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"
49 #include "urcu-utils.h"
52 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
54 #include <urcu/urcu.h>
58 * If a reader is really non-cooperative and refuses to commit its
59 * rcu_active_readers count to memory (there is no barrier in the reader
60 * per-se), kick it after 10 loops waiting for it.
62 #define KICK_READER_LOOPS 10
65 * Active attempts to check for reader Q.S. before calling futex().
67 #define RCU_QS_ACTIVE_ATTEMPTS 100
69 /* If the headers do not support membarrier system call, fall back on RCU_MB */
70 #ifdef __NR_membarrier
71 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
73 # define membarrier(...) -ENOSYS
77 MEMBARRIER_CMD_QUERY
= 0,
78 MEMBARRIER_CMD_SHARED
= (1 << 0),
79 /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
80 /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
81 MEMBARRIER_CMD_PRIVATE_EXPEDITED
= (1 << 3),
82 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
= (1 << 4),
87 static int urcu_memb_has_sys_membarrier_private_expedited
;
89 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
91 * Explicitly initialize to zero because we can't alias a non-static
92 * uninitialized variable.
94 int urcu_memb_has_sys_membarrier
= 0;
95 __attribute__((alias("urcu_memb_has_sys_membarrier")))
96 extern int rcu_has_sys_membarrier_memb
;
99 void __attribute__((constructor
)) rcu_init(void);
106 __attribute__((alias(urcu_stringify(rcu_init
))))
107 void alias_rcu_init(void);
111 static int init_done
;
113 void __attribute__((constructor
)) rcu_init(void);
114 void __attribute__((destructor
)) rcu_exit(void);
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
};
133 __attribute__((alias(urcu_stringify(rcu_gp
))))
134 extern struct urcu_gp alias_rcu_gp
;
137 * Written to only by each individual reader. Read by both the reader and the
140 DEFINE_URCU_TLS(struct urcu_reader
, rcu_reader
);
141 __attribute__((alias(urcu_stringify(rcu_reader
))))
142 extern struct urcu_reader alias_rcu_reader
;
144 static CDS_LIST_HEAD(registry
);
147 * Queue keeping threads awaiting to wait for a grace period. Contains
148 * struct gp_waiters_thread objects.
150 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
152 static void mutex_lock(pthread_mutex_t
*mutex
)
156 #ifndef DISTRUST_SIGNALS_EXTREME
157 ret
= pthread_mutex_lock(mutex
);
160 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
161 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
162 if (ret
!= EBUSY
&& ret
!= EINTR
)
164 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
166 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
169 (void) poll(NULL
, 0, 10);
171 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
174 static void mutex_unlock(pthread_mutex_t
*mutex
)
178 ret
= pthread_mutex_unlock(mutex
);
183 #ifdef RCU_MEMBARRIER
184 static void smp_mb_master(void)
186 if (caa_likely(urcu_memb_has_sys_membarrier
)) {
187 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited
?
188 MEMBARRIER_CMD_PRIVATE_EXPEDITED
:
189 MEMBARRIER_CMD_SHARED
, 0))
198 static void smp_mb_master(void)
205 static void force_mb_all_readers(void)
207 struct urcu_reader
*index
;
210 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
211 * compiler barriers around rcu read lock as real memory barriers.
213 if (cds_list_empty(®istry
))
216 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
217 * a cache flush on architectures with non-coherent cache. Let's play
218 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
219 * cache flush is enforced.
221 cds_list_for_each_entry(index
, ®istry
, node
) {
222 CMM_STORE_SHARED(index
->need_mb
, 1);
223 pthread_kill(index
->tid
, SIGRCU
);
226 * Wait for sighandler (and thus mb()) to execute on every thread.
228 * Note that the pthread_kill() will never be executed on systems
229 * that correctly deliver signals in a timely manner. However, it
230 * is not uncommon for kernels to have bugs that can result in
231 * lost or unduly delayed signals.
233 * If you are seeing the below pthread_kill() executing much at
234 * all, we suggest testing the underlying kernel and filing the
235 * relevant bug report. For Linux kernels, we recommend getting
236 * the Linux Test Project (LTP).
238 cds_list_for_each_entry(index
, ®istry
, node
) {
239 while (CMM_LOAD_SHARED(index
->need_mb
)) {
240 pthread_kill(index
->tid
, SIGRCU
);
241 (void) poll(NULL
, 0, 1);
244 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
247 static void smp_mb_master(void)
249 force_mb_all_readers();
251 #endif /* #ifdef RCU_SIGNAL */
254 * synchronize_rcu() waiting. Single thread.
255 * Always called with rcu_registry lock held. Releases this lock and
256 * grabs it again. Holds the lock when it returns.
