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.
39 #include "urcu/wfqueue.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
98 void __attribute__((constructor
)) rcu_bp_init(void);
100 void __attribute__((destructor
)) _rcu_bp_exit(void);
102 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
104 static pthread_mutex_t init_lock
= PTHREAD_MUTEX_INITIALIZER
;
105 static int initialized
;
107 static pthread_key_t urcu_bp_key
;
110 unsigned int yield_active
;
111 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield
);
115 * Global grace period counter.
116 * Contains the current RCU_GP_CTR_PHASE.
117 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
118 * Written to only by writer with mutex taken. Read by both writer and readers.
120 long rcu_gp_ctr
= RCU_GP_COUNT
;
123 * Pointer to registry elements. Written to only by each individual reader. Read
124 * by both the reader and the writers.
126 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader
*, rcu_reader
);
128 static CDS_LIST_HEAD(registry
);
130 struct registry_chunk
{
131 size_t data_len
; /* data length */
132 size_t used
; /* amount of data used */
133 struct cds_list_head node
; /* chunk_list node */
137 struct registry_arena
{
138 struct cds_list_head chunk_list
;
141 static struct registry_arena registry_arena
= {
142 .chunk_list
= CDS_LIST_HEAD_INIT(registry_arena
.chunk_list
),
145 /* Saved fork signal mask, protected by rcu_gp_lock */
146 static sigset_t saved_fork_signal_mask
;
148 static void mutex_lock(pthread_mutex_t
*mutex
)
152 #ifndef DISTRUST_SIGNALS_EXTREME
153 ret
= pthread_mutex_lock(mutex
);
156 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
157 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
158 if (ret
!= EBUSY
&& ret
!= EINTR
)
162 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
165 static void mutex_unlock(pthread_mutex_t
*mutex
)
169 ret
= pthread_mutex_unlock(mutex
);
174 void update_counter_and_wait(void)
176 CDS_LIST_HEAD(qsreaders
);
177 unsigned int wait_loops
= 0;
178 struct rcu_reader
*index
, *tmp
;
180 /* Switch parity: 0 -> 1, 1 -> 0 */
181 CMM_STORE_SHARED(rcu_gp_ctr
, rcu_gp_ctr
^ RCU_GP_CTR_PHASE
);
184 * Must commit qparity update to memory before waiting for other parity
185 * quiescent state. Failure to do so could result in the writer waiting
186 * forever while new readers are always accessing data (no progress).
187 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
191 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
192 * model easier to understand. It does not have a big performance impact
193 * anyway, given this is the write-side.
198 * Wait for each thread rcu_reader.ctr count to become 0.
201 if (wait_loops
< RCU_QS_ACTIVE_ATTEMPTS
)
204 cds_list_for_each_entry_safe(index
, tmp
, ®istry
, node
) {
205 if (!rcu_old_gp_ongoing(&index
->ctr
))
206 cds_list_move(&index
->node
, &qsreaders
);
209 if (cds_list_empty(®istry
)) {
212 if (wait_loops
>= RCU_QS_ACTIVE_ATTEMPTS
)
213 (void) poll(NULL
, 0, RCU_SLEEP_DELAY_MS
);
218 /* put back the reader list in the registry */
219 cds_list_splice(&qsreaders
, ®istry
);
222 void synchronize_rcu(void)
224 sigset_t newmask
, oldmask
;
227 ret
= sigfillset(&newmask
);
229 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
232 mutex_lock(&rcu_gp_lock
);
234 if (cds_list_empty(®istry
))
237 /* All threads should read qparity before accessing data structure
238 * where new ptr points to. */
239 /* Write new ptr before changing the qparity */
243 * Wait for previous parity to be empty of readers.
245 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
248 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
249 * model easier to understand. It does not have a big performance impact
250 * anyway, given this is the write-side.
255 * Wait for previous parity to be empty of readers.
257 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
260 * Finish waiting for reader threads before letting the old ptr being
265 mutex_unlock(&rcu_gp_lock
);
266 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
271 * library wrappers to be used by non-LGPL compatible source code.
274 void rcu_read_lock(void)
279 void rcu_read_unlock(void)
285 * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
286 * Else, try expanding the last chunk. If this fails, allocate a new
287 * chunk twice as big as the last chunk.
