xref: /netbsd-src/lib/libpthread/pthread.c (revision b5677b36047b601b9addaaa494a58ceae82c2a6c)
1 /*	$NetBSD: pthread.c,v 1.109 2009/04/01 10:13:24 drochner 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 1.109 2009/04/01 10:13:24 drochner 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(pthread_t, void *(*)(void *), 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 	makecontext(&newthread->pt_uc, pthread__create_tramp, 3,
410 	    newthread, startfunc, arg);
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 		free(name);
419 		newthread->pt_state = PT_STATE_DEAD;
420 		pthread_mutex_lock(&pthread__deadqueue_lock);
421 		PTQ_INSERT_HEAD(&pthread__deadqueue, newthread, pt_deadq);
422 		pthread_mutex_unlock(&pthread__deadqueue_lock);
423 		return ret;
424 	}
425 
426 	if ((nattr.pta_flags & PT_FLAG_EXPLICIT_SCHED) != 0) {
427 		if (p != NULL) {
428 			(void)pthread_setschedparam(newthread, p->ptap_policy,
429 			    &p->ptap_sp);
430 		}
431 		if ((newthread->pt_flags & PT_FLAG_SUSPENDED) == 0) {
432 			(void)_lwp_continue(newthread->pt_lid);
433 		}
434 	}
435 
436 	*thread = newthread;
437 
438 	return 0;
439 }
440 
441 
442 static void
443 pthread__create_tramp(pthread_t self, void *(*start)(void *), void *arg)
444 {
445 	void *retval;
446 
447 	/*
448 	 * Set up identity register.
449 	 * XXX Race: could receive a signal before this.
450 	 */
451 	(void)_lwp_setprivate(self);
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.  Note
458 	 * that pt_lid may not be set by this point, but we don't
459 	 * care.
460 	 */
461 	(void)alloca(((unsigned)self->pt_lid & 7) << 8);
462 
463 	if (self->pt_name != NULL) {
464 		pthread_mutex_lock(&self->pt_lock);
465 		if (self->pt_name != NULL)
466 			(void)_lwp_setname(0, self->pt_name);
467 		pthread_mutex_unlock(&self->pt_lock);
468 	}
469 
470 	if (_lwp_ctl(LWPCTL_FEATURE_CURCPU, &self->pt_lwpctl)) {
471 		err(1, "_lwp_ctl");
472 	}
473 
474 	retval = (*start)(arg);
475 
476 	pthread_exit(retval);
477 
478 	/*NOTREACHED*/
479 	pthread__abort();
480 }
481 
482 int
483 pthread_suspend_np(pthread_t thread)
484 {
485 	pthread_t self;
486 
487 	self = pthread__self();
488 	if (self == thread) {
489 		return EDEADLK;
490 	}
491 	if (pthread__find(thread) != 0)
492 		return ESRCH;
493 	if (_lwp_suspend(thread->pt_lid) == 0)
494 		return 0;
495 	return errno;
496 }
497 
498 int
499 pthread_resume_np(pthread_t thread)
500 {
501 
502 	if (pthread__find(thread) != 0)
503 		return ESRCH;
504 	if (_lwp_continue(thread->pt_lid) == 0)
505 		return 0;
506 	return errno;
507 }
508 
509 void
510 pthread_exit(void *retval)
511 {
512 	pthread_t self;
513 	struct pt_clean_t *cleanup;
514 	char *name;
515 
516 	self = pthread__self();
517 
518 	/* Disable cancellability. */
519 	pthread_mutex_lock(&self->pt_lock);
520 	self->pt_flags |= PT_FLAG_CS_DISABLED;
521 	self->pt_cancel = 0;
522 
523 	/* Call any cancellation cleanup handlers */
524 	if (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
525 		pthread_mutex_unlock(&self->pt_lock);
526 		while (!PTQ_EMPTY(&self->pt_cleanup_stack)) {
527 			cleanup = PTQ_FIRST(&self->pt_cleanup_stack);
528 			PTQ_REMOVE(&self->pt_cleanup_stack, cleanup, ptc_next);
529 			(*cleanup->ptc_cleanup)(cleanup->ptc_arg);
530 		}
531 		pthread_mutex_lock(&self->pt_lock);
532 	}
533 
