Merge 216da0b7 (automated)

11 files changed, 1754 insertions, 0 deletions

diff --git a/pkg/sentry/platform/ring0/pagetables/allocator.go b/pkg/sentry/platform/ring0/pagetables/allocator.go
new file mode 100644
index 000000000..23fd5c352
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/allocator.go
@@ -0,0 +1,122 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pagetables
+
+// Allocator is used to allocate and map PTEs.
+//
+// Note that allocators may be called concurrently.
+type Allocator interface {
+	// NewPTEs returns a new set of PTEs and their physical address.
+	NewPTEs() *PTEs
+
+	// PhysicalFor gives the physical address for a set of PTEs.
+	PhysicalFor(ptes *PTEs) uintptr
+
+	// LookupPTEs looks up PTEs by physical address.
+	LookupPTEs(physical uintptr) *PTEs
+
+	// FreePTEs marks a set of PTEs a freed, although they may not be available
+	// for use again until Recycle is called, below.
+	FreePTEs(ptes *PTEs)
+
+	// Recycle makes freed PTEs available for use again.
+	Recycle()
+}
+
+// RuntimeAllocator is a trivial allocator.
+type RuntimeAllocator struct {
+	// used is the set of PTEs that have been allocated. This includes any
+	// PTEs that may be in the pool below. PTEs are only freed from this
+	// map by the Drain call.
+	//
+	// This exists to prevent accidental garbage collection.
+	used map[*PTEs]struct{}
+
+	// pool is the set of free-to-use PTEs.
+	pool []*PTEs
+
+	// freed is the set of recently-freed PTEs.
+	freed []*PTEs
+}
+
+// NewRuntimeAllocator returns an allocator that uses runtime allocation.
+func NewRuntimeAllocator() *RuntimeAllocator {
+	return &RuntimeAllocator{
+		used: make(map[*PTEs]struct{}),
+	}
+}
+
+// Recycle returns freed pages to the pool.
+func (r *RuntimeAllocator) Recycle() {
+	r.pool = append(r.pool, r.freed...)
+	r.freed = r.freed[:0]
+}
+
+// Drain empties the pool.
+func (r *RuntimeAllocator) Drain() {
+	r.Recycle()
+	for i, ptes := range r.pool {
+		// Zap the entry in the underlying array to ensure that it can
+		// be properly garbage collected.
+		r.pool[i] = nil
+		// Similarly, free the reference held by the used map (these
+		// also apply for the pool entries).
+		delete(r.used, ptes)
+	}
+	r.pool = r.pool[:0]
+}
+
+// NewPTEs implements Allocator.NewPTEs.
+//
+// Note that the "physical" address here is actually the virtual address of the
+// PTEs structure. The entries are tracked only to avoid garbage collection.
+//
+// This is guaranteed not to split as long as the pool is sufficiently full.
+//
+//go:nosplit
+func (r *RuntimeAllocator) NewPTEs() *PTEs {
+	// Pull from the pool if we can.
+	if len(r.pool) > 0 {
+		ptes := r.pool[len(r.pool)-1]
+		r.pool = r.pool[:len(r.pool)-1]
+		return ptes
+	}
+
+	// Allocate a new entry.
+	ptes := newAlignedPTEs()
+	r.used[ptes] = struct{}{}
+	return ptes
+}
+
+// PhysicalFor returns the physical address for the given PTEs.
+//
+//go:nosplit
+func (r *RuntimeAllocator) PhysicalFor(ptes *PTEs) uintptr {
+	return physicalFor(ptes)
+}
+
+// LookupPTEs implements Allocator.LookupPTEs.
+//
+//go:nosplit
+func (r *RuntimeAllocator) LookupPTEs(physical uintptr) *PTEs {
+	return fromPhysical(physical)
+}
+
+// FreePTEs implements Allocator.FreePTEs.
+//
+//go:nosplit
+func (r *RuntimeAllocator) FreePTEs(ptes *PTEs) {
+	r.freed = append(r.freed, ptes)
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/allocator_unsafe.go b/pkg/sentry/platform/ring0/pagetables/allocator_unsafe.go
new file mode 100644
index 000000000..1b996b4e2
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/allocator_unsafe.go
@@ -0,0 +1,53 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pagetables
+
+import (
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// newAlignedPTEs returns a set of aligned PTEs.
