xref: /netbsd-src/lib/libpthread/pthread.c (revision 62a8debe1dc62962e18a1c918def78666141273b)
1 /*	$NetBSD: pthread.c,v 1.113 2009/10/03 23:49:50 christos 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 1.113 2009/10/03 23:49:50 christos 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 
239 	/* lwpctl state is not copied across fork. */
240 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &pthread__first->pt_lwpctl)) {
241 		err(1, "_lwp_ctl");
242 	}
243 }
244 
245 static void
246 pthread__child_callback(void)
247 {
248 
249 	/*
250 	 * Clean up data structures that a forked child process might
251 	 * trip over. Note that if threads have been created (causing
252 	 * this handler to be registered) the standards say that the
253 	 * child will trigger undefined behavior if it makes any
254 	 * pthread_* calls (or any other calls that aren't
255 	 * async-signal-safe), so we don't really have to clean up
256 	 * much. Anything that permits some pthread_* calls to work is
257 	 * merely being polite.
258 	 */
259 	pthread__started = 0;
260 }
261 
262 static void
263 pthread__start(void)
264 {
265 
266 	/*
267 	 * Per-process timers are cleared by fork(); despite the
268 	 * various restrictions on fork() and threads, it's legal to
269 	 * fork() before creating any threads.
270 	 */
271 	pthread_atfork(NULL, NULL, pthread__child_callback);
272 }
273 
274 
275 /* General-purpose thread data structure sanitization. */
276 /* ARGSUSED */
277 static void
278 pthread__initthread(pthread_t t)
279 {
280 
281 	t->pt_self = t;
282 	t->pt_magic = PT_MAGIC;
283 	t->pt_willpark = 0;
284 	t->pt_unpark = 0;
285 	t->pt_nwaiters = 0;
286 	t->pt_sleepobj = NULL;
287 	t->pt_signalled = 0;
288 	t->pt_havespecific = 0;
289 	t->pt_early = NULL;
290 	t->pt_lwpctl = &pthread__dummy_lwpctl;
291 	t->pt_blocking = 0;
292 	t->pt_droplock = NULL;
293 
294 	memcpy(&t->pt_lockops, pthread__lock_ops, sizeof(t->pt_lockops));
295 	pthread_mutex_init(&t->pt_lock, NULL);
296 	PTQ_INIT(&t->pt_cleanup_stack);
297 	pthread_cond_init(&t->pt_joiners, NULL);
298 	memset(&t->pt_specific, 0, sizeof(t->pt_specific));
299 }
300 
301 static void
302 pthread__scrubthread(pthread_t t, char *name, int flags)
303 {
304 
305 	t->pt_state = PT_STATE_RUNNING;
306 	t->pt_exitval = NULL;
307 	t->pt_flags = flags;
308 	t->pt_cancel = 0;
309 	t->pt_errno = 0;
310 	t->pt_name = name;
311 	t->pt_lid = 0;
312 }
313 
314 
315 int
316 pthread_create(pthread_t *thread, const pthread_attr_t *attr,
317 	    void *(*startfunc)(void *), void *arg)
318 {
319 	pthread_t newthread;
320 	pthread_attr_t nattr;
321 	struct pthread_attr_private *p;
322 	char * volatile name;
323 	unsigned long flag;
324 	int ret;
325 
326 	/*
327 	 * It's okay to check this without a lock because there can
328 	 * only be one thread before it becomes true.
329 	 */
330 	if (pthread__started == 0) {
331 		pthread__start();
332 		pthread__started = 1;
333 	}
334 
335 	if (attr == NULL)
336 		nattr = pthread_default_attr;
337 	else if (attr->pta_magic == PT_ATTR_MAGIC)
338 		nattr = *attr;
339 	else
340 		return EINVAL;
341 
342 	/* Fetch misc. attributes from the attr structure. */
343 	name = NULL;
344 	if ((p = nattr.pta_private) != NULL)
345 		if (p->ptap_name[0] != '\0')
346 			if ((name = strdup(p->ptap_name)) == NULL)
347 				return ENOMEM;
348 
349 	newthread = NULL;
350 
351 	/*
352 	 * Try to reclaim a dead thread.
353 	 */
354 	if (!PTQ_EMPTY(&pthread__deadqueue)) {
355 		pthread_mutex_lock(&pthread__deadqueue_lock);
356 		PTQ_FOREACH(newthread, &pthread__deadqueue, pt_deadq) {
357 			/* Still running? */
358 			if (newthread->pt_lwpctl->lc_curcpu ==
359 			    LWPCTL_CPU_EXITED ||
360 			    (_lwp_kill(newthread->pt_lid, 0) == -1 &&
361 			    errno == ESRCH))
362 				break;
363 		}
364 		if (newthread)
365 			PTQ_REMOVE(&pthread__deadqueue, newthread, pt_deadq);
366 		pthread_mutex_unlock(&pthread__deadqueue_lock);
367 	}
368 
369 	/*
370 	 * If necessary set up a stack, allocate space for a pthread_st,
371 	 * and initialize it.
