[mirror_ovs.git] / lib / ovs-atomic-msvc.h

/*
 * Copyright (c) 2014 Nicira, Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at:
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

/* This header implements atomic operation primitives for MSVC
 * on i586 or greater platforms (32 bit). */
#ifndef IN_OVS_ATOMIC_H
#error "This header should only be included indirectly via ovs-atomic.h."
#endif

/* From msdn documentation: With Visual Studio 2003, volatile to volatile
 * references are ordered; the compiler will not re-order volatile variable
 * access. With Visual Studio 2005, the compiler also uses acquire semantics
 * for read operations on volatile variables and release semantics for write
 * operations on volatile variables (when supported by the CPU).
 *
 * Though there is no clear documentation that states that anything greater
 * than VS 2005 has the same behavior as described above, looking through MSVCs
 * C++ atomics library in VS2013 shows that the compiler still takes
 * acquire/release semantics on volatile variables. */
#define ATOMIC(TYPE) TYPE volatile

typedef enum {
    memory_order_relaxed,
    memory_order_consume,
    memory_order_acquire,
    memory_order_release,
    memory_order_acq_rel,
    memory_order_seq_cst
} memory_order;

#define ATOMIC_BOOL_LOCK_FREE 2
#define ATOMIC_CHAR_LOCK_FREE 2
#define ATOMIC_SHORT_LOCK_FREE 2
#define ATOMIC_INT_LOCK_FREE 2
#define ATOMIC_LONG_LOCK_FREE 2
#define ATOMIC_LLONG_LOCK_FREE 2
#define ATOMIC_POINTER_LOCK_FREE 2

#define IS_LOCKLESS_ATOMIC(OBJECT)                      \
    (sizeof(OBJECT) <= 8 && IS_POW2(sizeof(OBJECT)))

#define ATOMIC_VAR_INIT(VALUE) (VALUE)
#define atomic_init(OBJECT, VALUE) (*(OBJECT) = (VALUE), (void) 0)

static inline void
atomic_compiler_barrier(memory_order order)
{
    /* In case of 'memory_order_consume', it is implicitly assumed that
     * the compiler will not move instructions that have data-dependency
     * on the variable in question before the barrier. */
    if (order > memory_order_consume) {
        _ReadWriteBarrier();
    }
}

static inline void
atomic_thread_fence(memory_order order)
{
    /* x86 is strongly ordered and acquire/release semantics come
     * automatically. */
    atomic_compiler_barrier(order);
    if (order == memory_order_seq_cst) {
        MemoryBarrier();
        atomic_compiler_barrier(order);
    }
}

static inline void
atomic_signal_fence(memory_order order)
{
    atomic_compiler_barrier(order);
}

/* 1, 2 and 4 bytes loads and stores are atomic on aligned memory. In addition,
 * since the compiler automatically takes acquire and release semantics on
 * volatile variables, for any order lesser than 'memory_order_seq_cst', we
 * can directly assign or read values. */

#define atomic_store32(DST, SRC, ORDER)                                 \
    if (ORDER == memory_order_seq_cst) {                                \
        InterlockedExchange((int32_t volatile *) (DST),                 \
                               (int32_t) (SRC));                        \
    } else {                                                            \
        *(DST) = (SRC);                                                 \
    }

/* 64 bit writes are atomic on i586 if 64 bit aligned. */
#define atomic_store64(DST, SRC, ORDER)                                    \
    if (((size_t) (DST) & (sizeof *(DST) - 1))                             \
        || ORDER == memory_order_seq_cst) {                                \
        InterlockedExchange64((int64_t volatile *) (DST),                  \
                              (int64_t) (SRC));                            \
    } else {                                                               \
        *(DST) = (SRC);                                                    \
    }

/* Used for 8 and 16 bit variations. */
#define atomic_storeX(X, DST, SRC, ORDER)                               \
    if (ORDER == memory_order_seq_cst) {                                \
        InterlockedExchange##X((int##X##_t volatile *) (DST),           \
                               (int##X##_t) (SRC));                     \
    } else {                                                            \
        *(DST) = (SRC);                                                 \
    }

#define atomic_store(DST, SRC)                               \
        atomic_store_explicit(DST, SRC, memory_order_seq_cst)