534 	/* Perform cleanup of thread-specific data */
535 	pthread__destroy_tsd(self);
536 
537 	/* Signal our exit. */
538 	self->pt_exitval = retval;
539 	if (self->pt_flags & PT_FLAG_DETACHED) {
540 		self->pt_state = PT_STATE_DEAD;
541 		name = self->pt_name;
542 		self->pt_name = NULL;
543 		pthread_mutex_unlock(&self->pt_lock);
544 		if (name != NULL)
545 			free(name);
546 		pthread_mutex_lock(&pthread__deadqueue_lock);
547 		PTQ_INSERT_TAIL(&pthread__deadqueue, self, pt_deadq);
548 		pthread_mutex_unlock(&pthread__deadqueue_lock);
549 		_lwp_exit();
550 	} else {
551 		self->pt_state = PT_STATE_ZOMBIE;
552 		pthread_cond_broadcast(&self->pt_joiners);
553 		pthread_mutex_unlock(&self->pt_lock);
554 		/* Note: name will be freed by the joiner. */
555 		_lwp_exit();
556 	}
557 
558 	/*NOTREACHED*/
559 	pthread__abort();
560 	exit(1);
561 }
562 
563 
564 int
565 pthread_join(pthread_t thread, void **valptr)
566 {
567 	pthread_t self;
568 	int error;
569 
570 	self = pthread__self();
571 
572 	if (pthread__find(thread) != 0)
573 		return ESRCH;
574 
575 	if (thread->pt_magic != PT_MAGIC)
576 		return EINVAL;
577 
578 	if (thread == self)
579 		return EDEADLK;
580 
581 	self->pt_droplock = &thread->pt_lock;
582 	pthread_mutex_lock(&thread->pt_lock);
583 	for (;;) {
584 		if (thread->pt_state == PT_STATE_ZOMBIE)
585 			break;
586 		if (thread->pt_state == PT_STATE_DEAD) {
587 			pthread_mutex_unlock(&thread->pt_lock);
588 			self->pt_droplock = NULL;
589 			return ESRCH;
590 		}
591 		if ((thread->pt_flags & PT_FLAG_DETACHED) != 0) {
592 			pthread_mutex_unlock(&thread->pt_lock);
593 			self->pt_droplock = NULL;
594 			return EINVAL;
595 		}
596 		error = pthread_cond_wait(&thread->pt_joiners,
597 		    &thread->pt_lock);
598 		if (error != 0) {
599 			pthread__errorfunc(__FILE__, __LINE__,
600 			    __func__, "unexpected return from cond_wait()");
601 		}
602 
603 	}
604 	pthread__testcancel(self);
605 	if (valptr != NULL)
606 		*valptr = thread->pt_exitval;
607 	/* pthread__reap() will drop the lock. */
608 	pthread__reap(thread);
609 	self->pt_droplock = NULL;
610 
611 	return 0;
612 }
613 
614 static void
615 pthread__reap(pthread_t thread)
616 {
617 	char *name;
618 
619 	name = thread->pt_name;
620 	thread->pt_name = NULL;
621 	thread->pt_state = PT_STATE_DEAD;
622 	pthread_mutex_unlock(&thread->pt_lock);
623 
624 	pthread_mutex_lock(&pthread__deadqueue_lock);
625 	PTQ_INSERT_HEAD(&pthread__deadqueue, thread, pt_deadq);
626 	pthread_mutex_unlock(&pthread__deadqueue_lock);
627 
628 	if (name != NULL)
629 		free(name);
630 }
631 
632 int
633 pthread_equal(pthread_t t1, pthread_t t2)
634 {
635 
636 	/* Nothing special here. */
637 	return (t1 == t2);
638 }
639 
640 
641 int
642 pthread_detach(pthread_t thread)
643 {
644 
645 	if (pthread__find(thread) != 0)
646 		return ESRCH;
647 
648 	if (thread->pt_magic != PT_MAGIC)
649 		return EINVAL;
650 
651 	pthread_mutex_lock(&thread->pt_lock);
652 	thread->pt_flags |= PT_FLAG_DETACHED;
653 	if (thread->pt_state == PT_STATE_ZOMBIE) {
654 		/* pthread__reap() will drop the lock. */
655 		pthread__reap(thread);
656 	} else {
657 		/*
658 		 * Not valid for threads to be waiting in
659 		 * pthread_join() (there are intractable
660 		 * sync issues from the application
661 		 * perspective), but give those threads
662 		 * a chance anyway.
