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.
26 #define URCU_NO_COMPAT_IDENTIFIERS
29 #define _DEFAULT_SOURCE
40 #include <urcu/config.h>
41 #include <urcu/assert.h>
42 #include <urcu/arch.h>
43 #include <urcu/wfcqueue.h>
44 #include <urcu/map/urcu.h>
45 #include <urcu/static/urcu.h>
46 #include <urcu/pointer.h>
47 #include <urcu/tls-compat.h>
50 #include "urcu-wait.h"
51 #include "urcu-utils.h"
54 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
56 #include <urcu/urcu.h>
60 * If a reader is really non-cooperative and refuses to commit its
61 * rcu_active_readers count to memory (there is no barrier in the reader
62 * per-se), kick it after 10 loops waiting for it.
64 #define KICK_READER_LOOPS 10
67 * Active attempts to check for reader Q.S. before calling futex().
69 #define RCU_QS_ACTIVE_ATTEMPTS 100
71 /* If the headers do not support membarrier system call, fall back on RCU_MB */
72 #ifdef __NR_membarrier
73 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
75 # define membarrier(...) -ENOSYS
79 MEMBARRIER_CMD_QUERY
= 0,
80 MEMBARRIER_CMD_SHARED
= (1 << 0),
81 /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
82 /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
83 MEMBARRIER_CMD_PRIVATE_EXPEDITED
= (1 << 3),
84 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
= (1 << 4),
89 static int urcu_memb_has_sys_membarrier_private_expedited
;
91 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
93 * Explicitly initialize to zero because we can't alias a non-static
94 * uninitialized variable.
96 int urcu_memb_has_sys_membarrier
= 0;
99 void __attribute__((constructor
)) rcu_init(void);
109 static int init_done
;
111 void __attribute__((constructor
)) rcu_init(void);
113 static DEFINE_URCU_TLS(int, rcu_signal_was_blocked
);
116 void __attribute__((destructor
)) rcu_exit(void);
117 static void urcu_call_rcu_exit(void);
120 * rcu_gp_lock ensures mutual exclusion between threads calling
123 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
125 * rcu_registry_lock ensures mutual exclusion between threads
126 * registering and unregistering themselves to/from the registry, and
127 * with threads reading that registry from synchronize_rcu(). However,
128 * this lock is not held all the way through the completion of awaiting
129 * for the grace period. It is sporadically released between iterations
131 * rcu_registry_lock may nest inside rcu_gp_lock.
133 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
134 struct urcu_gp rcu_gp
= { .ctr
= URCU_GP_COUNT
};
137 * Written to only by each individual reader. Read by both the reader and the
140 DEFINE_URCU_TLS(struct urcu_reader
, rcu_reader
);
142 static CDS_LIST_HEAD(registry
);
145 * Queue keeping threads awaiting to wait for a grace period. Contains
146 * struct gp_waiters_thread objects.
148 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
150 static void mutex_lock(pthread_mutex_t
*mutex
)
154 #ifndef DISTRUST_SIGNALS_EXTREME
155 ret
= pthread_mutex_lock(mutex
);
158 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
159 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
160 if (ret
!= EBUSY
&& ret
!= EINTR
)
162 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
164 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
167 (void) poll(NULL
, 0, 10);
169 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
172 static void mutex_unlock(pthread_mutex_t
*mutex
)
176 ret
= pthread_mutex_unlock(mutex
);
181 #ifdef RCU_MEMBARRIER
182 static void smp_mb_master(void)
184 if (caa_likely(urcu_memb_has_sys_membarrier
)) {
185 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited
?
186 MEMBARRIER_CMD_PRIVATE_EXPEDITED
:
187 MEMBARRIER_CMD_SHARED
, 0))
196 static void smp_mb_master(void)
203 static void force_mb_all_readers(void)
205 struct urcu_reader
*index
;
208 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
209 * compiler barriers around rcu read lock as real memory barriers.
211 if (cds_list_empty(®istry
))
214 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
215 * a cache flush on architectures with non-coherent cache. Let's play
216 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
217 * cache flush is enforced.
