xref: /netbsd-src/lib/libpthread/pthread.c (revision 313c6c94c424eed90c7b7e494aa83308a0a5d0ce)
1 /*	$NetBSD: pthread.c,v 1.110 2009/05/17 14:49:00 ad 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 1.110 2009/05/17 14:49:00 ad 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 		newthread = PTQ_FIRST(&pthread__deadqueue);
357 		if (newthread != NULL) {
358 			PTQ_REMOVE(&pthread__deadqueue, newthread, pt_deadq);
359 			pthread_mutex_unlock(&pthread__deadqueue_lock);
360 			/* Still running? */
361 			if (newthread->pt_lwpctl->lc_curcpu !=
362 			    LWPCTL_CPU_EXITED &&
363 			    (_lwp_kill(newthread->pt_lid, 0) == 0 ||
364 			    errno != ESRCH)) {
365 				pthread_mutex_lock(&pthread__deadqueue_lock);
366 				PTQ_INSERT_TAIL(&pthread__deadqueue,
367 				    newthread, pt_deadq);
368 				pthread_mutex_unlock(&pthread__deadqueue_lock);
369 				newthread = NULL;
370 			}
371 		} else
372 			pthread_mutex_unlock(&pthread__deadqueue_lock);
373 	}
374 
375 	/*
376 	 * If necessary set up a stack, allocate space for a pthread_st,
377 	 * and initialize it.
378 	 */
379 	if (newthread == NULL) {
380 		ret = pthread__stackalloc(&newthread);
381 		if (ret != 0) {
382 			if (name)
383 				free(name);
384 			return ret;
385 		}
386 
387 		/* This is used only when creating the thread. */
388 		_INITCONTEXT_U(&newthread->pt_uc);
389 #ifdef PTHREAD_MACHINE_HAS_ID_REGISTER
390 		pthread__uc_id(&newthread->pt_uc) = newthread;
391 #endif
392 		newthread->pt_uc.uc_stack = newthread->pt_stack;
393 		newthread->pt_uc.uc_link = NULL;
394 
395 		/* Add to list of all threads. */
396 		pthread_rwlock_wrlock(&pthread__alltree_lock);
397 		PTQ_INSERT_TAIL(&pthread__allqueue, newthread, pt_allq);
398 		RB_INSERT(__pthread__alltree, &pthread__alltree, newthread);
399 		pthread_rwlock_unlock(&pthread__alltree_lock);
400 
401 		/* Will be reset by the thread upon exit. */
402 		pthread__initthread(newthread);
403 	}
404 
405 	/*
406 	 * Create the new LWP.
407 	 */
408 	pthread__scrubthread(newthread, name, nattr.pta_flags);
409 	newthread->pt_func = startfunc;
410 	newthread->pt_arg = arg;
411 	_lwp_makecontext(&newthread->pt_uc, pthread__create_tramp,
412 	    newthread, newthread, newthread->pt_stack.ss_sp,
413 	    newthread->pt_stack.ss_size);
414 
415 	flag = LWP_DETACHED;
416 	if ((newthread->pt_flags & PT_FLAG_SUSPENDED) != 0 ||
417 	    (nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0)
418 		flag |= LWP_SUSPENDED;
419 	ret = _lwp_create(&newthread->pt_uc, flag, &newthread->pt_lid);
420 	if (ret != 0) {
421 		free(name);
422 		newthread->pt_state = PT_STATE_DEAD;
423 		pthread_mutex_lock(&pthread__deadqueue_lock);
424 		PTQ_INSERT_HEAD(&pthread__deadqueue, newthread, pt_deadq);
425 		pthread_mutex_unlock(&pthread__deadqueue_lock);
426 		return ret;
427 	}
428 
429 	if ((nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0) {
430 		if (p != NULL) {
431 			(void)pthread_setschedparam(newthread, p->ptap_policy,
432 			    &p->ptap_sp);
433 		}
434 		if ((newthread->pt_flags & PT_FLAG_SUSPENDED) == 0) {
435 			(void)_lwp_continue(newthread->pt_lid);
436 		}
437 	}
438 
439 	*thread = newthread;
440 
441 	return 0;
442 }
443 
444 
445 static void
446 pthread__create_tramp(void *cookie)
447 {
448 	pthread_t self;
449 	void *retval;
450 
451 	self = cookie;
452 
453 	/*
454 	 * Throw away some stack in a feeble attempt to reduce cache
455 	 * thrash.  May help for SMT processors.  XXX We should not
456 	 * be allocating stacks on fixed 2MB boundaries.  Needs a
457 	 * thread register or decent thread local storage.
