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 | ||
69915897 MD |
118 | if (likely(!rcu_gp_ongoing(&index->urcu_reader_status->qs_gp))) |
119 | continue; | |
120 | ||
8b25e300 MD |
121 | index->urcu_reader_status->gp_waiting = 1; |
122 | while (rcu_gp_ongoing(&index->urcu_reader_status->qs_gp)) { | |
bc6c15bb | 123 | if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) { |
ee72ef6e | 124 | /* adapted wait time, in us */ |
539718ae | 125 | usleep(LOAD_SHARED(index->urcu_reader_status->qs_time_delta_usec) / 4); |
b3c4dd1a | 126 | wait_loops = 0; |
bc6c15bb | 127 | } else { |
9f1621ca | 128 | #ifndef HAS_INCOHERENT_CACHES |
bc6c15bb | 129 | cpu_relax(); |
9f1621ca | 130 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
bc6c15bb | 131 | smp_mb(); |
9f1621ca | 132 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
bc6c15bb MD |
133 | } |
134 | } | |
8b25e300 | 135 | index->urcu_reader_status->gp_waiting = 0; |
9f1621ca MD |
136 | } |
137 | } | |
138 | ||
47d2f29e MD |
139 | /* |
140 | * Using a two-subphases algorithm for architectures with smaller than 64-bit | |
141 | * long-size to ensure we do not encounter an overflow bug. | |
142 | */ | |
143 | ||
144 | #if (BITS_PER_LONG < 64) | |
145 | /* | |
146 | * called with urcu_mutex held. | |
147 | */ | |
148 | static void switch_next_urcu_qparity(void) | |
149 | { | |
150 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR); | |
151 | } | |
152 | ||
153 | void synchronize_rcu(void) | |
154 | { | |
bc49c323 MD |
155 | unsigned long was_online; |
156 | ||
8b25e300 | 157 | was_online = urcu_reader_status.qs_gp; |
bc49c323 | 158 | |
47d2f29e MD |
159 | /* All threads should read qparity before accessing data structure |
160 | * where new ptr points to. | |
161 | */ | |
162 | /* Write new ptr before changing the qparity */ | |
163 | smp_mb(); | |
164 | ||
bc49c323 MD |
165 | /* |
166 | * Mark the writer thread offline to make sure we don't wait for | |
167 | * our own quiescent state. This allows using synchronize_rcu() in | |
168 | * threads registered as readers. | |
169 | */ | |
170 | if (was_online) | |
8b25e300 | 171 | STORE_SHARED(urcu_reader_status.qs_gp, 0); |
bc49c323 | 172 | |
47d2f29e MD |
173 | internal_urcu_lock(); |
174 | ||
ae62b5e8 MD |
175 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
176 | ||
47d2f29e MD |
177 | switch_next_urcu_qparity(); /* 0 -> 1 */ |
178 | ||
179 | /* | |
180 | * Must commit qparity update to memory before waiting for parity | |
181 | * 0 quiescent state. Failure to do so could result in the writer | |
182 | * waiting forever while new readers are always accessing data (no | |
183 | * progress). | |
184 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
185 | */ | |
186 | ||
187 | /* | |
188 | * Wait for previous parity to be empty of readers. | |
189 | */ | |
190 | wait_for_quiescent_state(); /* Wait readers in parity 0 */ | |
191 | ||
192 | /* | |
193 | * Must finish waiting for quiescent state for parity 0 before | |
194 | * committing qparity update to memory. Failure to do so could result in | |
195 | * the writer waiting forever while new readers are always accessing | |
196 | * data (no progress). | |
197 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
198 | */ | |
199 | ||
200 | switch_next_urcu_qparity(); /* 1 -> 0 */ | |
201 | ||
202 | /* | |
203 | * Must commit qparity update to memory before waiting for parity | |
204 | * 1 quiescent state. Failure to do so could result in the writer | |
205 | * waiting forever while new readers are always accessing data (no | |
206 | * progress). | |
207 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
208 | */ | |
209 | ||
210 | /* | |
211 | * Wait for previous parity to be empty of readers. | |
212 | */ | |
213 | wait_for_quiescent_state(); /* Wait readers in parity 1 */ | |
214 | ||
ae62b5e8 MD |
215 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
216 | ||
47d2f29e MD |
217 | internal_urcu_unlock(); |
218 | ||
bc49c323 MD |
219 | /* |
220 | * Finish waiting for reader threads before letting the old ptr being | |
47d2f29e MD |
221 | * freed. |
222 | */ | |
bc49c323 | 223 | if (was_online) |
8b25e300 MD |
224 | _STORE_SHARED(urcu_reader_status.qs_gp, |
225 | LOAD_SHARED(urcu_gp_ctr)); | |
47d2f29e MD |
226 | smp_mb(); |
227 | } | |
228 | #else /* !(BITS_PER_LONG < 64) */ | |
9f1621ca MD |
229 | void synchronize_rcu(void) |
230 | { | |
f0f7dbdd | 231 | unsigned long was_online; |
ff2f67a0 | 232 | |
8b25e300 | 233 | was_online = urcu_reader_status.qs_gp; |
ff2f67a0 MD |
234 | |
235 | /* | |
236 | * Mark the writer thread offline to make sure we don't wait for | |
