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 | * |
a5a9f428 PM |
9 | * TO BE INCLUDED ONLY IN CODE THAT IS TO BE RECOMPILED ON EACH LIBURCU |
10 | * RELEASE. See urcu.h for linking dynamically with the userspace rcu library. | |
adcfce54 | 11 | * |
6982d6d7 | 12 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
d2d23035 | 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 | 34 | #include <unistd.h> |
6d841bc2 | 35 | #include <stdint.h> |
adcfce54 | 36 | |
ec4e58a3 MD |
37 | #include <urcu/compiler.h> |
38 | #include <urcu/arch.h> | |
7e30abe3 | 39 | #include <urcu/system.h> |
a2e7bf9c | 40 | #include <urcu/uatomic.h> |
e3b0cef0 | 41 | #include <urcu/list.h> |
41849996 | 42 | #include <urcu/futex.h> |
bd252a04 | 43 | #include <urcu/tls-compat.h> |
e1c0b55c | 44 | #include <urcu/rand-compat.h> |
bc6c15bb | 45 | |
36bc70a8 MD |
46 | #ifdef __cplusplus |
47 | extern "C" { | |
48 | #endif | |
49 | ||
65f1e634 | 50 | /* Default is RCU_MEMBARRIER */ |
fdf01eed | 51 | #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL) |
65f1e634 MD |
52 | #define RCU_MEMBARRIER |
53 | #endif | |
54 | ||
adcfce54 MD |
55 | /* |
56 | * This code section can only be included in LGPL 2.1 compatible source code. | |
57 | * See below for the function call wrappers which can be used in code meant to | |
58 | * be only linked with the Userspace RCU library. This comes with a small | |
59 | * performance degradation on the read-side due to the added function calls. | |
60 | * This is required to permit relinking with newer versions of the library. | |
61 | */ | |
62 | ||
63 | /* | |
64 | * The signal number used by the RCU library can be overridden with | |
02be5561 | 65 | * -DSIGRCU= when compiling the library. |
ddf7eefb | 66 | * Provide backward compatibility for liburcu 0.3.x SIGURCU. |
adcfce54 | 67 | */ |
ddf7eefb MD |
68 | #ifdef SIGURCU |
69 | #define SIGRCU SIGURCU | |
70 | #endif | |
71 | ||
02be5561 MD |
72 | #ifndef SIGRCU |
73 | #define SIGRCU SIGUSR1 | |
adcfce54 MD |
74 | #endif |
75 | ||
fd189fa5 MD |
76 | enum rcu_state { |
77 | RCU_READER_ACTIVE_CURRENT, | |
78 | RCU_READER_ACTIVE_OLD, | |
79 | RCU_READER_INACTIVE, | |
80 | }; | |
81 | ||
7ac06cef MD |
82 | #ifdef DEBUG_RCU |
83 | #define rcu_assert(args...) assert(args) | |
84 | #else | |
85 | #define rcu_assert(args...) | |
86 | #endif | |
87 | ||
adcfce54 MD |
88 | #ifdef DEBUG_YIELD |
89 | #include <sched.h> | |
90 | #include <time.h> | |
91 | #include <pthread.h> | |
92 | #include <unistd.h> | |
93 | ||
1de4df4b MD |
94 | #define RCU_YIELD_READ (1 << 0) |
95 | #define RCU_YIELD_WRITE (1 << 1) | |
adcfce54 | 96 | |
b4ce1526 | 97 | /* |
fdf01eed | 98 | * Updates with RCU_SIGNAL are much slower. Account this in the delay. |
b4ce1526 | 99 | */ |
fdf01eed | 100 | #ifdef RCU_SIGNAL |
adcfce54 | 101 | /* maximum sleep delay, in us */ |
adcfce54 | 102 | #define MAX_SLEEP 30000 |
fdf01eed MD |
103 | #else |
104 | #define MAX_SLEEP 50 | |
adcfce54 MD |
105 | #endif |
106 | ||
1de4df4b MD |
107 | extern unsigned int rcu_yield_active; |
108 | extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield); | |
adcfce54 | 109 | |
1de4df4b | 110 | static inline void rcu_debug_yield_read(void) |
adcfce54 | 111 | { |
1de4df4b MD |
112 | if (rcu_yield_active & RCU_YIELD_READ) |
113 | if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) | |
114 | usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); | |
adcfce54 MD |
115 | } |
116 | ||
1de4df4b | 117 | static inline void rcu_debug_yield_write(void) |
adcfce54 | 118 | { |
1de4df4b MD |
119 | if (rcu_yield_active & RCU_YIELD_WRITE) |
120 | if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) | |
121 | usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); | |
adcfce54 MD |
122 | } |
123 | ||
1de4df4b | 124 | static inline void rcu_debug_yield_init(void) |
adcfce54 | 125 | { |
1de4df4b | 126 | URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self(); |
adcfce54 MD |
127 | } |