458 	 *
459 	 * Note that we may race with the kernel in _lwp_create(),
460 	 * and so pt_lid can be unset at this point, but we don't
461 	 * care.
462 	 */
463 	(void)alloca(((unsigned)self->pt_lid & 7) << 8);
464 
465 	if (self->pt_name != NULL) {
466 		pthread_mutex_lock(&self->pt_lock);
467 		if (self->pt_name != NULL)
468 			(void)_lwp_setname(0, self->pt_name);
469 		pthread_mutex_unlock(&self->pt_lock);
470 	}
471 
472 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &self->pt_lwpctl)) {
473 		err(1, "_lwp_ctl");
474 	}
475 
476 	retval = (*self->pt_func)(self->pt_arg);
477 
478 	pthread_exit(retval);
479 
480 	/*NOTREACHED*/
481 	pthread__abort();
482 }
483 
484 int
485 pthread_suspend_np(pthread_t thread)
486 {
487 	pthread_t self;
488 
489 	self = pthread__self();
490 	if (self == thread) {
491 		return EDEADLK;
492 	}
493 	if (pthread__find(thread) != 0)
494 		return ESRCH;
495 	if (_lwp_suspend(thread->pt_lid) == 0)
496 		return 0;
497 	return errno;
498 }
499 
500 int
501 pthread_resume_np(pthread_t thread)
502 {
503 
504 	if (pthread__find(thread) != 0)
505 		return ESRCH;
506 	if (_lwp_continue(thread->pt_lid) == 0)
507 		return 0;
508 	return errno;
509 }
510 
511 void
512 pthread_exit(void *retval)
513 {
514 	pthread_t self;
515 	struct pt_clean_t *cleanup;
516 	char *name;
517 
518 	self = pthread__self();
519 
520 	/* Disable cancellability. */
521 	pthread_mutex_lock(&self->pt_lock);
522 	self->pt_flags |= PT_FLAG_CS_DISABLED;
523 	self->pt_cancel = 0;
524 
525 	/* Call any cancellation cleanup handlers */
526 	if (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
527 		pthread_mutex_unlock(&self->pt_lock);
528 		while (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
529 			cleanup = PTQ_FIRST(&self->pt_cleanup_stack);
530 			PTQ_REMOVE(&self->pt_cleanup_stack, cleanup, ptc_next);
531 			(*cleanup->ptc_cleanup)(cleanup->ptc_arg);
532 		}
533 		pthread_mutex_lock(&self->pt_lock);
534 	}
535 
536 	/* Perform cleanup of thread-specific data */
537 	pthread__destroy_tsd(self);
538 
539 	/* Signal our exit. */
540 	self->pt_exitval = retval;
541 	if (self->pt_flags & PT_FLAG_DETACHED) {
542 		self->pt_state = PT_STATE_DEAD;
543 		name = self->pt_name;
544 		self->pt_name = NULL;
545 		pthread_mutex_unlock(&self->pt_lock);
546 		if (name != NULL)
547 			free(name);
548 		pthread_mutex_lock(&pthread__deadqueue_lock);
549 		PTQ_INSERT_TAIL(&pthread__deadqueue, self, pt_deadq);
550 		pthread_mutex_unlock(&pthread__deadqueue_lock);
551 		_lwp_exit();
552 	} else {
553 		self->pt_state = PT_STATE_ZOMBIE;
554 		pthread_cond_broadcast(&self->pt_joiners);
555 		pthread_mutex_unlock(&self->pt_lock);
556 		/* Note: name will be freed by the joiner. */
557 		_lwp_exit();
558 	}
559 
560 	/*NOTREACHED*/
561 	pthread__abort();
562 	exit(1);
563 }
564 
565 
566 int
567 pthread_join(pthread_t thread, void **valptr)
568 {
569 	pthread_t self;
570 	int error;
571 
572 	self = pthread__self();
573 
574 	if (pthread__find(thread) != 0)
575 		return ESRCH;
576 
577 	if (thread->pt_magic != PT_MAGIC)
578 		return EINVAL;
579 
580 	if (thread == self)
581 		return EDEADLK;
582 
583 	self->pt_droplock = &thread->pt_lock;
584 	pthread_mutex_lock(&thread->pt_lock);
585 	for (;;) {
586 		if (thread->pt_state == PT_STATE_ZOMBIE)
587 			break;
588 		if (thread->pt_state == PT_STATE_DEAD) {
589 			pthread_mutex_unlock(&thread->pt_lock);
