Rename urcu/urcu-futex.h to urcu/futex.h
[urcu.git] / urcu-bp.c
1 /*
2 * urcu-bp.c
3 *
4 * Userspace RCU library, "bulletproof" version.
5 *
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
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 #define _GNU_SOURCE
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/map/urcu-bp.h"
39
40 #include "urcu/static/urcu-bp.h"
41 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
42 #include "urcu-bp.h"
43
44 /* Sleep delay in us */
45 #define RCU_SLEEP_DELAY 1000
46 #define ARENA_INIT_ALLOC 16
47
48 void __attribute__((destructor)) rcu_bp_exit(void);
49
50 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
51
52 #ifdef DEBUG_YIELD
53 unsigned int yield_active;
54 unsigned int __thread rand_yield;
55 #endif
56
57 /*
58 * Global grace period counter.
59 * Contains the current RCU_GP_CTR_PHASE.
60 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
61 * Written to only by writer with mutex taken. Read by both writer and readers.
62 */
63 long rcu_gp_ctr = RCU_GP_COUNT;
64
65 /*
66 * Pointer to registry elements. Written to only by each individual reader. Read
67 * by both the reader and the writers.
68 */
69 struct rcu_reader __thread *rcu_reader;
70
71 static CDS_LIST_HEAD(registry);
72
73 struct registry_arena {
74 void *p;
75 size_t len;
76 size_t used;
77 };
78
79 static struct registry_arena registry_arena;
80
81 /* Saved fork signal mask, protected by rcu_gp_lock */
82 static sigset_t saved_fork_signal_mask;
83
84 static void rcu_gc_registry(void);
85
86 static void mutex_lock(pthread_mutex_t *mutex)
87 {
88 int ret;
89
90 #ifndef DISTRUST_SIGNALS_EXTREME
91 ret = pthread_mutex_lock(mutex);
92 if (ret) {
93 perror("Error in pthread mutex lock");
94 exit(-1);
95 }
96 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
97 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
98 if (ret != EBUSY && ret != EINTR) {
99 printf("ret = %d, errno = %d\n", ret, errno);
100 perror("Error in pthread mutex lock");
101 exit(-1);
102 }
103 poll(NULL,0,10);
104 }
105 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
106 }
107
108 static void mutex_unlock(pthread_mutex_t *mutex)
109 {
110 int ret;
111
112 ret = pthread_mutex_unlock(mutex);
113 if (ret) {
114 perror("Error in pthread mutex unlock");
115 exit(-1);
116 }
117 }
118
119 void update_counter_and_wait(void)
120 {
121 CDS_LIST_HEAD(qsreaders);
122 int wait_loops = 0;
123 struct rcu_reader *index, *tmp;
124
125 /* Switch parity: 0 -> 1, 1 -> 0 */
126 CMM_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 CMM_STORE_SHARED and CMM_LOAD_SHARED.
133 */
134
135 /*
136 * Adding a cmm_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 cmm_smp_mb();
141
142 /*
143 * Wait for each thread rcu_reader.ctr count to become 0.
144 */
145 for (;;) {
146 wait_loops++;
147 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
148 if (!rcu_old_gp_ongoing(&index->ctr))
149 cds_list_move(&index->node, &qsreaders);
150 }
151
152 if (cds_list_empty(&registry)) {
153 break;
154 } else {
155 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
156 usleep(RCU_SLEEP_DELAY);
157 else
158 caa_cpu_relax();
159 }
160 }
161 /* put back the reader list in the registry */
162 cds_list_splice(&qsreaders, &registry);
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
175 mutex_lock(&rcu_gp_lock);
176
177 if (cds_list_empty(&registry))
178 goto out;
179
180 /* All threads should read qparity before accessing data structure
181 * where new ptr points to. */
182 /* Write new ptr before changing the qparity */
183 cmm_smp_mb();
184
185 /* Remove old registry elements */
186 rcu_gc_registry();
187
188 /*
189 * Wait for previous parity to be empty of readers.