+func newAlignedPTEs() *PTEs {
+	ptes := new(PTEs)
+	offset := physicalFor(ptes) & (usermem.PageSize - 1)
+	if offset == 0 {
+		// Already aligned.
+		return ptes
+	}
+
+	// Need to force an aligned allocation.
+	unaligned := make([]byte, (2*usermem.PageSize)-1)
+	offset = uintptr(unsafe.Pointer(&unaligned[0])) & (usermem.PageSize - 1)
+	if offset != 0 {
+		offset = usermem.PageSize - offset
+	}
+	return (*PTEs)(unsafe.Pointer(&unaligned[offset]))
+}
+
+// physicalFor returns the "physical" address for PTEs.
+//
+//go:nosplit
+func physicalFor(ptes *PTEs) uintptr {
+	return uintptr(unsafe.Pointer(ptes))
+}
+
+// fromPhysical returns the PTEs from the "physical" address.
+//
+//go:nosplit
+func fromPhysical(physical uintptr) *PTEs {
+	return (*PTEs)(unsafe.Pointer(physical))
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables.go b/pkg/sentry/platform/ring0/pagetables/pagetables.go
new file mode 100644
index 000000000..e5dcaada7
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables.go
@@ -0,0 +1,221 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package pagetables provides a generic implementation of pagetables.
+//
+// The core functions must be safe to call from a nosplit context. Furthermore,
+// this pagetables implementation goes to lengths to ensure that all functions
+// are free from runtime allocation. Calls to NewPTEs/FreePTEs may be made
+// during walks, but these can be cached elsewhere if required.
+package pagetables
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// PageTables is a set of page tables.
+type PageTables struct {
+	// Allocator is used to allocate nodes.
+	Allocator Allocator
+
+	// root is the pagetable root.
+	root *PTEs
+
+	// rootPhysical is the cached physical address of the root.
+	//
+	// This is saved only to prevent constant translation.
+	rootPhysical uintptr
+
+	// archPageTables includes architecture-specific features.
+	archPageTables
+}
+
+// New returns new PageTables.
+func New(a Allocator) *PageTables {
+	p := new(PageTables)
+	p.Init(a)
+	return p
+}
+
+// Init initializes a set of PageTables.
+//
+//go:nosplit
+func (p *PageTables) Init(allocator Allocator) {
+	p.Allocator = allocator
+	p.root = p.Allocator.NewPTEs()
+	p.rootPhysical = p.Allocator.PhysicalFor(p.root)
+}
+
+// mapVisitor is used for map.
+type mapVisitor struct {
+	target   uintptr // Input.
+	physical uintptr // Input.
+	opts     MapOpts // Input.
+	prev     bool    // Output.
+}
+
+// visit is used for map.
+//
+//go:nosplit
+func (v *mapVisitor) visit(start uintptr, pte *PTE, align uintptr) {
+	p := v.physical + (start - uintptr(v.target))
+	if pte.Valid() && (pte.Address() != p || pte.Opts() != v.opts) {
+		v.prev = true
+	}
+	if p&align != 0 {
+		// We will install entries at a smaller granulaity if we don't
+		// install a valid entry here, however we must zap any existing
+		// entry to ensure this happens.
+		pte.Clear()
+		return
+	}
+	pte.Set(p, v.opts)
+}
+
+//go:nosplit
+func (*mapVisitor) requiresAlloc() bool { return true }
+
+//go:nosplit
+func (*mapVisitor) requiresSplit() bool { return true }
+
+// Map installs a mapping with the given physical address.
+//
+// True is returned iff there was a previous mapping in the range.
+//
+// Precondition: addr & length must be page-aligned, their sum must not overflow.
+//
+//go:nosplit
+func (p *PageTables) Map(addr usermem.Addr, length uintptr, opts MapOpts, physical uintptr) bool {
+	if !opts.AccessType.Any() {
+		return p.Unmap(addr, length)
+	}
+	w := mapWalker{
+		pageTables: p,
+		visitor: mapVisitor{
+			target:   uintptr(addr),
+			physical: physical,
+			opts:     opts,
+		},
+	}
+	w.iterateRange(uintptr(addr), uintptr(addr)+length)
+	return w.visitor.prev
+}
+
+// unmapVisitor is used for unmap.
