diff --git a/SConstruct b/SConstruct
index 592035d..25553da 100644
--- a/SConstruct
+++ b/SConstruct
@@ -54,7 +54,7 @@ env.Append(CFLAGS = [ "-Wshadow", "-Wno-typedef-redefinition" ])
 env.Append(CPPFLAGS = [ "-target", "x86_64-freebsd-freebsd-elf",
                         "-fno-builtin", "-fno-stack-protector",
                         "-fno-optimize-sibling-calls" ])
-env.Append(LINKFLAGS = [ "-no-pie" ])
+#env.Append(LINKFLAGS = [ "-no-pie" ])
 
 if (env["STRICT"] == "1"):
     env.Append(CPPFLAGS = [ "-Wformat=2", "-Wmissing-format-attribute",
diff --git a/lib/libc/posix/mman.c b/lib/libc/posix/mman.c
index 34159aa..8f51dc7 100644
--- a/lib/libc/posix/mman.c
+++ b/lib/libc/posix/mman.c
@@ -3,12 +3,8 @@
 #include <sys/types.h>
 #include <sys/mman.h>
 
-#if defined(__x86_64__)
 #include <sys/cdefs.h>
-#include <machine/amd64.h>
-#else
-#error "Unsupported Architecture!"
-#endif
+#include <machine/cpu.h>
 
 #include <syscall.h>
 
diff --git a/sys/amd64/debug.c b/sys/amd64/debug.c
index 5012510..83337b9 100644
--- a/sys/amd64/debug.c
+++ b/sys/amd64/debug.c
@@ -7,8 +7,8 @@
 #include <sys/kdebug.h>
 #include <sys/mp.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 #include <machine/atomic.h>
 #include <machine/trap.h>
 #include <machine/lapic.h>
diff --git a/sys/amd64/include/amd64.h b/sys/amd64/include/cpu.h
similarity index 99%
rename from sys/amd64/include/amd64.h
rename to sys/amd64/include/cpu.h
index d317ca0..c5c31ad 100644
--- a/sys/amd64/include/amd64.h
+++ b/sys/amd64/include/cpu.h
@@ -12,7 +12,7 @@
  * Page Tables
  */
 
-#define PGNUMMASK	0xFFFFFFFFFFFFF000ULL
+#define PGNUMMASK	0x7FFFFFFFFFFFF000ULL
 
 #define PGIDXSHIFT	9
 #define PGIDXMASK       (512 - 1)
@@ -223,7 +223,7 @@ typedef struct XSAVEArea
 #define MSR_CSTAR   0xC0000083
 #define MSR_SFMASK  0xC0000084
 
-#include "amd64op.h"
+#include "cpuop.h"
 
 #endif /* __AMD64_H__ */
 
diff --git a/sys/amd64/include/amd64op.h b/sys/amd64/include/cpuop.h
similarity index 100%
rename from sys/amd64/include/amd64op.h
rename to sys/amd64/include/cpuop.h
diff --git a/sys/amd64/include/pmap.h b/sys/amd64/include/pmap.h
index 24519c2..67909b1 100644
--- a/sys/amd64/include/pmap.h
+++ b/sys/amd64/include/pmap.h
@@ -2,7 +2,7 @@
 #ifndef __PMAP_H__
 #define __PMAP_H__
 
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 
 /*
  * +----------------------+
diff --git a/sys/amd64/include/thread.h b/sys/amd64/include/thread.h
index 3cdd7e1..f5d2bd5 100644
--- a/sys/amd64/include/thread.h
+++ b/sys/amd64/include/thread.h
@@ -2,8 +2,8 @@
 #ifndef __MACHINE_THREAD_H__
 #define __MACHINE_THREAD_H__
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 typedef struct ThreadArchStackFrame {
     uint64_t		r15;
diff --git a/sys/amd64/ioapic.c b/sys/amd64/ioapic.c
index 20bc7ac..8962e1a 100644
--- a/sys/amd64/ioapic.c
+++ b/sys/amd64/ioapic.c
@@ -9,7 +9,7 @@
 #include <sys/kdebug.h>
 #include <sys/cdefs.h>
 