219 cds_list_for_each_entry(index
, ®istry
, node
) {
220 CMM_STORE_SHARED(index
->need_mb
, 1);
221 pthread_kill(index
->tid
, SIGRCU
);
224 * Wait for sighandler (and thus mb()) to execute on every thread.
226 * Note that the pthread_kill() will never be executed on systems
227 * that correctly deliver signals in a timely manner. However, it
228 * is not uncommon for kernels to have bugs that can result in
229 * lost or unduly delayed signals.
231 * If you are seeing the below pthread_kill() executing much at
232 * all, we suggest testing the underlying kernel and filing the
233 * relevant bug report. For Linux kernels, we recommend getting
234 * the Linux Test Project (LTP).
236 cds_list_for_each_entry(index
, ®istry
, node
) {
237 while (CMM_LOAD_SHARED(index
->need_mb
)) {
238 pthread_kill(index
->tid
, SIGRCU
);
239 (void) poll(NULL
, 0, 1);
242 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
245 static void smp_mb_master(void)
247 force_mb_all_readers();
249 #endif /* #ifdef RCU_SIGNAL */
252 * synchronize_rcu() waiting. Single thread.
253 * Always called with rcu_registry lock held. Releases this lock and
254 * grabs it again. Holds the lock when it returns.
256 static void wait_gp(void)
259 * Read reader_gp before read futex. smp_mb_master() needs to
260 * be called with the rcu registry lock held in RCU_SIGNAL
264 /* Temporarily unlock the registry lock. */
265 mutex_unlock(&rcu_registry_lock
);
266 while (uatomic_read(&rcu_gp
.futex
) == -1) {
267 if (!futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
269 * Prior queued wakeups queued by unrelated code
270 * using the same address can cause futex wait to
271 * return 0 even through the futex value is still
272 * -1 (spurious wakeups). Check the value again
273 * in user-space to validate whether it really
280 /* Value already changed. */
283 /* Retry if interrupted by signal. */
284 break; /* Get out of switch. Check again. */
286 /* Unexpected error. */
292 * Re-lock the registry lock before the next loop.
294 mutex_lock(&rcu_registry_lock
);
298 * Always called with rcu_registry lock held. Releases this lock between
299 * iterations and grabs it again. Holds the lock when it returns.
301 static void wait_for_readers(struct cds_list_head
*input_readers
,
302 struct cds_list_head
*cur_snap_readers
,
303 struct cds_list_head
*qsreaders
)
305 unsigned int wait_loops
= 0;
306 struct urcu_reader
*index
, *tmp
;
307 #ifdef HAS_INCOHERENT_CACHES
308 unsigned int wait_gp_loops
= 0;
309 #endif /* HAS_INCOHERENT_CACHES */
312 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
313 * indicate quiescence (not nested), or observe the current
317 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
319 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
320 uatomic_dec(&rcu_gp
.futex
);
321 /* Write futex before read reader_gp */
325 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
326 switch (urcu_common_reader_state(&rcu_gp
, &index
->ctr
)) {
327 case URCU_READER_ACTIVE_CURRENT
:
328 if (cur_snap_readers
) {
329 cds_list_move(&index
->node
,
334 case URCU_READER_INACTIVE
:
335 cds_list_move(&index
->node
, qsreaders
);
337 case URCU_READER_ACTIVE_OLD
:
339 * Old snapshot. Leaving node in
340 * input_readers will make us busy-loop
341 * until the snapshot becomes current or
342 * the reader becomes inactive.
