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"
45 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
51 #define MAP_ANONYMOUS MAP_ANON
56 void *mremap_wrapper(void *old_address
, size_t old_size
,
57 size_t new_size
, int flags
)
59 return mremap(old_address
, old_size
, new_size
, flags
);
63 #define MREMAP_MAYMOVE 1
64 #define MREMAP_FIXED 2
67 * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping.
68 * This is not generic.
71 void *mremap_wrapper(void *old_address
, size_t old_size
,
72 size_t new_size
, int flags
)
76 assert(flags
& MREMAP_MAYMOVE
);
77 assert(!(flags
& MREMAP_FIXED
));
78 new_address
= mmap(old_address
, new_size
,
79 PROT_READ
| PROT_WRITE
,
80 MAP_ANONYMOUS
| MAP_PRIVATE
,
82 if (new_address
== MAP_FAILED
)
85 memcpy(new_address
, old_address
, old_size
);
86 munmap(old_address
, old_size
);
92 /* Sleep delay in us */
93 #define RCU_SLEEP_DELAY 1000
94 #define ARENA_INIT_ALLOC 16
97 * Active attempts to check for reader Q.S. before calling sleep().
99 #define RCU_QS_ACTIVE_ATTEMPTS 100
101 void __attribute__((destructor
)) rcu_bp_exit(void);
103 static pthread_mutex_t rcu_gp_lock
= PTHREAD_MUTEX_INITIALIZER
;
106 unsigned int yield_active
;
107 DEFINE_URCU_TLS(unsigned int, rand_yield
);
111 * Global grace period counter.
112 * Contains the current RCU_GP_CTR_PHASE.
113 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
114 * Written to only by writer with mutex taken. Read by both writer and readers.
116 long rcu_gp_ctr
= RCU_GP_COUNT
;
119 * Pointer to registry elements. Written to only by each individual reader. Read
120 * by both the reader and the writers.
122 DEFINE_URCU_TLS(struct rcu_reader
*, rcu_reader
);
124 static CDS_LIST_HEAD(registry
);
126 struct registry_arena
{
132 static struct registry_arena registry_arena
;
134 /* Saved fork signal mask, protected by rcu_gp_lock */
135 static sigset_t saved_fork_signal_mask
;
137 static void rcu_gc_registry(void);
139 static void mutex_lock(pthread_mutex_t
*mutex
)
143 #ifndef DISTRUST_SIGNALS_EXTREME
144 ret
= pthread_mutex_lock(mutex
);
146 perror("Error in pthread mutex lock");
149 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
150 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
151 if (ret
!= EBUSY
&& ret
!= EINTR
) {
152 printf("ret = %d, errno = %d\n", ret
, errno
);
153 perror("Error in pthread mutex lock");
158 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
161 static void mutex_unlock(pthread_mutex_t
*mutex
)
165 ret
= pthread_mutex_unlock(mutex
);
167 perror("Error in pthread mutex unlock");
172 void update_counter_and_wait(void)
174 CDS_LIST_HEAD(qsreaders
);
176 struct rcu_reader
*index
, *tmp
;
178 /* Switch parity: 0 -> 1, 1 -> 0 */
179 CMM_STORE_SHARED(rcu_gp_ctr
, rcu_gp_ctr
^ RCU_GP_CTR_PHASE
);
182 * Must commit qparity update to memory before waiting for other parity
183 * quiescent state. Failure to do so could result in the writer waiting
184 * forever while new readers are always accessing data (no progress).
185 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
189 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
190 * model easier to understand. It does not have a big performance impact
191 * anyway, given this is the write-side.
196 * Wait for each thread rcu_reader.ctr count to become 0.
200 cds_list_for_each_entry_safe(index
, tmp
, ®istry
, node
) {
201 if (!rcu_old_gp_ongoing(&index
->ctr
))
202 cds_list_move(&index
->node
, &qsreaders
);
205 if (cds_list_empty(®istry
)) {
208 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
)
209 usleep(RCU_SLEEP_DELAY
);
214 /* put back the reader list in the registry */
215 cds_list_splice(&qsreaders
, ®istry
);
218 void synchronize_rcu(void)
220 sigset_t newmask
, oldmask
;
223 ret
= sigemptyset(&newmask
);
225 ret
= pthread_sigmask(SIG_SETMASK
, &newmask
, &oldmask
);
228 mutex_lock(&rcu_gp_lock
);
230 if (cds_list_empty(®istry
))
233 /* All threads should read qparity before accessing data structure
234 * where new ptr points to. */
235 /* Write new ptr before changing the qparity */
238 /* Remove old registry elements */
242 * Wait for previous parity to be empty of readers.
244 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
247 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
248 * model easier to understand. It does not have a big performance impact
249 * anyway, given this is the write-side.
254 * Wait for previous parity to be empty of readers.
