| 1 | #ifndef _URCU_STATIC_H |
| 2 | #define _URCU_STATIC_H |
| 3 | |
| 4 | /* |
| 5 | * urcu-static.h |
| 6 | * |
| 7 | * Userspace RCU header. |
| 8 | * |
| 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. |
| 11 | * |
| 12 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 13 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. |
| 14 | * |
| 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 |
| 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 | #include <unistd.h> |
| 35 | #include <stdint.h> |
| 36 | |
| 37 | #include <urcu/compiler.h> |
| 38 | #include <urcu/arch.h> |
| 39 | #include <urcu/system.h> |
| 40 | #include <urcu/uatomic.h> |
| 41 | #include <urcu/list.h> |
| 42 | #include <urcu/futex.h> |
| 43 | #include <urcu/tls-compat.h> |
| 44 | |
| 45 | #ifdef __cplusplus |
| 46 | extern "C" { |
| 47 | #endif |
| 48 | |
| 49 | /* Default is RCU_MEMBARRIER */ |
| 50 | #if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL) |
| 51 | #define RCU_MEMBARRIER |
| 52 | #endif |
| 53 | |
| 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 |
| 64 | * -DSIGRCU= when compiling the library. |
| 65 | * Provide backward compatibility for liburcu 0.3.x SIGURCU. |
| 66 | */ |
| 67 | #ifdef SIGURCU |
| 68 | #define SIGRCU SIGURCU |
| 69 | #endif |
| 70 | |
| 71 | #ifndef SIGRCU |
| 72 | #define SIGRCU SIGUSR1 |
| 73 | #endif |
| 74 | |
| 75 | enum rcu_state { |
| 76 | RCU_READER_ACTIVE_CURRENT, |
| 77 | RCU_READER_ACTIVE_OLD, |
| 78 | RCU_READER_INACTIVE, |
| 79 | }; |
| 80 | |
| 81 | #ifdef DEBUG_RCU |
| 82 | #define rcu_assert(args...) assert(args) |
| 83 | #else |
| 84 | #define rcu_assert(args...) |
| 85 | #endif |
| 86 | |
| 87 | #ifdef DEBUG_YIELD |
| 88 | #include <sched.h> |
| 89 | #include <time.h> |
| 90 | #include <pthread.h> |
| 91 | #include <unistd.h> |
| 92 | |
| 93 | #define RCU_YIELD_READ (1 << 0) |
| 94 | #define RCU_YIELD_WRITE (1 << 1) |
| 95 | |
| 96 | /* |
| 97 | * Updates with RCU_SIGNAL are much slower. Account this in the delay. |
| 98 | */ |
| 99 | #ifdef RCU_SIGNAL |
| 100 | /* maximum sleep delay, in us */ |
| 101 | #define MAX_SLEEP 30000 |
| 102 | #else |
| 103 | #define MAX_SLEEP 50 |
| 104 | #endif |
| 105 | |
| 106 | extern unsigned int rcu_yield_active; |
| 107 | extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield); |
| 108 | |
| 109 | static inline void rcu_debug_yield_read(void) |
| 110 | { |
| 111 | if (rcu_yield_active & RCU_YIELD_READ) |
| 112 | if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) |
| 113 | usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); |
| 114 | } |
| 115 | |
| 116 | static inline void rcu_debug_yield_write(void) |
| 117 | { |
| 118 | if (rcu_yield_active & RCU_YIELD_WRITE) |
| 119 | if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) |
| 120 | usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); |
| 121 | } |
| 122 | |
| 123 | static inline void rcu_debug_yield_init(void) |
| 124 | { |
| 125 | URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self(); |
| 126 | } |
| 127 | #else |
| 128 | static inline void rcu_debug_yield_read(void) |
| 129 | { |
| 130 | } |
| 131 | |
| 132 | static inline void rcu_debug_yield_write(void) |
| 133 | { |
| 134 | } |
| 135 | |
| 136 | static inline void rcu_debug_yield_init(void) |
| 137 | { |
| 138 | |
| 139 | } |
| 140 | #endif |
| 141 | |
| 142 | /* |