#define atomic_store_explicit(DST, SRC, ORDER)                           \
    if (sizeof *(DST) == 1) {                                            \
        atomic_storeX(8, DST, SRC, ORDER)                                \
    } else if (sizeof *(DST) == 2) {                                     \
        atomic_storeX(16, DST, SRC, ORDER)                               \
    } else if (sizeof *(DST) == 4) {                                     \
        atomic_store32(DST, SRC, ORDER)                                  \
    } else if (sizeof *(DST) == 8) {                                     \
        atomic_store64(DST, SRC, ORDER)                                  \
    } else {                                                             \
        abort();                                                         \
    }

/* On x86, for 'memory_order_seq_cst', if stores are locked, the corresponding
 * reads don't need to be locked (based on the following in Intel Developers
 * manual:
 * “Locked operations are atomic with respect to all other memory operations
 * and all externally visible events. Only instruction fetch and page table
 * accesses can pass locked instructions. Locked instructions can be used to
 * synchronize data written by one processor and read by another processor.
 * For the P6 family processors, locked operations serialize all outstanding
 * load and store operations (that is, wait for them to complete). This rule
 * is also true for the Pentium 4 and Intel Xeon processors, with one
 * exception. Load operations that reference weakly ordered memory types
 * (such as the WC memory type) may not be serialized."). */

 /* For 8, 16 and 32 bit variations. */
#define atomic_readX(SRC, DST, ORDER)                                      \
    *(DST) = *(SRC);

/* 64 bit reads are atomic on i586 if 64 bit aligned. */
#define atomic_read64(SRC, DST, ORDER)                                     \
    if (((size_t) (SRC) & (sizeof *(SRC) - 1)) == 0) {                     \
        *(DST) = *(SRC);                                                   \
    } else {                                                               \
       *(DST) = InterlockedOr64((int64_t volatile *) (SRC), 0);            \
    }

#define atomic_read(SRC, DST)                               \
        atomic_read_explicit(SRC, DST, memory_order_seq_cst)

#define atomic_read_explicit(SRC, DST, ORDER)                             \
    if (sizeof *(DST) == 1 || sizeof *(DST) == 2 || sizeof *(DST) == 4) { \
        atomic_readX(SRC, DST, ORDER)                                     \
    } else if (sizeof *(DST) == 8) {                                      \
        atomic_read64(SRC, DST, ORDER)                                    \
    } else {                                                              \
        abort();                                                          \
    }

/* For add, sub, and logical operations, for 8, 16 and 64 bit data types,
 * functions for all the different memory orders does not exist
 * (though documentation exists for some of them).  The MSVC C++ library which
 * implements the c11 atomics simply calls the full memory barrier function
 * for everything in x86(see xatomic.h). So do the same here. */

/* For 8, 16 and 64 bit variations. */
#define atomic_op(OP, X, RMW, ARG, ORIG, ORDER)                         \
    atomic_##OP##_generic(X, RMW, ARG, ORIG, ORDER)

/* Arithmetic addition calls. */

#define atomic_add32(RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedExchangeAdd((int32_t volatile *) (RMW),   \
                                      (int32_t) (ARG));

/* For 8, 16 and 64 bit variations. */
#define atomic_add_generic(X, RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = _InterlockedExchangeAdd##X((int##X##_t volatile *) (RMW),     \
                                      (int##X##_t) (ARG));

#define atomic_add(RMW, ARG, ORIG)                               \
        atomic_add_explicit(RMW, ARG, ORIG, memory_order_seq_cst)

#define atomic_add_explicit(RMW, ARG, ORIG, ORDER)             \
    if (sizeof *(RMW) == 1) {                                  \
        atomic_op(add, 8, RMW, ARG, ORIG, ORDER)               \
    } else if (sizeof *(RMW) == 2) {                           \
        atomic_op(add, 16, RMW, ARG, ORIG, ORDER)              \
    } else if (sizeof *(RMW) == 4) {                           \
        atomic_add32(RMW, ARG, ORIG, ORDER)                    \
    } else if (sizeof *(RMW) == 8) {                           \
        atomic_op(add, 64, RMW, ARG, ORIG, ORDER)              \
    } else {                                                   \
        abort();                                               \
    }

