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.
28 #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),
77 /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
78 /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
79 MEMBARRIER_CMD_PRIVATE_EXPEDITED
= (1 << 3),
80 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
= (1 << 4),
85 static int has_sys_membarrier_private_expedited
;
87 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
88 int rcu_has_sys_membarrier_memb
;
91 void __attribute__((constructor
)) rcu_init(void);
101 static int init_done
;
103 void __attribute__((constructor
)) rcu_init(void);
104 void __attribute__((destructor
)) rcu_exit(void);
108 * rcu_gp_lock ensures mutual exclusion between threads calling
111 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
113 * rcu_registry_lock ensures mutual exclusion between threads
114 * registering and unregistering themselves to/from the registry, and
115 * with threads reading that registry from synchronize_rcu(). However,
116 * this lock is not held all the way through the completion of awaiting
117 * for the grace period. It is sporadically released between iterations
119 * rcu_registry_lock may nest inside rcu_gp_lock.
121 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
122 struct rcu_gp rcu_gp
= { .ctr
= RCU_GP_COUNT
};
125 * Written to only by each individual reader. Read by both the reader and the
128 DEFINE_URCU_TLS(struct rcu_reader
, rcu_reader
);
130 static CDS_LIST_HEAD(registry
);
133 * Queue keeping threads awaiting to wait for a grace period. Contains
134 * struct gp_waiters_thread objects.
136 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
138 static void mutex_lock(pthread_mutex_t
*mutex
)
142 #ifndef DISTRUST_SIGNALS_EXTREME
143 ret
= pthread_mutex_lock(mutex
);
146 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
147 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
148 if (ret
!= EBUSY
&& ret
!= EINTR
)
150 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
152 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
155 (void) poll(NULL
, 0, 10);
157 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
160 static void mutex_unlock(pthread_mutex_t
*mutex
)
164 ret
= pthread_mutex_unlock(mutex
);
169 #ifdef RCU_MEMBARRIER
170 static void smp_mb_master(void)
172 if (caa_likely(rcu_has_sys_membarrier_memb
)) {
173 if (membarrier(has_sys_membarrier_private_expedited
?
174 MEMBARRIER_CMD_PRIVATE_EXPEDITED
:
175 MEMBARRIER_CMD_SHARED
, 0))
184 static void smp_mb_master(void)
191 static void force_mb_all_readers(void)
193 struct rcu_reader
*index
;
196 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
197 * compiler barriers around rcu read lock as real memory barriers.
199 if (cds_list_empty(®istry
))
202 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
203 * a cache flush on architectures with non-coherent cache. Let's play
204 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
205 * cache flush is enforced.
207 cds_list_for_each_entry(index
, ®istry
, node
) {
208 CMM_STORE_SHARED(index
->need_mb
, 1);
209 pthread_kill(index
->tid
, SIGRCU
);
212 * Wait for sighandler (and thus mb()) to execute on every thread.
214 * Note that the pthread_kill() will never be executed on systems
215 * that correctly deliver signals in a timely manner. However, it
216 * is not uncommon for kernels to have bugs that can result in
217 * lost or unduly delayed signals.
219 * If you are seeing the below pthread_kill() executing much at
220 * all, we suggest testing the underlying kernel and filing the
221 * relevant bug report. For Linux kernels, we recommend getting
222 * the Linux Test Project (LTP).
224 cds_list_for_each_entry(index
, ®istry
, node
) {
225 while (CMM_LOAD_SHARED(index
->need_mb
)) {
226 pthread_kill(index
->tid
, SIGRCU
);
227 (void) poll(NULL
, 0, 1);
230 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
233 static void smp_mb_master(void)
235 force_mb_all_readers();
237 #endif /* #ifdef RCU_SIGNAL */
240 * synchronize_rcu() waiting. Single thread.
241 * Always called with rcu_registry lock held. Releases this lock and
242 * grabs it again. Holds the lock when it returns.
244 static void wait_gp(void)
247 * Read reader_gp before read futex. smp_mb_master() needs to
248 * be called with the rcu registry lock held in RCU_SIGNAL
252 /* Temporarily unlock the registry lock. */
253 mutex_unlock(&rcu_registry_lock
);
254 if (uatomic_read(&rcu_gp
.futex
) != -1)
256 while (futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1,
260 /* Value already changed. */
263 /* Retry if interrupted by signal. */
264 break; /* Get out of switch. */
266 /* Unexpected error. */
272 * Re-lock the registry lock before the next loop.
