Support architectures with non-coherent caches
[urcu.git] / urcu.c
1 /*
2 * urcu.c
3 *
4 * Userspace RCU library
5 *
6 * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
7 *
8 * Distributed under GPLv2
9 */
10
11 #include <stdio.h>
12 #include <pthread.h>
13 #include <signal.h>
14 #include <assert.h>
15 #include <stdlib.h>
16 #include <string.h>
17
18 #include "urcu.h"
19
20 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
21
22 /*
23 * Global grace period counter.
24 * Contains the current RCU_GP_CTR_BIT.
25 * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
26 */
27 long urcu_gp_ctr = RCU_GP_COUNT;
28
29 long __thread urcu_active_readers;
30
31 /* Thread IDs of registered readers */
32 #define INIT_NUM_THREADS 4
33
34 struct reader_data {
35 pthread_t tid;
36 long *urcu_active_readers;
37 };
38
39 #ifdef DEBUG_YIELD
40 unsigned int yield_active;
41 unsigned int __thread rand_yield;
42 #endif
43
44 static struct reader_data *reader_data;
45 static int num_readers, alloc_readers;
46 #ifndef DEBUG_FULL_MB
47 static int sig_done;
48 #endif
49
50 void internal_urcu_lock(void)
51 {
52 int ret;
53 ret = pthread_mutex_lock(&urcu_mutex);
54 if (ret) {
55 perror("Error in pthread mutex lock");
56 exit(-1);
57 }
58 }
59
60 void internal_urcu_unlock(void)
61 {
62 int ret;
63
64 ret = pthread_mutex_unlock(&urcu_mutex);
65 if (ret) {
66 perror("Error in pthread mutex unlock");
67 exit(-1);
68 }
69 }
70
71 /*
72 * called with urcu_mutex held.
73 */
74 static void switch_next_urcu_qparity(void)
75 {
76 urcu_gp_ctr ^= RCU_GP_CTR_BIT;
77 }
78
79 #ifdef DEBUG_FULL_MB
80 static void force_mb_single_thread(pthread_t tid)
81 {
82 smp_mb();
83 }
84
85 static void force_mb_all_threads(void)
86 {
87 smp_mb();
88 }
89 #else
90
91 static void force_mb_single_thread(pthread_t tid)
92 {
93 assert(reader_data);
94 sig_done = 0;
95 smp_mb(); /* write sig_done before sending the signals */
96 pthread_kill(tid, SIGURCU);
97 /*
98 * Wait for sighandler (and thus mb()) to execute on every thread.
99 * BUSY-LOOP.
100 */
101 while (LOAD_REMOTE(sig_done) < 1)
102 cpu_relax();
103 smp_mb(); /* read sig_done before ending the barrier */
104 }
105
106 static void force_mb_all_threads(void)
107 {
108 struct reader_data *index;
109 /*
110 * Ask for each threads to execute a smp_mb() so we can consider the
111 * compiler barriers around rcu read lock as real memory barriers.
112 */
113 if (!reader_data)
114 return;
115 sig_done = 0;
116 /*
117 * pthread_kill has a smp_mb(). But beware, we assume it performs
118 * a cache flush on architectures with non-coherent cache.
119 * smp_mb(); write sig_done before sending the signals
120 */
121 for (index = reader_data; index < reader_data + num_readers; index++)
122 pthread_kill(index->tid, SIGURCU);
123 /*
124 * Wait for sighandler (and thus mb()) to execute on every thread.
125 * BUSY-LOOP.
126 */
127 while (LOAD_REMOTE(sig_done) < num_readers)
128 cpu_relax();
129 smp_mb(); /* read sig_done before ending the barrier */
130 }
131 #endif
132
133 void wait_for_quiescent_state(void)
134 {
135 struct reader_data *index;
136
137 if (!reader_data)
138 return;
139 /*
140 * Wait for each thread urcu_active_readers count to become 0.
141 */
142 for (index = reader_data; index < reader_data + num_readers; index++) {
143 int wait_loops = 0;
144 /*
145 * BUSY-LOOP. Force the reader thread to commit its
146 * urcu_active_readers update to memory if we wait for too long.
147 */
148 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
149 if (wait_loops++ == KICK_READER_LOOPS) {
150 force_mb_single_thread(index->tid);
151 wait_loops = 0;
152 }
153 }
154 }
155 }
156
157 void synchronize_rcu(void)
158 {
159 internal_urcu_lock();
160
161 /* All threads should read qparity before accessing data structure
162 * where new ptr points to. Must be done within internal_urcu_lock
163 * because it iterates on reader threads.*/
164 /* Write new ptr before changing the qparity */
165 force_mb_all_threads();
166
167 switch_next_urcu_qparity(); /* 0 -> 1 */
168
169 /*
170 * Must commit qparity update to memory before waiting for parity
171 * 0 quiescent state. Failure to do so could result in the writer
172 * waiting forever while new readers are always accessing data (no
173 * progress).
