libglusterfs/atomic: Improved atomic support

This patch solves a detection problem in configure.ac that prevented
that compilation detects builtin __atomic or __sync functions.

It also adds more atomic types and support for other atomic functions.

An special case has been added to support 64-bit atomics on 32-bit
systems. The solution is to fallback to the mutex solution only for
64-bit atomics, but smaller atomic types will still take advantage
of builtins if available.

Change-Id: I6b9afc7cd6e66b28a33278715583552872278801
BUG: 1510397
Signed-off-by: Xavier Hernandez <jahernan@redhat.com>

1 files changed, 425 insertions, 72 deletions

diff --git a/libglusterfs/src/atomic.h b/libglusterfs/src/atomic.h
index 71fcb1ee972..ca7f919f0d1 100644
--- a/libglusterfs/src/atomic.h
+++ b/libglusterfs/src/atomic.h
@@ -12,98 +12,451 @@
 #define _ATOMIC_H
 
 #include <inttypes.h>
+#include <stdbool.h>
+
+#include "locking.h"
+
+/* Macros used to join two arguments and generate a new macro name. */
+#define GF_ATOMIC_MACRO_1(_macro) _macro
+#define GF_ATOMIC_MACRO(_base, _name) GF_ATOMIC_MACRO_1(_base##_name)
+
+/* There's a problem on 32-bit architectures when we try to use atomic
+ * builtins with 64-bit types. Only way to solve the problem is to use
+ * a mutex to protect the access to the atomic, but we don't want to
+ * use mutexes for other smaller types that could work with the atomic
+ * builtins.
+ *
+ * So on each atomic type we add a field for the mutex if atomic operation
+ * is not supported and a dummy zero size field if it's supported. This way
+ * we can have different atomic types, some with a mutex and some without.
+ *
+ * To define these types, we use two macros:
+ *
+ *     GF_ATOMIC_MUTEX_FIELD_0 = char lk[0]
+ *     GF_ATOMIC_MUTEX_FILED_1 = gf_lock_t lk
+ *
+ * Both macros define the 'lk' field that will be used in the atomic
+ * structure. One when the atomic is supported by the architecture and
+ * another when not. We need to define the field even if it won't be
+ * used. Otherwise the compiler will return an error.
+ *
+ * Now we need to take the mutex or not depending on the existence of
+ * the mutex field in the structure. To do so we check the size of the
+ * structure, and if it's bigger than uint64_t (all structures with a
+ * mutex will be bigger), we use the mutex-based version. Otherwise we
+ * use the atomic builtin. This check is easily optimized out by the
+ * compiler, leaving a clean and efficient compiled code. */
+
+#define GF_ATOMIC_MUTEX_FIELD_0 char lk[0]
+#define GF_ATOMIC_MUTEX_FIELD_1 gf_lock_t lk
+
+/* We'll use SIZEOF_LONG to determine the architecture. 32-bit machines
+ * will have 4 here, while 64-bit machines will have 8. If additional
+ * needs or restrictions appear on other platforms, these tests can be
+ * extended to handle them. */
+
+/* GF_ATOMIC_SIZE_X macros map each type size to one of the
+ * GF_ATOMIC_MUTEX_FIELD_X macros, depending on detected conditions. */
 
 #if defined(HAVE_ATOMIC_BUILTINS) || defined(HAVE_SYNC_BUILTINS)
-/* optimized implementation, macros only */
 
