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.
29 #define _DEFAULT_SOURCE
40 #include "urcu/wfcqueue.h"
41 #include "urcu/map/urcu.h"
42 #include "urcu/static/urcu.h"
43 #include "urcu-pointer.h"
44 #include "urcu/tls-compat.h"
47 #include "urcu-wait.h"
49 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
55 * If a reader is really non-cooperative and refuses to commit its
56 * rcu_active_readers count to memory (there is no barrier in the reader
57 * per-se), kick it after 10 loops waiting for it.
59 #define KICK_READER_LOOPS 10
62 * Active attempts to check for reader Q.S. before calling futex().
64 #define RCU_QS_ACTIVE_ATTEMPTS 100
67 * The ABI of sys_membarrier changed after its original implementation.
68 * Disable it for now. Use RCU_MB flavor instead.
70 #define membarrier(...) -ENOSYS
72 #define MEMBARRIER_EXPEDITED (1 << 0)
73 #define MEMBARRIER_DELAYED (1 << 1)
74 #define MEMBARRIER_QUERY (1 << 16)
78 int rcu_has_sys_membarrier
;
80 void __attribute__((constructor
)) rcu_init(void);
92 void __attribute__((constructor
)) rcu_init(void);
93 void __attribute__((destructor
)) rcu_exit(void);
97 * rcu_gp_lock ensures mutual exclusion between threads calling
100 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
102 * rcu_registry_lock ensures mutual exclusion between threads
103 * registering and unregistering themselves to/from the registry, and
104 * with threads reading that registry from synchronize_rcu(). However,
105 * this lock is not held all the way through the completion of awaiting
106 * for the grace period. It is sporadically released between iterations
108 * rcu_registry_lock may nest inside rcu_gp_lock.
110 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
111 struct rcu_gp rcu_gp
= { .ctr
= RCU_GP_COUNT
};
114 * Written to only by each individual reader. Read by both the reader and the
117 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader
, rcu_reader
);
119 static CDS_LIST_HEAD(registry
);
122 * Queue keeping threads awaiting to wait for a grace period. Contains
123 * struct gp_waiters_thread objects.
125 static DEFINE_URCU_WAIT_QUEUE(gp_waiters
);
127 static void mutex_lock(pthread_mutex_t
*mutex
)
131 #ifndef DISTRUST_SIGNALS_EXTREME
132 ret
= pthread_mutex_lock(mutex
);
135 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
136 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
137 if (ret
!= EBUSY
&& ret
!= EINTR
)
139 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
141 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
146 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
149 static void mutex_unlock(pthread_mutex_t
*mutex
)
153 ret
= pthread_mutex_unlock(mutex
);
158 #ifdef RCU_MEMBARRIER
159 static void smp_mb_master(int group
)
161 if (caa_likely(rcu_has_sys_membarrier
))
162 (void) membarrier(MEMBARRIER_EXPEDITED
);
169 static void smp_mb_master(int group
)
176 static void force_mb_all_readers(void)
178 struct rcu_reader
*index
;
181 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
182 * compiler barriers around rcu read lock as real memory barriers.
184 if (cds_list_empty(®istry
))
187 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
188 * a cache flush on architectures with non-coherent cache. Let's play
189 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
190 * cache flush is enforced.
192 cds_list_for_each_entry(index
, ®istry
, node
) {
193 CMM_STORE_SHARED(index
->need_mb
, 1);
194 pthread_kill(index
->tid
, SIGRCU
);
197 * Wait for sighandler (and thus mb()) to execute on every thread.
199 * Note that the pthread_kill() will never be executed on systems
200 * that correctly deliver signals in a timely manner. However, it
201 * is not uncommon for kernels to have bugs that can result in
202 * lost or unduly delayed signals.
204 * If you are seeing the below pthread_kill() executing much at
205 * all, we suggest testing the underlying kernel and filing the
206 * relevant bug report. For Linux kernels, we recommend getting
207 * the Linux Test Project (LTP).
209 cds_list_for_each_entry(index
, ®istry
, node
) {
210 while (CMM_LOAD_SHARED(index
->need_mb
)) {
211 pthread_kill(index
->tid
, SIGRCU
);
215 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
218 static void smp_mb_master(int group
)
220 force_mb_all_readers();
222 #endif /* #ifdef RCU_SIGNAL */
225 * synchronize_rcu() waiting. Single thread.
226 * Always called with rcu_registry lock held. Releases this lock and
227 * grabs it again. Holds the lock when it returns.
