diff --git a/sys/amd64/amd64/support.S b/sys/amd64/amd64/support.S
index 4e8efdfa5903..cea965f63b7b 100644
--- a/sys/amd64/amd64/support.S
+++ b/sys/amd64/amd64/support.S
@@ -342,10 +342,10 @@ ENTRY(casuword32)
 	ret
 
 /*
- * casuptr.  Compare and set user pointer.  Returns -1 or the current value.
+ * casuword.  Compare and set user word.  Returns -1 or the current value.
  *        dst = %rdi, old = %rsi, new = %rdx
  */
-ENTRY(casuptr)
+ENTRY(casuword)
 	movq	PCPU(CURPCB),%rcx
 	movq	$fusufault,PCB_ONFAULT(%rcx)
 
diff --git a/sys/arm/arm/fusu.S b/sys/arm/arm/fusu.S
index afd4663e9145..02e487083906 100644
--- a/sys/arm/arm/fusu.S
+++ b/sys/arm/arm/fusu.S
@@ -53,7 +53,7 @@ __FBSDID("$FreeBSD$");
  */
 
 ENTRY_NP(casuword32)
-ENTRY(casuptr)
+ENTRY(casuword)
 #ifdef MULTIPROCESSOR
 	/* XXX Probably not appropriate for non-Hydra SMPs */
 	stmfd	sp!, {r0, r14}
@@ -72,7 +72,7 @@ ENTRY(casuptr)
 	beq	.Lfusupcbfault
 #endif
 	stmfd	sp!, {r4, r5}
-	adr	r4, .Lcasuptrfault
+	adr	r4, .Lcasuwordfault
 	str	r4, [r3, #PCB_ONFAULT]
 	ldrt	r5, [r0]
 	cmp	r5, r1
@@ -85,10 +85,10 @@ ENTRY(casuptr)
 	RET
 
 /*
- * Handle faults from casuptr.  Clean up and return -1.
+ * Handle faults from casuword.  Clean up and return -1.
  */
 
-.Lcasuptrfault:
+.Lcasuwordfault:
 	mov	r0, #0x00000000
 	str	r0, [r3, #PCB_ONFAULT]
 	mvn	r0, #0x00000000
diff --git a/sys/i386/i386/support.s b/sys/i386/i386/support.s
index aee43a4f02a4..bacf506dee9d 100644
--- a/sys/i386/i386/support.s
+++ b/sys/i386/i386/support.s
@@ -1140,11 +1140,11 @@ fastmove_tail_fault:
 #endif /* I586_CPU && defined(DEV_NPX) */
 
 /*
- * casuptr.  Compare and set user pointer.  Returns -1 or the current value.
+ * casuword.  Compare and set user word.  Returns -1 or the current value.
  */
 
 ALTENTRY(casuword32)
-ENTRY(casuptr)
+ENTRY(casuword)
 	movl	PCPU(CURPCB),%ecx
 	movl	$fusufault,PCB_ONFAULT(%ecx)
 	movl	4(%esp),%edx			/* dst */
diff --git a/sys/ia64/ia64/support.S b/sys/ia64/ia64/support.S
index 4a212747307e..bc9ef91d280a 100644
--- a/sys/ia64/ia64/support.S
+++ b/sys/ia64/ia64/support.S
@@ -192,10 +192,10 @@ ENTRY(fusufault, 0)
 END(fusufault)
 
 /*
- * casuptr(intptr_t *p, intptr_t old, intptr_t new)
+ * casuword(u_long *p, u_long old, u_long new)
  *	Perform a compare-exchange in user space.
  */
-ENTRY(casuptr, 3)
+ENTRY(casuword, 3)
 {	.mlx
 	add		r15=PC_CURTHREAD,r13
 	movl		r14=VM_MAX_ADDRESS
@@ -239,10 +239,10 @@ ENTRY(casuptr, 3)
 	br.ret.sptk	rp
 	;;
 }
-END(casuptr)
+END(casuword)
 