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 #include <machine/pmap.h>
 #include <machine/trap.h>
 
diff --git a/sys/amd64/lapic.c b/sys/amd64/lapic.c
index c89effa..0e222d3 100644
--- a/sys/amd64/lapic.c
+++ b/sys/amd64/lapic.c
@@ -8,8 +8,8 @@
 #include <sys/kassert.h>
 #include <sys/kdebug.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 #include <machine/pmap.h>
 #include <machine/trap.h>
 
diff --git a/sys/amd64/machine.c b/sys/amd64/machine.c
index 7ab99f4..ca51465 100644
--- a/sys/amd64/machine.c
+++ b/sys/amd64/machine.c
@@ -8,7 +8,7 @@
 #include <sys/irq.h>
 #include <sys/spinlock.h>
 
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 #include <machine/ioapic.h>
 #include <machine/lapic.h>
 #include <machine/trap.h>
diff --git a/sys/amd64/mbentry.c b/sys/amd64/mbentry.c
index 5cc008f..fed1098 100644
--- a/sys/amd64/mbentry.c
+++ b/sys/amd64/mbentry.c
@@ -10,7 +10,7 @@
 
 #include "../dev/console.h"
 
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 #include <machine/pmap.h>
 #include <machine/multiboot.h>
 
diff --git a/sys/amd64/mp.c b/sys/amd64/mp.c
index d0ef06c..743eceb 100644
--- a/sys/amd64/mp.c
+++ b/sys/amd64/mp.c
@@ -10,8 +10,8 @@
 #include <sys/ktime.h>
 #include <sys/mp.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 #include <machine/pmap.h>
 #include <machine/lapic.h>
 #include <machine/mp.h>
diff --git a/sys/amd64/pci.c b/sys/amd64/pci.c
index f6b1ff9..d007752 100644
--- a/sys/amd64/pci.c
+++ b/sys/amd64/pci.c
@@ -4,8 +4,8 @@
 #include <sys/kassert.h>
 #include <sys/pci.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 #define PCI_PORT_ADDR		0xCF8
 #define PCI_PORT_DATABASE	0xCFC
diff --git a/sys/amd64/pmap.c b/sys/amd64/pmap.c
index 9cf7a84..2c94e5f 100644
--- a/sys/amd64/pmap.c
+++ b/sys/amd64/pmap.c
@@ -6,15 +6,40 @@
 #include <sys/kassert.h>
 #include <sys/kdebug.h>
 #include <sys/kmem.h>
+#include <sys/mman.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 #include <machine/mp.h>
 #include <machine/pmap.h>
 
 AS systemAS;
 AS *currentAS[MAX_CPUS];
 
+static uint64_t
+PMapProtToPTE(uint64_t flags)
+{
+    uint64_t pte = 0;
+
+    ASSERT(flags != 0);
+
+    if (flags & PROT_READ)
+	pte |= PTE_P;
+    if (flags & PROT_WRITE)
+	pte |= PTE_P|PTE_W;
+    if (flags & PROT_EXEC)
+	pte |= PTE_P;
+    else
+	pte |= PTE_NX;
+
+    ASSERT(pte != 0);
+
+    if (flags & MAP_NOCACHE)
+	pte |= PTE_PCD;
+
+    return pte;
+}
+
 void
 PMap_Init()
 {
@@ -52,11 +77,11 @@ PMap_Init()
 
     // Setup system mappings
     PMap_SystemLMap(0x0, MEM_DIRECTMAP_BASE + 0x0,
-		    3*512, 0); // 3GB RWX
+		    3*512, PROT_ALL); // 3GB RWX
     PMap_SystemLMap(0xC0000000, MEM_DIRECTMAP_BASE + 0xC0000000,
-		    512, PTE_NX|PTE_PCD); // 1GB RW + PCD
+		    512, PROT_ALL|MAP_NOCACHE); // 1GB RW + PCD
     PMap_SystemLMap(0x100000000, MEM_DIRECTMAP_BASE + 0x100000000,
-		    60*512, 0); // 60GB RWX
+		    60*512, PROT_ALL); // 60GB RWX
 
     PMap_LoadAS(&systemAS);
 
@@ -248,7 +273,7 @@ PMap_Translate(AS *space, uintptr_t va)
     if ((pte & PTE_PS) == PTE_PS) {
 	// Handle 2MB pages
 	entry = &table->entries[k];
-	return (*entry & ~(LARGE_PGMASK | PTE_NX)) + (va & LARGE_PGMASK);
+	return (*entry & PGNUMMASK) + (va & LARGE_PGMASK);
     }
     if (pte == 0) {
 	ASSERT(pte);
@@ -259,7 +284,7 @@ PMap_Translate(AS *space, uintptr_t va)
     // Handle 4KB pages
     entry = &table->entries[l];
 
