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.
38 #include "urcu/arch.h"
39 #include "urcu/wfcqueue.h"
40 #include "urcu/map/urcu-bp.h"
41 #include "urcu/static/urcu-bp.h"
42 #include "urcu-pointer.h"
43 #include "urcu/tls-compat.h"
47 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
53 #define MAP_ANONYMOUS MAP_ANON
58 void *mremap_wrapper(void *old_address
, size_t old_size
,
59 size_t new_size
, int flags
)
61 return mremap(old_address
, old_size
, new_size
, flags
);
65 #define MREMAP_MAYMOVE 1
66 #define MREMAP_FIXED 2
69 * mremap wrapper for non-Linux systems not allowing MAYMOVE.
70 * This is not generic.
73 void *mremap_wrapper(void *old_address
, size_t old_size
,
74 size_t new_size
, int flags
)
76 assert(!(flags
& MREMAP_MAYMOVE
));
82 /* Sleep delay in ms */
83 #define RCU_SLEEP_DELAY_MS 10
84 #define INIT_NR_THREADS 8
85 #define ARENA_INIT_ALLOC \
86 sizeof(struct registry_chunk) \
87 + INIT_NR_THREADS * sizeof(struct rcu_reader)
90 * Active attempts to check for reader Q.S. before calling sleep().
92 #define RCU_QS_ACTIVE_ATTEMPTS 100
97 /* If the headers do not support membarrier system call, fall back smp_mb. */
98 #ifdef __NR_membarrier
99 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
101 # define membarrier(...) -ENOSYS
104 enum membarrier_cmd
{
105 MEMBARRIER_CMD_QUERY
= 0,
106 MEMBARRIER_CMD_SHARED
= (1 << 0),
110 void __attribute__((constructor
)) rcu_bp_init(void);
112 void __attribute__((destructor
)) rcu_bp_exit(void);
114 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
115 int urcu_bp_has_sys_membarrier
;
119 * rcu_gp_lock ensures mutual exclusion between threads calling
122 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
124 * rcu_registry_lock ensures mutual exclusion between threads
125 * registering and unregistering themselves to/from the registry, and
126 * with threads reading that registry from synchronize_rcu(). However,
127 * this lock is not held all the way through the completion of awaiting
128 * for the grace period. It is sporadically released between iterations
130 * rcu_registry_lock may nest inside rcu_gp_lock.
132 static pthread_mutex_t rcu_registry_lock
= PTHREAD_MUTEX_INITIALIZER
;
134 static pthread_mutex_t init_lock
= PTHREAD_MUTEX_INITIALIZER
;
135 static int initialized
;
137 static pthread_key_t urcu_bp_key
;
139 struct rcu_gp rcu_gp
= { .ctr
= RCU_GP_COUNT
};
142 * Pointer to registry elements. Written to only by each individual reader. Read
143 * by both the reader and the writers.
145 DEFINE_URCU_TLS(struct rcu_reader
*, rcu_reader
);
147 static CDS_LIST_HEAD(registry
);
149 struct registry_chunk
{
150 size_t data_len
; /* data length */
151 size_t used
; /* amount of data used */
152 struct cds_list_head node
; /* chunk_list node */
156 struct registry_arena
{
157 struct cds_list_head chunk_list
;
160 static struct registry_arena registry_arena
= {
161 .chunk_list
= CDS_LIST_HEAD_INIT(registry_arena
.chunk_list
),
164 /* Saved fork signal mask, protected by rcu_gp_lock */
165 static sigset_t saved_fork_signal_mask
;
167 static void mutex_lock(pthread_mutex_t
*mutex
)
171 #ifndef DISTRUST_SIGNALS_EXTREME
172 ret
= pthread_mutex_lock(mutex
);
175 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
176 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
177 if (ret
!= EBUSY
&& ret
!= EINTR
)
181 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
184 static void mutex_unlock(pthread_mutex_t
*mutex
)
188 ret
= pthread_mutex_unlock(mutex
);
193 static void smp_mb_master(void)
195 if (caa_likely(urcu_bp_has_sys_membarrier
))
196 (void) membarrier(MEMBARRIER_CMD_SHARED
, 0);
202 * Always called with rcu_registry lock held. Releases this lock between
203 * iterations and grabs it again. Holds the lock when it returns.