590 			self->pt_droplock = NULL;
591 			return ESRCH;
592 		}
593 		if ((thread->pt_flags & PT_FLAG_DETACHED) != 0) {
594 			pthread_mutex_unlock(&thread->pt_lock);
595 			self->pt_droplock = NULL;
596 			return EINVAL;
597 		}
598 		error = pthread_cond_wait(&thread->pt_joiners,
599 		    &thread->pt_lock);
600 		if (error != 0) {
601 			pthread__errorfunc(__FILE__, __LINE__,
602 			    __func__, "unexpected return from cond_wait()");
603 		}
604 
605 	}
606 	pthread__testcancel(self);
607 	if (valptr != NULL)
608 		*valptr = thread->pt_exitval;
609 	/* pthread__reap() will drop the lock. */
610 	pthread__reap(thread);
611 	self->pt_droplock = NULL;
612 
613 	return 0;
614 }
615 
616 static void
617 pthread__reap(pthread_t thread)
618 {
619 	char *name;
620 
621 	name = thread->pt_name;
622 	thread->pt_name = NULL;
623 	thread->pt_state = PT_STATE_DEAD;
624 	pthread_mutex_unlock(&thread->pt_lock);
625 
626 	pthread_mutex_lock(&pthread__deadqueue_lock);
627 	PTQ_INSERT_HEAD(&pthread__deadqueue, thread, pt_deadq);
628 	pthread_mutex_unlock(&pthread__deadqueue_lock);
629 
630 	if (name != NULL)
631 		free(name);
632 }
633 
634 int
635 pthread_equal(pthread_t t1, pthread_t t2)
636 {
637 
638 	/* Nothing special here. */
639 	return (t1 == t2);
640 }
641 
642 
643 int
644 pthread_detach(pthread_t thread)
645 {
646 
647 	if (pthread__find(thread) != 0)
648 		return ESRCH;
649 
650 	if (thread->pt_magic != PT_MAGIC)
651 		return EINVAL;
652 
653 	pthread_mutex_lock(&thread->pt_lock);
654 	thread->pt_flags |= PT_FLAG_DETACHED;
655 	if (thread->pt_state == PT_STATE_ZOMBIE) {
656 		/* pthread__reap() will drop the lock. */
657 		pthread__reap(thread);
658 	} else {
659 		/*
660 		 * Not valid for threads to be waiting in
661 		 * pthread_join() (there are intractable
662 		 * sync issues from the application
663 		 * perspective), but give those threads
664 		 * a chance anyway.
665 		 */
666 		pthread_cond_broadcast(&thread->pt_joiners);
667 		pthread_mutex_unlock(&thread->pt_lock);
668 	}
669 
670 	return 0;
671 }
672 
673 
674 int
675 pthread_getname_np(pthread_t thread, char *name, size_t len)
676 {
677 
678 	if (pthread__find(thread) != 0)
679 		return ESRCH;
680 
681 	if (thread->pt_magic != PT_MAGIC)
682 		return EINVAL;
683 
684 	pthread_mutex_lock(&thread->pt_lock);
685 	if (thread->pt_name == NULL)
686 		name[0] = '\0';
687 	else
688 		strlcpy(name, thread->pt_name, len);
689 	pthread_mutex_unlock(&thread->pt_lock);
690 
691 	return 0;
692 }
693 
694 
695 int
696 pthread_setname_np(pthread_t thread, const char *name, void *arg)
697 {
698 	char *oldname, *cp, newname[PTHREAD_MAX_NAMELEN_NP];
699 	int namelen;
700 
701 	if (pthread__find(thread) != 0)
702 		return ESRCH;
703 
704 	if (thread->pt_magic != PT_MAGIC)
705 		return EINVAL;
706 
707 	namelen = snprintf(newname, sizeof(newname), name, arg);
708 	if (namelen >= PTHREAD_MAX_NAMELEN_NP)
709 		return EINVAL;
710 
711 	cp = strdup(newname);
712 	if (cp == NULL)
713 		return ENOMEM;
714 
715 	pthread_mutex_lock(&thread->pt_lock);
716 	oldname = thread->pt_name;
717 	thread->pt_name = cp;
718 	(void)_lwp_setname(thread->pt_lid, cp);
719 	pthread_mutex_unlock(&thread->pt_lock);
720 
721 	if (oldname != NULL)
722 		free(oldname);
723 
724 	return 0;
725 }
726 
727 
728 
729 /*
730  * XXX There should be a way for applications to use the efficent
731  *  inline version, but there are opacity/namespace issues.
