Commit | Line | Data |
---|---|---|
9f1621ca MD |
1 | /* |
2 | * urcu.c | |
3 | * | |
4 | * Userspace RCU library | |
5 | * | |
6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> | |
7 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. | |
8 | * | |
9 | * This library is free software; you can redistribute it and/or | |
10 | * modify it under the terms of the GNU Lesser General Public | |
11 | * License as published by the Free Software Foundation; either | |
12 | * version 2.1 of the License, or (at your option) any later version. | |
13 | * | |
14 | * This library is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * Lesser General Public License for more details. | |
18 | * | |
19 | * You should have received a copy of the GNU Lesser General Public | |
20 | * License along with this library; if not, write to the Free Software | |
21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
22 | * | |
23 | * IBM's contributions to this file may be relicensed under LGPLv2 or later. | |
24 | */ | |
25 | ||
26 | #include <stdio.h> | |
27 | #include <pthread.h> | |
28 | #include <signal.h> | |
29 | #include <assert.h> | |
30 | #include <stdlib.h> | |
31 | #include <string.h> | |
32 | #include <errno.h> | |
33 | #include <poll.h> | |
34 | ||
35 | #include "urcu-static.h" | |
36 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ | |
37 | #include "urcu.h" | |
38 | ||
39 | void __attribute__((constructor)) urcu_init(void); | |
40 | void __attribute__((destructor)) urcu_exit(void); | |
41 | ||
42 | int init_done; | |
43 | ||
44 | pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; | |
45 | ||
46 | /* | |
47 | * Global grace period counter. | |
48 | */ | |
49 | long urcu_gp_ctr = 0; | |
50 | ||
51 | /* | |
52 | * Written to only by each individual reader. Read by both the reader and the | |
53 | * writers. | |
54 | */ | |
55 | long __thread urcu_active_readers; | |
56 | ||
57 | /* Thread IDs of registered readers */ | |
58 | #define INIT_NUM_THREADS 4 | |
59 | ||
60 | struct reader_registry { | |
61 | pthread_t tid; | |
62 | long *urcu_active_readers; | |
63 | char *need_mb; | |
64 | }; | |
65 | ||
66 | #ifdef DEBUG_YIELD | |
67 | unsigned int yield_active; | |
68 | unsigned int __thread rand_yield; | |
69 | #endif | |
70 | ||
71 | static struct reader_registry *registry; | |
72 | static char __thread need_mb; | |
73 | static int num_readers, alloc_readers; | |
74 | ||
75 | void internal_urcu_lock(void) | |
76 | { | |
77 | int ret; | |
78 | ||
79 | #ifndef DISTRUST_SIGNALS_EXTREME | |
80 | ret = pthread_mutex_lock(&urcu_mutex); | |
81 | if (ret) { | |
82 | perror("Error in pthread mutex lock"); | |
83 | exit(-1); | |
84 | } | |
85 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
86 | while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { | |
87 | if (ret != EBUSY && ret != EINTR) { | |
88 | printf("ret = %d, errno = %d\n", ret, errno); | |
89 | perror("Error in pthread mutex lock"); | |
90 | exit(-1); | |
91 | } | |
92 | if (need_mb) { | |
93 | smp_mb(); | |
94 | need_mb = 0; | |
95 | smp_mb(); | |
96 | } | |
97 | poll(NULL,0,10); | |
98 | } | |
99 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
100 | } | |
101 | ||
102 | void internal_urcu_unlock(void) | |
103 | { | |
104 | int ret; | |
105 | ||
106 | ret = pthread_mutex_unlock(&urcu_mutex); | |
107 | if (ret) { | |
108 | perror("Error in pthread mutex unlock"); | |
109 | exit(-1); | |
110 | } | |
111 | } | |
112 | ||
113 | #ifdef DEBUG_FULL_MB | |
114 | #ifdef HAS_INCOHERENT_CACHES | |
115 | static void force_mb_single_thread(struct reader_registry *index) | |
116 | { | |
117 | smp_mb(); | |
118 | } | |
119 | #endif /* #ifdef HAS_INCOHERENT_CACHES */ | |
120 | ||
121 | static void force_mb_all_threads(void) | |
122 | { | |
123 | smp_mb(); | |
124 | } | |
125 | #else /* #ifdef DEBUG_FULL_MB */ | |
126 | #ifdef HAS_INCOHERENT_CACHES | |
127 | static void force_mb_single_thread(struct reader_registry *index) | |
128 | { | |
129 | assert(registry); | |
130 | /* | |
131 | * pthread_kill has a smp_mb(). But beware, we assume it performs | |
132 | * a cache flush on architectures with non-coherent cache. Let's play | |
133 | * safe and don't assume anything : we use smp_mc() to make sure the | |
