Commit | Line | Data |
---|---|---|
9f1621ca | 1 | /* |
7ac06cef | 2 | * urcu-qsbr.c |
9f1621ca | 3 | * |
7ac06cef | 4 | * Userspace RCU QSBR library |
9f1621ca MD |
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> | |
ee72ef6e | 34 | #include <unistd.h> |
9f1621ca | 35 | |
727f819d | 36 | #define BUILD_QSBR_LIB |
7ac06cef | 37 | #include "urcu-qsbr-static.h" |
9f1621ca | 38 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
7ac06cef | 39 | #include "urcu-qsbr.h" |
9f1621ca | 40 | |
81d1f1f5 | 41 | static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
9f1621ca MD |
42 | |
43 | /* | |
44 | * Global grace period counter. | |
45 | */ | |
ac258107 | 46 | unsigned long urcu_gp_ctr = RCU_GP_ONLINE; |
9f1621ca MD |
47 | |
48 | /* | |
49 | * Written to only by each individual reader. Read by both the reader and the | |
50 | * writers. | |
51 | */ | |
8b25e300 | 52 | struct urcu_reader_status __thread urcu_reader_status; |
9f1621ca MD |
53 | |
54 | /* Thread IDs of registered readers */ | |
55 | #define INIT_NUM_THREADS 4 | |
56 | ||
57 | struct reader_registry { | |
58 | pthread_t tid; | |
8b25e300 | 59 | struct urcu_reader_status *urcu_reader_status; |
9f1621ca MD |
60 | }; |
61 | ||
62 | #ifdef DEBUG_YIELD | |
63 | unsigned int yield_active; | |
64 | unsigned int __thread rand_yield; | |
65 | #endif | |
66 | ||
67 | static struct reader_registry *registry; | |
9f1621ca MD |
68 | static int num_readers, alloc_readers; |
69 | ||
90c1618a | 70 | static void internal_urcu_lock(void) |
9f1621ca MD |
71 | { |
72 | int ret; | |
73 | ||
74 | #ifndef DISTRUST_SIGNALS_EXTREME | |
75 | ret = pthread_mutex_lock(&urcu_mutex); | |
76 | if (ret) { | |
77 | perror("Error in pthread mutex lock"); | |
78 | exit(-1); | |
79 | } | |
80 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
81 | while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { | |
82 | if (ret != EBUSY && ret != EINTR) { | |
83 | printf("ret = %d, errno = %d\n", ret, errno); | |
84 | perror("Error in pthread mutex lock"); | |
85 | exit(-1); | |
86 | } | |
9f1621ca MD |
87 | poll(NULL,0,10); |
88 | } | |
89 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
90 | } | |
91 | ||
90c1618a | 92 | static void internal_urcu_unlock(void) |
9f1621ca MD |
93 | { |
94 | int ret; | |
95 | ||
96 | ret = pthread_mutex_unlock(&urcu_mutex); | |
97 | if (ret) { | |
98 | perror("Error in pthread mutex unlock"); | |
99 | exit(-1); | |
100 | } | |
101 | } | |
102 | ||
bc6c15bb MD |
103 | /* |
104 | * synchronize_rcu() waiting. Single thread. | |
105 | */ | |
90c1618a | 106 | static void wait_for_quiescent_state(void) |
9f1621ca MD |
107 | { |
108 | struct reader_registry *index; | |
109 | ||
110 | if (!registry) | |
111 | return; | |
112 | /* | |
8b25e300 | 113 | * Wait for each thread rcu_reader qs_gp count to become 0. |
9f1621ca MD |
114 | */ |
115 | for (index = registry; index < registry + num_readers; index++) { | |
bc6c15bb MD |
116 | int wait_loops = 0; |
117 | ||
8b25e300 MD |
118 | index->urcu_reader_status->gp_waiting = 1; |
119 | while (rcu_gp_ongoing(&index->urcu_reader_status->qs_gp)) { | |
bc6c15bb | 120 | if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) { |
ee72ef6e | 121 | /* adapted wait time, in us */ |
539718ae | 122 | usleep(LOAD_SHARED(index->urcu_reader_status->qs_time_delta_usec) / 4); |
b3c4dd1a | 123 | wait_loops = 0; |
bc6c15bb | 124 | } else { |
9f1621ca | 125 | #ifndef HAS_INCOHERENT_CACHES |
bc6c15bb | 126 | cpu_relax(); |
9f1621ca | 127 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
bc6c15bb | 128 | smp_mb(); |
9f1621ca | 129 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
bc6c15bb MD |
130 | } |
131 | } | |
8b25e300 | 132 | index->urcu_reader_status->gp_waiting = 0; |
9f1621ca MD |
133 | } |
134 | } | |
135 | ||
47d2f29e MD |
136 | /* |
137 | * Using a two-subphases algorithm for architectures with smaller than 64-bit | |
138 | * long-size to ensure we do not encounter an overflow bug. | |
139 | */ | |
140 | ||
141 | #if (BITS_PER_LONG < 64) | |