372 	 */
373 	if (newthread == NULL) {
374 		ret = pthread__stackalloc(&newthread);
375 		if (ret != 0) {
376 			if (name)
377 				free(name);
378 			return ret;
379 		}
380 
381 		/* This is used only when creating the thread. */
382 		_INITCONTEXT_U(&newthread->pt_uc);
383 #ifdef PTHREAD_MACHINE_HAS_ID_REGISTER
384 		pthread__uc_id(&newthread->pt_uc) = newthread;
385 #endif
386 		newthread->pt_uc.uc_stack = newthread->pt_stack;
387 		newthread->pt_uc.uc_link = NULL;
388 
389 		/* Add to list of all threads. */
390 		pthread_rwlock_wrlock(&pthread__alltree_lock);
391 		PTQ_INSERT_TAIL(&pthread__allqueue, newthread, pt_allq);
392 		RB_INSERT(__pthread__alltree, &pthread__alltree, newthread);
393 		pthread_rwlock_unlock(&pthread__alltree_lock);
394 
395 		/* Will be reset by the thread upon exit. */
396 		pthread__initthread(newthread);
397 	}
398 
399 	/*
400 	 * Create the new LWP.
401 	 */
402 	pthread__scrubthread(newthread, name, nattr.pta_flags);
403 	newthread->pt_func = startfunc;
404 	newthread->pt_arg = arg;
405 	_lwp_makecontext(&newthread->pt_uc, pthread__create_tramp,
406 	    newthread, newthread, newthread->pt_stack.ss_sp,
407 	    newthread->pt_stack.ss_size);
408 
409 	flag = LWP_DETACHED;
410 	if ((newthread->pt_flags & PT_FLAG_SUSPENDED) != 0 ||
411 	    (nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0)
412 		flag |= LWP_SUSPENDED;
413 	ret = _lwp_create(&newthread->pt_uc, flag, &newthread->pt_lid);
414 	if (ret != 0) {
415 		free(name);
416 		newthread->pt_state = PT_STATE_DEAD;
417 		pthread_mutex_lock(&pthread__deadqueue_lock);
418 		PTQ_INSERT_HEAD(&pthread__deadqueue, newthread, pt_deadq);
419 		pthread_mutex_unlock(&pthread__deadqueue_lock);
420 		return ret;
421 	}
422 
423 	if ((nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0) {
424 		if (p != NULL) {
425 			(void)pthread_setschedparam(newthread, p->ptap_policy,
426 			    &p->ptap_sp);
427 		}
428 		if ((newthread->pt_flags & PT_FLAG_SUSPENDED) == 0) {
429 			(void)_lwp_continue(newthread->pt_lid);
430 		}
431 	}
432 
433 	*thread = newthread;
434 
435 	return 0;
436 }
437 
438 
439 static void
440 pthread__create_tramp(void *cookie)
441 {
442 	pthread_t self;
443 	void *retval;
444 
445 	self = cookie;
446 
447 	/*
448 	 * Throw away some stack in a feeble attempt to reduce cache
449 	 * thrash.  May help for SMT processors.  XXX We should not
450 	 * be allocating stacks on fixed 2MB boundaries.  Needs a
451 	 * thread register or decent thread local storage.
452 	 *
453 	 * Note that we may race with the kernel in _lwp_create(),
454 	 * and so pt_lid can be unset at this point, but we don't
455 	 * care.