732  */
733 pthread_t
734 pthread_self(void)
735 {
736 
737 	return pthread__self();
738 }
739 
740 
741 int
742 pthread_cancel(pthread_t thread)
743 {
744 
745 	if (pthread__find(thread) != 0)
746 		return ESRCH;
747 	pthread_mutex_lock(&thread->pt_lock);
748 	thread->pt_flags |= PT_FLAG_CS_PENDING;
749 	if ((thread->pt_flags & PT_FLAG_CS_DISABLED) == 0) {
750 		thread->pt_cancel = 1;
751 		pthread_mutex_unlock(&thread->pt_lock);
752 		_lwp_wakeup(thread->pt_lid);
753 	} else
754 		pthread_mutex_unlock(&thread->pt_lock);
755 
756 	return 0;
757 }
758 
759 
760 int
761 pthread_setcancelstate(int state, int *oldstate)
762 {
763 	pthread_t self;
764 	int retval;
765 
766 	self = pthread__self();
767 	retval = 0;
768 
769 	pthread_mutex_lock(&self->pt_lock);
770 
771 	if (oldstate != NULL) {
772 		if (self->pt_flags & PT_FLAG_CS_DISABLED)
773 			*oldstate = PTHREAD_CANCEL_DISABLE;
774 		else
775 			*oldstate = PTHREAD_CANCEL_ENABLE;
776 	}
777 
778 	if (state == PTHREAD_CANCEL_DISABLE) {
779 		self->pt_flags |= PT_FLAG_CS_DISABLED;
780 		if (self->pt_cancel) {
781 			self->pt_flags |= PT_FLAG_CS_PENDING;
782 			self->pt_cancel = 0;
783 		}
784 	} else if (state == PTHREAD_CANCEL_ENABLE) {
785 		self->pt_flags &= ~PT_FLAG_CS_DISABLED;
786 		/*
787 		 * If a cancellation was requested while cancellation
788 		 * was disabled, note that fact for future
789 		 * cancellation tests.
790 		 */
791 		if (self->pt_flags & PT_FLAG_CS_PENDING) {
792 			self->pt_cancel = 1;
793 			/* This is not a deferred cancellation point. */
794 			if (self->pt_flags & PT_FLAG_CS_ASYNC) {
795 				pthread_mutex_unlock(&self->pt_lock);
796 				pthread__cancelled();
797 			}
798 		}
799 	} else
800 		retval = EINVAL;
801 
802 	pthread_mutex_unlock(&self->pt_lock);
803 
804 	return retval;
805 }
806 
807 
808 int
809 pthread_setcanceltype(int type, int *oldtype)
810 {
811 	pthread_t self;
812 	int retval;
813 
814 	self = pthread__self();
815 	retval = 0;
816 
817 	pthread_mutex_lock(&self->pt_lock);
818 
819 	if (oldtype != NULL) {
820 		if (self->pt_flags & PT_FLAG_CS_ASYNC)
821 			*oldtype = PTHREAD_CANCEL_ASYNCHRONOUS;
822 		else
823 			*oldtype = PTHREAD_CANCEL_DEFERRED;
824 	}
825 
826 	if (type == PTHREAD_CANCEL_ASYNCHRONOUS) {
827 		self->pt_flags |= PT_FLAG_CS_ASYNC;
828 		if (self->pt_cancel) {
829 			pthread_mutex_unlock(&self->pt_lock);
830 			pthread__cancelled();
831 		}
832 	} else if (type == PTHREAD_CANCEL_DEFERRED)
833 		self->pt_flags &= ~PT_FLAG_CS_ASYNC;
834 	else
835 		retval = EINVAL;
836 
837 	pthread_mutex_unlock(&self->pt_lock);
838 
839 	return retval;
840 }
841 
842 
843 void
844 pthread_testcancel(void)
845 {
846 	pthread_t self;
847 
848 	self = pthread__self();
849 	if (self->pt_cancel)
850 		pthread__cancelled();
851 }
852 
853 
854 /*
855  * POSIX requires that certain functions return an error rather than
856  * invoking undefined behavior even when handed completely bogus
857  * pthread_t values, e.g. stack garbage or (pthread_t)666. This
858  * utility routine searches the list of threads for the pthread_t
859  * value without dereferencing it.