663 		 */
664 		pthread_cond_broadcast(&thread->pt_joiners);
665 		pthread_mutex_unlock(&thread->pt_lock);
666 	}
667 
668 	return 0;
669 }
670 
671 
672 int
673 pthread_getname_np(pthread_t thread, char *name, size_t len)
674 {
675 
676 	if (pthread__find(thread) != 0)
677 		return ESRCH;
678 
679 	if (thread->pt_magic != PT_MAGIC)
680 		return EINVAL;
681 
682 	pthread_mutex_lock(&thread->pt_lock);
683 	if (thread->pt_name == NULL)
684 		name[0] = '\0';
685 	else
686 		strlcpy(name, thread->pt_name, len);
687 	pthread_mutex_unlock(&thread->pt_lock);
688 
689 	return 0;
690 }
691 
692 
693 int
694 pthread_setname_np(pthread_t thread, const char *name, void *arg)
695 {
696 	char *oldname, *cp, newname[PTHREAD_MAX_NAMELEN_NP];
697 	int namelen;
698 
699 	if (pthread__find(thread) != 0)
700 		return ESRCH;
701 
702 	if (thread->pt_magic != PT_MAGIC)
703 		return EINVAL;
704 
705 	namelen = snprintf(newname, sizeof(newname), name, arg);
706 	if (namelen >= PTHREAD_MAX_NAMELEN_NP)
707 		return EINVAL;
708 
709 	cp = strdup(newname);
710 	if (cp == NULL)
711 		return ENOMEM;
712 
713 	pthread_mutex_lock(&thread->pt_lock);
714 	oldname = thread->pt_name;
715 	thread->pt_name = cp;
716 	(void)_lwp_setname(thread->pt_lid, cp);
717 	pthread_mutex_unlock(&thread->pt_lock);
718 
719 	if (oldname != NULL)
720 		free(oldname);
721 
722 	return 0;
723 }
724 
725 
726 
727 /*
728  * XXX There should be a way for applications to use the efficent
729  *  inline version, but there are opacity/namespace issues.
730  */
731 pthread_t
732 pthread_self(void)
733 {
734 
735 	return pthread__self();
736 }
737 
738 
739 int
740 pthread_cancel(pthread_t thread)
741 {
742 
743 	if (pthread__find(thread) != 0)
744 		return ESRCH;
745 	pthread_mutex_lock(&thread->pt_lock);
746 	thread->pt_flags |= PT_FLAG_CS_PENDING;
747 	if ((thread->pt_flags & PT_FLAG_CS_DISABLED) == 0) {
748 		thread->pt_cancel = 1;
749 		pthread_mutex_unlock(&thread->pt_lock);
750 		_lwp_wakeup(thread->pt_lid);
751 	} else
752 		pthread_mutex_unlock(&thread->pt_lock);
753 
754 	return 0;
755 }
756 
757 
758 int
759 pthread_setcancelstate(int state, int *oldstate)
760 {
761 	pthread_t self;
762 	int retval;
763 
764 	self = pthread__self();
765 	retval = 0;
766 
767 	pthread_mutex_lock(&self->pt_lock);
768 
769 	if (oldstate != NULL) {
770 		if (self->pt_flags & PT_FLAG_CS_DISABLED)
771 			*oldstate = PTHREAD_CANCEL_DISABLE;
772 		else
773 			*oldstate = PTHREAD_CANCEL_ENABLE;
774 	}
775 
776 	if (state == PTHREAD_CANCEL_DISABLE) {
777 		self->pt_flags |= PT_FLAG_CS_DISABLED;
778 		if (self->pt_cancel) {
779 			self->pt_flags |= PT_FLAG_CS_PENDING;
780 			self->pt_cancel = 0;
781 		}
782 	} else if (state == PTHREAD_CANCEL_ENABLE) {
783 		self->pt_flags &= ~PT_FLAG_CS_DISABLED;
784 		/*
785 		 * If a cancellation was requested while cancellation
786 		 * was disabled, note that fact for future
787 		 * cancellation tests.
