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/list.h>
42 * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
44 #define _LOAD_SHARED(p) ACCESS_ONCE(p)
47 * Load a data from shared memory, doing a cache flush if required.
49 #define LOAD_SHARED(p) \
56 * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
58 #define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); })
61 * Store v into x, where x is located in shared memory. Performs the required
62 * cache flush after writing. Returns v.
64 #define STORE_SHARED(x, v) \
66 _STORE_SHARED(x, v); \
72 * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
73 * into a RCU read-side critical section. The pointer can later be safely
74 * dereferenced within the critical section.
76 * This ensures that the pointer copy is invariant thorough the whole critical
79 * Inserts memory barriers on architectures that require them (currently only
80 * Alpha) and documents which pointers are protected by RCU.
82 * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
83 * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
84 * data read before the pointer read by speculating the value of the pointer.
85 * Correct ordering is ensured because the pointer is read as a volatile access.
86 * This acts as a global side-effect operation, which forbids reordering of
87 * dependent memory operations. Note that such concern about dependency-breaking
88 * optimizations will eventually be taken care of by the "memory_order_consume"
89 * addition to forthcoming C++ standard.
91 * Should match rcu_assign_pointer() or rcu_xchg_pointer().
94 #define _rcu_dereference(p) ({ \
95 typeof(p) _________p1 = LOAD_SHARED(p); \
96 smp_read_barrier_depends(); \
100 #define futex(...) syscall(__NR_futex, __VA_ARGS__)
105 * This code section can only be included in LGPL 2.1 compatible source code.
106 * See below for the function call wrappers which can be used in code meant to
107 * be only linked with the Userspace RCU library. This comes with a small
108 * performance degradation on the read-side due to the added function calls.
109 * This is required to permit relinking with newer versions of the library.
113 * The signal number used by the RCU library can be overridden with
114 * -DSIGURCU= when compiling the library.
117 #define SIGURCU SIGUSR1
121 * If a reader is really non-cooperative and refuses to commit its
122 * urcu_active_readers count to memory (there is no barrier in the reader
123 * per-se), kick it after a few loops waiting for it.
125 #define KICK_READER_LOOPS 10000
128 * Active attempts to check for reader Q.S. before calling futex().
130 #define RCU_QS_ACTIVE_ATTEMPTS 100
133 #define rcu_assert(args...) assert(args)
135 #define rcu_assert(args...)
144 #define YIELD_READ (1 << 0)
145 #define YIELD_WRITE (1 << 1)
148 * Updates without URCU_MB are much slower. Account this in
152 /* maximum sleep delay, in us */
155 #define MAX_SLEEP 30000
158 extern unsigned int yield_active
;
159 extern unsigned int __thread rand_yield
;
161 static inline void debug_yield_read(void)
163 if (yield_active
& YIELD_READ
)
164 if (rand_r(&rand_yield
) & 0x1)
165 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
168 static inline void debug_yield_write(void)
170 if (yield_active
& YIELD_WRITE
)
171 if (rand_r(&rand_yield
) & 0x1)
172 usleep(rand_r(&rand_yield
) % MAX_SLEEP
);
175 static inline void debug_yield_init(void)
177 rand_yield
= time(NULL
) ^ pthread_self();
180 static inline void debug_yield_read(void)
184 static inline void debug_yield_write(void)
188 static inline void debug_yield_init(void)
195 static inline void reader_barrier()
200 static inline void reader_barrier()
207 * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
208 * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
210 #define RCU_GP_COUNT (1UL << 0)
211 /* Use the amount of bits equal to half of the architecture long size */
212 #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2))
213 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
216 * Global quiescent period counter with low-order bits unused.
217 * Using a int rather than a char to eliminate false register dependencies
218 * causing stalls on some architectures.
220 extern long urcu_gp_ctr
;
224 struct list_head head
;
229 extern struct urcu_reader __thread urcu_reader
;
234 * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
236 static inline void wake_up_gp(void)
238 if (unlikely(uatomic_read(&gp_futex
) == -1)) {
239 uatomic_set(&gp_futex
, 0);
240 futex(&gp_futex
, FUTEX_WAKE
, 1,
245 static inline int rcu_old_gp_ongoing(long *value
)
252 * Make sure both tests below are done on the same version of *value
253 * to insure consistency.
255 v
= LOAD_SHARED(*value
);
256 return (v
& RCU_GP_CTR_NEST_MASK
) &&
257 ((v
^ urcu_gp_ctr
) & RCU_GP_CTR_BIT
);
260 static inline void _rcu_read_lock(void)
264 tmp
= urcu_reader
.ctr
;
265 /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
266 if (likely(!(tmp
& RCU_GP_CTR_NEST_MASK
))) {
267 _STORE_SHARED(urcu_reader
.ctr
, _LOAD_SHARED(urcu_gp_ctr
));
269 * Set active readers count for outermost nesting level before
270 * accessing the pointer. See force_mb_all_threads().
274 _STORE_SHARED(urcu_reader
.ctr
, tmp
+ RCU_GP_COUNT
);
278 static inline void _rcu_read_unlock(void)
282 tmp
= urcu_reader
.ctr
;
284 * Finish using rcu before decrementing the pointer.
285 * See force_mb_all_threads().
287 if (likely((tmp
& RCU_GP_CTR_NEST_MASK
) == RCU_GP_COUNT
)) {
289 _STORE_SHARED(urcu_reader
.ctr
, urcu_reader
.ctr
- RCU_GP_COUNT
);
290 /* write urcu_reader.ctr before read futex */
294 _STORE_SHARED(urcu_reader
.ctr
, urcu_reader
.ctr
- RCU_GP_COUNT
);
299 * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
300 * meant to be read by RCU read-side critical sections. Returns the assigned
303 * Documents which pointers will be dereferenced by RCU read-side critical
304 * sections and adds the required memory barriers on architectures requiring
305 * them. It also makes sure the compiler does not reorder code initializing the
306 * data structure before its publication.
308 * Should match rcu_dereference_pointer().
311 #define _rcu_assign_pointer(p, v) \
313 if (!__builtin_constant_p(v) || \
316 STORE_SHARED(p, v); \
320 * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
321 * is as expected by "old". If succeeds, returns the previous pointer to the
322 * data structure, which can be safely freed after waiting for a quiescent state
323 * using synchronize_rcu(). If fails (unexpected value), returns old (which
324 * should not be freed !).
327 #define _rcu_cmpxchg_pointer(p, old, _new) \
329 if (!__builtin_constant_p(_new) || \
332 uatomic_cmpxchg(p, old, _new); \
336 * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
337 * pointer to the data structure, which can be safely freed after waiting for a
338 * quiescent state using synchronize_rcu().
341 #define _rcu_xchg_pointer(p, v) \
343 if (!__builtin_constant_p(v) || \
346 uatomic_xchg(p, v); \
350 * Exchanges the pointer and waits for quiescent state.
351 * The pointer returned can be freed.
353 #define _rcu_publish_content(p, v) \
356 oldptr = _rcu_xchg_pointer(p, v); \
361 #endif /* _URCU_STATIC_H */