-typedef struct gf_atomic_t {
-        int64_t    cnt;
-} gf_atomic_t;
+#define GF_ATOMIC_SIZE_1 GF_ATOMIC_MUTEX_FIELD_0
+#define GF_ATOMIC_SIZE_2 GF_ATOMIC_MUTEX_FIELD_0
+#define GF_ATOMIC_SIZE_4 GF_ATOMIC_MUTEX_FIELD_0
+
+#if SIZEOF_LONG >= 8
+#define GF_ATOMIC_SIZE_8 GF_ATOMIC_MUTEX_FIELD_0
+#endif
+
+#endif /* HAVE_(ATOMIC|SYNC)_BUILTINS */
+
+/* Any GF_ATOMIC_SIZE_X macro not yet defined will use the mutex version */
+#ifndef GF_ATOMIC_SIZE_1
+#define GF_ATOMIC_SIZE_1 GF_ATOMIC_MUTEX_FIELD_1
+#endif
+
+#ifndef GF_ATOMIC_SIZE_2
+#define GF_ATOMIC_SIZE_2 GF_ATOMIC_MUTEX_FIELD_1
+#endif
+
+#ifndef GF_ATOMIC_SIZE_4
+#define GF_ATOMIC_SIZE_4 GF_ATOMIC_MUTEX_FIELD_1
+#endif
+
+#ifndef GF_ATOMIC_SIZE_8
+#define GF_ATOMIC_SIZE_8 GF_ATOMIC_MUTEX_FIELD_1
+#endif
+
+/* This macro is used to define all atomic types supported. First field
+ * represents the size of the type in bytes, and the second one the name. */
+#define GF_ATOMIC_TYPE(_size, _name) \
+        typedef struct _gf_atomic_##_name##_t { \
+                GF_ATOMIC_MACRO(GF_ATOMIC_SIZE_, _size); \
+                _name##_t value; \
+        } gf_atomic_##_name##_t
+
+/* The atomic types we support */
+GF_ATOMIC_TYPE(1, int8);              /* gf_atomic_int8_t */
+GF_ATOMIC_TYPE(2, int16);             /* gf_atomic_int16_t */
+GF_ATOMIC_TYPE(4, int32);             /* gf_atomic_int32_t */
+GF_ATOMIC_TYPE(8, int64);             /* gf_atomic_int64_t */
+GF_ATOMIC_TYPE(SIZEOF_LONG, intptr);  /* gf_atomic_intptr_t */
+GF_ATOMIC_TYPE(1, uint8);             /* gf_atomic_uint8_t */
+GF_ATOMIC_TYPE(2, uint16);            /* gf_atomic_uint16_t */
+GF_ATOMIC_TYPE(4, uint32);            /* gf_atomic_uint32_t */
+GF_ATOMIC_TYPE(8, uint64);            /* gf_atomic_uint64_t */
+GF_ATOMIC_TYPE(SIZEOF_LONG, uintptr); /* gf_atomic_uintptr_t */
+
+/* Define the default atomic type as int64_t */
+#define gf_atomic_t gf_atomic_int64_t
+
+/* This macro will choose between the mutex based version and the atomic
+ * builtin version depending on the size of the atomic structure. */
+#define GF_ATOMIC_CHOOSE(_atomic, _op, _args...) \
+        ((sizeof(_atomic) > sizeof(uint64_t)) \
+            ? ({ GF_ATOMIC_MACRO(GF_ATOMIC_LOCK_, _op)(_atomic, ## _args); }) \
+            : ({ GF_ATOMIC_MACRO(GF_ATOMIC_BASE_, _op)(_atomic, ## _args); }))
+
+/* Macros to implement the mutex-based atomics. */
+#define GF_ATOMIC_OP_PREPARE(_atomic, _name) \
+        typeof(_atomic) *__atomic = &(_atomic); \
+        gf_lock_t *__lock = (gf_lock_t *)&__atomic->lk; \
+        LOCK(__lock); \
+        typeof(__atomic->value) _name = __atomic->value
+
+#define GF_ATOMIC_OP_STORE(_value) \
+        (__atomic->value = (_value))
+
+#define GF_ATOMIC_OP_RETURN(_value) \
+        ({ \
+                UNLOCK(__lock); \
+                _value; \
+        })
+
+#define GF_ATOMIC_LOCK_INIT(_atomic, _value) \
+        do { \
+                typeof(_atomic) *__atomic = &(_atomic); \
+                LOCK_INIT((gf_lock_t *)&__atomic->lk); \
+                __atomic->value = (_value); \
+        } while (0)
+
+#define GF_ATOMIC_LOCK_GET(_atomic) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_ADD(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value += (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_SUB(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value -= (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_AND(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value &= (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_OR(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value |= (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_XOR(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value ^= (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_NAND(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value = ~(__value & (_value))); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_FETCH_ADD(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value + (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_FETCH_SUB(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value - (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_FETCH_AND(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value & (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_FETCH_OR(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value | (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_FETCH_XOR(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(__value ^ (_value)); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_FETCH_NAND(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(~(__value & (_value))); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_SWAP(_atomic, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                GF_ATOMIC_OP_STORE(_value); \
+                GF_ATOMIC_OP_RETURN(__value); \
+        })
+
+#define GF_ATOMIC_LOCK_CMP_SWAP(_atomic, _expected, _value) \
+        ({ \
+                GF_ATOMIC_OP_PREPARE(_atomic, __value); \
+                bool __ret = (__value == (_expected)); \
+                if (__ret) { \
+                        GF_ATOMIC_OP_STORE(_value); \
+                } \
+                GF_ATOMIC_OP_RETURN(__ret); \
+        })
 
