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.
39 * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
41 #define _LOAD_SHARED(p) ACCESS_ONCE(p)
44 * Load a data from shared memory, doing a cache flush if required.
46 #define LOAD_SHARED(p) \
53 * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
55 #define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); })
58 * Store v into x, where x is located in shared memory. Performs the required
59 * cache flush after writing. Returns v.
61 #define STORE_SHARED(x, v) \
63 _STORE_SHARED(x, v); \
69 * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
70 * into a RCU read-side critical section. The pointer can later be safely
71 * dereferenced within the critical section.
73 * This ensures that the pointer copy is invariant thorough the whole critical
76 * Inserts memory barriers on architectures that require them (currently only
77 * Alpha) and documents which pointers are protected by RCU.
79 * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
80 * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
81 * data read before the pointer read by speculating the value of the pointer.
82 * Correct ordering is ensured because the pointer is read as a volatile access.
83 * This acts as a global side-effect operation, which forbids reordering of
84 * dependent memory operations. Note that such concern about dependency-breaking
85 * optimizations will eventually be taken care of by the "memory_order_consume"
86 * addition to forthcoming C++ standard.
88 * Should match rcu_assign_pointer() or rcu_xchg_pointer().
91 #define _rcu_dereference(p) ({ \
92 typeof(p) _________p1 = LOAD_SHARED(p); \
93 smp_read_barrier_depends(); \
98 * This code section can only be included in LGPL 2.1 compatible source code.
99 * See below for the function call wrappers which can be used in code meant to
100 * be only linked with the Userspace RCU library. This comes with a small
101 * performance degradation on the read-side due to the added function calls.
102 * This is required to permit relinking with newer versions of the library.
106 * The signal number used by the RCU library can be overridden with
107 * -DSIGURCU= when compiling the library.
110 #define SIGURCU SIGUSR1
114 * If a reader is really non-cooperative and refuses to commit its
115 * urcu_active_readers count to memory (there is no barrier in the reader
116 * per-se), kick it after a few loops waiting for it.
118 #define KICK_READER_LOOPS 10000
121 #define rcu_assert(args...) assert(args)
123 #define rcu_assert(args...)
132 #define YIELD_READ (1 << 0)
133 #define YIELD_WRITE (1 << 1)
136 * Updates without URCU_MB are much slower. Account this in
140 /* maximum sleep delay, in us */
143 #define MAX_SLEEP 30000
146 extern unsigned int yield_active
;
147 extern unsigned int __thread rand_yield
;
149 static inline void debug_yield_read(void)
151 if (yield_active
& YIELD_READ
)
152 if (rand_r(&rand_yield
) & 0x1)
153 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
156 static inline void debug_yield_write(void)
158 if (yield_active
& YIELD_WRITE
)
159 if (rand_r(&rand_yield
) & 0x1)
160 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
163 static inline void debug_yield_init(void)
165 rand_yield
= time(NULL
) ^ pthread_self();
168 static inline void debug_yield_read(void)
172 static inline void debug_yield_write(void)
176 static inline void debug_yield_init(void)
183 static inline void reader_barrier()
188 static inline void reader_barrier()
195 * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
196 * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
198 #define RCU_GP_COUNT (1UL << 0)
199 /* Use the amount of bits equal to half of the architecture long size */
200 #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2))
201 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
204 * Global quiescent period counter with low-order bits unused.
205 * Using a int rather than a char to eliminate false register dependencies
206 * causing stalls on some architectures.
208 extern long urcu_gp_ctr
;
210 extern long __thread urcu_active_readers
;
212 static inline int rcu_old_gp_ongoing(long *value
)
219 * Make sure both tests below are done on the same version of *value
220 * to insure consistency.
222 v
= LOAD_SHARED(*value
);
223 return (v
& RCU_GP_CTR_NEST_MASK
) &&
224 ((v
^ urcu_gp_ctr
) & RCU_GP_CTR_BIT
);
227 static inline void _rcu_read_lock(void)
231 tmp
= urcu_active_readers
;
232 /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
233 if (likely(!(tmp
& RCU_GP_CTR_NEST_MASK
))) {
234 _STORE_SHARED(urcu_active_readers
, _LOAD_SHARED(urcu_gp_ctr
));
236 * Set active readers count for outermost nesting level before
237 * accessing the pointer. See force_mb_all_threads().
241 _STORE_SHARED(urcu_active_readers
, tmp
+ RCU_GP_COUNT
);
245 static inline void _rcu_read_unlock(void)
249 * Finish using rcu before decrementing the pointer.
250 * See force_mb_all_threads().
251 * Formally only needed for outermost nesting level, but leave barrier
252 * in place for nested unlocks to remove a branch from the common case
255 _STORE_SHARED(urcu_active_readers
, urcu_active_readers
- RCU_GP_COUNT
);
259 * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
260 * meant to be read by RCU read-side critical sections. Returns the assigned
263 * Documents which pointers will be dereferenced by RCU read-side critical
264 * sections and adds the required memory barriers on architectures requiring
265 * them. It also makes sure the compiler does not reorder code initializing the
266 * data structure before its publication.
268 * Should match rcu_dereference_pointer().
271 #define _rcu_assign_pointer(p, v) \
273 if (!__builtin_constant_p(v) || \
276 STORE_SHARED(p, v); \
280 * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
281 * pointer to the data structure, which can be safely freed after waiting for a
282 * quiescent state using synchronize_rcu().
285 #define _rcu_xchg_pointer(p, v) \
287 if (!__builtin_constant_p(v) || \
294 * Exchanges the pointer and waits for quiescent state.
295 * The pointer returned can be freed.
297 #define _rcu_publish_content(p, v) \
300 oldptr = _rcu_xchg_pointer(p, v); \
305 #endif /* _URCU_STATIC_H */