/* Arithmetic subtraction calls. */

#define atomic_sub(RMW, ARG, ORIG)                             \
        atomic_add_explicit(RMW, (0 - (ARG)), ORIG, memory_order_seq_cst)

#define atomic_sub_explicit(RMW, ARG, ORIG, ORDER)           \
        atomic_add_explicit(RMW, (0 - (ARG)), ORIG, ORDER)

/* Logical 'and' calls. */

#define atomic_and32(RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedAnd((int32_t volatile *) (RMW), (int32_t) (ARG));

/* For 8, 16 and 64 bit variations. */
#define atomic_and_generic(X, RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedAnd##X((int##X##_t volatile *) (RMW),              \
                                (int##X##_t) (ARG));

#define atomic_and(RMW, ARG, ORIG)                               \
        atomic_and_explicit(RMW, ARG, ORIG, memory_order_seq_cst)

#define atomic_and_explicit(RMW, ARG, ORIG, ORDER)             \
    if (sizeof *(RMW) == 1) {                                  \
        atomic_op(and, 8, RMW, ARG, ORIG, ORDER)               \
    } else if (sizeof *(RMW) == 2) {                           \
        atomic_op(and, 16, RMW, ARG, ORIG, ORDER)              \
    } else if (sizeof *(RMW) == 4) {                           \
        atomic_and32(RMW, ARG, ORIG, ORDER)                    \
    } else if (sizeof *(RMW) == 8) {                           \
        atomic_op(and, 64, RMW, ARG, ORIG, ORDER)              \
    } else {                                                   \
        abort();                                               \
    }

/* Logical 'Or' calls. */

#define atomic_or32(RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedOr((int32_t volatile *) (RMW), (int32_t) (ARG));

/* For 8, 16 and 64 bit variations. */
#define atomic_or_generic(X, RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedOr##X((int##X##_t volatile *) (RMW),              \
                               (int##X##_t) (ARG));

#define atomic_or(RMW, ARG, ORIG)                               \
        atomic_or_explicit(RMW, ARG, ORIG, memory_order_seq_cst)

#define atomic_or_explicit(RMW, ARG, ORIG, ORDER)              \
    if (sizeof *(RMW) == 1) {                                  \
        atomic_op(or, 8, RMW, ARG, ORIG, ORDER)                \
    } else if (sizeof *(RMW) == 2) {                           \
        atomic_op(or, 16, RMW, ARG, ORIG, ORDER)               \
    } else if (sizeof *(RMW) == 4) {                           \
        atomic_or32(RMW, ARG, ORIG, ORDER)                     \
    } else if (sizeof *(RMW) == 8) {                           \
        atomic_op(or, 64, RMW, ARG, ORIG, ORDER)               \
    } else {                                                   \
        abort();                                               \
    }

/* Logical Xor calls. */

#define atomic_xor32(RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedXor((int32_t volatile *) (RMW), (int32_t) (ARG));

/* For 8, 16 and 64 bit variations. */
#define atomic_xor_generic(X, RMW, ARG, ORIG, ORDER)                        \
    *(ORIG) = InterlockedXor##X((int##X##_t volatile *) (RMW),              \
                                (int##X##_t) (ARG));

#define atomic_xor(RMW, ARG, ORIG)                               \
        atomic_xor_explicit(RMW, ARG, ORIG, memory_order_seq_cst)

#define atomic_xor_explicit(RMW, ARG, ORIG, ORDER)             \
    if (sizeof *(RMW) == 1) {                                  \
        atomic_op(xor, 8, RMW, ARG, ORIG, ORDER)               \
    } else if (sizeof *(RMW) == 2) {                           \
        atomic_op(xor, 16, RMW, ARG, ORIG, ORDER)              \
    } else if (sizeof *(RMW) == 4) {                           \
        atomic_xor32(RMW, ARG, ORIG, ORDER);                   \
    } else if (sizeof *(RMW) == 8) {                           \
        atomic_op(xor, 64, RMW, ARG, ORIG, ORDER)              \
    } else {                                                   \
        abort();                                               \
    }

#define atomic_compare_exchange_strong(DST, EXP, SRC)   \
    atomic_compare_exchange_strong_explicit(DST, EXP, SRC, \
                                            memory_order_seq_cst, \
                                            memory_order_seq_cst)

