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