274 mutex_lock(&rcu_registry_lock
);
278 * Always called with rcu_registry lock held. Releases this lock between
279 * iterations and grabs it again. Holds the lock when it returns.
281 static void wait_for_readers(struct cds_list_head
*input_readers
,
282 struct cds_list_head
*cur_snap_readers
,
283 struct cds_list_head
*qsreaders
)
285 unsigned int wait_loops
= 0;
286 struct rcu_reader
*index
, *tmp
;
287 #ifdef HAS_INCOHERENT_CACHES
288 unsigned int wait_gp_loops
= 0;
289 #endif /* HAS_INCOHERENT_CACHES */
292 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
293 * indicate quiescence (not nested), or observe the current
297 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
299 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
300 uatomic_dec(&rcu_gp
.futex
);
301 /* Write futex before read reader_gp */
305 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
306 switch (rcu_reader_state(&index
->ctr
)) {
307 case RCU_READER_ACTIVE_CURRENT
:
308 if (cur_snap_readers
) {
309 cds_list_move(&index
->node
,
314 case RCU_READER_INACTIVE
:
315 cds_list_move(&index
->node
, qsreaders
);
317 case RCU_READER_ACTIVE_OLD
:
319 * Old snapshot. Leaving node in
320 * input_readers will make us busy-loop
321 * until the snapshot becomes current or
322 * the reader becomes inactive.
328 #ifndef HAS_INCOHERENT_CACHES
329 if (cds_list_empty(input_readers
)) {
330 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
331 /* Read reader_gp before write futex */
333 uatomic_set(&rcu_gp
.futex
, 0);
337 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
338 /* wait_gp unlocks/locks registry lock. */
341 /* Temporarily unlock the registry lock. */
342 mutex_unlock(&rcu_registry_lock
);
345 * Re-lock the registry lock before the
348 mutex_lock(&rcu_registry_lock
);
351 #else /* #ifndef HAS_INCOHERENT_CACHES */
353 * BUSY-LOOP. Force the reader thread to commit its
354 * URCU_TLS(rcu_reader).ctr update to memory if we wait
357 if (cds_list_empty(input_readers
)) {
358 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
359 /* Read reader_gp before write futex */
361 uatomic_set(&rcu_gp
.futex
, 0);
365 if (wait_gp_loops
== KICK_READER_LOOPS
) {
369 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
370 /* wait_gp unlocks/locks registry lock. */
374 /* Temporarily unlock the registry lock. */
375 mutex_unlock(&rcu_registry_lock
);
378 * Re-lock the registry lock before the
381 mutex_lock(&rcu_registry_lock
);
384 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
388 void synchronize_rcu(void)
390 CDS_LIST_HEAD(cur_snap_readers
);
391 CDS_LIST_HEAD(qsreaders
);
392 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
393 struct urcu_waiters waiters
;
396 * Add ourself to gp_waiters queue of threads awaiting to wait
397 * for a grace period. Proceed to perform the grace period only
398 * if we are the first thread added into the queue.
399 * The implicit memory barrier before urcu_wait_add()
400 * orders prior memory accesses of threads put into the wait
401 * queue before their insertion into the wait queue.
403 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
404 /* Not first in queue: will be awakened by another thread. */
405 urcu_adaptative_busy_wait(&wait
);
406 /* Order following memory accesses after grace period. */
410 /* We won't need to wake ourself up */
411 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
413 mutex_lock(&rcu_gp_lock
);
416 * Move all waiters into our local queue.
418 urcu_move_waiters(&waiters
, &gp_waiters
);
420 mutex_lock(&rcu_registry_lock
);
422 if (cds_list_empty(®istry
))
426 * All threads should read qparity before accessing data structure
427 * where new ptr points to. Must be done within rcu_registry_lock
428 * because it iterates on reader threads.
430 /* Write new ptr before changing the qparity */
434 * Wait for readers to observe original parity or be quiescent.
