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
41 #include <urcu/config.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;
97 URCU_ATTR_ALIAS("urcu_memb_has_sys_membarrier")
98 extern int rcu_has_sys_membarrier_memb
;
101 void __attribute__((constructor
)) rcu_init(void);
108 URCU_ATTR_ALIAS(urcu_stringify(rcu_init
))
109 void alias_rcu_init(void);
113 static int init_done
;
115 void __attribute__((constructor
)) rcu_init(void);
116 void __attribute__((destructor
)) 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
};
135 URCU_ATTR_ALIAS(urcu_stringify(rcu_gp
))
136 extern struct urcu_gp alias_rcu_gp
;
139 * Written to only by each individual reader. Read by both the reader and the
142 DEFINE_URCU_TLS(struct urcu_reader
, rcu_reader
);
143 DEFINE_URCU_TLS_ALIAS(struct urcu_reader
, rcu_reader
, alias_rcu_reader
);
145 static CDS_LIST_HEAD(registry
);
148 * Queue keeping threads awaiting to wait for a grace period. Contains
149 * struct gp_waiters_thread objects.
151 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
153 static void mutex_lock(pthread_mutex_t
*mutex
)
157 #ifndef DISTRUST_SIGNALS_EXTREME
158 ret
= pthread_mutex_lock(mutex
);
161 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
162 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
163 if (ret
!= EBUSY
&& ret
!= EINTR
)
165 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
167 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
170 (void) poll(NULL
, 0, 10);
172 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
175 static void mutex_unlock(pthread_mutex_t
*mutex
)
179 ret
= pthread_mutex_unlock(mutex
);
184 #ifdef RCU_MEMBARRIER
185 static void smp_mb_master(void)
187 if (caa_likely(urcu_memb_has_sys_membarrier
)) {
188 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited
?
189 MEMBARRIER_CMD_PRIVATE_EXPEDITED
:
190 MEMBARRIER_CMD_SHARED
, 0))
199 static void smp_mb_master(void)
206 static void force_mb_all_readers(void)
208 struct urcu_reader
*index
;
211 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
212 * compiler barriers around rcu read lock as real memory barriers.
214 if (cds_list_empty(®istry
))
217 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
218 * a cache flush on architectures with non-coherent cache. Let's play
219 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
220 * cache flush is enforced.
222 cds_list_for_each_entry(index
, ®istry
, node
) {
223 CMM_STORE_SHARED(index
->need_mb
, 1);
224 pthread_kill(index
->tid
, SIGRCU
);
227 * Wait for sighandler (and thus mb()) to execute on every thread.
229 * Note that the pthread_kill() will never be executed on systems
230 * that correctly deliver signals in a timely manner. However, it
231 * is not uncommon for kernels to have bugs that can result in
232 * lost or unduly delayed signals.
234 * If you are seeing the below pthread_kill() executing much at
235 * all, we suggest testing the underlying kernel and filing the
236 * relevant bug report. For Linux kernels, we recommend getting
237 * the Linux Test Project (LTP).
239 cds_list_for_each_entry(index
, ®istry
, node
) {
240 while (CMM_LOAD_SHARED(index
->need_mb
)) {
241 pthread_kill(index
->tid
, SIGRCU
);
242 (void) poll(NULL
, 0, 1);
245 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
248 static void smp_mb_master(void)
250 force_mb_all_readers();
252 #endif /* #ifdef RCU_SIGNAL */
255 * synchronize_rcu() waiting. Single thread.
256 * Always called with rcu_registry lock held. Releases this lock and
257 * grabs it again. Holds the lock when it returns.
259 static void wait_gp(void)
262 * Read reader_gp before read futex. smp_mb_master() needs to
263 * be called with the rcu registry lock held in RCU_SIGNAL
267 /* Temporarily unlock the registry lock. */
268 mutex_unlock(&rcu_registry_lock
);
269 while (uatomic_read(&rcu_gp
.futex
) == -1) {
270 if (!futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
272 * Prior queued wakeups queued by unrelated code
273 * using the same address can cause futex wait to
274 * return 0 even through the futex value is still
275 * -1 (spurious wakeups). Check the value again
276 * in user-space to validate whether it really
283 /* Value already changed. */
286 /* Retry if interrupted by signal. */
287 break; /* Get out of switch. Check again. */
289 /* Unexpected error. */
295 * Re-lock the registry lock before the next loop.
297 mutex_lock(&rcu_registry_lock
);
301 * Always called with rcu_registry lock held. Releases this lock between
302 * iterations and grabs it again. Holds the lock when it returns.