 /*
- * casuword32(int32_t *p, int32_t old, int32_t new)
+ * casuword32(uint32_t *p, uint32_t old, uint32_t new)
  *	Perform a 32-bit compare-exchange in user space.
  */
 ENTRY(casuword32, 3)
diff --git a/sys/kern/kern_umtx.c b/sys/kern/kern_umtx.c
index 4c7f9d7bc38c..c74dda004a59 100644
--- a/sys/kern/kern_umtx.c
+++ b/sys/kern/kern_umtx.c
@@ -522,11 +522,11 @@ umtx_key_release(struct umtx_key *key)
  * Lock a umtx object.
  */
 static int
-_do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id, int timo)
+_do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id, int timo)
 {
 	struct umtx_q *uq;
-	intptr_t owner;
-	intptr_t old;
+	u_long owner;
+	u_long old;
 	int error = 0;
 
 	uq = td->td_umtxq;
@@ -539,7 +539,7 @@ _do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id, int timo)
 		/*
 		 * Try the uncontested case.  This should be done in userland.
 		 */
-		owner = casuptr((intptr_t *)&umtx->u_owner, UMTX_UNOWNED, id);
+		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
 
 		/* The acquire succeeded. */
 		if (owner == UMTX_UNOWNED)
@@ -551,7 +551,7 @@ _do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id, int timo)
 
 		/* If no one owns it but it is contested try to acquire it. */
 		if (owner == UMTX_CONTESTED) {
-			owner = casuptr((intptr_t *)&umtx->u_owner,
+			owner = casuword(&umtx->u_owner,
 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
 
 			if (owner == UMTX_CONTESTED)
@@ -588,8 +588,7 @@ _do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id, int timo)
 		 * either some one else has acquired the lock or it has been
 		 * released.
 		 */
-		old = casuptr((intptr_t *)&umtx->u_owner, owner,
-		    owner | UMTX_CONTESTED);
+		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
 
 		/* The address was invalid. */
 		if (old == -1) {
@@ -620,7 +619,7 @@ _do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id, int timo)
  * Lock a umtx object.
  */
 static int
-do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id,
+do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
 	struct timespec *timeout)
 {
 	struct timespec ts, ts2, ts3;
@@ -660,21 +659,18 @@ do_lock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id,
  * Unlock a umtx object.
  */
 static int
-do_unlock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id)
+do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
 {
 	struct umtx_key key;
-	intptr_t owner;
-	intptr_t old;
+	u_long owner;
+	u_long old;
 	int error;
 	int count;
 
 	/*
 	 * Make sure we own this mtx.
-	 *
-	 * XXX Need a {fu,su}ptr this is not correct on arch where
-	 * sizeof(intptr_t) != sizeof(long).
 	 */
-	owner = fuword(&umtx->u_owner);
+	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
 	if (owner == -1)
 		return (EFAULT);
 
@@ -683,8 +679,7 @@ do_unlock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id)
 
 	/* This should be done in userland */
 	if ((owner & UMTX_CONTESTED) == 0) {
-		old = casuptr((intptr_t *)&umtx->u_owner, owner,
-			UMTX_UNOWNED);
+		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
 		if (old == -1)
 			return (EFAULT);
 		if (old == owner)
@@ -707,8 +702,8 @@ do_unlock_umtx(struct thread *td, struct umtx *umtx, uintptr_t id)
 	 * there is zero or one thread only waiting for it.
 	 * Otherwise, it must be marked as contested.
 	 */
-	old = casuptr((intptr_t *)&umtx->u_owner, owner,
-			count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
+	old = casuword(&umtx->u_owner, owner,
+		count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
 	umtxq_lock(&key);
 	umtxq_signal(&key,1);
 	umtxq_unbusy(&key);
@@ -873,9 +868,6 @@ do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
 
