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;
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);
116 * rcu_gp_lock ensures mutual exclusion between threads calling
119 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
121 * rcu_registry_lock ensures mutual exclusion between threads
122 * registering and unregistering themselves to/from the registry, and
123 * with threads reading that registry from synchronize_rcu(). However,
124 * this lock is not held all the way through the completion of awaiting
125 * for the grace period. It is sporadically released between iterations
127 * rcu_registry_lock may nest inside rcu_gp_lock.
129 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
130 struct urcu_gp rcu_gp
= { .ctr
= URCU_GP_COUNT
};
133 * Written to only by each individual reader. Read by both the reader and the
136 DEFINE_URCU_TLS(struct urcu_reader
, rcu_reader
);
138 static CDS_LIST_HEAD(registry
);
141 * Queue keeping threads awaiting to wait for a grace period. Contains
142 * struct gp_waiters_thread objects.
144 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
146 static void mutex_lock(pthread_mutex_t
*mutex
)
150 #ifndef DISTRUST_SIGNALS_EXTREME
151 ret
= pthread_mutex_lock(mutex
);
154 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
155 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
156 if (ret
!= EBUSY
&& ret
!= EINTR
)
158 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
160 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
163 (void) poll(NULL
, 0, 10);
165 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
168 static void mutex_unlock(pthread_mutex_t
*mutex
)
172 ret
= pthread_mutex_unlock(mutex
);
177 #ifdef RCU_MEMBARRIER
178 static void smp_mb_master(void)
180 if (caa_likely(urcu_memb_has_sys_membarrier
)) {
181 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited
?
182 MEMBARRIER_CMD_PRIVATE_EXPEDITED
:
183 MEMBARRIER_CMD_SHARED
, 0))
192 static void smp_mb_master(void)
199 static void force_mb_all_readers(void)
201 struct urcu_reader
*index
;
204 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
205 * compiler barriers around rcu read lock as real memory barriers.
207 if (cds_list_empty(®istry
))
210 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
211 * a cache flush on architectures with non-coherent cache. Let's play
212 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
213 * cache flush is enforced.
215 cds_list_for_each_entry(index
, ®istry
, node
) {
216 CMM_STORE_SHARED(index
->need_mb
, 1);
217 pthread_kill(index
->tid
, SIGRCU
);
220 * Wait for sighandler (and thus mb()) to execute on every thread.
222 * Note that the pthread_kill() will never be executed on systems
223 * that correctly deliver signals in a timely manner. However, it
224 * is not uncommon for kernels to have bugs that can result in
225 * lost or unduly delayed signals.
227 * If you are seeing the below pthread_kill() executing much at
228 * all, we suggest testing the underlying kernel and filing the
229 * relevant bug report. For Linux kernels, we recommend getting
230 * the Linux Test Project (LTP).
232 cds_list_for_each_entry(index
, ®istry
, node
) {
233 while (CMM_LOAD_SHARED(index
->need_mb
)) {
234 pthread_kill(index
->tid
, SIGRCU
);
235 (void) poll(NULL
, 0, 1);
238 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
241 static void smp_mb_master(void)
243 force_mb_all_readers();
245 #endif /* #ifdef RCU_SIGNAL */
248 * synchronize_rcu() waiting. Single thread.
249 * Always called with rcu_registry lock held. Releases this lock and
250 * grabs it again. Holds the lock when it returns.
252 static void wait_gp(void)
255 * Read reader_gp before read futex. smp_mb_master() needs to
256 * be called with the rcu registry lock held in RCU_SIGNAL
260 /* Temporarily unlock the registry lock. */
261 mutex_unlock(&rcu_registry_lock
);
262 if (uatomic_read(&rcu_gp
.futex
) != -1)
264 while (futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1,
268 /* Value already changed. */
271 /* Retry if interrupted by signal. */
272 break; /* Get out of switch. */
274 /* Unexpected error. */
280 * Re-lock the registry lock before the next loop.
282 mutex_lock(&rcu_registry_lock
);
286 * Always called with rcu_registry lock held. Releases this lock between
287 * iterations and grabs it again. Holds the lock when it returns.