304 static void wait_for_readers(struct cds_list_head
*input_readers
,
305 struct cds_list_head
*cur_snap_readers
,
306 struct cds_list_head
*qsreaders
)
308 unsigned int wait_loops
= 0;
309 struct urcu_reader
*index
, *tmp
;
310 #ifdef HAS_INCOHERENT_CACHES
311 unsigned int wait_gp_loops
= 0;
312 #endif /* HAS_INCOHERENT_CACHES */
315 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
316 * indicate quiescence (not nested), or observe the current
320 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
322 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
323 uatomic_dec(&rcu_gp
.futex
);
324 /* Write futex before read reader_gp */
328 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
329 switch (urcu_common_reader_state(&rcu_gp
, &index
->ctr
)) {
330 case URCU_READER_ACTIVE_CURRENT
:
331 if (cur_snap_readers
) {
332 cds_list_move(&index
->node
,
337 case URCU_READER_INACTIVE
:
338 cds_list_move(&index
->node
, qsreaders
);
340 case URCU_READER_ACTIVE_OLD
:
342 * Old snapshot. Leaving node in
343 * input_readers will make us busy-loop
344 * until the snapshot becomes current or
345 * the reader becomes inactive.
351 #ifndef HAS_INCOHERENT_CACHES
352 if (cds_list_empty(input_readers
)) {
353 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
354 /* Read reader_gp before write futex */
356 uatomic_set(&rcu_gp
.futex
, 0);
360 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
361 /* wait_gp unlocks/locks registry lock. */
364 /* Temporarily unlock the registry lock. */
365 mutex_unlock(&rcu_registry_lock
);
368 * Re-lock the registry lock before the
371 mutex_lock(&rcu_registry_lock
);
374 #else /* #ifndef HAS_INCOHERENT_CACHES */
376 * BUSY-LOOP. Force the reader thread to commit its
377 * URCU_TLS(rcu_reader).ctr update to memory if we wait
380 if (cds_list_empty(input_readers
)) {
381 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
382 /* Read reader_gp before write futex */
384 uatomic_set(&rcu_gp
.futex
, 0);
388 if (wait_gp_loops
== KICK_READER_LOOPS
) {
392 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
393 /* wait_gp unlocks/locks registry lock. */
397 /* Temporarily unlock the registry lock. */
398 mutex_unlock(&rcu_registry_lock
);
401 * Re-lock the registry lock before the
404 mutex_lock(&rcu_registry_lock
);
407 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
411 void synchronize_rcu(void)
413 CDS_LIST_HEAD(cur_snap_readers
);
414 CDS_LIST_HEAD(qsreaders
);
415 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
416 struct urcu_waiters waiters
;
419 * Add ourself to gp_waiters queue of threads awaiting to wait
420 * for a grace period. Proceed to perform the grace period only
421 * if we are the first thread added into the queue.
422 * The implicit memory barrier before urcu_wait_add()
423 * orders prior memory accesses of threads put into the wait
424 * queue before their insertion into the wait queue.
426 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
427 /* Not first in queue: will be awakened by another thread. */
428 urcu_adaptative_busy_wait(&wait
);
429 /* Order following memory accesses after grace period. */
433 /* We won't need to wake ourself up */
434 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
436 mutex_lock(&rcu_gp_lock
);
439 * Move all waiters into our local queue.
441 urcu_move_waiters(&waiters
, &gp_waiters
);
443 mutex_lock(&rcu_registry_lock
);
445 if (cds_list_empty(®istry
))
449 * All threads should read qparity before accessing data structure
450 * where new ptr points to. Must be done within rcu_registry_lock
451 * because it iterates on reader threads.
453 /* Write new ptr before changing the qparity */
457 * Wait for readers to observe original parity or be quiescent.
458 * wait_for_readers() can release and grab again rcu_registry_lock
461 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
464 * Must finish waiting for quiescent state for original parity before
465 * committing next rcu_gp.ctr update to memory. Failure to do so could
466 * result in the writer waiting forever while new readers are always
467 * accessing data (no progress). Enforce compiler-order of load
468 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
473 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
474 * model easier to understand. It does not have a big performance impact
475 * anyway, given this is the write-side.
479 /* Switch parity: 0 -> 1, 1 -> 0 */
480 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ URCU_GP_CTR_PHASE
);
483 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
484 * state. Failure to do so could result in the writer waiting forever
485 * while new readers are always accessing data (no progress). Enforce
486 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
492 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
493 * model easier to understand. It does not have a big performance impact
494 * anyway, given this is the write-side.
499 * Wait for readers to observe new parity or be quiescent.
500 * wait_for_readers() can release and grab again rcu_registry_lock
503 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
506 * Put quiescent reader list back into registry.
508 cds_list_splice(&qsreaders
, ®istry
);
511 * Finish waiting for reader threads before letting the old ptr
512 * being freed. Must be done within rcu_registry_lock because it
513 * iterates on reader threads.
