| 1 | #ifndef _URCU_H |
| 2 | #define _URCU_H |
| 3 | |
| 4 | /* |
| 5 | * urcu.h |
| 6 | * |
| 7 | * Userspace RCU header |
| 8 | * |
| 9 | * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> |
| 10 | * |
| 11 | * Credits for Paul e. McKenney <paulmck@linux.vnet.ibm.com> |
| 12 | * for inspiration coming from the Linux kernel RCU and rcu-preempt. |
| 13 | * |
| 14 | * The barrier, mb, rmb, wmb, atomic_inc, smp_read_barrier_depends, ACCESS_ONCE |
| 15 | * and rcu_dereference primitives come from the Linux kernel. |
| 16 | * |
| 17 | * Distributed under GPLv2 |
| 18 | */ |
| 19 | |
| 20 | #include <stdlib.h> |
| 21 | #include <pthread.h> |
| 22 | |
| 23 | /* The "volatile" is due to gcc bugs */ |
| 24 | #define barrier() __asm__ __volatile__("": : :"memory") |
| 25 | |
| 26 | #define likely(x) __builtin_expect(!!(x), 1) |
| 27 | #define unlikely(x) __builtin_expect(!!(x), 0) |
| 28 | |
| 29 | /* x86 32/64 specific */ |
| 30 | #define mb() asm volatile("mfence":::"memory") |
| 31 | #define rmb() asm volatile("lfence":::"memory") |
| 32 | #define wmb() asm volatile("sfence" ::: "memory") |
| 33 | |
| 34 | /* Assume SMP machine, given we don't have this information */ |
| 35 | #define CONFIG_SMP 1 |
| 36 | |
| 37 | #ifdef CONFIG_SMP |
| 38 | #define smp_mb() mb() |
| 39 | #define smp_rmb() rmb() |
| 40 | #define smp_wmb() wmb() |
| 41 | #else |
| 42 | #define smp_mb() barrier() |
| 43 | #define smp_rmb() barrier() |
| 44 | #define smp_wmb() barrier() |
| 45 | #endif |
| 46 | |
| 47 | static inline void atomic_inc(int *v) |
| 48 | { |
| 49 | asm volatile("lock; incl %0" |
| 50 | : "+m" (*v)); |
| 51 | } |
| 52 | |
| 53 | #define xchg(ptr, v) \ |
| 54 | ((__typeof__(*(ptr)))__xchg((unsigned long)(v), (ptr), sizeof(*(ptr)))) |
| 55 | |
| 56 | struct __xchg_dummy { |
| 57 | unsigned long a[100]; |
| 58 | }; |
| 59 | #define __xg(x) ((struct __xchg_dummy *)(x)) |
| 60 | |
| 61 | /* |
| 62 | * Note: no "lock" prefix even on SMP: xchg always implies lock anyway |
| 63 | * Note 2: xchg has side effect, so that attribute volatile is necessary, |
| 64 | * but generally the primitive is invalid, *ptr is output argument. --ANK |
| 65 | */ |
| 66 | static inline unsigned long __xchg(unsigned long x, volatile void *ptr, |
| 67 | int size) |
| 68 | { |
| 69 | switch (size) { |
| 70 | case 1: |
| 71 | asm volatile("xchgb %b0,%1" |
| 72 | : "=q" (x) |
| 73 | : "m" (*__xg(ptr)), "0" (x) |
| 74 | : "memory"); |
| 75 | break; |
| 76 | case 2: |
| 77 | asm volatile("xchgw %w0,%1" |
| 78 | : "=r" (x) |
| 79 | : "m" (*__xg(ptr)), "0" (x) |
| 80 | : "memory"); |
| 81 | break; |
| 82 | case 4: |
| 83 | asm volatile("xchgl %k0,%1" |
| 84 | : "=r" (x) |
| 85 | : "m" (*__xg(ptr)), "0" (x) |
| 86 | : "memory"); |
| 87 | break; |
| 88 | case 8: |
| 89 | asm volatile("xchgq %0,%1" |
| 90 | : "=r" (x) |
| 91 | : "m" (*__xg(ptr)), "0" (x) |
| 92 | : "memory"); |
| 93 | break; |
| 94 | } |
| 95 | return x; |
| 96 | } |
| 97 | |
| 98 | /* Nop everywhere except on alpha. */ |
| 99 | #define smp_read_barrier_depends() |
| 100 | |
| 101 | /* |
| 102 | * Prevent the compiler from merging or refetching accesses. The compiler |
| 103 | * is also forbidden from reordering successive instances of ACCESS_ONCE(), |
| 104 | * but only when the compiler is aware of some particular ordering. One way |