#define atomic_compare_exchange_weak atomic_compare_exchange_strong
#define atomic_compare_exchange_weak_explicit \
        atomic_compare_exchange_strong_explicit

/* MSVCs c++ compiler implements c11 atomics and looking through its
 * implementation (in xatomic.h), orders are ignored for x86 platform.
 * Do the same here. */
static inline bool
atomic_compare_exchange8(int8_t volatile *dst, int8_t *expected, int8_t src)
{
    int8_t previous = _InterlockedCompareExchange8(dst, src, *expected);
    if (previous == *expected) {
        return true;
    } else {
        *expected = previous;
        return false;
    }
}

static inline bool
atomic_compare_exchange16(int16_t volatile *dst, int16_t *expected,
                          int16_t src)
{
    int16_t previous = InterlockedCompareExchange16(dst, src, *expected);
    if (previous == *expected) {
        return true;
    } else {
        *expected = previous;
        return false;
    }
}

static inline bool
atomic_compare_exchange32(int32_t volatile *dst, int32_t *expected,
                          int32_t src)
{
    int32_t previous = InterlockedCompareExchange(dst, src, *expected);
    if (previous == *expected) {
        return true;
    } else {
        *expected = previous;
        return false;
    }
}

static inline bool
atomic_compare_exchange64(int64_t volatile *dst, int64_t *expected,
                          int64_t src)
{
    int64_t previous = InterlockedCompareExchange64(dst, src, *expected);
    if (previous == *expected) {
        return true;
    } else {
        *expected = previous;
        return false;
    }
}

static inline bool
atomic_compare_unreachable()
{
    return true;
}

#define atomic_compare_exchange_strong_explicit(DST, EXP, SRC, ORD1, ORD2)    \
    (sizeof *(DST) == 1                                                       \
     ? atomic_compare_exchange8((int8_t volatile *) (DST), (int8_t *) (EXP),  \
                                (int8_t) (SRC))                               \
     : (sizeof *(DST) == 2                                                    \
     ? atomic_compare_exchange16((int16_t volatile *) (DST),                  \
                                 (int16_t *) (EXP), (int16_t) (SRC))          \
     : (sizeof *(DST) == 4                                                    \
     ? atomic_compare_exchange32((int32_t volatile *) (DST),                  \
                                 (int32_t *) (EXP), (int32_t) (SRC))          \
     : (sizeof *(DST) == 8                                                    \
     ? atomic_compare_exchange64((int64_t volatile *) (DST),                  \
                                 (int64_t *) (EXP), (int64_t) (SRC))          \
     : ovs_fatal(0, "atomic operation with size greater than 8 bytes"),       \
       atomic_compare_unreachable()))))

\f
/* atomic_flag */

typedef ATOMIC(int32_t) atomic_flag;
#define ATOMIC_FLAG_INIT 0

#define atomic_flag_test_and_set(FLAG)                 \
    (bool) InterlockedBitTestAndSet(FLAG, 0)

#define atomic_flag_test_and_set_explicit(FLAG, ORDER) \
        atomic_flag_test_and_set(FLAG)

