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>
49 /* Default is RCU_MEMBARRIER */
50 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
51 #define RCU_MEMBARRIER
55 * RCU_MEMBARRIER is only possibly available on Linux. Fallback to RCU_MB
58 #if !defined(__linux__) && defined(RCU_MEMBARRIER)
66 /* If the headers do not support SYS_membarrier, statically use RCU_MB */
68 # define MEMBARRIER_EXPEDITED (1 << 0)
69 # define MEMBARRIER_DELAYED (1 << 1)
70 # define MEMBARRIER_QUERY (1 << 16)
71 # define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__)
73 # undef RCU_MEMBARRIER
79 * This code section can only be included in LGPL 2.1 compatible source code.
80 * See below for the function call wrappers which can be used in code meant to
81 * be only linked with the Userspace RCU library. This comes with a small
82 * performance degradation on the read-side due to the added function calls.
83 * This is required to permit relinking with newer versions of the library.
87 * The signal number used by the RCU library can be overridden with
88 * -DSIGRCU= when compiling the library.
89 * Provide backward compatibility for liburcu 0.3.x SIGURCU.
92 #define SIGRCU SIGURCU
96 #define SIGRCU SIGUSR1
100 RCU_READER_ACTIVE_CURRENT
,
101 RCU_READER_ACTIVE_OLD
,
106 #define rcu_assert(args...) assert(args)
108 #define rcu_assert(args...)
117 #define RCU_YIELD_READ (1 << 0)
118 #define RCU_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 rcu_yield_active
;
131 extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield
);
133 static inline void rcu_debug_yield_read(void)
135 if (rcu_yield_active
& RCU_YIELD_READ
)
136 if (rand_r(&URCU_TLS(rcu_rand_yield
)) & 0x1)
137 usleep(rand_r(&URCU_TLS(rcu_rand_yield
)) % MAX_SLEEP
);
140 static inline void rcu_debug_yield_write(void)
142 if (rcu_yield_active
& RCU_YIELD_WRITE
)
143 if (rand_r(&URCU_TLS(rcu_rand_yield
)) & 0x1)
144 usleep(rand_r(&URCU_TLS(rcu_rand_yield
)) % MAX_SLEEP
);
147 static inline void rcu_debug_yield_init(void)
149 URCU_TLS(rcu_rand_yield
) = time(NULL
) ^ (unsigned long) pthread_self();
152 static inline void rcu_debug_yield_read(void)
156 static inline void rcu_debug_yield_write(void)
160 static inline void rcu_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 rcu_has_sys_membarrier
;
184 static inline void smp_mb_slave(int group
)
186 if (caa_likely(rcu_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)
218 * Global grace period counter.
219 * Contains the current RCU_GP_CTR_PHASE.
220 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
221 * Written to only by writer with mutex taken.
222 * Read by both writer and readers.
227 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
229 extern struct rcu_gp rcu_gp
;
232 /* Data used by both reader and synchronize_rcu() */
235 /* Data used for registry */
236 struct cds_list_head node
__attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
240 extern DECLARE_URCU_TLS(struct rcu_reader
, rcu_reader
);
243 * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
245 static inline void wake_up_gp(void)
247 if (caa_unlikely(uatomic_read(&rcu_gp
.futex
) == -1)) {
248 uatomic_set(&rcu_gp
.futex
, 0);
249 futex_async(&rcu_gp
.futex
, FUTEX_WAKE
, 1,
254 static inline enum rcu_state
rcu_reader_state(unsigned long *ctr
)
259 * Make sure both tests below are done on the same version of *value
260 * to insure consistency.
262 v
= CMM_LOAD_SHARED(*ctr
);
263 if (!(v
& RCU_GP_CTR_NEST_MASK
))
264 return RCU_READER_INACTIVE
;
265 if (!((v
^ rcu_gp
.ctr
) & RCU_GP_CTR_PHASE
))
266 return RCU_READER_ACTIVE_CURRENT
;
267 return RCU_READER_ACTIVE_OLD
;
271 * Helper for _rcu_read_lock(). The format of rcu_gp.ctr (as well as
272 * the per-thread rcu_reader.ctr) has the upper bits containing a count of
273 * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
274 * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in
275 * _rcu_read_lock() happen before the subsequent read-side critical section.
277 static inline void _rcu_read_lock_update(unsigned long tmp
)
279 if (caa_likely(!(tmp
& RCU_GP_CTR_NEST_MASK
))) {
280 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, _CMM_LOAD_SHARED(rcu_gp
.ctr
));
281 smp_mb_slave(RCU_MB_GROUP
);
283 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, tmp
+ RCU_GP_COUNT
);
287 * Enter an RCU read-side critical section.
289 * The first cmm_barrier() call ensures that the compiler does not reorder
290 * the body of _rcu_read_lock() with a mutex.
292 * This function and its helper are both less than 10 lines long. The
293 * intent is that this function meets the 10-line criterion in LGPL,
294 * allowing this function to be invoked directly from non-LGPL code.
296 static inline void _rcu_read_lock(void)
301 tmp
= URCU_TLS(rcu_reader
).ctr
;
302 _rcu_read_lock_update(tmp
);
306 * This is a helper function for _rcu_read_unlock().
308 * The first smp_mb_slave() call ensures that the critical section is
309 * seen to precede the store to rcu_reader.ctr.
310 * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr
311 * before reading the update-side futex.
313 static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp
)
315 if (caa_likely((tmp
& RCU_GP_CTR_NEST_MASK
) == RCU_GP_COUNT
)) {
316 smp_mb_slave(RCU_MB_GROUP
);
317 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, URCU_TLS(rcu_reader
).ctr
- RCU_GP_COUNT
);
318 smp_mb_slave(RCU_MB_GROUP
);
321 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).ctr
, URCU_TLS(rcu_reader
).ctr
- RCU_GP_COUNT
);
325 * Exit an RCU read-side crtical section. Both this function and its
326 * helper are smaller than 10 lines of code, and are intended to be
327 * usable by non-LGPL code, as called out in LGPL.
329 static inline void _rcu_read_unlock(void)
333 tmp
= URCU_TLS(rcu_reader
).ctr
;
334 _rcu_read_unlock_update_and_wakeup(tmp
);
335 cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
339 * Returns whether within a RCU read-side critical section.
341 * This function is less than 10 lines long. The intent is that this
342 * function meets the 10-line criterion for LGPL, allowing this function
343 * to be invoked directly from non-LGPL code.
345 static inline int _rcu_read_ongoing(void)
347 return URCU_TLS(rcu_reader
).ctr
& RCU_GP_CTR_NEST_MASK
;
354 #endif /* _URCU_STATIC_H */