456 	 */
457 	(void)alloca(((unsigned)self->pt_lid & 7) << 8);
458 
459 	if (self->pt_name != NULL) {
460 		pthread_mutex_lock(&self->pt_lock);
461 		if (self->pt_name != NULL)
462 			(void)_lwp_setname(0, self->pt_name);
463 		pthread_mutex_unlock(&self->pt_lock);
464 	}
465 
466 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &self->pt_lwpctl)) {
467 		err(1, "_lwp_ctl");
468 	}
469 
470 	retval = (*self->pt_func)(self->pt_arg);
471 
472 	pthread_exit(retval);
473 
474 	/*NOTREACHED*/
475 	pthread__abort();
476 }
477 
478 int
479 pthread_suspend_np(pthread_t thread)
480 {
481 	pthread_t self;
482 
483 	self = pthread__self();
484 	if (self == thread) {
485 		return EDEADLK;
486 	}
487 	if (pthread__find(thread) != 0)
488 		return ESRCH;
489 	if (_lwp_suspend(thread->pt_lid) == 0)
490 		return 0;
491 	return errno;
492 }
493 
494 int
495 pthread_resume_np(pthread_t thread)
496 {
497 
498 	if (pthread__find(thread) != 0)
499 		return ESRCH;
500 	if (_lwp_continue(thread->pt_lid) == 0)
501 		return 0;
502 	return errno;
503 }
504 
505 void
506 pthread_exit(void *retval)
507 {
508 	pthread_t self;
509 	struct pt_clean_t *cleanup;
510 	char *name;
511 
512 	self = pthread__self();
513 
514 	/* Disable cancellability. */
515 	pthread_mutex_lock(&self->pt_lock);
516 	self->pt_flags |= PT_FLAG_CS_DISABLED;
517 	self->pt_cancel = 0;
518 
519 	/* Call any cancellation cleanup handlers */
520 	if (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
521 		pthread_mutex_unlock(&self->pt_lock);
522 		while (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
523 			cleanup = PTQ_FIRST(&self->pt_cleanup_stack);
524 			PTQ_REMOVE(&self->pt_cleanup_stack, cleanup, ptc_next);
525 			(*cleanup->ptc_cleanup)(cleanup->ptc_arg);
526 		}
527 		pthread_mutex_lock(&self->pt_lock);
528 	}
529 
530 	/* Perform cleanup of thread-specific data */
531 	pthread__destroy_tsd(self);
532 
533 	/* Signal our exit. */
534 	self->pt_exitval = retval;
535 	if (self->pt_flags & PT_FLAG_DETACHED) {
536 		self->pt_state = PT_STATE_DEAD;
537 		name = self->pt_name;
538 		self->pt_name = NULL;
539 		pthread_mutex_unlock(&self->pt_lock);
540 		if (name != NULL)
541 			free(name);
542 		pthread_mutex_lock(&pthread__deadqueue_lock);
543 		PTQ_INSERT_TAIL(&pthread__deadqueue, self, pt_deadq);
544 		pthread_mutex_unlock(&pthread__deadqueue_lock);
545 		_lwp_exit();
546 	} else {
547 		self->pt_state = PT_STATE_ZOMBIE;
548 		pthread_cond_broadcast(&self->pt_joiners);
549 		pthread_mutex_unlock(&self->pt_lock);
550 		/* Note: name will be freed by the joiner. */
551 		_lwp_exit();
552 	}
553 
554 	/*NOTREACHED*/
555 	pthread__abort();
556 	exit(1);
557 }
558 
559 
560 int
561 pthread_join(pthread_t thread, void **valptr)
562 {
563 	pthread_t self;
564 	int error;
565 
566 	self = pthread__self();
567 
568 	if (pthread__find(thread) != 0)
569 		return ESRCH;
570 
571 	if (thread->pt_magic != PT_MAGIC)
572 		return EINVAL;
573 
574 	if (thread == self)
575 		return EDEADLK;
576 
577 	self->pt_droplock = &thread->pt_lock;
578 	pthread_mutex_lock(&thread->pt_lock);
579 	for (;;) {
580 		if (thread->pt_state == PT_STATE_ZOMBIE)
581 			break;
582 		if (thread->pt_state == PT_STATE_DEAD) {
583 			pthread_mutex_unlock(&thread->pt_lock);
584 			self->pt_droplock = NULL;
585 			return ESRCH;
586 		}
587 		if ((thread->pt_flags & PT_FLAG_DETACHED) != 0) {
588 			pthread_mutex_unlock(&thread->pt_lock);
589 			self->pt_droplock = NULL;
590 			return EINVAL;
591 		}
592 		error = pthread_cond_wait(&thread->pt_joiners,
593 		    &thread->pt_lock);
594 		if (error != 0) {
595 			pthread__errorfunc(__FILE__, __LINE__,
596 			    __func__, "unexpected return from cond_wait()");
597 		}
598 
599 	}
600 	pthread__testcancel(self);
601 	if (valptr != NULL)
602 		*valptr = thread->pt_exitval;
603 	/* pthread__reap() will drop the lock. */
604 	pthread__reap(thread);
605 	self->pt_droplock = NULL;
606 
607 	return 0;
608 }
609 
610 static void
611 pthread__reap(pthread_t thread)
612 {
613 	char *name;
614 
615 	name = thread->pt_name;
616 	thread->pt_name = NULL;
617 	thread->pt_state = PT_STATE_DEAD;
618 	pthread_mutex_unlock(&thread->pt_lock);
619 
620 	pthread_mutex_lock(&pthread__deadqueue_lock);
621 	PTQ_INSERT_HEAD(&pthread__deadqueue, thread, pt_deadq);
622 	pthread_mutex_unlock(&pthread__deadqueue_lock);
623 
624 	if (name != NULL)
625 		free(name);
626 }
627 
628 int
629 pthread_equal(pthread_t t1, pthread_t t2)
630 {
631 
632 	/* Nothing special here. */
633 	return (t1 == t2);
634 }
635 
636 
637 int
638 pthread_detach(pthread_t thread)
639 {
640 
641 	if (pthread__find(thread) != 0)
642 		return ESRCH;
643 
644 	if (thread->pt_magic != PT_MAGIC)
645 		return EINVAL;
646 
647 	pthread_mutex_lock(&thread->pt_lock);
648 	thread->pt_flags |= PT_FLAG_DETACHED;
649 	if (thread->pt_state == PT_STATE_ZOMBIE) {
650 		/* pthread__reap() will drop the lock. */
651 		pthread__reap(thread);
652 	} else {
653 		/*
654 		 * Not valid for threads to be waiting in
655 		 * pthread_join() (there are intractable
656 		 * sync issues from the application
657 		 * perspective), but give those threads
658 		 * a chance anyway.
