7 * Userspace RCU header.
9 * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking
10 * dynamically with the userspace rcu library.
12 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
13 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
15 * This library is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU Lesser General Public
17 * License as published by the Free Software Foundation; either
18 * version 2.1 of the License, or (at your option) any later version.
20 * This library is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * Lesser General Public License for more details.
25 * You should have received a copy of the GNU Lesser General Public
26 * License along with this library; if not, write to the Free Software
27 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
29 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
37 #include <urcu/compiler.h>
38 #include <urcu/arch.h>
39 #include <urcu/system.h>
40 #include <urcu/uatomic_arch.h>
41 #include <urcu/list.h>
42 #include <urcu/urcu-futex.h>
48 /* Default is RCU_MEMBARRIER */
49 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
50 #define RCU_MEMBARRIER
55 #include <sys/syscall.h>
57 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
59 # define MEMBARRIER_EXPEDITED (1 << 0)
60 # define MEMBARRIER_DELAYED (1 << 1)
61 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
63 # undef RCU_MEMBARRIER
69 * This code section can only be included in LGPL 2.1 compatible source code.
70 * See below for the function call wrappers which can be used in code meant to
71 * be only linked with the Userspace RCU library. This comes with a small
72 * performance degradation on the read-side due to the added function calls.
73 * This is required to permit relinking with newer versions of the library.
77 * The signal number used by the RCU library can be overridden with
78 * -DSIGRCU= when compiling the library.
81 #define SIGRCU SIGUSR1
85 * If a reader is really non-cooperative and refuses to commit its
86 * rcu_active_readers count to memory (there is no barrier in the reader
87 * per-se), kick it after a few loops waiting for it.
89 #define KICK_READER_LOOPS 10000
92 * Active attempts to check for reader Q.S. before calling futex().
94 #define RCU_QS_ACTIVE_ATTEMPTS 100
97 #define rcu_assert(args...) assert(args)
99 #define rcu_assert(args...)
108 #define YIELD_READ (1 << 0)
109 #define YIELD_WRITE (1 << 1)
112 * Updates with RCU_SIGNAL are much slower. Account this in the delay.
115 /* maximum sleep delay, in us */
116 #define MAX_SLEEP 30000
121 extern unsigned int yield_active
;
122 extern unsigned int __thread rand_yield
;
124 static inline void debug_yield_read(void)
126 if (yield_active
& YIELD_READ
)
127 if (rand_r(&rand_yield
) & 0x1)
128 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
131 static inline void debug_yield_write(void)
133 if (yield_active
& YIELD_WRITE
)
134 if (rand_r(&rand_yield
) & 0x1)
135 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
138 static inline void debug_yield_init(void)
140 rand_yield
= time(NULL
) ^ pthread_self();
143 static inline void debug_yield_read(void)
147 static inline void debug_yield_write(void)
151 static inline void debug_yield_init(void)
157 #ifdef RCU_MEMBARRIER
158 extern int has_sys_membarrier
;
160 static inline void smp_mb_light()
162 if (likely(has_sys_membarrier
))
170 static inline void smp_mb_light()
177 static inline void smp_mb_light()
184 * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
185 * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
187 #define RCU_GP_COUNT (1UL << 0)
188 /* Use the amount of bits equal to half of the architecture long size */
189 #define RCU_GP_CTR_PHASE (1UL << (sizeof(long) << 2))
190 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1)
193 * Global quiescent period counter with low-order bits unused.
194 * Using a int rather than a char to eliminate false register dependencies
195 * causing stalls on some architectures.
197 extern long rcu_gp_ctr
;
200 /* Data used by both reader and synchronize_rcu() */
203 /* Data used for registry */
204 struct list_head head
__attribute__((aligned(CACHE_LINE_SIZE
)));
208 extern struct rcu_reader __thread rcu_reader
;
213 * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
215 static inline void wake_up_gp(void)
217 if (unlikely(uatomic_read(&gp_futex
) == -1)) {
218 uatomic_set(&gp_futex
, 0);
219 futex_async(&gp_futex
, FUTEX_WAKE
, 1,
224 static inline int rcu_old_gp_ongoing(long *value
)
231 * Make sure both tests below are done on the same version of *value
232 * to insure consistency.
234 v
= LOAD_SHARED(*value
);
235 return (v
& RCU_GP_CTR_NEST_MASK
) &&
236 ((v
^ rcu_gp_ctr
) & RCU_GP_CTR_PHASE
);
239 static inline void _rcu_read_lock(void)
243 tmp
= rcu_reader
.ctr
;
246 * RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE)
248 if (likely(!(tmp
& RCU_GP_CTR_NEST_MASK
))) {
249 _STORE_SHARED(rcu_reader
.ctr
, _LOAD_SHARED(rcu_gp_ctr
));
251 * Set active readers count for outermost nesting level before
252 * accessing the pointer. See smp_mb_heavy().
256 _STORE_SHARED(rcu_reader
.ctr
, tmp
+ RCU_GP_COUNT
);
260 static inline void _rcu_read_unlock(void)
264 tmp
= rcu_reader
.ctr
;
266 * Finish using rcu before decrementing the pointer.
267 * See smp_mb_heavy().
269 if (likely((tmp
& RCU_GP_CTR_NEST_MASK
) == RCU_GP_COUNT
)) {
271 _STORE_SHARED(rcu_reader
.ctr
, rcu_reader
.ctr
- RCU_GP_COUNT
);
272 /* write rcu_reader.ctr before read futex */
276 _STORE_SHARED(rcu_reader
.ctr
, rcu_reader
.ctr
- RCU_GP_COUNT
);
284 #endif /* _URCU_STATIC_H */