258 static void wait_gp(void)
261 * Read reader_gp before read futex. smp_mb_master() needs to
262 * be called with the rcu registry lock held in RCU_SIGNAL
266 /* Temporarily unlock the registry lock. */
267 mutex_unlock(&rcu_registry_lock
);
268 if (uatomic_read(&rcu_gp
.futex
) != -1)
270 while (futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1,
274 /* Value already changed. */
277 /* Retry if interrupted by signal. */
278 break; /* Get out of switch. */
280 /* Unexpected error. */
286 * Re-lock the registry lock before the next loop.
288 mutex_lock(&rcu_registry_lock
);
292 * Always called with rcu_registry lock held. Releases this lock between
293 * iterations and grabs it again. Holds the lock when it returns.
295 static void wait_for_readers(struct cds_list_head
*input_readers
,
296 struct cds_list_head
*cur_snap_readers
,
297 struct cds_list_head
*qsreaders
)
299 unsigned int wait_loops
= 0;
300 struct urcu_reader
*index
, *tmp
;
301 #ifdef HAS_INCOHERENT_CACHES
302 unsigned int wait_gp_loops
= 0;
303 #endif /* HAS_INCOHERENT_CACHES */
306 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
307 * indicate quiescence (not nested), or observe the current
311 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
313 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
314 uatomic_dec(&rcu_gp
.futex
);
315 /* Write futex before read reader_gp */
319 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
320 switch (urcu_common_reader_state(&rcu_gp
, &index
->ctr
)) {
321 case URCU_READER_ACTIVE_CURRENT
:
322 if (cur_snap_readers
) {
323 cds_list_move(&index
->node
,
328 case URCU_READER_INACTIVE
:
329 cds_list_move(&index
->node
, qsreaders
);
331 case URCU_READER_ACTIVE_OLD
:
333 * Old snapshot. Leaving node in
334 * input_readers will make us busy-loop
335 * until the snapshot becomes current or
336 * the reader becomes inactive.
342 #ifndef HAS_INCOHERENT_CACHES
343 if (cds_list_empty(input_readers
)) {
344 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
345 /* Read reader_gp before write futex */
347 uatomic_set(&rcu_gp
.futex
, 0);
351 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
352 /* wait_gp unlocks/locks registry lock. */
355 /* Temporarily unlock the registry lock. */
356 mutex_unlock(&rcu_registry_lock
);
359 * Re-lock the registry lock before the
362 mutex_lock(&rcu_registry_lock
);
365 #else /* #ifndef HAS_INCOHERENT_CACHES */
367 * BUSY-LOOP. Force the reader thread to commit its
368 * URCU_TLS(rcu_reader).ctr update to memory if we wait
371 if (cds_list_empty(input_readers
)) {
372 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
373 /* Read reader_gp before write futex */
375 uatomic_set(&rcu_gp
.futex
, 0);
379 if (wait_gp_loops
== KICK_READER_LOOPS
) {
383 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
384 /* wait_gp unlocks/locks registry lock. */
388 /* Temporarily unlock the registry lock. */
389 mutex_unlock(&rcu_registry_lock
);
392 * Re-lock the registry lock before the
395 mutex_lock(&rcu_registry_lock
);
398 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
402 void synchronize_rcu(void)
404 CDS_LIST_HEAD(cur_snap_readers
);
405 CDS_LIST_HEAD(qsreaders
);
406 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
407 struct urcu_waiters waiters
;
410 * Add ourself to gp_waiters queue of threads awaiting to wait
411 * for a grace period. Proceed to perform the grace period only
412 * if we are the first thread added into the queue.
413 * The implicit memory barrier before urcu_wait_add()
414 * orders prior memory accesses of threads put into the wait
415 * queue before their insertion into the wait queue.
417 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
418 /* Not first in queue: will be awakened by another thread. */
419 urcu_adaptative_busy_wait(&wait
);
420 /* Order following memory accesses after grace period. */
424 /* We won't need to wake ourself up */
425 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
427 mutex_lock(&rcu_gp_lock
);
430 * Move all waiters into our local queue.
432 urcu_move_waiters(&waiters
, &gp_waiters
);
434 mutex_lock(&rcu_registry_lock
);
436 if (cds_list_empty(®istry
))
440 * All threads should read qparity before accessing data structure
441 * where new ptr points to. Must be done within rcu_registry_lock
442 * because it iterates on reader threads.
444 /* Write new ptr before changing the qparity */
448 * Wait for readers to observe original parity or be quiescent.
449 * wait_for_readers() can release and grab again rcu_registry_lock
452 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
455 * Must finish waiting for quiescent state for original parity before
456 * committing next rcu_gp.ctr update to memory. Failure to do so could
457 * result in the writer waiting forever while new readers are always
458 * accessing data (no progress). Enforce compiler-order of load
459 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
464 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
465 * model easier to understand. It does not have a big performance impact
466 * anyway, given this is the write-side.
470 /* Switch parity: 0 -> 1, 1 -> 0 */
471 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ URCU_GP_CTR_PHASE
);
474 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
475 * state. Failure to do so could result in the writer waiting forever
476 * while new readers are always accessing data (no progress). Enforce
477 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
483 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
484 * model easier to understand. It does not have a big performance impact
485 * anyway, given this is the write-side.