659 		 */
660 		pthread_cond_broadcast(&thread->pt_joiners);
661 		pthread_mutex_unlock(&thread->pt_lock);
662 	}
663 
664 	return 0;
665 }
666 
667 
668 int
669 pthread_getname_np(pthread_t thread, char *name, size_t len)
670 {
671 
672 	if (pthread__find(thread) != 0)
673 		return ESRCH;
674 
675 	if (thread->pt_magic != PT_MAGIC)
676 		return EINVAL;
677 
678 	pthread_mutex_lock(&thread->pt_lock);
679 	if (thread->pt_name == NULL)
680 		name[0] = '\0';
681 	else
682 		strlcpy(name, thread->pt_name, len);
683 	pthread_mutex_unlock(&thread->pt_lock);
684 
685 	return 0;
686 }
687 
688 
689 int
690 pthread_setname_np(pthread_t thread, const char *name, void *arg)
691 {
692 	char *oldname, *cp, newname[PTHREAD_MAX_NAMELEN_NP];
693 	int namelen;
694 
695 	if (pthread__find(thread) != 0)
696 		return ESRCH;
697 
698 	if (thread->pt_magic != PT_MAGIC)
699 		return EINVAL;
700 
701 	namelen = snprintf(newname, sizeof(newname), name, arg);
702 	if (namelen >= PTHREAD_MAX_NAMELEN_NP)
703 		return EINVAL;
704 
705 	cp = strdup(newname);
706 	if (cp == NULL)
707 		return ENOMEM;
708 
709 	pthread_mutex_lock(&thread->pt_lock);
710 	oldname = thread->pt_name;
711 	thread->pt_name = cp;
712 	(void)_lwp_setname(thread->pt_lid, cp);
713 	pthread_mutex_unlock(&thread->pt_lock);
714 
715 	if (oldname != NULL)
716 		free(oldname);
717 
718 	return 0;
719 }
720 
721 
722 
723 /*
724  * XXX There should be a way for applications to use the efficent
725  *  inline version, but there are opacity/namespace issues.
726  */
727 pthread_t
728 pthread_self(void)
729 {
730 
731 	return pthread__self();
732 }
733 
734 
735 int
736 pthread_cancel(pthread_t thread)
737 {
738 
739 	if (pthread__find(thread) != 0)
740 		return ESRCH;
741 	pthread_mutex_lock(&thread->pt_lock);
742 	thread->pt_flags |= PT_FLAG_CS_PENDING;
743 	if ((thread->pt_flags & PT_FLAG_CS_DISABLED) == 0) {
744 		thread->pt_cancel = 1;
745 		pthread_mutex_unlock(&thread->pt_lock);
746 		_lwp_wakeup(thread->pt_lid);
747 	} else
748 		pthread_mutex_unlock(&thread->pt_lock);
749 
750 	return 0;
751 }
752 
753 
754 int
755 pthread_setcancelstate(int state, int *oldstate)
756 {
757 	pthread_t self;
758 	int retval;
759 
760 	self = pthread__self();
761 	retval = 0;
762 
763 	pthread_mutex_lock(&self->pt_lock);
764 
765 	if (oldstate != NULL) {
766 		if (self->pt_flags & PT_FLAG_CS_DISABLED)
767 			*oldstate = PTHREAD_CANCEL_DISABLE;
768 		else
769 			*oldstate = PTHREAD_CANCEL_ENABLE;
770 	}
771 
772 	if (state == PTHREAD_CANCEL_DISABLE) {
773 		self->pt_flags |= PT_FLAG_CS_DISABLED;
774 		if (self->pt_cancel) {
775 			self->pt_flags |= PT_FLAG_CS_PENDING;
776 			self->pt_cancel = 0;
777 		}
778 	} else if (state == PTHREAD_CANCEL_ENABLE) {
779 		self->pt_flags &= ~PT_FLAG_CS_DISABLED;
780 		/*
781 		 * If a cancellation was requested while cancellation
782 		 * was disabled, note that fact for future
783 		 * cancellation tests.
