urcu-bp: improve 2-phase wait scheme
[urcu.git] / urcu-qsbr.c
1 /*
2 * urcu-qsbr.c
3 *
4 * Userspace RCU QSBR library
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 #define _LGPL_SOURCE
28 #include <stdio.h>
29 #include <pthread.h>
30 #include <signal.h>
31 #include <assert.h>
32 #include <stdlib.h>
33 #include <stdint.h>
34 #include <string.h>
35 #include <errno.h>
36 #include <poll.h>
37
38 #include "urcu/wfcqueue.h"
39 #include "urcu/map/urcu-qsbr.h"
40 #define BUILD_QSBR_LIB
41 #include "urcu/static/urcu-qsbr.h"
42 #include "urcu-pointer.h"
43 #include "urcu/tls-compat.h"
44
45 #include "urcu-die.h"
46
47 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
48 #undef _LGPL_SOURCE
49 #include "urcu-qsbr.h"
50 #define _LGPL_SOURCE
51
52 void __attribute__((destructor)) rcu_exit(void);
53
54 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
55
56 int32_t rcu_gp_futex;
57
58 /*
59 * Global grace period counter.
60 */
61 unsigned long rcu_gp_ctr = RCU_GP_ONLINE;
62
63 /*
64 * Active attempts to check for reader Q.S. before calling futex().
65 */
66 #define RCU_QS_ACTIVE_ATTEMPTS 100
67
68 /*
69 * Written to only by each individual reader. Read by both the reader and the
70 * writers.
71 */
72 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
73
74 #ifdef DEBUG_YIELD
75 unsigned int rcu_yield_active;
76 DEFINE_URCU_TLS(unsigned int, rcu_rand_yield);
77 #endif
78
79 static CDS_LIST_HEAD(registry);
80
81 static void mutex_lock(pthread_mutex_t *mutex)
82 {
83 int ret;
84
85 #ifndef DISTRUST_SIGNALS_EXTREME
86 ret = pthread_mutex_lock(mutex);
87 if (ret)
88 urcu_die(ret);
89 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
90 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
91 if (ret != EBUSY && ret != EINTR)
92 urcu_die(ret);
93 poll(NULL,0,10);
94 }
95 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
96 }
97
98 static void mutex_unlock(pthread_mutex_t *mutex)
99 {
100 int ret;
101
102 ret = pthread_mutex_unlock(mutex);
103 if (ret)
104 urcu_die(ret);
105 }
106
107 /*
108 * synchronize_rcu() waiting. Single thread.
109 */
110 static void wait_gp(void)
111 {
112 /* Read reader_gp before read futex */
113 cmm_smp_rmb();
114 if (uatomic_read(&rcu_gp_futex) == -1)
115 futex_noasync(&rcu_gp_futex, FUTEX_WAIT, -1,
116 NULL, NULL, 0);
117 }
118
119 static void wait_for_readers(void)
120 {
121 CDS_LIST_HEAD(qsreaders);
122 int wait_loops = 0;
123 struct rcu_reader *index, *tmp;
124
125 /*
126 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
127 * indicate quiescence (offline), or for them to observe the
128 * current rcu_gp_ctr value.
129 */
130 for (;;) {
131 wait_loops++;
132 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
133 uatomic_set(&rcu_gp_futex, -1);
134 /*
135 * Write futex before write waiting (the other side
136 * reads them in the opposite order).
137 */
138 cmm_smp_wmb();
139 cds_list_for_each_entry(index, &registry, node) {
140 _CMM_STORE_SHARED(index->waiting, 1);
141 }
142 /* Write futex before read reader_gp */
143 cmm_smp_mb();
144 }
145 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
146 if (!rcu_gp_ongoing(&index->ctr))
147 cds_list_move(&index->node, &qsreaders);
148 }
149
150 if (cds_list_empty(&registry)) {
151 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
152 /* Read reader_gp before write futex */
153 cmm_smp_mb();
154 uatomic_set(&rcu_gp_futex, 0);
155 }
156 break;
157 } else {
158 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
159 wait_gp();
160 } else {
161 #ifndef HAS_INCOHERENT_CACHES
162 caa_cpu_relax();
163 #else /* #ifndef HAS_INCOHERENT_CACHES */
164 cmm_smp_mb();
165 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
166 }
167 }
168 }
169 /* put back the reader list in the registry */
170 cds_list_splice(&qsreaders, &registry);
171 }
172
173 /*
174 * Using a two-subphases algorithm for architectures with smaller than 64-bit
175 * long-size to ensure we do not encounter an overflow bug.
176 */
177
178 #if (CAA_BITS_PER_LONG < 64)
179 void synchronize_rcu(void)
180 {
181 unsigned long was_online;
182
183 was_online = URCU_TLS(rcu_reader).ctr;
184
185 /* All threads should read qparity before accessing data structure
186 * where new ptr points to. In the "then" case, rcu_thread_offline
187 * includes a memory barrier.
188 *
189 * Mark the writer thread offline to make sure we don't wait for
190 * our own quiescent state. This allows using synchronize_rcu()
191 * in threads registered as readers.
