Add QSBR RCU timing tests
[urcu.git] / urcu-qsbr.h
1 #ifndef _URCU_QSBR_H
2 #define _URCU_QSBR_H
3
4 /*
5 * urcu-static.h
6 *
7 * Userspace RCU header.
8 *
9 * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking
10 * dynamically with the userspace rcu library.
11 *
12 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
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
35 #include <compiler.h>
36 #include <arch.h>
37
38 /*
39 * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
40 */
41 #define _LOAD_SHARED(p) ACCESS_ONCE(p)
42
43 /*
44 * Load a data from shared memory, doing a cache flush if required.
45 */
46 #define LOAD_SHARED(p) \
47 ({ \
48 smp_rmc(); \
49 _LOAD_SHARED(p); \
50 })
51
52 /*
53 * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
54 */
55 #define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); })
56
57 /*
58 * Store v into x, where x is located in shared memory. Performs the required
59 * cache flush after writing. Returns v.
60 */
61 #define STORE_SHARED(x, v) \
62 ({ \
63 _STORE_SHARED(x, v); \
64 smp_wmc(); \
65 (v); \
66 })
67
68 /**
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.
72 *
73 * This ensures that the pointer copy is invariant thorough the whole critical
74 * section.
75 *
76 * Inserts memory barriers on architectures that require them (currently only
77 * Alpha) and documents which pointers are protected by RCU.
78 *
79 * Should match rcu_assign_pointer() or rcu_xchg_pointer().
80 */
81
82 #define _rcu_dereference(p) ({ \
83 typeof(p) _________p1 = LOAD_SHARED(p); \
84 smp_read_barrier_depends(); \
85 (_________p1); \
86 })
87
88 /*
89 * This code section can only be included in LGPL 2.1 compatible source code.
90 * See below for the function call wrappers which can be used in code meant to
91 * be only linked with the Userspace RCU library. This comes with a small
92 * performance degradation on the read-side due to the added function calls.
93 * This is required to permit relinking with newer versions of the library.
94 */
95
96 /*
97 * The signal number used by the RCU library can be overridden with
98 * -DSIGURCU= when compiling the library.
99 */
100 #ifndef SIGURCU
101 #define SIGURCU SIGUSR1
102 #endif
103
104 /*
105 * If a reader is really non-cooperative and refuses to commit its
106 * urcu_active_readers count to memory (there is no barrier in the reader
107 * per-se), kick it after a few loops waiting for it.
108 */
109 #define KICK_READER_LOOPS 10000
110
111 #ifdef DEBUG_YIELD
112 #include <sched.h>
113 #include <time.h>
114 #include <pthread.h>
115 #include <unistd.h>
116
117 #define YIELD_READ (1 << 0)
118 #define YIELD_WRITE (1 << 1)
119
120 /* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */
121 #ifdef DEBUG_FULL_MB
122 /* maximum sleep delay, in us */
123 #define MAX_SLEEP 50
124 #else
125 #define MAX_SLEEP 30000
126 #endif
127
128 extern unsigned int yield_active;
129 extern unsigned int __thread rand_yield;
130
131 static inline void debug_yield_read(void)
132 {
133 if (yield_active & YIELD_READ)
134 if (rand_r(&rand_yield) & 0x1)
135 usleep(rand_r(&rand_yield) % MAX_SLEEP);
136 }
137
138 static inline void debug_yield_write(void)
139 {
140 if (yield_active & YIELD_WRITE)
141 if (rand_r(&rand_yield) & 0x1)
142 usleep(rand_r(&rand_yield) % MAX_SLEEP);
143 }
144
145 static inline void debug_yield_init(void)
146 {
147 rand_yield = time(NULL) ^ pthread_self();
148 }
149 #else
150 static inline void debug_yield_read(void)
151 {
152 }
153
154 static inline void debug_yield_write(void)
155 {
156 }
157
158 static inline void debug_yield_init(void)
159 {
160
161 }
162 #endif
163
164 #ifdef DEBUG_FULL_MB
165 static inline void reader_barrier()
166 {
167 smp_mb();
168 }
169 #else
170 static inline void reader_barrier()
171 {
172 barrier();
173 }
174 #endif
175
176 /*
177 * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
178 * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
179 */
180 #define RCU_GP_COUNT (1UL << 0)
181 /* Use the amount of bits equal to half of the architecture long size */
182 #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2))
183 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
184
185 /*
186 * Global quiescent period counter with low-order bits unused.
187 * Using a int rather than a char to eliminate false register dependencies
188 * causing stalls on some architectures.
189 */
190 extern long urcu_gp_ctr;
191
192 extern long __thread urcu_active_readers;
193
194 static inline int rcu_old_gp_ongoing(long *value)
195 {
196 if (value == NULL)
197 return 0;
198
199 return LOAD_SHARED(*value) & 1;
200 }
201
202 static inline void _rcu_read_lock(void)
203 {
204 }
205
206 static inline void _rcu_read_unlock(void)
207 {
208 }
209
210 static inline void _rcu_quiescent_state(void)
211 {
212 smp_mb();
213 urcu_active_readers = ACCESS_ONCE(urcu_gp_ctr) + 1;
214 smp_mb();
215 }
216
217 static inline void _rcu_thread_offline(void)
218 {
219 smp_mb();
220 urcu_active_readers = ACCESS_ONCE(urcu_gp_ctr);
221 }
222
223 static inline void _rcu_thread_online(void)
224 {
225 urcu_active_readers = ACCESS_ONCE(urcu_gp_ctr) + 1;
226 smp_mb();
227 }
228
229 /**
230 * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
231 * meant to be read by RCU read-side critical sections. Returns the assigned
232 * value.
233 *
234 * Documents which pointers will be dereferenced by RCU read-side critical
235 * sections and adds the required memory barriers on architectures requiring
236 * them. It also makes sure the compiler does not reorder code initializing the
237 * data structure before its publication.
238 *
239 * Should match rcu_dereference_pointer().
240 */
241
242 #define _rcu_assign_pointer(p, v) \
243 ({ \
244 if (!__builtin_constant_p(v) || \
245 ((v) != NULL)) \
246 wmb(); \
247 STORE_SHARED(p, v); \
248 })
249
250 /**
251 * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
252 * pointer to the data structure, which can be safely freed after waiting for a
253 * quiescent state using synchronize_rcu().
254 */
255
256 #define _rcu_xchg_pointer(p, v) \
257 ({ \
258 if (!__builtin_constant_p(v) || \
259 ((v) != NULL)) \
260 wmb(); \
261 xchg(p, v); \
262 })
263
264 /*
265 * Exchanges the pointer and waits for quiescent state.
266 * The pointer returned can be freed.
267 */
268 #define _rcu_publish_content(p, v) \
269 ({ \
270 void *oldptr; \
271 oldptr = _rcu_xchg_pointer(p, v); \
272 synchronize_rcu(); \
273 oldptr; \
274 })
275
276 extern void synchronize_rcu(void);
277
278 /*
279 * Reader thread registration.
280 */
281 extern void rcu_register_thread(void);
282 extern void rcu_unregister_thread(void);
283
284 #endif /* _URCU_QSBR_H */
This page took 0.04875 seconds and 4 git commands to generate.