qsbr: only mark reader thread as being waited for in contended case
[urcu.git] / urcu-qsbr-static.h
1 #ifndef _URCU_QSBR_STATIC_H
2 #define _URCU_QSBR_STATIC_H
3
4 /*
5 * urcu-qsbr-static.h
6 *
7 * Userspace RCU QSBR header.
8 *
9 * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu-qsbr.h for linking
10 * dynamically with the userspace rcu QSBR 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 #include <assert.h>
35 #include <limits.h>
36 #include <sched.h>
37
38 #include <compiler.h>
39 #include <arch.h>
40
41 /*
42 * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
43 */
44 #define _LOAD_SHARED(p) ACCESS_ONCE(p)
45
46 /*
47 * Load a data from shared memory, doing a cache flush if required.
48 */
49 #define LOAD_SHARED(p) \
50 ({ \
51 smp_rmc(); \
52 _LOAD_SHARED(p); \
53 })
54
55 /*
56 * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
57 */
58 #define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); })
59
60 /*
61 * Store v into x, where x is located in shared memory. Performs the required
62 * cache flush after writing. Returns v.
63 */
64 #define STORE_SHARED(x, v) \
65 ({ \
66 _STORE_SHARED(x, v); \
67 smp_wmc(); \
68 (v); \
69 })
70
71 /**
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.
75 *
76 * This ensures that the pointer copy is invariant thorough the whole critical
77 * section.
78 *
79 * Inserts memory barriers on architectures that require them (currently only
80 * Alpha) and documents which pointers are protected by RCU.
81 *
82 * Should match rcu_assign_pointer() or rcu_xchg_pointer().
83 */
84
85 #define _rcu_dereference(p) ({ \
86 typeof(p) _________p1 = LOAD_SHARED(p); \
87 smp_read_barrier_depends(); \
88 (_________p1); \
89 })
90
91 /*
92 * This code section can only be included in LGPL 2.1 compatible source code.
93 * See below for the function call wrappers which can be used in code meant to
94 * be only linked with the Userspace RCU library. This comes with a small
95 * performance degradation on the read-side due to the added function calls.
96 * This is required to permit relinking with newer versions of the library.
97 */
98
99 /*
100 * If a reader is really non-cooperative and refuses to commit its
101 * rcu_reader qs_gp count to memory (there is no barrier in the reader
102 * per-se), kick it after a few loops waiting for it.
103 */
104 #define KICK_READER_LOOPS 10000
105
106 /*
107 * Active attempts to check for reader Q.S. before calling sched_yield().
108 */
109 #define RCU_QS_ACTIVE_ATTEMPTS 100
110
111 #ifdef DEBUG_RCU
112 #define rcu_assert(args...) assert(args)
113 #else
114 #define rcu_assert(args...)
115 #endif
116
117 #ifdef DEBUG_YIELD
118 #include <sched.h>
119 #include <time.h>
120 #include <pthread.h>
121 #include <unistd.h>
122
123 #define YIELD_READ (1 << 0)
124 #define YIELD_WRITE (1 << 1)
125
126 /* maximum sleep delay, in us */
127 #define MAX_SLEEP 50
128
129 extern unsigned int yield_active;
130 extern unsigned int __thread rand_yield;
131
132 static inline void debug_yield_read(void)
133 {
134 if (yield_active & YIELD_READ)
135 if (rand_r(&rand_yield) & 0x1)
136 usleep(rand_r(&rand_yield) % MAX_SLEEP);
137 }
138
139 static inline void debug_yield_write(void)
140 {
141 if (yield_active & YIELD_WRITE)
142 if (rand_r(&rand_yield) & 0x1)
143 usleep(rand_r(&rand_yield) % MAX_SLEEP);
144 }
145
146 static inline void debug_yield_init(void)
147 {
148 rand_yield = time(NULL) ^ pthread_self();
149 }
150 #else
151 static inline void debug_yield_read(void)
152 {
153 }
154
155 static inline void debug_yield_write(void)
156 {
157 }
158
159 static inline void debug_yield_init(void)
160 {
161
162 }
163 #endif
164
165 static inline void reader_barrier()
166 {
167 smp_mb();
168 }
169
170 #define RCU_GP_ONLINE (1UL << 0)
171 #define RCU_GP_ONGOING (1UL << 1)
172 #define RCU_GP_CTR (1UL << 2)
173
174 /*
175 * Global quiescent period counter with low-order bits unused.
176 * Using a int rather than a char to eliminate false register dependencies
177 * causing stalls on some architectures.
