urcu-defer: Add fast path for empty queues
[urcu.git] / urcu-defer.c
1 /*
2 * urcu-defer.c
3 *
4 * Userspace RCU library - batch memory reclamation
5 *
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 #include <stdio.h>
24 #include <pthread.h>
25 #include <signal.h>
26 #include <assert.h>
27 #include <stdlib.h>
28 #include <string.h>
29 #include <errno.h>
30 #include <poll.h>
31
32 #include "urcu-defer-static.h"
33 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
34 #include "urcu-defer.h"
35
36 void __attribute__((destructor)) urcu_defer_exit(void);
37
38 extern void synchronize_rcu(void);
39
40 /*
41 * urcu_defer_mutex nests inside defer_thread_mutex.
42 */
43 static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
44 static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
45
46 /*
47 * Written to only by each individual deferer. Read by both the deferer and
48 * the reclamation tread.
49 */
50 struct defer_queue __thread defer_queue;
51
52 /* Thread IDs of registered deferers */
53 #define INIT_NUM_THREADS 4
54
55 struct deferer_registry {
56 pthread_t tid;
57 struct defer_queue *defer_queue;
58 unsigned long last_head;
59 };
60
61 static struct deferer_registry *registry;
62 static int num_deferers, alloc_deferers;
63
64 static pthread_t tid_defer;
65 static int exit_defer;
66
67 static void internal_urcu_lock(pthread_mutex_t *mutex)
68 {
69 int ret;
70
71 #ifndef DISTRUST_SIGNALS_EXTREME
72 ret = pthread_mutex_lock(mutex);
73 if (ret) {
74 perror("Error in pthread mutex lock");
75 exit(-1);
76 }
77 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
78 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
79 if (ret != EBUSY && ret != EINTR) {
80 printf("ret = %d, errno = %d\n", ret, errno);
81 perror("Error in pthread mutex lock");
82 exit(-1);
83 }
84 poll(NULL,0,10);
85 }
86 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
87 }
88
89 static void internal_urcu_unlock(pthread_mutex_t *mutex)
90 {
91 int ret;
92
93 ret = pthread_mutex_unlock(mutex);
94 if (ret) {
95 perror("Error in pthread mutex unlock");
96 exit(-1);
97 }
98 }
99
100 /*
101 * Must be called after Q.S. is reached.
102 */
103 static void rcu_defer_barrier_queue(struct defer_queue *queue,
104 unsigned long head)
105 {
106 unsigned long i;
107 void (*fct)(void *p);
108 void *p;
109
110 /*
111 * Tail is only modified when lock is held.
112 * Head is only modified by owner thread.
113 */
114
115 for (i = queue->tail; i != head;) {
116 smp_rmb(); /* read head before q[]. */
117 p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
118 if (unlikely(DQ_IS_FCT_BIT(p))) {
119 DQ_CLEAR_FCT_BIT(p);
120 queue->last_fct_out = p;
121 p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
122 } else if (unlikely(p == DQ_FCT_MARK)) {
123 p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
124 queue->last_fct_out = p;
125 p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
126 }
127 fct = queue->last_fct_out;
128 fct(p);
129 }
130 smp_mb(); /* push tail after having used q[] */
131 STORE_SHARED(queue->tail, i);
132 }
133
134 static void _rcu_defer_barrier_thread(void)
135 {
136 unsigned long head, num_items;
137
138 head = defer_queue.head;
139 num_items = head - defer_queue.tail;
140 if (unlikely(!num_items))
141 return;
142 synchronize_rcu();
143 rcu_defer_barrier_queue(&defer_queue, head);
144 }
145
146
147 void rcu_defer_barrier_thread(void)
148 {
149 internal_urcu_lock(&urcu_defer_mutex);
150 _rcu_defer_barrier_thread();
151 internal_urcu_unlock(&urcu_defer_mutex);
152 }
153
154 /*
155 * rcu_defer_barrier - Execute all queued rcu callbacks.
156 *
157 * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
158 * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
159 * are guaranteed to be executed.
160 * Callbacks queued by other threads concurrently with rcu_defer_barrier()
161 * execution are not guaranteed to be executed in the current batch (could
162 * be left for the next batch). These callbacks queued by other threads are only
163 * guaranteed to be executed if there is explicit synchronization between
164 * the thread adding to the queue and the thread issuing the defer_barrier call.