237 | * our own quiescent state. This allows using synchronize_rcu() in | |
238 | * threads registered as readers. | |
239 | */ | |
8f50d1ce | 240 | smp_mb(); |
ff2f67a0 | 241 | if (was_online) |
8b25e300 | 242 | STORE_SHARED(urcu_reader_status.qs_gp, 0); |
ff2f67a0 | 243 | |
4e27f058 | 244 | internal_urcu_lock(); |
ae62b5e8 | 245 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
55570466 | 246 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR); |
9f1621ca | 247 | wait_for_quiescent_state(); |
ae62b5e8 | 248 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING); |
9f1621ca | 249 | internal_urcu_unlock(); |
ff2f67a0 MD |
250 | |
251 | if (was_online) | |
8b25e300 MD |
252 | _STORE_SHARED(urcu_reader_status.qs_gp, |
253 | LOAD_SHARED(urcu_gp_ctr)); | |
8f50d1ce | 254 | smp_mb(); |
9f1621ca | 255 | } |
47d2f29e | 256 | #endif /* !(BITS_PER_LONG < 64) */ |
9f1621ca MD |
257 | |
258 | /* | |
259 | * library wrappers to be used by non-LGPL compatible source code. | |
260 | */ | |
261 | ||
262 | void rcu_read_lock(void) | |
263 | { | |
264 | _rcu_read_lock(); | |
265 | } | |
266 | ||
267 | void rcu_read_unlock(void) | |
268 | { | |
269 | _rcu_read_unlock(); | |
270 | } | |
271 | ||
272 | void *rcu_dereference(void *p) | |
273 | { | |
274 | return _rcu_dereference(p); | |
275 | } | |
276 | ||
277 | void *rcu_assign_pointer_sym(void **p, void *v) | |
278 | { | |
279 | wmb(); | |
280 | return STORE_SHARED(p, v); | |
281 | } | |
282 | ||
4d1ce26f MD |
283 | void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
284 | { | |
285 | wmb(); | |
286 | return cmpxchg(p, old, _new); | |
287 | } | |
288 | ||
9f1621ca MD |
289 | void *rcu_xchg_pointer_sym(void **p, void *v) |
290 | { | |
291 | wmb(); | |
292 | return xchg(p, v); | |
293 | } | |
294 | ||
295 | void *rcu_publish_content_sym(void **p, void *v) | |
296 | { | |
297 | void *oldptr; | |
298 | ||
299 | oldptr = _rcu_xchg_pointer(p, v); | |
300 | synchronize_rcu(); | |
301 | return oldptr; | |
302 | } | |
303 | ||
7ac06cef MD |
304 | void rcu_quiescent_state(void) |
305 | { | |
306 | _rcu_quiescent_state(); | |
307 | } | |
308 | ||
309 | void rcu_thread_offline(void) | |
310 | { | |
311 | _rcu_thread_offline(); | |
312 | } | |
313 | ||
314 | void rcu_thread_online(void) | |
315 | { | |
316 | _rcu_thread_online(); | |
317 | } | |
318 | ||
9f1621ca MD |
319 | static void rcu_add_reader(pthread_t id) |
320 | { | |
321 | struct reader_registry *oldarray; | |
322 | ||
323 | if (!registry) { | |
324 | alloc_readers = INIT_NUM_THREADS; | |
325 | num_readers = 0; | |
326 | registry = | |
327 | malloc(sizeof(struct reader_registry) * alloc_readers); | |
328 | } | |
329 | if (alloc_readers < num_readers + 1) { | |
330 | oldarray = registry; | |
331 | registry = malloc(sizeof(struct reader_registry) | |
332 | * (alloc_readers << 1)); | |
333 | memcpy(registry, oldarray, | |
334 | sizeof(struct reader_registry) * alloc_readers); | |
335 | alloc_readers <<= 1; | |
336 | free(oldarray); | |
337 | } | |
338 | registry[num_readers].tid = id; | |
339 | /* reference to the TLS of _this_ reader thread. */ | |
8b25e300 | 340 | registry[num_readers].urcu_reader_status = &urcu_reader_status; |
9f1621ca MD |
341 | num_readers++; |
342 | } | |
343 | ||
344 | /* | |
345 | * Never shrink (implementation limitation). | |
346 | * This is O(nb threads). Eventually use a hash table. | |
347 | */ | |
348 | static void rcu_remove_reader(pthread_t id) | |
349 | { | |
350 | struct reader_registry *index; | |
351 | ||
352 | assert(registry != NULL); | |
353 | for (index = registry; index < registry + num_readers; index++) { | |
354 | if (pthread_equal(index->tid, id)) { | |
355 | memcpy(index, ®istry[num_readers - 1], | |
356 | sizeof(struct reader_registry)); | |
357 | registry[num_readers - 1].tid = 0; | |
8b25e300 | 358 | registry[num_readers - 1].urcu_reader_status = NULL; |
9f1621ca MD |
359 | num_readers--; |
360 | return; | |
361 | } | |
362 | } | |
363 | /* Hrm not found, forgot to register ? */ | |
364 | assert(0); | |
365 | } | |
366 | ||
367 | void rcu_register_thread(void) | |
368 | { | |
369 | internal_urcu_lock(); | |
9f1621ca MD |
370 | rcu_add_reader(pthread_self()); |
371 | internal_urcu_unlock(); | |
5f373c84 | 372 | _rcu_thread_online(); |
9f1621ca MD |
373 | } |
374 | ||
375 | void rcu_unregister_thread(void) | |
376 | { | |
76f3022f MD |
377 | /* |
378 | * We have to make the thread offline otherwise we end up dealocking | |
379 | * with a waiting writer. | |
380 | */ | |
381 | _rcu_thread_offline(); | |
9f1621ca MD |
382 | internal_urcu_lock(); |
383 | rcu_remove_reader(pthread_self()); | |
384 | internal_urcu_unlock(); | |
385 | } |