435 * wait_for_readers() can release and grab again rcu_registry_lock
438 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
441 * Must finish waiting for quiescent state for original parity before
442 * committing next rcu_gp.ctr update to memory. Failure to do so could
443 * result in the writer waiting forever while new readers are always
444 * accessing data (no progress). Enforce compiler-order of load
445 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
450 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
451 * model easier to understand. It does not have a big performance impact
452 * anyway, given this is the write-side.
456 /* Switch parity: 0 -> 1, 1 -> 0 */
457 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ RCU_GP_CTR_PHASE
);
460 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
461 * state. Failure to do so could result in the writer waiting forever
462 * while new readers are always accessing data (no progress). Enforce
463 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
469 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
470 * model easier to understand. It does not have a big performance impact
471 * anyway, given this is the write-side.
476 * Wait for readers to observe new parity or be quiescent.
477 * wait_for_readers() can release and grab again rcu_registry_lock
480 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
483 * Put quiescent reader list back into registry.
485 cds_list_splice(&qsreaders
, ®istry
);
488 * Finish waiting for reader threads before letting the old ptr
489 * being freed. Must be done within rcu_registry_lock because it
490 * iterates on reader threads.
494 mutex_unlock(&rcu_registry_lock
);
495 mutex_unlock(&rcu_gp_lock
);
498 * Wakeup waiters only after we have completed the grace period
499 * and have ensured the memory barriers at the end of the grace
500 * period have been issued.
502 urcu_wake_all_waiters(&waiters
);
506 * library wrappers to be used by non-LGPL compatible source code.
509 void rcu_read_lock(void)
514 void rcu_read_unlock(void)
519 int rcu_read_ongoing(void)
521 return _rcu_read_ongoing();
524 void rcu_register_thread(void)
526 URCU_TLS(rcu_reader
).tid
= pthread_self();
527 assert(URCU_TLS(rcu_reader
).need_mb
== 0);
528 assert(!(URCU_TLS(rcu_reader
).ctr
& RCU_GP_CTR_NEST_MASK
));
530 mutex_lock(&rcu_registry_lock
);
531 assert(!URCU_TLS(rcu_reader
).registered
);
532 URCU_TLS(rcu_reader
).registered
= 1;
533 rcu_init(); /* In case gcc does not support constructor attribute */
534 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
535 mutex_unlock(&rcu_registry_lock
);
538 void rcu_unregister_thread(void)
540 mutex_lock(&rcu_registry_lock
);
541 assert(URCU_TLS(rcu_reader
).registered
);
542 URCU_TLS(rcu_reader
).registered
= 0;
543 cds_list_del(&URCU_TLS(rcu_reader
).node
);
544 mutex_unlock(&rcu_registry_lock
);
547 #ifdef RCU_MEMBARRIER
549 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
551 void rcu_sys_membarrier_status(bool available
)
558 void rcu_sys_membarrier_status(bool available
)
562 rcu_has_sys_membarrier_memb
= 1;
567 void rcu_sys_membarrier_init(void)
569 bool available
= false;
572 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
574 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
575 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
577 has_sys_membarrier_private_expedited
= 1;
579 } else if (mask
& MEMBARRIER_CMD_SHARED
) {
583 rcu_sys_membarrier_status(available
);
591 rcu_sys_membarrier_init();
596 static void sigrcu_handler(int signo
, siginfo_t
*siginfo
, void *context
)
599 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
600 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
604 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
609 * rcu_init constructor. Called when the library is linked, but also when
610 * reader threads are calling rcu_register_thread().
611 * Should only be called by a single thread at a given time. This is ensured by
612 * holing the rcu_registry_lock from rcu_register_thread() or by running
613 * at library load time, which should not be executed by multiple
614 * threads nor concurrently with rcu_register_thread() anyway.
618 struct sigaction act
;
625 act
.sa_sigaction
= sigrcu_handler
;
626 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
627 sigemptyset(&act
.sa_mask
);
628 ret
= sigaction(SIGRCU
, &act
, NULL
);
636 * Don't unregister the SIGRCU signal handler anymore, because
637 * call_rcu threads could still be using it shortly before the
639 * Assertion disabled because call_rcu threads are now rcu
640 * readers, and left running at exit.
641 * assert(cds_list_empty(®istry));
645 #endif /* #ifdef RCU_SIGNAL */
647 DEFINE_RCU_FLAVOR(rcu_flavor
);
649 #include "urcu-call-rcu-impl.h"
650 #include "urcu-defer-impl.h"