| 105 | * to make the compiler aware of ordering is to put the two invocations of |
| 106 | * ACCESS_ONCE() in different C statements. |
| 107 | * |
| 108 | * This macro does absolutely -nothing- to prevent the CPU from reordering, |
| 109 | * merging, or refetching absolutely anything at any time. Its main intended |
| 110 | * use is to mediate communication between process-level code and irq/NMI |
| 111 | * handlers, all running on the same CPU. |
| 112 | */ |
| 113 | #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x)) |
| 114 | |
| 115 | /** |
| 116 | * rcu_dereference - fetch an RCU-protected pointer in an |
| 117 | * RCU read-side critical section. This pointer may later |
| 118 | * be safely dereferenced. |
| 119 | * |
| 120 | * Inserts memory barriers on architectures that require them |
| 121 | * (currently only the Alpha), and, more importantly, documents |
| 122 | * exactly which pointers are protected by RCU. |
| 123 | */ |
| 124 | |
| 125 | #define rcu_dereference(p) ({ \ |
| 126 | typeof(p) _________p1 = ACCESS_ONCE(p); \ |
| 127 | smp_read_barrier_depends(); \ |
| 128 | (_________p1); \ |
| 129 | }) |
| 130 | |
| 131 | #define SIGURCU SIGUSR1 |
| 132 | |
| 133 | #ifdef DEBUG_YIELD |
| 134 | #include <sched.h> |
| 135 | #include <time.h> |
| 136 | #include <pthread.h> |
| 137 | #include <unistd.h> |
| 138 | |
| 139 | #define YIELD_READ (1 << 0) |
| 140 | #define YIELD_WRITE (1 << 1) |
| 141 | |
| 142 | /* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */ |
| 143 | #ifdef DEBUG_FULL_MB |
| 144 | /* maximum sleep delay, in us */ |
| 145 | #define MAX_SLEEP 50 |
| 146 | #else |
| 147 | #define MAX_SLEEP 30000 |
| 148 | #endif |
| 149 | |
| 150 | extern unsigned int yield_active; |
| 151 | extern unsigned int __thread rand_yield; |
| 152 | |
| 153 | static inline void debug_yield_read(void) |
| 154 | { |
| 155 | if (yield_active & YIELD_READ) |
| 156 | if (rand_r(&rand_yield) & 0x1) |
| 157 | usleep(rand_r(&rand_yield) % MAX_SLEEP); |
| 158 | } |
| 159 | |
| 160 | static inline void debug_yield_write(void) |
| 161 | { |
| 162 | if (yield_active & YIELD_WRITE) |
| 163 | if (rand_r(&rand_yield) & 0x1) |
| 164 | usleep(rand_r(&rand_yield) % MAX_SLEEP); |
| 165 | } |
| 166 | |
| 167 | static inline void debug_yield_init(void) |
| 168 | { |
| 169 | rand_yield = time(NULL) ^ pthread_self(); |
| 170 | } |
| 171 | #else |
| 172 | static inline void debug_yield_read(void) |
| 173 | { |
| 174 | } |
| 175 | |
| 176 | static inline void debug_yield_write(void) |
| 177 | { |
| 178 | } |
| 179 | |
| 180 | static inline void debug_yield_init(void) |
| 181 | { |
| 182 | |
| 183 | } |
| 184 | #endif |
| 185 | |
| 186 | #ifdef DEBUG_FULL_MB |
| 187 | static inline void read_barrier() |
| 188 | { |
| 189 | smp_mb(); |
| 190 | } |
| 191 | #else |
| 192 | static inline void read_barrier() |
| 193 | { |
| 194 | barrier(); |
| 195 | } |
| 196 | #endif |
| 197 | |
| 198 | /* |
| 199 | * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a |
| 200 | * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. |
| 201 | */ |
| 202 | #define RCU_GP_COUNT (1UL << 0) |
| 203 | /* Use the amount of bits equal to half of the architecture long size */ |
| 204 | #define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2)) |
| 205 | #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1) |