+type unmapVisitor struct {
+	count int
+}
+
+//go:nosplit
+func (*unmapVisitor) requiresAlloc() bool { return false }
+
+//go:nosplit
+func (*unmapVisitor) requiresSplit() bool { return true }
+
+// visit unmaps the given entry.
+//
+//go:nosplit
+func (v *unmapVisitor) visit(start uintptr, pte *PTE, align uintptr) {
+	pte.Clear()
+	v.count++
+}
+
+// Unmap unmaps the given range.
+//
+// True is returned iff there was a previous mapping in the range.
+//
+// Precondition: addr & length must be page-aligned.
+//
+//go:nosplit
+func (p *PageTables) Unmap(addr usermem.Addr, length uintptr) bool {
+	w := unmapWalker{
+		pageTables: p,
+		visitor: unmapVisitor{
+			count: 0,
+		},
+	}
+	w.iterateRange(uintptr(addr), uintptr(addr)+length)
+	return w.visitor.count > 0
+}
+
+// emptyVisitor is used for emptiness checks.
+type emptyVisitor struct {
+	count int
+}
+
+//go:nosplit
+func (*emptyVisitor) requiresAlloc() bool { return false }
+
+//go:nosplit
+func (*emptyVisitor) requiresSplit() bool { return false }
+
+// visit unmaps the given entry.
+//
+//go:nosplit
+func (v *emptyVisitor) visit(start uintptr, pte *PTE, align uintptr) {
+	v.count++
+}
+
+// IsEmpty checks if the given range is empty.
+//
+// Precondition: addr & length must be page-aligned.
+//
+//go:nosplit
+func (p *PageTables) IsEmpty(addr usermem.Addr, length uintptr) bool {
+	w := emptyWalker{
+		pageTables: p,
+	}
+	w.iterateRange(uintptr(addr), uintptr(addr)+length)
+	return w.visitor.count == 0
+}
+
+// lookupVisitor is used for lookup.
+type lookupVisitor struct {
+	target   uintptr // Input.
+	physical uintptr // Output.
+	opts     MapOpts // Output.
+}
+
+// visit matches the given address.
+//
+//go:nosplit
+func (v *lookupVisitor) visit(start uintptr, pte *PTE, align uintptr) {
+	if !pte.Valid() {
+		return
+	}
+	v.physical = pte.Address() + (start - uintptr(v.target))
+	v.opts = pte.Opts()
+}
+
+//go:nosplit
+func (*lookupVisitor) requiresAlloc() bool { return false }
+
+//go:nosplit
+func (*lookupVisitor) requiresSplit() bool { return false }
+
+// Lookup returns the physical address for the given virtual address.
+//
+//go:nosplit
+func (p *PageTables) Lookup(addr usermem.Addr) (physical uintptr, opts MapOpts) {
+	mask := uintptr(usermem.PageSize - 1)
+	offset := uintptr(addr) & mask
+	w := lookupWalker{
+		pageTables: p,
+		visitor: lookupVisitor{
+			target: uintptr(addr &^ usermem.Addr(mask)),
+		},
+	}
+	w.iterateRange(uintptr(addr), uintptr(addr)+1)
+	return w.visitor.physical + offset, w.visitor.opts
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_amd64.go b/pkg/sentry/platform/ring0/pagetables/pagetables_amd64.go
new file mode 100644
index 000000000..7aa6c524e
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_amd64.go
@@ -0,0 +1,45 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pagetables
+
+// Address constraints.
+//
+// The lowerTop and upperBottom currently apply to four-level pagetables;
+// additional refactoring would be necessary to support five-level pagetables.
+const (
+	lowerTop    = 0x00007fffffffffff
+	upperBottom = 0xffff800000000000
+
+	pteShift = 12
+	pmdShift = 21
+	pudShift = 30
+	pgdShift = 39
+
+	pteMask = 0x1ff << pteShift
+	pmdMask = 0x1ff << pmdShift
+	pudMask = 0x1ff << pudShift
+	pgdMask = 0x1ff << pgdShift
+
+	pteSize = 1 << pteShift
+	pmdSize = 1 << pmdShift
+	pudSize = 1 << pudShift
+	pgdSize = 1 << pgdShift
+
+	executeDisable = 1 << 63
+	entriesPerPage = 512
+)
+
+// PTEs is a collection of entries.
