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> | |
34 | ||
727f819d | 35 | #define BUILD_QSBR_LIB |
7ac06cef | 36 | #include "urcu-qsbr-static.h" |
9f1621ca | 37 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
7ac06cef | 38 | #include "urcu-qsbr.h" |
9f1621ca | 39 | |
81d1f1f5 | 40 | static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
9f1621ca MD |
41 | |
42 | /* | |
43 | * Global grace period counter. | |
44 | */ | |
ac258107 | 45 | unsigned long urcu_gp_ctr = RCU_GP_ONLINE; |
9f1621ca MD |
46 | |
47 | /* | |
48 | * Written to only by each individual reader. Read by both the reader and the | |
49 | * writers. | |
50 | */ | |
8b25e300 | 51 | struct urcu_reader_status __thread urcu_reader_status; |
9f1621ca MD |
52 | |
53 | /* Thread IDs of registered readers */ | |
54 | #define INIT_NUM_THREADS 4 | |
55 | ||
56 | struct reader_registry { | |
57 | pthread_t tid; | |
8b25e300 | 58 | struct urcu_reader_status *urcu_reader_status; |
9f1621ca MD |
59 | }; |
60 | ||
61 | #ifdef DEBUG_YIELD | |
62 | unsigned int yield_active; | |
63 | unsigned int __thread rand_yield; | |
64 | #endif | |
65 | ||
66 | static struct reader_registry *registry; | |
9f1621ca MD |
67 | static int num_readers, alloc_readers; |
68 | ||
90c1618a | 69 | static void internal_urcu_lock(void) |
9f1621ca MD |
70 | { |
71 | int ret; | |
72 | ||
73 | #ifndef DISTRUST_SIGNALS_EXTREME | |
74 | ret = pthread_mutex_lock(&urcu_mutex); | |
75 | if (ret) { | |
76 | perror("Error in pthread mutex lock"); | |
77 | exit(-1); | |
78 | } | |
79 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
80 | while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { | |
81 | if (ret != EBUSY && ret != EINTR) { | |
82 | printf("ret = %d, errno = %d\n", ret, errno); | |
83 | perror("Error in pthread mutex lock"); | |
84 | exit(-1); | |
85 | } | |
9f1621ca MD |
86 | poll(NULL,0,10); |
87 | } | |
88 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
89 | } | |
90 | ||
90c1618a | 91 | static void internal_urcu_unlock(void) |
9f1621ca MD |
92 | { |
93 | int ret; | |
94 | ||
95 | ret = pthread_mutex_unlock(&urcu_mutex); | |
96 | if (ret) { | |
97 | perror("Error in pthread mutex unlock"); | |
98 | exit(-1); | |
99 | } | |
100 | } | |
101 | ||
bc6c15bb MD |
102 | /* |
103 | * synchronize_rcu() waiting. Single thread. | |
104 | */ | |
90c1618a | 105 | static void wait_for_quiescent_state(void) |
9f1621ca MD |
106 | { |
107 | struct reader_registry *index; | |
108 | ||
109 | if (!registry) | |
110 | return; | |
111 | /* | |
8b25e300 | 112 | * Wait for each thread rcu_reader qs_gp count to become 0. |
9f1621ca MD |
113 | */ |
114 | for (index = registry; index < registry + num_readers; index++) { | |
bc6c15bb MD |
115 | int wait_loops = 0; |
116 | ||
2b7f706f MD |
117 | if (likely(!rcu_gp_ongoing(&index->urcu_reader_status->qs_gp))) |
118 | continue; | |
119 | ||
8b25e300 MD |
120 | index->urcu_reader_status->gp_waiting = 1; |
121 | while (rcu_gp_ongoing(&index->urcu_reader_status->qs_gp)) { | |
bc6c15bb | 122 | if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) { |
ae62b5e8 | 123 | sched_yield(); /* ideally sched_yield_to() */ |
b3c4dd1a | 124 | wait_loops = 0; |
bc6c15bb | 125 | } else { |
9f1621ca | 126 | #ifndef HAS_INCOHERENT_CACHES |
bc6c15bb | 127 | cpu_relax(); |
9f1621ca | 128 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
bc6c15bb | 129 | smp_mb(); |
9f1621ca | 130 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
bc6c15bb MD |
131 | } |
132 | } | |
8b25e300 | 133 | index->urcu_reader_status->gp_waiting = 0; |
9f1621ca MD |
134 | } |
135 | } | |
136 | ||
47d2f29e MD |
137 | /* |
138 | * Using a two-subphases algorithm for architectures with smaller than 64-bit | |
139 | * long-size to ensure we do not encounter an overflow bug. | |
140 | */ | |
141 | ||
142 | #if (BITS_PER_LONG < 64) | |
143 | /* | |
144 | * called with urcu_mutex held. | |
145 | */ | |
146 | static void switch_next_urcu_qparity(void) | |
