Commit | Line | Data |
---|---|---|
adcfce54 MD |
1 | #ifndef _URCU_STATIC_H |
2 | #define _URCU_STATIC_H | |
3 | ||
4 | /* | |
5 | * urcu-static.h | |
6 | * | |
d2d23035 | 7 | * Userspace RCU header. |
adcfce54 | 8 | * |
d2d23035 MD |
9 | * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking |
10 | * dynamically with the userspace rcu library. | |
adcfce54 | 11 | * |
d2d23035 MD |
12 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> |
13 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. | |
adcfce54 | 14 | * |
d2d23035 MD |
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 | |
adcfce54 MD |
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 | * | |
809f4fde MD |
79 | * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative |
80 | * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the | |
81 | * data read before the pointer read by speculating the value of the pointer. | |
82 | * Correct ordering is ensured because the pointer is read as a volatile access. | |
83 | * This acts as a global side-effect operation, which forbids reordering of | |
015c702f MD |
84 | * dependent memory operations. Note that such concern about dependency-breaking |
85 | * optimizations will eventually be taken care of by the "memory_order_consume" | |
86 | * addition to forthcoming C++ standard. | |
809f4fde | 87 | * |
adcfce54 MD |
88 | * Should match rcu_assign_pointer() or rcu_xchg_pointer(). |
89 | */ | |
90 | ||
91 | #define _rcu_dereference(p) ({ \ | |
92 | typeof(p) _________p1 = LOAD_SHARED(p); \ | |
93 | smp_read_barrier_depends(); \ | |
94 | (_________p1); \ | |
95 | }) | |
96 | ||
97 | /* | |
98 | * This code section can only be included in LGPL 2.1 compatible source code. | |
99 | * See below for the function call wrappers which can be used in code meant to | |
100 | * be only linked with the Userspace RCU library. This comes with a small | |
101 | * performance degradation on the read-side due to the added function calls. | |
102 | * This is required to permit relinking with newer versions of the library. | |
103 | */ | |
104 | ||
105 | /* | |
106 | * The signal number used by the RCU library can be overridden with | |
107 | * -DSIGURCU= when compiling the library. | |
108 | */ | |
109 | #ifndef SIGURCU | |
110 | #define SIGURCU SIGUSR1 | |
111 | #endif | |
112 | ||
113 | /* | |
114 | * If a reader is really non-cooperative and refuses to commit its | |
115 | * urcu_active_readers count to memory (there is no barrier in the reader | |
116 | * per-se), kick it after a few loops waiting for it. | |
117 | */ | |
118 | #define KICK_READER_LOOPS 10000 | |
119 | ||
120 | #ifdef DEBUG_YIELD | |
121 | #include <sched.h> | |
122 | #include <time.h> | |
123 | #include <pthread.h> | |
124 | #include <unistd.h> | |
125 | ||
126 | #define YIELD_READ (1 << 0) | |
127 | #define YIELD_WRITE (1 << 1) | |
128 | ||
b4ce1526 MD |
129 | /* |
130 | * Updates without CONFIG_URCU_AVOID_SIGNALS are much slower. Account this in | |
131 | * the delay. | |
132 | */ | |
133 | #ifdef CONFIG_URCU_AVOID_SIGNALS | |
adcfce54 MD |
134 | /* maximum sleep delay, in us */ |
135 | #define MAX_SLEEP 50 | |
136 | #else | |
137 | #define MAX_SLEEP 30000 | |
138 | #endif | |
139 | ||
140 | extern unsigned int yield_active; | |
141 | extern unsigned int __thread rand_yield; | |
142 | ||
143 | static inline void debug_yield_read(void) | |
144 | { | |
145 | if (yield_active & YIELD_READ) | |
146 | if (rand_r(&rand_yield) & 0x1) | |
147 | usleep(rand_r(&rand_yield) % MAX_SLEEP); | |
148 | } | |
149 | ||
150 | static inline void debug_yield_write(void) | |
151 | { | |
152 | if (yield_active & YIELD_WRITE) | |
153 | if (rand_r(&rand_yield) & 0x1) | |
154 | usleep(rand_r(&rand_yield) % MAX_SLEEP); | |
155 | } | |
156 | ||
157 | static inline void debug_yield_init(void) | |
158 | { | |
159 | rand_yield = time(NULL) ^ pthread_self(); | |
160 | } | |
161 | #else | |
162 | static inline void debug_yield_read(void) | |
163 | { | |
164 | } | |
165 | ||
166 | static inline void debug_yield_write(void) | |
167 | { | |
168 | } | |
169 | ||
170 | static inline void debug_yield_init(void) | |
171 | { | |
172 | ||
173 | } | |
174 | #endif | |
175 | ||
b4ce1526 | 176 | #ifdef CONFIG_URCU_AVOID_SIGNALS |
adcfce54 MD |
177 | static inline void reader_barrier() |