142 | /* | |
143 | * called with urcu_mutex held. | |
144 | */ | |
145 | static void switch_next_urcu_qparity(void) | |
146 | { | |
147 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR); | |
148 | } | |
149 | ||
150 | void synchronize_rcu(void) | |
151 | { | |
bc49c323 MD |
152 | unsigned long was_online; |
153 | ||
8b25e300 | 154 | was_online = urcu_reader_status.qs_gp; |
bc49c323 | 155 | |
47d2f29e MD |
156 | /* All threads should read qparity before accessing data structure |
157 | * where new ptr points to. | |
158 | */ | |
159 | /* Write new ptr before changing the qparity */ | |
160 | smp_mb(); | |
161 | ||
bc49c323 MD |
162 | /* |
163 | * Mark the writer thread offline to make sure we don't wait for | |
164 | * our own quiescent state. This allows using synchronize_rcu() in | |
165 | * threads registered as readers. | |
166 | */ | |
167 | if (was_online) | |
8b25e300 | 168 | STORE_SHARED(urcu_reader_status.qs_gp, 0); |
bc49c323 | 169 | |
47d2f29e MD |
170 | internal_urcu_lock(); |
171 | ||
ae62b5e8 MD |
172 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
173 | ||
47d2f29e MD |
174 | switch_next_urcu_qparity(); /* 0 -> 1 */ |
175 | ||
176 | /* | |
177 | * Must commit qparity update to memory before waiting for parity | |
178 | * 0 quiescent state. Failure to do so could result in the writer | |
179 | * waiting forever while new readers are always accessing data (no | |
180 | * progress). | |
181 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
182 | */ | |
183 | ||
184 | /* | |
185 | * Wait for previous parity to be empty of readers. | |
186 | */ | |
187 | wait_for_quiescent_state(); /* Wait readers in parity 0 */ | |
188 | ||
189 | /* | |
190 | * Must finish waiting for quiescent state for parity 0 before | |
191 | * committing qparity update to memory. Failure to do so could result in | |
192 | * the writer waiting forever while new readers are always accessing | |
193 | * data (no progress). | |
194 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
195 | */ | |
196 | ||
197 | switch_next_urcu_qparity(); /* 1 -> 0 */ | |
198 | ||
199 | /* | |
200 | * Must commit qparity update to memory before waiting for parity | |
201 | * 1 quiescent state. Failure to do so could result in the writer | |
202 | * waiting forever while new readers are always accessing data (no | |
203 | * progress). | |
204 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
205 | */ | |
206 | ||
207 | /* | |
208 | * Wait for previous parity to be empty of readers. | |
209 | */ | |
210 | wait_for_quiescent_state(); /* Wait readers in parity 1 */ | |
211 | ||
ae62b5e8 MD |
212 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
213 | ||
47d2f29e MD |
214 | internal_urcu_unlock(); |
215 | ||
bc49c323 MD |
216 | /* |
217 | * Finish waiting for reader threads before letting the old ptr being | |
47d2f29e MD |
218 | * freed. |
219 | */ | |
bc49c323 | 220 | if (was_online) |
8b25e300 MD |
221 | _STORE_SHARED(urcu_reader_status.qs_gp, |
222 | LOAD_SHARED(urcu_gp_ctr)); | |
47d2f29e MD |
223 | smp_mb(); |
224 | } | |
225 | #else /* !(BITS_PER_LONG < 64) */ | |
9f1621ca MD |
226 | void synchronize_rcu(void) |
227 | { | |
f0f7dbdd | 228 | unsigned long was_online; |
ff2f67a0 | 229 | |
8b25e300 | 230 | was_online = urcu_reader_status.qs_gp; |
ff2f67a0 MD |
231 | |
232 | /* | |
233 | * Mark the writer thread offline to make sure we don't wait for | |
234 | * our own quiescent state. This allows using synchronize_rcu() in | |
235 | * threads registered as readers. | |
236 | */ | |
8f50d1ce | 237 | smp_mb(); |
ff2f67a0 | 238 | if (was_online) |
8b25e300 | 239 | STORE_SHARED(urcu_reader_status.qs_gp, 0); |
ff2f67a0 | 240 | |
4e27f058 | 241 | internal_urcu_lock(); |
ae62b5e8 | 242 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
55570466 | 243 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR); |
9f1621ca | 244 | wait_for_quiescent_state(); |
ae62b5e8 | 245 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
9f1621ca | 246 | internal_urcu_unlock(); |
ff2f67a0 MD |
