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.
29 #define _DEFAULT_SOURCE
40 #include "urcu/arch.h"
41 #include "urcu/wfcqueue.h"
42 #include "urcu/map/urcu.h"
43 #include "urcu/static/urcu.h"
44 #include "urcu-pointer.h"
45 #include "urcu/tls-compat.h"
48 #include "urcu-wait.h"
50 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
56 * If a reader is really non-cooperative and refuses to commit its
57 * rcu_active_readers count to memory (there is no barrier in the reader
58 * per-se), kick it after 10 loops waiting for it.
60 #define KICK_READER_LOOPS 10
63 * Active attempts to check for reader Q.S. before calling futex().
65 #define RCU_QS_ACTIVE_ATTEMPTS 100
67 /* If the headers do not support membarrier system call, fall back on RCU_MB */
68 #ifdef __NR_membarrier
69 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
71 # define membarrier(...) -ENOSYS
75 MEMBARRIER_CMD_QUERY
= 0,
76 MEMBARRIER_CMD_SHARED
= (1 << 0),
81 int rcu_has_sys_membarrier
;
83 void __attribute__((constructor
)) rcu_init(void);
95 void __attribute__((constructor
)) rcu_init(void);
96 void __attribute__((destructor
)) rcu_exit(void);
100 * rcu_gp_lock ensures mutual exclusion between threads calling
103 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
105 * rcu_registry_lock ensures mutual exclusion between threads
106 * registering and unregistering themselves to/from the registry, and
107 * with threads reading that registry from synchronize_rcu(). However,
108 * this lock is not held all the way through the completion of awaiting
109 * for the grace period. It is sporadically released between iterations
111 * rcu_registry_lock may nest inside rcu_gp_lock.
113 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
114 struct rcu_gp rcu_gp
= { .ctr
= RCU_GP_COUNT
};
117 * Written to only by each individual reader. Read by both the reader and the
120 DEFINE_URCU_TLS(struct rcu_reader
, rcu_reader
);
122 static CDS_LIST_HEAD(registry
);
125 * Queue keeping threads awaiting to wait for a grace period. Contains
126 * struct gp_waiters_thread objects.
128 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
130 static void mutex_lock(pthread_mutex_t
*mutex
)
134 #ifndef DISTRUST_SIGNALS_EXTREME
135 ret
= pthread_mutex_lock(mutex
);
138 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
139 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
140 if (ret
!= EBUSY
&& ret
!= EINTR
)
142 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
144 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
147 (void) poll(NULL
, 0, 10);
149 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
152 static void mutex_unlock(pthread_mutex_t
*mutex
)
156 ret
= pthread_mutex_unlock(mutex
);
161 #ifdef RCU_MEMBARRIER
162 static void smp_mb_master(void)
164 if (caa_likely(rcu_has_sys_membarrier
))
165 (void) membarrier(MEMBARRIER_CMD_SHARED
, 0);
172 static void smp_mb_master(void)
179 static void force_mb_all_readers(void)
181 struct rcu_reader
*index
;
184 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
185 * compiler barriers around rcu read lock as real memory barriers.
187 if (cds_list_empty(®istry
))
190 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
191 * a cache flush on architectures with non-coherent cache. Let's play
192 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
193 * cache flush is enforced.
195 cds_list_for_each_entry(index
, ®istry
, node
) {
196 CMM_STORE_SHARED(index
->need_mb
, 1);
197 pthread_kill(index
->tid
, SIGRCU
);
200 * Wait for sighandler (and thus mb()) to execute on every thread.
202 * Note that the pthread_kill() will never be executed on systems
203 * that correctly deliver signals in a timely manner. However, it
204 * is not uncommon for kernels to have bugs that can result in
205 * lost or unduly delayed signals.
207 * If you are seeing the below pthread_kill() executing much at
208 * all, we suggest testing the underlying kernel and filing the
209 * relevant bug report. For Linux kernels, we recommend getting
210 * the Linux Test Project (LTP).
212 cds_list_for_each_entry(index
, ®istry
, node
) {
213 while (CMM_LOAD_SHARED(index
->need_mb
)) {
214 pthread_kill(index
->tid
, SIGRCU
);
215 (void) poll(NULL
, 0, 1);
218 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
221 static void smp_mb_master(void)
223 force_mb_all_readers();
225 #endif /* #ifdef RCU_SIGNAL */
228 * synchronize_rcu() waiting. Single thread.