229 static void wait_gp(void)
232 * Read reader_gp before read futex. smp_mb_master() needs to
233 * be called with the rcu registry lock held in RCU_SIGNAL
236 smp_mb_master(RCU_MB_GROUP
);
237 /* Temporarily unlock the registry lock. */
238 mutex_unlock(&rcu_registry_lock
);
239 if (uatomic_read(&rcu_gp
.futex
) != -1)
241 while (futex_async(&rcu_gp
.futex
, FUTEX_WAIT
, -1,
245 /* Value already changed. */
248 /* Retry if interrupted by signal. */
249 break; /* Get out of switch. */
251 /* Unexpected error. */
257 * Re-lock the registry lock before the next loop.
259 mutex_lock(&rcu_registry_lock
);
263 * Always called with rcu_registry lock held. Releases this lock between
264 * iterations and grabs it again. Holds the lock when it returns.
266 static void wait_for_readers(struct cds_list_head
*input_readers
,
267 struct cds_list_head
*cur_snap_readers
,
268 struct cds_list_head
*qsreaders
)
270 unsigned int wait_loops
= 0;
271 struct rcu_reader
*index
, *tmp
;
272 #ifdef HAS_INCOHERENT_CACHES
273 unsigned int wait_gp_loops
= 0;
274 #endif /* HAS_INCOHERENT_CACHES */
277 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
278 * indicate quiescence (not nested), or observe the current
282 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
284 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
285 uatomic_dec(&rcu_gp
.futex
);
286 /* Write futex before read reader_gp */
287 smp_mb_master(RCU_MB_GROUP
);
290 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
291 switch (rcu_reader_state(&index
->ctr
)) {
292 case RCU_READER_ACTIVE_CURRENT
:
293 if (cur_snap_readers
) {
294 cds_list_move(&index
->node
,
299 case RCU_READER_INACTIVE
:
300 cds_list_move(&index
->node
, qsreaders
);
302 case RCU_READER_ACTIVE_OLD
:
304 * Old snapshot. Leaving node in
305 * input_readers will make us busy-loop
306 * until the snapshot becomes current or
307 * the reader becomes inactive.
313 #ifndef HAS_INCOHERENT_CACHES
314 if (cds_list_empty(input_readers
)) {
315 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
316 /* Read reader_gp before write futex */
317 smp_mb_master(RCU_MB_GROUP
);
318 uatomic_set(&rcu_gp
.futex
, 0);
322 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
323 /* wait_gp unlocks/locks registry lock. */
326 /* Temporarily unlock the registry lock. */
327 mutex_unlock(&rcu_registry_lock
);
330 * Re-lock the registry lock before the
333 mutex_lock(&rcu_registry_lock
);
336 #else /* #ifndef HAS_INCOHERENT_CACHES */
338 * BUSY-LOOP. Force the reader thread to commit its
339 * URCU_TLS(rcu_reader).ctr update to memory if we wait
342 if (cds_list_empty(input_readers
)) {
343 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
344 /* Read reader_gp before write futex */
345 smp_mb_master(RCU_MB_GROUP
);
346 uatomic_set(&rcu_gp
.futex
, 0);
350 if (wait_gp_loops
== KICK_READER_LOOPS
) {
351 smp_mb_master(RCU_MB_GROUP
);
354 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
) {
355 /* 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 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
373 void synchronize_rcu(void)
375 CDS_LIST_HEAD(cur_snap_readers
);
376 CDS_LIST_HEAD(qsreaders
);
377 DEFINE_URCU_WAIT_NODE(wait
, URCU_WAIT_WAITING
);
378 struct urcu_waiters waiters
;
381 * Add ourself to gp_waiters queue of threads awaiting to wait
382 * for a grace period. Proceed to perform the grace period only
383 * if we are the first thread added into the queue.
384 * The implicit memory barrier before urcu_wait_add()
385 * orders prior memory accesses of threads put into the wait
386 * queue before their insertion into the wait queue.
388 if (urcu_wait_add(&gp_waiters
, &wait
) != 0) {
389 /* Not first in queue: will be awakened by another thread. */
390 urcu_adaptative_busy_wait(&wait
);
391 /* Order following memory accesses after grace period. */
395 /* We won't need to wake ourself up */
396 urcu_wait_set_state(&wait
, URCU_WAIT_RUNNING
);
398 mutex_lock(&rcu_gp_lock
);
401 * Move all waiters into our local queue.
