Index: sys/arm/include/atomic-v4.h
===================================================================
--- sys/arm/include/atomic-v4.h
+++ sys/arm/include/atomic-v4.h
@@ -521,8 +521,10 @@
 #define atomic_fcmpset_rel_32	atomic_fcmpset_32
 #define atomic_fcmpset_acq_32	atomic_fcmpset_32
 #ifdef _KERNEL
+#define atomic_fcmpset_8	atomic_fcmpset_8
 #define atomic_fcmpset_rel_8	atomic_fcmpset_8
 #define atomic_fcmpset_acq_8	atomic_fcmpset_8
+#define atomic_fcmpset_16	atomic_fcmpset_16
 #define atomic_fcmpset_rel_16	atomic_fcmpset_16
 #define atomic_fcmpset_acq_16	atomic_fcmpset_16
 #define atomic_fcmpset_rel_64	atomic_fcmpset_64
@@ -533,8 +535,10 @@
 #define atomic_cmpset_rel_32	atomic_cmpset_32
 #define atomic_cmpset_acq_32	atomic_cmpset_32
 #ifdef _KERNEL
+#define atomic_cmpset_8		atomic_cmpset_8
 #define atomic_cmpset_rel_8	atomic_cmpset_8
 #define atomic_cmpset_acq_8	atomic_cmpset_8
+#define atomic_cmpset_16	atomic_cmpset_16
 #define atomic_cmpset_rel_16	atomic_cmpset_16
 #define atomic_cmpset_acq_16	atomic_cmpset_16
 #define atomic_cmpset_rel_64	atomic_cmpset_64
Index: sys/arm/include/atomic-v6.h
===================================================================
--- sys/arm/include/atomic-v6.h
+++ sys/arm/include/atomic-v6.h
@@ -245,6 +245,7 @@
 	ATOMIC_FCMPSET_CODE(ret, uint8_t, "b");
 	return (ret);
 }
+#define	atomic_fcmpset_8	atomic_fcmpset_8
 
 static __inline int
 atomic_fcmpset_acq_8(volatile uint8_t *_ptr, uint8_t *_old, uint8_t _new)
@@ -274,6 +275,7 @@
 	ATOMIC_FCMPSET_CODE(ret, uint16_t, "h");
 	return (ret);
 }
+#define	atomic_fcmpset_16	atomic_fcmpset_16
 
 static __inline int
 atomic_fcmpset_acq_16(volatile uint16_t *_ptr, uint16_t *_old, uint16_t _new)
@@ -429,6 +431,7 @@
 	ATOMIC_CMPSET_CODE(ret, "b");
 	return (ret);
 }
+#define	atomic_cmpset_8		atomic_cmpset_8
 
 static __inline int
 atomic_cmpset_acq_8(volatile uint8_t *_ptr, uint8_t _old, uint8_t _new)
@@ -458,6 +461,7 @@
 	ATOMIC_CMPSET_CODE(ret, "h");
 	return (ret);
 }
+#define	atomic_cmpset_16	atomic_cmpset_16
 
 static __inline int
 atomic_cmpset_acq_16(volatile uint16_t *_ptr, uint16_t _old, uint16_t _new)
@@ -890,6 +894,7 @@
 
 	return (atomic_testandset_32((volatile uint32_t *)p, v));
 }
+#define	atomic_testandset_long	atomic_testandset_long
 
 static __inline int
 atomic_testandset_64(volatile uint64_t *p, u_int v)
Index: sys/arm/include/atomic.h
===================================================================
--- sys/arm/include/atomic.h
+++ sys/arm/include/atomic.h
@@ -103,4 +103,6 @@
 #define atomic_store_rel_int		atomic_store_rel_32
 #define atomic_swap_int			atomic_swap_32
 
+#include <sys/_atomic_subword.h>
+
 #endif /* _MACHINE_ATOMIC_H_ */
Index: sys/kern/subr_csan.c
===================================================================
--- sys/kern/subr_csan.c
+++ sys/kern/subr_csan.c
@@ -540,7 +540,7 @@
 CSAN_ATOMIC_FUNC_CLEAR(8, uint8_t)
 CSAN_ATOMIC_FUNC_CMPSET(8, uint8_t)
 CSAN_ATOMIC_FUNC_FCMPSET(8, uint8_t)
-_CSAN_ATOMIC_FUNC_LOAD(8, uint8_t)
+CSAN_ATOMIC_FUNC_LOAD(8, uint8_t)
 CSAN_ATOMIC_FUNC_SET(8, uint8_t)
 CSAN_ATOMIC_FUNC_SUBTRACT(8, uint8_t)
 _CSAN_ATOMIC_FUNC_STORE(8, uint8_t)
@@ -556,11 +556,7 @@
 CSAN_ATOMIC_FUNC_CLEAR(16, uint16_t)
 CSAN_ATOMIC_FUNC_CMPSET(16, uint16_t)
 CSAN_ATOMIC_FUNC_FCMPSET(16, uint16_t)
-#if defined(__aarch64__)
-_CSAN_ATOMIC_FUNC_LOAD(16, uint16_t)
-#else
 CSAN_ATOMIC_FUNC_LOAD(16, uint16_t)
-#endif
 CSAN_ATOMIC_FUNC_SET(16, uint16_t)
 CSAN_ATOMIC_FUNC_SUBTRACT(16, uint16_t)
 _CSAN_ATOMIC_FUNC_STORE(16, uint16_t)
Index: sys/sys/_atomic_subword.h
===================================================================
--- sys/sys/_atomic_subword.h
+++ sys/sys/_atomic_subword.h
@@ -41,6 +41,9 @@
 #endif
 