190 */
191 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
192
193 /*
194 * Adding a cmm_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 cmm_smp_mb();
199
200 /*
201 * Wait for previous parity to be empty of readers.
202 */
203 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
204
205 /*
206 * Finish waiting for reader threads before letting the old ptr being
207 * freed.
208 */
209 cmm_smp_mb();
210 out:
211 mutex_unlock(&rcu_gp_lock);
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
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
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 {
261 struct rcu_reader *rcu_reader_reg;
262
263 if (registry_arena.len
264 < registry_arena.used + sizeof(struct rcu_reader))
265 resize_arena(&registry_arena,
266 max(registry_arena.len << 1, ARENA_INIT_ALLOC));
267 /*
268 * Find a free spot.
269 */
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)
274 break;
275 }
276 rcu_reader_reg->alloc = 1;
277 registry_arena.used += sizeof(struct rcu_reader);
278
279 /* Add to registry */
280 rcu_reader_reg->tid = pthread_self();
281 assert(rcu_reader_reg->ctr == 0);
282 cds_list_add(&rcu_reader_reg->node, &registry);
283 rcu_reader = rcu_reader_reg;
284 }
285
286 /* Called with signals off and mutex locked */
287 static void rcu_gc_registry(void)
288 {
289 struct rcu_reader *rcu_reader_reg;
290 pthread_t tid;
291 int ret;
292
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)
297 continue;
298 tid = rcu_reader_reg->tid;
299 ret = pthread_kill(tid, 0);
300 assert(ret != EINVAL);
301 if (ret == ESRCH) {
302 cds_list_del(&rcu_reader_reg->node);
303 rcu_reader_reg->ctr = 0;
304 rcu_reader_reg->alloc = 0;
305 registry_arena.used -= sizeof(struct rcu_reader);
306 }
307 }
308 }
309
310 /* Disable signals, take mutex, add to registry */
311 void rcu_bp_register(void)
312 {
313 sigset_t newmask, oldmask;
314 int ret;
315
316 ret = sigemptyset(&newmask);
317 assert(!ret);
318 ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
319 assert(!ret);
320
321 /*
322 * Check if a signal concurrently registered our thread since
323 * the check in rcu_read_lock(). */
324 if (rcu_reader)
325 goto end;
326
327 mutex_lock(&rcu_gp_lock);
328 add_thread();
329 mutex_unlock(&rcu_gp_lock);
330 end:
331 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
332 assert(!ret);
333 }
334
335 void rcu_bp_exit()
336 {
337 munmap(registry_arena.p, registry_arena.len);
338 }
339
340 /*
341 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
342 * a concurrent thread executing with this same lock held. This ensures that the
343 * registry is in a coherent state in the child.
344 */
345 void rcu_bp_before_fork(void)
346 {
347 sigset_t newmask, oldmask;
348 int ret;
349
350 ret = sigemptyset(&newmask);
351 assert(!ret);
352 ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
353 assert(!ret);
354 mutex_lock(&rcu_gp_lock);
355 saved_fork_signal_mask = oldmask;
356 }
357
358 void rcu_bp_after_fork_parent(void)
359 {
360 sigset_t oldmask;
361 int ret;
362
363 oldmask = saved_fork_signal_mask;
364 mutex_unlock(&rcu_gp_lock);
365 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
366 assert(!ret);
367 }
368
369 void rcu_bp_after_fork_child(void)
370 {
371 sigset_t oldmask;
372 int ret;
373
374 rcu_gc_registry();
375 oldmask = saved_fork_signal_mask;
376 mutex_unlock(&rcu_gp_lock);
377 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
378 assert(!ret);
379 }
380
381 #include "urcu-call-rcu-impl.h"
382 #include "urcu-defer-impl.h"
This page took 0.035907 seconds and 4 git commands to generate.