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/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. Maps a RW, anonymous private mapping.
70 * This is not generic.
73 void *mremap_wrapper(void *old_address
, size_t old_size
,
74 size_t new_size
, int flags
)
78 assert(flags
& MREMAP_MAYMOVE
);
79 assert(!(flags
& MREMAP_FIXED
));
80 new_address
= mmap(old_address
, new_size
,
81 PROT_READ
| PROT_WRITE
,
82 MAP_ANONYMOUS
| MAP_PRIVATE
,
84 if (new_address
== MAP_FAILED
)
87 memcpy(new_address
, old_address
, old_size
);
88 munmap(old_address
, old_size
);
94 /* Sleep delay in us */
95 #define RCU_SLEEP_DELAY 1000
96 #define ARENA_INIT_ALLOC 16
99 * Active attempts to check for reader Q.S. before calling sleep().
101 #define RCU_QS_ACTIVE_ATTEMPTS 100
103 void __attribute__((destructor
)) rcu_bp_exit(void);
105 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
108 unsigned int rcu_yield_active
;
109 DEFINE_URCU_TLS(unsigned int, rcu_rand_yield
);
112 struct rcu_gp rcu_gp
= { .ctr
= RCU_GP_COUNT
};
115 * Pointer to registry elements. Written to only by each individual reader. Read
116 * by both the reader and the writers.
118 DEFINE_URCU_TLS(struct rcu_reader
*, rcu_reader
);
120 static CDS_LIST_HEAD(registry
);
122 struct registry_arena
{
128 static struct registry_arena registry_arena
;
130 /* Saved fork signal mask, protected by rcu_gp_lock */
131 static sigset_t saved_fork_signal_mask
;
133 static void rcu_gc_registry(void);
135 static void mutex_lock(pthread_mutex_t
*mutex
)
139 #ifndef DISTRUST_SIGNALS_EXTREME
140 ret
= pthread_mutex_lock(mutex
);
143 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
144 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
145 if (ret
!= EBUSY
&& ret
!= EINTR
)
149 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
152 static void mutex_unlock(pthread_mutex_t
*mutex
)
156 ret
= pthread_mutex_unlock(mutex
);
161 static void wait_for_readers(struct cds_list_head
*input_readers
,
162 struct cds_list_head
*cur_snap_readers
,
163 struct cds_list_head
*qsreaders
)
166 struct rcu_reader
*index
, *tmp
;
169 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
170 * indicate quiescence (not nested), or observe the current
175 cds_list_for_each_entry_safe(index
, tmp
, input_readers
, node
) {
176 switch (rcu_reader_state(&index
->ctr
)) {
177 case RCU_READER_ACTIVE_CURRENT
:
178 if (cur_snap_readers
) {
179 cds_list_move(&index
->node
,
184 case RCU_READER_INACTIVE
:
185 cds_list_move(&index
->node
, qsreaders
);
187 case RCU_READER_ACTIVE_OLD
:
189 * Old snapshot. Leaving node in
190 * input_readers will make us busy-loop
191 * until the snapshot becomes current or
192 * the reader becomes inactive.
198 if (cds_list_empty(input_readers
)) {
201 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
)
202 usleep(RCU_SLEEP_DELAY
);
209 void synchronize_rcu(void)
211 CDS_LIST_HEAD(cur_snap_readers
);
212 CDS_LIST_HEAD(qsreaders
);
213 sigset_t newmask
, oldmask
;
216 ret
= sigfillset(&newmask
);
218 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
221 mutex_lock(&rcu_gp_lock
);
223 if (cds_list_empty(®istry
))
226 /* All threads should read qparity before accessing data structure
227 * where new ptr points to. */
228 /* Write new ptr before changing the qparity */
231 /* Remove old registry elements */
235 * Wait for readers to observe original parity or be quiescent.
237 wait_for_readers(®istry
, &cur_snap_readers
, &qsreaders
);
240 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
241 * model easier to understand. It does not have a big performance impact
242 * anyway, given this is the write-side.
246 /* Switch parity: 0 -> 1, 1 -> 0 */
247 CMM_STORE_SHARED(rcu_gp
.ctr
, rcu_gp
.ctr
^ RCU_GP_CTR_PHASE
);
250 * Must commit qparity update to memory before waiting for other parity
251 * quiescent state. Failure to do so could result in the writer waiting
252 * forever while new readers are always accessing data (no progress).
253 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
257 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
258 * model easier to understand. It does not have a big performance impact
259 * anyway, given this is the write-side.
264 * Wait for readers to observe new parity or be quiescent.
266 wait_for_readers(&cur_snap_readers
, NULL
, &qsreaders
);
269 * Put quiescent reader list back into registry.
271 cds_list_splice(&qsreaders
, ®istry
);
274 * Finish waiting for reader threads before letting the old ptr being
279 mutex_unlock(&rcu_gp_lock
);
280 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
285 * library wrappers to be used by non-LGPL compatible source code.