178 | { | |
179 | smp_mb(); | |
180 | } | |
181 | #else | |
182 | static inline void reader_barrier() | |
183 | { | |
184 | barrier(); | |
185 | } | |
186 | #endif | |
187 | ||
188 | /* | |
189 | * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a | |
190 | * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. | |
191 | */ | |
192 | #define RCU_GP_COUNT (1UL << 0) | |
193 | /* Use the amount of bits equal to half of the architecture long size */ | |
194 | #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2)) | |
195 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1) | |
196 | ||
197 | /* | |
198 | * Global quiescent period counter with low-order bits unused. | |
199 | * Using a int rather than a char to eliminate false register dependencies | |
200 | * causing stalls on some architectures. | |
201 | */ | |
202 | extern long urcu_gp_ctr; | |
203 | ||
204 | extern long __thread urcu_active_readers; | |
205 | ||
206 | static inline int rcu_old_gp_ongoing(long *value) | |
207 | { | |
208 | long v; | |
209 | ||
210 | if (value == NULL) | |
211 | return 0; | |
212 | /* | |
213 | * Make sure both tests below are done on the same version of *value | |
214 | * to insure consistency. | |
215 | */ | |
216 | v = LOAD_SHARED(*value); | |
217 | return (v & RCU_GP_CTR_NEST_MASK) && | |
218 | ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT); | |
219 | } | |
220 | ||
221 | static inline void _rcu_read_lock(void) | |
222 | { | |
223 | long tmp; | |
224 | ||
225 | tmp = urcu_active_readers; | |
226 | /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */ | |
67ef1a2c | 227 | if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { |
adcfce54 | 228 | _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr)); |
67ef1a2c MD |
229 | /* |
230 | * Set active readers count for outermost nesting level before | |
231 | * accessing the pointer. See force_mb_all_threads(). | |
232 | */ | |
233 | reader_barrier(); | |
234 | } else { | |
adcfce54 | 235 | _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT); |
67ef1a2c | 236 | } |
adcfce54 MD |
237 | } |
238 | ||
239 | static inline void _rcu_read_unlock(void) | |
240 | { | |
241 | reader_barrier(); | |
242 | /* | |
243 | * Finish using rcu before decrementing the pointer. | |
244 | * See force_mb_all_threads(). | |
67ef1a2c MD |
245 | * Formally only needed for outermost nesting level, but leave barrier |
246 | * in place for nested unlocks to remove a branch from the common case | |
247 | * (no nesting). | |
adcfce54 MD |
248 | */ |
249 | _STORE_SHARED(urcu_active_readers, urcu_active_readers - RCU_GP_COUNT); | |
250 | } | |
251 | ||
252 | /** | |
253 | * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure | |
254 | * meant to be read by RCU read-side critical sections. Returns the assigned | |
255 | * value. | |
256 | * | |
257 | * Documents which pointers will be dereferenced by RCU read-side critical | |
258 | * sections and adds the required memory barriers on architectures requiring | |
259 | * them. It also makes sure the compiler does not reorder code initializing the | |
260 | * data structure before its publication. | |
261 | * | |
262 | * Should match rcu_dereference_pointer(). | |
263 | */ | |
264 | ||
265 | #define _rcu_assign_pointer(p, v) \ | |
266 | ({ \ | |
267 | if (!__builtin_constant_p(v) || \ | |
268 | ((v) != NULL)) \ | |
269 | wmb(); \ | |
270 | STORE_SHARED(p, v); \ | |
271 | }) | |
272 | ||
273 | /** | |
274 | * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous | |
67ef1a2c | 275 | * pointer to the data structure, which can be safely freed after waiting for a |
adcfce54 MD |
276 | * quiescent state using synchronize_rcu(). |
277 | */ | |
278 | ||
279 | #define _rcu_xchg_pointer(p, v) \ | |
280 | ({ \ | |
281 | if (!__builtin_constant_p(v) || \ | |
282 | ((v) != NULL)) \ | |
283 | wmb(); \ | |
284 | xchg(p, v); \ | |
285 | }) | |
286 | ||
287 | /* | |
288 | * Exchanges the pointer and waits for quiescent state. | |
289 | * The pointer returned can be freed. | |
290 | */ | |
291 | #define _rcu_publish_content(p, v) \ | |
292 | ({ \ | |
293 | void *oldptr; \ | |
294 | oldptr = _rcu_xchg_pointer(p, v); \ | |
295 | synchronize_rcu(); \ | |
296 | oldptr; \ | |
297 | }) | |
298 | ||
299 | #endif /* _URCU_STATIC_H */ |