256 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
259 * Finish waiting for reader threads before letting the old ptr being
264 mutex_unlock(&rcu_gp_lock
);
265 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
270 * library wrappers to be used by non-LGPL compatible source code.
273 void rcu_read_lock(void)
278 void rcu_read_unlock(void)
286 static void resize_arena(struct registry_arena
*arena
, size_t len
)
291 new_arena
= mmap(arena
->p
, len
,
292 PROT_READ
| PROT_WRITE
,
293 MAP_ANONYMOUS
| MAP_PRIVATE
,
296 new_arena
= mremap_wrapper(arena
->p
, arena
->len
,
297 len
, MREMAP_MAYMOVE
);
298 assert(new_arena
!= MAP_FAILED
);
301 * re-used the same region ?
303 if (new_arena
== arena
->p
)
306 bzero(new_arena
+ arena
->len
, len
- arena
->len
);
307 arena
->p
= new_arena
;
310 /* Called with signals off and mutex locked */
311 static void add_thread(void)
313 struct rcu_reader
*rcu_reader_reg
;
315 if (registry_arena
.len
316 < registry_arena
.used
+ sizeof(struct rcu_reader
))
317 resize_arena(®istry_arena
,
318 caa_max(registry_arena
.len
<< 1, ARENA_INIT_ALLOC
));
322 for (rcu_reader_reg
= registry_arena
.p
;
323 (void *)rcu_reader_reg
< registry_arena
.p
+ registry_arena
.len
;
325 if (!rcu_reader_reg
->alloc
)
328 rcu_reader_reg
->alloc
= 1;
329 registry_arena
.used
+= sizeof(struct rcu_reader
);
331 /* Add to registry */
332 rcu_reader_reg
->tid
= pthread_self();
333 assert(rcu_reader_reg
->ctr
== 0);
334 cds_list_add(&rcu_reader_reg
->node
, ®istry
);
335 URCU_TLS(rcu_reader
) = rcu_reader_reg
;
338 /* Called with signals off and mutex locked */
339 static void rcu_gc_registry(void)
341 struct rcu_reader
*rcu_reader_reg
;
345 for (rcu_reader_reg
= registry_arena
.p
;
346 (void *)rcu_reader_reg
< registry_arena
.p
+ registry_arena
.len
;
348 if (!rcu_reader_reg
->alloc
)
350 tid
= rcu_reader_reg
->tid
;
351 ret
= pthread_kill(tid
, 0);
352 assert(ret
!= EINVAL
);
354 cds_list_del(&rcu_reader_reg
->node
);
355 rcu_reader_reg
->ctr
= 0;
356 rcu_reader_reg
->alloc
= 0;
357 registry_arena
.used
-= sizeof(struct rcu_reader
);
362 /* Disable signals, take mutex, add to registry */
363 void rcu_bp_register(void)
365 sigset_t newmask
, oldmask
;
368 ret
= sigemptyset(&newmask
);
370 ret
= pthread_sigmask(SIG_SETMASK
, &newmask
, &oldmask
);
374 * Check if a signal concurrently registered our thread since
375 * the check in rcu_read_lock(). */
376 if (URCU_TLS(rcu_reader
))
379 mutex_lock(&rcu_gp_lock
);
381 mutex_unlock(&rcu_gp_lock
);
383 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
387 void rcu_bp_exit(void)
389 if (registry_arena
.p
)
390 munmap(registry_arena
.p
, registry_arena
.len
);
394 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
395 * a concurrent thread executing with this same lock held. This ensures that the
396 * registry is in a coherent state in the child.
398 void rcu_bp_before_fork(void)
400 sigset_t newmask
, oldmask
;
403 ret
= sigemptyset(&newmask
);
405 ret
= pthread_sigmask(SIG_SETMASK
, &newmask
, &oldmask
);
407 mutex_lock(&rcu_gp_lock
);
408 saved_fork_signal_mask
= oldmask
;
411 void rcu_bp_after_fork_parent(void)
416 oldmask
= saved_fork_signal_mask
;
417 mutex_unlock(&rcu_gp_lock
);
418 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
422 void rcu_bp_after_fork_child(void)
428 oldmask
= saved_fork_signal_mask
;
429 mutex_unlock(&rcu_gp_lock
);
430 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
434 void *rcu_dereference_sym_bp(void *p
)
436 return _rcu_dereference(p
);
439 void *rcu_set_pointer_sym_bp(void **p
, void *v
)
446 void *rcu_xchg_pointer_sym_bp(void **p
, void *v
)
449 return uatomic_xchg(p
, v
);
452 void *rcu_cmpxchg_pointer_sym_bp(void **p
, void *old
, void *_new
)
455 return uatomic_cmpxchg(p
, old
, _new
);
458 DEFINE_RCU_FLAVOR(rcu_flavor
);
460 #include "urcu-call-rcu-impl.h"
461 #include "urcu-defer-impl.h"