165 */
166
167 void rcu_defer_barrier(void)
168 {
169 struct deferer_registry *index;
170 unsigned long num_items = 0;
171
172 if (!registry)
173 return;
174
175 internal_urcu_lock(&urcu_defer_mutex);
176 for (index = registry; index < registry + num_deferers; index++) {
177 index->last_head = LOAD_SHARED(index->defer_queue->head);
178 num_items += index->last_head - index->defer_queue->tail;
179 }
180 if (likely(!num_items)) {
181 /*
182 * We skip the grace period because there are no queued
183 * callbacks to execute.
184 */
185 goto end;
186 }
187 synchronize_rcu();
188 for (index = registry; index < registry + num_deferers; index++)
189 rcu_defer_barrier_queue(index->defer_queue,
190 index->last_head);
191 end:
192 internal_urcu_unlock(&urcu_defer_mutex);
193 }
194
195 void *thr_defer(void *args)
196 {
197 for (;;) {
198 if (LOAD_SHARED(exit_defer))
199 break;
200 poll(NULL,0,100); /* wait for 100ms */
201 rcu_defer_barrier();
202 }
203
204 return NULL;
205 }
206
207 /*
208 * library wrappers to be used by non-LGPL compatible source code.
209 */
210
211 void rcu_defer_queue(void (*fct)(void *p), void *p)
212 {
213 _rcu_defer_queue(fct, p);
214 }
215
216 static void rcu_add_deferer(pthread_t id)
217 {
218 struct deferer_registry *oldarray;
219
220 if (!registry) {
221 alloc_deferers = INIT_NUM_THREADS;
222 num_deferers = 0;
223 registry =
224 malloc(sizeof(struct deferer_registry) * alloc_deferers);
225 }
226 if (alloc_deferers < num_deferers + 1) {
227 oldarray = registry;
228 registry = malloc(sizeof(struct deferer_registry)
229 * (alloc_deferers << 1));
230 memcpy(registry, oldarray,
231 sizeof(struct deferer_registry) * alloc_deferers);
232 alloc_deferers <<= 1;
233 free(oldarray);
234 }
235 registry[num_deferers].tid = id;
236 /* reference to the TLS of _this_ deferer thread. */
237 registry[num_deferers].defer_queue = &defer_queue;
238 registry[num_deferers].last_head = 0;
239 num_deferers++;
240 }
241
242 /*
243 * Never shrink (implementation limitation).
244 * This is O(nb threads). Eventually use a hash table.
245 */
246 static void rcu_remove_deferer(pthread_t id)
247 {
248 struct deferer_registry *index;
249
250 assert(registry != NULL);
251 for (index = registry; index < registry + num_deferers; index++) {
252 if (pthread_equal(index->tid, id)) {
253 memcpy(index, &registry[num_deferers - 1],
254 sizeof(struct deferer_registry));
255 registry[num_deferers - 1].tid = 0;
256 registry[num_deferers - 1].defer_queue = NULL;
257 registry[num_deferers - 1].last_head = 0;
258 num_deferers--;
259 return;
260 }
261 }
262 /* Hrm not found, forgot to register ? */
263 assert(0);
264 }
265
266 static void start_defer_thread(void)
267 {
268 int ret;
269
270 ret = pthread_create(&tid_defer, NULL, thr_defer,
271 NULL);
272 assert(!ret);
273 }
274
275 static void stop_defer_thread(void)
276 {
277 int ret;
278 void *tret;
279
280 STORE_SHARED(exit_defer, 1);
281 ret = pthread_join(tid_defer, &tret);
282 assert(!ret);
283 }
284
285 void rcu_defer_register_thread(void)
286 {
287 int deferers;
288
289 internal_urcu_lock(&defer_thread_mutex);
290 internal_urcu_lock(&urcu_defer_mutex);
291 defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
292 rcu_add_deferer(pthread_self());
293 deferers = num_deferers;
294 internal_urcu_unlock(&urcu_defer_mutex);
295
296 if (deferers == 1)
297 start_defer_thread();
298 internal_urcu_unlock(&defer_thread_mutex);
299 }
300
301 void rcu_defer_unregister_thread(void)
302 {
303 int deferers;
304
305 internal_urcu_lock(&defer_thread_mutex);
306 internal_urcu_lock(&urcu_defer_mutex);
307 rcu_remove_deferer(pthread_self());
308 _rcu_defer_barrier_thread();
309 free(defer_queue.q);
310 defer_queue.q = NULL;
311 deferers = num_deferers;
312 internal_urcu_unlock(&urcu_defer_mutex);
313
314 if (deferers == 0)
315 stop_defer_thread();
316 internal_urcu_unlock(&defer_thread_mutex);
317 }
318
319 void urcu_defer_exit(void)
320 {
321 free(registry);
322 }
This page took 0.035582 seconds and 5 git commands to generate.