Commit | Line | Data |
---|---|---|
adcfce54 MD |
1 | #ifndef _URCU_STATIC_H |
2 | #define _URCU_STATIC_H | |
3 | ||
4 | /* | |
5 | * urcu-static.h | |
6 | * | |
7 | * Userspace RCU header, to be included only in LGPL-compatible code. | |
8 | * | |
9 | * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> | |
10 | * | |
11 | * Credits for Paul E. McKenney <paulmck@linux.vnet.ibm.com> | |
12 | * for inspiration coming from the Linux kernel RCU and rcu-preempt. | |
13 | * | |
14 | * Distributed under LGPLv2.1 | |
15 | * | |
16 | * IBM's contributions to this file may be relicensed under LGPLv2 or later. | |
17 | */ | |
18 | ||
19 | #include <stdlib.h> | |
20 | #include <pthread.h> | |
21 | ||
22 | #include <compiler.h> | |
23 | #include <arch.h> | |
24 | ||
25 | /* | |
26 | * Identify a shared load. A smp_rmc() or smp_mc() should come before the load. | |
27 | */ | |
28 | #define _LOAD_SHARED(p) ACCESS_ONCE(p) | |
29 | ||
30 | /* | |
31 | * Load a data from shared memory, doing a cache flush if required. | |
32 | */ | |
33 | #define LOAD_SHARED(p) \ | |
34 | ({ \ | |
35 | smp_rmc(); \ | |
36 | _LOAD_SHARED(p); \ | |
37 | }) | |
38 | ||
39 | /* | |
40 | * Identify a shared store. A smp_wmc() or smp_mc() should follow the store. | |
41 | */ | |
42 | #define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); }) | |
43 | ||
44 | /* | |
45 | * Store v into x, where x is located in shared memory. Performs the required | |
46 | * cache flush after writing. Returns v. | |
47 | */ | |
48 | #define STORE_SHARED(x, v) \ | |
49 | ({ \ | |
50 | _STORE_SHARED(x, v); \ | |
51 | smp_wmc(); \ | |
52 | (v); \ | |
53 | }) | |
54 | ||
55 | /** | |
56 | * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable | |
57 | * into a RCU read-side critical section. The pointer can later be safely | |
58 | * dereferenced within the critical section. | |
59 | * | |
60 | * This ensures that the pointer copy is invariant thorough the whole critical | |
61 | * section. | |
62 | * | |
63 | * Inserts memory barriers on architectures that require them (currently only | |
64 | * Alpha) and documents which pointers are protected by RCU. | |
65 | * | |
66 | * Should match rcu_assign_pointer() or rcu_xchg_pointer(). | |
67 | */ | |
68 | ||
69 | #define _rcu_dereference(p) ({ \ | |
70 | typeof(p) _________p1 = LOAD_SHARED(p); \ | |
71 | smp_read_barrier_depends(); \ | |
72 | (_________p1); \ | |
73 | }) | |
74 | ||
75 | /* | |
76 | * This code section can only be included in LGPL 2.1 compatible source code. | |
77 | * See below for the function call wrappers which can be used in code meant to | |
78 | * be only linked with the Userspace RCU library. This comes with a small | |
79 | * performance degradation on the read-side due to the added function calls. | |
80 | * This is required to permit relinking with newer versions of the library. | |
81 | */ | |
82 | ||
83 | /* | |
84 | * The signal number used by the RCU library can be overridden with | |
85 | * -DSIGURCU= when compiling the library. | |
86 | */ | |
87 | #ifndef SIGURCU | |
88 | #define SIGURCU SIGUSR1 | |
89 | #endif | |
90 | ||
91 | /* | |
92 | * If a reader is really non-cooperative and refuses to commit its | |
93 | * urcu_active_readers count to memory (there is no barrier in the reader | |
94 | * per-se), kick it after a few loops waiting for it. | |
95 | */ | |
96 | #define KICK_READER_LOOPS 10000 | |
97 | ||
98 | #ifdef DEBUG_YIELD | |
99 | #include <sched.h> | |
100 | #include <time.h> | |
101 | #include <pthread.h> | |
102 | #include <unistd.h> | |
103 | ||
104 | #define YIELD_READ (1 << 0) | |
105 | #define YIELD_WRITE (1 << 1) | |
106 | ||
107 | /* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */ | |
108 | #ifdef DEBUG_FULL_MB | |
109 | /* maximum sleep delay, in us */ | |
110 | #define MAX_SLEEP 50 | |
111 | #else | |
112 | #define MAX_SLEEP 30000 | |
113 | #endif | |
114 | ||
115 | extern unsigned int yield_active; | |
116 | extern unsigned int __thread rand_yield; | |
117 | ||
118 | static inline void debug_yield_read(void) | |
119 | { | |
120 | if (yield_active & YIELD_READ) | |
121 | if (rand_r(&rand_yield) & 0x1) | |
122 | usleep(rand_r(&rand_yield) % MAX_SLEEP); | |
123 | } | |
124 | ||
125 | static inline void debug_yield_write(void) | |
126 | { | |
127 | if (yield_active & YIELD_WRITE) | |
128 | if (rand_r(&rand_yield) & 0x1) | |
129 | usleep(rand_r(&rand_yield) % MAX_SLEEP); | |
130 | } | |
131 | ||
132 | static inline void debug_yield_init(void) | |
133 | { | |
134 | rand_yield = time(NULL) ^ pthread_self(); | |
135 | } | |
136 | #else | |
137 | static inline void debug_yield_read(void) | |
