| 1 | #ifndef _URCU_BP_STATIC_H |
| 2 | #define _URCU_BP_STATIC_H |
| 3 | |
| 4 | /* |
| 5 | * urcu-bp-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 | |
| 36 | #include <urcu/compiler.h> |
| 37 | #include <urcu/arch.h> |
| 38 | #include <urcu/system.h> |
| 39 | #include <urcu/uatomic.h> |
| 40 | #include <urcu/list.h> |
| 41 | #include <urcu/tls-compat.h> |
| 42 | |
| 43 | /* |
| 44 | * This code section can only be included in LGPL 2.1 compatible source code. |
| 45 | * See below for the function call wrappers which can be used in code meant to |
| 46 | * be only linked with the Userspace RCU library. This comes with a small |
| 47 | * performance degradation on the read-side due to the added function calls. |
| 48 | * This is required to permit relinking with newer versions of the library. |
| 49 | */ |
| 50 | |
| 51 | #ifdef __cplusplus |
| 52 | extern "C" { |
| 53 | #endif |
| 54 | |
| 55 | #ifdef DEBUG_RCU |
| 56 | #define rcu_assert(args...) assert(args) |
| 57 | #else |
| 58 | #define rcu_assert(args...) |
| 59 | #endif |
| 60 | |
| 61 | enum rcu_state { |
| 62 | RCU_READER_ACTIVE_CURRENT, |
| 63 | RCU_READER_ACTIVE_OLD, |
| 64 | RCU_READER_INACTIVE, |
| 65 | }; |
| 66 | |
| 67 | #ifdef DEBUG_YIELD |
| 68 | #include <sched.h> |
| 69 | #include <time.h> |
| 70 | #include <pthread.h> |
| 71 | #include <unistd.h> |
| 72 | |
| 73 | #define RCU_YIELD_READ (1 << 0) |
| 74 | #define RCU_YIELD_WRITE (1 << 1) |
| 75 | |
| 76 | /* |
| 77 | * Updates without RCU_MB are much slower. Account this in |
| 78 | * the delay. |
| 79 | */ |
| 80 | /* maximum sleep delay, in us */ |
| 81 | #define MAX_SLEEP 50 |
| 82 | |
| 83 | extern unsigned int rcu_yield_active; |
| 84 | extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield); |
| 85 | |
| 86 | static inline void rcu_debug_yield_read(void) |
| 87 | { |
| 88 | if (rcu_yield_active & RCU_YIELD_READ) |
| 89 | if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) |
| 90 | usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); |
| 91 | } |
| 92 | |
| 93 | static inline void rcu_debug_yield_write(void) |
| 94 | { |
| 95 | if (rcu_yield_active & RCU_YIELD_WRITE) |
| 96 | if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) |
| 97 | usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); |
| 98 | } |
| 99 | |
| 100 | static inline void rcu_debug_yield_init(void) |
| 101 | { |
| 102 | URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self(); |
| 103 | } |
| 104 | #else |
| 105 | static inline void rcu_debug_yield_read(void) |
| 106 | { |
| 107 | } |
| 108 | |
| 109 | static inline void rcu_debug_yield_write(void) |
| 110 | { |
| 111 | } |
| 112 | |
| 113 | static inline void rcu_debug_yield_init(void) |
| 114 | { |
| 115 | |
| 116 | } |
| 117 | #endif |
| 118 | |
| 119 | /* |
| 120 | * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use a |
| 121 | * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. |
| 122 | */ |
| 123 | #define RCU_GP_COUNT (1UL << 0) |
| 124 | /* Use the amount of bits equal to half of the architecture long size */ |
| 125 | #define RCU_GP_CTR_PHASE (1UL << (sizeof(long) << 2)) |
| 126 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1) |
| 127 | |
| 128 | /* |
| 129 | * Used internally by _rcu_read_lock. |
| 130 | */ |
| 131 | extern void rcu_bp_register(void); |
| 132 | |
| 133 | struct rcu_gp { |
| 134 | /* |
| 135 | * Global grace period counter. |
| 136 | * Contains the current RCU_GP_CTR_PHASE. |
| 137 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
| 138 | * Written to only by writer with mutex taken. |
| 139 | * Read by both writer and readers. |
| 140 | */ |
| 141 | unsigned long ctr; |
| 142 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 143 | |
| 144 | extern struct rcu_gp rcu_gp; |
| 145 | |
| 146 | struct rcu_reader { |
| 147 | /* Data used by both reader and synchronize_rcu() */ |
| 148 | unsigned long ctr; |
| 149 | /* Data used for registry */ |
