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. 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 yield_active
;
109 DEFINE_URCU_TLS(unsigned int, rand_yield
);
113 * Global grace period counter.
114 * Contains the current RCU_GP_CTR_PHASE.
115 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
116 * Written to only by writer with mutex taken. Read by both writer and readers.
118 long rcu_gp_ctr
= RCU_GP_COUNT
;
121 * Pointer to registry elements. Written to only by each individual reader. Read
122 * by both the reader and the writers.
124 DEFINE_URCU_TLS(struct rcu_reader
*, rcu_reader
);
126 static CDS_LIST_HEAD(registry
);
128 struct registry_arena
{
134 static struct registry_arena registry_arena
;
136 /* Saved fork signal mask, protected by rcu_gp_lock */
137 static sigset_t saved_fork_signal_mask
;
139 static void rcu_gc_registry(void);
141 static void mutex_lock(pthread_mutex_t
*mutex
)
145 #ifndef DISTRUST_SIGNALS_EXTREME
146 ret
= pthread_mutex_lock(mutex
);
149 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
150 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
151 if (ret
!= EBUSY
&& ret
!= EINTR
)
155 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
158 static void mutex_unlock(pthread_mutex_t
*mutex
)
162 ret
= pthread_mutex_unlock(mutex
);
167 void update_counter_and_wait(void)
169 CDS_LIST_HEAD(qsreaders
);
171 struct rcu_reader
*index
, *tmp
;
173 /* Switch parity: 0 -> 1, 1 -> 0 */
174 CMM_STORE_SHARED(rcu_gp_ctr
, rcu_gp_ctr
^ RCU_GP_CTR_PHASE
);
177 * Must commit qparity update to memory before waiting for other parity
178 * quiescent state. Failure to do so could result in the writer waiting
179 * forever while new readers are always accessing data (no progress).
180 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
184 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
185 * model easier to understand. It does not have a big performance impact
186 * anyway, given this is the write-side.
191 * Wait for each thread rcu_reader.ctr count to become 0.
195 cds_list_for_each_entry_safe(index
, tmp
, ®istry
, node
) {
196 if (!rcu_old_gp_ongoing(&index
->ctr
))
197 cds_list_move(&index
->node
, &qsreaders
);
200 if (cds_list_empty(®istry
)) {
203 if (wait_loops
== RCU_QS_ACTIVE_ATTEMPTS
)
204 usleep(RCU_SLEEP_DELAY
);
209 /* put back the reader list in the registry */
210 cds_list_splice(&qsreaders
, ®istry
);
213 void synchronize_rcu(void)
215 sigset_t newmask
, oldmask
;
218 ret
= sigfillset(&newmask
);
220 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
223 mutex_lock(&rcu_gp_lock
);
225 if (cds_list_empty(®istry
))
228 /* All threads should read qparity before accessing data structure
229 * where new ptr points to. */
230 /* Write new ptr before changing the qparity */
233 /* Remove old registry elements */
237 * Wait for previous parity to be empty of readers.
239 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
242 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
243 * model easier to understand. It does not have a big performance impact
244 * anyway, given this is the write-side.
249 * Wait for previous parity to be empty of readers.
251 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
254 * Finish waiting for reader threads before letting the old ptr being
259 mutex_unlock(&rcu_gp_lock
);
260 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
265 * library wrappers to be used by non-LGPL compatible source code.
268 void rcu_read_lock(void)
273 void rcu_read_unlock(void)
281 static void resize_arena(struct registry_arena
*arena
, size_t len
)
286 old_len
= arena
->len
;
289 new_p
= mmap(arena
->p
, len
,
290 PROT_READ
| PROT_WRITE
,
291 MAP_ANONYMOUS
| MAP_PRIVATE
,
294 new_p
= mremap_wrapper(arena
->p
, old_len
,
295 len
, MREMAP_MAYMOVE
);
296 assert(new_p
!= MAP_FAILED
);
299 * Zero the newly allocated memory. Since mmap() does not
300 * clearly specify if memory is zeroed or not (although it is
301 * very likely that it is), be extra careful by not expecting
302 * the new range to be zeroed by mremap.