788 		 */
789 		if (self->pt_flags & PT_FLAG_CS_PENDING) {
790 			self->pt_cancel = 1;
791 			/* This is not a deferred cancellation point. */
792 			if (self->pt_flags & PT_FLAG_CS_ASYNC) {
793 				pthread_mutex_unlock(&self->pt_lock);
794 				pthread__cancelled();
795 			}
796 		}
797 	} else
798 		retval = EINVAL;
799 
800 	pthread_mutex_unlock(&self->pt_lock);
801 
802 	return retval;
803 }
804 
805 
806 int
807 pthread_setcanceltype(int type, int *oldtype)
808 {
809 	pthread_t self;
810 	int retval;
811 
812 	self = pthread__self();
813 	retval = 0;
814 
815 	pthread_mutex_lock(&self->pt_lock);
816 
817 	if (oldtype != NULL) {
818 		if (self->pt_flags & PT_FLAG_CS_ASYNC)
819 			*oldtype = PTHREAD_CANCEL_ASYNCHRONOUS;
820 		else
821 			*oldtype = PTHREAD_CANCEL_DEFERRED;
822 	}
823 
824 	if (type == PTHREAD_CANCEL_ASYNCHRONOUS) {
825 		self->pt_flags |= PT_FLAG_CS_ASYNC;
826 		if (self->pt_cancel) {
827 			pthread_mutex_unlock(&self->pt_lock);
828 			pthread__cancelled();
829 		}
830 	} else if (type == PTHREAD_CANCEL_DEFERRED)
831 		self->pt_flags &= ~PT_FLAG_CS_ASYNC;
832 	else
833 		retval = EINVAL;
834 
835 	pthread_mutex_unlock(&self->pt_lock);
836 
837 	return retval;
838 }
839 
840 
841 void
842 pthread_testcancel(void)
843 {
844 	pthread_t self;
845 
846 	self = pthread__self();
847 	if (self->pt_cancel)
848 		pthread__cancelled();
849 }
850 
851 
852 /*
853  * POSIX requires that certain functions return an error rather than
854  * invoking undefined behavior even when handed completely bogus
855  * pthread_t values, e.g. stack garbage or (pthread_t)666. This
856  * utility routine searches the list of threads for the pthread_t
857  * value without dereferencing it.
