1 /*===-- atomic.c - Implement support functions for atomic operations.------=== 2 * 3 * The LLVM Compiler Infrastructure 4 * 5 * This file is dual licensed under the MIT and the University of Illinois Open 6 * Source Licenses. See LICENSE.TXT for details. 7 * 8 *===----------------------------------------------------------------------=== 9 * 10 * atomic.c defines a set of functions for performing atomic accesses on 11 * arbitrary-sized memory locations. This design uses locks that should 12 * be fast in the uncontended case, for two reasons: 13 * 14 * 1) This code must work with C programs that do not link to anything 15 * (including pthreads) and so it should not depend on any pthread 16 * functions. 17 * 2) Atomic operations, rather than explicit mutexes, are most commonly used 18 * on code where contended operations are rate. 19 * 20 * To avoid needing a per-object lock, this code allocates an array of 21 * locks and hashes the object pointers to find the one that it should use. 22 * For operations that must be atomic on two locations, the lower lock is 23 * always acquired first, to avoid deadlock. 24 * 25 *===----------------------------------------------------------------------=== 26 */ 27 28 #include <stdint.h> 29 #include <string.h> 30 31 // Clang objects if you redefine a builtin. This little hack allows us to 32 // define a function with the same name as an intrinsic. 33 #pragma redefine_extname __atomic_load_n __atomic_load 34 #pragma redefine_extname __atomic_store_n __atomic_store 35 #pragma redefine_extname __atomic_exchange_n __atomic_exchange 36 #pragma redefine_extname __atomic_compare_exchange_n __atomic_compare_exchange 37 38 /// Number of locks. This allocates one page on 32-bit platforms, two on 39 /// 64-bit. This can be specified externally if a different trade between 40 /// memory usage and contention probability is required for a given platform. 41 #ifndef SPINLOCK_COUNT 42 #define SPINLOCK_COUNT (1<<10) 43 #endif 44 static const long SPINLOCK_MASK = SPINLOCK_COUNT - 1; 45 46 //////////////////////////////////////////////////////////////////////////////// 47 // Platform-specific lock implementation. Falls back to spinlocks if none is 48 // defined. Each platform should define the Lock type, and corresponding 49 // lock() and unlock() functions. 50 //////////////////////////////////////////////////////////////////////////////// 51 #ifdef __FreeBSD__ 52 #include <errno.h> 53 #include <sys/types.h> 54 #include <machine/atomic.h> 55 #include <sys/umtx.h> 56 typedef struct _usem Lock; 57 inline static void unlock(Lock *l) { 58 __atomic_store((_Atomic(uint32_t)*)&l->_count, 1, __ATOMIC_RELEASE); 59 __atomic_thread_fence(__ATOMIC_SEQ_CST); 60 if (l->_has_waiters) 61 _umtx_op(l, UMTX_OP_SEM_WAKE, 1, 0, 0); 62 } 63 inline static void lock(Lock *l) { 64 uint32_t old = 1; 65 while (!__atomic_compare_exchange_weak((_Atomic(uint32_t)*)&l->_count, &old, 66 0, __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { 67 _umtx_op(l, UMTX_OP_SEM_WAIT, 0, 0, 0); 68 old = 1; 69 } 70 } 71 /// locks for atomic operations 72 static Lock locks[SPINLOCK_COUNT] = { [0 ... SPINLOCK_COUNT-1] = {0,1,0} }; 73 #else 74 typedef _Atomic(uintptr_t) Lock; 75 /// Unlock a lock. This is a release operation. 76 inline static void unlock(Lock *l) { 77 __atomic_store(l, 0, __ATOMIC_RELEASE); 78 } 79 /// Locks a lock. In the current implementation, this is potentially 80 /// unbounded in the contended case. 81 inline static void lock(Lock *l) { 82 uintptr_t old = 0; 83 while (!__atomic_compare_exchange_weak(l, &old, 1, __ATOMIC_ACQUIRE, 84 __ATOMIC_RELAXED)) 85 old = 0; 86 } 87 /// locks for atomic operations 88 static Lock locks[SPINLOCK_COUNT]; 89 #endif 90 91 92 /// Returns a lock to use for a given pointer. 