Disable sys_membarrier
[userspace-rcu.git] / urcu / static / urcu.h
1 #ifndef _URCU_STATIC_H
2 #define _URCU_STATIC_H
3
4 /*
5 * urcu-static.h
6 *
7 * Userspace RCU header.
8 *
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.
11 *
12 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
13 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
14 *
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.
19 *
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.
24 *
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
28 *
29 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
30 */
31
32 #include <stdlib.h>
33 #include <pthread.h>
34 #include <unistd.h>
35 #include <stdint.h>
36
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
45 #ifdef __cplusplus
46 extern "C" {
47 #endif
48
49 /* Default is RCU_MEMBARRIER */
50 #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL)
51 #define RCU_MEMBARRIER
52 #endif
53
54 /*
55 * RCU_MEMBARRIER is only possibly available on Linux. Fallback to RCU_MB
56 * otherwise.
57 */
58 #if !defined(__linux__) && defined(RCU_MEMBARRIER)
59 #undef RCU_MEMBARRIER
60 #define RCU_MB
61 #endif
62
63 /*
64 * The ABI of sys_membarrier changed after its original implementation.
65 * Disable it for now. Use RCU_MB flavor instead.
66 */
67 #ifdef RCU_MEMBARRIER
68 #undef RCU_MEMBARRIER
69 #define RCU_MB
70 #endif
71
72 /*
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.
78 */
79
80 /*
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.
84 */
85 #ifdef SIGURCU
86 #define SIGRCU SIGURCU
87 #endif
88
89 #ifndef SIGRCU
90 #define SIGRCU SIGUSR1
91 #endif
92
93 #ifdef DEBUG_RCU
94 #define rcu_assert(args...) assert(args)
95 #else
96 #define rcu_assert(args...)
97 #endif
98
99 #ifdef DEBUG_YIELD
100 #include <sched.h>
101 #include <time.h>
102 #include <pthread.h>
103 #include <unistd.h>
104
105 #define YIELD_READ (1 << 0)
106 #define YIELD_WRITE (1 << 1)
107
108 /*
109 * Updates with RCU_SIGNAL are much slower. Account this in the delay.
110 */
111 #ifdef RCU_SIGNAL
112 /* maximum sleep delay, in us */
113 #define MAX_SLEEP 30000
114 #else
115 #define MAX_SLEEP 50
116 #endif
117
118 extern unsigned int yield_active;
119 extern DECLARE_URCU_TLS(unsigned int, rand_yield);
120
121 static inline void debug_yield_read(void)
122 {
123 if (yield_active & YIELD_READ)
124 if (rand_r(&URCU_TLS(rand_yield)) & 0x1)
125 usleep(rand_r(&URCU_TLS(rand_yield)) % MAX_SLEEP);
126 }
127
128 static inline void debug_yield_write(void)
129 {
130 if (yield_active & YIELD_WRITE)
131 if (rand_r(&URCU_TLS(rand_yield)) & 0x1)
132 usleep(rand_r(&URCU_TLS(rand_yield)) % MAX_SLEEP);
133 }
134
135 static inline void debug_yield_init(void)
136 {
137 URCU_TLS(rand_yield) = time(NULL) ^ (unsigned long) pthread_self();
138 }
139 #else
140 static inline void debug_yield_read(void)
141 {
142 }
143
144 static inline void debug_yield_write(void)
145 {
146 }
147
148 static inline void debug_yield_init(void)
149 {
150
151 }
152 #endif
153
154 /*
155 * RCU memory barrier broadcast group. Currently, only broadcast to all process
156 * threads is supported (group 0).
157 *
158 * Slave barriers are only guaranteed to be ordered wrt master barriers.
159 *
160 * The pair ordering is detailed as (O: ordered, X: not ordered) :
161 * slave master
162 * slave X O
163 * master O O
164 */
165
166 #define MB_GROUP_ALL 0
167 #define RCU_MB_GROUP MB_GROUP_ALL
168
169 #ifdef RCU_MEMBARRIER
170 extern int has_sys_membarrier;
171
172 static inline void smp_mb_slave(int group)
173 {
174 if (caa_likely(has_sys_membarrier))
175 cmm_barrier();
176 else
177 cmm_smp_mb();
178 }
179 #endif
180
181 #ifdef RCU_MB
182 static inline void smp_mb_slave(int group)
183 {
184 cmm_smp_mb();
185 }
186 #endif
187
188 #ifdef RCU_SIGNAL
189 static inline void smp_mb_slave(int group)
190 {
191 cmm_barrier();
192 }
193 #endif
194
195 /*
196 * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use
197 * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
198 */
199 #define RCU_GP_COUNT (1UL << 0)
200 /* Use the amount of bits equal to half of the architecture long size */
201 #define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2))
202 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1)
203
204 /*
205 * Global quiescent period counter with low-order bits unused.