128 | #else | |
1de4df4b | 129 | static inline void rcu_debug_yield_read(void) |
adcfce54 MD |
130 | { |
131 | } | |
132 | ||
1de4df4b | 133 | static inline void rcu_debug_yield_write(void) |
adcfce54 MD |
134 | { |
135 | } | |
136 | ||
1de4df4b | 137 | static inline void rcu_debug_yield_init(void) |
adcfce54 MD |
138 | { |
139 | ||
140 | } | |
141 | #endif | |
142 | ||
25cc6d18 MD |
143 | /* |
144 | * RCU memory barrier broadcast group. Currently, only broadcast to all process | |
145 | * threads is supported (group 0). | |
146 | * | |
147 | * Slave barriers are only guaranteed to be ordered wrt master barriers. | |
148 | * | |
149 | * The pair ordering is detailed as (O: ordered, X: not ordered) : | |
150 | * slave master | |
151 | * slave X O | |
152 | * master O O | |
153 | */ | |
154 | ||
155 | #define MB_GROUP_ALL 0 | |
156 | #define RCU_MB_GROUP MB_GROUP_ALL | |
157 | ||
fdf01eed | 158 | #ifdef RCU_MEMBARRIER |
1de4df4b | 159 | extern int rcu_has_sys_membarrier; |
fdf01eed | 160 | |
25cc6d18 | 161 | static inline void smp_mb_slave(int group) |
fdf01eed | 162 | { |
1de4df4b | 163 | if (caa_likely(rcu_has_sys_membarrier)) |
5481ddb3 | 164 | cmm_barrier(); |
fdf01eed | 165 | else |
5481ddb3 | 166 | cmm_smp_mb(); |
fdf01eed MD |
167 | } |
168 | #endif | |
169 | ||
02be5561 | 170 | #ifdef RCU_MB |
25cc6d18 | 171 | static inline void smp_mb_slave(int group) |
adcfce54 | 172 | { |
5481ddb3 | 173 | cmm_smp_mb(); |
adcfce54 | 174 | } |
fdf01eed MD |
175 | #endif |
176 | ||
177 | #ifdef RCU_SIGNAL | |
25cc6d18 | 178 | static inline void smp_mb_slave(int group) |
adcfce54 | 179 | { |
5481ddb3 | 180 | cmm_barrier(); |
adcfce54 MD |
181 | } |
182 | #endif | |
183 | ||
184 | /* | |
02be5561 MD |
185 | * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use |
186 | * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. | |
adcfce54 MD |
187 | */ |
188 | #define RCU_GP_COUNT (1UL << 0) | |
189 | /* Use the amount of bits equal to half of the architecture long size */ | |
27d65bc5 | 190 | #define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2)) |
02be5561 | 191 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1) |
adcfce54 | 192 | |
4de0cd31 | 193 | struct rcu_gp { |
ed1b099e LJ |
194 | /* |
195 | * Global grace period counter. | |
196 | * Contains the current RCU_GP_CTR_PHASE. | |
197 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. | |
198 | * Written to only by writer with mutex taken. | |
199 | * Read by both writer and readers. | |
200 | */ | |
201 | unsigned long ctr; | |
202 | ||
203 | int32_t futex; | |
204 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); | |
205 | ||
4de0cd31 | 206 | extern struct rcu_gp rcu_gp; |
adcfce54 | 207 | |
02be5561 | 208 | struct rcu_reader { |
bd1a5e15 | 209 | /* Data used by both reader and synchronize_rcu() */ |
27d65bc5 | 210 | unsigned long ctr; |
e3b0cef0 | 211 | char need_mb; |
bd1a5e15 | 212 | /* Data used for registry */ |
16aa9ee8 | 213 | struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
bd1a5e15 | 214 | pthread_t tid; |
e3b0cef0 MD |
215 | }; |
216 | ||
bd252a04 | 217 | extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader); |
adcfce54 | 218 | |
bc6c15bb MD |
219 | /* |
220 | * Wake-up waiting synchronize_rcu(). Called from many concurrent threads. | |
221 | */ | |
222 | static inline void wake_up_gp(void) | |
223 | { | |
ed1b099e LJ |
224 | if (caa_unlikely(uatomic_read(&rcu_gp.futex) == -1)) { |
225 | uatomic_set(&rcu_gp.futex, 0); | |
226 | futex_async(&rcu_gp.futex, FUTEX_WAKE, 1, | |
bc6c15bb MD |
227 | NULL, NULL, 0); |
228 | } | |
229 | } | |
230 | ||
fd189fa5 | 231 | static inline enum rcu_state rcu_reader_state(unsigned long *ctr) |
adcfce54 | 232 | { |
27d65bc5 | 233 | unsigned long v; |
adcfce54 | 234 | |
adcfce54 MD |
235 | /* |
236 | * Make sure both tests below are done on the same version of *value | |
237 | * to insure consistency. | |
238 | */ | |
6cf3827c | 239 | v = CMM_LOAD_SHARED(*ctr); |
