1 /* $NetBSD: pthread_atfork.c,v 1.27 2025/04/09 22:10:59 kre Exp $ */
2
3 /*-
4 * Copyright (c) 2002 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.
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 #if defined(LIBC_SCCS) && !defined(lint)
34 __RCSID("$NetBSD: pthread_atfork.c,v 1.27 2025/04/09 22:10:59 kre Exp $");
35 #endif /* LIBC_SCCS and not lint */
36
37 #include "namespace.h"
38
39 #include <errno.h>
40 #include <stdlib.h>
41 #include <unistd.h>
42
43 #include <sys/mman.h>
44 #include <sys/param.h>
45 #include <sys/queue.h>
46 #include <sys/sysctl.h>
47
48 #include "extern.h"
49 #include "reentrant.h"
50
51 #ifdef __weak_alias
__weak_alias(pthread_atfork,_pthread_atfork)52 __weak_alias(pthread_atfork, _pthread_atfork)
53 __weak_alias(fork, _fork)
54 #endif /* __weak_alias */
55
56 pid_t
57 __locked_fork(int *my_errno)
58 {
59 return __fork();
60 }
61
62 struct atfork_callback {
63 SIMPLEQ_ENTRY(atfork_callback) next;
64 void (*fn)(void);
65 };
66
67 struct atfork_cb_header {
68 uint16_t entries;
69 uint16_t used;
70 };
71
72 struct atfork_cb_block {
73 union {
74 struct atfork_callback block;
75 struct atfork_cb_header hdr;
76 } u;
77 };
78
79 #define cb_blocks(bp) (&(bp)->u.block)
80 #define cb_ents(bp) (bp)->u.hdr.entries
81 #define cb_used(bp) (bp)->u.hdr.used
82
83 /*
84 * We need to keep a cache for of at least 6, one for prepare, one for parent,
85 * one for child x 2 bexause of the two uses in the libpthread (pthread_init,
86 * pthread_tsd_init) constructors, where it is too early to call malloc(3).
87 * This does not guarantee that we will have enough, because other libraries
88 * can also call pthread_atfork() from their own constructors, so this is not
89 * a complete solution and will need to be fixed properly. For now a keep
90 * space for 16 since it is just 256 bytes.
91 */
92 static struct atfork_callback atfork_builtin[16];
93 static struct atfork_cb_block *atfork_storage = NULL;
94 static int hw_pagesize = 0;
95
96 static const int hw_pagesize_sysctl[2] = { CTL_HW, HW_PAGESIZE };
97
98 /*
99 * Hypothetically, we could protect the queues with a rwlock which is
100 * write-locked by pthread_atfork() and read-locked by fork(), but
101 * since the intended use of the functions is obtaining locks to hold
102 * across the fork, forking is going to be serialized anyway.
103 */
104 #ifdef _REENTRANT
105 static mutex_t atfork_lock = MUTEX_INITIALIZER;
106 #endif
107 SIMPLEQ_HEAD(atfork_callback_q, atfork_callback);
108
109 static struct atfork_callback_q prepareq = SIMPLEQ_HEAD_INITIALIZER(prepareq);
110 static struct atfork_callback_q parentq = SIMPLEQ_HEAD_INITIALIZER(parentq);
111 static struct atfork_callback_q childq = SIMPLEQ_HEAD_INITIALIZER(childq);
112
113 /*
114 * Nb: nothing allocated by this allocator is ever freed.
115 * (there is no API to free anything, and no need for one)
116 *
117 * The code relies upon this.
118 */
119 static struct atfork_callback *
af_alloc(unsigned int blocks)120 af_alloc(unsigned int blocks)
121 {
122 struct atfork_callback *result;
123
124 if (__predict_false(blocks == 0))
125 return NULL;
126
127 if (__predict_true(atfork_storage == NULL)) {
128 for (size_t i = 0; i < __arraycount(atfork_builtin); i++) {
129 if (atfork_builtin[i].fn == NULL) {
130 if (i + blocks <= __arraycount(atfork_builtin))
131 return &atfork_builtin[i];
132 else
133 break;
134 }
135 }
136 }
137
138 if (__predict_false(atfork_storage == NULL ||
139 cb_used(atfork_storage) + blocks > cb_ents(atfork_storage))) {
140 if (__predict_false(hw_pagesize == 0)) {
141 size_t len = sizeof(hw_pagesize);
142
143 if (sysctl(hw_pagesize_sysctl, 2, &hw_pagesize,
144 &len, NULL, 0) != 0)
145 return NULL;
146 if (len != sizeof(hw_pagesize))
147 return NULL;
148 if (hw_pagesize == 0 || (hw_pagesize & 0xFF) != 0)
149 return NULL;
150 }
151 atfork_storage = mmap(0, hw_pagesize, PROT_READ|PROT_WRITE,
152 MAP_PRIVATE | MAP_ANON, -1, 0);
153 if (__predict_false(atfork_storage == NULL))
154 return NULL;
155 cb_used(atfork_storage) = 1;
156 cb_ents(atfork_storage) =
157 (uint16_t)(hw_pagesize / sizeof(struct atfork_cb_block));
158 if (__predict_false(cb_ents(atfork_storage) < blocks + 1))
159 return NULL;
160 }
161
162 result = cb_blocks(atfork_storage) + cb_used(atfork_storage);
163 cb_used(atfork_storage) += blocks;
164
165 return result;
166 }
167
168 int
pthread_atfork(void (* prepare)(void),void (* parent)(void),void (* child)(void))169 pthread_atfork(void (*prepare)(void), void (*parent)(void),
170 void (*child)(void))
171 {
172 struct atfork_callback *newprepare, *newparent, *newchild;
173 sigset_t mask, omask;
174 int error;
175
176 sigfillset(&mask);
177 thr_sigsetmask(SIG_SETMASK, &mask, &omask);
178
179 mutex_lock(&atfork_lock);
180
181 /*
182 * Note here that we either get all the blocks
183 * we need, in one call, or we get NULL.