230 static void wait_gp(void)
232 /* Read reader_gp before read futex */
234 if (uatomic_read(&rcu_gp
.futex
) != -1)
236 while (futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1,
240 /* Value already changed. */
243 /* Retry if interrupted by signal. */
244 break; /* Get out of switch. */
246 /* Unexpected error. */
253 * Always called with rcu_registry lock held. Releases this lock between
254 * iterations and grabs it again. Holds the lock when it returns.
256 static void wait_for_readers(struct cds_list_head
*input_readers
,
257 struct cds_list_head
*cur_snap_readers
,
258 struct cds_list_head
*qsreaders
)
260 unsigned int wait_loops
= 0;
261 struct rcu_reader
*index
, *tmp
;
262 #ifdef HAS_INCOHERENT_CACHES
263 unsigned int wait_gp_loops
= 0;
264 #endif /* HAS_INCOHERENT_CACHES */
267 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
268 * indicate quiescence (not nested), or observe the current
272 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
274 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
275 uatomic_dec(&rcu_gp
.futex
);
276 /* Write futex before read reader_gp */
280 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
281 switch (rcu_reader_state(&index
->ctr
)) {
282 case RCU_READER_ACTIVE_CURRENT
:
283 if (cur_snap_readers
) {
284 cds_list_move(&index
->node
,
289 case RCU_READER_INACTIVE
:
290 cds_list_move(&index
->node
, qsreaders
);
292 case RCU_READER_ACTIVE_OLD
:
294 * Old snapshot. Leaving node in
295 * input_readers will make us busy-loop
296 * until the snapshot becomes current or
297 * the reader becomes inactive.
303 #ifndef HAS_INCOHERENT_CACHES
304 if (cds_list_empty(input_readers
)) {
305 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
306 /* Read reader_gp before write futex */
308 uatomic_set(&rcu_gp
.futex
, 0);
312 /* Temporarily unlock the registry lock. */
313 mutex_unlock(&rcu_registry_lock
);
314 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
)
318 /* Re-lock the registry lock before the next loop. */
319 mutex_lock(&rcu_registry_lock
);
321 #else /* #ifndef HAS_INCOHERENT_CACHES */
323 * BUSY-LOOP. Force the reader thread to commit its
324 * URCU_TLS(rcu_reader).ctr update to memory if we wait
327 if (cds_list_empty(input_readers
)) {
328 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
329 /* Read reader_gp before write futex */
331 uatomic_set(&rcu_gp
.futex
, 0);
335 if (wait_gp_loops
== KICK_READER_LOOPS
) {
339 /* Temporarily unlock the registry lock. */
340 mutex_unlock(&rcu_registry_lock
);
341 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
347 /* Re-lock the registry lock before the next loop. */
348 mutex_lock(&rcu_registry_lock
);
350 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
354 void synchronize_rcu(void)
356 CDS_LIST_HEAD(cur_snap_readers
);
357 CDS_LIST_HEAD(qsreaders
);
358 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
359 struct urcu_waiters waiters
;
362 * Add ourself to gp_waiters queue of threads awaiting to wait
363 * for a grace period. Proceed to perform the grace period only
364 * if we are the first thread added into the queue.
365 * The implicit memory barrier before urcu_wait_add()
366 * orders prior memory accesses of threads put into the wait
367 * queue before their insertion into the wait queue.
369 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
370 /* Not first in queue: will be awakened by another thread. */
371 urcu_adaptative_busy_wait(&wait
);
372 /* Order following memory accesses after grace period. */
376 /* We won't need to wake ourself up */
377 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
379 mutex_lock(&rcu_gp_lock
);
382 * Move all waiters into our local queue.
384 urcu_move_waiters(&waiters
, &gp_waiters
);
386 mutex_lock(&rcu_registry_lock
);
388 if (cds_list_empty(®istry
))
392 * All threads should read qparity before accessing data structure
393 * where new ptr points to. Must be done within rcu_registry_lock
394 * because it iterates on reader threads.
