| 1 | /* |
| 2 | * Simple interface for atomic operations. |
| 3 | * |
| 4 | * Copyright (C) 2013 Red Hat, Inc. |
| 5 | * |
| 6 | * Author: Paolo Bonzini <pbonzini@redhat.com> |
| 7 | * |
| 8 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
| 9 | * See the COPYING file in the top-level directory. |
| 10 | * |
| 11 | * See docs/devel/atomics.rst for discussion about the guarantees each |
| 12 | * atomic primitive is meant to provide. |
| 13 | */ |
| 14 | |
| 15 | #ifndef QEMU_ATOMIC_H |
| 16 | #define QEMU_ATOMIC_H |
| 17 | |
| 18 | #include "compiler.h" |
| 19 | |
| 20 | /* Compiler barrier */ |
| 21 | #define barrier() ({ asm volatile("" ::: "memory"); (void)0; }) |
| 22 | |
| 23 | #ifndef __ATOMIC_RELAXED |
| 24 | #error "Expecting C11 atomic ops" |
| 25 | #endif |
| 26 | |
| 27 | /* Manual memory barriers |
| 28 | * |
| 29 | *__atomic_thread_fence does not include a compiler barrier; instead, |
| 30 | * the barrier is part of __atomic_load/__atomic_store's "volatile-like" |
| 31 | * semantics. If smp_wmb() is a no-op, absence of the barrier means that |
| 32 | * the compiler is free to reorder stores on each side of the barrier. |
| 33 | * Add one here, and similarly in smp_rmb() and smp_read_barrier_depends(). |
| 34 | */ |
| 35 | |
| 36 | #define smp_mb() ({ barrier(); __atomic_thread_fence(__ATOMIC_SEQ_CST); }) |
| 37 | #define smp_mb_release() ({ barrier(); __atomic_thread_fence(__ATOMIC_RELEASE); }) |
| 38 | #define smp_mb_acquire() ({ barrier(); __atomic_thread_fence(__ATOMIC_ACQUIRE); }) |
| 39 | |
| 40 | /* Most compilers currently treat consume and acquire the same, but really |
| 41 | * no processors except Alpha need a barrier here. Leave it in if |
| 42 | * using Thread Sanitizer to avoid warnings, otherwise optimize it away. |
| 43 | */ |
| 44 | #ifdef QEMU_SANITIZE_THREAD |
| 45 | #define smp_read_barrier_depends() ({ barrier(); __atomic_thread_fence(__ATOMIC_CONSUME); }) |
| 46 | #elif defined(__alpha__) |
| 47 | #define smp_read_barrier_depends() asm volatile("mb":::"memory") |
| 48 | #else |
| 49 | #define smp_read_barrier_depends() barrier() |
| 50 | #endif |
| 51 | |
| 52 | /* |
| 53 | * A signal barrier forces all pending local memory ops to be observed before |
| 54 | * a SIGSEGV is delivered to the *same* thread. In practice this is exactly |
| 55 | * the same as barrier(), but since we have the correct builtin, use it. |
| 56 | */ |
| 57 | #define signal_barrier() __atomic_signal_fence(__ATOMIC_SEQ_CST) |
| 58 | |
| 59 | /* |
| 60 | * Sanity check that the size of an atomic operation isn't "overly large". |
| 61 | * Despite the fact that e.g. x86-64 has 128-bit atomic operations, we do not |
| 62 | * want to use them because we ought not need them, and this lets us do a |
| 63 | * bit of sanity checking that other 32- and 64-bit hosts might build. |
| 64 | */ |
| 65 | #define ATOMIC_REG_SIZE sizeof(uint64_t) |
| 66 | |
| 67 | /* Weak atomic operations prevent the compiler moving other |
| 68 | * loads/stores past the atomic operation load/store. However there is |
| 69 | * no explicit memory barrier for the processor. |
| 70 | * |
| 71 | * The C11 memory model says that variables that are accessed from |
| 72 | * different threads should at least be done with __ATOMIC_RELAXED |
| 73 | * primitives or the result is undefined. Generally this has little to |
| 74 | * no effect on the generated code but not using the atomic primitives |
| 75 | * will get flagged by sanitizers as a violation. |