147 | { | |
148 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR); | |
149 | } | |
150 | ||
151 | void synchronize_rcu(void) | |
152 | { | |
bc49c323 MD |
153 | unsigned long was_online; |
154 | ||
8b25e300 | 155 | was_online = urcu_reader_status.qs_gp; |
bc49c323 | 156 | |
47d2f29e MD |
157 | /* All threads should read qparity before accessing data structure |
158 | * where new ptr points to. | |
159 | */ | |
160 | /* Write new ptr before changing the qparity */ | |
161 | smp_mb(); | |
162 | ||
bc49c323 MD |
163 | /* |
164 | * Mark the writer thread offline to make sure we don't wait for | |
165 | * our own quiescent state. This allows using synchronize_rcu() in | |
166 | * threads registered as readers. | |
167 | */ | |
168 | if (was_online) | |
8b25e300 | 169 | STORE_SHARED(urcu_reader_status.qs_gp, 0); |
bc49c323 | 170 | |
47d2f29e MD |
171 | internal_urcu_lock(); |
172 | ||
ae62b5e8 MD |
173 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
174 | ||
47d2f29e MD |
175 | switch_next_urcu_qparity(); /* 0 -> 1 */ |
176 | ||
177 | /* | |
178 | * Must commit qparity update to memory before waiting for parity | |
179 | * 0 quiescent state. Failure to do so could result in the writer | |
180 | * waiting forever while new readers are always accessing data (no | |
181 | * progress). | |
182 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
183 | */ | |
184 | ||
185 | /* | |
186 | * Wait for previous parity to be empty of readers. | |
187 | */ | |
188 | wait_for_quiescent_state(); /* Wait readers in parity 0 */ | |
189 | ||
190 | /* | |
191 | * Must finish waiting for quiescent state for parity 0 before | |
192 | * committing qparity update to memory. Failure to do so could result in | |
193 | * the writer waiting forever while new readers are always accessing | |
194 | * data (no progress). | |
195 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
196 | */ | |
197 | ||
198 | switch_next_urcu_qparity(); /* 1 -> 0 */ | |
199 | ||
200 | /* | |
201 | * Must commit qparity update to memory before waiting for parity | |
202 | * 1 quiescent state. Failure to do so could result in the writer | |
203 | * waiting forever while new readers are always accessing data (no | |
204 | * progress). | |
205 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
206 | */ | |
207 | ||
208 | /* | |
209 | * Wait for previous parity to be empty of readers. | |
210 | */ | |
211 | wait_for_quiescent_state(); /* Wait readers in parity 1 */ | |
212 | ||
ae62b5e8 MD |
213 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
214 | ||
47d2f29e MD |
215 | internal_urcu_unlock(); |
216 | ||
bc49c323 MD |
217 | /* |
218 | * Finish waiting for reader threads before letting the old ptr being | |
47d2f29e MD |
219 | * freed. |
220 | */ | |
bc49c323 | 221 | if (was_online) |
8b25e300 MD |
222 | _STORE_SHARED(urcu_reader_status.qs_gp, |
223 | LOAD_SHARED(urcu_gp_ctr)); | |
47d2f29e MD |
224 | smp_mb(); |
225 | } | |
226 | #else /* !(BITS_PER_LONG < 64) */ | |
9f1621ca MD |
227 | void synchronize_rcu(void) |
228 | { | |
f0f7dbdd | 229 | unsigned long was_online; |
ff2f67a0 | 230 | |
8b25e300 | 231 | was_online = urcu_reader_status.qs_gp; |
ff2f67a0 MD |
232 | |
233 | /* | |
234 | * Mark the writer thread offline to make sure we don't wait for | |
235 | * our own quiescent state. This allows using synchronize_rcu() in | |
236 | * threads registered as readers. | |
237 | */ | |
8f50d1ce | 238 | smp_mb(); |
ff2f67a0 | 239 | if (was_online) |
8b25e300 | 240 | STORE_SHARED(urcu_reader_status.qs_gp, 0); |
ff2f67a0 | 241 | |
4e27f058 | 242 | internal_urcu_lock(); |
ae62b5e8 | 243 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
55570466 | 244 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR); |
9f1621ca | 245 | wait_for_quiescent_state(); |
ae62b5e8 | 246 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
9f1621ca | 247 | internal_urcu_unlock(); |
ff2f67a0 MD |
248 | |
249 | if (was_online) | |
8b25e300 MD |
