Index: sys/arm64/include/atomic.h
===================================================================
--- sys/arm64/include/atomic.h	(revision 312643)
+++ sys/arm64/include/atomic.h	(working copy)
@@ -98,6 +98,61 @@
 ATOMIC(set,      orr)
 ATOMIC(subtract, sub)
 
+#define	ATOMIC_FCMPSET(bar, a, l)					\
+static __inline int							\
+atomic_fcmpset_##bar##32(volatile uint32_t *p, uint32_t *cmpval,	\
+    uint32_t newval)		 					\
+{									\
+	uint32_t tmp;							\
+	uint32_t _cmpval = *cmpval;					\
+	int res;							\
+									\
+	__asm __volatile(						\
+	    "1: mov      %w1, #1        \n"				\
+	    "   ld"#a"xr %w0, [%2]      \n"				\
+	    "   cmp      %w0, %w3       \n"				\
+	    "   b.ne     2f             \n"				\
+	    "   st"#l"xr %w1, %w4, [%2] \n"				\
+	    "2:"							\
+	    : "=&r"(tmp), "=&r"(res)					\
+	    : "r" (p), "r" (_cmpval), "r" (newval)			\
+	    : "cc", "memory"						\
+	);								\
+	*cmpval = tmp;							\
+									\
+	return (!res);							\
+}									\
+									\
+static __inline int							\
+atomic_fcmpset_##bar##64(volatile uint64_t *p, uint64_t *cmpval,	\
+    uint64_t newval)							\
+{									\
+	uint64_t tmp;							\
+	uint64_t _cmpval = *cmpval;					\
+	int res;							\
+									\
+	__asm __volatile(						\
+	    "1: mov      %w1, #1       \n"				\
+	    "   ld"#a"xr %0, [%2]      \n"				\
+	    "   cmp      %0, %3        \n"				\
+	    "   b.ne     2f            \n"				\
+	    "   st"#l"xr %w1, %4, [%2] \n"				\
+	    "2:"							\
+	    : "=&r"(tmp), "=&r"(res)					\
+	    : "r" (p), "r" (_cmpval), "r" (newval)			\
+	    : "cc", "memory"						\
+	);								\
+	*cmpval = tmp;							\
+									\
+	return (!res);							\
+}
+
+ATOMIC_FCMPSET(    ,  , )
+ATOMIC_FCMPSET(acq_, a, )
+ATOMIC_FCMPSET(rel_,  ,l)
+
+#undef ATOMIC_FCMPSET
+
 #define	ATOMIC_CMPSET(bar, a, l)					\
 static __inline int							\
 atomic_cmpset_##bar##32(volatile uint32_t *p, uint32_t cmpval,		\
@@ -311,6 +366,7 @@
 
 
 #define	atomic_add_int			atomic_add_32
+#define	atomic_fcmpset_int		atomic_fcmpset_32
 #define	atomic_clear_int		atomic_clear_32
 #define	atomic_cmpset_int		atomic_cmpset_32
 #define	atomic_fetchadd_int		atomic_fetchadd_32
@@ -320,6 +376,7 @@
 #define	atomic_subtract_int		atomic_subtract_32
 
 #define	atomic_add_acq_int		atomic_add_acq_32
+#define	atomic_fcmpset_acq_int		atomic_fcmpset_acq_32
 #define	atomic_clear_acq_int		atomic_clear_acq_32
 #define	atomic_cmpset_acq_int		atomic_cmpset_acq_32
 #define	atomic_load_acq_int		atomic_load_acq_32
@@ -327,6 +384,7 @@
 #define	atomic_subtract_acq_int		atomic_subtract_acq_32
 
 #define	atomic_add_rel_int		atomic_add_rel_32
+#define	atomic_fcmpset_rel_int		atomic_fcmpset_rel_32
 #define	atomic_clear_rel_int		atomic_add_rel_32
 #define	atomic_cmpset_rel_int		atomic_cmpset_rel_32
 #define	atomic_set_rel_int		atomic_set_rel_32
@@ -334,6 +392,7 @@
 #define	atomic_store_rel_int		atomic_store_rel_32
 
 #define	atomic_add_long			atomic_add_64
+#define	atomic_fcmpset_long		atomic_fcmpset_64
 #define	atomic_clear_long		atomic_clear_64
 #define	atomic_cmpset_long		atomic_cmpset_64
 #define	atomic_fetchadd_long		atomic_fetchadd_64
@@ -343,6 +402,7 @@
 #define	atomic_subtract_long		atomic_subtract_64
 
 #define	atomic_add_ptr			atomic_add_64
+#define	atomic_fcmpset_ptr		atomic_fcmpset_64
 #define	atomic_clear_ptr		atomic_clear_64
 #define	atomic_cmpset_ptr		atomic_cmpset_64
 #define	atomic_fetchadd_ptr		atomic_fetchadd_64
@@ -352,6 +412,7 @@
 #define	atomic_subtract_ptr		atomic_subtract_64
 