348 #ifndef HAS_INCOHERENT_CACHES
349 if (cds_list_empty(input_readers
)) {
350 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
351 /* Read reader_gp before write futex */
353 uatomic_set(&rcu_gp
.futex
, 0);
357 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
358 /* wait_gp unlocks/locks registry lock. */
361 /* Temporarily unlock the registry lock. */
362 mutex_unlock(&rcu_registry_lock
);
365 * Re-lock the registry lock before the
368 mutex_lock(&rcu_registry_lock
);
371 #else /* #ifndef HAS_INCOHERENT_CACHES */
373 * BUSY-LOOP. Force the reader thread to commit its
374 * URCU_TLS(rcu_reader).ctr update to memory if we wait
377 if (cds_list_empty(input_readers
)) {
378 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
379 /* Read reader_gp before write futex */
381 uatomic_set(&rcu_gp
.futex
, 0);
385 if (wait_gp_loops
== KICK_READER_LOOPS
) {
389 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
390 /* wait_gp unlocks/locks registry lock. */
394 /* Temporarily unlock the registry lock. */
395 mutex_unlock(&rcu_registry_lock
);
398 * Re-lock the registry lock before the
401 mutex_lock(&rcu_registry_lock
);
404 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
408 void synchronize_rcu(void)
410 CDS_LIST_HEAD(cur_snap_readers
);
411 CDS_LIST_HEAD(qsreaders
);
412 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
413 struct urcu_waiters waiters
;
416 * Add ourself to gp_waiters queue of threads awaiting to wait
417 * for a grace period. Proceed to perform the grace period only
418 * if we are the first thread added into the queue.
419 * The implicit memory barrier before urcu_wait_add()
420 * orders prior memory accesses of threads put into the wait
421 * queue before their insertion into the wait queue.
423 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
424 /* Not first in queue: will be awakened by another thread. */
425 urcu_adaptative_busy_wait(&wait
);
426 /* Order following memory accesses after grace period. */
430 /* We won't need to wake ourself up */
431 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
433 mutex_lock(&rcu_gp_lock
);
436 * Move all waiters into our local queue.
438 urcu_move_waiters(&waiters
, &gp_waiters
);
440 mutex_lock(&rcu_registry_lock
);
442 if (cds_list_empty(®istry
))
446 * All threads should read qparity before accessing data structure
447 * where new ptr points to. Must be done within rcu_registry_lock
448 * because it iterates on reader threads.
450 /* Write new ptr before changing the qparity */
454 * Wait for readers to observe original parity or be quiescent.
455 * wait_for_readers() can release and grab again rcu_registry_lock
458 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
461 * Must finish waiting for quiescent state for original parity before
462 * committing next rcu_gp.ctr update to memory. Failure to do so could
463 * result in the writer waiting forever while new readers are always
464 * accessing data (no progress). Enforce compiler-order of load
465 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
470 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
471 * model easier to understand. It does not have a big performance impact
472 * anyway, given this is the write-side.
476 /* Switch parity: 0 -> 1, 1 -> 0 */
477 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ URCU_GP_CTR_PHASE
);
480 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
481 * state. Failure to do so could result in the writer waiting forever
482 * while new readers are always accessing data (no progress). Enforce
483 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
489 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
490 * model easier to understand. It does not have a big performance impact
491 * anyway, given this is the write-side.
496 * Wait for readers to observe new parity or be quiescent.
497 * wait_for_readers() can release and grab again rcu_registry_lock
500 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
503 * Put quiescent reader list back into registry.
505 cds_list_splice(&qsreaders
, ®istry
);
508 * Finish waiting for reader threads before letting the old ptr
509 * being freed. Must be done within rcu_registry_lock because it
510 * iterates on reader threads.
514 mutex_unlock(&rcu_registry_lock
);
515 mutex_unlock(&rcu_gp_lock
);
518 * Wakeup waiters only after we have completed the grace period
519 * and have ensured the memory barriers at the end of the grace
520 * period have been issued.
522 urcu_wake_all_waiters(&waiters
);
526 * library wrappers to be used by non-LGPL compatible source code.
529 void rcu_read_lock(void)
534 void rcu_read_unlock(void)
539 int rcu_read_ongoing(void)
541 return _rcu_read_ongoing();
546 * Make sure the signal used by the urcu-signal flavor is unblocked
547 * while the thread is registered.
