4 * Userspace RCU library, "bulletproof" version.
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
40 #include <urcu/arch.h>
41 #include <urcu/wfcqueue.h>
42 #include <urcu/map/urcu-bp.h>
43 #include <urcu/static/urcu-bp.h>
44 #include <urcu/pointer.h>
45 #include <urcu/tls-compat.h>
48 #include "urcu-utils.h"
51 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
53 #include <urcu/urcu-bp.h>
57 #define MAP_ANONYMOUS MAP_ANON
62 void *mremap_wrapper(void *old_address
, size_t old_size
,
63 size_t new_size
, int flags
)
65 return mremap(old_address
, old_size
, new_size
, flags
);
69 #define MREMAP_MAYMOVE 1
70 #define MREMAP_FIXED 2
73 * mremap wrapper for non-Linux systems not allowing MAYMOVE.
74 * This is not generic.
77 void *mremap_wrapper(void *old_address
, size_t old_size
,
78 size_t new_size
, int flags
)
80 assert(!(flags
& MREMAP_MAYMOVE
));
86 /* Sleep delay in ms */
87 #define RCU_SLEEP_DELAY_MS 10
88 #define INIT_NR_THREADS 8
89 #define ARENA_INIT_ALLOC \
90 sizeof(struct registry_chunk) \
91 + INIT_NR_THREADS * sizeof(struct urcu_bp_reader)
94 * Active attempts to check for reader Q.S. before calling sleep().
96 #define RCU_QS_ACTIVE_ATTEMPTS 100
101 /* If the headers do not support membarrier system call, fall back smp_mb. */
102 #ifdef __NR_membarrier
103 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
105 # define membarrier(...) -ENOSYS
108 enum membarrier_cmd
{
109 MEMBARRIER_CMD_QUERY
= 0,
110 MEMBARRIER_CMD_SHARED
= (1 << 0),
111 /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
112 /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
113 MEMBARRIER_CMD_PRIVATE_EXPEDITED
= (1 << 3),
114 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
= (1 << 4),
118 void __attribute__((constructor
)) _urcu_bp_init(void);
120 void __attribute__((destructor
)) urcu_bp_exit(void);
122 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
123 int urcu_bp_has_sys_membarrier
;
127 * rcu_gp_lock ensures mutual exclusion between threads calling
130 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
132 * rcu_registry_lock ensures mutual exclusion between threads
133 * registering and unregistering themselves to/from the registry, and
134 * with threads reading that registry from synchronize_rcu(). However,
135 * this lock is not held all the way through the completion of awaiting
136 * for the grace period. It is sporadically released between iterations
138 * rcu_registry_lock may nest inside rcu_gp_lock.
140 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
142 static pthread_mutex_t init_lock
= PTHREAD_MUTEX_INITIALIZER
;
143 static int initialized
;
145 static pthread_key_t urcu_bp_key
;
147 struct urcu_bp_gp urcu_bp_gp
= { .ctr
= URCU_BP_GP_COUNT
};
148 URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp
;
151 * Pointer to registry elements. Written to only by each individual reader. Read
152 * by both the reader and the writers.
154 DEFINE_URCU_TLS(struct urcu_bp_reader
*, urcu_bp_reader
);
155 DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader
*, urcu_bp_reader
, rcu_reader_bp
);
157 static CDS_LIST_HEAD(registry
);
159 struct registry_chunk
{
160 size_t data_len
; /* data length */
161 size_t used
; /* amount of data used */
162 struct cds_list_head node
; /* chunk_list node */
166 struct registry_arena
{
167 struct cds_list_head chunk_list
;
170 static struct registry_arena registry_arena
= {
171 .chunk_list
= CDS_LIST_HEAD_INIT(registry_arena
.chunk_list
),
174 /* Saved fork signal mask, protected by rcu_gp_lock */
175 static sigset_t saved_fork_signal_mask
;
177 static void mutex_lock(pthread_mutex_t
*mutex
)
181 #ifndef DISTRUST_SIGNALS_EXTREME
182 ret
= pthread_mutex_lock(mutex
);
185 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
186 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
187 if (ret
!= EBUSY
&& ret
!= EINTR
)
191 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
194 static void mutex_unlock(pthread_mutex_t
*mutex
)
198 ret
= pthread_mutex_unlock(mutex
);
203 static void smp_mb_master(void)
205 if (caa_likely(urcu_bp_has_sys_membarrier
)) {
206 if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED
, 0))
214 * Always called with rcu_registry lock held. Releases this lock between
215 * iterations and grabs it again. Holds the lock when it returns.
