7 * Userspace RCU header.
9 * TO BE INCLUDED ONLY IN CODE THAT IS TO BE RECOMPILED ON EACH LIBURCU
10 * RELEASE. See urcu.h for linking 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>
43 #include <urcu/tls-compat.h>
44 #include <urcu/rand-compat.h>
50 /* Default is RCU_MEMBARRIER */
51 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
52 #define RCU_MEMBARRIER
56 * This code section can only be included in LGPL 2.1 compatible source code.
57 * See below for the function call wrappers which can be used in code meant to
58 * be only linked with the Userspace RCU library. This comes with a small
59 * performance degradation on the read-side due to the added function calls.
60 * This is required to permit relinking with newer versions of the library.
64 * The signal number used by the RCU library can be overridden with
65 * -DSIGRCU= when compiling the library.
66 * Provide backward compatibility for liburcu 0.3.x SIGURCU.
69 #define SIGRCU SIGURCU
73 #define SIGRCU SIGUSR1
77 RCU_READER_ACTIVE_CURRENT
,
78 RCU_READER_ACTIVE_OLD
,
83 #define rcu_assert(args...) assert(args)
85 #define rcu_assert(args...)
94 #define RCU_YIELD_READ (1 << 0)
95 #define RCU_YIELD_WRITE (1 << 1)
98 * Updates with RCU_SIGNAL are much slower. Account this in the delay.
101 /* maximum sleep delay, in us */
102 #define MAX_SLEEP 30000
107 extern unsigned int rcu_yield_active
;
108 extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield
);
110 static inline void rcu_debug_yield_read(void)
112 if (rcu_yield_active
& RCU_YIELD_READ
)
113 if (rand_r(&URCU_TLS(rcu_rand_yield
)) & 0x1)
114 usleep(rand_r(&URCU_TLS(rcu_rand_yield
)) % MAX_SLEEP
);
117 static inline void rcu_debug_yield_write(void)
119 if (rcu_yield_active
& RCU_YIELD_WRITE
)
120 if (rand_r(&URCU_TLS(rcu_rand_yield
)) & 0x1)
121 usleep(rand_r(&URCU_TLS(rcu_rand_yield
)) % MAX_SLEEP
);
124 static inline void rcu_debug_yield_init(void)
126 URCU_TLS(rcu_rand_yield
) = time(NULL
) ^ (unsigned long) pthread_self();
129 static inline void rcu_debug_yield_read(void)
133 static inline void rcu_debug_yield_write(void)
137 static inline void rcu_debug_yield_init(void)
144 * RCU memory barrier broadcast group. Currently, only broadcast to all process
145 * threads is supported (group 0).
147 * Slave barriers are only guaranteed to be ordered wrt master barriers.
149 * The pair ordering is detailed as (O: ordered, X: not ordered) :
155 #define MB_GROUP_ALL 0
156 #define RCU_MB_GROUP MB_GROUP_ALL
158 #ifdef RCU_MEMBARRIER
159 extern int rcu_has_sys_membarrier
;
161 static inline void smp_mb_slave(int group
)
163 if (caa_likely(rcu_has_sys_membarrier
))
171 static inline void smp_mb_slave(int group
)
178 static inline void smp_mb_slave(int group
)
185 * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
186 * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
188 #define RCU_GP_COUNT (1UL << 0)
189 /* Use the amount of bits equal to half of the architecture long size */
190 #define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2))
191 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1)
195 * Global grace period counter.
196 * Contains the current RCU_GP_CTR_PHASE.
197 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
198 * Written to only by writer with mutex taken.
199 * Read by both writer and readers.
204 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
206 extern struct rcu_gp rcu_gp
;
209 /* Data used by both reader and synchronize_rcu() */
212 /* Data used for registry */
213 struct cds_list_head node
__attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
217 extern DECLARE_URCU_TLS(struct rcu_reader
, rcu_reader
);
220 * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
222 static inline void wake_up_gp(void)
224 if (caa_unlikely(uatomic_read(&rcu_gp
.futex
) == -1)) {
225 uatomic_set(&rcu_gp
.futex
, 0);
226 futex_async(&rcu_gp
.futex
, FUTEX_WAKE
, 1,
231 static inline enum rcu_state
rcu_reader_state(unsigned long *ctr
)
236 * Make sure both tests below are done on the same version of *value
237 * to insure consistency.
239 v
= CMM_LOAD_SHARED(*ctr
);
240 if (!(v
& RCU_GP_CTR_NEST_MASK
))
241 return RCU_READER_INACTIVE
;
242 if (!((v
^ rcu_gp
.ctr
) & RCU_GP_CTR_PHASE
))
243 return RCU_READER_ACTIVE_CURRENT
;
244 return RCU_READER_ACTIVE_OLD
;
248 * Helper for _rcu_read_lock(). The format of rcu_gp.ctr (as well as
249 * the per-thread rcu_reader.ctr) has the upper bits containing a count of
250 * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
251 * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in
252 * _rcu_read_lock() happen before the subsequent read-side critical section.
254 static inline void _rcu_read_lock_update(unsigned long tmp
)
256 if (caa_likely(!(tmp
& RCU_GP_CTR_NEST_MASK
))) {
257 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, _CMM_LOAD_SHARED(rcu_gp
.ctr
));
258 smp_mb_slave(RCU_MB_GROUP
);
260 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, tmp
+ RCU_GP_COUNT
);
264 * Enter an RCU read-side critical section.
266 * The first cmm_barrier() call ensures that the compiler does not reorder
267 * the body of _rcu_read_lock() with a mutex.
269 * This function and its helper are both less than 10 lines long. The
270 * intent is that this function meets the 10-line criterion in LGPL,
271 * allowing this function to be invoked directly from non-LGPL code.
273 static inline void _rcu_read_lock(void)
278 tmp
= URCU_TLS(rcu_reader
).ctr
;
279 _rcu_read_lock_update(tmp
);
283 * This is a helper function for _rcu_read_unlock().
285 * The first smp_mb_slave() call ensures that the critical section is
286 * seen to precede the store to rcu_reader.ctr.
287 * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr
288 * before reading the update-side futex.
290 static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp
)
292 if (caa_likely((tmp
& RCU_GP_CTR_NEST_MASK
) == RCU_GP_COUNT
)) {
293 smp_mb_slave(RCU_MB_GROUP
);
294 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, URCU_TLS(rcu_reader
).ctr
- RCU_GP_COUNT
);
295 smp_mb_slave(RCU_MB_GROUP
);
298 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, URCU_TLS(rcu_reader
).ctr
- RCU_GP_COUNT
);
302 * Exit an RCU read-side crtical section. Both this function and its
303 * helper are smaller than 10 lines of code, and are intended to be
304 * usable by non-LGPL code, as called out in LGPL.
306 static inline void _rcu_read_unlock(void)
310 tmp
= URCU_TLS(rcu_reader
).ctr
;
311 _rcu_read_unlock_update_and_wakeup(tmp
);
312 cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
316 * Returns whether within a RCU read-side critical section.
318 * This function is less than 10 lines long. The intent is that this
319 * function meets the 10-line criterion for LGPL, allowing this function
320 * to be invoked directly from non-LGPL code.
322 static inline int _rcu_read_ongoing(void)
324 return URCU_TLS(rcu_reader
).ctr
& RCU_GP_CTR_NEST_MASK
;
331 #endif /* _URCU_STATIC_H */