 #define	atomic_add_acq_long		atomic_add_acq_64
+#define	atomic_fcmpset_acq_long		atomic_fcmpset_acq_64
 #define	atomic_clear_acq_long		atomic_add_acq_64
 #define	atomic_cmpset_acq_long		atomic_cmpset_acq_64
 #define	atomic_load_acq_long		atomic_load_acq_64
@@ -359,6 +420,7 @@
 #define	atomic_subtract_acq_long	atomic_subtract_acq_64
 
 #define	atomic_add_acq_ptr		atomic_add_acq_64
+#define	atomic_fcmpset_acq_ptr		atomic_fcmpset_acq_64
 #define	atomic_clear_acq_ptr		atomic_add_acq_64
 #define	atomic_cmpset_acq_ptr		atomic_cmpset_acq_64
 #define	atomic_load_acq_ptr		atomic_load_acq_64
@@ -366,6 +428,7 @@
 #define	atomic_subtract_acq_ptr		atomic_subtract_acq_64
 
 #define	atomic_add_rel_long		atomic_add_rel_64
+#define	atomic_fcmpset_rel_long		atomic_fcmpset_rel_64
 #define	atomic_clear_rel_long		atomic_clear_rel_64
 #define	atomic_cmpset_rel_long		atomic_cmpset_rel_64
 #define	atomic_set_rel_long		atomic_set_rel_64
@@ -373,6 +436,7 @@
 #define	atomic_store_rel_long		atomic_store_rel_64
 
 #define	atomic_add_rel_ptr		atomic_add_rel_64
+#define	atomic_fcmpset_rel_ptr		atomic_fcmpset_rel_64
 #define	atomic_clear_rel_ptr		atomic_clear_rel_64
 #define	atomic_cmpset_rel_ptr		atomic_cmpset_rel_64
 #define	atomic_set_rel_ptr		atomic_set_rel_64
Index: sys/arm/include/atomic-v4.h
===================================================================
--- sys/arm/include/atomic-v4.h	(revision 312643)
+++ sys/arm/include/atomic-v4.h	(working copy)
@@ -112,6 +112,43 @@
 	__with_interrupts_disabled(*address &= ~clearmask);
 }
 
+static __inline int
+atomic_fcmpset_32(volatile u_int32_t *p, volatile u_int32_t *cmpval, volatile u_int32_t newval)
+{
+	u_int32_t ret;
+
+	__with_interrupts_disabled(
+	 {
+	 	ret = *p;
+	    	if (*p == *cmpval) {
+			*p = newval;
+			ret = 1;
+		} else {
+			*cmpval = *p;
+			ret = 0;
+		}
+	});
+	return (ret);
+}
+
+static __inline int
+atomic_fcmpset_64(volatile u_int64_t *p, volatile u_int64_t *cmpval, volatile u_int64_t newval)
+{
+	u_int64_t ret;
+
+	__with_interrupts_disabled(
+	 {
+	    	if (*p == *cmpval) {
+			*p = newval;
+			ret = 1;
+		} else {
+			*cmpval = *p;
+			ret = 0;
+		}
+	});
+	return (ret);
+}
+
 static __inline u_int32_t
 atomic_cmpset_32(volatile u_int32_t *p, volatile u_int32_t cmpval, volatile u_int32_t newval)
 {
@@ -370,6 +407,12 @@
 	return (__swp(v, p));
 }
 
+#define atomic_fcmpset_rel_32	atomic_fcmpset_32
+#define atomic_fcmpset_acq_32	atomic_fcmpset_32
+#define atomic_fcmpset_rel_64	atomic_fcmpset_64
+#define atomic_fcmpset_acq_64	atomic_fcmpset_64
+#define atomic_fcmpset_acq_long	atomic_fcmpset_long
+#define atomic_fcmpset_rel_long	atomic_fcmpset_long
 #define atomic_cmpset_rel_32	atomic_cmpset_32
 #define atomic_cmpset_acq_32	atomic_cmpset_32
 #define atomic_cmpset_rel_64	atomic_cmpset_64
@@ -421,6 +464,14 @@
 }
 
 static __inline u_long
+atomic_fcmpset_long(volatile u_long *dst, u_long *old, u_long newe)
+{
+
+	return (atomic_fcmpset_32((volatile uint32_t *)dst,
+	    (uint32_t *)old, newe));
+}
+
+static __inline u_long
 atomic_fetchadd_long(volatile u_long *p, u_long v)
 {
 
Index: sys/arm/include/atomic-v6.h
===================================================================
--- sys/arm/include/atomic-v6.h	(revision 312643)
+++ sys/arm/include/atomic-v6.h	(working copy)
@@ -190,7 +190,117 @@
 ATOMIC_ACQ_REL(clear, 64)
 ATOMIC_ACQ_REL_LONG(clear)
 