| 150 | struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 151 | pthread_t tid; |
| 152 | int alloc; /* registry entry allocated */ |
| 153 | }; |
| 154 | |
| 155 | /* |
| 156 | * Bulletproof version keeps a pointer to a registry not part of the TLS. |
| 157 | * Adds a pointer dereference on the read-side, but won't require to unregister |
| 158 | * the reader thread. |
| 159 | */ |
| 160 | extern DECLARE_URCU_TLS(struct rcu_reader *, rcu_reader); |
| 161 | |
| 162 | static inline enum rcu_state rcu_reader_state(unsigned long *ctr) |
| 163 | { |
| 164 | unsigned long v; |
| 165 | |
| 166 | if (ctr == NULL) |
| 167 | return RCU_READER_INACTIVE; |
| 168 | /* |
| 169 | * Make sure both tests below are done on the same version of *value |
| 170 | * to insure consistency. |
| 171 | */ |
| 172 | v = CMM_LOAD_SHARED(*ctr); |
| 173 | if (!(v & RCU_GP_CTR_NEST_MASK)) |
| 174 | return RCU_READER_INACTIVE; |
| 175 | if (!((v ^ rcu_gp.ctr) & RCU_GP_CTR_PHASE)) |
| 176 | return RCU_READER_ACTIVE_CURRENT; |
| 177 | return RCU_READER_ACTIVE_OLD; |
| 178 | } |
| 179 | |
| 180 | /* |
| 181 | * Helper for _rcu_read_lock(). The format of rcu_gp.ctr (as well as |
| 182 | * the per-thread rcu_reader.ctr) has the upper bits containing a count of |
| 183 | * _rcu_read_lock() nesting, and a lower-order bit that contains either zero |
| 184 | * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in |
| 185 | * _rcu_read_lock() happen before the subsequent read-side critical section. |
| 186 | */ |
| 187 | static inline void _rcu_read_lock_update(unsigned long tmp) |
| 188 | { |
| 189 | if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { |
| 190 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, _CMM_LOAD_SHARED(rcu_gp.ctr)); |
| 191 | cmm_smp_mb(); |
| 192 | } else |
| 193 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp + RCU_GP_COUNT); |
| 194 | } |
| 195 | |
| 196 | /* |
| 197 | * Enter an RCU read-side critical section. |
| 198 | * |
| 199 | * The first cmm_barrier() call ensures that the compiler does not reorder |
| 200 | * the body of _rcu_read_lock() with a mutex. |
| 201 | * |
| 202 | * This function and its helper are both less than 10 lines long. The |
| 203 | * intent is that this function meets the 10-line criterion in LGPL, |
| 204 | * allowing this function to be invoked directly from non-LGPL code. |
| 205 | */ |
| 206 | static inline void _rcu_read_lock(void) |
| 207 | { |
| 208 | unsigned long tmp; |
| 209 | |
| 210 | if (caa_unlikely(!URCU_TLS(rcu_reader))) |
| 211 | rcu_bp_register(); /* If not yet registered. */ |
| 212 | cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ |
| 213 | tmp = URCU_TLS(rcu_reader)->ctr; |
| 214 | _rcu_read_lock_update(tmp); |
| 215 | } |
| 216 | |
| 217 | /* |
| 218 | * Exit an RCU read-side critical section. This function is less than |
| 219 | * 10 lines of code, and is intended to be usable by non-LGPL code, as |
| 220 | * called out in LGPL. |
| 221 | */ |
| 222 | static inline void _rcu_read_unlock(void) |
| 223 | { |
| 224 | /* |
| 225 | * Finish using rcu before decrementing the pointer. |
| 226 | */ |
| 227 | cmm_smp_mb(); |
| 228 | _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, URCU_TLS(rcu_reader)->ctr - RCU_GP_COUNT); |
| 229 | cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ |
| 230 | } |
| 231 | |
| 232 | /* |
| 233 | * Returns whether within a RCU read-side critical section. |
| 234 | * |
| 235 | * This function is less than 10 lines long. The intent is that this |
| 236 | * function meets the 10-line criterion for LGPL, allowing this function |
| 237 | * to be invoked directly from non-LGPL code. |
| 238 | */ |
| 239 | static inline int _rcu_read_ongoing(void) |
| 240 | { |
| 241 | if (caa_unlikely(!URCU_TLS(rcu_reader))) |
| 242 | rcu_bp_register(); /* If not yet registered. */ |
| 243 | return URCU_TLS(rcu_reader)->ctr & RCU_GP_CTR_NEST_MASK; |
| 244 | } |
| 245 | |
| 246 | #ifdef __cplusplus |
| 247 | } |
| 248 | #endif |
| 249 | |
| 250 | #endif /* _URCU_BP_STATIC_H */ |