138 | { | |
139 | } | |
140 | ||
141 | static inline void debug_yield_write(void) | |
142 | { | |
143 | } | |
144 | ||
145 | static inline void debug_yield_init(void) | |
146 | { | |
147 | ||
148 | } | |
149 | #endif | |
150 | ||
151 | #ifdef DEBUG_FULL_MB | |
152 | static inline void reader_barrier() | |
153 | { | |
154 | smp_mb(); | |
155 | } | |
156 | #else | |
157 | static inline void reader_barrier() | |
158 | { | |
159 | barrier(); | |
160 | } | |
161 | #endif | |
162 | ||
163 | /* | |
164 | * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a | |
165 | * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. | |
166 | */ | |
167 | #define RCU_GP_COUNT (1UL << 0) | |
168 | /* Use the amount of bits equal to half of the architecture long size */ | |
169 | #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2)) | |
170 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1) | |
171 | ||
172 | /* | |
173 | * Global quiescent period counter with low-order bits unused. | |
174 | * Using a int rather than a char to eliminate false register dependencies | |
175 | * causing stalls on some architectures. | |
176 | */ | |
177 | extern long urcu_gp_ctr; | |
178 | ||
179 | extern long __thread urcu_active_readers; | |
180 | ||
181 | static inline int rcu_old_gp_ongoing(long *value) | |
182 | { | |
183 | long v; | |
184 | ||
185 | if (value == NULL) | |
186 | return 0; | |
187 | /* | |
188 | * Make sure both tests below are done on the same version of *value | |
189 | * to insure consistency. | |
190 | */ | |
191 | v = LOAD_SHARED(*value); | |
192 | return (v & RCU_GP_CTR_NEST_MASK) && | |
193 | ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT); | |
194 | } | |
195 | ||
196 | static inline void _rcu_read_lock(void) | |
197 | { | |
198 | long tmp; | |
199 | ||
200 | tmp = urcu_active_readers; | |
201 | /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */ | |
202 | /* | |
203 | * The data dependency "read urcu_gp_ctr, write urcu_active_readers", | |
204 | * serializes those two memory operations. The memory barrier in the | |
205 | * signal handler ensures we receive the proper memory commit barriers | |
206 | * required by _STORE_SHARED and _LOAD_SHARED whenever communication | |
207 | * with the writer is needed. | |
208 | */ | |
209 | if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) | |
210 | _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr)); | |
211 | else | |
212 | _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT); | |
213 | /* | |
214 | * Increment active readers count before accessing the pointer. | |
215 | * See force_mb_all_threads(). | |
216 | */ | |
217 | reader_barrier(); | |
218 | } | |
219 | ||
220 | static inline void _rcu_read_unlock(void) | |
221 | { | |
222 | reader_barrier(); | |
223 | /* | |
224 | * Finish using rcu before decrementing the pointer. | |
225 | * See force_mb_all_threads(). | |
226 | */ | |
227 | _STORE_SHARED(urcu_active_readers, urcu_active_readers - RCU_GP_COUNT); | |
228 | } | |
229 | ||
230 | /** | |
231 | * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure | |
232 | * meant to be read by RCU read-side critical sections. Returns the assigned | |
233 | * value. | |
234 | * | |
235 | * Documents which pointers will be dereferenced by RCU read-side critical | |
236 | * sections and adds the required memory barriers on architectures requiring | |
237 | * them. It also makes sure the compiler does not reorder code initializing the | |
238 | * data structure before its publication. | |
239 | * | |
240 | * Should match rcu_dereference_pointer(). | |
241 | */ | |
242 | ||
243 | #define _rcu_assign_pointer(p, v) \ | |
244 | ({ \ | |
245 | if (!__builtin_constant_p(v) || \ | |
246 | ((v) != NULL)) \ | |
247 | wmb(); \ | |
248 | STORE_SHARED(p, v); \ | |
249 | }) | |
250 | ||
251 | /** | |
252 | * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous | |
253 | * pointer to the data structure, which can be safely freed after waitin for a | |
254 | * quiescent state using synchronize_rcu(). | |
255 | */ | |
256 | ||
257 | #define _rcu_xchg_pointer(p, v) \ | |
258 | ({ \ | |
259 | if (!__builtin_constant_p(v) || \ | |
260 | ((v) != NULL)) \ | |
261 | wmb(); \ | |
262 | xchg(p, v); \ | |
263 | }) | |
264 | ||
265 | /* | |
266 | * Exchanges the pointer and waits for quiescent state. | |
267 | * The pointer returned can be freed. | |
268 | */ | |
269 | #define _rcu_publish_content(p, v) \ | |
270 | ({ \ | |
271 | void *oldptr; \ | |
272 | oldptr = _rcu_xchg_pointer(p, v); \ | |
273 | synchronize_rcu(); \ | |
274 | oldptr; \ | |
275 | }) | |
276 | ||
277 | #endif /* _URCU_STATIC_H */ |