Commit | Line | Data |
---|---|---|
fdee2e6d MD |
1 | /* |
2 | * urcu-bp.c | |
3 | * | |
4 | * Userspace RCU library, "bulletproof" version. | |
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 | ||
0617bf4c | 26 | #define _GNU_SOURCE |
fdee2e6d MD |
27 | #include <stdio.h> |
28 | #include <pthread.h> | |
29 | #include <signal.h> | |
30 | #include <assert.h> | |
31 | #include <stdlib.h> | |
32 | #include <string.h> | |
33 | #include <errno.h> | |
34 | #include <poll.h> | |
35 | #include <unistd.h> | |
36 | #include <sys/mman.h> | |
37 | ||
38 | #include "urcu-bp-static.h" | |
39 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ | |
40 | #include "urcu-bp.h" | |
41 | ||
42 | /* Sleep delay in us */ | |
43 | #define RCU_SLEEP_DELAY 1000 | |
44 | #define ARENA_INIT_ALLOC 16 | |
45 | ||
46 | void __attribute__((destructor)) urcu_bp_exit(void); | |
47 | ||
48 | static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; | |
49 | ||
50 | #ifdef DEBUG_YIELD | |
51 | unsigned int yield_active; | |
52 | unsigned int __thread rand_yield; | |
53 | #endif | |
54 | ||
55 | /* | |
56 | * Global grace period counter. | |
57 | * Contains the current RCU_GP_CTR_BIT. | |
58 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. | |
59 | * Written to only by writer with mutex taken. Read by both writer and readers. | |
60 | */ | |
61 | long urcu_gp_ctr = RCU_GP_COUNT; | |
62 | ||
63 | /* | |
64 | * Pointer to registry elements. Written to only by each individual reader. Read | |
65 | * by both the reader and the writers. | |
66 | */ | |
67 | struct urcu_reader __thread *urcu_reader; | |
68 | ||
69 | static LIST_HEAD(registry); | |
70 | ||
71 | struct registry_arena { | |
72 | void *p; | |
73 | size_t len; | |
74 | size_t used; | |
75 | }; | |
76 | ||
77 | static struct registry_arena registry_arena; | |
78 | ||
79 | static void rcu_gc_registry(void); | |
80 | ||
81 | static void internal_urcu_lock(void) | |
82 | { | |
83 | int ret; | |
84 | ||
85 | #ifndef DISTRUST_SIGNALS_EXTREME | |
86 | ret = pthread_mutex_lock(&urcu_mutex); | |
87 | if (ret) { | |
88 | perror("Error in pthread mutex lock"); | |
89 | exit(-1); | |
90 | } | |
91 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
92 | while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { | |
93 | if (ret != EBUSY && ret != EINTR) { | |
94 | printf("ret = %d, errno = %d\n", ret, errno); | |
95 | perror("Error in pthread mutex lock"); | |
96 | exit(-1); | |
97 | } | |
98 | if (urcu_reader.need_mb) { | |
99 | smp_mb(); | |
100 | urcu_reader.need_mb = 0; | |
101 | smp_mb(); | |
102 | } | |
103 | poll(NULL,0,10); | |
104 | } | |
105 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
106 | } | |
107 | ||
108 | static void internal_urcu_unlock(void) | |
109 | { | |
110 | int ret; | |
111 | ||
112 | ret = pthread_mutex_unlock(&urcu_mutex); | |
113 | if (ret) { | |
114 | perror("Error in pthread mutex unlock"); | |
115 | exit(-1); | |
116 | } | |
117 | } | |
118 | ||
119 | /* | |
120 | * called with urcu_mutex held. | |
121 | */ | |
122 | static void switch_next_urcu_qparity(void) | |
123 | { | |
124 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT); | |
125 | } | |
126 | ||
127 | void wait_for_quiescent_state(void) | |
128 | { | |
129 | LIST_HEAD(qsreaders); | |
130 | int wait_loops = 0; | |
131 | struct urcu_reader *index, *tmp; | |
132 | ||
133 | if (list_empty(®istry)) | |
134 | return; | |
135 | /* | |
136 | * Wait for each thread urcu_reader.ctr count to become 0. | |
137 | */ | |
138 | for (;;) { | |
139 | wait_loops++; | |
140 | list_for_each_entry_safe(index, tmp, ®istry, head) { | |
141 | if (!rcu_old_gp_ongoing(&index->ctr)) | |
142 | list_move(&index->head, &qsreaders); | |