288 * Memory used by chunks _never_ moves. A chunk could theoretically be
289 * freed when all "used" slots are released, but we don't do it at this
293 void expand_arena(struct registry_arena
*arena
)
295 struct registry_chunk
*new_chunk
, *last_chunk
;
296 size_t old_chunk_len
, new_chunk_len
;
299 if (cds_list_empty(&arena
->chunk_list
)) {
300 assert(ARENA_INIT_ALLOC
>=
301 sizeof(struct registry_chunk
)
302 + sizeof(struct rcu_reader
));
303 new_chunk_len
= ARENA_INIT_ALLOC
;
304 new_chunk
= mmap(NULL
, new_chunk_len
,
305 PROT_READ
| PROT_WRITE
,
306 MAP_ANONYMOUS
| MAP_PRIVATE
,
308 if (new_chunk
== MAP_FAILED
)
310 bzero(new_chunk
, new_chunk_len
);
311 new_chunk
->data_len
=
312 new_chunk_len
- sizeof(struct registry_chunk
);
313 cds_list_add_tail(&new_chunk
->node
, &arena
->chunk_list
);
314 return; /* We're done. */
317 /* Try expanding last chunk. */
318 last_chunk
= cds_list_entry(arena
->chunk_list
.prev
,
319 struct registry_chunk
, node
);
321 last_chunk
->data_len
+ sizeof(struct registry_chunk
);
322 new_chunk_len
= old_chunk_len
<< 1;
324 /* Don't allow memory mapping to move, just expand. */
325 new_chunk
= mremap_wrapper(last_chunk
, old_chunk_len
,
327 if (new_chunk
!= MAP_FAILED
) {
328 /* Should not have moved. */
329 assert(new_chunk
== last_chunk
);
330 bzero((char *) last_chunk
+ old_chunk_len
,
331 new_chunk_len
- old_chunk_len
);
332 last_chunk
->data_len
=
333 new_chunk_len
- sizeof(struct registry_chunk
);
334 return; /* We're done. */
337 /* Remap did not succeed, we need to add a new chunk. */
338 new_chunk
= mmap(NULL
, new_chunk_len
,
339 PROT_READ
| PROT_WRITE
,
340 MAP_ANONYMOUS
| MAP_PRIVATE
,
342 if (new_chunk
== MAP_FAILED
)
344 bzero(new_chunk
, new_chunk_len
);
345 new_chunk
->data_len
=
346 new_chunk_len
- sizeof(struct registry_chunk
);
347 cds_list_add_tail(&new_chunk
->node
, &arena
->chunk_list
);
351 struct rcu_reader
*arena_alloc(struct registry_arena
*arena
)
353 struct registry_chunk
*chunk
;
354 struct rcu_reader
*rcu_reader_reg
;
355 int expand_done
= 0; /* Only allow to expand once per alloc */
356 size_t len
= sizeof(struct rcu_reader
);
359 cds_list_for_each_entry(chunk
, &arena
->chunk_list
, node
) {
360 if (chunk
->data_len
- chunk
->used
< len
)
363 for (rcu_reader_reg
= (struct rcu_reader
*) &chunk
->data
[0];
364 rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
];
366 if (!rcu_reader_reg
->alloc
) {
367 rcu_reader_reg
->alloc
= 1;
369 return rcu_reader_reg
;
383 /* Called with signals off and mutex locked */
385 void add_thread(void)
387 struct rcu_reader
*rcu_reader_reg
;
390 rcu_reader_reg
= arena_alloc(®istry_arena
);
393 ret
= pthread_setspecific(urcu_bp_key
, rcu_reader_reg
);
397 /* Add to registry */
398 rcu_reader_reg
->tid
= pthread_self();
399 assert(rcu_reader_reg
->ctr
== 0);
400 cds_list_add(&rcu_reader_reg
->node
, ®istry
);
402 * Reader threads are pointing to the reader registry. This is
403 * why its memory should never be relocated.
405 URCU_TLS(rcu_reader
) = rcu_reader_reg
;
408 /* Called with mutex locked */
410 void cleanup_thread(struct registry_chunk
*chunk
,
411 struct rcu_reader
*rcu_reader_reg
)
413 rcu_reader_reg
->ctr
= 0;
414 cds_list_del(&rcu_reader_reg
->node
);
415 rcu_reader_reg
->tid
= 0;
416 rcu_reader_reg
->alloc
= 0;
417 chunk
->used
-= sizeof(struct rcu_reader
);
421 struct registry_chunk
*find_chunk(struct rcu_reader
*rcu_reader_reg
)
423 struct registry_chunk
*chunk
;
425 cds_list_for_each_entry(chunk
, ®istry_arena
.chunk_list
, node
) {
426 if (rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[0])
428 if (rcu_reader_reg
>= (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
])
435 /* Called with signals off and mutex locked */
437 void remove_thread(struct rcu_reader
*rcu_reader_reg
)
439 cleanup_thread(find_chunk(rcu_reader_reg
), rcu_reader_reg
);
440 URCU_TLS(rcu_reader
) = NULL
;
443 /* Disable signals, take mutex, add to registry */
444 void rcu_bp_register(void)
446 sigset_t newmask
, oldmask
;
449 ret
= sigfillset(&newmask
);
452 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
457 * Check if a signal concurrently registered our thread since
458 * the check in rcu_read_lock().