784 		 */
785 		if (self->pt_flags & PT_FLAG_CS_PENDING) {
786 			self->pt_cancel = 1;
787 			/* This is not a deferred cancellation point. */
788 			if (self->pt_flags & PT_FLAG_CS_ASYNC) {
789 				pthread_mutex_unlock(&self->pt_lock);
790 				pthread__cancelled();
791 			}
792 		}
793 	} else
794 		retval = EINVAL;
795 
796 	pthread_mutex_unlock(&self->pt_lock);
797 
798 	return retval;
799 }
800 
801 
802 int
803 pthread_setcanceltype(int type, int *oldtype)
804 {
805 	pthread_t self;
806 	int retval;
807 
808 	self = pthread__self();
809 	retval = 0;
810 
811 	pthread_mutex_lock(&self->pt_lock);
812 
813 	if (oldtype != NULL) {
814 		if (self->pt_flags & PT_FLAG_CS_ASYNC)
815 			*oldtype = PTHREAD_CANCEL_ASYNCHRONOUS;
816 		else
817 			*oldtype = PTHREAD_CANCEL_DEFERRED;
818 	}
819 
820 	if (type == PTHREAD_CANCEL_ASYNCHRONOUS) {
821 		self->pt_flags |= PT_FLAG_CS_ASYNC;
822 		if (self->pt_cancel) {
823 			pthread_mutex_unlock(&self->pt_lock);
824 			pthread__cancelled();
825 		}
826 	} else if (type == PTHREAD_CANCEL_DEFERRED)
827 		self->pt_flags &= ~PT_FLAG_CS_ASYNC;
828 	else
829 		retval = EINVAL;
830 
831 	pthread_mutex_unlock(&self->pt_lock);
832 
833 	return retval;
834 }
835 
836 
837 void
838 pthread_testcancel(void)
839 {
840 	pthread_t self;
841 
842 	self = pthread__self();
843 	if (self->pt_cancel)
844 		pthread__cancelled();
845 }
846 
847 
848 /*
849  * POSIX requires that certain functions return an error rather than
850  * invoking undefined behavior even when handed completely bogus
851  * pthread_t values, e.g. stack garbage or (pthread_t)666. This
852  * utility routine searches the list of threads for the pthread_t
853  * value without dereferencing it.
854  */
855 int
856 pthread__find(pthread_t id)
857 {
858 	pthread_t target;
859 
860 	pthread_rwlock_rdlock(&pthread__alltree_lock);
861 	/* LINTED */
862 	target = RB_FIND(__pthread__alltree, &pthread__alltree, id);
863 	pthread_rwlock_unlock(&pthread__alltree_lock);
864 
865 	if (target == NULL || target->pt_state == PT_STATE_DEAD)
866 		return ESRCH;
867 
868 	return 0;
869 }
870 
871 
872 void
873 pthread__testcancel(pthread_t self)
874 {
875 
876 	if (self->pt_cancel)
877 		pthread__cancelled();
878 }
879 
880 
881 void
882 pthread__cancelled(void)
883 {
884 	pthread_mutex_t *droplock;
885 	pthread_t self;
886 
887 	self = pthread__self();
888 	droplock = self->pt_droplock;
889 	self->pt_droplock = NULL;
890 
891 	if (droplock != NULL && pthread_mutex_held_np(droplock))
892 		pthread_mutex_unlock(droplock);
893 
894 	pthread_exit(PTHREAD_CANCELED);
895 }
896 
897 
898 void
899 pthread__cleanup_push(void (*cleanup)(void *), void *arg, void *store)
900 {
901 	pthread_t self;
902 	struct pt_clean_t *entry;
903 
904 	self = pthread__self();
905 	entry = store;
906 	entry->ptc_cleanup = cleanup;
907 	entry->ptc_arg = arg;
908 	PTQ_INSERT_HEAD(&self->pt_cleanup_stack, entry, ptc_next);
909 }
910 
911 
912 void
913 pthread__cleanup_pop(int ex, void *store)
914 {
915 	pthread_t self;
916 	struct pt_clean_t *entry;
917 
918 	self = pthread__self();
919 	entry = store;
920 
921 	PTQ_REMOVE(&self->pt_cleanup_stack, entry, ptc_next);
922 	if (ex)
923 		(*entry->ptc_cleanup)(entry->ptc_arg);
924 }
925 
926 
927 int *
928 pthread__errno(void)
929 {
930 	pthread_t self;
931 
932 	self = pthread__self();
933 
934 	return &(self->pt_errno);
935 }
936 
937 ssize_t	_sys_write(int, const void *, size_t);
938 
939 void
940 pthread__assertfunc(const char *file, int line, const char *function,
941 		    const char *expr)
942 {
943 	char buf[1024];
944 	int len;
945 
946 	/*
947 	 * snprintf should not acquire any locks, or we could
948 	 * end up deadlocked if the assert caller held locks.