143 | } | |
144 | ||
145 | if (list_empty(®istry)) { | |
146 | break; | |
147 | } else { | |
148 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) | |
149 | usleep(RCU_SLEEP_DELAY); | |
150 | else | |
151 | cpu_relax(); | |
152 | } | |
153 | } | |
154 | /* put back the reader list in the registry */ | |
155 | list_splice(&qsreaders, ®istry); | |
156 | } | |
157 | ||
158 | void synchronize_rcu(void) | |
159 | { | |
160 | sigset_t newmask, oldmask; | |
161 | int ret; | |
162 | ||
163 | ret = sigemptyset(&newmask); | |
164 | assert(!ret); | |
165 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
166 | assert(!ret); | |
167 | ||
168 | internal_urcu_lock(); | |
169 | ||
170 | /* Remove old registry elements */ | |
171 | rcu_gc_registry(); | |
172 | ||
173 | /* All threads should read qparity before accessing data structure | |
174 | * where new ptr points to. Must be done within internal_urcu_lock | |
175 | * because it iterates on reader threads.*/ | |
176 | /* Write new ptr before changing the qparity */ | |
177 | smp_mb(); | |
178 | ||
179 | switch_next_urcu_qparity(); /* 0 -> 1 */ | |
180 | ||
181 | /* | |
182 | * Must commit qparity update to memory before waiting for parity | |
183 | * 0 quiescent state. Failure to do so could result in the writer | |
184 | * waiting forever while new readers are always accessing data (no | |
185 | * progress). | |
186 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
187 | */ | |
188 | ||
189 | /* | |
190 | * Adding a smp_mb() which is _not_ formally required, but makes the | |
191 | * model easier to understand. It does not have a big performance impact | |
192 | * anyway, given this is the write-side. | |
193 | */ | |
194 | smp_mb(); | |
195 | ||
196 | /* | |
197 | * Wait for previous parity to be empty of readers. | |
198 | */ | |
199 | wait_for_quiescent_state(); /* Wait readers in parity 0 */ | |
200 | ||
201 | /* | |
202 | * Must finish waiting for quiescent state for parity 0 before | |
203 | * committing qparity update to memory. Failure to do so could result in | |
204 | * the writer waiting forever while new readers are always accessing | |
205 | * data (no progress). | |
206 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
207 | */ | |
208 | ||
209 | /* | |
210 | * Adding a smp_mb() which is _not_ formally required, but makes the | |
211 | * model easier to understand. It does not have a big performance impact | |
212 | * anyway, given this is the write-side. | |
213 | */ | |
214 | smp_mb(); | |
215 | ||
216 | switch_next_urcu_qparity(); /* 1 -> 0 */ | |
217 | ||
218 | /* | |
219 | * Must commit qparity update to memory before waiting for parity | |
220 | * 1 quiescent state. Failure to do so could result in the writer | |
221 | * waiting forever while new readers are always accessing data (no | |
222 | * progress). | |
223 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
224 | */ | |
225 | ||
226 | /* | |
227 | * Adding a smp_mb() which is _not_ formally required, but makes the | |
228 | * model easier to understand. It does not have a big performance impact | |
229 | * anyway, given this is the write-side. | |
230 | */ | |
231 | smp_mb(); | |
232 | ||
233 | /* | |
234 | * Wait for previous parity to be empty of readers. | |
235 | */ | |
236 | wait_for_quiescent_state(); /* Wait readers in parity 1 */ | |
237 | ||
238 | /* Finish waiting for reader threads before letting the old ptr being | |
239 | * freed. Must be done within internal_urcu_lock because it iterates on | |
240 | * reader threads. */ | |
241 | smp_mb(); | |
242 | ||
243 | internal_urcu_unlock(); | |
244 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
245 | assert(!ret); | |