 #if defined(HAVE_ATOMIC_BUILTINS)
 
-/* all macros have a 'gf_atomic_t' as 1st argument */
-#define GF_ATOMIC_INIT(op, n)     __atomic_store (&(op.cnt), __ATOMIC_RELEASE)
-#define GF_ATOMIC_GET(op)         __atomic_load (&(op.cnt), __ATOMIC_ACQUIRE)
-#define GF_ATOMIC_INC(op)         __atomic_add_and_fetch (&(op.cnt), 1, \
-                                                          __ATOMIC_ACQ_REL)
-#define GF_ATOMIC_DEC(op)         __atomic_sub_and_fetch (&(op.cnt), 1, \
-                                                          __ATOMIC_ACQ_REL)
-#define GF_ATOMIC_ADD(op, n)      __atomic_add_and_fetch (&(op.cnt), n, \
-                                                          __ATOMIC_ACQ_REL)
-#define GF_ATOMIC_SUB(op, n)      __atomic_sub_and_fetch (&(op.cnt), n, \
-                                                          __ATOMIC_ACQ_REL)
-
-#else /* !HAVE_ATOMIC_BUILTINS, but HAVE_SYNC_BUILTINS */
-
-/* all macros have a 'gf_atomic_t' as 1st argument */
-#define GF_ATOMIC_INIT(op, n)     ({ op.cnt = n; __sync_synchronize (); })
-#define GF_ATOMIC_GET(op)         __sync_add_and_fetch (&(op.cnt), 0)
-#define GF_ATOMIC_INC(op)         __sync_add_and_fetch (&(op.cnt), 1)
-#define GF_ATOMIC_DEC(op)         __sync_sub_and_fetch (&(op.cnt), 1)
-#define GF_ATOMIC_ADD(op, n)      __sync_add_and_fetch (&(op.cnt), n)
-#define GF_ATOMIC_SUB(op, n)      __sync_sub_and_fetch (&(op.cnt), n)
-
-#endif /* HAVE_ATOMIC_BUILTINS || HAVE_SYNC_BUILTINS */
-
-#else /* no HAVE_(ATOMIC|SYNC)_BUILTINS */
-/* fallback implementation, using small inline functions to improve type
- * checking while compiling */
+/* If compiler supports __atomic builtins, we use them. */
 
