4 * Userspace RCU library
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
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.
35 #include "urcu-static.h"
36 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
39 void __attribute__((constructor
)) urcu_init(void);
40 void __attribute__((destructor
)) urcu_exit(void);
44 pthread_mutex_t urcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
47 * Global grace period counter.
48 * Contains the current RCU_GP_CTR_BIT.
49 * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
50 * Written to only by writer with mutex taken. Read by both writer and readers.
52 long urcu_gp_ctr
= RCU_GP_COUNT
;
55 * Written to only by each individual reader. Read by both the reader and the
58 long __thread urcu_active_readers
;
60 /* Thread IDs of registered readers */
61 #define INIT_NUM_THREADS 4
63 struct reader_registry
{
65 long *urcu_active_readers
;
70 unsigned int yield_active
;
71 unsigned int __thread rand_yield
;
74 static struct reader_registry
*registry
;
75 static char __thread need_mb
;
76 static int num_readers
, alloc_readers
;
78 void internal_urcu_lock(void)
82 #ifndef DISTRUST_SIGNALS_EXTREME
83 ret
= pthread_mutex_lock(&urcu_mutex
);
85 perror("Error in pthread mutex lock");
88 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
89 while ((ret
= pthread_mutex_trylock(&urcu_mutex
)) != 0) {
90 if (ret
!= EBUSY
&& ret
!= EINTR
) {
91 printf("ret = %d, errno = %d\n", ret
, errno
);
92 perror("Error in pthread mutex lock");
102 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
105 void internal_urcu_unlock(void)
109 ret
= pthread_mutex_unlock(&urcu_mutex
);
111 perror("Error in pthread mutex unlock");
117 * called with urcu_mutex held.
119 static void switch_next_urcu_qparity(void)
121 STORE_SHARED(urcu_gp_ctr
, urcu_gp_ctr
^ RCU_GP_CTR_BIT
);
125 #ifdef HAS_INCOHERENT_CACHES
126 static void force_mb_single_thread(struct reader_registry
*index
)
130 #endif /* #ifdef HAS_INCOHERENT_CACHES */
132 static void force_mb_all_threads(void)
136 #else /* #ifdef DEBUG_FULL_MB */
137 #ifdef HAS_INCOHERENT_CACHES
138 static void force_mb_single_thread(struct reader_registry
*index
)
142 * pthread_kill has a smp_mb(). But beware, we assume it performs
143 * a cache flush on architectures with non-coherent cache. Let's play
144 * safe and don't assume anything : we use smp_mc() to make sure the
145 * cache flush is enforced.
148 smp_mc(); /* write ->need_mb before sending the signals */
149 pthread_kill(index
->tid
, SIGURCU
);
152 * Wait for sighandler (and thus mb()) to execute on every thread.
155 while (*index
->need_mb
) {
158 smp_mb(); /* read ->need_mb before ending the barrier */
160 #endif /* #ifdef HAS_INCOHERENT_CACHES */
162 static void force_mb_all_threads(void)
164 struct reader_registry
*index
;
166 * Ask for each threads to execute a smp_mb() so we can consider the
167 * compiler barriers around rcu read lock as real memory barriers.
172 * pthread_kill has a smp_mb(). But beware, we assume it performs
173 * a cache flush on architectures with non-coherent cache. Let's play
174 * safe and don't assume anything : we use smp_mc() to make sure the
175 * cache flush is enforced.
177 for (index
= registry
; index
< registry
+ num_readers
; index
++) {
179 smp_mc(); /* write need_mb before sending the signal */
180 pthread_kill(index
->tid
, SIGURCU
);
183 * Wait for sighandler (and thus mb()) to execute on every thread.
185 * Note that the pthread_kill() will never be executed on systems
186 * that correctly deliver signals in a timely manner. However, it
187 * is not uncommon for kernels to have bugs that can result in
188 * lost or unduly delayed signals.
190 * If you are seeing the below pthread_kill() executing much at
191 * all, we suggest testing the underlying kernel and filing the
192 * relevant bug report. For Linux kernels, we recommend getting
193 * the Linux Test Project (LTP).
195 for (index
= registry
; index
< registry
+ num_readers
; index
++) {
196 while (*index
->need_mb
) {
197 pthread_kill(index
->tid
, SIGURCU
);
201 smp_mb(); /* read ->need_mb before ending the barrier */
203 #endif /* #else #ifdef DEBUG_FULL_MB */
205 void wait_for_quiescent_state(void)
207 struct reader_registry
*index
;
212 * Wait for each thread urcu_active_readers count to become 0.
214 for (index
= registry
; index
< registry
+ num_readers
; index
++) {
215 #ifndef HAS_INCOHERENT_CACHES
216 while (rcu_old_gp_ongoing(index
->urcu_active_readers
))
218 #else /* #ifndef HAS_INCOHERENT_CACHES */
221 * BUSY-LOOP. Force the reader thread to commit its
222 * urcu_active_readers update to memory if we wait for too long.
