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);
112 void __attribute__((destructor
)) rcu_exit(void);
114 static DEFINE_URCU_TLS(int, rcu_signal_was_blocked
);
118 * rcu_gp_lock ensures mutual exclusion between threads calling
121 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
123 * rcu_registry_lock ensures mutual exclusion between threads
124 * registering and unregistering themselves to/from the registry, and
125 * with threads reading that registry from synchronize_rcu(). However,
126 * this lock is not held all the way through the completion of awaiting
127 * for the grace period. It is sporadically released between iterations
129 * rcu_registry_lock may nest inside rcu_gp_lock.
131 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
132 struct urcu_gp rcu_gp
= { .ctr
= URCU_GP_COUNT
};
135 * Written to only by each individual reader. Read by both the reader and the
138 DEFINE_URCU_TLS(struct urcu_reader
, rcu_reader
);
140 static CDS_LIST_HEAD(registry
);
143 * Queue keeping threads awaiting to wait for a grace period. Contains
144 * struct gp_waiters_thread objects.
146 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
148 static void mutex_lock(pthread_mutex_t
*mutex
)
152 #ifndef DISTRUST_SIGNALS_EXTREME
153 ret
= pthread_mutex_lock(mutex
);
156 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
157 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
158 if (ret
!= EBUSY
&& ret
!= EINTR
)
160 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
162 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
165 (void) poll(NULL
, 0, 10);
167 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
170 static void mutex_unlock(pthread_mutex_t
*mutex
)
174 ret
= pthread_mutex_unlock(mutex
);
179 #ifdef RCU_MEMBARRIER
180 static void smp_mb_master(void)
182 if (caa_likely(urcu_memb_has_sys_membarrier
)) {
183 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited
?
184 MEMBARRIER_CMD_PRIVATE_EXPEDITED
:
185 MEMBARRIER_CMD_SHARED
, 0))
194 static void smp_mb_master(void)
201 static void force_mb_all_readers(void)
203 struct urcu_reader
*index
;
206 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
207 * compiler barriers around rcu read lock as real memory barriers.
209 if (cds_list_empty(®istry
))
212 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
213 * a cache flush on architectures with non-coherent cache. Let's play
214 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
215 * cache flush is enforced.
217 cds_list_for_each_entry(index
, ®istry
, node
) {
218 CMM_STORE_SHARED(index
->need_mb
, 1);
219 pthread_kill(index
->tid
, SIGRCU
);
222 * Wait for sighandler (and thus mb()) to execute on every thread.
224 * Note that the pthread_kill() will never be executed on systems
225 * that correctly deliver signals in a timely manner. However, it
226 * is not uncommon for kernels to have bugs that can result in
227 * lost or unduly delayed signals.
229 * If you are seeing the below pthread_kill() executing much at
230 * all, we suggest testing the underlying kernel and filing the
231 * relevant bug report. For Linux kernels, we recommend getting
232 * the Linux Test Project (LTP).
234 cds_list_for_each_entry(index
, ®istry
, node
) {
235 while (CMM_LOAD_SHARED(index
->need_mb
)) {
236 pthread_kill(index
->tid
, SIGRCU
);
237 (void) poll(NULL
, 0, 1);
240 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
243 static void smp_mb_master(void)
245 force_mb_all_readers();
247 #endif /* #ifdef RCU_SIGNAL */
250 * synchronize_rcu() waiting. Single thread.
251 * Always called with rcu_registry lock held. Releases this lock and
252 * grabs it again. Holds the lock when it returns.
254 static void wait_gp(void)
257 * Read reader_gp before read futex. smp_mb_master() needs to
258 * be called with the rcu registry lock held in RCU_SIGNAL
262 /* Temporarily unlock the registry lock. */
263 mutex_unlock(&rcu_registry_lock
);
264 while (uatomic_read(&rcu_gp
.futex
) == -1) {
265 if (!futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
267 * Prior queued wakeups queued by unrelated code
268 * using the same address can cause futex wait to
269 * return 0 even through the futex value is still
270 * -1 (spurious wakeups). Check the value again
271 * in user-space to validate whether it really
278 /* Value already changed. */
281 /* Retry if interrupted by signal. */
282 break; /* Get out of switch. Check again. */
284 /* Unexpected error. */
290 * Re-lock the registry lock before the next loop.
292 mutex_lock(&rcu_registry_lock
);
296 * Always called with rcu_registry lock held. Releases this lock between
297 * iterations and grabs it again. Holds the lock when it returns.