-#include "locking.h"
+#define GF_ATOMIC_BASE_INIT(_atomic, _value) \
+        __atomic_store_n(&(_atomic).value, (_value), __ATOMIC_RELEASE)
+
+#define GF_ATOMIC_BASE_GET(_atomic) \
+        __atomic_load_n(&(_atomic).value, __ATOMIC_ACQUIRE)
+
+#define GF_ATOMIC_BASE_ADD(_atomic, _value) \
+        __atomic_add_fetch(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_SUB(_atomic, _value) \
+        __atomic_sub_fetch(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_AND(_atomic, _value) \
+        __atomic_and_fetch(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_OR(_atomic, _value) \
+        __atomic_or_fetch(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_XOR(_atomic, _value) \
+        __atomic_xor_fetch(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_NAND(_atomic, _value) \
+        __atomic_nand_fetch(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_FETCH_ADD(_atomic, _value) \
+        __atomic_fetch_add(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_FETCH_SUB(_atomic, _value) \
+        __atomic_fetch_sub(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_FETCH_AND(_atomic, _value) \
+        __atomic_fetch_and(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_FETCH_OR(_atomic, _value) \
+        __atomic_fetch_or(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_FETCH_XOR(_atomic, _value) \
+        __atomic_fetch_xor(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_FETCH_NAND(_atomic, _value) \
+        __atomic_fetch_nand(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_SWAP(_atomic, _value) \
+        __atomic_exchange_n(&(_atomic).value, (_value), __ATOMIC_ACQ_REL)
+
+#define GF_ATOMIC_BASE_CMP_SWAP(_atomic, _expected, _value) \
+        ({ \
+                typeof((_atomic).value) __expected = (_expected); \
+                __atomic_compare_exchange_n(&(_atomic).value, &__expected, \
+                                            (_value), 0, __ATOMIC_ACQ_REL, \
+                                            __ATOMIC_ACQUIRE); \
+        })
+
+#elif defined(HAVE_SYNC_BUILTINS)
+
+/* If compiler doesn't support __atomic builtins but supports __sync builtins,
+ * we use them. */
+
+#define GF_ATOMIC_BASE_INIT(_atomic, _value) \
+        do { \
+                (_atomic).value = (_value); \
+                __sync_synchronize(); \
+        } while (0)
+
+#define GF_ATOMIC_BASE_ADD(_atomic, _value) \
+        __sync_add_and_fetch(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_SUB(_atomic, _value) \
+        __sync_sub_and_fetch(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_AND(_atomic, _value) \
+        __sync_and_and_fetch(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_OR(_atomic, _value) \
+        __sync_or_and_fetch(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_XOR(_atomic, _value) \
+        __sync_xor_and_fetch(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_NAND(_atomic, _value) \
+        __sync_nand_and_fetch(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_FETCH_ADD(_atomic, _value) \
+        __sync_fetch_and_add(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_FETCH_SUB(_atomic, _value) \
+        __sync_fetch_and_sub(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_FETCH_AND(_atomic, _value) \
+        __sync_fetch_and_and(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_FETCH_OR(_atomic, _value) \
+        __sync_fetch_and_or(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_FETCH_XOR(_atomic, _value) \
+        __sync_fetch_and_xor(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_FETCH_NAND(_atomic, _value) \
+        __sync_fetch_and_nand(&(_atomic).value, (_value))
+
+#define GF_ATOMIC_BASE_SWAP(_atomic, _value) \
+        ({ \
+                __sync_synchronize(); \
+                __sync_lock_test_and_set(&(_atomic).value, (_value)); \
+        })
+
+#define GF_ATOMIC_BASE_CMP_SWAP(_atomic, _expected, _value) \
+        __sync_bool_compare_and_swap(&(_atomic).value, (_expected), (_value))
+
+#define GF_ATOMIC_BASE_GET(_atomic) GF_ATOMIC_BASE_ADD(_atomic, 0)
+
+#else /* !HAVE_ATOMIC_BUILTINS && !HAVE_SYNC_BUILTINS */
+
+/* The compiler doesn't support any atomic builtin. We fallback to the
+ * mutex-based implementation. */
+
+#define GF_ATOMIC_BASE_INIT(_atomic, _value) \
+        GF_ATOMIC_LOCK_INIT(_atomic, _value)
+
+#define GF_ATOMIC_BASE_GET(_atomic) \
+        GF_ATOMIC_LOCK_GET(_atomic)
+
+#define GF_ATOMIC_BASE_ADD(_atomic, _value) \
+        GF_ATOMIC_LOCK_ADD(_atomic, _value)
+
+#define GF_ATOMIC_BASE_SUB(_atomic, _value) \
+        GF_ATOMIC_LOCK_SUB(_atomic, _value)
+
+#define GF_ATOMIC_BASE_AND(_atomic, _value) \
+        GF_ATOMIC_LOCK_AND(_atomic, _value)
+
+#define GF_ATOMIC_BASE_OR(_atomic, _value) \
+        GF_ATOMIC_LOCK_OR(_atomic, _value)
+
+#define GF_ATOMIC_BASE_XOR(_atomic, _value) \
+        GF_ATOMIC_LOCK_XOR(_atomic, _value)
+
+#define GF_ATOMIC_BASE_NAND(_atomic, _value) \
+        GF_ATOMIC_LOCK_NAND(_atomic, _value)
+
+#define GF_ATOMIC_BASE_FETCH_ADD(_atomic, _value) \
+        GF_ATOMIC_LOCK_FETCH_ADD(_atomic, _value)
+
+#define GF_ATOMIC_BASE_FETCH_SUB(_atomic, _value) \
+        GF_ATOMIC_LOCK_FETCH_SUB(_atomic, _value)
+
+#define GF_ATOMIC_BASE_FETCH_AND(_atomic, _value) \
+        GF_ATOMIC_LOCK_FETCH_AND(_atomic, _value)
+
+#define GF_ATOMIC_BASE_FETCH_OR(_atomic, _value) \
+        GF_ATOMIC_LOCK_FETCH_OR(_atomic, _value)
+
+#define GF_ATOMIC_BASE_FETCH_XOR(_atomic, _value) \
+        GF_ATOMIC_LOCK_FETCH_XOR(_atomic, _value)
+
+#define GF_ATOMIC_BASE_FETCH_NAND(_atomic, _value) \
+        GF_ATOMIC_LOCK_FETCH_NAND(_atomic, _value)
+
+#define GF_ATOMIC_BASE_SWAP(_atomic, _value) \
+        GF_ATOMIC_LOCK_SWAP(_atomic, _value)
 