250 | _STORE_SHARED(urcu_reader_status.qs_gp, |
251 | LOAD_SHARED(urcu_gp_ctr)); | |
8f50d1ce | 252 | smp_mb(); |
9f1621ca | 253 | } |
47d2f29e | 254 | #endif /* !(BITS_PER_LONG < 64) */ |
9f1621ca MD |
255 | |
256 | /* | |
257 | * library wrappers to be used by non-LGPL compatible source code. | |
258 | */ | |
259 | ||
260 | void rcu_read_lock(void) | |
261 | { | |
262 | _rcu_read_lock(); | |
263 | } | |
264 | ||
265 | void rcu_read_unlock(void) | |
266 | { | |
267 | _rcu_read_unlock(); | |
268 | } | |
269 | ||
270 | void *rcu_dereference(void *p) | |
271 | { | |
272 | return _rcu_dereference(p); | |
273 | } | |
274 | ||
275 | void *rcu_assign_pointer_sym(void **p, void *v) | |
276 | { | |
277 | wmb(); | |
278 | return STORE_SHARED(p, v); | |
279 | } | |
280 | ||
4d1ce26f MD |
281 | void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
282 | { | |
283 | wmb(); | |
284 | return cmpxchg(p, old, _new); | |
285 | } | |
286 | ||
9f1621ca MD |
287 | void *rcu_xchg_pointer_sym(void **p, void *v) |
288 | { | |
289 | wmb(); | |
290 | return xchg(p, v); | |
291 | } | |
292 | ||
293 | void *rcu_publish_content_sym(void **p, void *v) | |
294 | { | |
295 | void *oldptr; | |
296 | ||
297 | oldptr = _rcu_xchg_pointer(p, v); | |
298 | synchronize_rcu(); | |
299 | return oldptr; | |
300 | } | |
301 | ||
7ac06cef MD |
302 | void rcu_quiescent_state(void) |
303 | { | |
304 | _rcu_quiescent_state(); | |
305 | } | |
306 | ||
307 | void rcu_thread_offline(void) | |
308 | { | |
309 | _rcu_thread_offline(); | |
310 | } | |
311 | ||
312 | void rcu_thread_online(void) | |
313 | { | |
314 | _rcu_thread_online(); | |
315 | } | |
316 | ||
9f1621ca MD |
317 | static void rcu_add_reader(pthread_t id) |
318 | { | |
319 | struct reader_registry *oldarray; | |
320 | ||
321 | if (!registry) { | |
322 | alloc_readers = INIT_NUM_THREADS; | |
323 | num_readers = 0; | |
324 | registry = | |
325 | malloc(sizeof(struct reader_registry) * alloc_readers); | |
326 | } | |
327 | if (alloc_readers < num_readers + 1) { | |
328 | oldarray = registry; | |
329 | registry = malloc(sizeof(struct reader_registry) | |
330 | * (alloc_readers << 1)); | |
331 | memcpy(registry, oldarray, | |
332 | sizeof(struct reader_registry) * alloc_readers); | |
333 | alloc_readers <<= 1; | |
334 | free(oldarray); | |
335 | } | |
336 | registry[num_readers].tid = id; | |
337 | /* reference to the TLS of _this_ reader thread. */ | |
8b25e300 | 338 | registry[num_readers].urcu_reader_status = &urcu_reader_status; |
9f1621ca MD |
339 | num_readers++; |
340 | } | |
341 | ||
342 | /* | |
343 | * Never shrink (implementation limitation). | |
344 | * This is O(nb threads). Eventually use a hash table. | |
345 | */ | |
346 | static void rcu_remove_reader(pthread_t id) | |
347 | { | |
348 | struct reader_registry *index; | |
349 | ||
350 | assert(registry != NULL); | |
351 | for (index = registry; index < registry + num_readers; index++) { | |
352 | if (pthread_equal(index->tid, id)) { | |
353 | memcpy(index, ®istry[num_readers - 1], | |
354 | sizeof(struct reader_registry)); | |
355 | registry[num_readers - 1].tid = 0; | |
8b25e300 | 356 | registry[num_readers - 1].urcu_reader_status = NULL; |
9f1621ca MD |
357 | num_readers--; |
358 | return; | |
359 | } | |
360 | } | |
361 | /* Hrm not found, forgot to register ? */ | |
362 | assert(0); | |
363 | } | |
364 | ||
365 | void rcu_register_thread(void) | |
366 | { | |
367 | internal_urcu_lock(); | |
9f1621ca MD |
368 | rcu_add_reader(pthread_self()); |
369 | internal_urcu_unlock(); | |
5f373c84 | 370 | _rcu_thread_online(); |
9f1621ca MD |
371 | } |
372 | ||
373 | void rcu_unregister_thread(void) | |
374 | { | |
76f3022f MD |
375 | /* |
376 | * We have to make the thread offline otherwise we end up dealocking | |
377 | * with a waiting writer. | |
378 | */ | |
379 | _rcu_thread_offline(); | |
9f1621ca MD |
380 | internal_urcu_lock(); |
381 | rcu_remove_reader(pthread_self()); | |
382 | internal_urcu_unlock(); | |
383 | } |