| 143 | * RCU memory barrier broadcast group. Currently, only broadcast to all process |
| 144 | * threads is supported (group 0). |
| 145 | * |
| 146 | * Slave barriers are only guaranteed to be ordered wrt master barriers. |
| 147 | * |
| 148 | * The pair ordering is detailed as (O: ordered, X: not ordered) : |
| 149 | * slave master |
| 150 | * slave X O |
| 151 | * master O O |
| 152 | */ |
| 153 | |
| 154 | #define MB_GROUP_ALL 0 |
| 155 | #define RCU_MB_GROUP MB_GROUP_ALL |
| 156 | |
| 157 | #ifdef RCU_MEMBARRIER |
| 158 | extern int rcu_has_sys_membarrier; |
| 159 | |
| 160 | static inline void smp_mb_slave(int group) |
| 161 | { |
| 162 | if (caa_likely(rcu_has_sys_membarrier)) |
| 163 | cmm_barrier(); |
| 164 | else |
| 165 | cmm_smp_mb(); |
| 166 | } |
| 167 | #endif |
| 168 | |
| 169 | #ifdef RCU_MB |
| 170 | static inline void smp_mb_slave(int group) |
| 171 | { |
| 172 | cmm_smp_mb(); |
| 173 | } |
| 174 | #endif |
| 175 | |
| 176 | #ifdef RCU_SIGNAL |
| 177 | static inline void smp_mb_slave(int group) |
| 178 | { |
| 179 | cmm_barrier(); |
| 180 | } |
| 181 | #endif |
| 182 | |
| 183 | /* |
| 184 | * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use |
| 185 | * a full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. |
| 186 | */ |
| 187 | #define RCU_GP_COUNT (1UL << 0) |
| 188 | /* Use the amount of bits equal to half of the architecture long size */ |
| 189 | #define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2)) |
| 190 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1) |
| 191 | |
| 192 | struct rcu_gp { |
| 193 | /* |
| 194 | * Global grace period counter. |
| 195 | * Contains the current RCU_GP_CTR_PHASE. |
| 196 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
| 197 | * Written to only by writer with mutex taken. |
| 198 | * Read by both writer and readers. |
| 199 | */ |
| 200 | unsigned long ctr; |
| 201 | |
| 202 | int32_t futex; |
| 203 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 204 | |
| 205 | extern struct rcu_gp rcu_gp; |
| 206 | |
| 207 | struct rcu_reader { |
| 208 | /* Data used by both reader and synchronize_rcu() */ |
| 209 | unsigned long ctr; |
| 210 | char need_mb; |
| 211 | /* Data used for registry */ |
| 212 | struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 213 | pthread_t tid; |
| 214 | }; |
| 215 | |
| 216 | extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader); |
| 217 | |
| 218 | /* |
| 219 | * Wake-up waiting synchronize_rcu(). Called from many concurrent threads. |
| 220 | */ |
| 221 | static inline void wake_up_gp(void) |
| 222 | { |
| 223 | if (caa_unlikely(uatomic_read(&rcu_gp.futex) == -1)) { |
| 224 | uatomic_set(&rcu_gp.futex, 0); |
| 225 | futex_async(&rcu_gp.futex, FUTEX_WAKE, 1, |
| 226 | NULL, NULL, 0); |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | static inline enum rcu_state rcu_reader_state(unsigned long *ctr) |
| 231 | { |
| 232 | unsigned long v; |
| 233 | |
| 234 | /* |
| 235 | * Make sure both tests below are done on the same version of *value |
| 236 | * to insure consistency. |
| 237 | */ |
| 238 | v = CMM_LOAD_SHARED(*ctr); |
| 239 | if (!(v & RCU_GP_CTR_NEST_MASK)) |
| 240 | return RCU_READER_INACTIVE; |
| 241 | if (!((v ^ rcu_gp.ctr) & RCU_GP_CTR_PHASE)) |
| 242 | return RCU_READER_ACTIVE_CURRENT; |
| 243 | return RCU_READER_ACTIVE_OLD; |
| 244 | } |
| 245 | |
| 246 | /* |