403 urcu_move_waiters(&waiters
, &gp_waiters
);
405 mutex_lock(&rcu_registry_lock
);
407 if (cds_list_empty(®istry
))
411 * All threads should read qparity before accessing data structure
412 * where new ptr points to. Must be done within rcu_registry_lock
413 * because it iterates on reader threads.
415 /* Write new ptr before changing the qparity */
416 smp_mb_master(RCU_MB_GROUP
);
419 * Wait for readers to observe original parity or be quiescent.
420 * wait_for_readers() can release and grab again rcu_registry_lock
423 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
426 * Must finish waiting for quiescent state for original parity before
427 * committing next rcu_gp.ctr update to memory. Failure to do so could
428 * result in the writer waiting forever while new readers are always
429 * accessing data (no progress). Enforce compiler-order of load
430 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
435 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
436 * model easier to understand. It does not have a big performance impact
437 * anyway, given this is the write-side.
441 /* Switch parity: 0 -> 1, 1 -> 0 */
442 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ RCU_GP_CTR_PHASE
);
445 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
446 * state. Failure to do so could result in the writer waiting forever
447 * while new readers are always accessing data (no progress). Enforce
448 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
454 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
455 * model easier to understand. It does not have a big performance impact
456 * anyway, given this is the write-side.
461 * Wait for readers to observe new parity or be quiescent.
462 * wait_for_readers() can release and grab again rcu_registry_lock
465 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
468 * Put quiescent reader list back into registry.
470 cds_list_splice(&qsreaders
, ®istry
);
473 * Finish waiting for reader threads before letting the old ptr
474 * being freed. Must be done within rcu_registry_lock because it
475 * iterates on reader threads.
477 smp_mb_master(RCU_MB_GROUP
);
479 mutex_unlock(&rcu_registry_lock
);
480 mutex_unlock(&rcu_gp_lock
);
483 * Wakeup waiters only after we have completed the grace period
484 * and have ensured the memory barriers at the end of the grace
485 * period have been issued.
487 urcu_wake_all_waiters(&waiters
);
491 * library wrappers to be used by non-LGPL compatible source code.
494 void rcu_read_lock(void)
499 void rcu_read_unlock(void)
504 int rcu_read_ongoing(void)
506 return _rcu_read_ongoing();
509 void rcu_register_thread(void)
511 URCU_TLS(rcu_reader
).tid
= pthread_self();
512 assert(URCU_TLS(rcu_reader
).need_mb
== 0);
513 assert(!(URCU_TLS(rcu_reader
).ctr
& RCU_GP_CTR_NEST_MASK
));
515 mutex_lock(&rcu_registry_lock
);
516 rcu_init(); /* In case gcc does not support constructor attribute */
517 cds_list_add(&URCU_TLS(rcu_reader
).node
, ®istry
);
518 mutex_unlock(&rcu_registry_lock
);
521 void rcu_unregister_thread(void)
523 mutex_lock(&rcu_registry_lock
);
524 cds_list_del(&URCU_TLS(rcu_reader
).node
);
525 mutex_unlock(&rcu_registry_lock
);
528 #ifdef RCU_MEMBARRIER
534 if (!membarrier(MEMBARRIER_EXPEDITED
| MEMBARRIER_QUERY
))
535 rcu_has_sys_membarrier
= 1;
540 static void sigrcu_handler(int signo
, siginfo_t
*siginfo
, void *context
)
543 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
544 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
548 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
553 * rcu_init constructor. Called when the library is linked, but also when
554 * reader threads are calling rcu_register_thread().
555 * Should only be called by a single thread at a given time. This is ensured by
556 * holing the rcu_registry_lock from rcu_register_thread() or by running
557 * at library load time, which should not be executed by multiple
558 * threads nor concurrently with rcu_register_thread() anyway.
562 struct sigaction act
;
569 act
.sa_sigaction
= sigrcu_handler
;
570 act
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
571 sigemptyset(&act
.sa_mask
);
572 ret
= sigaction(SIGRCU
, &act
, NULL
);
580 * Don't unregister the SIGRCU signal handler anymore, because
581 * call_rcu threads could still be using it shortly before the
583 * Assertion disabled because call_rcu threads are now rcu
584 * readers, and left running at exit.
585 * assert(cds_list_empty(®istry));
589 #endif /* #ifdef RCU_SIGNAL */
591 DEFINE_RCU_FLAVOR(rcu_flavor
);
593 #include "urcu-call-rcu-impl.h"
594 #include "urcu-defer-impl.h"