217 static void wait_for_readers(struct cds_list_head
*input_readers
,
218 struct cds_list_head
*cur_snap_readers
,
219 struct cds_list_head
*qsreaders
)
221 unsigned int wait_loops
= 0;
222 struct urcu_bp_reader
*index
, *tmp
;
225 * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
226 * indicate quiescence (not nested), or observe the current
230 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
233 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
234 switch (urcu_bp_reader_state(&index
->ctr
)) {
235 case URCU_BP_READER_ACTIVE_CURRENT
:
236 if (cur_snap_readers
) {
237 cds_list_move(&index
->node
,
242 case URCU_BP_READER_INACTIVE
:
243 cds_list_move(&index
->node
, qsreaders
);
245 case URCU_BP_READER_ACTIVE_OLD
:
247 * Old snapshot. Leaving node in
248 * input_readers will make us busy-loop
249 * until the snapshot becomes current or
250 * the reader becomes inactive.
256 if (cds_list_empty(input_readers
)) {
259 /* Temporarily unlock the registry lock. */
260 mutex_unlock(&rcu_registry_lock
);
261 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
)
262 (void) poll(NULL
, 0, RCU_SLEEP_DELAY_MS
);
265 /* Re-lock the registry lock before the next loop. */
266 mutex_lock(&rcu_registry_lock
);
271 void urcu_bp_synchronize_rcu(void)
273 CDS_LIST_HEAD(cur_snap_readers
);
274 CDS_LIST_HEAD(qsreaders
);
275 sigset_t newmask
, oldmask
;
278 ret
= sigfillset(&newmask
);
280 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
283 mutex_lock(&rcu_gp_lock
);
285 mutex_lock(&rcu_registry_lock
);
287 if (cds_list_empty(®istry
))
290 /* All threads should read qparity before accessing data structure
291 * where new ptr points to. */
292 /* Write new ptr before changing the qparity */
296 * Wait for readers to observe original parity or be quiescent.
297 * wait_for_readers() can release and grab again rcu_registry_lock
300 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
303 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
304 * model easier to understand. It does not have a big performance impact
305 * anyway, given this is the write-side.
309 /* Switch parity: 0 -> 1, 1 -> 0 */
310 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ URCU_BP_GP_CTR_PHASE
);
313 * Must commit qparity update to memory before waiting for other parity
314 * quiescent state. Failure to do so could result in the writer waiting
315 * forever while new readers are always accessing data (no progress).
316 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
320 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
321 * model easier to understand. It does not have a big performance impact
322 * anyway, given this is the write-side.
327 * Wait for readers to observe new parity or be quiescent.
328 * wait_for_readers() can release and grab again rcu_registry_lock
331 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
334 * Put quiescent reader list back into registry.
336 cds_list_splice(&qsreaders
, ®istry
);
339 * Finish waiting for reader threads before letting the old ptr being
344 mutex_unlock(&rcu_registry_lock
);
345 mutex_unlock(&rcu_gp_lock
);
346 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
349 URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp();
352 * library wrappers to be used by non-LGPL compatible source code.
355 void urcu_bp_read_lock(void)
357 _urcu_bp_read_lock();
359 URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp();
361 void urcu_bp_read_unlock(void)
363 _urcu_bp_read_unlock();
365 URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp();
367 int urcu_bp_read_ongoing(void)
369 return _urcu_bp_read_ongoing();
371 URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp();
374 * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
375 * Else, try expanding the last chunk. If this fails, allocate a new
376 * chunk twice as big as the last chunk.
377 * Memory used by chunks _never_ moves. A chunk could theoretically be
378 * freed when all "used" slots are released, but we don't do it at this
382 void expand_arena(struct registry_arena
*arena
)
384 struct registry_chunk
*new_chunk
, *last_chunk
;
385 size_t old_chunk_len
, new_chunk_len
;
388 if (cds_list_empty(&arena
->chunk_list
)) {
389 assert(ARENA_INIT_ALLOC
>=
390 sizeof(struct registry_chunk
)
391 + sizeof(struct rcu_reader
));
392 new_chunk_len
= ARENA_INIT_ALLOC
;
393 new_chunk
= (struct registry_chunk
*) mmap(NULL
,
395 PROT_READ
| PROT_WRITE
,
396 MAP_ANONYMOUS
| MAP_PRIVATE
,
398 if (new_chunk
== MAP_FAILED
)
400 memset(new_chunk
, 0, new_chunk_len
);