+type PTEs [entriesPerPage]PTE
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_state_autogen.go b/pkg/sentry/platform/ring0/pagetables/pagetables_state_autogen.go
new file mode 100755
index 000000000..ac1ccf3d3
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_state_autogen.go
@@ -0,0 +1,4 @@
+// automatically generated by stateify.
+
+package pagetables
+
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go
new file mode 100644
index 000000000..ff427fbe9
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go
@@ -0,0 +1,180 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package pagetables
+
+import (
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// archPageTables is architecture-specific data.
+type archPageTables struct {
+	// pcid is the value assigned by PCIDs.Assign.
+	//
+	// Note that zero is a valid PCID.
+	pcid uint16
+}
+
+// CR3 returns the CR3 value for these tables.
+//
+// This may be called in interrupt contexts. A PCID of zero always implies a
+// flush and should be passed when PCIDs are not enabled. See pcids_x86.go for
+// more information.
+//
+//go:nosplit
+func (p *PageTables) CR3(noFlush bool, pcid uint16) uint64 {
+	// Bit 63 is set to avoid flushing the PCID (per SDM 4.10.4.1).
+	const noFlushBit uint64 = 0x8000000000000000
+	if noFlush && pcid != 0 {
+		return noFlushBit | uint64(p.rootPhysical) | uint64(pcid)
+	}
+	return uint64(p.rootPhysical) | uint64(pcid)
+}
+
+// Bits in page table entries.
+const (
+	present      = 0x001
+	writable     = 0x002
+	user         = 0x004
+	writeThrough = 0x008
+	cacheDisable = 0x010
+	accessed     = 0x020
+	dirty        = 0x040
+	super        = 0x080
+	global       = 0x100
+	optionMask   = executeDisable | 0xfff
+)
+
+// MapOpts are x86 options.
+type MapOpts struct {
+	// AccessType defines permissions.
+	AccessType usermem.AccessType
+
+	// Global indicates the page is globally accessible.
+	Global bool
+
+	// User indicates the page is a user page.
+	User bool
+}
+
+// PTE is a page table entry.
+type PTE uintptr
+
+// Clear clears this PTE, including super page information.
+//
+//go:nosplit
+func (p *PTE) Clear() {
+	atomic.StoreUintptr((*uintptr)(p), 0)
+}
+
+// Valid returns true iff this entry is valid.
+//
+//go:nosplit
+func (p *PTE) Valid() bool {
+	return atomic.LoadUintptr((*uintptr)(p))&present != 0
+}
+
+// Opts returns the PTE options.
+//
+// These are all options except Valid and Super.
+//
+//go:nosplit
+func (p *PTE) Opts() MapOpts {
+	v := atomic.LoadUintptr((*uintptr)(p))
+	return MapOpts{
+		AccessType: usermem.AccessType{
+			Read:    v&present != 0,
+			Write:   v&writable != 0,
+			Execute: v&executeDisable == 0,
+		},
+		Global: v&global != 0,
+		User:   v&user != 0,
+	}
+}
+
+// SetSuper sets this page as a super page.
+//
+// The page must not be valid or a panic will result.
+//
+//go:nosplit
+func (p *PTE) SetSuper() {
+	if p.Valid() {
+		// This is not allowed.
+		panic("SetSuper called on valid page!")
+	}
+	atomic.StoreUintptr((*uintptr)(p), super)
+}
+
+// IsSuper returns true iff this page is a super page.
+//
+//go:nosplit
+func (p *PTE) IsSuper() bool {
+	return atomic.LoadUintptr((*uintptr)(p))&super != 0
+}
+
+// Set sets this PTE value.
+//
+// This does not change the super page property.