-    return (*entry & ~(PGMASK | PTE_NX)) + (va & PGMASK);
+    return (*entry & PGNUMMASK) + (va & PGMASK);
 }
 
 /**
@@ -356,6 +381,7 @@ PMap_Map(AS *as, uint64_t phys, uint64_t virt, uint64_t pages, uint64_t flags)
 {
     int i;
     PageEntry *entry;
+    uint64_t pteflags = PMapProtToPTE(flags);
 
     for (i = 0; i < pages; i++) {
 	uint64_t va = virt + PGSIZE * i;
@@ -365,7 +391,7 @@ PMap_Map(AS *as, uint64_t phys, uint64_t virt, uint64_t pages, uint64_t flags)
 	    return false;
 	}
 
-	*entry = (phys + PGSIZE * i) | PTE_P | PTE_W | PTE_U | flags;
+	*entry = (phys + PGSIZE * i) | PTE_U | pteflags;
     }
 
     return true;
@@ -425,6 +451,7 @@ PMap_AllocMap(AS *as, uint64_t virt, uint64_t len, uint64_t flags)
     int i;
     uint64_t pages = (len + PGSIZE - 1) / PGSIZE;
     PageEntry *entry;
+    uint64_t pteflags = PMapProtToPTE(flags);
 
     ASSERT((virt & PGMASK) == 0);
 
@@ -438,7 +465,7 @@ PMap_AllocMap(AS *as, uint64_t virt, uint64_t len, uint64_t flags)
 
 	if ((*entry & PTE_P) != PTE_P) {
 	    void *pg = PAlloc_AllocPage();
-	    *entry = (uint64_t)DMVA2PA(pg) | PTE_P | PTE_U | flags;
+	    *entry = (uint64_t)DMVA2PA(pg) | PTE_U | pteflags;
 	}
     }
 
@@ -481,6 +508,7 @@ PMap_SystemLMap(uint64_t phys, uint64_t virt, uint64_t lpages, uint64_t flags)
 {
     int i;
     PageEntry *entry;
+    uint64_t pteflags = PMapProtToPTE(flags);
 
     for (i = 0; i < lpages; i++) {
 	uint64_t va = virt + LARGE_PGSIZE * i;
@@ -490,7 +518,7 @@ PMap_SystemLMap(uint64_t phys, uint64_t virt, uint64_t lpages, uint64_t flags)
 	    return false;
 	}
 
-	*entry = (phys + LARGE_PGSIZE * i) | PTE_P | PTE_W | PTE_PS | flags;
+	*entry = (phys + LARGE_PGSIZE * i) | PTE_PS | pteflags;
     }
 
     return true;
@@ -515,6 +543,7 @@ PMap_SystemMap(uint64_t phys, uint64_t virt, uint64_t pages, uint64_t flags)
 {
     int i;
     PageEntry *entry;
+    uint64_t pteflags = PMapProtToPTE(flags);
 
     for (i = 0; i < pages; i++) {
 	uint64_t va = virt + PGSIZE * i;
@@ -524,7 +553,7 @@ PMap_SystemMap(uint64_t phys, uint64_t virt, uint64_t pages, uint64_t flags)
 	    return false;
 	}
 
-	*entry = (phys + PGSIZE * i) | PTE_P | PTE_W | flags;
+	*entry = (phys + PGSIZE * i) | pteflags;
     }
 
     return true;
diff --git a/sys/amd64/thread.c b/sys/amd64/thread.c
index a36fe1f..ca48948 100644
--- a/sys/amd64/thread.c
+++ b/sys/amd64/thread.c
@@ -8,8 +8,8 @@
 #include <sys/kmem.h>
 #include <sys/mp.h>
 #include <sys/thread.h>
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 #include <machine/trap.h>
 #include <machine/pmap.h>
 
diff --git a/sys/amd64/time.c b/sys/amd64/time.c
index 03a7d6f..c32183d 100644
--- a/sys/amd64/time.c
+++ b/sys/amd64/time.c
@@ -6,8 +6,8 @@
 #include <sys/kdebug.h>
 #include <sys/ktime.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 uint64_t
 Time_GetTSC()
diff --git a/sys/amd64/trap.c b/sys/amd64/trap.c
index ea70440..867ab05 100644
--- a/sys/amd64/trap.c
+++ b/sys/amd64/trap.c
@@ -12,7 +12,7 @@
 #include <sys/syscall.h>
 #include <sys/mp.h>
 