246 | } | |
247 | ||
248 | /* | |
249 | * library wrappers to be used by non-LGPL compatible source code. | |
250 | */ | |
251 | ||
252 | void rcu_read_lock(void) | |
253 | { | |
254 | _rcu_read_lock(); | |
255 | } | |
256 | ||
257 | void rcu_read_unlock(void) | |
258 | { | |
259 | _rcu_read_unlock(); | |
260 | } | |
261 | ||
262 | /* | |
263 | * only grow for now. | |
264 | */ | |
265 | static void resize_arena(struct registry_arena *arena, size_t len) | |
266 | { | |
267 | void *new_arena; | |
268 | ||
0617bf4c MD |
269 | if (!arena->p) |
270 | new_arena = mmap(arena->p, len, | |
271 | PROT_READ | PROT_WRITE, | |
272 | MAP_ANONYMOUS | MAP_PRIVATE, | |
273 | -1, 0); | |
274 | else | |
275 | new_arena = mremap(arena->p, arena->len, | |
276 | len, MREMAP_MAYMOVE); | |
277 | assert(new_arena != MAP_FAILED); | |
278 | ||
fdee2e6d MD |
279 | /* |
280 | * re-used the same region ? | |
281 | */ | |
282 | if (new_arena == arena->p) | |
283 | return; | |
284 | ||
285 | memcpy(new_arena, arena->p, arena->len); | |
286 | bzero(new_arena + arena->len, len - arena->len); | |
287 | arena->p = new_arena; | |
288 | } | |
289 | ||
290 | /* Called with signals off and mutex locked */ | |
291 | static void add_thread(void) | |
292 | { | |
293 | struct urcu_reader *urcu_reader_reg; | |
294 | ||
295 | if (registry_arena.len | |
296 | < registry_arena.used + sizeof(struct urcu_reader)) | |
297 | resize_arena(®istry_arena, | |
298 | max(registry_arena.len << 1, ARENA_INIT_ALLOC)); | |
299 | /* | |
300 | * Find a free spot. | |
301 | */ | |
302 | for (urcu_reader_reg = registry_arena.p; | |
303 | (void *)urcu_reader_reg < registry_arena.p + registry_arena.len; | |
304 | urcu_reader_reg++) { | |
305 | if (!urcu_reader_reg->alloc) | |
306 | break; | |
307 | } | |
308 | urcu_reader_reg->alloc = 1; | |
309 | registry_arena.used += sizeof(struct urcu_reader); | |
310 | ||
311 | /* Add to registry */ | |
312 | urcu_reader_reg->tid = pthread_self(); | |
313 | assert(urcu_reader_reg->ctr == 0); | |
314 | list_add(&urcu_reader_reg->head, ®istry); | |
315 | urcu_reader = urcu_reader_reg; | |
316 | } | |
317 | ||
318 | /* Called with signals off and mutex locked */ | |
319 | static void rcu_gc_registry(void) | |
320 | { | |
321 | struct urcu_reader *urcu_reader_reg; | |
322 | pthread_t tid; | |
323 | int ret; | |
324 | ||
325 | for (urcu_reader_reg = registry_arena.p; | |
326 | (void *)urcu_reader_reg < registry_arena.p + registry_arena.len; | |
327 | urcu_reader_reg++) { | |
328 | if (!urcu_reader_reg->alloc) | |
329 | continue; | |
330 | tid = urcu_reader_reg->tid; | |
331 | ret = pthread_kill(tid, 0); | |
332 | assert(ret != EINVAL); | |
333 | if (ret == ESRCH) { | |
334 | list_del(&urcu_reader_reg->head); | |
335 | urcu_reader_reg->alloc = 0; | |
336 | registry_arena.used -= sizeof(struct urcu_reader); | |
337 | } | |
338 | } | |
339 | } | |
340 | ||
341 | /* Disable signals, take mutex, add to registry */ | |
342 | void rcu_bp_register(void) | |
343 | { | |
344 | sigset_t newmask, oldmask; | |
345 | int ret; | |
346 | ||
347 | ret = sigemptyset(&newmask); | |
348 | assert(!ret); | |
349 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
350 | assert(!ret); | |
351 | ||
352 | /* | |
353 | * Check if a signal concurrently registered our thread since | |
354 | * the check in rcu_read_lock(). */ | |
355 | if (urcu_reader) | |
356 | goto end; | |
357 | ||
358 | internal_urcu_lock(); | |
359 | add_thread(); | |
360 | internal_urcu_unlock(); | |
361 | end: | |
362 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
363 | assert(!ret); | |
364 | } | |
365 | ||
366 | void urcu_bp_exit() | |
367 | { | |
368 | munmap(registry_arena.p, registry_arena.len); | |
369 | } |