4 * Userspace RCU library
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
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.
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.
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
23 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
36 #include "urcu-static.h"
37 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
42 int has_sys_membarrier
;
44 void __attribute__((constructor
)) rcu_init(void);
56 void __attribute__((constructor
)) rcu_init(void);
57 void __attribute__((destructor
)) rcu_exit(void);
60 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
65 * Global grace period counter.
66 * Contains the current RCU_GP_CTR_PHASE.
67 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
68 * Written to only by writer with mutex taken. Read by both writer and readers.
70 unsigned long rcu_gp_ctr
= RCU_GP_COUNT
;
73 * Written to only by each individual reader. Read by both the reader and the
76 struct rcu_reader __thread rcu_reader
;
79 unsigned int yield_active
;
80 unsigned int __thread rand_yield
;
83 static CDS_LIST_HEAD(registry
);
85 static void mutex_lock(pthread_mutex_t
*mutex
)
89 #ifndef DISTRUST_SIGNALS_EXTREME
90 ret
= pthread_mutex_lock(mutex
);
92 perror("Error in pthread mutex lock");
95 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
96 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
97 if (ret
!= EBUSY
&& ret
!= EINTR
) {
98 printf("ret = %d, errno = %d\n", ret
, errno
);
99 perror("Error in pthread mutex lock");
102 if (CMM_LOAD_SHARED(rcu_reader
.need_mb
)) {
104 _CMM_STORE_SHARED(rcu_reader
.need_mb
, 0);
109 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
112 static void mutex_unlock(pthread_mutex_t
*mutex
)
116 ret
= pthread_mutex_unlock(mutex
);
118 perror("Error in pthread mutex unlock");
123 #ifdef RCU_MEMBARRIER
124 static void smp_mb_master(int group
)
126 if (likely(has_sys_membarrier
))
127 membarrier(MEMBARRIER_EXPEDITED
);
134 static void smp_mb_master(int group
)
141 static void force_mb_all_readers(void)
143 struct rcu_reader
*index
;
146 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
147 * compiler barriers around rcu read lock as real memory barriers.
149 if (cds_list_empty(®istry
))
152 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
153 * a cache flush on architectures with non-coherent cache. Let's play
154 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
155 * cache flush is enforced.
157 cds_list_for_each_entry(index
, ®istry
, node
) {
158 CMM_STORE_SHARED(index
->need_mb
, 1);
159 pthread_kill(index
->tid
, SIGRCU
);
162 * Wait for sighandler (and thus mb()) to execute on every thread.
164 * Note that the pthread_kill() will never be executed on systems
165 * that correctly deliver signals in a timely manner. However, it
166 * is not uncommon for kernels to have bugs that can result in
167 * lost or unduly delayed signals.
169 * If you are seeing the below pthread_kill() executing much at
170 * all, we suggest testing the underlying kernel and filing the
171 * relevant bug report. For Linux kernels, we recommend getting
172 * the Linux Test Project (LTP).
174 cds_list_for_each_entry(index
, ®istry
, node
) {
175 while (CMM_LOAD_SHARED(index
->need_mb
)) {
176 pthread_kill(index
->tid
, SIGRCU
);
180 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
183 static void smp_mb_master(int group
)
185 force_mb_all_readers();
187 #endif /* #ifdef RCU_SIGNAL */
190 * synchronize_rcu() waiting. Single thread.
192 static void wait_gp(void)
194 /* Read reader_gp before read futex */
195 smp_mb_master(RCU_MB_GROUP
);
196 if (uatomic_read(&gp_futex
) == -1)
197 futex_async(&gp_futex
, FUTEX_WAIT
, -1,
201 void update_counter_and_wait(void)
203 CDS_LIST_HEAD(qsreaders
);
205 struct rcu_reader
*index
, *tmp
;
207 /* Switch parity: 0 -> 1, 1 -> 0 */
208 CMM_STORE_SHARED(rcu_gp_ctr
, rcu_gp_ctr
^ RCU_GP_CTR_PHASE
);
211 * Must commit rcu_gp_ctr update to memory before waiting for quiescent
212 * state. Failure to do so could result in the writer waiting forever
213 * while new readers are always accessing data (no progress). Enforce
214 * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
220 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
221 * model easier to understand. It does not have a big performance impact
222 * anyway, given this is the write-side.
227 * Wait for each thread rcu_reader.ctr count to become 0.