+//
+//go:nosplit
+func (p *PTE) Set(addr uintptr, opts MapOpts) {
+	if !opts.AccessType.Any() {
+		p.Clear()
+		return
+	}
+	v := (addr &^ optionMask) | present | accessed
+	if opts.User {
+		v |= user
+	}
+	if opts.Global {
+		v |= global
+	}
+	if !opts.AccessType.Execute {
+		v |= executeDisable
+	}
+	if opts.AccessType.Write {
+		v |= writable | dirty
+	}
+	if p.IsSuper() {
+		// Note that this is inherited from the previous instance. Set
+		// does not change the value of Super. See above.
+		v |= super
+	}
+	atomic.StoreUintptr((*uintptr)(p), v)
+}
+
+// setPageTable sets this PTE value and forces the write bit and super bit to
+// be cleared. This is used explicitly for breaking super pages.
+//
+//go:nosplit
+func (p *PTE) setPageTable(pt *PageTables, ptes *PTEs) {
+	addr := pt.Allocator.PhysicalFor(ptes)
+	if addr&^optionMask != addr {
+		// This should never happen.
+		panic("unaligned physical address!")
+	}
+	v := addr | present | user | writable | accessed | dirty
+	atomic.StoreUintptr((*uintptr)(p), v)
+}
+
+// Address extracts the address. This should only be used if Valid returns true.
+//
+//go:nosplit
+func (p *PTE) Address() uintptr {
+	return atomic.LoadUintptr((*uintptr)(p)) &^ optionMask
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go
new file mode 100644
index 000000000..0f029f25d
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go
@@ -0,0 +1,109 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package pagetables
+
+import (
+	"sync"
+)
+
+// limitPCID is the number of valid PCIDs.
+const limitPCID = 4096
+
+// PCIDs is a simple PCID database.
+//
+// This is not protected by locks and is thus suitable for use only with a
+// single CPU at a time.
+type PCIDs struct {
+	// mu protects below.
+	mu sync.Mutex
+
+	// cache are the assigned page tables.
+	cache map[*PageTables]uint16
+
+	// avail are available PCIDs.
+	avail []uint16
+}
+
+// NewPCIDs returns a new PCID database.
+//
+// start is the first index to assign. Typically this will be one, as the zero
+// pcid will always be flushed on transition (see pagetables_x86.go). This may
+// be more than one if specific PCIDs are reserved.
+//
+// Nil is returned iff the start and size are out of range.
+func NewPCIDs(start, size uint16) *PCIDs {
+	if start+uint16(size) >= limitPCID {
+		return nil // See comment.
+	}
+	p := &PCIDs{
+		cache: make(map[*PageTables]uint16),
+	}
+	for pcid := start; pcid < start+size; pcid++ {
+		p.avail = append(p.avail, pcid)
+	}
+	return p
+}
+
+// Assign assigns a PCID to the given PageTables.
+//
+// This may overwrite any previous assignment provided. If this in the case,
+// true is returned to indicate that the PCID should be flushed.
+func (p *PCIDs) Assign(pt *PageTables) (uint16, bool) {
+	p.mu.Lock()
+	if pcid, ok := p.cache[pt]; ok {
+		p.mu.Unlock()
+		return pcid, false // No flush.
+	}
+
+	// Is there something available?
+	if len(p.avail) > 0 {
+		pcid := p.avail[len(p.avail)-1]
+		p.avail = p.avail[:len(p.avail)-1]
+		p.cache[pt] = pcid
+
+		// We need to flush because while this is in the available
+		// pool, it may have been used previously.
+		p.mu.Unlock()
+		return pcid, true
+	}
+
+	// Evict an existing table.
+	for old, pcid := range p.cache {
+		delete(p.cache, old)
+		p.cache[pt] = pcid
+
+		// A flush is definitely required in this case, these page
+		// tables may still be active. (They will just be assigned some
+		// other PCID if and when they hit the given CPU again.)
+		p.mu.Unlock()
+		return pcid, true
+	}
+
+	// No PCID.
+	p.mu.Unlock()
+	return 0, false
+}
+
+// Drop drops references to a set of page tables.
+func (p *PCIDs) Drop(pt *PageTables) {
+	p.mu.Lock()
+	if pcid, ok := p.cache[pt]; ok {
+		delete(p.cache, pt)
+		p.avail = append(p.avail, pcid)
+	}
+	p.mu.Unlock()
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/walker_empty.go b/pkg/sentry/platform/ring0/pagetables/walker_empty.go
new file mode 100755
index 000000000..417784e17
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/walker_empty.go
@@ -0,0 +1,255 @@
+package pagetables
+
+// Walker walks page tables.