192 */
193 if (was_online)
194 rcu_thread_offline();
195 else
196 cmm_smp_mb();
197
198 mutex_lock(&rcu_gp_lock);
199
200 if (cds_list_empty(&registry))
201 goto out;
202
203 /*
204 * Wait for readers to observe original parity or be quiescent.
205 */
206 wait_for_readers();
207
208 /*
209 * Must finish waiting for quiescent state for original parity
210 * before committing next rcu_gp_ctr update to memory. Failure
211 * to do so could result in the writer waiting forever while new
212 * readers are always accessing data (no progress). Enforce
213 * compiler-order of load URCU_TLS(rcu_reader).ctr before store
214 * to rcu_gp_ctr.
215 */
216 cmm_barrier();
217
218 /*
219 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
220 * model easier to understand. It does not have a big performance impact
221 * anyway, given this is the write-side.
222 */
223 cmm_smp_mb();
224
225 /* Switch parity: 0 -> 1, 1 -> 0 */
226 CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
227
228 /*
229 * Must commit rcu_gp_ctr update to memory before waiting for
230 * quiescent state. Failure to do so could result in the writer
231 * waiting forever while new readers are always accessing data
232 * (no progress). Enforce compiler-order of store to rcu_gp_ctr
233 * before load URCU_TLS(rcu_reader).ctr.
234 */
235 cmm_barrier();
236
237 /*
238 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
239 * model easier to understand. It does not have a big performance impact
240 * anyway, given this is the write-side.
241 */
242 cmm_smp_mb();
243
244 /*
245 * Wait for readers to observe new parity or be quiescent.
246 */
247 wait_for_readers();
248 out:
249 mutex_unlock(&rcu_gp_lock);
250
251 /*
252 * Finish waiting for reader threads before letting the old ptr being
253 * freed.
254 */
255 if (was_online)
256 rcu_thread_online();
257 else
258 cmm_smp_mb();
259 }
260 #else /* !(CAA_BITS_PER_LONG < 64) */
261 void synchronize_rcu(void)
262 {
263 unsigned long was_online;
264
265 was_online = URCU_TLS(rcu_reader).ctr;
266
267 /*
268 * Mark the writer thread offline to make sure we don't wait for
269 * our own quiescent state. This allows using synchronize_rcu()
270 * in threads registered as readers.
271 */
272 if (was_online)
273 rcu_thread_offline();
274 else
275 cmm_smp_mb();
276
277 mutex_lock(&rcu_gp_lock);
278 if (cds_list_empty(&registry))
279 goto out;
280
281 /* Increment current G.P. */
282 CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
283
284 /*
285 * Must commit rcu_gp_ctr update to memory before waiting for
286 * quiescent state. Failure to do so could result in the writer
287 * waiting forever while new readers are always accessing data
288 * (no progress). Enforce compiler-order of store to rcu_gp_ctr
289 * before load URCU_TLS(rcu_reader).ctr.
290 */
291 cmm_barrier();
292
293 /*
294 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
295 * model easier to understand. It does not have a big performance impact
296 * anyway, given this is the write-side.
297 */
298 cmm_smp_mb();
299
300 /*
301 * Wait for readers to observe new count of be quiescent.
302 */
303 wait_for_readers();
304 out:
305 mutex_unlock(&rcu_gp_lock);
306
307 if (was_online)
308 rcu_thread_online();
309 else
310 cmm_smp_mb();
311 }
312 #endif /* !(CAA_BITS_PER_LONG < 64) */
313
314 /*
315 * library wrappers to be used by non-LGPL compatible source code.
316 */
317
318 void rcu_read_lock(void)
319 {
320 _rcu_read_lock();
321 }
322
323 void rcu_read_unlock(void)
324 {
325 _rcu_read_unlock();
326 }
327
328 void rcu_quiescent_state(void)
329 {
330 _rcu_quiescent_state();
331 }
332
333 void rcu_thread_offline(void)
334 {
335 _rcu_thread_offline();
336 }
337
338 void rcu_thread_online(void)
339 {
340 _rcu_thread_online();
341 }
342
343 void rcu_register_thread(void)
344 {
345 URCU_TLS(rcu_reader).tid = pthread_self();
346 assert(URCU_TLS(rcu_reader).ctr == 0);
347
348 mutex_lock(&rcu_gp_lock);
349 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
350 mutex_unlock(&rcu_gp_lock);
351 _rcu_thread_online();
352 }
353
354 void rcu_unregister_thread(void)
355 {
356 /*
357 * We have to make the thread offline otherwise we end up dealocking
358 * with a waiting writer.
359 */
360 _rcu_thread_offline();
361 mutex_lock(&rcu_gp_lock);
362 cds_list_del(&URCU_TLS(rcu_reader).node);
363 mutex_unlock(&rcu_gp_lock);
364 }
365
366 void rcu_exit(void)
367 {
368 /*
369 * Assertion disabled because call_rcu threads are now rcu
370 * readers, and left running at exit.
371 * assert(cds_list_empty(&registry));
372 */
373 }
374
375 DEFINE_RCU_FLAVOR(rcu_flavor);
376
377 #include "urcu-call-rcu-impl.h"
378 #include "urcu-defer-impl.h"
This page took 0.036427 seconds and 5 git commands to generate.