134 | * cache flush is enforced. | |
135 | */ | |
136 | *index->need_mb = 1; | |
137 | smp_mc(); /* write ->need_mb before sending the signals */ | |
138 | pthread_kill(index->tid, SIGURCU); | |
139 | smp_mb(); | |
140 | /* | |
141 | * Wait for sighandler (and thus mb()) to execute on every thread. | |
142 | * BUSY-LOOP. | |
143 | */ | |
144 | while (*index->need_mb) { | |
145 | poll(NULL, 0, 1); | |
146 | } | |
147 | smp_mb(); /* read ->need_mb before ending the barrier */ | |
148 | } | |
149 | #endif /* #ifdef HAS_INCOHERENT_CACHES */ | |
150 | ||
151 | static void force_mb_all_threads(void) | |
152 | { | |
153 | struct reader_registry *index; | |
154 | /* | |
155 | * Ask for each threads to execute a smp_mb() so we can consider the | |
156 | * compiler barriers around rcu read lock as real memory barriers. | |
157 | */ | |
158 | if (!registry) | |
159 | return; | |
160 | /* | |
161 | * pthread_kill has a smp_mb(). But beware, we assume it performs | |
162 | * a cache flush on architectures with non-coherent cache. Let's play | |
163 | * safe and don't assume anything : we use smp_mc() to make sure the | |
164 | * cache flush is enforced. | |
165 | */ | |
166 | for (index = registry; index < registry + num_readers; index++) { | |
167 | *index->need_mb = 1; | |
168 | smp_mc(); /* write need_mb before sending the signal */ | |
169 | pthread_kill(index->tid, SIGURCU); | |
170 | } | |
171 | /* | |
172 | * Wait for sighandler (and thus mb()) to execute on every thread. | |
173 | * | |
174 | * Note that the pthread_kill() will never be executed on systems | |
175 | * that correctly deliver signals in a timely manner. However, it | |
176 | * is not uncommon for kernels to have bugs that can result in | |
177 | * lost or unduly delayed signals. | |
178 | * | |
179 | * If you are seeing the below pthread_kill() executing much at | |
180 | * all, we suggest testing the underlying kernel and filing the | |
181 | * relevant bug report. For Linux kernels, we recommend getting | |
182 | * the Linux Test Project (LTP). | |
183 | */ | |
184 | for (index = registry; index < registry + num_readers; index++) { | |
185 | while (*index->need_mb) { | |
186 | pthread_kill(index->tid, SIGURCU); | |
187 | poll(NULL, 0, 1); | |
188 | } | |
189 | } | |
190 | smp_mb(); /* read ->need_mb before ending the barrier */ | |
191 | } | |
192 | #endif /* #else #ifdef DEBUG_FULL_MB */ | |
193 | ||
194 | void wait_for_quiescent_state(void) | |
195 | { | |
196 | struct reader_registry *index; | |
197 | ||
198 | if (!registry) | |
199 | return; | |
200 | /* | |
201 | * Wait for each thread urcu_active_readers count to become 0. | |
202 | */ | |
203 | for (index = registry; index < registry + num_readers; index++) { | |
204 | #ifndef HAS_INCOHERENT_CACHES | |
205 | while (rcu_old_gp_ongoing(index->urcu_active_readers)) | |
206 | cpu_relax(); | |
207 | #else /* #ifndef HAS_INCOHERENT_CACHES */ | |
208 | int wait_loops = 0; | |
209 | /* | |
210 | * BUSY-LOOP. Force the reader thread to commit its | |
211 | * urcu_active_readers update to memory if we wait for too long. | |
212 | */ | |
213 | while (rcu_old_gp_ongoing(index->urcu_active_readers)) { | |
214 | if (wait_loops++ == KICK_READER_LOOPS) { | |
215 | force_mb_single_thread(index); | |
216 | wait_loops = 0; | |
217 | } else { | |
218 | cpu_relax(); | |
219 | } | |
220 | } | |
221 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ | |
222 | } | |
223 | } | |
224 | ||
225 | void synchronize_rcu(void) | |
226 | { | |
227 | internal_urcu_lock(); | |
228 | force_mb_all_threads(); | |
229 | urcu_gp_ctr += 2; | |
230 | wait_for_quiescent_state(); | |
231 | force_mb_all_threads(); | |
232 | internal_urcu_unlock(); | |
233 | } | |
234 | ||
235 | /* | |
236 | * library wrappers to be used by non-LGPL compatible source code. | |
237 | */ | |
238 | ||
239 | void rcu_read_lock(void) | |
240 | { | |
241 | _rcu_read_lock(); | |
242 | } | |
243 | ||
244 | void rcu_read_unlock(void) | |
245 | { | |
246 | _rcu_read_unlock(); | |
247 | } | |
248 | ||
249 | void *rcu_dereference(void *p) | |
250 | { | |
251 | return _rcu_dereference(p); | |
252 | } | |
253 | ||
254 | void *rcu_assign_pointer_sym(void **p, void *v) | |
255 | { | |
256 | wmb(); | |