#define atomic_flag_clear_explicit(FLAG, ORDER) \
        atomic_flag_clear()
#define atomic_flag_clear(FLAG)                 \
    InterlockedBitTestAndReset(FLAG, 0)
Commit	Line	Data
ec2d2b5f GS	1	/*
	2	* Copyright (c) 2014 Nicira, Inc.
	3	*
	4	* Licensed under the Apache License, Version 2.0 (the "License");
	5	* you may not use this file except in compliance with the License.
	6	* You may obtain a copy of the License at:
	7	*
	8	* http://www.apache.org/licenses/LICENSE-2.0
	9	*
	10	* Unless required by applicable law or agreed to in writing, software
	11	* distributed under the License is distributed on an "AS IS" BASIS,
	12	* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	13	* See the License for the specific language governing permissions and
	14	* limitations under the License.
	15	*/
	16
	17	/* This header implements atomic operation primitives for MSVC
	18	* on i586 or greater platforms (32 bit). */
	19	#ifndef IN_OVS_ATOMIC_H
	20	#error "This header should only be included indirectly via ovs-atomic.h."
	21	#endif
	22
	23	/* From msdn documentation: With Visual Studio 2003, volatile to volatile
	24	* references are ordered; the compiler will not re-order volatile variable
	25	* access. With Visual Studio 2005, the compiler also uses acquire semantics
	26	* for read operations on volatile variables and release semantics for write
	27	* operations on volatile variables (when supported by the CPU).
	28	*
	29	* Though there is no clear documentation that states that anything greater
	30	* than VS 2005 has the same behavior as described above, looking through MSVCs
	31	* C++ atomics library in VS2013 shows that the compiler still takes
	32	* acquire/release semantics on volatile variables. */
	33	#define ATOMIC(TYPE) TYPE volatile
	34
	35	typedef enum {
	36	memory_order_relaxed,
	37	memory_order_consume,
	38	memory_order_acquire,
	39	memory_order_release,
	40	memory_order_acq_rel,
	41	memory_order_seq_cst
	42	} memory_order;
	43
	44	#define ATOMIC_BOOL_LOCK_FREE 2
	45	#define ATOMIC_CHAR_LOCK_FREE 2
	46	#define ATOMIC_SHORT_LOCK_FREE 2
	47	#define ATOMIC_INT_LOCK_FREE 2
	48	#define ATOMIC_LONG_LOCK_FREE 2
	49	#define ATOMIC_LLONG_LOCK_FREE 2
	50	#define ATOMIC_POINTER_LOCK_FREE 2
	51
	52	#define IS_LOCKLESS_ATOMIC(OBJECT) \
	53	(sizeof(OBJECT) <= 8 && IS_POW2(sizeof(OBJECT)))
	54
	55	#define ATOMIC_VAR_INIT(VALUE) (VALUE)
	56	#define atomic_init(OBJECT, VALUE) (*(OBJECT) = (VALUE), (void) 0)
	57
	58	static inline void
	59	atomic_compiler_barrier(memory_order order)
	60	{
	61	/* In case of 'memory_order_consume', it is implicitly assumed that
	62	* the compiler will not move instructions that have data-dependency
	63	* on the variable in question before the barrier. */
	64	if (order > memory_order_consume) {
65	_ReadWriteBarrier();
66	}
67	}
68
69	static inline void
70	atomic_thread_fence(memory_order order)
71	{
72	/* x86 is strongly ordered and acquire/release semantics come
73	* automatically. */
74	atomic_compiler_barrier(order);
75	if (order == memory_order_seq_cst) {
76	MemoryBarrier();
77	atomic_compiler_barrier(order);
78	}
79	}
80
81	static inline void
82	atomic_signal_fence(memory_order order)
83	{
84	atomic_compiler_barrier(order);
85	}
86
87	/* 1, 2 and 4 bytes loads and stores are atomic on aligned memory. In addition,
88	* since the compiler automatically takes acquire and release semantics on
89	* volatile variables, for any order lesser than 'memory_order_seq_cst', we