304 bzero(new_p
+ old_len
, len
- old_len
);
307 * If we did not re-use the same region, we need to update the
310 if (new_p
!= arena
->p
)
316 /* Called with signals off and mutex locked */
317 static void add_thread(void)
319 struct rcu_reader
*rcu_reader_reg
;
321 if (registry_arena
.len
322 < registry_arena
.used
+ sizeof(struct rcu_reader
))
323 resize_arena(®istry_arena
,
324 caa_max(registry_arena
.len
<< 1, ARENA_INIT_ALLOC
));
328 for (rcu_reader_reg
= registry_arena
.p
;
329 (void *)rcu_reader_reg
< registry_arena
.p
+ registry_arena
.len
;
331 if (!rcu_reader_reg
->alloc
)
334 rcu_reader_reg
->alloc
= 1;
335 registry_arena
.used
+= sizeof(struct rcu_reader
);
337 /* Add to registry */
338 rcu_reader_reg
->tid
= pthread_self();
339 assert(rcu_reader_reg
->ctr
== 0);
340 cds_list_add(&rcu_reader_reg
->node
, ®istry
);
341 URCU_TLS(rcu_reader
) = rcu_reader_reg
;
344 /* Called with signals off and mutex locked */
345 static void rcu_gc_registry(void)
347 struct rcu_reader
*rcu_reader_reg
;
351 for (rcu_reader_reg
= registry_arena
.p
;
352 (void *)rcu_reader_reg
< registry_arena
.p
+ registry_arena
.len
;
354 if (!rcu_reader_reg
->alloc
)
356 tid
= rcu_reader_reg
->tid
;
357 ret
= pthread_kill(tid
, 0);
358 assert(ret
!= EINVAL
);
360 cds_list_del(&rcu_reader_reg
->node
);
361 rcu_reader_reg
->ctr
= 0;
362 rcu_reader_reg
->alloc
= 0;
363 registry_arena
.used
-= sizeof(struct rcu_reader
);
368 /* Disable signals, take mutex, add to registry */
369 void rcu_bp_register(void)
371 sigset_t newmask
, oldmask
;
374 ret
= sigfillset(&newmask
);
376 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
380 * Check if a signal concurrently registered our thread since
381 * the check in rcu_read_lock(). */
382 if (URCU_TLS(rcu_reader
))
385 mutex_lock(&rcu_gp_lock
);
387 mutex_unlock(&rcu_gp_lock
);
389 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
393 void rcu_bp_exit(void)
395 if (registry_arena
.p
)
396 munmap(registry_arena
.p
, registry_arena
.len
);
400 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
401 * a concurrent thread executing with this same lock held. This ensures that the
402 * registry is in a coherent state in the child.
404 void rcu_bp_before_fork(void)
406 sigset_t newmask
, oldmask
;
409 ret
= sigfillset(&newmask
);
411 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
413 mutex_lock(&rcu_gp_lock
);
414 saved_fork_signal_mask
= oldmask
;
417 void rcu_bp_after_fork_parent(void)
422 oldmask
= saved_fork_signal_mask
;
423 mutex_unlock(&rcu_gp_lock
);
424 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
428 void rcu_bp_after_fork_child(void)
434 oldmask
= saved_fork_signal_mask
;
435 mutex_unlock(&rcu_gp_lock
);
436 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
440 void *rcu_dereference_sym_bp(void *p
)
442 return _rcu_dereference(p
);
445 void *rcu_set_pointer_sym_bp(void **p
, void *v
)
452 void *rcu_xchg_pointer_sym_bp(void **p
, void *v
)
455 return uatomic_xchg(p
, v
);
458 void *rcu_cmpxchg_pointer_sym_bp(void **p
, void *old
, void *_new
)
461 return uatomic_cmpxchg(p
, old
, _new
);
464 DEFINE_RCU_FLAVOR(rcu_flavor
);
466 #include "urcu-call-rcu-impl.h"
467 #include "urcu-defer-impl.h"