 	/*
 	 * Make sure we own this mtx.
-	 *
-	 * XXX Need a {fu,su}ptr this is not correct on arch where
-	 * sizeof(intptr_t) != sizeof(long).
 	 */
 	owner = fuword32(m);
 	if (owner == -1)
@@ -928,13 +920,13 @@ do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
  * Fetch and compare value, sleep on the address if value is not changed.
  */
 static int
-do_wait(struct thread *td, void *addr, uintptr_t id,
+do_wait(struct thread *td, void *addr, u_long id,
 	struct timespec *timeout, int compat32)
 {
 	struct umtx_q *uq;
 	struct timespec ts, ts2, ts3;
 	struct timeval tv;
-	uintptr_t tmp;
+	u_long tmp;
 	int error = 0;
 
 	uq = td->td_umtxq;
@@ -1132,7 +1124,7 @@ do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags)
 	/*
 	 * Make sure we own this mtx.
 	 */
-	owner = fuword32(&m->m_owner);
+	owner = fuword32(__DEVOLATILE(uint32_t *, &m->m_owner));
 	if (owner == -1)
 		return (EFAULT);
 
@@ -1729,7 +1721,7 @@ do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags)
 	/*
 	 * Make sure we own this mtx.
 	 */
-	owner = fuword32(&m->m_owner);
+	owner = fuword32(__DEVOLATILE(uint32_t *, &m->m_owner));
 	if (owner == -1)
 		return (EFAULT);
 
@@ -1945,7 +1937,7 @@ do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags)
 	/*
 	 * Make sure we own this mtx.
 	 */
-	owner = fuword32(&m->m_owner);
+	owner = fuword32(__DEVOLATILE(uint32_t *, &m->m_owner));
 	if (owner == -1)
 		return (EFAULT);
 
@@ -1977,7 +1969,8 @@ do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags)
 	 * to lock the mutex, it is necessary because thread priority
 	 * has to be adjusted for such mutex.
 	 */
-	error = suword32(&m->m_owner, UMUTEX_CONTESTED);
+	error = suword32(__DEVOLATILE(uint32_t *, &m->m_owner),
+		UMUTEX_CONTESTED);
 
 	umtxq_lock(&key);
 	if (error == 0)
@@ -2040,7 +2033,8 @@ do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
 
 		if (owner == UMUTEX_CONTESTED) {
 			suword32(&m->m_ceilings[0], ceiling);
-			suword32(&m->m_owner, UMUTEX_CONTESTED);
+			suword32(__DEVOLATILE(uint32_t *, &m->m_owner),
+				UMUTEX_CONTESTED);
 			error = 0;
 			break;
 		}
diff --git a/sys/kern/syscalls.master b/sys/kern/syscalls.master
index 1dc61ca7a8b4..54e7c391ada6 100644
--- a/sys/kern/syscalls.master
+++ b/sys/kern/syscalls.master
@@ -796,7 +796,7 @@
 				    u_int length); }
 453	AUE_AUDITCTL	STD	{ int auditctl(char *path); }
 454	AUE_NULL	STD	{ int _umtx_op(void *obj, int op, \
-				    uintptr_t val, void *uaddr1, void *uaddr2); }
+				    u_long val, void *uaddr1, void *uaddr2); }
 455	AUE_NULL	STD	{ int thr_new(struct thr_param *param, \
 				    int param_size); }
 456	AUE_NULL	STD	{ int sigqueue(pid_t pid, int signum, void *value); }
diff --git a/sys/powerpc/aim/copyinout.c b/sys/powerpc/aim/copyinout.c
index 4c19cbfe9252..a5683eb7fe7b 100644
--- a/sys/powerpc/aim/copyinout.c
+++ b/sys/powerpc/aim/copyinout.c
@@ -322,23 +322,23 @@ fuword32(const void *addr)
 	return ((int32_t)fuword(addr));
 }
 
-int32_t
-casuword32(int32_t *base, int32_t oldval, int32_t newval)
+uint32_t
+casuword32(volatile uint32_t *base, uint32_t oldval, uint32_t newval)
 {
-	return (casuptr(base, oldval, newval));
+	return (casuword(base, oldval, newval));
 }
 