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 #include <machine/lapic.h>
 #include <machine/trap.h>
 #include <machine/mp.h>
diff --git a/sys/dev/x86/debugcons.c b/sys/dev/x86/debugcons.c
index 45cf030..00716b6 100644
--- a/sys/dev/x86/debugcons.c
+++ b/sys/dev/x86/debugcons.c
@@ -7,8 +7,8 @@
 
 #include <sys/cdefs.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 void DebugConsole_Init()
 {
diff --git a/sys/dev/x86/ps2.c b/sys/dev/x86/ps2.c
index 7e93eda..e458181 100644
--- a/sys/dev/x86/ps2.c
+++ b/sys/dev/x86/ps2.c
@@ -7,8 +7,8 @@
 #include <sys/kdebug.h>
 #include <sys/irq.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 #include "../console.h"
 #include "ps2.h"
diff --git a/sys/dev/x86/vgacons.c b/sys/dev/x86/vgacons.c
index 8957a46..5d6217a 100644
--- a/sys/dev/x86/vgacons.c
+++ b/sys/dev/x86/vgacons.c
@@ -8,7 +8,7 @@
 #include <sys/cdefs.h>
 #include <sys/kdebug.h>
 
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 #include <machine/pmap.h>
 
 #include "vgacons.h"
diff --git a/sys/include/mman.h b/sys/include/mman.h
index 3f2fc14..0756f86 100644
--- a/sys/include/mman.h
+++ b/sys/include/mman.h
@@ -6,11 +6,16 @@
 #define PROT_READ	0x01
 #define PROT_WRITE	0x02
 #define PROT_EXEC	0x04
+#define PROT_ALL	(PROT_READ|PROT_WRITE|PROT_EXEC)
 
 #define MAP_FILE	0x0010
 #define MAP_ANON	0x0020
 #define MAP_FIXED	0x0040
 
+#ifdef _KERNEL
+#define MAP_NOCACHE	0x1000
+#endif
+
 
 #ifdef _KERNEL
 #else /* _KERNEL */
diff --git a/sys/kern/critical.c b/sys/kern/critical.c
index 4164b2a..6f133ab 100644
--- a/sys/kern/critical.c
+++ b/sys/kern/critical.c
@@ -8,8 +8,8 @@
 #include <sys/mp.h>
 #include <sys/spinlock.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 uint32_t lockLevel[MAX_CPUS];
 
diff --git a/sys/kern/loader.c b/sys/kern/loader.c
index d0b9777..3318474 100644
--- a/sys/kern/loader.c
+++ b/sys/kern/loader.c
@@ -13,8 +13,8 @@
 #include <sys/queue.h>
 #include <sys/disk.h>
 #include <sys/elf64.h>
+#include <sys/mman.h>
 
-#include <machine/amd64.h>
 #include <machine/trap.h>
 #include <machine/pmap.h>
 #include <sys/thread.h>
@@ -158,7 +158,7 @@ Loader_Load(Thread *thr, VNode *vn, void *buf, uint64_t len)
 	    memsz += phdr[i].p_vaddr - va;
 
 	    Log(loader, "AllocMap %016llx %08llx\n", va, memsz);
-	    if (!PMap_AllocMap(as, va, memsz, PTE_W)) {
+	    if (!PMap_AllocMap(as, va, memsz, PROT_ALL)) {
 		// XXX: Cleanup!
 		ASSERT(false);
 		return false;
@@ -166,7 +166,7 @@ Loader_Load(Thread *thr, VNode *vn, void *buf, uint64_t len)
 	}
     }
 
-    PMap_AllocMap(as, MEM_USERSPACE_STKBASE, MEM_USERSPACE_STKLEN, PTE_W);
+    PMap_AllocMap(as, MEM_USERSPACE_STKBASE, MEM_USERSPACE_STKLEN, PROT_READ|PROT_WRITE);
 
     for (i = 0; i < ehdr->e_phnum; i++)
     {
@@ -238,6 +238,7 @@ Loader_LoadInit()
      */
     PMap_LoadAS(thr->space); // Reload CR3
 