396 /* Write new ptr before changing the qparity */
400 * Wait for readers to observe original parity or be quiescent.
401 * wait_for_readers() can release and grab again rcu_registry_lock
404 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
407 * Must finish waiting for quiescent state for original parity before
408 * committing next rcu_gp.ctr update to memory. Failure to do so could
409 * result in the writer waiting forever while new readers are always
410 * accessing data (no progress). Enforce compiler-order of load
411 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
416 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
417 * model easier to understand. It does not have a big performance impact
418 * anyway, given this is the write-side.
422 /* Switch parity: 0 -> 1, 1 -> 0 */
423 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ RCU_GP_CTR_PHASE
);
426 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
427 * state. Failure to do so could result in the writer waiting forever
428 * while new readers are always accessing data (no progress). Enforce
429 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
435 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
436 * model easier to understand. It does not have a big performance impact
437 * anyway, given this is the write-side.
442 * Wait for readers to observe new parity or be quiescent.
443 * wait_for_readers() can release and grab again rcu_registry_lock
446 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
449 * Put quiescent reader list back into registry.
451 cds_list_splice(&qsreaders
, ®istry
);
454 * Finish waiting for reader threads before letting the old ptr
455 * being freed. Must be done within rcu_registry_lock because it
456 * iterates on reader threads.
460 mutex_unlock(&rcu_registry_lock
);
461 mutex_unlock(&rcu_gp_lock
);
464 * Wakeup waiters only after we have completed the grace period
465 * and have ensured the memory barriers at the end of the grace
466 * period have been issued.
468 urcu_wake_all_waiters(&waiters
);
472 * library wrappers to be used by non-LGPL compatible source code.
475 void rcu_read_lock(void)
480 void rcu_read_unlock(void)
485 int rcu_read_ongoing(void)
487 return _rcu_read_ongoing();
490 void rcu_register_thread(void)
492 URCU_TLS(rcu_reader
).tid
= pthread_self();
493 assert(URCU_TLS(rcu_reader
).need_mb
== 0);
494 assert(!(URCU_TLS(rcu_reader
).ctr
& RCU_GP_CTR_NEST_MASK
));
496 mutex_lock(&rcu_registry_lock
);
497 assert(!URCU_TLS(rcu_reader
).registered
);
498 URCU_TLS(rcu_reader
).registered
= 1;
499 rcu_init(); /* In case gcc does not support constructor attribute */
500 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
501 mutex_unlock(&rcu_registry_lock
);
504 void rcu_unregister_thread(void)
506 mutex_lock(&rcu_registry_lock
);
507 assert(URCU_TLS(rcu_reader
).registered
);
508 URCU_TLS(rcu_reader
).registered
= 0;
509 cds_list_del(&URCU_TLS(rcu_reader
).node
);
510 mutex_unlock(&rcu_registry_lock
);
513 #ifdef RCU_MEMBARRIER
521 ret
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
522 if (ret
>= 0 && (ret
& MEMBARRIER_CMD_SHARED
)) {
523 rcu_has_sys_membarrier
= 1;
529 static void sigrcu_handler(int signo
, siginfo_t
*siginfo
, void *context
)
532 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
533 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
537 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
542 * rcu_init constructor. Called when the library is linked, but also when
543 * reader threads are calling rcu_register_thread().
544 * Should only be called by a single thread at a given time. This is ensured by
545 * holing the rcu_registry_lock from rcu_register_thread() or by running
546 * at library load time, which should not be executed by multiple
547 * threads nor concurrently with rcu_register_thread() anyway.
551 struct sigaction act
;
558 act
.sa_sigaction
= sigrcu_handler
;
559 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
560 sigemptyset(&act
.sa_mask
);
561 ret
= sigaction(SIGRCU
, &act
, NULL
);
569 * Don't unregister the SIGRCU signal handler anymore, because
570 * call_rcu threads could still be using it shortly before the
572 * Assertion disabled because call_rcu threads are now rcu
573 * readers, and left running at exit.
574 * assert(cds_list_empty(®istry));
578 #endif /* #ifdef RCU_SIGNAL */
580 DEFINE_RCU_FLAVOR(rcu_flavor
);
582 #include "urcu-call-rcu-impl.h"
583 #include "urcu-defer-impl.h"