include/qemu/atomic.h

   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  */
  12
  13 #ifndef __QEMU_ATOMIC_H
  14 #define __QEMU_ATOMIC_H 1
  15
  16 #include "qemu/compiler.h"
  17
  18 /* For C11 atomic ops */
  19
  20 /* Compiler barrier */
  21 #define barrier()   ({ asm volatile("" ::: "memory"); (void)0; })
  22
  23 #ifndef __ATOMIC_RELAXED
  24
  25 /*
  26  * We use GCC builtin if it's available, as that can use mfence on
  27  * 32-bit as well, e.g. if built with -march=pentium-m. However, on
  28  * i386 the spec is buggy, and the implementation followed it until
  29  * 4.3 (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36793).
  30  */
  31 #if defined(__i386__) || defined(__x86_64__)
  32 #if !QEMU_GNUC_PREREQ(4, 4)
  33 #if defined __x86_64__
  34 #define smp_mb()    ({ asm volatile("mfence" ::: "memory"); (void)0; })
  35 #else
  36 #define smp_mb()    ({ asm volatile("lock; addl $0,0(%%esp) " ::: "memory"); (void)0; })
  37 #endif
  38 #endif
  39 #endif
  40
  41
  42 #ifdef __alpha__
  43 #define smp_read_barrier_depends()   asm volatile("mb":::"memory")
  44 #endif
  45
  46 #if defined(__i386__) || defined(__x86_64__) || defined(__s390x__)
  47
  48 /*
  49  * Because of the strongly ordered storage model, wmb() and rmb() are nops
  50  * here (a compiler barrier only).  QEMU doesn't do accesses to write-combining
  51  * qemu memory or non-temporal load/stores from C code.
  52  */
  53 #define smp_wmb()   barrier()
  54 #define smp_rmb()   barrier()
  55
  56 /*
  57  * __sync_lock_test_and_set() is documented to be an acquire barrier only,
  58  * but it is a full barrier at the hardware level.  Add a compiler barrier
  59  * to make it a full barrier also at the compiler level.
  60  */
  61 #define atomic_xchg(ptr, i)    (barrier(), __sync_lock_test_and_set(ptr, i))
  62
  63 /*
  64  * Load/store with Java volatile semantics.
  65  */
  66 #define atomic_mb_set(ptr, i)  ((void)atomic_xchg(ptr, i))
  67
  68 #elif defined(_ARCH_PPC)
  69
  70 /*
  71  * We use an eieio() for wmb() on powerpc.  This assumes we don't
  72  * need to order cacheable and non-cacheable stores with respect to
  73  * each other.
  74  *
  75  * smp_mb has the same problem as on x86 for not-very-new GCC
  76  * (http://patchwork.ozlabs.org/patch/126184/, Nov 2011).
  77  */
  78 #define smp_wmb()   ({ asm volatile("eieio" ::: "memory"); (void)0; })
  79 #if defined(__powerpc64__)
  80 #define smp_rmb()   ({ asm volatile("lwsync" ::: "memory"); (void)0; })
  81 #else
  82 #define smp_rmb()   ({ asm volatile("sync" ::: "memory"); (void)0; })
  83 #endif
  84 #define smp_mb()    ({ asm volatile("sync" ::: "memory"); (void)0; })
  85
  86 #endif /* _ARCH_PPC */
  87
  88 #endif /* C11 atomics */
  89
  90 /*
  91  * For (host) platforms we don't have explicit barrier definitions
  92  * for, we use the gcc __sync_synchronize() primitive to generate a
  93  * full barrier.  This should be safe on all platforms, though it may
  94  * be overkill for smp_wmb() and smp_rmb().
  95  */
  96 #ifndef smp_mb
  97 #define smp_mb()    __sync_synchronize()
  98 #endif
  99
 100 #ifndef smp_wmb
 101 #ifdef __ATOMIC_RELEASE
 102 /* __atomic_thread_fence does not include a compiler barrier; instead,
 103  * the barrier is part of __atomic_load/__atomic_store's "volatile-like"
 104  * semantics. If smp_wmb() is a no-op, absence of the barrier means that
 105  * the compiler is free to reorder stores on each side of the barrier.