860  */
861 int
862 pthread__find(pthread_t id)
863 {
864 	pthread_t target;
865 
866 	pthread_rwlock_rdlock(&pthread__alltree_lock);
867 	/* LINTED */
868 	target = RB_FIND(__pthread__alltree, &pthread__alltree, id);
869 	pthread_rwlock_unlock(&pthread__alltree_lock);
870 
871 	if (target == NULL || target->pt_state == PT_STATE_DEAD)
872 		return ESRCH;
873 
874 	return 0;
875 }
876 
877 
878 void
879 pthread__testcancel(pthread_t self)
880 {
881 
882 	if (self->pt_cancel)
883 		pthread__cancelled();
884 }
885 
886 
887 void
888 pthread__cancelled(void)
889 {
890 	pthread_mutex_t *droplock;
891 	pthread_t self;
892 
893 	self = pthread__self();
894 	droplock = self->pt_droplock;
895 	self->pt_droplock = NULL;
896 
897 	if (droplock != NULL && pthread_mutex_held_np(droplock))
898 		pthread_mutex_unlock(droplock);
899 
900 	pthread_exit(PTHREAD_CANCELED);
901 }
902 
903 
904 void
905 pthread__cleanup_push(void (*cleanup)(void *), void *arg, void *store)
906 {
907 	pthread_t self;
908 	struct pt_clean_t *entry;
909 
910 	self = pthread__self();
911 	entry = store;
912 	entry->ptc_cleanup = cleanup;
913 	entry->ptc_arg = arg;
914 	PTQ_INSERT_HEAD(&self->pt_cleanup_stack, entry, ptc_next);
915 }
916 
917 
918 void
919 pthread__cleanup_pop(int ex, void *store)
920 {
921 	pthread_t self;
922 	struct pt_clean_t *entry;
923 
924 	self = pthread__self();
925 	entry = store;
926 
927 	PTQ_REMOVE(&self->pt_cleanup_stack, entry, ptc_next);
928 	if (ex)
929 		(*entry->ptc_cleanup)(entry->ptc_arg);
930 }
931 
932 
933 int *
934 pthread__errno(void)
935 {
936 	pthread_t self;
937 
938 	self = pthread__self();
939 
940 	return &(self->pt_errno);
941 }
942 
943 ssize_t	_sys_write(int, const void *, size_t);
944 
945 void
946 pthread__assertfunc(const char *file, int line, const char *function,
947 		    const char *expr)
948 {
949 	char buf[1024];
950 	int len;
951 
952 	/*
953 	 * snprintf should not acquire any locks, or we could
954 	 * end up deadlocked if the assert caller held locks.
955 	 */
956 	len = snprintf(buf, 1024,
957 	    "assertion \"%s\" failed: file \"%s\", line %d%s%s%s\n",
958 	    expr, file, line,
959 	    function ? ", function \"" : "",
960 	    function ? function : "",
961 	    function ? "\"" : "");
962 
963 	_sys_write(STDERR_FILENO, buf, (size_t)len);
964 	(void)kill(getpid(), SIGABRT);
965 
966 	_exit(1);
967 }
968 
969 
970 void
971 pthread__errorfunc(const char *file, int line, const char *function,
972 		   const char *msg)
973 {
974 	char buf[1024];
975 	size_t len;
976 
977 	if (pthread__diagassert == 0)
978 		return;
979 
980 	/*
981 	 * snprintf should not acquire any locks, or we could
982 	 * end up deadlocked if the assert caller held locks.