90	* can directly assign or read values. */
91
92	#define atomic_store32(DST, SRC, ORDER) \
93	if (ORDER == memory_order_seq_cst) { \
94	InterlockedExchange((int32_t volatile *) (DST), \
95	(int32_t) (SRC)); \
96	} else { \
97	*(DST) = (SRC); \
98	}
99
100	/* 64 bit writes are atomic on i586 if 64 bit aligned. */
101	#define atomic_store64(DST, SRC, ORDER) \
102	if (((size_t) (DST) & (sizeof *(DST) - 1)) \
103	\|\| ORDER == memory_order_seq_cst) { \
104	InterlockedExchange64((int64_t volatile *) (DST), \
105	(int64_t) (SRC)); \
106	} else { \
107	*(DST) = (SRC); \
108	}
109
110	/* Used for 8 and 16 bit variations. */
111	#define atomic_storeX(X, DST, SRC, ORDER) \
112	if (ORDER == memory_order_seq_cst) { \
113	InterlockedExchange##X((int##X##_t volatile *) (DST), \
114	(int##X##_t) (SRC)); \
115	} else { \
116	*(DST) = (SRC); \
117	}
118
119	#define atomic_store(DST, SRC) \
120	atomic_store_explicit(DST, SRC, memory_order_seq_cst)
121
122	#define atomic_store_explicit(DST, SRC, ORDER) \
123	if (sizeof *(DST) == 1) { \
124	atomic_storeX(8, DST, SRC, ORDER) \
125	} else if (sizeof *(DST) == 2) { \
126	atomic_storeX(16, DST, SRC, ORDER) \
127	} else if (sizeof *(DST) == 4) { \
128	atomic_store32(DST, SRC, ORDER) \
129	} else if (sizeof *(DST) == 8) { \
130	atomic_store64(DST, SRC, ORDER) \
131	} else { \
132	abort(); \
133	}
134
135	/* On x86, for 'memory_order_seq_cst', if stores are locked, the corresponding
136	* reads don't need to be locked (based on the following in Intel Developers
137	* manual:
138	* “Locked operations are atomic with respect to all other memory operations
139	* and all externally visible events. Only instruction fetch and page table
140	* accesses can pass locked instructions. Locked instructions can be used to
141	* synchronize data written by one processor and read by another processor.
142	* For the P6 family processors, locked operations serialize all outstanding
143	* load and store operations (that is, wait for them to complete). This rule
144	* is also true for the Pentium 4 and Intel Xeon processors, with one
145	* exception. Load operations that reference weakly ordered memory types
146	* (such as the WC memory type) may not be serialized."). */
147
148	/* For 8, 16 and 32 bit variations. */
149	#define atomic_readX(SRC, DST, ORDER) \
150	(DST) = (SRC);
151
152	/* 64 bit reads are atomic on i586 if 64 bit aligned. */
153	#define atomic_read64(SRC, DST, ORDER) \
154	if (((size_t) (SRC) & (sizeof *(SRC) - 1)) == 0) { \
155	(DST) = (SRC); \
156	} else { \
157	(DST) = InterlockedOr64((int64_t volatile ) (SRC), 0); \
158	}
159
160	#define atomic_read(SRC, DST) \
161	atomic_read_explicit(SRC, DST, memory_order_seq_cst)
162
163	#define atomic_read_explicit(SRC, DST, ORDER) \
164	if (sizeof (DST) == 1 \|\| sizeof (DST) == 2 \|\| sizeof *(DST) == 4) { \
165	atomic_readX(SRC, DST, ORDER) \
166	} else if (sizeof *(DST) == 8) { \
167	atomic_read64(SRC, DST, ORDER) \
168	} else { \
169	abort(); \
170	}
171
172	/* For add, sub, and logical operations, for 8, 16 and 64 bit data types,
173	* functions for all the different memory orders does not exist
174	* (though documentation exists for some of them). The MSVC C++ library which
175	* implements the c11 atomics simply calls the full memory barrier function
176	* for everything in x86(see xatomic.h). So do the same here. */
177
178	/* For 8, 16 and 64 bit variations. */
179	#define atomic_op(OP, X, RMW, ARG, ORIG, ORDER) \