 106  * Add one here, and similarly in smp_rmb() and smp_read_barrier_depends().
 107  */
 108 #define smp_wmb()   ({ barrier(); __atomic_thread_fence(__ATOMIC_RELEASE); barrier(); })
 109 #else
 110 #define smp_wmb()   __sync_synchronize()
 111 #endif
 112 #endif
 113
 114 #ifndef smp_rmb
 115 #ifdef __ATOMIC_ACQUIRE
 116 #define smp_rmb()   ({ barrier(); __atomic_thread_fence(__ATOMIC_ACQUIRE); barrier(); })
 117 #else
 118 #define smp_rmb()   __sync_synchronize()
 119 #endif
 120 #endif
 121
 122 #ifndef smp_read_barrier_depends
 123 #ifdef __ATOMIC_CONSUME
 124 #define smp_read_barrier_depends()   ({ barrier(); __atomic_thread_fence(__ATOMIC_CONSUME); barrier(); })
 125 #else
 126 #define smp_read_barrier_depends()   barrier()
 127 #endif
 128 #endif
 129
 130 #ifndef atomic_read
 131 #define atomic_read(ptr)       (*(__typeof__(*ptr) volatile*) (ptr))
 132 #endif
 133
 134 #ifndef atomic_set
 135 #define atomic_set(ptr, i)     ((*(__typeof__(*ptr) volatile*) (ptr)) = (i))
 136 #endif
 137
 138 /**
 139  * atomic_rcu_read - reads a RCU-protected pointer to a local variable
 140  * into a RCU read-side critical section. The pointer can later be safely
 141  * dereferenced within the critical section.
 142  *
 143  * This ensures that the pointer copy is invariant thorough the whole critical
 144  * section.
 145  *
 146  * Inserts memory barriers on architectures that require them (currently only
 147  * Alpha) and documents which pointers are protected by RCU.
 148  *
 149  * Unless the __ATOMIC_CONSUME memory order is available, atomic_rcu_read also
 150  * includes a compiler barrier to ensure that value-speculative optimizations
 151  * (e.g. VSS: Value Speculation Scheduling) does not perform the data read
 152  * before the pointer read by speculating the value of the pointer.  On new
 153  * enough compilers, atomic_load takes care of such concern about
 154  * dependency-breaking optimizations.
 155  *
 156  * Should match atomic_rcu_set(), atomic_xchg(), atomic_cmpxchg().
 157  */
 158 #ifndef atomic_rcu_read
 159 #ifdef __ATOMIC_CONSUME
 160 #define atomic_rcu_read(ptr)    ({                \
 161     typeof(*ptr) _val;                            \
 162      __atomic_load(ptr, &_val, __ATOMIC_CONSUME); \
 163     _val;                                         \
 164 })
 165 #else
 166 #define atomic_rcu_read(ptr)    ({                \
 167     typeof(*ptr) _val = atomic_read(ptr);         \
 168     smp_read_barrier_depends();                   \
 169     _val;                                         \
 170 })
 171 #endif
 172 #endif
 173
 174 /**
 175  * atomic_rcu_set - assigns (publicizes) a pointer to a new data structure
 176  * meant to be read by RCU read-side critical sections.
 177  *
 178  * Documents which pointers will be dereferenced by RCU read-side critical
 179  * sections and adds the required memory barriers on architectures requiring
 180  * them. It also makes sure the compiler does not reorder code initializing the
 181  * data structure before its publication.
 182  *
 183  * Should match atomic_rcu_read().
 184  */
 185 #ifndef atomic_rcu_set
 186 #ifdef __ATOMIC_RELEASE
 187 #define atomic_rcu_set(ptr, i)  do {              \
 188     typeof(*ptr) _val = (i);                      \
 189     __atomic_store(ptr, &_val, __ATOMIC_RELEASE); \
 190 } while(0)
 191 #else
 192 #define atomic_rcu_set(ptr, i)  do {              \
 193     smp_wmb();                                    \
 194     atomic_set(ptr, i);                           \
 195 } while (0)
 196 #endif
 197 #endif
 198
 199 /* These have the same semantics as Java volatile variables.