206 * Using a int rather than a char to eliminate false register dependencies
207 * causing stalls on some architectures.
208 */
209 extern unsigned long rcu_gp_ctr;
210
211 struct rcu_reader {
212 /* Data used by both reader and synchronize_rcu() */
213 unsigned long ctr;
214 char need_mb;
215 /* Data used for registry */
216 struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
217 pthread_t tid;
218 };
219
220 extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader);
221
222 extern int32_t gp_futex;
223
224 /*
225 * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
226 */
227 static inline void wake_up_gp(void)
228 {
229 if (caa_unlikely(uatomic_read(&gp_futex) == -1)) {
230 uatomic_set(&gp_futex, 0);
231 /*
232 * Ignoring return value until we can make this function
233 * return something (because urcu_die() is not publicly
234 * exposed).
235 */
236 (void) futex_async(&gp_futex, FUTEX_WAKE, 1,
237 NULL, NULL, 0);
238 }
239 }
240
241 static inline int rcu_gp_ongoing(unsigned long *ctr)
242 {
243 unsigned long v;
244
245 /*
246 * Make sure both tests below are done on the same version of *value
247 * to insure consistency.
248 */
249 v = CMM_LOAD_SHARED(*ctr);
250 return (v & RCU_GP_CTR_NEST_MASK) &&
251 ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE);
252 }
253
254 /*
255 * Helper for _rcu_read_lock(). The format of rcu_gp_ctr (as well as
256 * the per-thread rcu_reader.ctr) has the upper bits containing a count of
257 * _rcu_read_lock() nesting, and a lower-order bit that contains either zero
258 * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in
259 * _rcu_read_lock() happen before the subsequent read-side critical section.
260 */
261 static inline void _rcu_read_lock_update(unsigned long tmp)
262 {
263 if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
264 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
265 smp_mb_slave(RCU_MB_GROUP);
266 } else
267 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT);
268 }
269
270 /*
271 * Enter an RCU read-side critical section.
272 *
273 * The first cmm_barrier() call ensures that the compiler does not reorder
274 * the body of _rcu_read_lock() with a mutex.
275 *
276 * This function and its helper are both less than 10 lines long. The
277 * intent is that this function meets the 10-line criterion in LGPL,
278 * allowing this function to be invoked directly from non-LGPL code.
279 */
280 static inline void _rcu_read_lock(void)
281 {
282 unsigned long tmp;
283
284 cmm_barrier();
285 tmp = URCU_TLS(rcu_reader).ctr;
286 _rcu_read_lock_update(tmp);
287 }
288
289 /*
290 * This is a helper function for _rcu_read_unlock().
291 *
292 * The first smp_mb_slave() call ensures that the critical section is
293 * seen to precede the store to rcu_reader.ctr.
294 * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr
295 * before reading the update-side futex.
296 */
297 static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp)
298 {
299 if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
300 smp_mb_slave(RCU_MB_GROUP);
301 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
302 smp_mb_slave(RCU_MB_GROUP);
303 wake_up_gp();
304 } else
305 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT);
306 }
307
308 /*
309 * Exit an RCU read-side crtical section. Both this function and its
310 * helper are smaller than 10 lines of code, and are intended to be
311 * usable by non-LGPL code, as called out in LGPL.
312 */
313 static inline void _rcu_read_unlock(void)
314 {
315 unsigned long tmp;
316
317 tmp = URCU_TLS(rcu_reader).ctr;
318 _rcu_read_unlock_update_and_wakeup(tmp);
319 cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */
320 }
321
322 #ifdef __cplusplus
323 }
324 #endif
325
326 #endif /* _URCU_STATIC_H */
This page took 0.03578 seconds and 4 git commands to generate.