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@efficios.com>
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.h>
41 #include <urcu/list.h>
42 #include <urcu/futex.h>
48 /* Default is RCU_MEMBARRIER on linux */
49 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
51 # define RCU_MEMBARRIER
60 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
62 # define MEMBARRIER_EXPEDITED (1 << 0)
63 # define MEMBARRIER_DELAYED (1 << 1)
64 # define MEMBARRIER_QUERY (1 << 16)
65 # define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__)
67 # undef RCU_MEMBARRIER
73 * This code section can only be included in LGPL 2.1 compatible source code.
74 * See below for the function call wrappers which can be used in code meant to
75 * be only linked with the Userspace RCU library. This comes with a small
76 * performance degradation on the read-side due to the added function calls.
77 * This is required to permit relinking with newer versions of the library.
81 * The signal number used by the RCU library can be overridden with
82 * -DSIGRCU= when compiling the library.
83 * Provide backward compatibility for liburcu 0.3.x SIGURCU.
86 #define SIGRCU SIGURCU
90 #define SIGRCU SIGUSR1
94 * If a reader is really non-cooperative and refuses to commit its
95 * rcu_active_readers count to memory (there is no barrier in the reader
96 * per-se), kick it after a few loops waiting for it.
98 #define KICK_READER_LOOPS 10000
101 * Active attempts to check for reader Q.S. before calling futex().
103 #define RCU_QS_ACTIVE_ATTEMPTS 100
106 #define rcu_assert(args...) assert(args)
108 #define rcu_assert(args...)
117 #define YIELD_READ (1 << 0)
118 #define YIELD_WRITE (1 << 1)
121 * Updates with RCU_SIGNAL are much slower. Account this in the delay.
124 /* maximum sleep delay, in us */
125 #define MAX_SLEEP 30000
130 extern unsigned int yield_active
;
131 extern unsigned int __thread rand_yield
;
133 static inline void debug_yield_read(void)
135 if (yield_active
& YIELD_READ
)
136 if (rand_r(&rand_yield
) & 0x1)
137 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
140 static inline void debug_yield_write(void)
142 if (yield_active
& YIELD_WRITE
)
143 if (rand_r(&rand_yield
) & 0x1)
144 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
147 static inline void debug_yield_init(void)
149 rand_yield
= time(NULL
) ^ pthread_self();
152 static inline void debug_yield_read(void)
156 static inline void debug_yield_write(void)
160 static inline void debug_yield_init(void)
167 * RCU memory barrier broadcast group. Currently, only broadcast to all process
168 * threads is supported (group 0).
170 * Slave barriers are only guaranteed to be ordered wrt master barriers.
172 * The pair ordering is detailed as (O: ordered, X: not ordered) :
178 #define MB_GROUP_ALL 0
179 #define RCU_MB_GROUP MB_GROUP_ALL
181 #ifdef RCU_MEMBARRIER
182 extern int has_sys_membarrier
;
184 static inline void smp_mb_slave(int group
)
186 if (likely(has_sys_membarrier
))
194 static inline void smp_mb_slave(int group
)
201 static inline void smp_mb_slave(int group
)
208 * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
209 * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
211 #define RCU_GP_COUNT (1UL << 0)
212 /* Use the amount of bits equal to half of the architecture long size */
213 #define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2))
214 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1)
217 * Global quiescent period counter with low-order bits unused.
218 * Using a int rather than a char to eliminate false register dependencies
219 * causing stalls on some architectures.
221 extern unsigned long rcu_gp_ctr
;
224 /* Data used by both reader and synchronize_rcu() */
227 /* Data used for registry */
228 struct cds_list_head node
__attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
232 extern struct rcu_reader __thread rcu_reader
;
234 extern int32_t gp_futex
;
237 * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
239 static inline void wake_up_gp(void)
241 if (unlikely(uatomic_read(&gp_futex
) == -1)) {
242 uatomic_set(&gp_futex
, 0);
243 futex_async(&gp_futex
, FUTEX_WAKE
, 1,
248 static inline int rcu_gp_ongoing(unsigned long *ctr
)
253 * Make sure both tests below are done on the same version of *value
254 * to insure consistency.
256 v
= CMM_LOAD_SHARED(*ctr
);
257 return (v
& RCU_GP_CTR_NEST_MASK
) &&
258 ((v
^ rcu_gp_ctr
) & RCU_GP_CTR_PHASE
);
261 static inline void _rcu_read_lock(void)
265 cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
266 tmp
= rcu_reader
.ctr
;
269 * RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE)
271 if (likely(!(tmp
& RCU_GP_CTR_NEST_MASK
))) {
272 _CMM_STORE_SHARED(rcu_reader
.ctr
, _CMM_LOAD_SHARED(rcu_gp_ctr
));
274 * Set active readers count for outermost nesting level before
275 * accessing the pointer. See smp_mb_master().
277 smp_mb_slave(RCU_MB_GROUP
);
279 _CMM_STORE_SHARED(rcu_reader
.ctr
, tmp
+ RCU_GP_COUNT
);
283 static inline void _rcu_read_unlock(void)
287 tmp
= rcu_reader
.ctr
;
289 * Finish using rcu before decrementing the pointer.
290 * See smp_mb_master().
292 if (likely((tmp
& RCU_GP_CTR_NEST_MASK
) == RCU_GP_COUNT
)) {
293 smp_mb_slave(RCU_MB_GROUP
);
294 _CMM_STORE_SHARED(rcu_reader
.ctr
, rcu_reader
.ctr
- RCU_GP_COUNT
);
295 /* write rcu_reader.ctr before read futex */
296 smp_mb_slave(RCU_MB_GROUP
);
299 _CMM_STORE_SHARED(rcu_reader
.ctr
, rcu_reader
.ctr
- RCU_GP_COUNT
);
301 cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
308 #endif /* _URCU_STATIC_H */