289 static void wait_for_readers(struct cds_list_head
*input_readers
,
290 struct cds_list_head
*cur_snap_readers
,
291 struct cds_list_head
*qsreaders
)
293 unsigned int wait_loops
= 0;
294 struct urcu_reader
*index
, *tmp
;
295 #ifdef HAS_INCOHERENT_CACHES
296 unsigned int wait_gp_loops
= 0;
297 #endif /* HAS_INCOHERENT_CACHES */
300 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
301 * indicate quiescence (not nested), or observe the current
305 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
307 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
308 uatomic_dec(&rcu_gp
.futex
);
309 /* Write futex before read reader_gp */
313 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
314 switch (urcu_common_reader_state(&rcu_gp
, &index
->ctr
)) {
315 case URCU_READER_ACTIVE_CURRENT
:
316 if (cur_snap_readers
) {
317 cds_list_move(&index
->node
,
322 case URCU_READER_INACTIVE
:
323 cds_list_move(&index
->node
, qsreaders
);
325 case URCU_READER_ACTIVE_OLD
:
327 * Old snapshot. Leaving node in
328 * input_readers will make us busy-loop
329 * until the snapshot becomes current or
330 * the reader becomes inactive.
336 #ifndef HAS_INCOHERENT_CACHES
337 if (cds_list_empty(input_readers
)) {
338 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
339 /* Read reader_gp before write futex */
341 uatomic_set(&rcu_gp
.futex
, 0);
345 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
346 /* wait_gp unlocks/locks registry lock. */
349 /* Temporarily unlock the registry lock. */
350 mutex_unlock(&rcu_registry_lock
);
353 * Re-lock the registry lock before the
356 mutex_lock(&rcu_registry_lock
);
359 #else /* #ifndef HAS_INCOHERENT_CACHES */
361 * BUSY-LOOP. Force the reader thread to commit its
362 * URCU_TLS(rcu_reader).ctr update to memory if we wait
365 if (cds_list_empty(input_readers
)) {
366 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
367 /* Read reader_gp before write futex */
369 uatomic_set(&rcu_gp
.futex
, 0);
373 if (wait_gp_loops
== KICK_READER_LOOPS
) {
377 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
378 /* wait_gp unlocks/locks registry lock. */
382 /* Temporarily unlock the registry lock. */
383 mutex_unlock(&rcu_registry_lock
);
386 * Re-lock the registry lock before the
389 mutex_lock(&rcu_registry_lock
);
392 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
396 void synchronize_rcu(void)
398 CDS_LIST_HEAD(cur_snap_readers
);
399 CDS_LIST_HEAD(qsreaders
);
400 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
401 struct urcu_waiters waiters
;
404 * Add ourself to gp_waiters queue of threads awaiting to wait
405 * for a grace period. Proceed to perform the grace period only
406 * if we are the first thread added into the queue.
407 * The implicit memory barrier before urcu_wait_add()
408 * orders prior memory accesses of threads put into the wait
409 * queue before their insertion into the wait queue.
411 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
412 /* Not first in queue: will be awakened by another thread. */
413 urcu_adaptative_busy_wait(&wait
);
414 /* Order following memory accesses after grace period. */
418 /* We won't need to wake ourself up */
419 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
421 mutex_lock(&rcu_gp_lock
);
424 * Move all waiters into our local queue.
426 urcu_move_waiters(&waiters
, &gp_waiters
);
428 mutex_lock(&rcu_registry_lock
);
430 if (cds_list_empty(®istry
))
434 * All threads should read qparity before accessing data structure
435 * where new ptr points to. Must be done within rcu_registry_lock
436 * because it iterates on reader threads.
438 /* Write new ptr before changing the qparity */
442 * Wait for readers to observe original parity or be quiescent.
443 * wait_for_readers() can release and grab again rcu_registry_lock
446 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
449 * Must finish waiting for quiescent state for original parity before
450 * committing next rcu_gp.ctr update to memory. Failure to do so could
451 * result in the writer waiting forever while new readers are always
452 * accessing data (no progress). Enforce compiler-order of load
453 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
458 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
459 * model easier to understand. It does not have a big performance impact
460 * anyway, given this is the write-side.