174 */
175 smp_mc();
176
177 /*
178 * Wait for previous parity to be empty of readers.
179 */
180 wait_for_quiescent_state(); /* Wait readers in parity 0 */
181
182 /*
183 * Must finish waiting for quiescent state for parity 0 before
184 * committing qparity update to memory. Failure to do so could result in
185 * the writer waiting forever while new readers are always accessing
186 * data (no progress).
187 */
188 smp_mc();
189
190 switch_next_urcu_qparity(); /* 1 -> 0 */
191
192 /*
193 * Must commit qparity update to memory before waiting for parity
194 * 1 quiescent state. Failure to do so could result in the writer
195 * waiting forever while new readers are always accessing data (no
196 * progress).
197 */
198 smp_mc();
199
200 /*
201 * Wait for previous parity to be empty of readers.
202 */
203 wait_for_quiescent_state(); /* Wait readers in parity 1 */
204
205 /* Finish waiting for reader threads before letting the old ptr being
206 * freed. Must be done within internal_urcu_lock because it iterates on
207 * reader threads. */
208 force_mb_all_threads();
209
210 internal_urcu_unlock();
211 }
212
213 void urcu_add_reader(pthread_t id)
214 {
215 struct reader_data *oldarray;
216
217 if (!reader_data) {
218 alloc_readers = INIT_NUM_THREADS;
219 num_readers = 0;
220 reader_data =
221 malloc(sizeof(struct reader_data) * alloc_readers);
222 }
223 if (alloc_readers < num_readers + 1) {
224 oldarray = reader_data;
225 reader_data = malloc(sizeof(struct reader_data)
226 * (alloc_readers << 1));
227 memcpy(reader_data, oldarray,
228 sizeof(struct reader_data) * alloc_readers);
229 alloc_readers <<= 1;
230 free(oldarray);
231 }
232 reader_data[num_readers].tid = id;
233 /* reference to the TLS of _this_ reader thread. */
234 reader_data[num_readers].urcu_active_readers = &urcu_active_readers;
235 num_readers++;
236 }
237
238 /*
239 * Never shrink (implementation limitation).
240 * This is O(nb threads). Eventually use a hash table.
241 */
242 void urcu_remove_reader(pthread_t id)
243 {
244 struct reader_data *index;
245
246 assert(reader_data != NULL);
247 for (index = reader_data; index < reader_data + num_readers; index++) {
248 if (pthread_equal(index->tid, id)) {
249 memcpy(index, &reader_data[num_readers - 1],
250 sizeof(struct reader_data));
251 reader_data[num_readers - 1].tid = 0;
252 reader_data[num_readers - 1].urcu_active_readers = NULL;
253 num_readers--;
254 return;
255 }
256 }
257 /* Hrm not found, forgot to register ? */
258 assert(0);
259 }
260
261 void urcu_register_thread(void)
262 {
263 internal_urcu_lock();
264 urcu_add_reader(pthread_self());
265 internal_urcu_unlock();
266 }
267
268 void urcu_unregister_thread(void)
269 {
270 internal_urcu_lock();
271 urcu_remove_reader(pthread_self());
272 internal_urcu_unlock();
273 }
274
275 #ifndef DEBUG_FULL_MB
276 void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
277 {
278 /*
279 * Executing this smp_mb() is the only purpose of this signal handler.
280 * It punctually promotes barrier() into smp_mb() on every thread it is
281 * executed on.
282 */
283 smp_mb();
284 atomic_inc(&sig_done);
285 }
286
287 void __attribute__((constructor)) urcu_init(void)
288 {
289 struct sigaction act;
290 int ret;
291
292 act.sa_sigaction = sigurcu_handler;
293 ret = sigaction(SIGURCU, &act, NULL);
294 if (ret) {
295 perror("Error in sigaction");
296 exit(-1);
297 }
298 }
299
300 void __attribute__((destructor)) urcu_exit(void)
301 {
302 struct sigaction act;
303 int ret;
304
305 ret = sigaction(SIGURCU, NULL, &act);
306 if (ret) {
307 perror("Error in sigaction");
308 exit(-1);
309 }
310 assert(act.sa_sigaction == sigurcu_handler);
311 free(reader_data);
312 }
313 #endif
This page took 0.035342 seconds and 5 git commands to generate.