xref: /freebsd-src/lib/libthr/thread/thr_stack.c (revision 31f688a26d82ce00d1ec7ca9ed17b9914bf5176f)
15e53a4f9SPedro F. Giffuni /*-
24d846d26SWarner Losh  * SPDX-License-Identifier: BSD-2-Clause
35e53a4f9SPedro F. Giffuni  *
4bb535300SJeff Roberson  * Copyright (c) 2001 Daniel Eischen <deischen@freebsd.org>
5bb535300SJeff Roberson  * Copyright (c) 2000-2001 Jason Evans <jasone@freebsd.org>
6bb535300SJeff Roberson  * All rights reserved.
7bb535300SJeff Roberson  *
8bb535300SJeff Roberson  * Redistribution and use in source and binary forms, with or without
9bb535300SJeff Roberson  * modification, are permitted provided that the following conditions
10bb535300SJeff Roberson  * are met:
11bb535300SJeff Roberson  * 1. Redistributions of source code must retain the above copyright
12bb535300SJeff Roberson  *    notice, this list of conditions and the following disclaimer.
13bb535300SJeff Roberson  * 2. Redistributions in binary form must reproduce the above copyright
14bb535300SJeff Roberson  *    notice, this list of conditions and the following disclaimer in the
15bb535300SJeff Roberson  *    documentation and/or other materials provided with the distribution.
16bb535300SJeff Roberson  *
17bb535300SJeff Roberson  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
18bb535300SJeff Roberson  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19bb535300SJeff Roberson  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20bb535300SJeff Roberson  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
21bb535300SJeff Roberson  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22bb535300SJeff Roberson  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23bb535300SJeff Roberson  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24bb535300SJeff Roberson  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25bb535300SJeff Roberson  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26bb535300SJeff Roberson  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27bb535300SJeff Roberson  * SUCH DAMAGE.
28bb535300SJeff Roberson  */
29a091d823SDavid Xu 
30ebf7a015SKonstantin Belousov #include <sys/param.h>
31e2879eceSKonstantin Belousov #include <sys/auxv.h>
32bb535300SJeff Roberson #include <sys/mman.h>
33bb535300SJeff Roberson #include <sys/queue.h>
34fad128dbSKonstantin Belousov #include <sys/resource.h>
35fad128dbSKonstantin Belousov #include <sys/sysctl.h>
36bb535300SJeff Roberson #include <stdlib.h>
37bb535300SJeff Roberson #include <pthread.h>
38da2fcff7SKonstantin Belousov #include <link.h>
39a091d823SDavid Xu 
40bb535300SJeff Roberson #include "thr_private.h"
41bb535300SJeff Roberson 
42bb535300SJeff Roberson /* Spare thread stack. */
43bb535300SJeff Roberson struct stack {
44bb535300SJeff Roberson 	LIST_ENTRY(stack)	qe;		/* Stack queue linkage. */
45bb535300SJeff Roberson 	size_t			stacksize;	/* Stack size (rounded up). */
46bb535300SJeff Roberson 	size_t			guardsize;	/* Guard size. */
47bb535300SJeff Roberson 	void			*stackaddr;	/* Stack address. */
48bb535300SJeff Roberson };
49bb535300SJeff Roberson 
50bb535300SJeff Roberson /*
51a091d823SDavid Xu  * Default sized (stack and guard) spare stack queue.  Stacks are cached
52a091d823SDavid Xu  * to avoid additional complexity managing mmap()ed stack regions.  Spare
53a091d823SDavid Xu  * stacks are used in LIFO order to increase cache locality.
54bb535300SJeff Roberson  */
55a091d823SDavid Xu static LIST_HEAD(, stack)	dstackq = LIST_HEAD_INITIALIZER(dstackq);
56bb535300SJeff Roberson 
57bb535300SJeff Roberson /*
58bb535300SJeff Roberson  * Miscellaneous sized (non-default stack and/or guard) spare stack queue.
59a091d823SDavid Xu  * Stacks are cached to avoid additional complexity managing mmap()ed
60a091d823SDavid Xu  * stack regions.  This list is unordered, since ordering on both stack
61a091d823SDavid Xu  * size and guard size would be more trouble than it's worth.  Stacks are
62a091d823SDavid Xu  * allocated from this cache on a first size match basis.