205 static void wait_for_readers(struct cds_list_head
*input_readers
,
206 struct cds_list_head
*cur_snap_readers
,
207 struct cds_list_head
*qsreaders
)
209 unsigned int wait_loops
= 0;
210 struct rcu_reader
*index
, *tmp
;
213 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
214 * indicate quiescence (not nested), or observe the current
218 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
221 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
222 switch (rcu_reader_state(&index
->ctr
)) {
223 case RCU_READER_ACTIVE_CURRENT
:
224 if (cur_snap_readers
) {
225 cds_list_move(&index
->node
,
230 case RCU_READER_INACTIVE
:
231 cds_list_move(&index
->node
, qsreaders
);
233 case RCU_READER_ACTIVE_OLD
:
235 * Old snapshot. Leaving node in
236 * input_readers will make us busy-loop
237 * until the snapshot becomes current or
238 * the reader becomes inactive.
244 if (cds_list_empty(input_readers
)) {
247 /* Temporarily unlock the registry lock. */
248 mutex_unlock(&rcu_registry_lock
);
249 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
)
250 (void) poll(NULL
, 0, RCU_SLEEP_DELAY_MS
);
253 /* Re-lock the registry lock before the next loop. */
254 mutex_lock(&rcu_registry_lock
);
259 void synchronize_rcu(void)
261 CDS_LIST_HEAD(cur_snap_readers
);
262 CDS_LIST_HEAD(qsreaders
);
263 sigset_t newmask
, oldmask
;
266 ret
= sigfillset(&newmask
);
268 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
271 mutex_lock(&rcu_gp_lock
);
273 mutex_lock(&rcu_registry_lock
);
275 if (cds_list_empty(®istry
))
278 /* All threads should read qparity before accessing data structure
279 * where new ptr points to. */
280 /* Write new ptr before changing the qparity */
284 * Wait for readers to observe original parity or be quiescent.
285 * wait_for_readers() can release and grab again rcu_registry_lock
288 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
291 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
292 * model easier to understand. It does not have a big performance impact
293 * anyway, given this is the write-side.
297 /* Switch parity: 0 -> 1, 1 -> 0 */
298 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ RCU_GP_CTR_PHASE
);
301 * Must commit qparity update to memory before waiting for other parity
302 * quiescent state. Failure to do so could result in the writer waiting
303 * forever while new readers are always accessing data (no progress).
304 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
308 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
309 * model easier to understand. It does not have a big performance impact
310 * anyway, given this is the write-side.
315 * Wait for readers to observe new parity or be quiescent.
316 * wait_for_readers() can release and grab again rcu_registry_lock
319 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
322 * Put quiescent reader list back into registry.
324 cds_list_splice(&qsreaders
, ®istry
);
327 * Finish waiting for reader threads before letting the old ptr being
332 mutex_unlock(&rcu_registry_lock
);
333 mutex_unlock(&rcu_gp_lock
);
334 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
339 * library wrappers to be used by non-LGPL compatible source code.
342 void rcu_read_lock(void)
347 void rcu_read_unlock(void)
352 int rcu_read_ongoing(void)
354 return _rcu_read_ongoing();
358 * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
359 * Else, try expanding the last chunk. If this fails, allocate a new
360 * chunk twice as big as the last chunk.