949 	 */
950 	len = snprintf(buf, 1024,
951 	    "assertion \"%s\" failed: file \"%s\", line %d%s%s%s\n",
952 	    expr, file, line,
953 	    function ? ", function \"" : "",
954 	    function ? function : "",
955 	    function ? "\"" : "");
956 
957 	_sys_write(STDERR_FILENO, buf, (size_t)len);
958 	(void)kill(getpid(), SIGABRT);
959 
960 	_exit(1);
961 }
962 
963 
964 void
965 pthread__errorfunc(const char *file, int line, const char *function,
966 		   const char *msg)
967 {
968 	char buf[1024];
969 	size_t len;
970 
971 	if (pthread__diagassert == 0)
972 		return;
973 
974 	/*
975 	 * snprintf should not acquire any locks, or we could
976 	 * end up deadlocked if the assert caller held locks.
977 	 */
978 	len = snprintf(buf, 1024,
979 	    "%s: Error detected by libpthread: %s.\n"
980 	    "Detected by file \"%s\", line %d%s%s%s.\n"
981 	    "See pthread(3) for information.\n",
982 	    getprogname(), msg, file, line,
983 	    function ? ", function \"" : "",
984 	    function ? function : "",
985 	    function ? "\"" : "");
986 
987 	if (pthread__diagassert & DIAGASSERT_STDERR)
988 		_sys_write(STDERR_FILENO, buf, len);
989 
990 	if (pthread__diagassert & DIAGASSERT_SYSLOG)
991 		syslog(LOG_DEBUG | LOG_USER, "%s", buf);
992 
993 	if (pthread__diagassert & DIAGASSERT_ABORT) {
994 		(void)kill(getpid(), SIGABRT);
995 		_exit(1);
996 	}
997 }
998 
999 /*
1000  * Thread park/unpark operations.  The kernel operations are
1001  * modelled after a brief description from "Multithreading in
1002  * the Solaris Operating Environment":
1003  *
1004  * http://www.sun.com/software/whitepapers/solaris9/multithread.pdf
1005  */
1006 
1007 #define	OOPS(msg)			\
1008     pthread__errorfunc(__FILE__, __LINE__, __func__, msg)
1009 
1010 int
1011 pthread__park(pthread_t self, pthread_mutex_t *lock,
1012 	      pthread_queue_t *queue, const struct timespec *abstime,
1013 	      int cancelpt, const void *hint)
1014 {
1015 	int rv, error;
1016 	void *obj;
1017 
1018 	/*
1019 	 * For non-interlocked release of mutexes we need a store
1020 	 * barrier before incrementing pt_blocking away from zero.
1021 	 * This is provided by pthread_mutex_unlock().
1022 	 */
1023 	self->pt_willpark = 1;
1024 	pthread_mutex_unlock(lock);
1025 	self->pt_willpark = 0;
1026 	self->pt_blocking++;
1027 
1028 	/*
1029 	 * Wait until we are awoken by a pending unpark operation,
1030 	 * a signal, an unpark posted after we have gone asleep,
1031 	 * or an expired timeout.
1032 	 *
1033 	 * It is fine to test the value of pt_sleepobj without
1034 	 * holding any locks, because:
1035 	 *
1036 	 * o Only the blocking thread (this thread) ever sets them
1037 	 *   to a non-NULL value.
1038 	 *
1039 	 * o Other threads may set them NULL, but if they do so they
1040 	 *   must also make this thread return from _lwp_park.
1041 	 *
1042 	 * o _lwp_park, _lwp_unpark and _lwp_unpark_all are system
1043 	 *   calls and all make use of spinlocks in the kernel.  So
1044 	 *   these system calls act as full memory barriers, and will
1045 	 *   ensure that the calling CPU's store buffers are drained.
1046 	 *   In combination with the spinlock release before unpark,
1047 	 *   this means that modification of pt_sleepobj/onq by another
1048 	 *   thread will become globally visible before that thread
1049 	 *   schedules an unpark operation on this thread.
1050 	 *
1051 	 * Note: the test in the while() statement dodges the park op if
1052 	 * we have already been awoken, unless there is another thread to
1053 	 * awaken.  This saves a syscall - if we were already awakened,
1054 	 * the next call to _lwp_park() would need to return early in order
1055 	 * to eat the previous wakeup.