247 | |
248 | if (was_online) | |
8b25e300 MD |
249 | _STORE_SHARED(urcu_reader_status.qs_gp, |
250 | LOAD_SHARED(urcu_gp_ctr)); | |
8f50d1ce | 251 | smp_mb(); |
9f1621ca | 252 | } |
47d2f29e | 253 | #endif /* !(BITS_PER_LONG < 64) */ |
9f1621ca MD |
254 | |
255 | /* | |
256 | * library wrappers to be used by non-LGPL compatible source code. | |
257 | */ | |
258 | ||
259 | void rcu_read_lock(void) | |
260 | { | |
261 | _rcu_read_lock(); | |
262 | } | |
263 | ||
264 | void rcu_read_unlock(void) | |
265 | { | |
266 | _rcu_read_unlock(); | |
267 | } | |
268 | ||
269 | void *rcu_dereference(void *p) | |
270 | { | |
271 | return _rcu_dereference(p); | |
272 | } | |
273 | ||
274 | void *rcu_assign_pointer_sym(void **p, void *v) | |
275 | { | |
276 | wmb(); | |
277 | return STORE_SHARED(p, v); | |
278 | } | |
279 | ||
4d1ce26f MD |
280 | void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
281 | { | |
282 | wmb(); | |
283 | return cmpxchg(p, old, _new); | |
284 | } | |
285 | ||
9f1621ca MD |
286 | void *rcu_xchg_pointer_sym(void **p, void *v) |
287 | { | |
288 | wmb(); | |
289 | return xchg(p, v); | |
290 | } | |
291 | ||
292 | void *rcu_publish_content_sym(void **p, void *v) | |
293 | { | |
294 | void *oldptr; | |
295 | ||
296 | oldptr = _rcu_xchg_pointer(p, v); | |
297 | synchronize_rcu(); | |
298 | return oldptr; | |
299 | } | |
300 | ||
7ac06cef MD |
301 | void rcu_quiescent_state(void) |
302 | { | |
303 | _rcu_quiescent_state(); | |
304 | } | |
305 | ||
306 | void rcu_thread_offline(void) | |
307 | { | |
308 | _rcu_thread_offline(); | |
309 | } | |
310 | ||
311 | void rcu_thread_online(void) | |
312 | { | |
313 | _rcu_thread_online(); | |
314 | } | |
315 | ||
9f1621ca MD |
316 | static void rcu_add_reader(pthread_t id) |
317 | { | |
318 | struct reader_registry *oldarray; | |
319 | ||
320 | if (!registry) { | |
321 | alloc_readers = INIT_NUM_THREADS; | |
322 | num_readers = 0; | |
323 | registry = | |
324 | malloc(sizeof(struct reader_registry) * alloc_readers); | |
325 | } | |
326 | if (alloc_readers < num_readers + 1) { | |
327 | oldarray = registry; | |
328 | registry = malloc(sizeof(struct reader_registry) | |
329 | * (alloc_readers << 1)); | |
330 | memcpy(registry, oldarray, | |
331 | sizeof(struct reader_registry) * alloc_readers); | |
332 | alloc_readers <<= 1; | |
333 | free(oldarray); | |
334 | } | |
335 | registry[num_readers].tid = id; | |
336 | /* reference to the TLS of _this_ reader thread. */ | |
8b25e300 | 337 | registry[num_readers].urcu_reader_status = &urcu_reader_status; |
9f1621ca MD |
338 | num_readers++; |
339 | } | |
340 | ||
341 | /* | |
342 | * Never shrink (implementation limitation). | |
343 | * This is O(nb threads). Eventually use a hash table. | |
344 | */ | |
345 | static void rcu_remove_reader(pthread_t id) | |
346 | { | |
347 | struct reader_registry *index; | |
348 | ||
349 | assert(registry != NULL); | |
350 | for (index = registry; index < registry + num_readers; index++) { | |
351 | if (pthread_equal(index->tid, id)) { | |
352 | memcpy(index, ®istry[num_readers - 1], | |
353 | sizeof(struct reader_registry)); | |
354 | registry[num_readers - 1].tid = 0; | |
8b25e300 | 355 | registry[num_readers - 1].urcu_reader_status = NULL; |
9f1621ca MD |
356 | num_readers--; |
357 | return; | |
358 | } | |
359 | } | |
360 | /* Hrm not found, forgot to register ? */ | |
361 | assert(0); | |
362 | } | |
363 | ||
364 | void rcu_register_thread(void) | |
365 | { | |
366 | internal_urcu_lock(); | |
9f1621ca MD |
367 | rcu_add_reader(pthread_self()); |
368 | internal_urcu_unlock(); | |
5f373c84 | 369 | _rcu_thread_online(); |
9f1621ca MD |
370 | } |
371 | ||
372 | void rcu_unregister_thread(void) | |
373 | { | |
76f3022f MD |
374 | /* |
375 | * We have to make the thread offline otherwise we end up dealocking | |
376 | * with a waiting writer. | |
377 | */ | |
378 | _rcu_thread_offline(); | |
9f1621ca MD |
379 | internal_urcu_lock(); |
380 | rcu_remove_reader(pthread_self()); | |
381 | internal_urcu_unlock(); | |
382 | } |