288 void rcu_read_lock(void)
293 void rcu_read_unlock(void)
298 int rcu_read_ongoing(void)
300 return _rcu_read_ongoing();
306 static void resize_arena(struct registry_arena
*arena
, size_t len
)
311 old_len
= arena
->len
;
314 new_p
= mmap(arena
->p
, len
,
315 PROT_READ
| PROT_WRITE
,
316 MAP_ANONYMOUS
| MAP_PRIVATE
,
319 new_p
= mremap_wrapper(arena
->p
, old_len
,
320 len
, MREMAP_MAYMOVE
);
321 assert(new_p
!= MAP_FAILED
);
324 * Zero the newly allocated memory. Since mmap() does not
325 * clearly specify if memory is zeroed or not (although it is
326 * very likely that it is), be extra careful by not expecting
327 * the new range to be zeroed by mremap.
329 bzero(new_p
+ old_len
, len
- old_len
);
332 * If we did not re-use the same region, we need to update the
335 if (new_p
!= arena
->p
)
341 /* Called with signals off and mutex locked */
342 static void add_thread(void)
344 struct rcu_reader
*rcu_reader_reg
;
346 if (registry_arena
.len
347 < registry_arena
.used
+ sizeof(struct rcu_reader
))
348 resize_arena(®istry_arena
,
349 caa_max(registry_arena
.len
<< 1, ARENA_INIT_ALLOC
));
353 for (rcu_reader_reg
= registry_arena
.p
;
354 (void *)rcu_reader_reg
< registry_arena
.p
+ registry_arena
.len
;
356 if (!rcu_reader_reg
->alloc
)
359 rcu_reader_reg
->alloc
= 1;
360 registry_arena
.used
+= sizeof(struct rcu_reader
);
362 /* Add to registry */
363 rcu_reader_reg
->tid
= pthread_self();
364 assert(rcu_reader_reg
->ctr
== 0);
365 cds_list_add(&rcu_reader_reg
->node
, ®istry
);
366 URCU_TLS(rcu_reader
) = rcu_reader_reg
;
369 /* Called with signals off and mutex locked */
370 static void rcu_gc_registry(void)
372 struct rcu_reader
*rcu_reader_reg
;
376 for (rcu_reader_reg
= registry_arena
.p
;
377 (void *)rcu_reader_reg
< registry_arena
.p
+ registry_arena
.len
;
379 if (!rcu_reader_reg
->alloc
)
381 tid
= rcu_reader_reg
->tid
;
382 ret
= pthread_kill(tid
, 0);
383 assert(ret
!= EINVAL
);
385 cds_list_del(&rcu_reader_reg
->node
);
386 rcu_reader_reg
->ctr
= 0;
387 rcu_reader_reg
->alloc
= 0;
388 registry_arena
.used
-= sizeof(struct rcu_reader
);
393 /* Disable signals, take mutex, add to registry */
394 void rcu_bp_register(void)
396 sigset_t newmask
, oldmask
;
399 ret
= sigfillset(&newmask
);
401 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
405 * Check if a signal concurrently registered our thread since
406 * the check in rcu_read_lock(). */
407 if (URCU_TLS(rcu_reader
))
410 mutex_lock(&rcu_gp_lock
);
412 mutex_unlock(&rcu_gp_lock
);
414 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
418 void rcu_bp_exit(void)
420 if (registry_arena
.p
)
421 munmap(registry_arena
.p
, registry_arena
.len
);
425 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
426 * a concurrent thread executing with this same lock held. This ensures that the
427 * registry is in a coherent state in the child.
429 void rcu_bp_before_fork(void)
431 sigset_t newmask
, oldmask
;
434 ret
= sigfillset(&newmask
);
436 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
438 mutex_lock(&rcu_gp_lock
);
439 saved_fork_signal_mask
= oldmask
;
442 void rcu_bp_after_fork_parent(void)
447 oldmask
= saved_fork_signal_mask
;
448 mutex_unlock(&rcu_gp_lock
);
449 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
453 void rcu_bp_after_fork_child(void)
459 oldmask
= saved_fork_signal_mask
;
460 mutex_unlock(&rcu_gp_lock
);
461 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
465 void *rcu_dereference_sym_bp(void *p
)
467 return _rcu_dereference(p
);
470 void *rcu_set_pointer_sym_bp(void **p
, void *v
)
477 void *rcu_xchg_pointer_sym_bp(void **p
, void *v
)
480 return uatomic_xchg(p
, v
);
483 void *rcu_cmpxchg_pointer_sym_bp(void **p
, void *old
, void *_new
)
486 return uatomic_cmpxchg(p
, old
, _new
);
489 DEFINE_RCU_FLAVOR(rcu_flavor
);
491 #include "urcu-call-rcu-impl.h"
492 #include "urcu-defer-impl.h"