180	atomic_##OP##_generic(X, RMW, ARG, ORIG, ORDER)
181
182	/* Arithmetic addition calls. */
183
184	#define atomic_add32(RMW, ARG, ORIG, ORDER) \
185	(ORIG) = InterlockedExchangeAdd((int32_t volatile ) (RMW), \
186	(int32_t) (ARG));
187
188	/* For 8, 16 and 64 bit variations. */
189	#define atomic_add_generic(X, RMW, ARG, ORIG, ORDER) \
190	(ORIG) = _InterlockedExchangeAdd##X((int##X##_t volatile ) (RMW), \
191	(int##X##_t) (ARG));
192
193	#define atomic_add(RMW, ARG, ORIG) \
194	atomic_add_explicit(RMW, ARG, ORIG, memory_order_seq_cst)
195
196	#define atomic_add_explicit(RMW, ARG, ORIG, ORDER) \
197	if (sizeof *(RMW) == 1) { \
198	atomic_op(add, 8, RMW, ARG, ORIG, ORDER) \
199	} else if (sizeof *(RMW) == 2) { \
200	atomic_op(add, 16, RMW, ARG, ORIG, ORDER) \
201	} else if (sizeof *(RMW) == 4) { \
202	atomic_add32(RMW, ARG, ORIG, ORDER) \
203	} else if (sizeof *(RMW) == 8) { \
204	atomic_op(add, 64, RMW, ARG, ORIG, ORDER) \
205	} else { \
206	abort(); \
207	}
208
209	/* Arithmetic subtraction calls. */
210
211	#define atomic_sub(RMW, ARG, ORIG) \
212	atomic_add_explicit(RMW, (0 - (ARG)), ORIG, memory_order_seq_cst)
213
214	#define atomic_sub_explicit(RMW, ARG, ORIG, ORDER) \
215	atomic_add_explicit(RMW, (0 - (ARG)), ORIG, ORDER)
216
217	/* Logical 'and' calls. */
218
219	#define atomic_and32(RMW, ARG, ORIG, ORDER) \
220	(ORIG) = InterlockedAnd((int32_t volatile ) (RMW), (int32_t) (ARG));
221
222	/* For 8, 16 and 64 bit variations. */
223	#define atomic_and_generic(X, RMW, ARG, ORIG, ORDER) \
224	(ORIG) = InterlockedAnd##X((int##X##_t volatile ) (RMW), \
225	(int##X##_t) (ARG));
226
227	#define atomic_and(RMW, ARG, ORIG) \
228	atomic_and_explicit(RMW, ARG, ORIG, memory_order_seq_cst)
229
230	#define atomic_and_explicit(RMW, ARG, ORIG, ORDER) \
231	if (sizeof *(RMW) == 1) { \
232	atomic_op(and, 8, RMW, ARG, ORIG, ORDER) \
233	} else if (sizeof *(RMW) == 2) { \
234	atomic_op(and, 16, RMW, ARG, ORIG, ORDER) \
235	} else if (sizeof *(RMW) == 4) { \
236	atomic_and32(RMW, ARG, ORIG, ORDER) \
237	} else if (sizeof *(RMW) == 8) { \
238	atomic_op(and, 64, RMW, ARG, ORIG, ORDER) \
239	} else { \
240	abort(); \
241	}
242
243	/* Logical 'Or' calls. */
244
245	#define atomic_or32(RMW, ARG, ORIG, ORDER) \
246	(ORIG) = InterlockedOr((int32_t volatile ) (RMW), (int32_t) (ARG));
247
248	/* For 8, 16 and 64 bit variations. */
249	#define atomic_or_generic(X, RMW, ARG, ORIG, ORDER) \
250	(ORIG) = InterlockedOr##X((int##X##_t volatile ) (RMW), \
251	(int##X##_t) (ARG));
252
253	#define atomic_or(RMW, ARG, ORIG) \
254	atomic_or_explicit(RMW, ARG, ORIG, memory_order_seq_cst)
255
256	#define atomic_or_explicit(RMW, ARG, ORIG, ORDER) \
257	if (sizeof *(RMW) == 1) { \
258	atomic_op(or, 8, RMW, ARG, ORIG, ORDER) \
259	} else if (sizeof *(RMW) == 2) { \
260	atomic_op(or, 16, RMW, ARG, ORIG, ORDER) \
261	} else if (sizeof *(RMW) == 4) { \
262	atomic_or32(RMW, ARG, ORIG, ORDER) \
263	} else if (sizeof *(RMW) == 8) { \
264	atomic_op(or, 64, RMW, ARG, ORIG, ORDER) \
265	} else { \
266	abort(); \
267	}
268
269	/* Logical Xor calls. */
270
271	#define atomic_xor32(RMW, ARG, ORIG, ORDER) \
272	(ORIG) = InterlockedXor((int32_t volatile ) (RMW), (int32_t) (ARG));
273
274	/* For 8, 16 and 64 bit variations. */
275	#define atomic_xor_generic(X, RMW, ARG, ORIG, ORDER) \
276	(ORIG) = InterlockedXor##X((int##X##_t volatile ) (RMW), \
277	(int##X##_t) (ARG));
278
279	#define atomic_xor(RMW, ARG, ORIG) \
280	atomic_xor_explicit(RMW, ARG, ORIG, memory_order_seq_cst)