63bb535300SJeff Roberson  */
64a091d823SDavid Xu static LIST_HEAD(, stack)	mstackq = LIST_HEAD_INITIALIZER(mstackq);
65bb535300SJeff Roberson 
66bb535300SJeff Roberson /**
67a091d823SDavid Xu  * Base address of the last stack allocated (including its red zone, if
68a091d823SDavid Xu  * there is one).  Stacks are allocated contiguously, starting beyond the
69a091d823SDavid Xu  * top of the main stack.  When a new stack is created, a red zone is
70a091d823SDavid Xu  * typically created (actually, the red zone is mapped with PROT_NONE) above
71a091d823SDavid Xu  * the top of the stack, such that the stack will not be able to grow all
72a091d823SDavid Xu  * the way to the bottom of the next stack.  This isn't fool-proof.  It is
73a091d823SDavid Xu  * possible for a stack to grow by a large amount, such that it grows into
74a091d823SDavid Xu  * the next stack, and as long as the memory within the red zone is never
75a091d823SDavid Xu  * accessed, nothing will prevent one thread stack from trouncing all over
76a091d823SDavid Xu  * the next.
77bb535300SJeff Roberson  *
78bb535300SJeff Roberson  * low memory
79bb535300SJeff Roberson  *     . . . . . . . . . . . . . . . . . .
80bb535300SJeff Roberson  *    |                                   |
81bb535300SJeff Roberson  *    |             stack 3               | start of 3rd thread stack
82bb535300SJeff Roberson  *    +-----------------------------------+
83bb535300SJeff Roberson  *    |                                   |
84bb535300SJeff Roberson  *    |       Red Zone (guard page)       | red zone for 2nd thread
85bb535300SJeff Roberson  *    |                                   |
86bb535300SJeff Roberson  *    +-----------------------------------+
87a091d823SDavid Xu  *    |  stack 2 - _thr_stack_default     | top of 2nd thread stack
88bb535300SJeff Roberson  *    |                                   |
89bb535300SJeff Roberson  *    |                                   |
90bb535300SJeff Roberson  *    |                                   |
91bb535300SJeff Roberson  *    |                                   |
92bb535300SJeff Roberson  *    |             stack 2               |
93bb535300SJeff Roberson  *    +-----------------------------------+ <-- start of 2nd thread stack
94bb535300SJeff Roberson  *    |                                   |
95bb535300SJeff Roberson  *    |       Red Zone                    | red zone for 1st thread
96bb535300SJeff Roberson  *    |                                   |
97bb535300SJeff Roberson  *    +-----------------------------------+
98a091d823SDavid Xu  *    |  stack 1 - _thr_stack_default     | top of 1st thread stack
99bb535300SJeff Roberson  *    |                                   |
100bb535300SJeff Roberson  *    |                                   |
101bb535300SJeff Roberson  *    |                                   |
102bb535300SJeff Roberson  *    |                                   |
103bb535300SJeff Roberson  *    |             stack 1               |
104bb535300SJeff Roberson  *    +-----------------------------------+ <-- start of 1st thread stack
105bb535300SJeff Roberson  *    |                                   |   (initial value of last_stack)
106bb535300SJeff Roberson  *    |       Red Zone                    |
107bb535300SJeff Roberson  *    |                                   | red zone for main thread
108bb535300SJeff Roberson  *    +-----------------------------------+
109a091d823SDavid Xu  *    | USRSTACK - _thr_stack_initial     | top of main thread stack
110bb535300SJeff Roberson  *    |                                   | ^
111bb535300SJeff Roberson  *    |                                   | |
112bb535300SJeff Roberson  *    |                                   | |
113bb535300SJeff Roberson  *    |                                   | | stack growth
114bb535300SJeff Roberson  *    |                                   |
115bb535300SJeff Roberson  *    +-----------------------------------+ <-- start of main thread stack
116bb535300SJeff Roberson  *                                              (USRSTACK)
117bb535300SJeff Roberson  * high memory
118bb535300SJeff Roberson  *
119bb535300SJeff Roberson  */
12037a6356bSDavid Xu static char *last_stack = NULL;
121bb535300SJeff Roberson 
122bb535300SJeff Roberson /*
123a091d823SDavid Xu  * Round size up to the nearest multiple of
124a091d823SDavid Xu  * _thr_page_size.