 200  * See http://gee.cs.oswego.edu/dl/jmm/cookbook.html:
 201  * "1. Issue a StoreStore barrier (wmb) before each volatile store."
 202  *  2. Issue a StoreLoad barrier after each volatile store.
 203  *     Note that you could instead issue one before each volatile load, but
 204  *     this would be slower for typical programs using volatiles in which
 205  *     reads greatly outnumber writes. Alternatively, if available, you
 206  *     can implement volatile store as an atomic instruction (for example
 207  *     XCHG on x86) and omit the barrier. This may be more efficient if
 208  *     atomic instructions are cheaper than StoreLoad barriers.
 209  *  3. Issue LoadLoad and LoadStore barriers after each volatile load."
 210  *
 211  * If you prefer to think in terms of "pairing" of memory barriers,
 212  * an atomic_mb_read pairs with an atomic_mb_set.
 213  *
 214  * And for the few ia64 lovers that exist, an atomic_mb_read is a ld.acq,
 215  * while an atomic_mb_set is a st.rel followed by a memory barrier.
 216  *
 217  * These are a bit weaker than __atomic_load/store with __ATOMIC_SEQ_CST
 218  * (see docs/atomics.txt), and I'm not sure that __ATOMIC_ACQ_REL is enough.
 219  * Just always use the barriers manually by the rules above.
 220  */
 221 #ifndef atomic_mb_read
 222 #define atomic_mb_read(ptr)    ({           \
 223     typeof(*ptr) _val = atomic_read(ptr);   \
 224     smp_rmb();                              \
 225     _val;                                   \
 226 })
 227 #endif
 228
 229 #ifndef atomic_mb_set
 230 #define atomic_mb_set(ptr, i)  do {         \
 231     smp_wmb();                              \
 232     atomic_set(ptr, i);                     \
 233     smp_mb();                               \
 234 } while (0)
 235 #endif
 236
 237 #ifndef atomic_xchg
 238 #if defined(__clang__)
 239 #define atomic_xchg(ptr, i)    __sync_swap(ptr, i)
 240 #elif defined(__ATOMIC_SEQ_CST)
 241 #define atomic_xchg(ptr, i)    ({                           \
 242     typeof(*ptr) _new = (i), _old;                          \
 243     __atomic_exchange(ptr, &_new, &_old, __ATOMIC_SEQ_CST); \
 244     _old;                                                   \
 245 })
 246 #else
 247 /* __sync_lock_test_and_set() is documented to be an acquire barrier only.  */
 248 #define atomic_xchg(ptr, i)    (smp_mb(), __sync_lock_test_and_set(ptr, i))
 249 #endif
 250 #endif
 251
 252 /* Provide shorter names for GCC atomic builtins.  */
 253 #define atomic_fetch_inc(ptr)  __sync_fetch_and_add(ptr, 1)
 254 #define atomic_fetch_dec(ptr)  __sync_fetch_and_add(ptr, -1)
 255 #define atomic_fetch_add       __sync_fetch_and_add
 256 #define atomic_fetch_sub       __sync_fetch_and_sub
 257 #define atomic_fetch_and       __sync_fetch_and_and
 258 #define atomic_fetch_or        __sync_fetch_and_or
 259 #define atomic_cmpxchg         __sync_val_compare_and_swap
 260
 261 /* And even shorter names that return void.  */
 262 #define atomic_inc(ptr)        ((void) __sync_fetch_and_add(ptr, 1))
 263 #define atomic_dec(ptr)        ((void) __sync_fetch_and_add(ptr, -1))
 264 #define atomic_add(ptr, n)     ((void) __sync_fetch_and_add(ptr, n))
 265 #define atomic_sub(ptr, n)     ((void) __sync_fetch_and_sub(ptr, n))
 266 #define atomic_and(ptr, n)     ((void) __sync_fetch_and_and(ptr, n))
 267 #define atomic_or(ptr, n)      ((void) __sync_fetch_and_or(ptr, n))
 268
 269 #endif