93 static inline Lock *lock_for_pointer(void *ptr) { 94 intptr_t hash = (intptr_t)ptr; 95 // Disregard the lowest 4 bits. We want all values that may be part of the 96 // same memory operation to hash to the same value and therefore use the same 97 // lock. 98 hash >>= 4; 99 // Use the next bits as the basis for the hash 100 intptr_t low = hash & SPINLOCK_MASK; 101 // Now use the high(er) set of bits to perturb the hash, so that we don't 102 // get collisions from atomic fields in a single object 103 hash >>= 16; 104 hash ^= low; 105 // Return a pointer to the word to use 106 return locks + (hash & SPINLOCK_MASK); 107 } 108 109 /// Macros for determining whether a size is lock free. Clang can not yet 110 /// codegen __atomic_is_lock_free(16), so for now we assume 16-byte values are 111 /// not lock free. 112 #define IS_LOCK_FREE_1 __atomic_is_lock_free(1) 113 #define IS_LOCK_FREE_2 __atomic_is_lock_free(2) 114 #define IS_LOCK_FREE_4 __atomic_is_lock_free(4) 115 #define IS_LOCK_FREE_8 __atomic_is_lock_free(8) 116 #define IS_LOCK_FREE_16 0 117 118 /// Macro that calls the compiler-generated lock-free versions of functions 119 /// when they exist. 120 #define LOCK_FREE_CASES() \ 121 do {\ 122 switch (size) {\ 123 case 2:\ 124 if (IS_LOCK_FREE_2) {\ 125 LOCK_FREE_ACTION(uint16_t);\ 126 }\ 127 case 4:\ 128 if (IS_LOCK_FREE_4) {\ 129 LOCK_FREE_ACTION(uint32_t);\ 130 }\ 131 case 8:\ 132 if (IS_LOCK_FREE_8) {\ 133 LOCK_FREE_ACTION(uint64_t);\ 134 }\ 135 case 16:\ 136 if (IS_LOCK_FREE_16) {\ 137 /* FIXME: __uint128_t isn't available on 32 bit platforms. 138 LOCK_FREE_ACTION(__uint128_t);*/\ 139 }\ 140 }\ 141 } while (0) 142 143 144 /// An atomic load operation. This is atomic with respect to the source 145 /// pointer only. 146 void __atomic_load_n(int size, void *src, void *dest, int model) { 147 #define LOCK_FREE_ACTION(type) \ 148 *((type*)dest) = __atomic_load((_Atomic(type)*)src, model);\ 149 return; 150 LOCK_FREE_CASES(); 151 #undef LOCK_FREE_ACTION 152 Lock *l = lock_for_pointer(src); 153 lock(l); 154 memcpy(dest, src, size); 155 unlock(l); 156 } 157 158 /// An atomic store operation. This is atomic with respect to the destination 159 /// pointer only. 160 void __atomic_store_n(int size, void *dest, void *src, int model) { 161 #define LOCK_FREE_ACTION(type) \ 162 __atomic_store((_Atomic(type)*)dest, *(type*)dest, model);\ 163 return; 164 LOCK_FREE_CASES(); 165 #undef LOCK_FREE_ACTION 166 Lock *l = lock_for_pointer(dest); 167 lock(l); 168 memcpy(dest, src, size); 169 unlock(l); 170 } 171 172 /// Atomic compare and exchange operation. If the value at *ptr is identical 173 /// to the value at *expected, then this copies value at *desired to *ptr. If 174 /// they are not, then this stores the current value from *ptr in *expected. 175 /// 176 /// This function returns 1 if the exchange takes place or 0 if it fails. 177 int __atomic_compare_exchange_n(int size, void *ptr, void *expected, 178 void *desired, int success, int failure) { 179 #define LOCK_FREE_ACTION(type) \ 180 return __atomic_compare_exchange_strong((_Atomic(type)*)ptr, (type*)expected,\ 181 *(type*)desired, success, failure) 182 LOCK_FREE_CASES(); 183 #undef LOCK_FREE_ACTION 184 Lock *l = lock_for_pointer(ptr); 185 lock(l); 186 if (memcmp(ptr, expected, size) == 0) { 187 memcpy(ptr, desired, size); 188 unlock(l); 189 return 1; 190 } 191 memcpy(expected, ptr, size); 192 unlock(l); 193 return 0; 194 } 195 196 /// Performs an atomic exchange operation between two pointers. This is atomic 197 /// with respect to the target address. 