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> | |
bc6c15bb MD |
34 | #include <syscall.h> |
35 | #include <unistd.h> | |
adcfce54 MD |
36 | |
37 | #include <compiler.h> | |
38 | #include <arch.h> | |
39 | ||
40 | /* | |
41 | * Identify a shared load. A smp_rmc() or smp_mc() should come before the load. | |
42 | */ | |
43 | #define _LOAD_SHARED(p) ACCESS_ONCE(p) | |
44 | ||
45 | /* | |
46 | * Load a data from shared memory, doing a cache flush if required. | |
47 | */ | |
48 | #define LOAD_SHARED(p) \ | |
49 | ({ \ | |
50 | smp_rmc(); \ | |
51 | _LOAD_SHARED(p); \ | |
52 | }) | |
53 | ||
54 | /* | |
55 | * Identify a shared store. A smp_wmc() or smp_mc() should follow the store. | |
56 | */ | |
57 | #define _STORE_SHARED(x, v) ({ ACCESS_ONCE(x) = (v); }) | |
58 | ||
59 | /* | |
60 | * Store v into x, where x is located in shared memory. Performs the required | |
61 | * cache flush after writing. Returns v. | |
62 | */ | |
63 | #define STORE_SHARED(x, v) \ | |
64 | ({ \ | |
65 | _STORE_SHARED(x, v); \ | |
66 | smp_wmc(); \ | |
67 | (v); \ | |
68 | }) | |
69 | ||
70 | /** | |
71 | * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable | |
72 | * into a RCU read-side critical section. The pointer can later be safely | |
73 | * dereferenced within the critical section. | |
74 | * | |
75 | * This ensures that the pointer copy is invariant thorough the whole critical | |
76 | * section. | |
77 | * | |
78 | * Inserts memory barriers on architectures that require them (currently only | |
79 | * Alpha) and documents which pointers are protected by RCU. | |
80 | * | |
809f4fde MD |
81 | * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative |
82 | * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the | |
83 | * data read before the pointer read by speculating the value of the pointer. | |
84 | * Correct ordering is ensured because the pointer is read as a volatile access. | |
85 | * This acts as a global side-effect operation, which forbids reordering of | |
015c702f MD |
86 | * dependent memory operations. Note that such concern about dependency-breaking |
87 | * optimizations will eventually be taken care of by the "memory_order_consume" | |
88 | * addition to forthcoming C++ standard. | |
809f4fde | 89 | * |
adcfce54 MD |
90 | * Should match rcu_assign_pointer() or rcu_xchg_pointer(). |
91 | */ | |
92 | ||
93 | #define _rcu_dereference(p) ({ \ | |
94 | typeof(p) _________p1 = LOAD_SHARED(p); \ | |
95 | smp_read_barrier_depends(); \ | |
96 | (_________p1); \ | |
97 | }) | |
98 | ||
bc6c15bb MD |
99 | #define futex(...) syscall(__NR_futex, __VA_ARGS__) |
100 | #define FUTEX_WAIT 0 | |
101 | #define FUTEX_WAKE 1 | |
102 | ||
adcfce54 MD |
103 | /* |
104 | * This code section can only be included in LGPL 2.1 compatible source code. | |
105 | * See below for the function call wrappers which can be used in code meant to | |
106 | * be only linked with the Userspace RCU library. This comes with a small | |
107 | * performance degradation on the read-side due to the added function calls. | |
108 | * This is required to permit relinking with newer versions of the library. | |
109 | */ | |
110 | ||
111 | /* | |
112 | * The signal number used by the RCU library can be overridden with | |
113 | * -DSIGURCU= when compiling the library. | |
114 | */ | |
115 | #ifndef SIGURCU | |
116 | #define SIGURCU SIGUSR1 | |
117 | #endif | |
118 | ||
119 | /* | |
120 | * If a reader is really non-cooperative and refuses to commit its | |
121 | * urcu_active_readers count to memory (there is no barrier in the reader | |
122 | * per-se), kick it after a few loops waiting for it. | |
123 | */ | |
124 | #define KICK_READER_LOOPS 10000 | |
125 | ||
bc6c15bb MD |
126 | /* |