1056 	 */
1057 	rv = 0;
1058 	do {
1059 		/*
1060 		 * If we deferred unparking a thread, arrange to
1061 		 * have _lwp_park() restart it before blocking.
1062 		 */
1063 		error = _lwp_park(abstime, self->pt_unpark, hint, hint);
1064 		self->pt_unpark = 0;
1065 		if (error != 0) {
1066 			switch (rv = errno) {
1067 			case EINTR:
1068 			case EALREADY:
1069 				rv = 0;
1070 				break;
1071 			case ETIMEDOUT:
1072 				break;
1073 			default:
1074 				OOPS("_lwp_park failed");
1075 				break;
1076 			}
1077 		}
1078 		/* Check for cancellation. */
1079 		if (cancelpt && self->pt_cancel)
1080 			rv = EINTR;
1081 	} while (self->pt_sleepobj != NULL && rv == 0);
1082 
1083 	/*
1084 	 * If we have been awoken early but are still on the queue,
1085 	 * then remove ourself.  Again, it's safe to do the test
1086 	 * without holding any locks.
1087 	 */
1088 	if (__predict_false(self->pt_sleepobj != NULL)) {
1089 		pthread_mutex_lock(lock);
1090 		if ((obj = self->pt_sleepobj) != NULL) {
1091 			PTQ_REMOVE(queue, self, pt_sleep);
1092 			self->pt_sleepobj = NULL;
1093 			if (obj != NULL && self->pt_early != NULL)
1094 				(*self->pt_early)(obj);
1095 		}
1096 		pthread_mutex_unlock(lock);
1097 	}
1098 	self->pt_early = NULL;
1099 	self->pt_blocking--;
1100 	membar_sync();
1101 
1102 	return rv;
1103 }
1104 
1105 void
1106 pthread__unpark(pthread_queue_t *queue, pthread_t self,
1107 		pthread_mutex_t *interlock)
1108 {
1109 	pthread_t target;
1110 	u_int max;
1111 	size_t nwaiters;
1112 
1113 	max = pthread__unpark_max;
1114 	nwaiters = self->pt_nwaiters;
1115 	target = PTQ_FIRST(queue);
1116 	if (nwaiters == max) {
1117 		/* Overflow. */
1118 		(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1119 		    __UNVOLATILE(&interlock->ptm_waiters));
1120 		nwaiters = 0;
1121 	}
1122 	target->pt_sleepobj = NULL;
1123 	self->pt_waiters[nwaiters++] = target->pt_lid;
1124 	PTQ_REMOVE(queue, target, pt_sleep);
1125 	self->pt_nwaiters = nwaiters;
1126 	pthread__mutex_deferwake(self, interlock);
1127 }
1128 
1129 void
1130 pthread__unpark_all(pthread_queue_t *queue, pthread_t self,
1131 		    pthread_mutex_t *interlock)
1132 {
1133 	pthread_t target;
1134 	u_int max;
1135 	size_t nwaiters;
1136 
1137 	max = pthread__unpark_max;
1138 	nwaiters = self->pt_nwaiters;
1139 	PTQ_FOREACH(target, queue, pt_sleep) {
1140 		if (nwaiters == max) {
1141 			/* Overflow. */
1142 			(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1143 			    __UNVOLATILE(&interlock->ptm_waiters));
1144 			nwaiters = 0;
1145 		}
1146 		target->pt_sleepobj = NULL;
1147 		self->pt_waiters[nwaiters++] = target->pt_lid;
1148 	}
1149 	self->pt_nwaiters = nwaiters;
1150 	PTQ_INIT(queue);
1151 	pthread__mutex_deferwake(self, interlock);
1152 }
1153 
1154 #undef	OOPS
1155 
1156 /*
1157  * Allocate a stack for a thread, and set it up. It needs to be aligned, so
1158  * that a thread can find itself by its stack pointer.
1159  */
1160 static int
1161 pthread__stackalloc(pthread_t *newt)
1162 {
1163 	void *addr;
1164 
1165 	addr = mmap(NULL, pthread__stacksize, PROT_READ|PROT_WRITE,
1166 	    MAP_ANON|MAP_PRIVATE | MAP_ALIGNED(pthread__stacksize_lg),
1167 	    -1, (off_t)0);
1168 
1169 	if (addr == MAP_FAILED)
1170 		return ENOMEM;
1171 
1172 	pthread__assert(((intptr_t)addr & pthread__stackmask) == 0);
1173 
1174 	return pthread__stackid_setup(addr, pthread__stacksize, newt);
1175 }
1176 
1177 
1178 /*
1179  * Set up the slightly special stack for the "initial" thread, which
1180  * runs on the normal system stack, and thus gets slightly different
1181  * treatment.