460 if (URCU_TLS(rcu_reader
))
464 * Take care of early registration before urcu_bp constructor.
468 mutex_lock(&rcu_gp_lock
);
470 mutex_unlock(&rcu_gp_lock
);
472 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
477 /* Disable signals, take mutex, remove from registry */
479 void rcu_bp_unregister(struct rcu_reader
*rcu_reader_reg
)
481 sigset_t newmask
, oldmask
;
484 ret
= sigfillset(&newmask
);
487 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
491 mutex_lock(&rcu_gp_lock
);
492 remove_thread(rcu_reader_reg
);
493 mutex_unlock(&rcu_gp_lock
);
494 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
501 * Remove thread from the registry when it exits, and flag it as
502 * destroyed so garbage collection can take care of it.
505 void urcu_bp_thread_exit_notifier(void *rcu_key
)
507 rcu_bp_unregister(rcu_key
);
511 void rcu_bp_init(void)
513 mutex_lock(&init_lock
);
514 if (!rcu_bp_refcount
++) {
517 ret
= pthread_key_create(&urcu_bp_key
,
518 urcu_bp_thread_exit_notifier
);
523 mutex_unlock(&init_lock
);
527 void _rcu_bp_exit(void)
529 mutex_lock(&init_lock
);
530 if (!--rcu_bp_refcount
) {
531 struct registry_chunk
*chunk
, *tmp
;
534 cds_list_for_each_entry_safe(chunk
, tmp
,
535 ®istry_arena
.chunk_list
, node
) {
536 munmap(chunk
, chunk
->data_len
537 + sizeof(struct registry_chunk
));
539 ret
= pthread_key_delete(urcu_bp_key
);
543 mutex_unlock(&init_lock
);
547 * Keep ABI compability within stable versions. This has never been
548 * exposed through a header, but needs to stay in the .so until the
551 void rcu_bp_exit(void)
556 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
557 * a concurrent thread executing with this same lock held. This ensures that the
558 * registry is in a coherent state in the child.
560 void rcu_bp_before_fork(void)
562 sigset_t newmask
, oldmask
;
565 ret
= sigfillset(&newmask
);
567 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
569 mutex_lock(&rcu_gp_lock
);
570 saved_fork_signal_mask
= oldmask
;
573 void rcu_bp_after_fork_parent(void)
578 oldmask
= saved_fork_signal_mask
;
579 mutex_unlock(&rcu_gp_lock
);
580 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
585 * Prune all entries from registry except our own thread. Fits the Linux
586 * fork behavior. Called with rcu_gp_lock held.
589 void urcu_bp_prune_registry(void)
591 struct registry_chunk
*chunk
;
592 struct rcu_reader
*rcu_reader_reg
;
594 cds_list_for_each_entry(chunk
, ®istry_arena
.chunk_list
, node
) {
595 for (rcu_reader_reg
= (struct rcu_reader
*) &chunk
->data
[0];
596 rcu_reader_reg
< (struct rcu_reader
*) &chunk
->data
[chunk
->data_len
];
598 if (!rcu_reader_reg
->alloc
)
600 if (rcu_reader_reg
->tid
== pthread_self())
602 cleanup_thread(chunk
, rcu_reader_reg
);
607 void rcu_bp_after_fork_child(void)
612 urcu_bp_prune_registry();
613 oldmask
= saved_fork_signal_mask
;
614 mutex_unlock(&rcu_gp_lock
);
615 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
619 void *rcu_dereference_sym_bp(void *p
)
621 return _rcu_dereference(p
);
624 void *rcu_set_pointer_sym_bp(void **p
, void *v
)
631 void *rcu_xchg_pointer_sym_bp(void **p
, void *v
)
634 return uatomic_xchg(p
, v
);
637 void *rcu_cmpxchg_pointer_sym_bp(void **p
, void *old
, void *_new
)
640 return uatomic_cmpxchg(p
, old
, _new
);
643 DEFINE_RCU_FLAVOR(rcu_flavor
);
645 #include "urcu-call-rcu-impl.h"
646 #include "urcu-defer-impl.h"