550 void urcu_signal_unblock(void)
552 sigset_t mask
, oldmask
;
555 ret
= sigemptyset(&mask
);
556 urcu_posix_assert(!ret
);
557 ret
= sigaddset(&mask
, SIGRCU
);
558 urcu_posix_assert(!ret
);
559 ret
= pthread_sigmask(SIG_UNBLOCK
, &mask
, &oldmask
);
560 urcu_posix_assert(!ret
);
561 URCU_TLS(rcu_signal_was_blocked
) = sigismember(&oldmask
, SIGRCU
);
565 void urcu_signal_restore(void)
570 if (!URCU_TLS(rcu_signal_was_blocked
))
572 ret
= sigemptyset(&mask
);
573 urcu_posix_assert(!ret
);
574 ret
= sigaddset(&mask
, SIGRCU
);
575 urcu_posix_assert(!ret
);
576 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
577 urcu_posix_assert(!ret
);
581 void urcu_signal_unblock(void) { }
583 void urcu_signal_restore(void) { }
586 void rcu_register_thread(void)
588 urcu_signal_unblock();
590 URCU_TLS(rcu_reader
).tid
= pthread_self();
591 urcu_posix_assert(URCU_TLS(rcu_reader
).need_mb
== 0);
592 urcu_posix_assert(!(URCU_TLS(rcu_reader
).ctr
& URCU_GP_CTR_NEST_MASK
));
594 mutex_lock(&rcu_registry_lock
);
595 urcu_posix_assert(!URCU_TLS(rcu_reader
).registered
);
596 URCU_TLS(rcu_reader
).registered
= 1;
597 rcu_init(); /* In case gcc does not support constructor attribute */
598 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
599 mutex_unlock(&rcu_registry_lock
);
602 void rcu_unregister_thread(void)
604 mutex_lock(&rcu_registry_lock
);
605 urcu_posix_assert(URCU_TLS(rcu_reader
).registered
);
606 URCU_TLS(rcu_reader
).registered
= 0;
607 cds_list_del(&URCU_TLS(rcu_reader
).node
);
608 mutex_unlock(&rcu_registry_lock
);
610 urcu_signal_restore();
613 #ifdef RCU_MEMBARRIER
615 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
617 void rcu_sys_membarrier_status(bool available
)
624 void rcu_sys_membarrier_status(bool available
)
628 urcu_memb_has_sys_membarrier
= 1;
633 void rcu_sys_membarrier_init(void)
635 bool available
= false;
638 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
640 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
641 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
643 urcu_memb_has_sys_membarrier_private_expedited
= 1;
645 } else if (mask
& MEMBARRIER_CMD_SHARED
) {
649 rcu_sys_membarrier_status(available
);
657 rcu_sys_membarrier_init();
662 static void sigrcu_handler(int signo
__attribute__((unused
)),
663 siginfo_t
*siginfo
__attribute__((unused
)),
664 void *context
__attribute__((unused
)))
667 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
668 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
672 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
677 * rcu_init constructor. Called when the library is linked, but also when
678 * reader threads are calling rcu_register_thread().
679 * Should only be called by a single thread at a given time. This is ensured by
680 * holing the rcu_registry_lock from rcu_register_thread() or by running
681 * at library load time, which should not be executed by multiple
682 * threads nor concurrently with rcu_register_thread() anyway.
686 struct sigaction act
;
693 act
.sa_sigaction
= sigrcu_handler
;
694 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
695 sigemptyset(&act
.sa_mask
);
696 ret
= sigaction(SIGRCU
, &act
, NULL
);
702 * Don't unregister the SIGRCU signal handler anymore, because
703 * call_rcu threads could still be using it shortly before the
705 * Assertion disabled because call_rcu threads are now rcu
706 * readers, and left running at exit.
707 * urcu_posix_assert(cds_list_empty(®istry));
710 #endif /* #ifdef RCU_SIGNAL */
714 urcu_call_rcu_exit();
717 DEFINE_RCU_FLAVOR(rcu_flavor
);
719 #include "urcu-call-rcu-impl.h"
720 #include "urcu-defer-impl.h"
721 #include "urcu-poll-impl.h"