1 /*-
2 * Copyright 1999, 2000 John D. Polstra.
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
15 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
16 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
17 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
18 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
19 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
20 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
21 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
22 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
23 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
24 *
25 * from: FreeBSD: src/libexec/rtld-elf/sparc64/lockdflt.c,v 1.3 2002/10/09
26 * $FreeBSD$
27 */
28
29 /*
30 * Thread locking implementation for the dynamic linker.
31 *
32 * We use the "simple, non-scalable reader-preference lock" from:
33 *
34 * J. M. Mellor-Crummey and M. L. Scott. "Scalable Reader-Writer
35 * Synchronization for Shared-Memory Multiprocessors." 3rd ACM Symp. on
36 * Principles and Practice of Parallel Programming, April 1991.
37 *
38 * In this algorithm the lock is a single word. Its low-order bit is
39 * set when a writer holds the lock. The remaining high-order bits
40 * contain a count of readers desiring the lock. The algorithm requires
41 * atomic "compare_and_store" and "add" operations, which we take
42 * from machine/atomic.h.
43 */
44
45 #include <sys/param.h>
46 #include <signal.h>
47 #include <stdlib.h>
48 #include <time.h>
49
50 #include "debug.h"
51 #include "rtld.h"
52 #include "rtld_machdep.h"
53
54 void _rtld_thread_init(struct RtldLockInfo *) __exported;
55 void _rtld_atfork_pre(int *) __exported;
56 void _rtld_atfork_post(int *) __exported;
57
58 #define WAFLAG 0x1 /* A writer holds the lock */
59 #define RC_INCR 0x2 /* Adjusts count of readers desiring lock */
60
61 typedef struct Struct_Lock {
62 volatile u_int lock;
63 void *base;
64 } Lock;
65
66 static sigset_t fullsigmask, oldsigmask;
67 static int thread_flag, wnested;
68
69 static void *
def_lock_create(void)70 def_lock_create(void)
71 {
72 void *base;
73 char *p;
74 uintptr_t r;
75 Lock *l;
76
77 /*
78 * Arrange for the lock to occupy its own cache line. First, we
79 * optimistically allocate just a cache line, hoping that malloc
80 * will give us a well-aligned block of memory. If that doesn't
81 * work, we allocate a larger block and take a well-aligned cache
82 * line from it.
83 */
84 base = xmalloc(CACHE_LINE_SIZE);
85 p = (char *)base;
86 if ((uintptr_t)p % CACHE_LINE_SIZE != 0) {
87 free(base);
88 base = xmalloc(2 * CACHE_LINE_SIZE);
89 p = (char *)base;
90 if ((r = (uintptr_t)p % CACHE_LINE_SIZE) != 0)
91 p += CACHE_LINE_SIZE - r;
92 }
93 l = (Lock *)p;
94 l->base = base;
95 l->lock = 0;
96 return l;
97 }
98
99 static void
def_lock_destroy(void * lock)100 def_lock_destroy(void *lock)
101 {
102 Lock *l = (Lock *)lock;
103
104 free(l->base);
105 }
106
107 static void
def_rlock_acquire(void * lock)108 def_rlock_acquire(void *lock)
109 {
110 Lock *l = (Lock *)lock;
111
112 atomic_add_acq_int(&l->lock, RC_INCR);
113 while (l->lock & WAFLAG)
114 ; /* Spin */
115 }
116
117 static void
def_wlock_acquire(void * lock)118 def_wlock_acquire(void *lock)
119 {
120 Lock *l;
121 sigset_t tmp_oldsigmask;
122
123 l = (Lock *)lock;
124 for (;;) {
125 sigprocmask(SIG_BLOCK, &fullsigmask, &tmp_oldsigmask);
126 if (atomic_cmpset_acq_int(&l->lock, 0, WAFLAG))
127 break;
128 sigprocmask(SIG_SETMASK, &tmp_oldsigmask, NULL);
129 }
130 if (atomic_fetchadd_int(&wnested, 1) == 0)
131 oldsigmask = tmp_oldsigmask;
132 }
133
134 static void
def_lock_release(void * lock)135 def_lock_release(void *lock)
136 {
137 Lock *l;
138
139 l = (Lock *)lock;
140 if ((l->lock & WAFLAG) == 0)
141 atomic_add_rel_int(&l->lock, -RC_INCR);
142 else {
143 assert(wnested > 0);
144 atomic_add_rel_int(&l->lock, -WAFLAG);
145 if (atomic_fetchadd_int(&wnested, -1) == 1)
146 sigprocmask(SIG_SETMASK, &oldsigmask, NULL);
147 }
148 }
149
150 static int
def_thread_set_flag(int mask)151 def_thread_set_flag(int mask)
152 {
153 int old_val = thread_flag;
154 thread_flag |= mask;
155 return (old_val);
156 }
157
158 static int
def_thread_clr_flag(int mask)159 def_thread_clr_flag(int mask)
160 {
161 int old_val = thread_flag;
162 thread_flag &= ~mask;
163 return (old_val);
164 }
165
166 /*
167 * Public interface exposed to the rest of the dynamic linker.