361 * Memory used by chunks _never_ moves. A chunk could theoretically be
362 * freed when all "used" slots are released, but we don't do it at this
366 void expand_arena(struct registry_arena
*arena
)
368 struct registry_chunk
*new_chunk
, *last_chunk
;
369 size_t old_chunk_len
, new_chunk_len
;
372 if (cds_list_empty(&arena
->chunk_list
)) {
373 assert(ARENA_INIT_ALLOC
>=
374 sizeof(struct registry_chunk
)
375 + sizeof(struct rcu_reader
));
376 new_chunk_len
= ARENA_INIT_ALLOC
;
377 new_chunk
= mmap(NULL
, new_chunk_len
,
378 PROT_READ
| PROT_WRITE
,
379 MAP_ANONYMOUS
| MAP_PRIVATE
,
381 if (new_chunk
== MAP_FAILED
)
383 memset(new_chunk
, 0, new_chunk_len
);
384 new_chunk
->data_len
=
385 new_chunk_len
- sizeof(struct registry_chunk
);
386 cds_list_add_tail(&new_chunk
->node
, &arena
->chunk_list
);
387 return; /* We're done. */
390 /* Try expanding last chunk. */
391 last_chunk
= cds_list_entry(arena
->chunk_list
.prev
,
392 struct registry_chunk
, node
);
394 last_chunk
->data_len
+ sizeof(struct registry_chunk
);
395 new_chunk_len
= old_chunk_len
<< 1;
397 /* Don't allow memory mapping to move, just expand. */
398 new_chunk
= mremap_wrapper(last_chunk
, old_chunk_len
,
400 if (new_chunk
!= MAP_FAILED
) {
401 /* Should not have moved. */
402 assert(new_chunk
== last_chunk
);
403 memset((char *) last_chunk
+ old_chunk_len
, 0,
404 new_chunk_len
- old_chunk_len
);
405 last_chunk
->data_len
=
406 new_chunk_len
- sizeof(struct registry_chunk
);
407 return; /* We're done. */
410 /* Remap did not succeed, we need to add a new chunk. */
411 new_chunk
= mmap(NULL
, new_chunk_len
,
412 PROT_READ
| PROT_WRITE
,
413 MAP_ANONYMOUS
| MAP_PRIVATE
,
415 if (new_chunk
== MAP_FAILED
)
417 memset(new_chunk
, 0, new_chunk_len
);
418 new_chunk
->data_len
=
419 new_chunk_len
- sizeof(struct registry_chunk
);
420 cds_list_add_tail(&new_chunk
->node
, &arena
->chunk_list
);
424 struct rcu_reader
*arena_alloc(struct registry_arena
*arena
)
426 struct registry_chunk
*chunk
;
427 struct rcu_reader
*rcu_reader_reg
;
428 int expand_done
= 0; /* Only allow to expand once per alloc */
429 size_t len
= sizeof(struct rcu_reader
);
432 cds_list_for_each_entry(chunk
, &arena
->chunk_list
, node
) {
433 if (chunk
->data_len
- chunk
->used
< len
)
436 for (rcu_reader_reg
= (struct rcu_reader
*) &chunk
->data
[0];
437 rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
];
439 if (!rcu_reader_reg
->alloc
) {
440 rcu_reader_reg
->alloc
= 1;
442 return rcu_reader_reg
;
456 /* Called with signals off and mutex locked */
458 void add_thread(void)
460 struct rcu_reader
*rcu_reader_reg
;
463 rcu_reader_reg
= arena_alloc(®istry_arena
);
466 ret
= pthread_setspecific(urcu_bp_key
, rcu_reader_reg
);
470 /* Add to registry */
471 rcu_reader_reg
->tid
= pthread_self();
472 assert(rcu_reader_reg
->ctr
== 0);
473 cds_list_add(&rcu_reader_reg
->node
, ®istry
);
475 * Reader threads are pointing to the reader registry. This is
476 * why its memory should never be relocated.
478 URCU_TLS(rcu_reader
) = rcu_reader_reg
;
481 /* Called with mutex locked */
483 void cleanup_thread(struct registry_chunk
*chunk
,
484 struct rcu_reader
*rcu_reader_reg
)
486 rcu_reader_reg
->ctr
= 0;
487 cds_list_del(&rcu_reader_reg
->node
);
488 rcu_reader_reg
->tid
= 0;
489 rcu_reader_reg
->alloc
= 0;
490 chunk
->used
-= sizeof(struct rcu_reader
);
494 struct registry_chunk
*find_chunk(struct rcu_reader
*rcu_reader_reg
)
496 struct registry_chunk
*chunk
;
498 cds_list_for_each_entry(chunk
, ®istry_arena
.chunk_list
, node
) {
499 if (rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[0])
501 if (rcu_reader_reg
>= (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
])
508 /* Called with signals off and mutex locked */
510 void remove_thread(struct rcu_reader
*rcu_reader_reg
)
512 cleanup_thread(find_chunk(rcu_reader_reg
), rcu_reader_reg
);
513 URCU_TLS(rcu_reader
) = NULL
;
516 /* Disable signals, take mutex, add to registry */
517 void rcu_bp_register(void)
519 sigset_t newmask
, oldmask
;
522 ret
= sigfillset(&newmask
);
525 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
530 * Check if a signal concurrently registered our thread since
531 * the check in rcu_read_lock().
533 if (URCU_TLS(rcu_reader
))
537 * Take care of early registration before urcu_bp constructor.