858  */
859 int
860 pthread__find(pthread_t id)
861 {
862 	pthread_t target;
863 
864 	pthread_rwlock_rdlock(&pthread__alltree_lock);
865 	/* LINTED */
866 	target = RB_FIND(__pthread__alltree, &pthread__alltree, id);
867 	pthread_rwlock_unlock(&pthread__alltree_lock);
868 
869 	if (target == NULL || target->pt_state == PT_STATE_DEAD)
870 		return ESRCH;
871 
872 	return 0;
873 }
874 
875 
876 void
877 pthread__testcancel(pthread_t self)
878 {
879 
880 	if (self->pt_cancel)
881 		pthread__cancelled();
882 }
883 
884 
885 void
886 pthread__cancelled(void)
887 {
888 	pthread_mutex_t *droplock;
889 	pthread_t self;
890 
891 	self = pthread__self();
892 	droplock = self->pt_droplock;
893 	self->pt_droplock = NULL;
894 
895 	if (droplock != NULL && pthread_mutex_held_np(droplock))
896 		pthread_mutex_unlock(droplock);
897 
898 	pthread_exit(PTHREAD_CANCELED);
899 }
900 
901 
902 void
903 pthread__cleanup_push(void (*cleanup)(void *), void *arg, void *store)
904 {
905 	pthread_t self;
906 	struct pt_clean_t *entry;
907 
908 	self = pthread__self();
909 	entry = store;
910 	entry->ptc_cleanup = cleanup;
911 	entry->ptc_arg = arg;
912 	PTQ_INSERT_HEAD(&self->pt_cleanup_stack, entry, ptc_next);
913 }
914 
915 
916 void
917 pthread__cleanup_pop(int ex, void *store)
918 {
919 	pthread_t self;
920 	struct pt_clean_t *entry;
921 
922 	self = pthread__self();
923 	entry = store;
924 
925 	PTQ_REMOVE(&self->pt_cleanup_stack, entry, ptc_next);
926 	if (ex)
927 		(*entry->ptc_cleanup)(entry->ptc_arg);
928 }
929 
930 
931 int *
932 pthread__errno(void)
933 {
934 	pthread_t self;
935 
936 	self = pthread__self();
937 
938 	return &(self->pt_errno);
939 }
940 
941 ssize_t	_sys_write(int, const void *, size_t);
942 
943 void
944 pthread__assertfunc(const char *file, int line, const char *function,
945 		    const char *expr)
946 {
947 	char buf[1024];
948 	int len;
949 
950 	/*
951 	 * snprintf should not acquire any locks, or we could
952 	 * end up deadlocked if the assert caller held locks.
953 	 */
954 	len = snprintf(buf, 1024,
955 	    "assertion \"%s\" failed: file \"%s\", line %d%s%s%s\n",
956 	    expr, file, line,
957 	    function ? ", function \"" : "",
958 	    function ? function : "",
959 	    function ? "\"" : "");
960 
961 	_sys_write(STDERR_FILENO, buf, (size_t)len);
962 	(void)kill(getpid(), SIGABRT);
963 
964 	_exit(1);
965 }
966 
967 
968 void
969 pthread__errorfunc(const char *file, int line, const char *function,
970 		   const char *msg)
971 {
972 	char buf[1024];
973 	size_t len;
974 
975 	if (pthread__diagassert == 0)
976 		return;
977 
978 	/*
979 	 * snprintf should not acquire any locks, or we could
980 	 * end up deadlocked if the assert caller held locks.
981 	 */
982 	len = snprintf(buf, 1024,
983 	    "%s: Error detected by libpthread: %s.\n"
984 	    "Detected by file \"%s\", line %d%s%s%s.\n"
985 	    "See pthread(3) for information.\n",
986 	    getprogname(), msg, file, line,
987 	    function ? ", function \"" : "",
988 	    function ? function : "",
989 	    function ? "\"" : "");
990 
991 	if (pthread__diagassert & DIAGASSERT_STDERR)
992 		_sys_write(STDERR_FILENO, buf, len);
993 
994 	if (pthread__diagassert & DIAGASSERT_SYSLOG)
995 		syslog(LOG_DEBUG | LOG_USER, "%s", buf);
996 
997 	if (pthread__diagassert & DIAGASSERT_ABORT) {
998 		(void)kill(getpid(), SIGABRT);
999 		_exit(1);
1000 	}
1001 }
1002 
1003 /*
1004  * Thread park/unpark operations.  The kernel operations are
1005  * modelled after a brief description from "Multithreading in
1006  * the Solaris Operating Environment":
1007  *
1008  * http://www.sun.com/software/whitepapers/solaris9/multithread.pdf
1009  */
1010 
1011 #define	OOPS(msg)			\
1012     pthread__errorfunc(__FILE__, __LINE__, __func__, msg)
1013 
1014 int
1015 pthread__park(pthread_t self, pthread_mutex_t *lock,
1016 	      pthread_queue_t *queue, const struct timespec *abstime,
1017 	      int cancelpt, const void *hint)
1018 {
1019 	int rv, error;
1020 	void *obj;
1021 
1022 	/*
1023 	 * For non-interlocked release of mutexes we need a store
1024 	 * barrier before incrementing pt_blocking away from zero.