517 mutex_unlock(&rcu_registry_lock
);
518 mutex_unlock(&rcu_gp_lock
);
521 * Wakeup waiters only after we have completed the grace period
522 * and have ensured the memory barriers at the end of the grace
523 * period have been issued.
525 urcu_wake_all_waiters(&waiters
);
527 URCU_ATTR_ALIAS(urcu_stringify(synchronize_rcu
))
528 void alias_synchronize_rcu();
531 * library wrappers to be used by non-LGPL compatible source code.
534 void rcu_read_lock(void)
538 URCU_ATTR_ALIAS(urcu_stringify(rcu_read_lock
))
539 void alias_rcu_read_lock();
541 void rcu_read_unlock(void)
545 URCU_ATTR_ALIAS(urcu_stringify(rcu_read_unlock
))
546 void alias_rcu_read_unlock();
548 int rcu_read_ongoing(void)
550 return _rcu_read_ongoing();
552 URCU_ATTR_ALIAS(urcu_stringify(rcu_read_ongoing
))
553 void alias_rcu_read_ongoing();
555 void rcu_register_thread(void)
557 URCU_TLS(rcu_reader
).tid
= pthread_self();
558 assert(URCU_TLS(rcu_reader
).need_mb
== 0);
559 assert(!(URCU_TLS(rcu_reader
).ctr
& URCU_GP_CTR_NEST_MASK
));
561 mutex_lock(&rcu_registry_lock
);
562 assert(!URCU_TLS(rcu_reader
).registered
);
563 URCU_TLS(rcu_reader
).registered
= 1;
564 rcu_init(); /* In case gcc does not support constructor attribute */
565 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
566 mutex_unlock(&rcu_registry_lock
);
568 URCU_ATTR_ALIAS(urcu_stringify(rcu_register_thread
))
569 void alias_rcu_register_thread();
571 void rcu_unregister_thread(void)
573 mutex_lock(&rcu_registry_lock
);
574 assert(URCU_TLS(rcu_reader
).registered
);
575 URCU_TLS(rcu_reader
).registered
= 0;
576 cds_list_del(&URCU_TLS(rcu_reader
).node
);
577 mutex_unlock(&rcu_registry_lock
);
579 URCU_ATTR_ALIAS(urcu_stringify(rcu_unregister_thread
))
580 void alias_rcu_unregister_thread();
582 #ifdef RCU_MEMBARRIER
584 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
586 void rcu_sys_membarrier_status(bool available
)
593 void rcu_sys_membarrier_status(bool available
)
597 urcu_memb_has_sys_membarrier
= 1;
602 void rcu_sys_membarrier_init(void)
604 bool available
= false;
607 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
609 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
610 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
612 urcu_memb_has_sys_membarrier_private_expedited
= 1;
614 } else if (mask
& MEMBARRIER_CMD_SHARED
) {
618 rcu_sys_membarrier_status(available
);
626 rcu_sys_membarrier_init();
628 URCU_ATTR_ALIAS(urcu_stringify(rcu_init
))
629 void alias_rcu_init(void);
633 static void sigrcu_handler(int signo
__attribute__((unused
)),
634 siginfo_t
*siginfo
__attribute__((unused
)),
635 void *context
__attribute__((unused
)))
638 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
639 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
643 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
648 * rcu_init constructor. Called when the library is linked, but also when
649 * reader threads are calling rcu_register_thread().
650 * Should only be called by a single thread at a given time. This is ensured by
651 * holing the rcu_registry_lock from rcu_register_thread() or by running
652 * at library load time, which should not be executed by multiple
653 * threads nor concurrently with rcu_register_thread() anyway.
657 struct sigaction act
;
664 act
.sa_sigaction
= sigrcu_handler
;
665 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
666 sigemptyset(&act
.sa_mask
);
667 ret
= sigaction(SIGRCU
, &act
, NULL
);
671 URCU_ATTR_ALIAS(urcu_stringify(rcu_init
))
672 void alias_rcu_init(void);
677 * Don't unregister the SIGRCU signal handler anymore, because
678 * call_rcu threads could still be using it shortly before the
680 * Assertion disabled because call_rcu threads are now rcu
681 * readers, and left running at exit.
682 * assert(cds_list_empty(®istry));
685 URCU_ATTR_ALIAS(urcu_stringify(rcu_exit
))
686 void alias_rcu_exit(void);
688 #endif /* #ifdef RCU_SIGNAL */
690 DEFINE_RCU_FLAVOR(rcu_flavor
);
691 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor
, alias_rcu_flavor
);
693 #include "urcu-call-rcu-impl.h"
694 #include "urcu-defer-impl.h"