fd189fa5 MD |
240 | if (!(v & RCU_GP_CTR_NEST_MASK)) |
241 | return RCU_READER_INACTIVE; | |
ed1b099e | 242 | if (!((v ^ rcu_gp.ctr) & RCU_GP_CTR_PHASE)) |
fd189fa5 MD |
243 | return RCU_READER_ACTIVE_CURRENT; |
244 | return RCU_READER_ACTIVE_OLD; | |
adcfce54 MD |
245 | } |
246 | ||
a5a9f428 | 247 | /* |
ed1b099e | 248 | * Helper for _rcu_read_lock(). The format of rcu_gp.ctr (as well as |
a5a9f428 PM |
249 | * the per-thread rcu_reader.ctr) has the upper bits containing a count of |
250 | * _rcu_read_lock() nesting, and a lower-order bit that contains either zero | |
251 | * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in | |
252 | * _rcu_read_lock() happen before the subsequent read-side critical section. | |
253 | */ | |
254 | static inline void _rcu_read_lock_update(unsigned long tmp) | |
adcfce54 | 255 | { |
a0b7f7ea | 256 | if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { |
ed1b099e | 257 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp.ctr)); |
25cc6d18 | 258 | smp_mb_slave(RCU_MB_GROUP); |
a5a9f428 | 259 | } else |
bd252a04 | 260 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT); |
adcfce54 MD |
261 | } |
262 | ||
a5a9f428 PM |
263 | /* |
264 | * Enter an RCU read-side critical section. | |
265 | * | |
266 | * The first cmm_barrier() call ensures that the compiler does not reorder | |
267 | * the body of _rcu_read_lock() with a mutex. | |
268 | * | |
269 | * This function and its helper are both less than 10 lines long. The | |
270 | * intent is that this function meets the 10-line criterion in LGPL, | |
271 | * allowing this function to be invoked directly from non-LGPL code. | |
272 | */ | |
273 | static inline void _rcu_read_lock(void) | |
adcfce54 | 274 | { |
27d65bc5 | 275 | unsigned long tmp; |
bc6c15bb | 276 | |
a5a9f428 | 277 | cmm_barrier(); |
bd252a04 | 278 | tmp = URCU_TLS(rcu_reader).ctr; |
a5a9f428 PM |
279 | _rcu_read_lock_update(tmp); |
280 | } | |
281 | ||
282 | /* | |
283 | * This is a helper function for _rcu_read_unlock(). | |
284 | * | |
285 | * The first smp_mb_slave() call ensures that the critical section is | |
286 | * seen to precede the store to rcu_reader.ctr. | |
287 | * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr | |
288 | * before reading the update-side futex. | |
289 | */ | |
290 | static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp) | |
291 | { | |
a0b7f7ea | 292 | if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) { |
25cc6d18 | 293 | smp_mb_slave(RCU_MB_GROUP); |
bd252a04 | 294 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT); |
25cc6d18 | 295 | smp_mb_slave(RCU_MB_GROUP); |
bc6c15bb | 296 | wake_up_gp(); |
a5a9f428 | 297 | } else |
bd252a04 | 298 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT); |
a5a9f428 PM |
299 | } |
300 | ||
301 | /* | |
302 | * Exit an RCU read-side crtical section. Both this function and its | |
303 | * helper are smaller than 10 lines of code, and are intended to be | |
304 | * usable by non-LGPL code, as called out in LGPL. | |
305 | */ | |
306 | static inline void _rcu_read_unlock(void) | |
307 | { | |
308 | unsigned long tmp; | |
309 | ||
310 | tmp = URCU_TLS(rcu_reader).ctr; | |
311 | _rcu_read_unlock_update_and_wakeup(tmp); | |
5481ddb3 | 312 | cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ |
adcfce54 MD |
313 | } |
314 | ||
882f3357 MD |
315 | /* |
316 | * Returns whether within a RCU read-side critical section. | |
317 | * | |
318 | * This function is less than 10 lines long. The intent is that this | |
319 | * function meets the 10-line criterion for LGPL, allowing this function | |
320 | * to be invoked directly from non-LGPL code. | |
321 | */ | |
322 | static inline int _rcu_read_ongoing(void) | |
323 | { | |
324 | return URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK; | |
325 | } | |
326 | ||
ad918eeb | 327 | #ifdef __cplusplus |
36bc70a8 MD |
328 | } |
329 | #endif | |
330 | ||
adcfce54 | 331 | #endif /* _URCU_STATIC_H */ |