464 /* Switch parity: 0 -> 1, 1 -> 0 */
465 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ URCU_GP_CTR_PHASE
);
468 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
469 * state. Failure to do so could result in the writer waiting forever
470 * while new readers are always accessing data (no progress). Enforce
471 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
477 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
478 * model easier to understand. It does not have a big performance impact
479 * anyway, given this is the write-side.
484 * Wait for readers to observe new parity or be quiescent.
485 * wait_for_readers() can release and grab again rcu_registry_lock
488 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
491 * Put quiescent reader list back into registry.
493 cds_list_splice(&qsreaders
, ®istry
);
496 * Finish waiting for reader threads before letting the old ptr
497 * being freed. Must be done within rcu_registry_lock because it
498 * iterates on reader threads.
502 mutex_unlock(&rcu_registry_lock
);
503 mutex_unlock(&rcu_gp_lock
);
506 * Wakeup waiters only after we have completed the grace period
507 * and have ensured the memory barriers at the end of the grace
508 * period have been issued.
510 urcu_wake_all_waiters(&waiters
);
514 * library wrappers to be used by non-LGPL compatible source code.
517 void rcu_read_lock(void)
522 void rcu_read_unlock(void)
527 int rcu_read_ongoing(void)
529 return _rcu_read_ongoing();
532 void rcu_register_thread(void)
534 URCU_TLS(rcu_reader
).tid
= pthread_self();
535 assert(URCU_TLS(rcu_reader
).need_mb
== 0);
536 assert(!(URCU_TLS(rcu_reader
).ctr
& URCU_GP_CTR_NEST_MASK
));
538 mutex_lock(&rcu_registry_lock
);
539 assert(!URCU_TLS(rcu_reader
).registered
);
540 URCU_TLS(rcu_reader
).registered
= 1;
541 rcu_init(); /* In case gcc does not support constructor attribute */
542 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
543 mutex_unlock(&rcu_registry_lock
);
546 void rcu_unregister_thread(void)
548 mutex_lock(&rcu_registry_lock
);
549 assert(URCU_TLS(rcu_reader
).registered
);
550 URCU_TLS(rcu_reader
).registered
= 0;
551 cds_list_del(&URCU_TLS(rcu_reader
).node
);
552 mutex_unlock(&rcu_registry_lock
);
555 #ifdef RCU_MEMBARRIER
557 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
559 void rcu_sys_membarrier_status(bool available
)
566 void rcu_sys_membarrier_status(bool available
)
570 urcu_memb_has_sys_membarrier
= 1;
575 void rcu_sys_membarrier_init(void)
577 bool available
= false;
580 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
582 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
583 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
585 urcu_memb_has_sys_membarrier_private_expedited
= 1;
587 } else if (mask
& MEMBARRIER_CMD_SHARED
) {
591 rcu_sys_membarrier_status(available
);
599 rcu_sys_membarrier_init();
604 static void sigrcu_handler(int signo
__attribute__((unused
)),
605 siginfo_t
*siginfo
__attribute__((unused
)),
606 void *context
__attribute__((unused
)))
609 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
610 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
614 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
619 * rcu_init constructor. Called when the library is linked, but also when
620 * reader threads are calling rcu_register_thread().
621 * Should only be called by a single thread at a given time. This is ensured by
622 * holing the rcu_registry_lock from rcu_register_thread() or by running
623 * at library load time, which should not be executed by multiple
624 * threads nor concurrently with rcu_register_thread() anyway.
628 struct sigaction act
;
635 act
.sa_sigaction
= sigrcu_handler
;
636 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
637 sigemptyset(&act
.sa_mask
);
638 ret
= sigaction(SIGRCU
, &act
, NULL
);
646 * Don't unregister the SIGRCU signal handler anymore, because
647 * call_rcu threads could still be using it shortly before the
649 * Assertion disabled because call_rcu threads are now rcu
650 * readers, and left running at exit.
651 * assert(cds_list_empty(®istry));
655 #endif /* #ifdef RCU_SIGNAL */
657 DEFINE_RCU_FLAVOR(rcu_flavor
);
659 #include "urcu-call-rcu-impl.h"
660 #include "urcu-defer-impl.h"