299 static void wait_for_readers(struct cds_list_head
*input_readers
,
300 struct cds_list_head
*cur_snap_readers
,
301 struct cds_list_head
*qsreaders
)
303 unsigned int wait_loops
= 0;
304 struct urcu_reader
*index
, *tmp
;
305 #ifdef HAS_INCOHERENT_CACHES
306 unsigned int wait_gp_loops
= 0;
307 #endif /* HAS_INCOHERENT_CACHES */
310 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
311 * indicate quiescence (not nested), or observe the current
315 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
317 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
318 uatomic_dec(&rcu_gp
.futex
);
319 /* Write futex before read reader_gp */
323 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
324 switch (urcu_common_reader_state(&rcu_gp
, &index
->ctr
)) {
325 case URCU_READER_ACTIVE_CURRENT
:
326 if (cur_snap_readers
) {
327 cds_list_move(&index
->node
,
332 case URCU_READER_INACTIVE
:
333 cds_list_move(&index
->node
, qsreaders
);
335 case URCU_READER_ACTIVE_OLD
:
337 * Old snapshot. Leaving node in
338 * input_readers will make us busy-loop
339 * until the snapshot becomes current or
340 * the reader becomes inactive.
346 #ifndef HAS_INCOHERENT_CACHES
347 if (cds_list_empty(input_readers
)) {
348 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
349 /* Read reader_gp before write futex */
351 uatomic_set(&rcu_gp
.futex
, 0);
355 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
356 /* wait_gp unlocks/locks registry lock. */
359 /* Temporarily unlock the registry lock. */
360 mutex_unlock(&rcu_registry_lock
);
363 * Re-lock the registry lock before the
366 mutex_lock(&rcu_registry_lock
);
369 #else /* #ifndef HAS_INCOHERENT_CACHES */
371 * BUSY-LOOP. Force the reader thread to commit its
372 * URCU_TLS(rcu_reader).ctr update to memory if we wait
375 if (cds_list_empty(input_readers
)) {
376 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
377 /* Read reader_gp before write futex */
379 uatomic_set(&rcu_gp
.futex
, 0);
383 if (wait_gp_loops
== KICK_READER_LOOPS
) {
387 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
388 /* wait_gp unlocks/locks registry lock. */
392 /* Temporarily unlock the registry lock. */
393 mutex_unlock(&rcu_registry_lock
);
396 * Re-lock the registry lock before the
399 mutex_lock(&rcu_registry_lock
);
402 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
406 void synchronize_rcu(void)
408 CDS_LIST_HEAD(cur_snap_readers
);
409 CDS_LIST_HEAD(qsreaders
);
410 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
411 struct urcu_waiters waiters
;
414 * Add ourself to gp_waiters queue of threads awaiting to wait
415 * for a grace period. Proceed to perform the grace period only
416 * if we are the first thread added into the queue.
417 * The implicit memory barrier before urcu_wait_add()
418 * orders prior memory accesses of threads put into the wait
419 * queue before their insertion into the wait queue.
421 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
422 /* Not first in queue: will be awakened by another thread. */
423 urcu_adaptative_busy_wait(&wait
);
424 /* Order following memory accesses after grace period. */
428 /* We won't need to wake ourself up */
429 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
431 mutex_lock(&rcu_gp_lock
);
434 * Move all waiters into our local queue.
436 urcu_move_waiters(&waiters
, &gp_waiters
);
438 mutex_lock(&rcu_registry_lock
);
440 if (cds_list_empty(®istry
))
444 * All threads should read qparity before accessing data structure
445 * where new ptr points to. Must be done within rcu_registry_lock
446 * because it iterates on reader threads.
448 /* Write new ptr before changing the qparity */
452 * Wait for readers to observe original parity or be quiescent.
453 * wait_for_readers() can release and grab again rcu_registry_lock
456 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
459 * Must finish waiting for quiescent state for original parity before
460 * committing next rcu_gp.ctr update to memory. Failure to do so could
461 * result in the writer waiting forever while new readers are always
462 * accessing data (no progress). Enforce compiler-order of load
463 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
468 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
469 * model easier to understand. It does not have a big performance impact
470 * anyway, given this is the write-side.
474 /* Switch parity: 0 -> 1, 1 -> 0 */
475 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ URCU_GP_CTR_PHASE
);
478 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
479 * state. Failure to do so could result in the writer waiting forever
480 * while new readers are always accessing data (no progress). Enforce
481 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
487 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
488 * model easier to understand. It does not have a big performance impact
489 * anyway, given this is the write-side.
494 * Wait for readers to observe new parity or be quiescent.
495 * wait_for_readers() can release and grab again rcu_registry_lock
498 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
501 * Put quiescent reader list back into registry.