257 | return STORE_SHARED(p, v); | |
258 | } | |
259 | ||
260 | void *rcu_xchg_pointer_sym(void **p, void *v) | |
261 | { | |
262 | wmb(); | |
263 | return xchg(p, v); | |
264 | } | |
265 | ||
266 | void *rcu_publish_content_sym(void **p, void *v) | |
267 | { | |
268 | void *oldptr; | |
269 | ||
270 | oldptr = _rcu_xchg_pointer(p, v); | |
271 | synchronize_rcu(); | |
272 | return oldptr; | |
273 | } | |
274 | ||
275 | static void rcu_add_reader(pthread_t id) | |
276 | { | |
277 | struct reader_registry *oldarray; | |
278 | ||
279 | if (!registry) { | |
280 | alloc_readers = INIT_NUM_THREADS; | |
281 | num_readers = 0; | |
282 | registry = | |
283 | malloc(sizeof(struct reader_registry) * alloc_readers); | |
284 | } | |
285 | if (alloc_readers < num_readers + 1) { | |
286 | oldarray = registry; | |
287 | registry = malloc(sizeof(struct reader_registry) | |
288 | * (alloc_readers << 1)); | |
289 | memcpy(registry, oldarray, | |
290 | sizeof(struct reader_registry) * alloc_readers); | |
291 | alloc_readers <<= 1; | |
292 | free(oldarray); | |
293 | } | |
294 | registry[num_readers].tid = id; | |
295 | /* reference to the TLS of _this_ reader thread. */ | |
296 | registry[num_readers].urcu_active_readers = &urcu_active_readers; | |
297 | registry[num_readers].need_mb = &need_mb; | |
298 | num_readers++; | |
299 | } | |
300 | ||
301 | /* | |
302 | * Never shrink (implementation limitation). | |
303 | * This is O(nb threads). Eventually use a hash table. | |
304 | */ | |
305 | static void rcu_remove_reader(pthread_t id) | |
306 | { | |
307 | struct reader_registry *index; | |
308 | ||
309 | assert(registry != NULL); | |
310 | for (index = registry; index < registry + num_readers; index++) { | |
311 | if (pthread_equal(index->tid, id)) { | |
312 | memcpy(index, ®istry[num_readers - 1], | |
313 | sizeof(struct reader_registry)); | |
314 | registry[num_readers - 1].tid = 0; | |
315 | registry[num_readers - 1].urcu_active_readers = NULL; | |
316 | num_readers--; | |
317 | return; | |
318 | } | |
319 | } | |
320 | /* Hrm not found, forgot to register ? */ | |
321 | assert(0); | |
322 | } | |
323 | ||
324 | void rcu_register_thread(void) | |
325 | { | |
326 | internal_urcu_lock(); | |
327 | urcu_init(); /* In case gcc does not support constructor attribute */ | |
328 | rcu_add_reader(pthread_self()); | |
329 | internal_urcu_unlock(); | |
330 | } | |
331 | ||
332 | void rcu_unregister_thread(void) | |
333 | { | |
334 | internal_urcu_lock(); | |
335 | rcu_remove_reader(pthread_self()); | |
336 | internal_urcu_unlock(); | |
337 | } | |
338 | ||
339 | #ifndef DEBUG_FULL_MB | |
340 | static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) | |
341 | { | |
342 | /* | |
343 | * Executing this smp_mb() is the only purpose of this signal handler. | |
344 | * It punctually promotes barrier() into smp_mb() on every thread it is | |
345 | * executed on. | |
346 | */ | |
347 | smp_mb(); | |
348 | need_mb = 0; | |
349 | smp_mb(); | |
350 | } | |
351 | ||
352 | /* | |
353 | * urcu_init constructor. Called when the library is linked, but also when | |
354 | * reader threads are calling rcu_register_thread(). | |
355 | * Should only be called by a single thread at a given time. This is ensured by | |
356 | * holing the internal_urcu_lock() from rcu_register_thread() or by running at | |
357 | * library load time, which should not be executed by multiple threads nor | |
358 | * concurrently with rcu_register_thread() anyway. | |
359 | */ | |
360 | void urcu_init(void) | |
361 | { | |
362 | struct sigaction act; | |
363 | int ret; | |
364 | ||
365 | if (init_done) | |
366 | return; | |
367 | init_done = 1; | |
368 | ||
369 | act.sa_sigaction = sigurcu_handler; | |
370 | ret = sigaction(SIGURCU, &act, NULL); | |
371 | if (ret) { | |
372 | perror("Error in sigaction"); | |
373 | exit(-1); | |
374 | } | |
375 | } | |
376 | ||
377 | void urcu_exit(void) | |
378 | { | |
379 | struct sigaction act; | |
380 | int ret; | |
381 | ||
382 | ret = sigaction(SIGURCU, NULL, &act); | |
383 | if (ret) { | |
384 | perror("Error in sigaction"); | |
385 | exit(-1); | |
386 | } | |
387 | assert(act.sa_sigaction == sigurcu_handler); | |
388 | free(registry); | |
389 | } | |
390 | #endif /* #ifndef DEBUG_FULL_MB */ |