168 */
169 static struct RtldLockInfo lockinfo;
170 static struct RtldLockInfo deflockinfo;
171
172 static __inline int
thread_mask_set(int mask)173 thread_mask_set(int mask)
174 {
175 return lockinfo.thread_set_flag(mask);
176 }
177
178 static __inline void
thread_mask_clear(int mask)179 thread_mask_clear(int mask)
180 {
181 lockinfo.thread_clr_flag(mask);
182 }
183
184 #define RTLD_LOCK_CNT 3
185 struct rtld_lock {
186 void *handle;
187 int mask;
188 } rtld_locks[RTLD_LOCK_CNT];
189
190 rtld_lock_t rtld_bind_lock = &rtld_locks[0];
191 rtld_lock_t rtld_libc_lock = &rtld_locks[1];
192 rtld_lock_t rtld_phdr_lock = &rtld_locks[2];
193
194 void
rlock_acquire(rtld_lock_t lock,RtldLockState * lockstate)195 rlock_acquire(rtld_lock_t lock, RtldLockState *lockstate)
196 {
197
198 if (lockstate == NULL)
199 return;
200
201 if (thread_mask_set(lock->mask) & lock->mask) {
202 dbg("rlock_acquire: recursed");
203 lockstate->lockstate = RTLD_LOCK_UNLOCKED;
204 return;
205 }
206 lockinfo.rlock_acquire(lock->handle);
207 lockstate->lockstate = RTLD_LOCK_RLOCKED;
208 }
209
210 void
wlock_acquire(rtld_lock_t lock,RtldLockState * lockstate)211 wlock_acquire(rtld_lock_t lock, RtldLockState *lockstate)
212 {
213
214 if (lockstate == NULL)
215 return;
216
217 if (thread_mask_set(lock->mask) & lock->mask) {
218 dbg("wlock_acquire: recursed");
219 lockstate->lockstate = RTLD_LOCK_UNLOCKED;
220 return;
221 }
222 lockinfo.wlock_acquire(lock->handle);
223 lockstate->lockstate = RTLD_LOCK_WLOCKED;
224 }
225
226 void
lock_release(rtld_lock_t lock,RtldLockState * lockstate)227 lock_release(rtld_lock_t lock, RtldLockState *lockstate)
228 {
229
230 if (lockstate == NULL)
231 return;
232
233 switch (lockstate->lockstate) {
234 case RTLD_LOCK_UNLOCKED:
235 break;
236 case RTLD_LOCK_RLOCKED:
237 case RTLD_LOCK_WLOCKED:
238 thread_mask_clear(lock->mask);
239 lockinfo.lock_release(lock->handle);
240 break;
241 default:
242 assert(0);
243 }
244 }
245
246 void
lock_upgrade(rtld_lock_t lock,RtldLockState * lockstate)247 lock_upgrade(rtld_lock_t lock, RtldLockState *lockstate)
248 {
249
250 if (lockstate == NULL)
251 return;
252
253 lock_release(lock, lockstate);
254 wlock_acquire(lock, lockstate);
255 }
256
257 void
lock_restart_for_upgrade(RtldLockState * lockstate)258 lock_restart_for_upgrade(RtldLockState *lockstate)
259 {
260
261 if (lockstate == NULL)
262 return;
263
264 switch (lockstate->lockstate) {
265 case RTLD_LOCK_UNLOCKED:
266 case RTLD_LOCK_WLOCKED:
267 break;
268 case RTLD_LOCK_RLOCKED:
269 siglongjmp(lockstate->env, 1);
270 break;
271 default:
272 assert(0);
273 }
274 }
275
276 void
lockdflt_init(void)277 lockdflt_init(void)
278 {
279 int i;
280
281 deflockinfo.rtli_version = RTLI_VERSION;
282 deflockinfo.lock_create = def_lock_create;
283 deflockinfo.lock_destroy = def_lock_destroy;
284 deflockinfo.rlock_acquire = def_rlock_acquire;
285 deflockinfo.wlock_acquire = def_wlock_acquire;
286 deflockinfo.lock_release = def_lock_release;
287 deflockinfo.thread_set_flag = def_thread_set_flag;
288 deflockinfo.thread_clr_flag = def_thread_clr_flag;
289 deflockinfo.at_fork = NULL;
290
291 for (i = 0; i < RTLD_LOCK_CNT; i++) {
292 rtld_locks[i].mask = (1 << i);
293 rtld_locks[i].handle = NULL;
294 }
295
296 memcpy(&lockinfo, &deflockinfo, sizeof(lockinfo));
297 _rtld_thread_init(NULL);
298 /*
299 * Construct a mask to block all signals except traps which might
300 * conceivably be generated within the dynamic linker itself.