1025 	 * This is provided by pthread_mutex_unlock().
1026 	 */
1027 	self->pt_willpark = 1;
1028 	pthread_mutex_unlock(lock);
1029 	self->pt_willpark = 0;
1030 	self->pt_blocking++;
1031 
1032 	/*
1033 	 * Wait until we are awoken by a pending unpark operation,
1034 	 * a signal, an unpark posted after we have gone asleep,
1035 	 * or an expired timeout.
1036 	 *
1037 	 * It is fine to test the value of pt_sleepobj without
1038 	 * holding any locks, because:
1039 	 *
1040 	 * o Only the blocking thread (this thread) ever sets them
1041 	 *   to a non-NULL value.
1042 	 *
1043 	 * o Other threads may set them NULL, but if they do so they
1044 	 *   must also make this thread return from _lwp_park.
1045 	 *
1046 	 * o _lwp_park, _lwp_unpark and _lwp_unpark_all are system
1047 	 *   calls and all make use of spinlocks in the kernel.  So
1048 	 *   these system calls act as full memory barriers, and will
1049 	 *   ensure that the calling CPU's store buffers are drained.
1050 	 *   In combination with the spinlock release before unpark,
1051 	 *   this means that modification of pt_sleepobj/onq by another
1052 	 *   thread will become globally visible before that thread
1053 	 *   schedules an unpark operation on this thread.
1054 	 *
1055 	 * Note: the test in the while() statement dodges the park op if
1056 	 * we have already been awoken, unless there is another thread to
1057 	 * awaken.  This saves a syscall - if we were already awakened,
1058 	 * the next call to _lwp_park() would need to return early in order
1059 	 * to eat the previous wakeup.
1060 	 */
1061 	rv = 0;
1062 	do {
1063 		/*
1064 		 * If we deferred unparking a thread, arrange to
1065 		 * have _lwp_park() restart it before blocking.
1066 		 */
1067 		error = _lwp_park(abstime, self->pt_unpark, hint, hint);
1068 		self->pt_unpark = 0;
1069 		if (error != 0) {
1070 			switch (rv = errno) {
1071 			case EINTR:
1072 			case EALREADY:
1073 				rv = 0;
1074 				break;
1075 			case ETIMEDOUT:
1076 				break;
1077 			default:
1078 				OOPS("_lwp_park failed");
1079 				break;
1080 			}
1081 		}
1082 		/* Check for cancellation. */
1083 		if (cancelpt && self->pt_cancel)
1084 			rv = EINTR;
1085 	} while (self->pt_sleepobj != NULL && rv == 0);
1086 
1087 	/*
1088 	 * If we have been awoken early but are still on the queue,
1089 	 * then remove ourself.  Again, it's safe to do the test
1090 	 * without holding any locks.