198 void __atomic_exchange_n(int size, void *ptr, void *val, void *old, int model) { 199 #define LOCK_FREE_ACTION(type) \ 200 *(type*)old = __atomic_exchange((_Atomic(type)*)ptr, *(type*)val,\ 201 model);\ 202 return; 203 LOCK_FREE_CASES(); 204 #undef LOCK_FREE_ACTION 205 Lock *l = lock_for_pointer(ptr); 206 lock(l); 207 memcpy(old, ptr, size); 208 memcpy(ptr, val, size); 209 unlock(l); 210 } 211 212 //////////////////////////////////////////////////////////////////////////////// 213 // Where the size is known at compile time, the compiler may emit calls to 214 // specialised versions of the above functions. 215 //////////////////////////////////////////////////////////////////////////////// 216 #define OPTIMISED_CASES\ 217 OPTIMISED_CASE(1, IS_LOCK_FREE_1, uint8_t)\ 218 OPTIMISED_CASE(2, IS_LOCK_FREE_2, uint16_t)\ 219 OPTIMISED_CASE(4, IS_LOCK_FREE_4, uint32_t)\ 220 OPTIMISED_CASE(8, IS_LOCK_FREE_8, uint64_t)\ 221 /* FIXME: __uint128_t isn't available on 32 bit platforms. 222 OPTIMISED_CASE(16, IS_LOCK_FREE_16, __uint128_t)*/\ 223 224 #define OPTIMISED_CASE(n, lockfree, type)\ 225 type __atomic_load_##n(type *src, int model) {\ 226 if (lockfree)\ 227 return __atomic_load((_Atomic(type)*)src, model);\ 228 Lock *l = lock_for_pointer(src);\ 229 lock(l);\ 230 type val = *src;\ 231 unlock(l);\ 232 return val;\ 233 } 234 OPTIMISED_CASES 235 #undef OPTIMISED_CASE 236 237 #define OPTIMISED_CASE(n, lockfree, type)\ 238 void __atomic_store_##n(type *dest, type val, int model) {\ 239 if (lockfree) {\ 240 __atomic_store((_Atomic(type)*)dest, val, model);\ 241 return;\ 242 }\ 243 Lock *l = lock_for_pointer(dest);\ 244 lock(l);\ 245 *dest = val;\ 246 unlock(l);\ 247 return;\ 248 } 249 OPTIMISED_CASES 250 #undef OPTIMISED_CASE 251 252 #define OPTIMISED_CASE(n, lockfree, type)\ 253 type __atomic_exchange_##n(type *dest, type val, int model) {\ 254 if (lockfree)\ 255 return __atomic_exchange((_Atomic(type)*)dest, val, model);\ 256 Lock *l = lock_for_pointer(dest);\ 257 lock(l);\ 258 type tmp = *dest;\ 259 *dest = val;\ 260 unlock(l);\ 261 return tmp;\ 262 } 263 OPTIMISED_CASES 264 #undef OPTIMISED_CASE 265 266 #define OPTIMISED_CASE(n, lockfree, type)\ 267 int __atomic_compare_exchange_##n(type *ptr, type *expected, type desired,\ 268 int success, int failure) {\ 269 if (lockfree)\ 270 return __atomic_compare_exchange_strong((_Atomic(type)*)ptr, expected, desired,\ 271 success, failure);\ 272 Lock *l = lock_for_pointer(ptr);\ 273 lock(l);\ 274 if (*ptr == *expected) {\ 275 *ptr = desired;\ 276 unlock(l);\ 277 return 1;\ 278 }\ 279 *expected = *ptr;\ 280 unlock(l);\ 281 return 0;\ 282 } 283 OPTIMISED_CASES 284 #undef OPTIMISED_CASE 285 286 //////////////////////////////////////////////////////////////////////////////// 287 // Atomic read-modify-write operations for integers of various sizes. 288 //////////////////////////////////////////////////////////////////////////////// 289 #define ATOMIC_RMW(n, lockfree, type, opname, op) \ 290 type __atomic_fetch_##opname##_##n(type *ptr, type val, int model) {\ 291 if (lockfree) \ 292 return __atomic_fetch_##opname((_Atomic(type)*)ptr, val, model);\ 293 Lock *l = lock_for_pointer(ptr);\ 294 lock(l);\ 295 type tmp = *ptr;\ 296 *ptr = tmp op val;\ 297 unlock(l);\ 298 return tmp;\ 299 } 300 301 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, add, +) 302 OPTIMISED_CASES 303 #undef OPTIMISED_CASE 304 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, sub, -) 305 OPTIMISED_CASES 306 #undef OPTIMISED_CASE 307 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, and, &) 308 OPTIMISED_CASES 309 #undef OPTIMISED_CASE 310 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, or, |) 311 OPTIMISED_CASES 312 #undef OPTIMISED_CASE 313 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, xor, ^) 314 OPTIMISED_CASES 315 #undef OPTIMISED_CASE 316