 #include <machine/endian.h>
+#ifndef _KERNEL
+#include <stdbool.h>
+#endif
 
 #ifndef NBBY
 #define	NBBY	8
@@ -113,6 +116,7 @@
 }
 #endif
 
+#ifndef atomic_cmpset_8
 static __inline int
 atomic_cmpset_8(__volatile uint8_t *addr, uint8_t old, uint8_t val)
 {
@@ -123,7 +127,9 @@
 	return (_atomic_cmpset_masked_word(_ATOMIC_WORD_ALIGNED(addr),
 	    old << shift, val << shift, 0xff << shift));
 }
+#endif
 
+#ifndef atomic_fcmpset_8
 static __inline int
 atomic_fcmpset_8(__volatile uint8_t *addr, uint8_t *old, uint8_t val)
 {
@@ -138,7 +144,9 @@
 		*old = (wold >> shift) & 0xff;
 	return (ret);
 }
+#endif
 
+#ifndef atomic_cmpset_16
 static __inline int
 atomic_cmpset_16(__volatile uint16_t *addr, uint16_t old, uint16_t val)
 {
@@ -149,7 +157,9 @@
 	return (_atomic_cmpset_masked_word(_ATOMIC_WORD_ALIGNED(addr),
 	    old << shift, val << shift, 0xffff << shift));
 }
+#endif
 
+#ifndef atomic_fcmpset_16
 static __inline int
 atomic_fcmpset_16(__volatile uint16_t *addr, uint16_t *old, uint16_t val)
 {
@@ -164,9 +174,83 @@
 		*old = (wold >> shift) & 0xffff;
 	return (ret);
 }
+#endif
+
+#ifndef atomic_load_acq_8
+static __inline uint8_t
+atomic_load_acq_8(volatile uint8_t *p)
+{
+	int shift;
+	uint8_t ret;
+
+	shift = _ATOMIC_BYTE_SHIFT(p);
+	ret = (atomic_load_acq_32(_ATOMIC_WORD_ALIGNED(p)) >> shift) & 0xff;
+	return (ret);
+}
+#endif
+
+#ifndef atomic_load_acq_16
+static __inline uint16_t
+atomic_load_acq_16(volatile uint16_t *p)
+{
+	int shift;
+	uint16_t ret;
+
+	shift = _ATOMIC_HWORD_SHIFT(p);
+	ret = (atomic_load_acq_32(_ATOMIC_WORD_ALIGNED(p)) >> shift) &
+	    0xffff;
+	return (ret);
+}
+#endif
 
 #undef _ATOMIC_WORD_ALIGNED
 #undef _ATOMIC_BYTE_SHIFT
 #undef _ATOMIC_HWORD_SHIFT
 
+/*
+ * Provide generic testandset_long implementation based on fcmpset long
+ * primitive.  It may not be ideal for any given arch, so machine/atomic.h
+ * should define the macro atomic_testandset_long to override with an
+ * MD-specific version.
+ *
+ * (Organizationally, this isn't really subword atomics.  But atomic_common is
+ * included too early in machine/atomic.h, so it isn't a good place for derived
+ * primitives like this.)
+ */
+#ifndef atomic_testandset_long
+static __inline int
+atomic_testandset_long(volatile u_long *p, u_int v)
+{
+	u_long bit, old;
+	bool ret;
+
+	bit = (1ul << (v % (sizeof(*p) * NBBY)));
+
+	old = atomic_load_acq_long(p);
+	ret = false;
+	while (!ret && (old & bit) == 0)
+		ret = atomic_fcmpset_acq_long(p, &old, old | bit);
+
+	return (!ret);
+}
+#endif
+
+#ifndef atomic_testandclear_long
+static __inline int
+atomic_testandclear_long(volatile u_long *p, u_int v)
+{
+	u_long bit, old;
+	bool ret;
+
+	bit = (1ul << (v % (sizeof(*p) * NBBY)));
+
+	old = atomic_load_acq_long(p);
+	ret = false;
+	while (!ret && (old & bit) != 0)
+		ret = atomic_fcmpset_acq_long(p, &old, old & ~bit);
+
+	return (ret);
+}
+#endif
+
 #endif	/* _SYS__ATOMIC_SUBWORD_H_ */