1091 	 */
1092 	if (__predict_false(self->pt_sleepobj != NULL)) {
1093 		pthread_mutex_lock(lock);
1094 		if ((obj = self->pt_sleepobj) != NULL) {
1095 			PTQ_REMOVE(queue, self, pt_sleep);
1096 			self->pt_sleepobj = NULL;
1097 			if (obj != NULL && self->pt_early != NULL)
1098 				(*self->pt_early)(obj);
1099 		}
1100 		pthread_mutex_unlock(lock);
1101 	}
1102 	self->pt_early = NULL;
1103 	self->pt_blocking--;
1104 	membar_sync();
1105 
1106 	return rv;
1107 }
1108 
1109 void
1110 pthread__unpark(pthread_queue_t *queue, pthread_t self,
1111 		pthread_mutex_t *interlock)
1112 {
1113 	pthread_t target;
1114 	u_int max;
1115 	size_t nwaiters;
1116 
1117 	max = pthread__unpark_max;
1118 	nwaiters = self->pt_nwaiters;
1119 	target = PTQ_FIRST(queue);
1120 	if (nwaiters == max) {
1121 		/* Overflow. */
1122 		(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1123 		    __UNVOLATILE(&interlock->ptm_waiters));
1124 		nwaiters = 0;
1125 	}
1126 	target->pt_sleepobj = NULL;
1127 	self->pt_waiters[nwaiters++] = target->pt_lid;
1128 	PTQ_REMOVE(queue, target, pt_sleep);
1129 	self->pt_nwaiters = nwaiters;
1130 	pthread__mutex_deferwake(self, interlock);
1131 }
1132 
1133 void
1134 pthread__unpark_all(pthread_queue_t *queue, pthread_t self,
1135 		    pthread_mutex_t *interlock)
1136 {
1137 	pthread_t target;
1138 	u_int max;
1139 	size_t nwaiters;
1140 
1141 	max = pthread__unpark_max;
1142 	nwaiters = self->pt_nwaiters;
1143 	PTQ_FOREACH(target, queue, pt_sleep) {
1144 		if (nwaiters == max) {
1145 			/* Overflow. */
1146 			(void)_lwp_unpark_all(self->pt_waiters, nwaiters,
1147 			    __UNVOLATILE(&interlock->ptm_waiters));
1148 			nwaiters = 0;
1149 		}
1150 		target->pt_sleepobj = NULL;
1151 		self->pt_waiters[nwaiters++] = target->pt_lid;
1152 	}
1153 	self->pt_nwaiters = nwaiters;
1154 	PTQ_INIT(queue);
1155 	pthread__mutex_deferwake(self, interlock);
1156 }
1157 
1158 #undef	OOPS
1159 
1160 /*
1161  * Allocate a stack for a thread, and set it up. It needs to be aligned, so
1162  * that a thread can find itself by its stack pointer.
1163  */
1164 static int
1165 pthread__stackalloc(pthread_t *newt)
1166 {
1167 	void *addr;
1168 
1169 	addr = mmap(NULL, pthread__stacksize, PROT_READ|PROT_WRITE,
1170 	    MAP_ANON|MAP_PRIVATE | MAP_ALIGNED(pthread__stacksize_lg),
1171 	    -1, (off_t)0);
1172 
1173 	if (addr == MAP_FAILED)
1174 		return ENOMEM;
1175 
1176 	pthread__assert(((intptr_t)addr & pthread__stackmask) == 0);
1177 
1178 	return pthread__stackid_setup(addr, pthread__stacksize, newt);
1179 }
1180 
1181 
1182 /*
1183  * Set up the slightly special stack for the "initial" thread, which
1184  * runs on the normal system stack, and thus gets slightly different
1185  * treatment.