541 mutex_lock(&rcu_registry_lock
);
543 mutex_unlock(&rcu_registry_lock
);
545 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
550 /* Disable signals, take mutex, remove from registry */
552 void rcu_bp_unregister(struct rcu_reader
*rcu_reader_reg
)
554 sigset_t newmask
, oldmask
;
557 ret
= sigfillset(&newmask
);
560 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
564 mutex_lock(&rcu_registry_lock
);
565 remove_thread(rcu_reader_reg
);
566 mutex_unlock(&rcu_registry_lock
);
567 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
574 * Remove thread from the registry when it exits, and flag it as
575 * destroyed so garbage collection can take care of it.
578 void urcu_bp_thread_exit_notifier(void *rcu_key
)
580 rcu_bp_unregister(rcu_key
);
583 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
585 void rcu_sys_membarrier_status(int available
)
592 void rcu_sys_membarrier_status(int available
)
595 urcu_bp_has_sys_membarrier
= 1;
600 void rcu_bp_init(void)
602 mutex_lock(&init_lock
);
603 if (!rcu_bp_refcount
++) {
606 ret
= pthread_key_create(&urcu_bp_key
,
607 urcu_bp_thread_exit_notifier
);
610 ret
= membarrier(MEMBARRIER_CMD_QUERY
, 0);
611 rcu_sys_membarrier_status(ret
>= 0
612 && (ret
& MEMBARRIER_CMD_SHARED
));
615 mutex_unlock(&init_lock
);
619 void rcu_bp_exit(void)
621 mutex_lock(&init_lock
);
622 if (!--rcu_bp_refcount
) {
623 struct registry_chunk
*chunk
, *tmp
;
626 cds_list_for_each_entry_safe(chunk
, tmp
,
627 ®istry_arena
.chunk_list
, node
) {
628 munmap(chunk
, chunk
->data_len
629 + sizeof(struct registry_chunk
));
631 CDS_INIT_LIST_HEAD(®istry_arena
.chunk_list
);
632 ret
= pthread_key_delete(urcu_bp_key
);
636 mutex_unlock(&init_lock
);
640 * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
641 * sure we fork() don't race with a concurrent thread executing with
642 * any of those locks held. This ensures that the registry and data
643 * protected by rcu_gp_lock are in a coherent state in the child.
645 void rcu_bp_before_fork(void)
647 sigset_t newmask
, oldmask
;
650 ret
= sigfillset(&newmask
);
652 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
654 mutex_lock(&rcu_gp_lock
);
655 mutex_lock(&rcu_registry_lock
);
656 saved_fork_signal_mask
= oldmask
;
659 void rcu_bp_after_fork_parent(void)
664 oldmask
= saved_fork_signal_mask
;
665 mutex_unlock(&rcu_registry_lock
);
666 mutex_unlock(&rcu_gp_lock
);
667 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
672 * Prune all entries from registry except our own thread. Fits the Linux
673 * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
676 void urcu_bp_prune_registry(void)
678 struct registry_chunk
*chunk
;
679 struct rcu_reader
*rcu_reader_reg
;
681 cds_list_for_each_entry(chunk
, ®istry_arena
.chunk_list
, node
) {
682 for (rcu_reader_reg
= (struct rcu_reader
*) &chunk
->data
[0];
683 rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
];
685 if (!rcu_reader_reg
->alloc
)
687 if (rcu_reader_reg
->tid
== pthread_self())
689 cleanup_thread(chunk
, rcu_reader_reg
);
694 void rcu_bp_after_fork_child(void)
699 urcu_bp_prune_registry();
700 oldmask
= saved_fork_signal_mask
;
701 mutex_unlock(&rcu_registry_lock
);
702 mutex_unlock(&rcu_gp_lock
);
703 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
707 void *rcu_dereference_sym_bp(void *p
)
709 return _rcu_dereference(p
);
712 void *rcu_set_pointer_sym_bp(void **p
, void *v
)
719 void *rcu_xchg_pointer_sym_bp(void **p
, void *v
)
722 return uatomic_xchg(p
, v
);
725 void *rcu_cmpxchg_pointer_sym_bp(void **p
, void *old
, void *_new
)
728 return uatomic_cmpxchg(p
, old
, _new
);
731 DEFINE_RCU_FLAVOR(rcu_flavor
);
733 #include "urcu-call-rcu-impl.h"
734 #include "urcu-defer-impl.h"