+type emptyWalker struct {
+	// pageTables are the tables to walk.
+	pageTables *PageTables
+
+	// Visitor is the set of arguments.
+	visitor emptyVisitor
+}
+
+// iterateRange iterates over all appropriate levels of page tables for the given range.
+//
+// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The
+// exception is super pages. If a valid super page (huge or jumbo) cannot be
+// installed, then the walk will continue to individual entries.
+//
+// This algorithm will attempt to maximize the use of super pages whenever
+// possible. Whether a super page is provided will be clear through the range
+// provided in the callback.
+//
+// Note that if requiresAlloc is true, then no gaps will be present. However,
+// if alloc is not set, then the iteration will likely be full of gaps.
+//
+// Note that this function should generally be avoided in favor of Map, Unmap,
+// etc. when not necessary.
+//
+// Precondition: start must be page-aligned.
+//
+// Precondition: start must be less than end.
+//
+// Precondition: If requiresAlloc is true, then start and end should not span
+// non-canonical ranges. If they do, a panic will result.
+//
+//go:nosplit
+func (w *emptyWalker) iterateRange(start, end uintptr) {
+	if start%pteSize != 0 {
+		panic("unaligned start")
+	}
+	if end < start {
+		panic("start > end")
+	}
+	if start < lowerTop {
+		if end <= lowerTop {
+			w.iterateRangeCanonical(start, end)
+		} else if end > lowerTop && end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else if start < upperBottom {
+		if end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else {
+		w.iterateRangeCanonical(start, end)
+	}
+}
+
+// next returns the next address quantized by the given size.
+//
+//go:nosplit
+func emptynext(start uintptr, size uintptr) uintptr {
+	start &= ^(size - 1)
+	start += size
+	return start
+}
+
+// iterateRangeCanonical walks a canonical range.
+//
+//go:nosplit
+func (w *emptyWalker) iterateRangeCanonical(start, end uintptr) {
+	for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ {
+		var (
+			pgdEntry   = &w.pageTables.root[pgdIndex]
+			pudEntries *PTEs
+		)
+		if !pgdEntry.Valid() {
+			if !w.visitor.requiresAlloc() {
+
+				start = emptynext(start, pgdSize)
+				continue
+			}
+
+			pudEntries = w.pageTables.Allocator.NewPTEs()
+			pgdEntry.setPageTable(w.pageTables, pudEntries)
+		} else {
+			pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address())
+		}
+
+		clearPUDEntries := uint16(0)
+
+		for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ {
+			var (
+				pudEntry   = &pudEntries[pudIndex]
+				pmdEntries *PTEs
+			)
+			if !pudEntry.Valid() {
+				if !w.visitor.requiresAlloc() {
+
+					clearPUDEntries++
+					start = emptynext(start, pudSize)
+					continue
+				}
+
+				if start&(pudSize-1) == 0 && end-start >= pudSize {
+					pudEntry.SetSuper()
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+					if pudEntry.Valid() {
+						start = emptynext(start, pudSize)
+						continue
+					}
+				}
+
+				pmdEntries = w.pageTables.Allocator.NewPTEs()
+				pudEntry.setPageTable(w.pageTables, pmdEntries)
+
+			} else if pudEntry.IsSuper() {
+
+				if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < emptynext(start, pudSize)) {
+
+					pmdEntries = w.pageTables.Allocator.NewPTEs()
+					for index := uint16(0); index < entriesPerPage; index++ {
+						pmdEntries[index].SetSuper()
+						pmdEntries[index].Set(
+							pudEntry.Address()+(pmdSize*uintptr(index)),
+							pudEntry.Opts())
+					}
+					pudEntry.setPageTable(w.pageTables, pmdEntries)
+				} else {
+
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+
+					if !pudEntry.Valid() {
+						clearPUDEntries++
+					}
+
+					start = emptynext(start, pudSize)
+					continue
+				}
+			} else {
+				pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address())
+			}
+
+			clearPMDEntries := uint16(0)
+
+			for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ {
+				var (
+					pmdEntry   = &pmdEntries[pmdIndex]