983 	 */
984 	len = snprintf(buf, 1024,
985 	    "%s: Error detected by libpthread: %s.\n"
986 	    "Detected by file \"%s\", line %d%s%s%s.\n"
987 	    "See pthread(3) for information.\n",
988 	    getprogname(), msg, file, line,
989 	    function ? ", function \"" : "",
990 	    function ? function : "",
991 	    function ? "\"" : "");
992 
993 	if (pthread__diagassert & DIAGASSERT_STDERR)
994 		_sys_write(STDERR_FILENO, buf, len);
995 
996 	if (pthread__diagassert & DIAGASSERT_SYSLOG)
997 		syslog(LOG_DEBUG | LOG_USER, "%s", buf);
998 
999 	if (pthread__diagassert & DIAGASSERT_ABORT) {
1000 		(void)kill(getpid(), SIGABRT);
1001 		_exit(1);
1002 	}
1003 }
1004 
1005 /*
1006  * Thread park/unpark operations.  The kernel operations are
1007  * modelled after a brief description from "Multithreading in
1008  * the Solaris Operating Environment":
1009  *
1010  * http://www.sun.com/software/whitepapers/solaris9/multithread.pdf
1011  */
1012 
1013 #define	OOPS(msg)			\
1014     pthread__errorfunc(__FILE__, __LINE__, __func__, msg)
1015 
1016 int
1017 pthread__park(pthread_t self, pthread_mutex_t *lock,
1018 	      pthread_queue_t *queue, const struct timespec *abstime,
1019 	      int cancelpt, const void *hint)
1020 {
1021 	int rv, error;
1022 	void *obj;
1023 
1024 	/*
1025 	 * For non-interlocked release of mutexes we need a store
1026 	 * barrier before incrementing pt_blocking away from zero.
1027 	 * This is provided by pthread_mutex_unlock().
1028 	 */
1029 	self->pt_willpark = 1;
1030 	pthread_mutex_unlock(lock);
1031 	self->pt_willpark = 0;
1032 	self->pt_blocking++;
1033 
1034 	/*
1035 	 * Wait until we are awoken by a pending unpark operation,
1036 	 * a signal, an unpark posted after we have gone asleep,
1037 	 * or an expired timeout.
1038 	 *
1039 	 * It is fine to test the value of pt_sleepobj without
1040 	 * holding any locks, because:
1041 	 *
1042 	 * o Only the blocking thread (this thread) ever sets them
1043 	 *   to a non-NULL value.
1044 	 *
1045 	 * o Other threads may set them NULL, but if they do so they
1046 	 *   must also make this thread return from _lwp_park.
1047 	 *
1048 	 * o _lwp_park, _lwp_unpark and _lwp_unpark_all are system
1049 	 *   calls and all make use of spinlocks in the kernel.  So
1050 	 *   these system calls act as full memory barriers, and will
1051 	 *   ensure that the calling CPU's store buffers are drained.
1052 	 *   In combination with the spinlock release before unpark,
1053 	 *   this means that modification of pt_sleepobj/onq by another
1054 	 *   thread will become globally visible before that thread
1055 	 *   schedules an unpark operation on this thread.
1056 	 *
1057 	 * Note: the test in the while() statement dodges the park op if
1058 	 * we have already been awoken, unless there is another thread to
1059 	 * awaken.  This saves a syscall - if we were already awakened,
1060 	 * the next call to _lwp_park() would need to return early in order
1061 	 * to eat the previous wakeup.
1062 	 */
1063 	rv = 0;
1064 	do {
1065 		/*
1066 		 * If we deferred unparking a thread, arrange to
1067 		 * have _lwp_park() restart it before blocking.
1068 		 */
1069 		error = _lwp_park(abstime, self->pt_unpark, hint, hint);
1070 		self->pt_unpark = 0;
1071 		if (error != 0) {
1072 			switch (rv = errno) {
1073 			case EINTR:
1074 			case EALREADY:
1075 				rv = 0;
1076 				break;
1077 			case ETIMEDOUT:
1078 				break;
1079 			default:
1080 				OOPS("_lwp_park failed");
1081 				break;
1082 			}
1083 		}
1084 		/* Check for cancellation. */
1085 		if (cancelpt && self->pt_cancel)
1086 			rv = EINTR;
1087 	} while (self->pt_sleepobj != NULL && rv == 0);
1088 
1089 	/*
1090 	 * If we have been awoken early but are still on the queue,
1091 	 * then remove ourself.  Again, it's safe to do the test
1092 	 * without holding any locks.
