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 | ||
02be5561 | 46 | void __attribute__((destructor)) rcu_bp_exit(void); |
fdee2e6d | 47 | |
6abb4bd5 | 48 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
fdee2e6d MD |
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. | |
02be5561 | 57 | * Contains the current RCU_GP_CTR_PHASE. |
fdee2e6d MD |
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 | */ | |
02be5561 | 61 | long rcu_gp_ctr = RCU_GP_COUNT; |
fdee2e6d MD |
62 | |
63 | /* | |
64 | * Pointer to registry elements. Written to only by each individual reader. Read | |
65 | * by both the reader and the writers. | |
66 | */ | |
02be5561 | 67 | struct rcu_reader __thread *rcu_reader; |
fdee2e6d MD |
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 | ||
6abb4bd5 | 81 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
82 | { |
83 | int ret; | |
84 | ||
85 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 86 | ret = pthread_mutex_lock(mutex); |
fdee2e6d MD |
87 | if (ret) { |
88 | perror("Error in pthread mutex lock"); | |
89 | exit(-1); | |
90 | } | |
91 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
6abb4bd5 | 92 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
fdee2e6d MD |
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 | } | |
02be5561 | 98 | if (rcu_reader.need_mb) { |
fdee2e6d | 99 | smp_mb(); |
02be5561 | 100 | rcu_reader.need_mb = 0; |
fdee2e6d MD |
101 | smp_mb(); |
102 | } | |
103 | poll(NULL,0,10); | |
104 | } | |
105 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
106 | } | |
107 | ||
6abb4bd5 | 108 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
109 | { |
110 | int ret; | |
111 | ||
6abb4bd5 | 112 | ret = pthread_mutex_unlock(mutex); |
fdee2e6d MD |
113 | if (ret) { |
114 | perror("Error in pthread mutex unlock"); | |
115 | exit(-1); | |
116 | } | |
117 | } | |
118 | ||
2dfb8b5e | 119 | void update_counter_and_wait(void) |
fdee2e6d MD |
120 | { |
121 | LIST_HEAD(qsreaders); | |
122 | int wait_loops = 0; | |
02be5561 | 123 | struct rcu_reader *index, *tmp; |
fdee2e6d | 124 | |
32c15e4e | 125 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
2dfb8b5e MD |
126 | STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); |
127 | ||
128 | /* | |
129 | * Must commit qparity update to memory before waiting for other parity | |
130 | * quiescent state. Failure to do so could result in the writer waiting | |
131 | * forever while new readers are always accessing data (no progress). | |
132 | * Ensured by STORE_SHARED and LOAD_SHARED. | |
133 | */ | |
134 | ||
135 | /* | |
136 | * Adding a smp_mb() which is _not_ formally required, but makes the | |
137 | * model easier to understand. It does not have a big performance impact | |
138 | * anyway, given this is the write-side. | |
139 | */ | |
140 | smp_mb(); | |
141 | ||
fdee2e6d | 142 | /* |
02be5561 | 143 | * Wait for each thread rcu_reader.ctr count to become 0. |
fdee2e6d MD |
144 | */ |
145 | for (;;) { | |
146 | wait_loops++; | |
147 | list_for_each_entry_safe(index, tmp, ®istry, head) { | |
148 | if (!rcu_old_gp_ongoing(&index->ctr)) | |
149 | list_move(&index->head, &qsreaders); | |
150 | } | |
151 | ||
152 | if (list_empty(®istry)) { | |
153 | break; | |
154 | } else { | |
155 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) | |
156 | usleep(RCU_SLEEP_DELAY); | |
157 | else | |
158 | cpu_relax(); | |
159 | } | |
160 | } | |
161 | /* put back the reader list in the registry */ | |
162 | list_splice(&qsreaders, ®istry); | |
163 | } | |
164 | ||
165 | void synchronize_rcu(void) | |
166 | { | |
167 | sigset_t newmask, oldmask; | |
168 | int ret; | |
169 | ||
170 | ret = sigemptyset(&newmask); | |
171 | assert(!ret); | |
172 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
173 | assert(!ret); | |
174 | ||
6abb4bd5 | 175 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 176 | |
2dfb8b5e MD |
177 | if (list_empty(®istry)) |
178 | goto out; | |
fdee2e6d MD |
179 | |
180 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 181 | * where new ptr points to. */ |
fdee2e6d MD |
182 | /* Write new ptr before changing the qparity */ |
183 | smp_mb(); | |
184 | ||
2dfb8b5e MD |
185 | /* Remove old registry elements */ |
186 | rcu_gc_registry(); | |
fdee2e6d MD |
187 | |
188 | /* | |
189 | * Wait for previous parity to be empty of readers. | |
190 | */ | |
2dfb8b5e | 191 | update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ |
fdee2e6d MD |
192 | |
193 | /* | |
194 | * Adding a smp_mb() which is _not_ formally required, but makes the | |
195 | * model easier to understand. It does not have a big performance impact | |