1182  */
1183 static void
1184 pthread__initmain(pthread_t *newt)
1185 {
1186 	struct rlimit slimit;
1187 	size_t pagesize;
1188 	pthread_t t;
1189 	void *base;
1190 	size_t size;
1191 	int error, ret;
1192 	char *value;
1193 
1194 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1195 	pthread__stacksize = 0;
1196 	ret = getrlimit(RLIMIT_STACK, &slimit);
1197 	if (ret == -1)
1198 		err(1, "Couldn't get stack resource consumption limits");
1199 
1200 	value = pthread__getenv("PTHREAD_STACKSIZE");
1201 	if (value != NULL) {
1202 		pthread__stacksize = atoi(value) * 1024;
1203 		if (pthread__stacksize > slimit.rlim_cur)
1204 			pthread__stacksize = (size_t)slimit.rlim_cur;
1205 	}
1206 	if (pthread__stacksize == 0)
1207 		pthread__stacksize = (size_t)slimit.rlim_cur;
1208 	if (pthread__stacksize < 4 * pagesize)
1209 		errx(1, "Stacksize limit is too low, minimum %zd kbyte.",
1210 		    4 * pagesize / 1024);
1211 
1212 	pthread__stacksize_lg = -1;
1213 	while (pthread__stacksize) {
1214 		pthread__stacksize >>= 1;
1215 		pthread__stacksize_lg++;
1216 	}
1217 
1218 	pthread__stacksize = (1 << pthread__stacksize_lg);
1219 	pthread__stackmask = pthread__stacksize - 1;
1220 	pthread__threadmask = ~pthread__stackmask;
1221 
1222 	base = (void *)(pthread__sp() & pthread__threadmask);
1223 	size = pthread__stacksize;
1224 
1225 	error = pthread__stackid_setup(base, size, &t);
1226 	if (error) {
1227 		/* XXX */
1228 		errx(2, "failed to setup main thread: error=%d", error);
1229 	}
1230 
1231 	*newt = t;
1232 
1233 	/* Set up identity register. */
1234 	(void)_lwp_setprivate(t);
1235 }
1236 
1237 static int
1238 /*ARGSUSED*/
1239 pthread__stackid_setup(void *base, size_t size, pthread_t *tp)
1240 {
1241 	pthread_t t;
1242 	void *redaddr;
1243 	size_t pagesize;
1244 	int ret;
1245 
1246 	t = base;
1247 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1248 
1249 	/*
1250 	 * Put a pointer to the pthread in the bottom (but
1251          * redzone-protected section) of the stack.
1252 	 */
1253 	redaddr = STACK_SHRINK(STACK_MAX(base, size), pagesize);
1254 	t->pt_stack.ss_size = size - 2 * pagesize;
1255 #ifdef __MACHINE_STACK_GROWS_UP
1256 	t->pt_stack.ss_sp = (char *)(void *)base + pagesize;
1257 #else
1258 	t->pt_stack.ss_sp = (char *)(void *)base + 2 * pagesize;
1259 #endif
1260 
1261 	/* Protect the next-to-bottom stack page as a red zone. */
1262 	ret = mprotect(redaddr, pagesize, PROT_NONE);
1263 	if (ret == -1) {
1264 		return errno;
1265 	}
1266 	*tp = t;
1267 	return 0;
1268 }
1269 
1270 #ifndef lint
1271 static int
1272 pthread__cmp(struct __pthread_st *a, struct __pthread_st *b)
1273 {
1274 
1275 	if ((uintptr_t)a < (uintptr_t)b)
1276 		return (-1);
1277 	else if (a == b)
1278 		return 0;
1279 	else
1280 		return 1;
1281 }
1282 RB_GENERATE_STATIC(__pthread__alltree, __pthread_st, pt_alltree, pthread__cmp)
1283 #endif
1284 
1285 /* Because getenv() wants to use locks. */
1286 char *
1287 pthread__getenv(const char *name)
1288 {
1289 	extern char *__findenv(const char *, int *);
1290 	int off;
1291 
1292 	return __findenv(name, &off);
1293 }
1294 
1295 pthread_mutex_t *
1296 pthread__hashlock(volatile const void *p)
1297 {
1298 	uintptr_t v;
1299 
1300 	v = (uintptr_t)p;
1301 	return &hashlocks[((v >> 9) ^ (v >> 3)) & (NHASHLOCK - 1)].mutex;
1302 }
1303 
1304 int
1305 pthread__checkpri(int pri)
1306 {
1307 	static int havepri;
1308 	static long min, max;
1309 
1310 	if (!havepri) {
1311 		min = sysconf(_SC_SCHED_PRI_MIN);
1312 		max = sysconf(_SC_SCHED_PRI_MAX);
1313 		havepri = 1;
1314 	}
1315 	return (pri < min || pri > max) ? EINVAL : 0;
1316 }
1317