| 76 | */ |
| 77 | #define qatomic_read__nocheck(ptr) \ |
| 78 | __atomic_load_n(ptr, __ATOMIC_RELAXED) |
| 79 | |
| 80 | #define qatomic_read(ptr) \ |
| 81 | ({ \ |
| 82 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 83 | qatomic_read__nocheck(ptr); \ |
| 84 | }) |
| 85 | |
| 86 | #define qatomic_set__nocheck(ptr, i) \ |
| 87 | __atomic_store_n(ptr, i, __ATOMIC_RELAXED) |
| 88 | |
| 89 | #define qatomic_set(ptr, i) do { \ |
| 90 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 91 | qatomic_set__nocheck(ptr, i); \ |
| 92 | } while(0) |
| 93 | |
| 94 | /* See above: most compilers currently treat consume and acquire the |
| 95 | * same, but this slows down qatomic_rcu_read unnecessarily. |
| 96 | */ |
| 97 | #ifdef QEMU_SANITIZE_THREAD |
| 98 | #define qatomic_rcu_read__nocheck(ptr, valptr) \ |
| 99 | __atomic_load(ptr, valptr, __ATOMIC_CONSUME); |
| 100 | #else |
| 101 | #define qatomic_rcu_read__nocheck(ptr, valptr) \ |
| 102 | __atomic_load(ptr, valptr, __ATOMIC_RELAXED); \ |
| 103 | smp_read_barrier_depends(); |
| 104 | #endif |
| 105 | |
| 106 | /* |
| 107 | * Preprocessor sorcery ahead: use a different identifier for the |
| 108 | * local variable in each expansion, so we can nest macro calls |
| 109 | * without shadowing variables. |
| 110 | */ |
| 111 | #define qatomic_rcu_read_internal(ptr, _val) \ |
| 112 | ({ \ |
| 113 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 114 | typeof_strip_qual(*ptr) _val; \ |
| 115 | qatomic_rcu_read__nocheck(ptr, &_val); \ |
| 116 | _val; \ |
| 117 | }) |
| 118 | #define qatomic_rcu_read(ptr) \ |
| 119 | qatomic_rcu_read_internal((ptr), MAKE_IDENTIFIER(_val)) |
| 120 | |
| 121 | #define qatomic_rcu_set(ptr, i) do { \ |
| 122 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 123 | __atomic_store_n(ptr, i, __ATOMIC_RELEASE); \ |
| 124 | } while(0) |
| 125 | |
| 126 | #define qatomic_load_acquire(ptr) \ |
| 127 | ({ \ |
| 128 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 129 | typeof_strip_qual(*ptr) _val; \ |
| 130 | __atomic_load(ptr, &_val, __ATOMIC_ACQUIRE); \ |
| 131 | _val; \ |
| 132 | }) |
| 133 | |
| 134 | #define qatomic_store_release(ptr, i) do { \ |
| 135 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 136 | __atomic_store_n(ptr, i, __ATOMIC_RELEASE); \ |
| 137 | } while(0) |
| 138 | |
| 139 | |
| 140 | /* All the remaining operations are fully sequentially consistent */ |
| 141 | |
| 142 | #define qatomic_xchg__nocheck(ptr, i) ({ \ |
| 143 | __atomic_exchange_n(ptr, (i), __ATOMIC_SEQ_CST); \ |
| 144 | }) |
| 145 | |
| 146 | #define qatomic_xchg(ptr, i) ({ \ |
| 147 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 148 | qatomic_xchg__nocheck(ptr, i); \ |
| 149 | }) |
| 150 | |
| 151 | /* Returns the old value of '*ptr' (whether the cmpxchg failed or not) */ |
| 152 | #define qatomic_cmpxchg__nocheck(ptr, old, new) ({ \ |
| 153 | typeof_strip_qual(*ptr) _old = (old); \ |
| 154 | (void)__atomic_compare_exchange_n(ptr, &_old, new, false, \ |
| 155 | __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); \ |
| 156 | _old; \ |
| 157 | }) |
| 158 | |
| 159 | #define qatomic_cmpxchg(ptr, old, new) ({ \ |
| 160 | qemu_build_assert(sizeof(*ptr) <= ATOMIC_REG_SIZE); \ |
| 161 | qatomic_cmpxchg__nocheck(ptr, old, new); \ |
| 162 | }) |
| 163 | |
| 164 | /* Provide shorter names for GCC atomic builtins, return old value */ |
| 165 | #define qatomic_fetch_inc(ptr) __atomic_fetch_add(ptr, 1, __ATOMIC_SEQ_CST) |