1186  */
1187 static void
1188 pthread__initmain(pthread_t *newt)
1189 {
1190 	struct rlimit slimit;
1191 	size_t pagesize;
1192 	pthread_t t;
1193 	void *base;
1194 	size_t size;
1195 	int error, ret;
1196 	char *value;
1197 
1198 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1199 	pthread__stacksize = 0;
1200 	ret = getrlimit(RLIMIT_STACK, &slimit);
1201 	if (ret == -1)
1202 		err(1, "Couldn't get stack resource consumption limits");
1203 
1204 	value = pthread__getenv("PTHREAD_STACKSIZE");
1205 	if (value != NULL) {
1206 		pthread__stacksize = atoi(value) * 1024;
1207 		if (pthread__stacksize > slimit.rlim_cur)
1208 			pthread__stacksize = (size_t)slimit.rlim_cur;
1209 	}
1210 	if (pthread__stacksize == 0)
1211 		pthread__stacksize = (size_t)slimit.rlim_cur;
1212 	if (pthread__stacksize < 4 * pagesize)
1213 		errx(1, "Stacksize limit is too low, minimum %zd kbyte.",
1214 		    4 * pagesize / 1024);
1215 
1216 	pthread__stacksize_lg = -1;
1217 	while (pthread__stacksize) {
1218 		pthread__stacksize >>= 1;
1219 		pthread__stacksize_lg++;
1220 	}
1221 
1222 	pthread__stacksize = (1 << pthread__stacksize_lg);
1223 	pthread__stackmask = pthread__stacksize - 1;
1224 	pthread__threadmask = ~pthread__stackmask;
1225 
1226 	base = (void *)(pthread__sp() & pthread__threadmask);
1227 	size = pthread__stacksize;
1228 
1229 	error = pthread__stackid_setup(base, size, &t);
1230 	if (error) {
1231 		/* XXX */
1232 		errx(2, "failed to setup main thread: error=%d", error);
1233 	}
1234 
1235 	*newt = t;
1236 
1237 	/* Set up identity register. */
1238 	(void)_lwp_setprivate(t);
1239 }
1240 
1241 static int
1242 /*ARGSUSED*/
1243 pthread__stackid_setup(void *base, size_t size, pthread_t *tp)
1244 {
1245 	pthread_t t;
1246 	void *redaddr;
1247 	size_t pagesize;
1248 	int ret;
1249 
1250 	t = base;
1251 	pagesize = (size_t)sysconf(_SC_PAGESIZE);
1252 
1253 	/*
1254 	 * Put a pointer to the pthread in the bottom (but
1255          * redzone-protected section) of the stack.
1256 	 */
1257 	redaddr = STACK_SHRINK(STACK_MAX(base, size), pagesize);
1258 	t->pt_stack.ss_size = size - 2 * pagesize;
1259 #ifdef __MACHINE_STACK_GROWS_UP
1260 	t->pt_stack.ss_sp = (char *)(void *)base + pagesize;
1261 #else
1262 	t->pt_stack.ss_sp = (char *)(void *)base + 2 * pagesize;
1263 #endif
1264 
1265 	/* Protect the next-to-bottom stack page as a red zone. */
1266 	ret = mprotect(redaddr, pagesize, PROT_NONE);
1267 	if (ret == -1) {
1268 		return errno;
1269 	}
1270 	*tp = t;
1271 	return 0;
1272 }
1273 
1274 #ifndef lint
1275 static int
1276 pthread__cmp(struct __pthread_st *a, struct __pthread_st *b)
1277 {
1278 
1279 	if ((uintptr_t)a < (uintptr_t)b)
1280 		return (-1);
1281 	else if (a == b)
1282 		return 0;
1283 	else
1284 		return 1;
1285 }
1286 RB_GENERATE_STATIC(__pthread__alltree, __pthread_st, pt_alltree, pthread__cmp)
1287 #endif
1288 
1289 /* Because getenv() wants to use locks. */
1290 char *
1291 pthread__getenv(const char *name)
1292 {
1293 	extern char *__findenv(const char *, int *);
1294 	int off;
1295 
1296 	return __findenv(name, &off);
1297 }
1298 
1299 pthread_mutex_t *
1300 pthread__hashlock(volatile const void *p)
1301 {
1302 	uintptr_t v;
1303 
1304 	v = (uintptr_t)p;
1305 	return &hashlocks[((v >> 9) ^ (v >> 3)) & (NHASHLOCK - 1)].mutex;
1306 }
1307 
1308 int
1309 pthread__checkpri(int pri)
1310 {
1311 	static int havepri;
1312 	static long min, max;
1313 
1314 	if (!havepri) {
1315 		min = sysconf(_SC_SCHED_PRI_MIN);
1316 		max = sysconf(_SC_SCHED_PRI_MAX);
1317 		havepri = 1;
1318 	}
1319 	return (pri < min || pri > max) ? EINVAL : 0;
1320 }
1321