401 new_chunk
->data_len
=
402 new_chunk_len
- sizeof(struct registry_chunk
);
403 cds_list_add_tail(&new_chunk
->node
, &arena
->chunk_list
);
404 return; /* We're done. */
407 /* Try expanding last chunk. */
408 last_chunk
= cds_list_entry(arena
->chunk_list
.prev
,
409 struct registry_chunk
, node
);
411 last_chunk
->data_len
+ sizeof(struct registry_chunk
);
412 new_chunk_len
= old_chunk_len
<< 1;
414 /* Don't allow memory mapping to move, just expand. */
415 new_chunk
= mremap_wrapper(last_chunk
, old_chunk_len
,
417 if (new_chunk
!= MAP_FAILED
) {
418 /* Should not have moved. */
419 assert(new_chunk
== last_chunk
);
420 memset((char *) last_chunk
+ old_chunk_len
, 0,
421 new_chunk_len
- old_chunk_len
);
422 last_chunk
->data_len
=
423 new_chunk_len
- sizeof(struct registry_chunk
);
424 return; /* We're done. */
427 /* Remap did not succeed, we need to add a new chunk. */
428 new_chunk
= (struct registry_chunk
*) mmap(NULL
,
430 PROT_READ
| PROT_WRITE
,
431 MAP_ANONYMOUS
| MAP_PRIVATE
,
433 if (new_chunk
== MAP_FAILED
)
435 memset(new_chunk
, 0, new_chunk_len
);
436 new_chunk
->data_len
=
437 new_chunk_len
- sizeof(struct registry_chunk
);
438 cds_list_add_tail(&new_chunk
->node
, &arena
->chunk_list
);
442 struct rcu_reader
*arena_alloc(struct registry_arena
*arena
)
444 struct registry_chunk
*chunk
;
445 struct rcu_reader
*rcu_reader_reg
;
446 int expand_done
= 0; /* Only allow to expand once per alloc */
447 size_t len
= sizeof(struct rcu_reader
);
450 cds_list_for_each_entry(chunk
, &arena
->chunk_list
, node
) {
451 if (chunk
->data_len
- chunk
->used
< len
)
454 for (rcu_reader_reg
= (struct rcu_reader
*) &chunk
->data
[0];
455 rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
];
457 if (!rcu_reader_reg
->alloc
) {
458 rcu_reader_reg
->alloc
= 1;
460 return rcu_reader_reg
;
474 /* Called with signals off and mutex locked */
476 void add_thread(void)
478 struct rcu_reader
*rcu_reader_reg
;
481 rcu_reader_reg
= arena_alloc(®istry_arena
);
484 ret
= pthread_setspecific(urcu_bp_key
, rcu_reader_reg
);
488 /* Add to registry */
489 rcu_reader_reg
->tid
= pthread_self();
490 assert(rcu_reader_reg
->ctr
== 0);
491 cds_list_add(&rcu_reader_reg
->node
, ®istry
);
493 * Reader threads are pointing to the reader registry. This is
494 * why its memory should never be relocated.
496 URCU_TLS(urcu_bp_reader
) = rcu_reader_reg
;
499 /* Called with mutex locked */
501 void cleanup_thread(struct registry_chunk
*chunk
,
502 struct rcu_reader
*rcu_reader_reg
)
504 rcu_reader_reg
->ctr
= 0;
505 cds_list_del(&rcu_reader_reg
->node
);
506 rcu_reader_reg
->tid
= 0;
507 rcu_reader_reg
->alloc
= 0;
508 chunk
->used
-= sizeof(struct rcu_reader
);
512 struct registry_chunk
*find_chunk(struct rcu_reader
*rcu_reader_reg
)
514 struct registry_chunk
*chunk
;
516 cds_list_for_each_entry(chunk
, ®istry_arena
.chunk_list
, node
) {
517 if (rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[0])
519 if (rcu_reader_reg
>= (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
])
526 /* Called with signals off and mutex locked */
528 void remove_thread(struct rcu_reader
*rcu_reader_reg
)
530 cleanup_thread(find_chunk(rcu_reader_reg
), rcu_reader_reg
);
531 URCU_TLS(urcu_bp_reader
) = NULL
;
534 /* Disable signals, take mutex, add to registry */
535 void urcu_bp_register(void)
537 sigset_t newmask
, oldmask
;
540 ret
= sigfillset(&newmask
);
543 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
548 * Check if a signal concurrently registered our thread since
549 * the check in rcu_read_lock().
551 if (URCU_TLS(urcu_bp_reader
))
555 * Take care of early registration before urcu_bp constructor.
559 mutex_lock(&rcu_registry_lock
);
561 mutex_unlock(&rcu_registry_lock
);
563 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
567 URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register();
569 void urcu_bp_register_thread(void)
571 if (caa_unlikely(!URCU_TLS(urcu_bp_reader
)))
572 urcu_bp_register(); /* If not yet registered. */
575 /* Disable signals, take mutex, remove from registry */
577 void urcu_bp_unregister(struct rcu_reader
*rcu_reader_reg
)
579 sigset_t newmask
, oldmask
;
582 ret
= sigfillset(&newmask
);
585 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
589 mutex_lock(&rcu_registry_lock
);
590 remove_thread(rcu_reader_reg
);
591 mutex_unlock(&rcu_registry_lock
);
592 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
599 * Remove thread from the registry when it exits, and flag it as
600 * destroyed so garbage collection can take care of it.