125bb535300SJeff Roberson  */
126a091d823SDavid Xu static inline size_t
127a091d823SDavid Xu round_up(size_t size)
128a091d823SDavid Xu {
129*31f688a2SKonstantin Belousov 	return (roundup2(size, _thr_page_size));
130a091d823SDavid Xu }
131bb535300SJeff Roberson 
132da2fcff7SKonstantin Belousov void
133da2fcff7SKonstantin Belousov _thr_stack_fix_protection(struct pthread *thrd)
134da2fcff7SKonstantin Belousov {
135da2fcff7SKonstantin Belousov 
136da2fcff7SKonstantin Belousov 	mprotect((char *)thrd->attr.stackaddr_attr +
137da2fcff7SKonstantin Belousov 	    round_up(thrd->attr.guardsize_attr),
138da2fcff7SKonstantin Belousov 	    round_up(thrd->attr.stacksize_attr),
139da2fcff7SKonstantin Belousov 	    _rtld_get_stack_prot());
140da2fcff7SKonstantin Belousov }
141da2fcff7SKonstantin Belousov 
142fad128dbSKonstantin Belousov static void
143fad128dbSKonstantin Belousov singlethread_map_stacks_exec(void)
144fad128dbSKonstantin Belousov {
145e03c7f50SKonstantin Belousov 	char *usrstack;
146e03c7f50SKonstantin Belousov 	size_t stacksz;
147fad128dbSKonstantin Belousov 
148e03c7f50SKonstantin Belousov 	if (!__thr_get_main_stack_base(&usrstack) ||
149e03c7f50SKonstantin Belousov 	    !__thr_get_main_stack_lim(&stacksz))
150fad128dbSKonstantin Belousov 		return;
151e03c7f50SKonstantin Belousov 	mprotect(usrstack - stacksz, stacksz, _rtld_get_stack_prot());
152fad128dbSKonstantin Belousov }
153fad128dbSKonstantin Belousov 
154da2fcff7SKonstantin Belousov void
155bd43f069SKonstantin Belousov __thr_map_stacks_exec(void)
156da2fcff7SKonstantin Belousov {
157da2fcff7SKonstantin Belousov 	struct pthread *curthread, *thrd;
158da2fcff7SKonstantin Belousov 	struct stack *st;
159da2fcff7SKonstantin Belousov 
160fad128dbSKonstantin Belousov 	if (!_thr_is_inited()) {
161fad128dbSKonstantin Belousov 		singlethread_map_stacks_exec();
162fad128dbSKonstantin Belousov 		return;
163fad128dbSKonstantin Belousov 	}
164da2fcff7SKonstantin Belousov 	curthread = _get_curthread();
165da2fcff7SKonstantin Belousov 	THREAD_LIST_RDLOCK(curthread);
166da2fcff7SKonstantin Belousov 	LIST_FOREACH(st, &mstackq, qe)
167da2fcff7SKonstantin Belousov 		mprotect((char *)st->stackaddr + st->guardsize, st->stacksize,
168da2fcff7SKonstantin Belousov 		    _rtld_get_stack_prot());
169da2fcff7SKonstantin Belousov 	LIST_FOREACH(st, &dstackq, qe)
170da2fcff7SKonstantin Belousov 		mprotect((char *)st->stackaddr + st->guardsize, st->stacksize,
171da2fcff7SKonstantin Belousov 		    _rtld_get_stack_prot());
172da2fcff7SKonstantin Belousov 	TAILQ_FOREACH(thrd, &_thread_gc_list, gcle)
173da2fcff7SKonstantin Belousov 		_thr_stack_fix_protection(thrd);
174da2fcff7SKonstantin Belousov 	TAILQ_FOREACH(thrd, &_thread_list, tle)
175da2fcff7SKonstantin Belousov 		_thr_stack_fix_protection(thrd);
176da2fcff7SKonstantin Belousov 	THREAD_LIST_UNLOCK(curthread);
177da2fcff7SKonstantin Belousov }
178da2fcff7SKonstantin Belousov 
179a091d823SDavid Xu int
180a091d823SDavid Xu _thr_stack_alloc(struct pthread_attr *attr)
181a091d823SDavid Xu {
182a091d823SDavid Xu 	struct pthread *curthread = _get_curthread();
183a091d823SDavid Xu 	struct stack *spare_stack;
184a091d823SDavid Xu 	size_t stacksize;
185a091d823SDavid Xu 	size_t guardsize;
186a091d823SDavid Xu 	char *stackaddr;
187a091d823SDavid Xu 
188a091d823SDavid Xu 	/*
189a091d823SDavid Xu 	 * Round up stack size to nearest multiple of _thr_page_size so
190a091d823SDavid Xu 	 * that mmap() * will work.  If the stack size is not an even
191a091d823SDavid Xu 	 * multiple, we end up initializing things such that there is
192a091d823SDavid Xu 	 * unused space above the beginning of the stack, so the stack
193a091d823SDavid Xu 	 * sits snugly against its guard.