301 */
302 sigfillset(&fullsigmask);
303 sigdelset(&fullsigmask, SIGILL);
304 sigdelset(&fullsigmask, SIGTRAP);
305 sigdelset(&fullsigmask, SIGABRT);
306 sigdelset(&fullsigmask, SIGEMT);
307 sigdelset(&fullsigmask, SIGFPE);
308 sigdelset(&fullsigmask, SIGBUS);
309 sigdelset(&fullsigmask, SIGSEGV);
310 sigdelset(&fullsigmask, SIGSYS);
311 }
312
313 /*
314 * Callback function to allow threads implementation to
315 * register their own locking primitives if the default
316 * one is not suitable.
317 * The current context should be the only context
318 * executing at the invocation time.
319 */
320 void
_rtld_thread_init(struct RtldLockInfo * pli)321 _rtld_thread_init(struct RtldLockInfo *pli)
322 {
323 int flags, i;
324 void *locks[RTLD_LOCK_CNT];
325
326 /* disable all locking while this function is running */
327 flags = thread_mask_set(~0);
328
329 if (pli == NULL)
330 pli = &deflockinfo;
331
332
333 for (i = 0; i < RTLD_LOCK_CNT; i++)
334 if ((locks[i] = pli->lock_create()) == NULL)
335 break;
336
337 if (i < RTLD_LOCK_CNT) {
338 while (--i >= 0)
339 pli->lock_destroy(locks[i]);
340 abort();
341 }
342
343 for (i = 0; i < RTLD_LOCK_CNT; i++) {
344 if (rtld_locks[i].handle == NULL)
345 continue;
346 if (flags & rtld_locks[i].mask)
347 lockinfo.lock_release(rtld_locks[i].handle);
348 lockinfo.lock_destroy(rtld_locks[i].handle);
349 }
350
351 for (i = 0; i < RTLD_LOCK_CNT; i++) {
352 rtld_locks[i].handle = locks[i];
353 if (flags & rtld_locks[i].mask)
354 pli->wlock_acquire(rtld_locks[i].handle);
355 }
356
357 lockinfo.lock_create = pli->lock_create;
358 lockinfo.lock_destroy = pli->lock_destroy;
359 lockinfo.rlock_acquire = pli->rlock_acquire;
360 lockinfo.wlock_acquire = pli->wlock_acquire;
361 lockinfo.lock_release = pli->lock_release;
362 lockinfo.thread_set_flag = pli->thread_set_flag;
363 lockinfo.thread_clr_flag = pli->thread_clr_flag;
364 lockinfo.at_fork = pli->at_fork;
365
366 /* restore thread locking state, this time with new locks */
367 thread_mask_clear(~0);
368 thread_mask_set(flags);
369 dbg("_rtld_thread_init: done");
370 }
371
372 void
_rtld_atfork_pre(int * locks)373 _rtld_atfork_pre(int *locks)
374 {
375 RtldLockState ls[2];
376
377 if (locks == NULL)
378 return;
379
380 /*
381 * Warning: this did not worked well with the rtld compat
382 * locks above, when the thread signal mask was corrupted (set
383 * to all signals blocked) if two locks were taken
384 * simultaneously in the write mode. The caller of the
385 * _rtld_atfork_pre() must provide the working implementation
386 * of the locks anyway, and libthr locks are fine.
387 */
388 wlock_acquire(rtld_phdr_lock, &ls[0]);
389 wlock_acquire(rtld_bind_lock, &ls[1]);
390
391 /* XXXKIB: I am really sorry for this. */
392 locks[0] = ls[1].lockstate;
393 locks[2] = ls[0].lockstate;
394 }
395
396 void
_rtld_atfork_post(int * locks)397 _rtld_atfork_post(int *locks)
398 {
399 RtldLockState ls[2];
400
401 if (locks == NULL)
402 return;
403
404 bzero(ls, sizeof(ls));
405 ls[0].lockstate = locks[2];
406 ls[1].lockstate = locks[0];
407 lock_release(rtld_bind_lock, &ls[1]);
408 lock_release(rtld_phdr_lock, &ls[0]);
409 }
410