127 | * Active attempts to check for reader Q.S. before calling futex(). | |
128 | */ | |
129 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
130 | ||
7ac06cef MD |
131 | #ifdef DEBUG_RCU |
132 | #define rcu_assert(args...) assert(args) | |
133 | #else | |
134 | #define rcu_assert(args...) | |
135 | #endif | |
136 | ||
adcfce54 MD |
137 | #ifdef DEBUG_YIELD |
138 | #include <sched.h> | |
139 | #include <time.h> | |
140 | #include <pthread.h> | |
141 | #include <unistd.h> | |
142 | ||
143 | #define YIELD_READ (1 << 0) | |
144 | #define YIELD_WRITE (1 << 1) | |
145 | ||
b4ce1526 | 146 | /* |
0a1d290b | 147 | * Updates without URCU_MB are much slower. Account this in |
b4ce1526 MD |
148 | * the delay. |
149 | */ | |
0a1d290b | 150 | #ifdef URCU_MB |
adcfce54 MD |
151 | /* maximum sleep delay, in us */ |
152 | #define MAX_SLEEP 50 | |
153 | #else | |
154 | #define MAX_SLEEP 30000 | |
155 | #endif | |
156 | ||
157 | extern unsigned int yield_active; | |
158 | extern unsigned int __thread rand_yield; | |
159 | ||
160 | static inline void debug_yield_read(void) | |
161 | { | |
162 | if (yield_active & YIELD_READ) | |
163 | if (rand_r(&rand_yield) & 0x1) | |
164 | usleep(rand_r(&rand_yield) % MAX_SLEEP); | |
165 | } | |
166 | ||
167 | static inline void debug_yield_write(void) | |
168 | { | |
169 | if (yield_active & YIELD_WRITE) | |
170 | if (rand_r(&rand_yield) & 0x1) | |
171 | usleep(rand_r(&rand_yield) % MAX_SLEEP); | |
172 | } | |
173 | ||
174 | static inline void debug_yield_init(void) | |
175 | { | |
176 | rand_yield = time(NULL) ^ pthread_self(); | |
177 | } | |
178 | #else | |
179 | static inline void debug_yield_read(void) | |
180 | { | |
181 | } | |
182 | ||
183 | static inline void debug_yield_write(void) | |
184 | { | |
185 | } | |
186 | ||
187 | static inline void debug_yield_init(void) | |
188 | { | |
189 | ||
190 | } | |
191 | #endif | |
192 | ||
0a1d290b | 193 | #ifdef URCU_MB |
adcfce54 MD |
194 | static inline void reader_barrier() |
195 | { | |
196 | smp_mb(); | |
197 | } | |
198 | #else | |
199 | static inline void reader_barrier() | |
200 | { | |
201 | barrier(); | |
202 | } | |
203 | #endif | |
204 | ||
205 | /* | |
206 | * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a | |
207 | * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. | |
208 | */ | |
209 | #define RCU_GP_COUNT (1UL << 0) | |
210 | /* Use the amount of bits equal to half of the architecture long size */ | |
211 | #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2)) | |
212 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1) | |
213 | ||
214 | /* | |
215 | * Global quiescent period counter with low-order bits unused. | |
216 | * Using a int rather than a char to eliminate false register dependencies | |
217 | * causing stalls on some architectures. | |
218 | */ | |
219 | extern long urcu_gp_ctr; | |
220 | ||
221 | extern long __thread urcu_active_readers; | |
222 | ||
bc6c15bb MD |
223 | extern int gp_futex; |
224 | ||
225 | /* | |
226 | * Wake-up waiting synchronize_rcu(). Called from many concurrent threads. | |
227 | */ | |
228 | static inline void wake_up_gp(void) | |
229 | { | |
230 | if (unlikely(atomic_read(&gp_futex) == -1)) { | |
231 | atomic_set(&gp_futex, 0); | |
232 | futex(&gp_futex, FUTEX_WAKE, 1, | |
233 | NULL, NULL, 0); | |
234 | } | |
235 | } | |
236 | ||
adcfce54 MD |
237 | static inline int rcu_old_gp_ongoing(long *value) |
238 | { | |
239 | long v; | |
240 | ||
241 | if (value == NULL) | |
242 | return 0; | |
243 | /* | |
244 | * Make sure both tests below are done on the same version of *value | |
245 | * to insure consistency. | |
246 | */ | |
247 | v = LOAD_SHARED(*value); | |
248 | return (v & RCU_GP_CTR_NEST_MASK) && | |
249 | ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT); | |
