nwaiters = nwaiters;
1151 	PTQ_INIT(queue);
1152 	pthread__mutex_deferwake(self, interlock);
1153 }
1154 
1155 #undef	OOPS
1156 
1157 /*
1158  * Allocate a stack for a thread, and set it up. It needs to be aligned, so
1159  * that a thread can find itself by its stack pointer.
1160  */
1161 static int
1162 pthread__stackalloc(pthread_t *newt)
1163 {
1164 	void *addr;
1165 
1166 	addr = mmap(NULL, pthread__stacksize, PROT_READ|PROT_WRITE,
1167 	    MAP_ANON|MAP_PRIVATE | MAP_ALIGNED(pthread__stacksize_lg),
1168 	    -1, (off_t)0);
1169 
1170 	if (addr == MAP_FAILED)
1171 		return ENOMEM;
1172 
1173 	pthread__assert(((intptr_t)addr & pthread__stackmask) == 0);
1174 
1175 	return pthread__stackid_setup(addr, pthread__stacksize, newt);
1176 }
1177 
1178 
1179 /*
1180  * Set up the slightly special stack for the "initial" thread, which
1181  * runs on the normal system stack, and thus gets slightly different
1182  * treatment.
1183  */
1184 static void
1185 pthread__initmain(pthread_t *newt)
1186 {
1187 	struct rlimit slimit;
1188 	size_t pagesize;
1189 	pthread_t t;
1190 	void *base;
1191 	size_t size;
1192 	int error, ret;
1193 	char *value;
1194 
1195 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1196 	pthread__stacksize = 0;
1197 	ret = getrlimit(RLIMIT_STACK, &slimit);
1198 	if (ret == -1)
1199 		err(1, "Couldn't get stack resource consumption limits");
1200 
1201 	value = pthread__getenv("PTHREAD_STACKSIZE");
1202 	if (value != NULL) {
1203 		pthread__stacksize = atoi(value) * 1024;
1204 		if (pthread__stacksize > slimit.rlim_cur)
1205 			pthread__stacksize = (size_t)slimit.rlim_cur;
1206 	}
1207 	if (pthread__stacksize == 0)
1208 		pthread__stacksize = (size_t)slimit.rlim_cur;
1209 	if (pthread__stacksize < 4 * pagesize)
1210 		errx(1, "Stacksize limit is too low, minimum %zd kbyte.",
1211 		    4 * pagesize / 1024);
1212 
1213 	pthread__stacksize_lg = -1;
1214 	while (pthread__stacksize) {
1215 		pthread__stacksize >>= 1;
1216 		pthread__stacksize_lg++;
1217 	}
1218 
1219 	pthread__stacksize = (1 << pthread__stacksize_lg);
1220 	pthread__stackmask = pthread__stacksize - 1;
1221 	pthread__threadmask = ~pthread__stackmask;
1222 
1223 	base = (void *)(pthread__sp() & pthread__threadmask);
1224 	size = pthread__stacksize;
1225 
1226 	error = pthread__stackid_setup(base, size, &t);
1227 	if (error) {
1228 		/* XXX */
1229 		errx(2, "failed to setup main thread: error=%d", error);
1230 	}
1231 
1232 	*newt = t;
1233 
1234 	/* Set up identity register. */
1235 	(void)_lwp_setprivate(t);
1236 }
1237 
1238 static int
1239 /*ARGSUSED*/
1240 pthread__stackid_setup(void *base, size_t size, pthread_t *tp)
1241 {
1242 	pthread_t t;
1243 	void *redaddr;
1244 	size_t pagesize;
1245 	int ret;
1246 
1247 	t = base;
1248 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1249 
1250 	/*
1251 	 * Put a pointer to the pthread in the bottom (but
1252          * redzone-protected section) of the stack.
1253 	 */
1254 	redaddr = STACK_SHRINK(STACK_MAX(base, size), pagesize);
1255 	t->pt_stack.ss_size = size - 2 * pagesize;
1256 #ifdef __MACHINE_STACK_GROWS_UP
1257 	t->pt_stack.ss_sp = (char *)(void *)base + pagesize;
1258 #else
1259 	t->pt_stack.ss_sp = (char *)(void *)base + 2 * pagesize;
1260 #endif
1261 
1262 	/* Protect the next-to-bottom stack page as a red zone. */
1263 	ret = mprotect(redaddr, pagesize, PROT_NONE);
1264 	if (ret == -1) {
1265 		return errno;
1266 	}
1267 	*tp = t;
1268 	return 0;
1269 }
1270 
1271 #ifndef lint
1272 static int
1273 pthread__cmp(struct __pthread_st *a, struct __pthread_st *b)
1274 {
1275 
1276 	if ((uintptr_t)a < (uintptr_t)b)
1277 		return (-1);
1278 	else if (a == b)
1279 		return 0;
1280 	else
1281 		return 1;
1282 }
1283 RB_GENERATE_STATIC(__pthread__alltree, __pthread_st, pt_alltree, pthread__cmp)
1284 #endif
1285 
1286 /* Because getenv() wants to use locks. */
1287 char *
1288 pthread__getenv(const char *name)
1289 {
1290 	extern char **environ;
1291 	size_t l_name, offset;
1292 
1293 	l_name = strlen(name);
1294 	for (offset = 0; environ[offset] != NULL; offset++) {
1295 		if (strncmp(name, environ[offset], l_name) == 0 &&
1296 		    environ[offset][l_name] == '=') {
1297 			return environ[offset] + l_name + 1;
1298 		}
1299 	}
1300 
1301 	return NULL;
1302 }
1303 
1304 pthread_mutex_t *
1305 pthread__hashlock(volatile const void *p)
1306 {
1307 	uintptr_t v;
1308 
1309 	v = (uintptr_t)p;
1310 	return &hashlocks[((v >> 9) ^ (v >> 3)) & (NHASHLOCK - 1)].mutex;
1311 }
1312 
1313 int
1314 pthread__checkpri(int pri)
1315 {
1316 	static int havepri;
1317 	static long min, max;
1318 
1319 	if (!havepri) {
1320 		min = sysconf(_SC_SCHED_PRI_MIN);
1321 		max = sysconf(_SC_SCHED_PRI_MAX);
1322 		havepri = 1;
1323 	}
1324 	return (pri < min || pri > max) ? EINVAL : 0;
1325 }
1326