-intptr_t
-casuptr(intptr_t *addr, intptr_t old, intptr_t new)
+u_long
+casuword(volatile u_long *addr, u_long old, u_long new)
 {
 	struct thread *td;
 	pmap_t pm;
 	faultbuf env;
-	intptr_t *p, val;
+	u_long *p, val;
 
 	td = PCPU_GET(curthread);
 	pm = &td->td_proc->p_vmspace->vm_pmap;
-	p = (intptr_t *)((u_int)USER_ADDR + ((u_int)addr & ~SEGMENT_MASK));
+	p = (u_long *)((u_int)USER_ADDR + ((u_int)addr & ~SEGMENT_MASK));
 
 	set_user_sr(pm->pm_sr[(u_int)addr >> ADDR_SR_SHFT]);
 
diff --git a/sys/powerpc/powerpc/copyinout.c b/sys/powerpc/powerpc/copyinout.c
index 4c19cbfe9252..a5683eb7fe7b 100644
--- a/sys/powerpc/powerpc/copyinout.c
+++ b/sys/powerpc/powerpc/copyinout.c
@@ -322,23 +322,23 @@ fuword32(const void *addr)
 	return ((int32_t)fuword(addr));
 }
 
-int32_t
-casuword32(int32_t *base, int32_t oldval, int32_t newval)
+uint32_t
+casuword32(volatile uint32_t *base, uint32_t oldval, uint32_t newval)
 {
-	return (casuptr(base, oldval, newval));
+	return (casuword(base, oldval, newval));
 }
 
-intptr_t
-casuptr(intptr_t *addr, intptr_t old, intptr_t new)
+u_long
+casuword(volatile u_long *addr, u_long old, u_long new)
 {
 	struct thread *td;
 	pmap_t pm;
 	faultbuf env;
-	intptr_t *p, val;
+	u_long *p, val;
 
 	td = PCPU_GET(curthread);
 	pm = &td->td_proc->p_vmspace->vm_pmap;
-	p = (intptr_t *)((u_int)USER_ADDR + ((u_int)addr & ~SEGMENT_MASK));
+	p = (u_long *)((u_int)USER_ADDR + ((u_int)addr & ~SEGMENT_MASK));
 
 	set_user_sr(pm->pm_sr[(u_int)addr >> ADDR_SR_SHFT]);
 
diff --git a/sys/sparc64/sparc64/support.S b/sys/sparc64/sparc64/support.S
index d554f63f7963..7dda888dbdbc 100644
--- a/sys/sparc64/sparc64/support.S
+++ b/sys/sparc64/sparc64/support.S
@@ -402,8 +402,8 @@ fs_nofault_begin:
 	.set	susword, suword16
 	.set	suword, suword64
 
-	.globl	casuword32, casuptr, fuptr, suptr
-	.set	casuptr, casuword64
+	.globl	casuword32, casuword, fuptr, suptr
+	.set	casuword, casuword64
 	.set	fuptr, fuword64
 	.set	suptr, suword64
 
diff --git a/sys/sys/systm.h b/sys/sys/systm.h
index 150df1758f99..2ad88589ac06 100644
--- a/sys/sys/systm.h
+++ b/sys/sys/systm.h
@@ -203,8 +203,8 @@ int	suword(void *base, long word);
 int	suword16(void *base, int word);
 int	suword32(void *base, int32_t word);
 int	suword64(void *base, int64_t word);
-int32_t	casuword32(int32_t *base, int32_t oldval, int32_t newval);
-intptr_t casuptr(intptr_t *p, intptr_t oldval, intptr_t newval);
+uint32_t casuword32(volatile uint32_t *base, uint32_t oldval, uint32_t newval);
+u_long	 casuword(volatile u_long *p, u_long oldval, u_long newval);
 
 void	realitexpire(void *);
 
diff --git a/sys/sys/umtx.h b/sys/sys/umtx.h
index 1a501ff430b8..3156b22dec9b 100644
--- a/sys/sys/umtx.h
+++ b/sys/sys/umtx.h
@@ -41,7 +41,7 @@
 #define	UMTX_CONTESTED	LONG_MIN
 
 struct umtx {
-	uintptr_t	u_owner;	/* Owner of the mutex. */
+	volatile u_long	u_owner;	/* Owner of the mutex. */
 };
 