224 while (rcu_old_gp_ongoing(index
->urcu_active_readers
)) {
225 if (wait_loops
++ == KICK_READER_LOOPS
) {
226 force_mb_single_thread(index
);
232 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
236 void synchronize_rcu(void)
238 internal_urcu_lock();
240 /* All threads should read qparity before accessing data structure
241 * where new ptr points to. Must be done within internal_urcu_lock
242 * because it iterates on reader threads.*/
243 /* Write new ptr before changing the qparity */
244 force_mb_all_threads();
246 switch_next_urcu_qparity(); /* 0 -> 1 */
249 * Must commit qparity update to memory before waiting for parity
250 * 0 quiescent state. Failure to do so could result in the writer
251 * waiting forever while new readers are always accessing data (no
253 * Ensured by STORE_SHARED and LOAD_SHARED.
257 * Adding a smp_mb() which is _not_ formally required, but makes the
258 * model easier to understand. It does not have a big performance impact
259 * anyway, given this is the write-side.
264 * Wait for previous parity to be empty of readers.
266 wait_for_quiescent_state(); /* Wait readers in parity 0 */
269 * Must finish waiting for quiescent state for parity 0 before
270 * committing qparity update to memory. Failure to do so could result in
271 * the writer waiting forever while new readers are always accessing
272 * data (no progress).
273 * Ensured by STORE_SHARED and LOAD_SHARED.
277 * Adding a smp_mb() which is _not_ formally required, but makes the
278 * model easier to understand. It does not have a big performance impact
279 * anyway, given this is the write-side.
283 switch_next_urcu_qparity(); /* 1 -> 0 */
286 * Must commit qparity update to memory before waiting for parity
287 * 1 quiescent state. Failure to do so could result in the writer
288 * waiting forever while new readers are always accessing data (no
290 * Ensured by STORE_SHARED and LOAD_SHARED.
294 * Adding a 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.
301 * Wait for previous parity to be empty of readers.
303 wait_for_quiescent_state(); /* Wait readers in parity 1 */
305 /* Finish waiting for reader threads before letting the old ptr being
306 * freed. Must be done within internal_urcu_lock because it iterates on
308 force_mb_all_threads();
310 internal_urcu_unlock();
314 * library wrappers to be used by non-LGPL compatible source code.
317 void rcu_read_lock(void)
322 void rcu_read_unlock(void)
327 void *rcu_dereference(void *p
)
329 return _rcu_dereference(p
);
332 void *rcu_assign_pointer_sym(void **p
, void *v
)
335 return STORE_SHARED(p
, v
);
338 void *rcu_xchg_pointer_sym(void **p
, void *v
)
344 void *rcu_publish_content_sym(void **p
, void *v
)
348 oldptr
= _rcu_xchg_pointer(p
, v
);
353 static void rcu_add_reader(pthread_t id
)
355 struct reader_registry
*oldarray
;
358 alloc_readers
= INIT_NUM_THREADS
;
361 malloc(sizeof(struct reader_registry
) * alloc_readers
);
363 if (alloc_readers
< num_readers
+ 1) {
365 registry
= malloc(sizeof(struct reader_registry
)
366 * (alloc_readers
<< 1));
367 memcpy(registry
, oldarray
,
368 sizeof(struct reader_registry
) * alloc_readers
);
372 registry
[num_readers
].tid
= id
;
373 /* reference to the TLS of _this_ reader thread. */
374 registry
[num_readers
].urcu_active_readers
= &urcu_active_readers
;
375 registry
[num_readers
].need_mb
= &need_mb
;
380 * Never shrink (implementation limitation).
381 * This is O(nb threads). Eventually use a hash table.
383 static void rcu_remove_reader(pthread_t id
)
385 struct reader_registry
*index
;
387 assert(registry
!= NULL
);
388 for (index
= registry
; index
< registry
+ num_readers
; index
++) {
389 if (pthread_equal(index
->tid
, id
)) {
390 memcpy(index
, ®istry
[num_readers
- 1],
391 sizeof(struct reader_registry
));
392 registry
[num_readers
- 1].tid
= 0;
393 registry
[num_readers
- 1].urcu_active_readers
= NULL
;
398 /* Hrm not found, forgot to register ? */
402 void rcu_register_thread(void)
404 internal_urcu_lock();
405 urcu_init(); /* In case gcc does not support constructor attribute */
406 rcu_add_reader(pthread_self());
407 internal_urcu_unlock();
410 void rcu_unregister_thread(void)
412 internal_urcu_lock();
413 rcu_remove_reader(pthread_self());
414 internal_urcu_unlock();
417 #ifndef DEBUG_FULL_MB
418 static void sigurcu_handler(int signo
, siginfo_t
*siginfo
, void *context
)
421 * Executing this smp_mb() is the only purpose of this signal handler.
422 * It punctually promotes barrier() into smp_mb() on every thread it is
431 * urcu_init constructor. Called when the library is linked, but also when
432 * reader threads are calling rcu_register_thread().
433 * Should only be called by a single thread at a given time. This is ensured by
434 * holing the internal_urcu_lock() from rcu_register_thread() or by running at
435 * library load time, which should not be executed by multiple threads nor
436 * concurrently with rcu_register_thread() anyway.
440 struct sigaction act
;
447 act
.sa_sigaction
= sigurcu_handler
;
448 ret
= sigaction(SIGURCU
, &act
, NULL
);
450 perror("Error in sigaction");
457 struct sigaction act
;
460 ret
= sigaction(SIGURCU
, NULL
, &act
);
462 perror("Error in sigaction");
465 assert(act
.sa_sigaction
== sigurcu_handler
);
468 #endif /* #ifndef DEBUG_FULL_MB */