490 * Wait for readers to observe new parity or be quiescent.
491 * wait_for_readers() can release and grab again rcu_registry_lock
494 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
497 * Put quiescent reader list back into registry.
499 cds_list_splice(&qsreaders
, ®istry
);
502 * Finish waiting for reader threads before letting the old ptr
503 * being freed. Must be done within rcu_registry_lock because it
504 * iterates on reader threads.
508 mutex_unlock(&rcu_registry_lock
);
509 mutex_unlock(&rcu_gp_lock
);
512 * Wakeup waiters only after we have completed the grace period
513 * and have ensured the memory barriers at the end of the grace
514 * period have been issued.
516 urcu_wake_all_waiters(&waiters
);
518 __attribute__((alias(urcu_stringify(synchronize_rcu
))))
519 void alias_synchronize_rcu();
522 * library wrappers to be used by non-LGPL compatible source code.
525 void rcu_read_lock(void)
529 __attribute__((alias(urcu_stringify(rcu_read_lock
))))
530 void alias_rcu_read_lock();
532 void rcu_read_unlock(void)
536 __attribute__((alias(urcu_stringify(rcu_read_unlock
))))
537 void alias_rcu_read_unlock();
539 int rcu_read_ongoing(void)
541 return _rcu_read_ongoing();
543 __attribute__((alias(urcu_stringify(rcu_read_ongoing
))))
544 void alias_rcu_read_ongoing();
546 void rcu_register_thread(void)
548 URCU_TLS(rcu_reader
).tid
= pthread_self();
549 assert(URCU_TLS(rcu_reader
).need_mb
== 0);
550 assert(!(URCU_TLS(rcu_reader
).ctr
& URCU_GP_CTR_NEST_MASK
));
552 mutex_lock(&rcu_registry_lock
);
553 assert(!URCU_TLS(rcu_reader
).registered
);
554 URCU_TLS(rcu_reader
).registered
= 1;
555 rcu_init(); /* In case gcc does not support constructor attribute */
556 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
557 mutex_unlock(&rcu_registry_lock
);
559 __attribute__((alias(urcu_stringify(rcu_register_thread
))))
560 void alias_rcu_register_thread();
562 void rcu_unregister_thread(void)
564 mutex_lock(&rcu_registry_lock
);
565 assert(URCU_TLS(rcu_reader
).registered
);
566 URCU_TLS(rcu_reader
).registered
= 0;
567 cds_list_del(&URCU_TLS(rcu_reader
).node
);
568 mutex_unlock(&rcu_registry_lock
);
570 __attribute__((alias(urcu_stringify(rcu_unregister_thread
))))
571 void alias_rcu_unregister_thread();
573 #ifdef RCU_MEMBARRIER
575 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
577 void rcu_sys_membarrier_status(bool available
)
584 void rcu_sys_membarrier_status(bool available
)
588 urcu_memb_has_sys_membarrier
= 1;
593 void rcu_sys_membarrier_init(void)
595 bool available
= false;
598 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
600 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
601 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
603 urcu_memb_has_sys_membarrier_private_expedited
= 1;
605 } else if (mask
& MEMBARRIER_CMD_SHARED
) {
609 rcu_sys_membarrier_status(available
);
617 rcu_sys_membarrier_init();
619 __attribute__((alias(urcu_stringify(rcu_init
))))
620 void alias_rcu_init(void);
624 static void sigrcu_handler(int signo
, siginfo_t
*siginfo
, void *context
)
627 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
628 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
632 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
637 * rcu_init constructor. Called when the library is linked, but also when
638 * reader threads are calling rcu_register_thread().
639 * Should only be called by a single thread at a given time. This is ensured by
640 * holing the rcu_registry_lock from rcu_register_thread() or by running
641 * at library load time, which should not be executed by multiple
642 * threads nor concurrently with rcu_register_thread() anyway.
646 struct sigaction act
;
653 act
.sa_sigaction
= sigrcu_handler
;
654 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
655 sigemptyset(&act
.sa_mask
);
656 ret
= sigaction(SIGRCU
, &act
, NULL
);
660 __attribute__((alias(urcu_stringify(rcu_init
))))
661 void alias_rcu_init(void);
666 * Don't unregister the SIGRCU signal handler anymore, because
667 * call_rcu threads could still be using it shortly before the
669 * Assertion disabled because call_rcu threads are now rcu
670 * readers, and left running at exit.
671 * assert(cds_list_empty(®istry));
674 __attribute__((alias(urcu_stringify(rcu_exit
))))
675 void alias_rcu_exit(void);
677 #endif /* #ifdef RCU_SIGNAL */
679 DEFINE_RCU_FLAVOR(rcu_flavor
);
680 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor
, alias_rcu_flavor
);
682 #include "urcu-call-rcu-impl.h"
683 #include "urcu-defer-impl.h"