194a091d823SDavid Xu 	 */
195a091d823SDavid Xu 	stacksize = round_up(attr->stacksize_attr);
196a091d823SDavid Xu 	guardsize = round_up(attr->guardsize_attr);
197a091d823SDavid Xu 
198a091d823SDavid Xu 	attr->stackaddr_attr = NULL;
199a091d823SDavid Xu 	attr->flags &= ~THR_STACK_USER;
200a091d823SDavid Xu 
201a091d823SDavid Xu 	/*
202a091d823SDavid Xu 	 * Use the garbage collector lock for synchronization of the
203f75b1ff6SMark Johnston 	 * spare stack lists and allocations from usrstack.
204a091d823SDavid Xu 	 */
205a9b764e2SDavid Xu 	THREAD_LIST_WRLOCK(curthread);
206bb535300SJeff Roberson 	/*
207bb535300SJeff Roberson 	 * If the stack and guard sizes are default, try to allocate a stack
208bb535300SJeff Roberson 	 * from the default-size stack cache:
209bb535300SJeff Roberson 	 */
210a091d823SDavid Xu 	if ((stacksize == THR_STACK_DEFAULT) &&
211a091d823SDavid Xu 	    (guardsize == _thr_guard_default)) {
212a091d823SDavid Xu 		if ((spare_stack = LIST_FIRST(&dstackq)) != NULL) {
213bb535300SJeff Roberson 			/* Use the spare stack. */
214bb535300SJeff Roberson 			LIST_REMOVE(spare_stack, qe);
215a091d823SDavid Xu 			attr->stackaddr_attr = spare_stack->stackaddr;
216bb535300SJeff Roberson 		}
217bb535300SJeff Roberson 	}
218bb535300SJeff Roberson 	/*
219bb535300SJeff Roberson 	 * The user specified a non-default stack and/or guard size, so try to
220bb535300SJeff Roberson 	 * allocate a stack from the non-default size stack cache, using the
221bb535300SJeff Roberson 	 * rounded up stack size (stack_size) in the search:
222bb535300SJeff Roberson 	 */
223bb535300SJeff Roberson 	else {
224a091d823SDavid Xu 		LIST_FOREACH(spare_stack, &mstackq, qe) {
225a091d823SDavid Xu 			if (spare_stack->stacksize == stacksize &&
226bb535300SJeff Roberson 			    spare_stack->guardsize == guardsize) {
227bb535300SJeff Roberson 				LIST_REMOVE(spare_stack, qe);
228a091d823SDavid Xu 				attr->stackaddr_attr = spare_stack->stackaddr;
229bb535300SJeff Roberson 				break;
230bb535300SJeff Roberson 			}
231bb535300SJeff Roberson 		}
232bb535300SJeff Roberson 	}
233a091d823SDavid Xu 	if (attr->stackaddr_attr != NULL) {
234a091d823SDavid Xu 		/* A cached stack was found.  Release the lock. */
235a091d823SDavid Xu 		THREAD_LIST_UNLOCK(curthread);
236a091d823SDavid Xu 	}
237a091d823SDavid Xu 	else {
2386c8ce3bfSKonstantin Belousov 		/*
239f75b1ff6SMark Johnston 		 * Allocate a stack from or below usrstack, depending
2406c8ce3bfSKonstantin Belousov 		 * on the LIBPTHREAD_BIGSTACK_MAIN env variable.