| 206 | |
| 207 | /* |
| 208 | * Global quiescent period counter with low-order bits unused. |
| 209 | * Using a int rather than a char to eliminate false register dependencies |
| 210 | * causing stalls on some architectures. |
| 211 | */ |
| 212 | extern long urcu_gp_ctr; |
| 213 | |
| 214 | extern long __thread urcu_active_readers; |
| 215 | |
| 216 | static inline int rcu_old_gp_ongoing(long *value) |
| 217 | { |
| 218 | long v; |
| 219 | |
| 220 | if (value == NULL) |
| 221 | return 0; |
| 222 | /* |
| 223 | * Make sure both tests below are done on the same version of *value |
| 224 | * to insure consistency. |
| 225 | */ |
| 226 | v = ACCESS_ONCE(*value); |
| 227 | return (v & RCU_GP_CTR_NEST_MASK) && |
| 228 | ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT); |
| 229 | } |
| 230 | |
| 231 | static inline void rcu_read_lock(void) |
| 232 | { |
| 233 | long tmp; |
| 234 | |
| 235 | tmp = urcu_active_readers; |
| 236 | /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */ |
| 237 | if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) |
| 238 | urcu_active_readers = urcu_gp_ctr; |
| 239 | else |
| 240 | urcu_active_readers = tmp + RCU_GP_COUNT; |
| 241 | /* |
| 242 | * Increment active readers count before accessing the pointer. |
| 243 | * See force_mb_all_threads(). |
| 244 | */ |
| 245 | read_barrier(); |
| 246 | } |
| 247 | |
| 248 | static inline void rcu_read_unlock(void) |
| 249 | { |
| 250 | read_barrier(); |
| 251 | /* |
| 252 | * Finish using rcu before decrementing the pointer. |
| 253 | * See force_mb_all_threads(). |
| 254 | */ |
| 255 | urcu_active_readers -= RCU_GP_COUNT; |
| 256 | } |
| 257 | |
| 258 | /** |
| 259 | * rcu_assign_pointer - assign (publicize) a pointer to a newly |
| 260 | * initialized structure that will be dereferenced by RCU read-side |
| 261 | * critical sections. Returns the value assigned. |
| 262 | * |
| 263 | * Inserts memory barriers on architectures that require them |
| 264 | * (pretty much all of them other than x86), and also prevents |
| 265 | * the compiler from reordering the code that initializes the |
| 266 | * structure after the pointer assignment. More importantly, this |
| 267 | * call documents which pointers will be dereferenced by RCU read-side |
| 268 | * code. |
| 269 | */ |
| 270 | |
| 271 | #define rcu_assign_pointer(p, v) \ |
| 272 | ({ \ |
| 273 | if (!__builtin_constant_p(v) || \ |
| 274 | ((v) != NULL)) \ |
| 275 | wmb(); \ |
| 276 | (p) = (v); \ |
| 277 | }) |
| 278 | |
| 279 | #define rcu_xchg_pointer(p, v) \ |
| 280 | ({ \ |
| 281 | if (!__builtin_constant_p(v) || \ |
| 282 | ((v) != NULL)) \ |
| 283 | wmb(); \ |
| 284 | xchg(p, v); \ |
| 285 | }) |
| 286 | |
| 287 | extern void synchronize_rcu(void); |
| 288 | |
| 289 | /* |
| 290 | * Exchanges the pointer and waits for quiescent state. |
| 291 | * The pointer returned can be freed. |
| 292 | */ |
| 293 | #define urcu_publish_content(p, v) \ |
| 294 | ({ \ |
| 295 | void *oldptr; \ |
| 296 | oldptr = rcu_xchg_pointer(p, v); \ |
| 297 | synchronize_rcu(); \ |
| 298 | oldptr; \ |
| 299 | }) |
| 300 | |
| 301 | /* |
| 302 | * Reader thread registration. |
| 303 | */ |
| 304 | extern void urcu_register_thread(void); |
| 305 | extern void urcu_unregister_thread(void); |
| 306 | |
| 307 | #endif /* _URCU_H */ |