+#if defined(__x86_64__)
     /*
      * Pass in zero arguments with null pointers to init
      */
@@ -263,6 +264,9 @@ Loader_LoadInit()
     tf.rflags = RFLAGS_IF;
     tf.rdi = rsp;
     Trap_Pop(&tf);
+#elif defined(__aarch64__)
+    NOT_IMPLEMENTED();
+#endif
 
     /*
      * We should never reach this point!
diff --git a/sys/kern/palloc.c b/sys/kern/palloc.c
index 375542d..0301bf1 100644
--- a/sys/kern/palloc.c
+++ b/sys/kern/palloc.c
@@ -16,7 +16,7 @@
 #include <sys/spinlock.h>
 
 // PGSIZE
-#include <machine/amd64.h>
+#include <machine/cpu.h>
 #include <machine/pmap.h>
 
 /* 'FREEPAGE' */
diff --git a/sys/kern/spinlock.c b/sys/kern/spinlock.c
index 061ce50..84fdca2 100644
--- a/sys/kern/spinlock.c
+++ b/sys/kern/spinlock.c
@@ -15,8 +15,8 @@
 #include <sys/spinlock.h>
 
 #include <machine/atomic.h>
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 
 Spinlock lockListLock = {
     0, 0, 0, 0, 0, 0, 0,
diff --git a/sys/kern/syscall.c b/sys/kern/syscall.c
index 856ed4e..a271241 100644
--- a/sys/kern/syscall.c
+++ b/sys/kern/syscall.c
@@ -228,7 +228,7 @@ Syscall_MMap(uint64_t addr, uint64_t len, uint64_t prot)
     Thread *cur = Sched_Current();
     bool status;
 
-    status = PMap_AllocMap(cur->space, addr, len, PTE_W);
+    status = PMap_AllocMap(cur->space, addr, len, prot);
     Thread_Release(cur);
     if (!status) {
 	// XXX: Need to unmap PMap_Unmap(cur->space, addr, pgs);
diff --git a/sys/kern/thread.c b/sys/kern/thread.c
index 2d54bcc..3094156 100644
--- a/sys/kern/thread.c
+++ b/sys/kern/thread.c
@@ -15,6 +15,7 @@
 #include <sys/kdebug.h>
 #include <sys/kmem.h>
 #include <sys/ktime.h>
+#include <sys/mman.h>
 #include <sys/mp.h>
 #include <sys/spinlock.h>
 #include <sys/thread.h>
@@ -171,7 +172,7 @@ Thread_UThreadCreate(Thread *oldThr, uint64_t rip, uint64_t arg)
     proc->ustackNext += MEM_USERSPACE_STKLEN;
     Spinlock_Unlock(&proc->lock);
 
-    PMap_AllocMap(thr->space, thr->ustack, MEM_USERSPACE_STKLEN, PTE_W);
+    PMap_AllocMap(thr->space, thr->ustack, MEM_USERSPACE_STKLEN, PROT_READ|PROT_WRITE);
     // XXX: Check failure
 
     Thread_InitArch(thr);
diff --git a/sys/kern/xmem.c b/sys/kern/xmem.c
index 13b373d..d227cdf 100644
--- a/sys/kern/xmem.c
+++ b/sys/kern/xmem.c
@@ -6,9 +6,10 @@
 #include <sys/kassert.h>
 #include <sys/kdebug.h>
 #include <sys/kmem.h>
+#include <sys/mman.h>
 
-#include <machine/amd64.h>
-#include <machine/amd64op.h>
+#include <machine/cpu.h>
+#include <machine/cpuop.h>
 #include <machine/pmap.h>
 #include <machine/mp.h>
 
@@ -106,7 +107,7 @@ XMem_Allocate(XMem *xmem, uintptr_t length)
 	if (pg == NULL)
 	    return false;
 
-	PMap_SystemMap(DMVA2PA((uint64_t)pg), xmem->base + off, 1, 0);
+	PMap_SystemMap(DMVA2PA((uint64_t)pg), xmem->base + off, 1, PROT_ALL);
 
 	xmem->length += PGSIZE;
     }