250 | } | |
251 | ||
252 | static inline void _rcu_read_lock(void) | |
253 | { | |
254 | long tmp; | |
255 | ||
256 | tmp = urcu_active_readers; | |
257 | /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */ | |
67ef1a2c | 258 | if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { |
adcfce54 | 259 | _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr)); |
67ef1a2c MD |
260 | /* |
261 | * Set active readers count for outermost nesting level before | |
262 | * accessing the pointer. See force_mb_all_threads(). | |
263 | */ | |
264 | reader_barrier(); | |
265 | } else { | |
adcfce54 | 266 | _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT); |
67ef1a2c | 267 | } |
adcfce54 MD |
268 | } |
269 | ||
270 | static inline void _rcu_read_unlock(void) | |
271 | { | |
bc6c15bb MD |
272 | long tmp; |
273 | ||
274 | tmp = urcu_active_readers; | |
adcfce54 MD |
275 | /* |
276 | * Finish using rcu before decrementing the pointer. | |
277 | * See force_mb_all_threads(). | |
278 | */ | |
bc6c15bb MD |
279 | if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) { |
280 | reader_barrier(); | |
281 | _STORE_SHARED(urcu_active_readers, | |
282 | urcu_active_readers - RCU_GP_COUNT); | |
283 | /* write urcu_active_readers before read futex */ | |
284 | reader_barrier(); | |
285 | wake_up_gp(); | |
286 | } else { | |
287 | _STORE_SHARED(urcu_active_readers, | |
288 | urcu_active_readers - RCU_GP_COUNT); | |
289 | } | |
adcfce54 MD |
290 | } |
291 | ||
292 | /** | |
293 | * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure | |
294 | * meant to be read by RCU read-side critical sections. Returns the assigned | |
295 | * value. | |
296 | * | |
297 | * Documents which pointers will be dereferenced by RCU read-side critical | |
298 | * sections and adds the required memory barriers on architectures requiring | |
299 | * them. It also makes sure the compiler does not reorder code initializing the | |
300 | * data structure before its publication. | |
301 | * | |
302 | * Should match rcu_dereference_pointer(). | |
303 | */ | |
304 | ||
305 | #define _rcu_assign_pointer(p, v) \ | |
306 | ({ \ | |
307 | if (!__builtin_constant_p(v) || \ | |
308 | ((v) != NULL)) \ | |
309 | wmb(); \ | |
310 | STORE_SHARED(p, v); \ | |
311 | }) | |
312 | ||
4d1ce26f MD |
313 | /** |
314 | * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer | |
315 | * is as expected by "old". If succeeds, returns the previous pointer to the | |
316 | * data structure, which can be safely freed after waiting for a quiescent state | |
317 | * using synchronize_rcu(). If fails (unexpected value), returns old (which | |
318 | * should not be freed !). | |
319 | */ | |
320 | ||
321 | #define _rcu_cmpxchg_pointer(p, old, _new) \ | |
322 | ({ \ | |
323 | if (!__builtin_constant_p(_new) || \ | |
324 | ((_new) != NULL)) \ | |
325 | wmb(); \ | |
326 | cmpxchg(p, old, _new); \ | |
327 | }) | |
328 | ||
adcfce54 MD |
329 | /** |
330 | * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous | |
67ef1a2c | 331 | * pointer to the data structure, which can be safely freed after waiting for a |
adcfce54 MD |
332 | * quiescent state using synchronize_rcu(). |
333 | */ | |
334 | ||
335 | #define _rcu_xchg_pointer(p, v) \ | |
336 | ({ \ | |
337 | if (!__builtin_constant_p(v) || \ | |
338 | ((v) != NULL)) \ | |
339 | wmb(); \ | |
340 | xchg(p, v); \ | |
341 | }) | |
342 | ||
343 | /* | |
344 | * Exchanges the pointer and waits for quiescent state. | |
345 | * The pointer returned can be freed. | |
346 | */ | |
347 | #define _rcu_publish_content(p, v) \ | |
348 | ({ \ | |
349 | void *oldptr; \ | |
350 | oldptr = _rcu_xchg_pointer(p, v); \ | |
351 | synchronize_rcu(); \ | |
352 | oldptr; \ | |
353 | }) | |
354 | ||
355 | #endif /* _URCU_STATIC_H */ |