| 166 | #define qatomic_fetch_dec(ptr) __atomic_fetch_sub(ptr, 1, __ATOMIC_SEQ_CST) |
| 167 | |
| 168 | #define qatomic_fetch_add(ptr, n) __atomic_fetch_add(ptr, n, __ATOMIC_SEQ_CST) |
| 169 | #define qatomic_fetch_sub(ptr, n) __atomic_fetch_sub(ptr, n, __ATOMIC_SEQ_CST) |
| 170 | #define qatomic_fetch_and(ptr, n) __atomic_fetch_and(ptr, n, __ATOMIC_SEQ_CST) |
| 171 | #define qatomic_fetch_or(ptr, n) __atomic_fetch_or(ptr, n, __ATOMIC_SEQ_CST) |
| 172 | #define qatomic_fetch_xor(ptr, n) __atomic_fetch_xor(ptr, n, __ATOMIC_SEQ_CST) |
| 173 | |
| 174 | #define qatomic_inc_fetch(ptr) __atomic_add_fetch(ptr, 1, __ATOMIC_SEQ_CST) |
| 175 | #define qatomic_dec_fetch(ptr) __atomic_sub_fetch(ptr, 1, __ATOMIC_SEQ_CST) |
| 176 | #define qatomic_add_fetch(ptr, n) __atomic_add_fetch(ptr, n, __ATOMIC_SEQ_CST) |
| 177 | #define qatomic_sub_fetch(ptr, n) __atomic_sub_fetch(ptr, n, __ATOMIC_SEQ_CST) |
| 178 | #define qatomic_and_fetch(ptr, n) __atomic_and_fetch(ptr, n, __ATOMIC_SEQ_CST) |
| 179 | #define qatomic_or_fetch(ptr, n) __atomic_or_fetch(ptr, n, __ATOMIC_SEQ_CST) |
| 180 | #define qatomic_xor_fetch(ptr, n) __atomic_xor_fetch(ptr, n, __ATOMIC_SEQ_CST) |
| 181 | |
| 182 | /* And even shorter names that return void. */ |
| 183 | #define qatomic_inc(ptr) \ |
| 184 | ((void) __atomic_fetch_add(ptr, 1, __ATOMIC_SEQ_CST)) |
| 185 | #define qatomic_dec(ptr) \ |
| 186 | ((void) __atomic_fetch_sub(ptr, 1, __ATOMIC_SEQ_CST)) |
| 187 | #define qatomic_add(ptr, n) \ |
| 188 | ((void) __atomic_fetch_add(ptr, n, __ATOMIC_SEQ_CST)) |
| 189 | #define qatomic_sub(ptr, n) \ |
| 190 | ((void) __atomic_fetch_sub(ptr, n, __ATOMIC_SEQ_CST)) |
| 191 | #define qatomic_and(ptr, n) \ |
| 192 | ((void) __atomic_fetch_and(ptr, n, __ATOMIC_SEQ_CST)) |
| 193 | #define qatomic_or(ptr, n) \ |
| 194 | ((void) __atomic_fetch_or(ptr, n, __ATOMIC_SEQ_CST)) |
| 195 | #define qatomic_xor(ptr, n) \ |
| 196 | ((void) __atomic_fetch_xor(ptr, n, __ATOMIC_SEQ_CST)) |
| 197 | |
| 198 | #define smp_wmb() smp_mb_release() |
| 199 | #define smp_rmb() smp_mb_acquire() |
| 200 | |
| 201 | /* |
| 202 | * SEQ_CST is weaker than the older __sync_* builtins and Linux |
| 203 | * kernel read-modify-write atomics. Provide a macro to obtain |
| 204 | * the same semantics. |
| 205 | */ |
| 206 | #if !defined(QEMU_SANITIZE_THREAD) && \ |
| 207 | (defined(__x86_64__) || defined(__s390x__)) |
| 208 | # define smp_mb__before_rmw() signal_barrier() |
| 209 | # define smp_mb__after_rmw() signal_barrier() |
| 210 | #else |
| 211 | # define smp_mb__before_rmw() smp_mb() |
| 212 | # define smp_mb__after_rmw() smp_mb() |
| 213 | #endif |
| 214 | |
| 215 | /* |
| 216 | * On some architectures, qatomic_set_mb is more efficient than a store |
| 217 | * plus a fence. |
| 218 | */ |
| 219 | |
| 220 | #if !defined(QEMU_SANITIZE_THREAD) && \ |
| 221 | (defined(__x86_64__) || defined(__s390x__)) |
| 222 | # define qatomic_set_mb(ptr, i) \ |
| 223 | ({ (void)qatomic_xchg(ptr, i); smp_mb__after_rmw(); }) |
| 224 | #else |
| 225 | # define qatomic_set_mb(ptr, i) \ |
| 226 | ({ qatomic_store_release(ptr, i); smp_mb(); }) |
| 227 | #endif |
| 228 | |
| 229 | #define qatomic_fetch_inc_nonzero(ptr) ({ \ |
| 230 | typeof_strip_qual(*ptr) _oldn = qatomic_read(ptr); \ |
| 231 | while (_oldn && qatomic_cmpxchg(ptr, _oldn, _oldn + 1) != _oldn) { \ |
| 232 | _oldn = qatomic_read(ptr); \ |
| 233 | } \ |
| 234 | _oldn; \ |
| 235 | }) |
| 236 | |
| 237 | #endif /* QEMU_ATOMIC_H */ |