| 247 | * Helper for _rcu_read_lock(). The format of rcu_gp.ctr (as well as |
| 248 | * the per-thread rcu_reader.ctr) has the upper bits containing a count of |
| 249 | * _rcu_read_lock() nesting, and a lower-order bit that contains either zero |
| 250 | * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in |
| 251 | * _rcu_read_lock() happen before the subsequent read-side critical section. |
| 252 | */ |
| 253 | static inline void _rcu_read_lock_update(unsigned long tmp) |
| 254 | { |
| 255 | if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { |
| 256 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp.ctr)); |
| 257 | smp_mb_slave(RCU_MB_GROUP); |
| 258 | } else |
| 259 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT); |
| 260 | } |
| 261 | |
| 262 | /* |
| 263 | * Enter an RCU read-side critical section. |
| 264 | * |
| 265 | * The first cmm_barrier() call ensures that the compiler does not reorder |
| 266 | * the body of _rcu_read_lock() with a mutex. |
| 267 | * |
| 268 | * This function and its helper are both less than 10 lines long. The |
| 269 | * intent is that this function meets the 10-line criterion in LGPL, |
| 270 | * allowing this function to be invoked directly from non-LGPL code. |
| 271 | */ |
| 272 | static inline void _rcu_read_lock(void) |
| 273 | { |
| 274 | unsigned long tmp; |
| 275 | |
| 276 | cmm_barrier(); |
| 277 | tmp = URCU_TLS(rcu_reader).ctr; |
| 278 | _rcu_read_lock_update(tmp); |
| 279 | } |
| 280 | |
| 281 | /* |
| 282 | * This is a helper function for _rcu_read_unlock(). |
| 283 | * |
| 284 | * The first smp_mb_slave() call ensures that the critical section is |
| 285 | * seen to precede the store to rcu_reader.ctr. |
| 286 | * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr |
| 287 | * before reading the update-side futex. |
| 288 | */ |
| 289 | static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp) |
| 290 | { |
| 291 | if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) { |
| 292 | smp_mb_slave(RCU_MB_GROUP); |
| 293 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT); |
| 294 | smp_mb_slave(RCU_MB_GROUP); |
| 295 | wake_up_gp(); |
| 296 | } else |
| 297 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT); |
| 298 | } |
| 299 | |
| 300 | /* |
| 301 | * Exit an RCU read-side crtical section. Both this function and its |
| 302 | * helper are smaller than 10 lines of code, and are intended to be |
| 303 | * usable by non-LGPL code, as called out in LGPL. |
| 304 | */ |
| 305 | static inline void _rcu_read_unlock(void) |
| 306 | { |
| 307 | unsigned long tmp; |
| 308 | |
| 309 | tmp = URCU_TLS(rcu_reader).ctr; |
| 310 | _rcu_read_unlock_update_and_wakeup(tmp); |
| 311 | cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ |
| 312 | } |
| 313 | |
| 314 | /* |
| 315 | * Returns whether within a RCU read-side critical section. |
| 316 | * |
| 317 | * This function is less than 10 lines long. The intent is that this |
| 318 | * function meets the 10-line criterion for LGPL, allowing this function |
| 319 | * to be invoked directly from non-LGPL code. |
| 320 | */ |
| 321 | static inline int _rcu_read_ongoing(void) |
| 322 | { |
| 323 | return URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK; |
| 324 | } |
| 325 | |
| 326 | #ifdef __cplusplus |
| 327 | } |
| 328 | #endif |
| 329 | |
| 330 | #endif /* _URCU_STATIC_H */ |