-typedef struct gf_atomic_t {
-        int64_t    cnt;
-        gf_lock_t  lk;
-} gf_atomic_t;
+#define GF_ATOMIC_BASE_CMP_SWAP(_atomic, _expected, _value) \
+        GF_ATOMIC_LOCK_CMP_SWAP(_atomic, _expected, _value)
 
+#endif /* HAVE_(ATOMIC|SYNC)_BUILTINS */
 
-static inline void
-gf_atomic_init (gf_atomic_t *op, int64_t cnt)
-{
-        LOCK_INIT (&op->lk);
-        op->cnt = cnt;
-}
+/* Here we declare the real atomic macros available to the user. */
 
+/* All macros have a 'gf_atomic_xxx' as 1st argument */
 
-static inline uint64_t
-gf_atomic_get (gf_atomic_t *op)
-{
-        uint64_t ret;
+#define GF_ATOMIC_INIT(_atomic, _value) GF_ATOMIC_CHOOSE(_atomic, INIT, _value)
+#define GF_ATOMIC_GET(_atomic)          GF_ATOMIC_CHOOSE(_atomic, GET)
+#define GF_ATOMIC_ADD(_atomic, _value)  GF_ATOMIC_CHOOSE(_atomic, ADD, _value)
+#define GF_ATOMIC_SUB(_atomic, _value)  GF_ATOMIC_CHOOSE(_atomic, SUB, _value)
+#define GF_ATOMIC_AND(_atomic, _value)  GF_ATOMIC_CHOOSE(_atomic, AND, _value)
+#define GF_ATOMIC_OR(_atomic, _value)   GF_ATOMIC_CHOOSE(_atomic, OR, _value)
+#define GF_ATOMIC_XOR(_atomic, _value)  GF_ATOMIC_CHOOSE(_atomic, XOR, _value)
+#define GF_ATOMIC_NAND(_atomic, _value) GF_ATOMIC_CHOOSE(_atomic, NAND, _value)
 
-        LOCK (&op->lk);
-        {
-               ret = op->cnt;
-        }
-        UNLOCK (&op->lk);
+#define GF_ATOMIC_FETCH_ADD(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, FETCH_ADD, _value)
 
-        return ret;
-}
+#define GF_ATOMIC_FETCH_SUB(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, FETCH_SUB, _value)
 
+#define GF_ATOMIC_FETCH_AND(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, FETCH_AND, _value)
 
-static inline int64_t
-gf_atomic_add (gf_atomic_t *op, int64_t n)
-{
-        uint64_t ret;
+#define GF_ATOMIC_FETCH_OR(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, FETCH_OR, _value)
 
-        LOCK (&op->lk);
-        {
-                op->cnt += n;
-                ret = op->cnt;
-        }
-        UNLOCK (&op->lk);
+#define GF_ATOMIC_FETCH_XOR(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, FETCH_XOR, _value)
 
-        return ret;
-}
+#define GF_ATOMIC_FETCH_NAND(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, FETCH_NAND, _value)
 
+#define GF_ATOMIC_SWAP(_atomic, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, SWAP, _value)
 
-#define GF_ATOMIC_INIT(op, cnt)   gf_atomic_init (&op, cnt)
-#define GF_ATOMIC_GET(op)         gf_atomic_get (&op)
-#define GF_ATOMIC_INC(op)         gf_atomic_add (&op, 1)
-#define GF_ATOMIC_DEC(op)         gf_atomic_add (&op, -1)
-#define GF_ATOMIC_ADD(op, n)      gf_atomic_add (&op, n)
-#define GF_ATOMIC_SUB(op, n)      gf_atomic_add (&op, -n)
+#define GF_ATOMIC_CMP_SWAP(_atomic, _expected, _value) \
+        GF_ATOMIC_CHOOSE(_atomic, CMP_SWAP, _expected, _value)
 
-#endif /* HAVE_ATOMIC_SYNC_OPS */
+#define GF_ATOMIC_INC(_atomic)       GF_ATOMIC_ADD(_atomic, 1)
+#define GF_ATOMIC_DEC(_atomic)       GF_ATOMIC_SUB(_atomic, 1)
+#define GF_ATOMIC_FETCH_INC(_atomic) GF_ATOMIC_FETCH_ADD(_atomic, 1)
+#define GF_ATOMIC_FETCH_DEC(_atomic) GF_ATOMIC_FETCH_SUB(_atomic, 1)
 
 #endif /* _ATOMIC_H */