281
282	#define atomic_xor_explicit(RMW, ARG, ORIG, ORDER) \
283	if (sizeof *(RMW) == 1) { \
284	atomic_op(xor, 8, RMW, ARG, ORIG, ORDER) \
285	} else if (sizeof *(RMW) == 2) { \
286	atomic_op(xor, 16, RMW, ARG, ORIG, ORDER) \
287	} else if (sizeof *(RMW) == 4) { \
288	atomic_xor32(RMW, ARG, ORIG, ORDER); \
289	} else if (sizeof *(RMW) == 8) { \
290	atomic_op(xor, 64, RMW, ARG, ORIG, ORDER) \
291	} else { \
292	abort(); \
293	}
294
295	#define atomic_compare_exchange_strong(DST, EXP, SRC) \
296	atomic_compare_exchange_strong_explicit(DST, EXP, SRC, \
297	memory_order_seq_cst, \
298	memory_order_seq_cst)
299
300	#define atomic_compare_exchange_weak atomic_compare_exchange_strong
301	#define atomic_compare_exchange_weak_explicit \
302	atomic_compare_exchange_strong_explicit
303
304	/* MSVCs c++ compiler implements c11 atomics and looking through its
305	* implementation (in xatomic.h), orders are ignored for x86 platform.
306	* Do the same here. */
307	static inline bool
308	atomic_compare_exchange8(int8_t volatile dst, int8_t expected, int8_t src)
309	{
310	int8_t previous = _InterlockedCompareExchange8(dst, src, *expected);
311	if (previous == *expected) {
312	return true;
313	} else {
314	*expected = previous;
315	return false;
316	}
317	}
318
319	static inline bool
320	atomic_compare_exchange16(int16_t volatile dst, int16_t expected,
321	int16_t src)
322	{
323	int16_t previous = InterlockedCompareExchange16(dst, src, *expected);
324	if (previous == *expected) {
325	return true;
326	} else {
327	*expected = previous;
328	return false;
329	}
330	}
331
332	static inline bool
333	atomic_compare_exchange32(int32_t volatile dst, int32_t expected,
334	int32_t src)
335	{
336	int32_t previous = InterlockedCompareExchange(dst, src, *expected);
337	if (previous == *expected) {
338	return true;
339	} else {
340	*expected = previous;
341	return false;
342	}
343	}
344
345	static inline bool
346	atomic_compare_exchange64(int64_t volatile dst, int64_t expected,
347	int64_t src)
348	{
349	int64_t previous = InterlockedCompareExchange64(dst, src, *expected);
350	if (previous == *expected) {
351	return true;
352	} else {
353	*expected = previous;
354	return false;
355	}
356	}
357
358	static inline bool
359	atomic_compare_unreachable()
360	{
361	return true;
362	}
363
364	#define atomic_compare_exchange_strong_explicit(DST, EXP, SRC, ORD1, ORD2) \
365	(sizeof *(DST) == 1 \
366	? atomic_compare_exchange8((int8_t volatile ) (DST), (int8_t ) (EXP), \
367	(int8_t) (SRC)) \
368	: (sizeof *(DST) == 2 \
369	? atomic_compare_exchange16((int16_t volatile *) (DST), \
370	(int16_t *) (EXP), (int16_t) (SRC)) \
371	: (sizeof *(DST) == 4 \
372	? atomic_compare_exchange32((int32_t volatile *) (DST), \
373	(int32_t *) (EXP), (int32_t) (SRC)) \
374	: (sizeof *(DST) == 8 \
375	? atomic_compare_exchange64((int64_t volatile *) (DST), \
376	(int64_t *) (EXP), (int64_t) (SRC)) \
377	: ovs_fatal(0, "atomic operation with size greater than 8 bytes"), \
378	atomic_compare_unreachable()))))
379
380	\f
381	/* atomic_flag */
382
383	typedef ATOMIC(int32_t) atomic_flag;
384	#define ATOMIC_FLAG_INIT 0
385
386	#define atomic_flag_test_and_set(FLAG) \
387	(bool) InterlockedBitTestAndSet(FLAG, 0)
388
389	#define atomic_flag_test_and_set_explicit(FLAG, ORDER) \
390	atomic_flag_test_and_set(FLAG)
391
392	#define atomic_flag_clear_explicit(FLAG, ORDER) \
393	atomic_flag_clear()
394	#define atomic_flag_clear(FLAG) \
395	InterlockedBitTestAndReset(FLAG, 0)