196 | * anyway, given this is the write-side. | |
197 | */ | |
198 | smp_mb(); | |
199 | ||
fdee2e6d | 200 | /* |
2dfb8b5e | 201 | * Wait for previous parity to be empty of readers. |
fdee2e6d | 202 | */ |
2dfb8b5e | 203 | update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ |
fdee2e6d MD |
204 | |
205 | /* | |
2dfb8b5e MD |
206 | * Finish waiting for reader threads before letting the old ptr being |
207 | * freed. | |
fdee2e6d | 208 | */ |
fdee2e6d | 209 | smp_mb(); |
2dfb8b5e | 210 | out: |
6abb4bd5 | 211 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
212 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
213 | assert(!ret); | |
214 | } | |
215 | ||
216 | /* | |
217 | * library wrappers to be used by non-LGPL compatible source code. | |
218 | */ | |
219 | ||
220 | void rcu_read_lock(void) | |
221 | { | |
222 | _rcu_read_lock(); | |
223 | } | |
224 | ||
225 | void rcu_read_unlock(void) | |
226 | { | |
227 | _rcu_read_unlock(); | |
228 | } | |
229 | ||
230 | /* | |
231 | * only grow for now. | |
232 | */ | |
233 | static void resize_arena(struct registry_arena *arena, size_t len) | |
234 | { | |
235 | void *new_arena; | |
236 | ||
0617bf4c MD |
237 | if (!arena->p) |
238 | new_arena = mmap(arena->p, len, | |
239 | PROT_READ | PROT_WRITE, | |
240 | MAP_ANONYMOUS | MAP_PRIVATE, | |
241 | -1, 0); | |
242 | else | |
243 | new_arena = mremap(arena->p, arena->len, | |
244 | len, MREMAP_MAYMOVE); | |
245 | assert(new_arena != MAP_FAILED); | |
246 | ||
fdee2e6d MD |
247 | /* |
248 | * re-used the same region ? | |
249 | */ | |
250 | if (new_arena == arena->p) | |
251 | return; | |
252 | ||
253 | memcpy(new_arena, arena->p, arena->len); | |
254 | bzero(new_arena + arena->len, len - arena->len); | |
255 | arena->p = new_arena; | |
256 | } | |
257 | ||
258 | /* Called with signals off and mutex locked */ | |
259 | static void add_thread(void) | |
260 | { | |
02be5561 | 261 | struct rcu_reader *rcu_reader_reg; |
fdee2e6d MD |
262 | |
263 | if (registry_arena.len | |
02be5561 | 264 | < registry_arena.used + sizeof(struct rcu_reader)) |
fdee2e6d MD |
265 | resize_arena(®istry_arena, |
266 | max(registry_arena.len << 1, ARENA_INIT_ALLOC)); | |
267 | /* | |
268 | * Find a free spot. | |
269 | */ | |
02be5561 MD |
270 | for (rcu_reader_reg = registry_arena.p; |
271 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; | |
272 | rcu_reader_reg++) { | |
273 | if (!rcu_reader_reg->alloc) | |
fdee2e6d MD |
274 | break; |
275 | } | |
02be5561 MD |
276 | rcu_reader_reg->alloc = 1; |
277 | registry_arena.used += sizeof(struct rcu_reader); | |
fdee2e6d MD |
278 | |
279 | /* Add to registry */ | |
02be5561 MD |
280 | rcu_reader_reg->tid = pthread_self(); |
281 | assert(rcu_reader_reg->ctr == 0); | |
282 | list_add(&rcu_reader_reg->head, ®istry); | |
283 | rcu_reader = rcu_reader_reg; | |
fdee2e6d MD |
284 | } |
285 | ||
286 | /* Called with signals off and mutex locked */ | |
287 | static void rcu_gc_registry(void) | |
288 | { | |
02be5561 | 289 | struct rcu_reader *rcu_reader_reg; |
fdee2e6d MD |
290 | pthread_t tid; |
291 | int ret; | |
292 | ||
02be5561 MD |
293 | for (rcu_reader_reg = registry_arena.p; |
294 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; | |
295 | rcu_reader_reg++) { | |
296 | if (!rcu_reader_reg->alloc) | |
fdee2e6d | 297 | continue; |
02be5561 | 298 | tid = rcu_reader_reg->tid; |
fdee2e6d MD |
299 | ret = pthread_kill(tid, 0); |
300 | assert(ret != EINVAL); | |
301 | if (ret == ESRCH) { | |
02be5561 MD |
302 | list_del(&rcu_reader_reg->head); |
303 | rcu_reader_reg->alloc = 0; | |
304 | registry_arena.used -= sizeof(struct rcu_reader); | |
fdee2e6d MD |
305 | } |
306 | } | |
307 | } | |
308 | ||
309 | /* Disable signals, take mutex, add to registry */ | |
310 | void rcu_bp_register(void) | |
311 | { | |
312 | sigset_t newmask, oldmask; | |
313 | int ret; | |
314 | ||
315 | ret = sigemptyset(&newmask); | |
316 | assert(!ret); | |
317 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
318 | assert(!ret); | |
319 | ||
320 | /* | |
321 | * Check if a signal concurrently registered our thread since | |
322 | * the check in rcu_read_lock(). */ | |
02be5561 | 323 | if (rcu_reader) |
fdee2e6d MD |
324 | goto end; |
325 | ||
6abb4bd5 | 326 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 327 | add_thread(); |
6abb4bd5 | 328 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
329 | end: |
330 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
331 | assert(!ret); | |
332 | } | |
333 | ||
02be5561 | 334 | void rcu_bp_exit() |
fdee2e6d MD |
335 | { |
336 | munmap(registry_arena.p, registry_arena.len); | |
337 | } |