1093 	 */
1094 	if (__predict_false(self->pt_sleepobj != NULL)) {
1095 		pthread_mutex_lock(lock);
1096 		if ((obj = self->pt_sleepobj) != NULL) {
1097 			PTQ_REMOVE(queue, self, pt_sleep);
1098 			self->pt_sleepobj = NULL;
1099 			if (obj != NULL && self->pt_early != NULL)
1100 				(*self->pt_early)(obj);
1101 		}
1102 		pthread_mutex_unlock(lock);
1103 	}
1104 	self->pt_early = NULL;
1105 	self->pt_blocking--;
1106 	membar_sync();
1107 
1108 	return rv;
1109 }
1110 
1111 void
1112 pthread__unpark(pthread_queue_t *queue, pthread_t self,
1113 		pthread_mutex_t *interlock)
1114 {
1115 	pthread_t target;
1116 	u_int max;
1117 	size_t nwaiters;
1118 
1119 	max = pthread__unpark_max;
1120 	nwaiters = self->pt_nwaiters;
1121 	target = PTQ_FIRST(queue);
1122 	if (nwaiters == max) {
1123 		/* Overflow. */
1124 		(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1125 		    __UNVOLATILE(&interlock->ptm_waiters));
1126 		nwaiters = 0;
1127 	}
1128 	target->pt_sleepobj = NULL;
1129 	self->pt_waiters[nwaiters++] = target->pt_lid;
1130 	PTQ_REMOVE(queue, target, pt_sleep);
1131 	self->pt_nwaiters = nwaiters;
1132 	pthread__mutex_deferwake(self, interlock);
1133 }
1134 
1135 void
1136 pthread__unpark_all(pthread_queue_t *queue, pthread_t self,
1137 		    pthread_mutex_t *interlock)
1138 {
1139 	pthread_t target;
1140 	u_int max;
1141 	size_t nwaiters;
1142 
1143 	max = pthread__unpark_max;
1144 	nwaiters = self->pt_nwaiters;
1145 	PTQ_FOREACH(target, queue, pt_sleep) {
1146 		if (nwaiters == max) {
1147 			/* Overflow. */
1148 			(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1149 			    __UNVOLATILE(&interlock->ptm_waiters));
1150 			nwaiters = 0;
1151 		}
1152 		target->pt_sleepobj = NULL;
1153 		self->pt_waiters[nwaiters++] = target->pt_lid;
1154 	}
1155 	self->pt_nwaiters = nwaiters;
1156 	PTQ_INIT(queue);
1157 	pthread__mutex_deferwake(self, interlock);
1158 }
1159 
1160 #undef	OOPS
1161 
1162 /*
1163  * Allocate a stack for a thread, and set it up. It needs to be aligned, so
1164  * that a thread can find itself by its stack pointer.
1165  */
1166 static int
1167 pthread__stackalloc(pthread_t *newt)
1168 {
1169 	void *addr;
1170 
1171 	addr = mmap(NULL, pthread__stacksize, PROT_READ|PROT_WRITE,
1172 	    MAP_ANON|MAP_PRIVATE | MAP_ALIGNED(pthread__stacksize_lg),
1173 	    -1, (off_t)0);
1174 
1175 	if (addr == MAP_FAILED)
1176 		return ENOMEM;
1177 
1178 	pthread__assert(((intptr_t)addr & pthread__stackmask) == 0);
1179 
1180 	return pthread__stackid_setup(addr, pthread__stacksize, newt);
1181 }
1182 
1183 
1184 /*
1185  * Set up the slightly special stack for the "initial" thread, which
1186  * runs on the normal system stack, and thus gets slightly different
1187  * treatment.