231 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
) {
232 uatomic_dec(&gp_futex
);
233 /* Write futex before read reader_gp */
234 smp_mb_master(RCU_MB_GROUP
);
237 cds_list_for_each_entry_safe(index
, tmp
, ®istry
, node
) {
238 if (!rcu_gp_ongoing(&index
->ctr
))
239 cds_list_move(&index
->node
, &qsreaders
);
242 #ifndef HAS_INCOHERENT_CACHES
243 if (cds_list_empty(®istry
)) {
244 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
) {
245 /* Read reader_gp before write futex */
246 smp_mb_master(RCU_MB_GROUP
);
247 uatomic_set(&gp_futex
, 0);
251 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
)
256 #else /* #ifndef HAS_INCOHERENT_CACHES */
258 * BUSY-LOOP. Force the reader thread to commit its
259 * rcu_reader.ctr update to memory if we wait for too long.
261 if (cds_list_empty(®istry
)) {
262 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
) {
263 /* Read reader_gp before write futex */
264 smp_mb_master(RCU_MB_GROUP
);
265 uatomic_set(&gp_futex
, 0);
269 switch (wait_loops
) {
270 case RCU_QS_ACTIVE_ATTEMPTS
:
272 break; /* only escape switch */
273 case KICK_READER_LOOPS
:
274 smp_mb_master(RCU_MB_GROUP
);
276 break; /* only escape switch */
281 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
283 /* put back the reader list in the registry */
284 cds_list_splice(&qsreaders
, ®istry
);
287 void synchronize_rcu(void)
289 mutex_lock(&rcu_gp_lock
);
291 if (cds_list_empty(®istry
))
294 /* All threads should read qparity before accessing data structure
295 * where new ptr points to. Must be done within rcu_gp_lock because it
296 * iterates on reader threads.*/
297 /* Write new ptr before changing the qparity */
298 smp_mb_master(RCU_MB_GROUP
);
301 * Wait for previous parity to be empty of readers.
303 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
306 * Must finish waiting for quiescent state for parity 0 before
307 * committing next rcu_gp_ctr update to memory. Failure to do so could
308 * result in the writer waiting forever while new readers are always
309 * accessing data (no progress). Enforce compiler-order of load
310 * rcu_reader ctr before store to rcu_gp_ctr.
315 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
316 * model easier to understand. It does not have a big performance impact
317 * anyway, given this is the write-side.
322 * Wait for previous parity to be empty of readers.
324 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
326 /* Finish waiting for reader threads before letting the old ptr being
327 * freed. Must be done within rcu_gp_lock because it iterates on reader
329 smp_mb_master(RCU_MB_GROUP
);
331 mutex_unlock(&rcu_gp_lock
);
335 * library wrappers to be used by non-LGPL compatible source code.
338 void rcu_read_lock(void)
343 void rcu_read_unlock(void)
348 void rcu_register_thread(void)
350 rcu_reader
.tid
= pthread_self();
351 assert(rcu_reader
.need_mb
== 0);
352 assert(!(rcu_reader
.ctr
& RCU_GP_CTR_NEST_MASK
));
354 mutex_lock(&rcu_gp_lock
);
355 rcu_init(); /* In case gcc does not support constructor attribute */
356 cds_list_add(&rcu_reader
.node
, ®istry
);
357 mutex_unlock(&rcu_gp_lock
);
360 void rcu_unregister_thread(void)
362 mutex_lock(&rcu_gp_lock
);
363 cds_list_del(&rcu_reader
.node
);
364 mutex_unlock(&rcu_gp_lock
);
367 #ifdef RCU_MEMBARRIER
373 if (!membarrier(MEMBARRIER_EXPEDITED
| MEMBARRIER_QUERY
))
374 has_sys_membarrier
= 1;
379 static void sigrcu_handler(int signo
, siginfo_t
*siginfo
, void *context
)
382 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
383 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
387 _CMM_STORE_SHARED(rcu_reader
.need_mb
, 0);
392 * rcu_init constructor. Called when the library is linked, but also when
393 * reader threads are calling rcu_register_thread().
394 * Should only be called by a single thread at a given time. This is ensured by
395 * holing the rcu_gp_lock from rcu_register_thread() or by running at library
396 * load time, which should not be executed by multiple threads nor concurrently
397 * with rcu_register_thread() anyway.
401 struct sigaction act
;
408 act
.sa_sigaction
= sigrcu_handler
;
409 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
410 sigemptyset(&act
.sa_mask
);
411 ret
= sigaction(SIGRCU
, &act
, NULL
);
413 perror("Error in sigaction");
420 struct sigaction act
;
423 ret
= sigaction(SIGRCU
, NULL
, &act
);
425 perror("Error in sigaction");
428 assert(act
.sa_sigaction
== sigrcu_handler
);
429 assert(cds_list_empty(®istry
));
431 #endif /* #ifdef RCU_SIGNAL */