184 *
185 * Note also that a NULL return is not an error
186 * if no blocks were required (all args == NULL)
187 */
188 newprepare = af_alloc((prepare != NULL) +
189 (parent != NULL) + (child != NULL));
190
191 error = ENOMEM; /* in case of "goto out" */
192
193 newparent = newprepare;
194 if (prepare != NULL) {
195 if (__predict_false(newprepare == NULL))
196 goto out;
197 newprepare->fn = prepare;
198 newparent++;
199 }
200
201 newchild = newparent;
202 if (parent != NULL) {
203 if (__predict_false(newparent == NULL))
204 goto out;
205 newparent->fn = parent;
206 newchild++;
207 }
208
209 if (child != NULL) {
210 if (__predict_false(newchild == NULL))
211 goto out;
212 newchild->fn = child;
213 }
214
215 /*
216 * The order in which the functions are called is specified as
217 * LIFO for the prepare handler and FIFO for the others; insert
218 * at the head and tail as appropriate so that SIMPLEQ_FOREACH()
219 * produces the right order.
220 */
221 if (prepare)
222 SIMPLEQ_INSERT_HEAD(&prepareq, newprepare, next);
223 if (parent)
224 SIMPLEQ_INSERT_TAIL(&parentq, newparent, next);
225 if (child)
226 SIMPLEQ_INSERT_TAIL(&childq, newchild, next);
227
228 error = 0;
229
230 out:;
231 mutex_unlock(&atfork_lock);
232 thr_sigsetmask(SIG_SETMASK, &omask, NULL);
233 return error;
234 }
235
236 pid_t
fork(void)237 fork(void)
238 {
239 struct atfork_callback *iter;
240 pid_t ret;
241
242 mutex_lock(&atfork_lock);
243 SIMPLEQ_FOREACH(iter, &prepareq, next)
244 (*iter->fn)();
245 _malloc_prefork();
246
247 ret = __locked_fork(&errno);
248
249 if (ret != 0) {
250 /*
251 * We are the parent. It doesn't matter here whether
252 * the fork call succeeded or failed.
253 */
254 _malloc_postfork();
255 SIMPLEQ_FOREACH(iter, &parentq, next)
256 (*iter->fn)();
257 mutex_unlock(&atfork_lock);
258 } else {
259 /* We are the child */
260 _malloc_postfork_child();
261 SIMPLEQ_FOREACH(iter, &childq, next)
262 (*iter->fn)();
263 /*
264 * Note: We are explicitly *not* unlocking
265 * atfork_lock. Unlocking atfork_lock is problematic,
266 * because if any threads in the parent blocked on it
267 * between the initial lock and the fork() syscall,
268 * unlocking in the child will try to schedule
269 * threads, and either the internal mutex interlock or
270 * the runqueue spinlock could have been held at the
271 * moment of fork(). Since the other threads do not
272 * exist in this process, the spinlock will never be
273 * unlocked, and we would wedge.
274 * Instead, we reinitialize atfork_lock, since we know
275 * that the state of the atfork lists is consistent here,
276 * and that there are no other threads to be affected by
277 * the forcible cleaning of the queue.
278 * This permits double-forking to work, although
279 * it requires knowing that it's "safe" to initialize
280 * a locked mutex in this context.
281 *
282 * The problem exists for users of this interface,
283 * too, since the intended use of pthread_atfork() is
284 * to acquire locks across the fork call to ensure
285 * that the child sees consistent state. There's not
286 * much that can usefully be done in a child handler,
287 * and conventional wisdom discourages using them, but
288 * they're part of the interface, so here we are...
289 */
290 mutex_init(&atfork_lock, NULL);
291 }
292
293 return ret;
294 }
295