+static __inline int
+atomic_fcmpset_32(volatile uint32_t *p, uint32_t *cmpval, uint32_t newval)
+{
+	uint32_t tmp;
+	uint32_t _cmpval = *cmpval;
+	int ret;
+
+	__asm __volatile(
+	    "1: mov 	%0, #1		\n"
+	    "   ldrex	%1, [%2]	\n"
+	    "   cmp	%1, %3		\n"
+	    "   it	ne		\n"
+	    "   bne	2f		\n"
+	    "   strex	%0, %4, [%2]	\n"
+	    "2:"
+	    : "=&r" (ret), "=&r" (tmp), "+r" (p), "+r" (_cmpval), "+r" (newval)
+	    : : "cc", "memory");
+	*cmpval = tmp;
+	return (!ret);
+}
+
+static __inline uint64_t
+atomic_fcmpset_64(volatile uint64_t *p, uint64_t *cmpval, uint64_t newval)
+{
+	uint64_t tmp;
+	uint64_t _cmpval = *cmpval;
+	int ret;
+
+	__asm __volatile(
+	    "1:	mov	%[ret], #1				\n"
+	    "   ldrexd	%Q[tmp], %R[tmp], [%[ptr]]		\n"
+	    "   teq	%Q[tmp], %Q[_cmpval]			\n"
+	    "   itee	eq					\n"
+	    "   teqeq	%R[tmp], %R[_cmpval]			\n"
+	    "   bne	2f					\n"
+	    "   strexd	%[ret], %Q[newval], %R[newval], [%[ptr]]\n"
+	    "2:							\n"
+	    : [ret]    "=&r" (ret),
+	      [tmp]    "=&r" (tmp)
+	    : [ptr]    "r"   (p),
+	      [_cmpval] "r"   (_cmpval),
+	      [newval] "r"   (newval)
+	    : "cc", "memory");
+	*cmpval = tmp;
+	return (!ret);
+}
+
+static __inline u_long
+atomic_fcmpset_long(volatile u_long *p, u_long *cmpval, u_long newval)
+{
+
+	return (atomic_fcmpset_32((volatile uint32_t *)p, 
+	    (uint32_t *)cmpval, newval));
+}
+
+static __inline uint64_t
+atomic_fcmpset_acq_64(volatile uint64_t *p, uint64_t *cmpval, uint64_t newval)
+{
+	uint64_t ret;
+
+	ret = atomic_fcmpset_64(p, cmpval, newval);
+	dmb();
+	return (ret);
+}
+
+static __inline u_long
+atomic_fcmpset_acq_long(volatile u_long *p, u_long *cmpval, u_long newval)
+{
+	u_long ret;
+
+	ret = atomic_fcmpset_long(p, cmpval, newval);
+	dmb();
+	return (ret);
+}
+
 static __inline uint32_t
+atomic_fcmpset_acq_32(volatile uint32_t *p, uint32_t *cmpval, uint32_t newval)
+{
+
+	uint32_t ret;
+
+	ret = atomic_fcmpset_32(p, cmpval, newval);
+	dmb();
+	return (ret);
+}
+
+static __inline uint32_t
+atomic_fcmpset_rel_32(volatile uint32_t *p, uint32_t *cmpval, uint32_t newval)
+{
+
+	dmb();
+	return (atomic_fcmpset_32(p, cmpval, newval));
+}
+
+static __inline uint64_t
+atomic_fcmpset_rel_64(volatile uint64_t *p, uint64_t *cmpval, uint64_t newval)
+{
+
+	dmb();
+	return (atomic_fcmpset_64(p, cmpval, newval));
+}
+
+static __inline u_long
+atomic_fcmpset_rel_long(volatile u_long *p, u_long *cmpval, u_long newval)
+{
+
+	dmb();
+	return (atomic_fcmpset_long(p, cmpval, newval));
+}
+
+static __inline uint32_t
 atomic_cmpset_32(volatile uint32_t *p, uint32_t cmpval, uint32_t newval)
 {
 	uint32_t ret;
Index: sys/arm/include/atomic.h
===================================================================
--- sys/arm/include/atomic.h	(revision 312643)
+++ sys/arm/include/atomic.h	(working copy)
@@ -84,6 +84,9 @@
 #define atomic_set_ptr			atomic_set_32
 #define atomic_set_acq_ptr		atomic_set_acq_32
 #define atomic_set_rel_ptr		atomic_set_rel_32
+#define atomic_fcmpset_ptr		atomic_fcmpset_32
+#define atomic_fcmpset_rel_ptr		atomic_fcmpset_rel_32
+#define atomic_fcmpset_acq_ptr		atomic_fcmpset_acq_32
 #define atomic_cmpset_ptr		atomic_cmpset_32
 #define atomic_cmpset_acq_ptr		atomic_cmpset_acq_32
 #define atomic_cmpset_rel_ptr		atomic_cmpset_rel_32
@@ -105,6 +108,9 @@
 #define atomic_set_int			atomic_set_32
 #define atomic_set_acq_int		atomic_set_acq_32
 #define atomic_set_rel_int		atomic_set_rel_32
+#define atomic_fcmpset_int		atomic_fcmpset_32
+#define atomic_fcmpset_acq_int		atomic_fcmpset_acq_32
+#define atomic_fcmpset_rel_int		atomic_fcmpset_rel_32
 #define atomic_cmpset_int		atomic_cmpset_32
 #define atomic_cmpset_acq_int		atomic_cmpset_acq_32
 #define atomic_cmpset_rel_int		atomic_cmpset_rel_32