178 */
179 extern unsigned long urcu_gp_ctr;
180
181 struct urcu_reader_status {
182 unsigned long qs_gp;
183 unsigned long gp_waiting;
184 };
185
186 extern struct urcu_reader_status __thread urcu_reader_status;
187
188 #if (BITS_PER_LONG < 64)
189 static inline int rcu_gp_ongoing(unsigned long *value)
190 {
191 unsigned long reader_gp;
192
193 if (value == NULL)
194 return 0;
195 reader_gp = LOAD_SHARED(*value);
196 return reader_gp && ((reader_gp ^ urcu_gp_ctr) & RCU_GP_CTR);
197 }
198 #else /* !(BITS_PER_LONG < 64) */
199 static inline int rcu_gp_ongoing(unsigned long *value)
200 {
201 unsigned long reader_gp;
202
203 if (value == NULL)
204 return 0;
205 reader_gp = LOAD_SHARED(*value);
206 return reader_gp && (reader_gp - urcu_gp_ctr > ULONG_MAX / 2);
207 }
208 #endif /* !(BITS_PER_LONG < 64) */
209
210 static inline void _rcu_read_lock(void)
211 {
212 rcu_assert(urcu_reader_status.qs_gp);
213 }
214
215 static inline void _rcu_read_unlock(void)
216 {
217 }
218
219 static inline void _rcu_quiescent_state(void)
220 {
221 long gp_ctr;
222
223 smp_mb();
224 /*
225 * volatile accesses can be reordered by the compiler when put in the
226 * same expression.
227 */
228 if (unlikely((gp_ctr = LOAD_SHARED(urcu_gp_ctr)) & RCU_GP_ONGOING) &&
229 unlikely(urcu_reader_status.gp_waiting)) {
230 _STORE_SHARED(urcu_reader_status.qs_gp, gp_ctr);
231 sched_yield();
232 } else {
233 _STORE_SHARED(urcu_reader_status.qs_gp, gp_ctr);
234 }
235 smp_mb();
236 }
237
238 static inline void _rcu_thread_offline(void)
239 {
240 smp_mb();
241 STORE_SHARED(urcu_reader_status.qs_gp, 0);
242 if (unlikely(LOAD_SHARED(urcu_gp_ctr) & RCU_GP_ONGOING) &&
243 unlikely(urcu_reader_status.gp_waiting)) {
244 sched_yield();
245 }
246 }
247
248 static inline void _rcu_thread_online(void)
249 {
250 _STORE_SHARED(urcu_reader_status.qs_gp, LOAD_SHARED(urcu_gp_ctr));
251 smp_mb();
252 }
253
254 /**
255 * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
256 * meant to be read by RCU read-side critical sections. Returns the assigned
257 * value.
258 *
259 * Documents which pointers will be dereferenced by RCU read-side critical
260 * sections and adds the required memory barriers on architectures requiring
261 * them. It also makes sure the compiler does not reorder code initializing the
262 * data structure before its publication.
263 *
264 * Should match rcu_dereference_pointer().
265 */
266
267 #define _rcu_assign_pointer(p, v) \
268 ({ \
269 if (!__builtin_constant_p(v) || \
270 ((v) != NULL)) \
271 wmb(); \
272 STORE_SHARED(p, v); \
273 })
274
275 /**
276 * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
277 * is as expected by "old". If succeeds, returns the previous pointer to the
278 * data structure, which can be safely freed after waiting for a quiescent state
279 * using synchronize_rcu(). If fails (unexpected value), returns old (which
280 * should not be freed !).
281 */
282
283 #define _rcu_cmpxchg_pointer(p, old, _new) \
284 ({ \
285 if (!__builtin_constant_p(_new) || \
286 ((_new) != NULL)) \
287 wmb(); \
288 cmpxchg(p, old, _new); \
289 })
290
291 /**
292 * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
293 * pointer to the data structure, which can be safely freed after waiting for a
294 * quiescent state using synchronize_rcu().
295 */
296
297 #define _rcu_xchg_pointer(p, v) \
298 ({ \
299 if (!__builtin_constant_p(v) || \
300 ((v) != NULL)) \
301 wmb(); \
302 xchg(p, v); \
303 })
304
305 /*
306 * Exchanges the pointer and waits for quiescent state.
307 * The pointer returned can be freed.
308 */
309 #define _rcu_publish_content(p, v) \
310 ({ \
311 void *oldptr; \
312 oldptr = _rcu_xchg_pointer(p, v); \
313 synchronize_rcu(); \
314 oldptr; \
315 })
316
317 #endif /* _URCU_QSBR_STATIC_H */
This page took 0.0382 seconds and 4 git commands to generate.