1188  */
1189 static void
1190 pthread__initmain(pthread_t *newt)
1191 {
1192 	struct rlimit slimit;
1193 	size_t pagesize;
1194 	pthread_t t;
1195 	void *base;
1196 	size_t size;
1197 	int error, ret;
1198 	char *value;
1199 
1200 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1201 	pthread__stacksize = 0;
1202 	ret = getrlimit(RLIMIT_STACK, &slimit);
1203 	if (ret == -1)
1204 		err(1, "Couldn't get stack resource consumption limits");
1205 
1206 	value = pthread__getenv("PTHREAD_STACKSIZE");
1207 	if (value != NULL) {
1208 		pthread__stacksize = atoi(value) * 1024;
1209 		if (pthread__stacksize > slimit.rlim_cur)
1210 			pthread__stacksize = (size_t)slimit.rlim_cur;
1211 	}
1212 	if (pthread__stacksize == 0)
1213 		pthread__stacksize = (size_t)slimit.rlim_cur;
1214 	if (pthread__stacksize < 4 * pagesize)
1215 		errx(1, "Stacksize limit is too low, minimum %zd kbyte.",
1216 		    4 * pagesize / 1024);
1217 
1218 	pthread__stacksize_lg = -1;
1219 	while (pthread__stacksize) {
1220 		pthread__stacksize >>= 1;
1221 		pthread__stacksize_lg++;
1222 	}
1223 
1224 	pthread__stacksize = (1 << pthread__stacksize_lg);
1225 	pthread__stackmask = pthread__stacksize - 1;
1226 	pthread__threadmask = ~pthread__stackmask;
1227 
1228 	base = (void *)(pthread__sp() & pthread__threadmask);
1229 	size = pthread__stacksize;
1230 
1231 	error = pthread__stackid_setup(base, size, &t);
1232 	if (error) {
1233 		/* XXX */
1234 		errx(2, "failed to setup main thread: error=%d", error);
1235 	}
1236 
1237 	*newt = t;
1238 
1239 	/* Set up identity register. */
1240 	(void)_lwp_setprivate(t);
1241 }
1242 
1243 static int
1244 /*ARGSUSED*/
1245 pthread__stackid_setup(void *base, size_t size, pthread_t *tp)
1246 {
1247 	pthread_t t;
1248 	void *redaddr;
1249 	size_t pagesize;
1250 	int ret;
1251 
1252 	t = base;
1253 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1254 
1255 	/*
1256 	 * Put a pointer to the pthread in the bottom (but
1257          * redzone-protected section) of the stack.
1258 	 */
1259 	redaddr = STACK_SHRINK(STACK_MAX(base, size), pagesize);
1260 	t->pt_stack.ss_size = size - 2 * pagesize;
1261 #ifdef __MACHINE_STACK_GROWS_UP
1262 	t->pt_stack.ss_sp = (char *)(void *)base + pagesize;
1263 #else
1264 	t->pt_stack.ss_sp = (char *)(void *)base + 2 * pagesize;
1265 #endif
1266 
1267 	/* Protect the next-to-bottom stack page as a red zone. */
1268 	ret = mprotect(redaddr, pagesize, PROT_NONE);
1269 	if (ret == -1) {
1270 		return errno;
1271 	}
1272 	*tp = t;
1273 	return 0;
1274 }
1275 
1276 #ifndef lint
1277 static int
1278 pthread__cmp(struct __pthread_st *a, struct __pthread_st *b)
1279 {
1280 
1281 	if ((uintptr_t)a < (uintptr_t)b)
1282 		return (-1);
1283 	else if (a == b)
1284 		return 0;
1285 	else
1286 		return 1;
1287 }
1288 RB_GENERATE_STATIC(__pthread__alltree, __pthread_st, pt_alltree, pthread__cmp)
1289 #endif
1290 
1291 /* Because getenv() wants to use locks. */
1292 char *
1293 pthread__getenv(const char *name)
1294 {
1295 	extern char *__findenv(const char *, int *);
1296 	int off;
1297 
1298 	return __findenv(name, &off);
1299 }
1300 
1301 pthread_mutex_t *
1302 pthread__hashlock(volatile const void *p)
1303 {
1304 	uintptr_t v;
1305 
1306 	v = (uintptr_t)p;
1307 	return &hashlocks[((v >> 9) ^ (v >> 3)) & (NHASHLOCK - 1)].mutex;
1308 }
1309 
1310 int
1311 pthread__checkpri(int pri)
1312 {
1313 	static int havepri;
1314 	static long min, max;
1315 
1316 	if (!havepri) {
1317 		min = sysconf(_SC_SCHED_PRI_MIN);
1318 		max = sysconf(_SC_SCHED_PRI_MAX);
1319 		havepri = 1;
1320 	}
1321 	return (pri < min || pri > max) ? EINVAL : 0;
1322 }
1323