503 cds_list_splice(&qsreaders
, ®istry
);
506 * Finish waiting for reader threads before letting the old ptr
507 * being freed. Must be done within rcu_registry_lock because it
508 * iterates on reader threads.
512 mutex_unlock(&rcu_registry_lock
);
513 mutex_unlock(&rcu_gp_lock
);
516 * Wakeup waiters only after we have completed the grace period
517 * and have ensured the memory barriers at the end of the grace
518 * period have been issued.
520 urcu_wake_all_waiters(&waiters
);
524 * library wrappers to be used by non-LGPL compatible source code.
527 void rcu_read_lock(void)
532 void rcu_read_unlock(void)
537 int rcu_read_ongoing(void)
539 return _rcu_read_ongoing();
544 * Make sure the signal used by the urcu-signal flavor is unblocked
545 * while the thread is registered.
548 void urcu_signal_unblock(void)
550 sigset_t mask
, oldmask
;
553 ret
= sigemptyset(&mask
);
554 urcu_posix_assert(!ret
);
555 ret
= sigaddset(&mask
, SIGRCU
);
556 urcu_posix_assert(!ret
);
557 ret
= pthread_sigmask(SIG_UNBLOCK
, &mask
, &oldmask
);
558 urcu_posix_assert(!ret
);
559 URCU_TLS(rcu_signal_was_blocked
) = sigismember(&oldmask
, SIGRCU
);
563 void urcu_signal_restore(void)
568 if (!URCU_TLS(rcu_signal_was_blocked
))
570 ret
= sigemptyset(&mask
);
571 urcu_posix_assert(!ret
);
572 ret
= sigaddset(&mask
, SIGRCU
);
573 urcu_posix_assert(!ret
);
574 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
575 urcu_posix_assert(!ret
);
579 void urcu_signal_unblock(void) { }
581 void urcu_signal_restore(void) { }
584 void rcu_register_thread(void)
586 urcu_signal_unblock();
588 URCU_TLS(rcu_reader
).tid
= pthread_self();
589 urcu_posix_assert(URCU_TLS(rcu_reader
).need_mb
== 0);
590 urcu_posix_assert(!(URCU_TLS(rcu_reader
).ctr
& URCU_GP_CTR_NEST_MASK
));
592 mutex_lock(&rcu_registry_lock
);
593 urcu_posix_assert(!URCU_TLS(rcu_reader
).registered
);
594 URCU_TLS(rcu_reader
).registered
= 1;
595 rcu_init(); /* In case gcc does not support constructor attribute */
596 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
597 mutex_unlock(&rcu_registry_lock
);
600 void rcu_unregister_thread(void)
602 mutex_lock(&rcu_registry_lock
);
603 urcu_posix_assert(URCU_TLS(rcu_reader
).registered
);
604 URCU_TLS(rcu_reader
).registered
= 0;
605 cds_list_del(&URCU_TLS(rcu_reader
).node
);
606 mutex_unlock(&rcu_registry_lock
);
608 urcu_signal_restore();
611 #ifdef RCU_MEMBARRIER
613 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
615 void rcu_sys_membarrier_status(bool available
)
622 void rcu_sys_membarrier_status(bool available
)
626 urcu_memb_has_sys_membarrier
= 1;
631 void rcu_sys_membarrier_init(void)
633 bool available
= false;
636 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
638 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
639 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
641 urcu_memb_has_sys_membarrier_private_expedited
= 1;
643 } else if (mask
& MEMBARRIER_CMD_SHARED
) {
647 rcu_sys_membarrier_status(available
);
655 rcu_sys_membarrier_init();
660 static void sigrcu_handler(int signo
__attribute__((unused
)),
661 siginfo_t
*siginfo
__attribute__((unused
)),
662 void *context
__attribute__((unused
)))
665 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
666 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
670 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
675 * rcu_init constructor. Called when the library is linked, but also when
676 * reader threads are calling rcu_register_thread().
677 * Should only be called by a single thread at a given time. This is ensured by
678 * holing the rcu_registry_lock from rcu_register_thread() or by running
679 * at library load time, which should not be executed by multiple
680 * threads nor concurrently with rcu_register_thread() anyway.
684 struct sigaction act
;
691 act
.sa_sigaction
= sigrcu_handler
;
692 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
693 sigemptyset(&act
.sa_mask
);
694 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));
711 #endif /* #ifdef RCU_SIGNAL */
713 DEFINE_RCU_FLAVOR(rcu_flavor
);
715 #include "urcu-call-rcu-impl.h"
716 #include "urcu-defer-impl.h"