2416c8ce3bfSKonstantin Belousov 		 */
242bb535300SJeff Roberson 		if (last_stack == NULL)
243f75b1ff6SMark Johnston 			last_stack = _usrstack - _thr_stack_initial -
244a091d823SDavid Xu 			    _thr_guard_default;
245bb535300SJeff Roberson 
246bb535300SJeff Roberson 		/* Allocate a new stack. */
247a091d823SDavid Xu 		stackaddr = last_stack - stacksize - guardsize;
248bb535300SJeff Roberson 
249bb535300SJeff Roberson 		/*
250a091d823SDavid Xu 		 * Even if stack allocation fails, we don't want to try to
251a091d823SDavid Xu 		 * use this location again, so unconditionally decrement
252bb535300SJeff Roberson 		 * last_stack.  Under normal operating conditions, the most
253a091d823SDavid Xu 		 * likely reason for an mmap() error is a stack overflow of
254a091d823SDavid Xu 		 * the adjacent thread stack.
255bb535300SJeff Roberson 		 */
256a091d823SDavid Xu 		last_stack -= (stacksize + guardsize);
257bb535300SJeff Roberson 
258a091d823SDavid Xu 		/* Release the lock before mmap'ing it. */
259a091d823SDavid Xu 		THREAD_LIST_UNLOCK(curthread);
260a091d823SDavid Xu 
261a091d823SDavid Xu 		/* Map the stack and guard page together, and split guard
262a091d823SDavid Xu 		   page from allocated space: */
263a091d823SDavid Xu 		if ((stackaddr = mmap(stackaddr, stacksize + guardsize,
264da2fcff7SKonstantin Belousov 		     _rtld_get_stack_prot(), MAP_STACK,
265a091d823SDavid Xu 		     -1, 0)) != MAP_FAILED &&
266a091d823SDavid Xu 		    (guardsize == 0 ||
267a091d823SDavid Xu 		     mprotect(stackaddr, guardsize, PROT_NONE) == 0)) {
268a091d823SDavid Xu 			stackaddr += guardsize;
269a091d823SDavid Xu 		} else {
270a091d823SDavid Xu 			if (stackaddr != MAP_FAILED)
271a091d823SDavid Xu 				munmap(stackaddr, stacksize + guardsize);
272a091d823SDavid Xu 			stackaddr = NULL;
273a091d823SDavid Xu 		}
274a091d823SDavid Xu 		attr->stackaddr_attr = stackaddr;
275a091d823SDavid Xu 	}
276a091d823SDavid Xu 	if (attr->stackaddr_attr != NULL)
277a091d823SDavid Xu 		return (0);
278a091d823SDavid Xu 	else
279a091d823SDavid Xu 		return (-1);
280bb535300SJeff Roberson }
281bb535300SJeff Roberson 
282a091d823SDavid Xu /* This function must be called with _thread_list_lock held. */
283bb535300SJeff Roberson void
284a091d823SDavid Xu _thr_stack_free(struct pthread_attr *attr)
285bb535300SJeff Roberson {
286bb535300SJeff Roberson 	struct stack *spare_stack;
287bb535300SJeff Roberson 
288a091d823SDavid Xu 	if ((attr != NULL) && ((attr->flags & THR_STACK_USER) == 0)
289a091d823SDavid Xu 	    && (attr->stackaddr_attr != NULL)) {
29037a6356bSDavid Xu 		spare_stack = (struct stack *)
29137a6356bSDavid Xu 			((char *)attr->stackaddr_attr +
29237a6356bSDavid Xu 			attr->stacksize_attr - sizeof(struct stack));
293a091d823SDavid Xu 		spare_stack->stacksize = round_up(attr->stacksize_attr);
294a091d823SDavid Xu 		spare_stack->guardsize = round_up(attr->guardsize_attr);
295a091d823SDavid Xu 		spare_stack->stackaddr = attr->stackaddr_attr;
296bb535300SJeff Roberson 
297a091d823SDavid Xu 		if (spare_stack->stacksize == THR_STACK_DEFAULT &&
298a091d823SDavid Xu 		    spare_stack->guardsize == _thr_guard_default) {
299bb535300SJeff Roberson 			/* Default stack/guard size. */
300a091d823SDavid Xu 			LIST_INSERT_HEAD(&dstackq, spare_stack, qe);
301bb535300SJeff Roberson 		} else {
302bb535300SJeff Roberson 			/* Non-default stack/guard size. */
303a091d823SDavid Xu 			LIST_INSERT_HEAD(&mstackq, spare_stack, qe);
304a091d823SDavid Xu 		}
305a091d823SDavid Xu 		attr->stackaddr_attr = NULL;
306bb535300SJeff Roberson 	}
307bb535300SJeff Roberson }
308