 #define USYNC_PROCESS_SHARED	0x0001	/* Process shared sync objs */
@@ -53,7 +53,7 @@ struct umtx {
 #define UMUTEX_PRIO_INHERIT	0x0004	/* Priority inherited mutex */
 #define UMUTEX_PRIO_PROTECT	0x0008	/* Priority protect mutex */
 struct umutex {
-	__lwpid_t	m_owner;	/* Owner of the mutex */
+	volatile __lwpid_t	m_owner;	/* Owner of the mutex */
 	uint32_t	m_flags;	/* Flags of the mutex */
 	uint32_t	m_ceilings[2];	/* Priority protect ceiling */
 	uint32_t	m_spare[4];	/* Spare space */
@@ -72,11 +72,7 @@ struct umutex {
 
 #ifndef _KERNEL
 
-/*
- * System call for userland mutex operations.
- * Bug: assume sizeof(uintptr_t) == sizeof(long)
- */
-int _umtx_op(void *obj, int op, uintptr_t val, void *uaddr, void *uaddr2);
+int _umtx_op(void *obj, int op, u_long val, void *uaddr, void *uaddr2);
 
 /*
  * Old (deprecated) userland mutex system calls.
@@ -94,33 +90,33 @@ umtx_init(struct umtx *umtx)
 	umtx->u_owner = UMTX_UNOWNED;
 }
 
-static __inline uintptr_t
+static __inline u_long
 umtx_owner(struct umtx *umtx)
 {
 	return (umtx->u_owner & ~LONG_MIN);
 }
 
 static __inline int
-umtx_lock(struct umtx *umtx, uintptr_t id)
+umtx_lock(struct umtx *umtx, u_long id)
 {
-	if (atomic_cmpset_acq_ptr(&umtx->u_owner, UMTX_UNOWNED, id) == 0)
+	if (atomic_cmpset_acq_long(&umtx->u_owner, UMTX_UNOWNED, id) == 0)
 		if (_umtx_lock(umtx) == -1)
 			return (errno);
 	return (0);
 }
 
 static __inline int
-umtx_trylock(struct umtx *umtx, uintptr_t id)
+umtx_trylock(struct umtx *umtx, u_long id)
 {
-	if (atomic_cmpset_acq_ptr(&umtx->u_owner, UMTX_UNOWNED, id) == 0)
+	if (atomic_cmpset_acq_long(&umtx->u_owner, UMTX_UNOWNED, id) == 0)
 		return (EBUSY);
 	return (0);
 }
 
 static __inline int
-umtx_timedlock(struct umtx *umtx, uintptr_t id, const struct timespec *timeout)
+umtx_timedlock(struct umtx *umtx, u_long id, const struct timespec *timeout)
 {
-	if (atomic_cmpset_acq_ptr(&umtx->u_owner, UMTX_UNOWNED, id) == 0)
+	if (atomic_cmpset_acq_long(&umtx->u_owner, UMTX_UNOWNED, id) == 0)
 		if (_umtx_op(umtx, UMTX_OP_LOCK, id, 0,
 		    __DECONST(void *, timeout)) == -1)
 			return (errno);
@@ -128,16 +124,16 @@ umtx_timedlock(struct umtx *umtx, uintptr_t id, const struct timespec *timeout)
 }
 
 static __inline int
-umtx_unlock(struct umtx *umtx, uintptr_t id)
+umtx_unlock(struct umtx *umtx, u_long id)
 {
-	if (atomic_cmpset_rel_ptr(&umtx->u_owner, id, UMTX_UNOWNED) == 0)
+	if (atomic_cmpset_rel_long(&umtx->u_owner, id, UMTX_UNOWNED) == 0)
 		if (_umtx_unlock(umtx) == -1)
 			return (errno);
 	return (0);
 }
 
 static __inline int
-umtx_wait(long *p, long val, const struct timespec *timeout)
+umtx_wait(u_long *p, long val, const struct timespec *timeout)
 {
 	if (_umtx_op(p, UMTX_OP_WAIT, val, 0,
 	    __DECONST(void *, timeout)) == -1)
@@ -147,7 +143,7 @@ umtx_wait(long *p, long val, const struct timespec *timeout)
 
 /* Wake threads waiting on a user address. */
 static __inline int
-umtx_wake(long *p, int nr_wakeup)
+umtx_wake(u_long *p, int nr_wakeup)
 {
 	if (_umtx_op(p, UMTX_OP_WAKE, nr_wakeup, 0, 0) == -1)
 		return (errno);