603 void urcu_bp_thread_exit_notifier(void *rcu_key
)
605 urcu_bp_unregister(rcu_key
);
608 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
610 void urcu_bp_sys_membarrier_status(bool available
)
617 void urcu_bp_sys_membarrier_status(bool available
)
621 urcu_bp_has_sys_membarrier
= 1;
626 void urcu_bp_sys_membarrier_init(void)
628 bool available
= false;
631 mask
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
633 if (mask
& MEMBARRIER_CMD_PRIVATE_EXPEDITED
) {
634 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED
, 0))
639 urcu_bp_sys_membarrier_status(available
);
643 void _urcu_bp_init(void)
645 mutex_lock(&init_lock
);
646 if (!urcu_bp_refcount
++) {
649 ret
= pthread_key_create(&urcu_bp_key
,
650 urcu_bp_thread_exit_notifier
);
653 urcu_bp_sys_membarrier_init();
656 mutex_unlock(&init_lock
);
660 void urcu_bp_exit(void)
662 mutex_lock(&init_lock
);
663 if (!--urcu_bp_refcount
) {
664 struct registry_chunk
*chunk
, *tmp
;
667 cds_list_for_each_entry_safe(chunk
, tmp
,
668 ®istry_arena
.chunk_list
, node
) {
669 munmap((void *) chunk
, chunk
->data_len
670 + sizeof(struct registry_chunk
));
672 CDS_INIT_LIST_HEAD(®istry_arena
.chunk_list
);
673 ret
= pthread_key_delete(urcu_bp_key
);
677 mutex_unlock(&init_lock
);
681 * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
682 * sure we fork() don't race with a concurrent thread executing with
683 * any of those locks held. This ensures that the registry and data
684 * protected by rcu_gp_lock are in a coherent state in the child.
686 void urcu_bp_before_fork(void)
688 sigset_t newmask
, oldmask
;
691 ret
= sigfillset(&newmask
);
693 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
695 mutex_lock(&rcu_gp_lock
);
696 mutex_lock(&rcu_registry_lock
);
697 saved_fork_signal_mask
= oldmask
;
699 URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork();
701 void urcu_bp_after_fork_parent(void)
706 oldmask
= saved_fork_signal_mask
;
707 mutex_unlock(&rcu_registry_lock
);
708 mutex_unlock(&rcu_gp_lock
);
709 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
712 URCU_ATTR_ALIAS("urcu_bp_after_fork_parent")
713 void rcu_bp_after_fork_parent(void);
716 * Prune all entries from registry except our own thread. Fits the Linux
717 * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
720 void urcu_bp_prune_registry(void)
722 struct registry_chunk
*chunk
;
723 struct urcu_bp_reader
*rcu_reader_reg
;
725 cds_list_for_each_entry(chunk
, ®istry_arena
.chunk_list
, node
) {
726 for (rcu_reader_reg
= (struct urcu_bp_reader
*) &chunk
->data
[0];
727 rcu_reader_reg
< (struct urcu_bp_reader
*) &chunk
->data
[chunk
->data_len
];
729 if (!rcu_reader_reg
->alloc
)
731 if (rcu_reader_reg
->tid
== pthread_self())
733 cleanup_thread(chunk
, rcu_reader_reg
);
738 void urcu_bp_after_fork_child(void)
743 urcu_bp_prune_registry();
744 oldmask
= saved_fork_signal_mask
;
745 mutex_unlock(&rcu_registry_lock
);
746 mutex_unlock(&rcu_gp_lock
);
747 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
750 URCU_ATTR_ALIAS("urcu_bp_after_fork_child")
751 void rcu_bp_after_fork_child(void);
753 void *urcu_bp_dereference_sym(void *p
)
755 return _rcu_dereference(p
);
757 URCU_ATTR_ALIAS("urcu_bp_dereference_sym")
758 void *rcu_dereference_sym_bp();
760 void *urcu_bp_set_pointer_sym(void **p
, void *v
)
766 URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym")
767 void *rcu_set_pointer_sym_bp();
769 void *urcu_bp_xchg_pointer_sym(void **p
, void *v
)
772 return uatomic_xchg(p
, v
);
774 URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym")
775 void *rcu_xchg_pointer_sym_bp();
777 void *urcu_bp_cmpxchg_pointer_sym(void **p
, void *old
, void *_new
)
780 return uatomic_cmpxchg(p
, old
, _new
);
782 URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym")
783 void *rcu_cmpxchg_pointer_sym_bp();
785 DEFINE_RCU_FLAVOR(rcu_flavor
);
786 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor
, alias_rcu_flavor
);
788 #include "urcu-call-rcu-impl.h"
789 #include "urcu-defer-impl.h"