+					pteEntries *PTEs
+				)
+				if !pmdEntry.Valid() {
+					if !w.visitor.requiresAlloc() {
+
+						clearPMDEntries++
+						start = emptynext(start, pmdSize)
+						continue
+					}
+
+					if start&(pmdSize-1) == 0 && end-start >= pmdSize {
+						pmdEntry.SetSuper()
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+						if pmdEntry.Valid() {
+							start = emptynext(start, pmdSize)
+							continue
+						}
+					}
+
+					pteEntries = w.pageTables.Allocator.NewPTEs()
+					pmdEntry.setPageTable(w.pageTables, pteEntries)
+
+				} else if pmdEntry.IsSuper() {
+
+					if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < emptynext(start, pmdSize)) {
+
+						pteEntries = w.pageTables.Allocator.NewPTEs()
+						for index := uint16(0); index < entriesPerPage; index++ {
+							pteEntries[index].Set(
+								pmdEntry.Address()+(pteSize*uintptr(index)),
+								pmdEntry.Opts())
+						}
+						pmdEntry.setPageTable(w.pageTables, pteEntries)
+					} else {
+
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+
+						if !pmdEntry.Valid() {
+							clearPMDEntries++
+						}
+
+						start = emptynext(start, pmdSize)
+						continue
+					}
+				} else {
+					pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address())
+				}
+
+				clearPTEEntries := uint16(0)
+
+				for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ {
+					var (
+						pteEntry = &pteEntries[pteIndex]
+					)
+					if !pteEntry.Valid() && !w.visitor.requiresAlloc() {
+						clearPTEEntries++
+						start += pteSize
+						continue
+					}
+
+					w.visitor.visit(uintptr(start), pteEntry, pteSize-1)
+					if !pteEntry.Valid() {
+						if w.visitor.requiresAlloc() {
+							panic("PTE not set after iteration with requiresAlloc!")
+						}
+						clearPTEEntries++
+					}
+
+					start += pteSize
+					continue
+				}
+
+				if clearPTEEntries == entriesPerPage {
+					pmdEntry.Clear()
+					w.pageTables.Allocator.FreePTEs(pteEntries)
+					clearPMDEntries++
+				}
+			}
+
+			if clearPMDEntries == entriesPerPage {
+				pudEntry.Clear()
+				w.pageTables.Allocator.FreePTEs(pmdEntries)
+				clearPUDEntries++
+			}
+		}
+
+		if clearPUDEntries == entriesPerPage {
+			pgdEntry.Clear()
+			w.pageTables.Allocator.FreePTEs(pudEntries)
+		}
+	}
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/walker_lookup.go b/pkg/sentry/platform/ring0/pagetables/walker_lookup.go
new file mode 100755
index 000000000..906c9c50f
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/walker_lookup.go
@@ -0,0 +1,255 @@
+package pagetables
+
+// Walker walks page tables.
+type lookupWalker struct {
+	// pageTables are the tables to walk.
+	pageTables *PageTables
+
+	// Visitor is the set of arguments.
+	visitor lookupVisitor
+}
+
+// iterateRange iterates over all appropriate levels of page tables for the given range.
+//
+// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The
+// exception is super pages. If a valid super page (huge or jumbo) cannot be
+// installed, then the walk will continue to individual entries.
+//
+// This algorithm will attempt to maximize the use of super pages whenever
+// possible. Whether a super page is provided will be clear through the range
+// provided in the callback.
+//
+// Note that if requiresAlloc is true, then no gaps will be present. However,
+// if alloc is not set, then the iteration will likely be full of gaps.
+//
+// Note that this function should generally be avoided in favor of Map, Unmap,
+// etc. when not necessary.
+//
+// Precondition: start must be page-aligned.
+//
+// Precondition: start must be less than end.
+//
+// Precondition: If requiresAlloc is true, then start and end should not span
+// non-canonical ranges. If they do, a panic will result.
+//
+//go:nosplit
+func (w *lookupWalker) iterateRange(start, end uintptr) {
+	if start%pteSize != 0 {
+		panic("unaligned start")
+	}
+	if end < start {
+		panic("start > end")
+	}
+	if start < lowerTop {
+		if end <= lowerTop {
+			w.iterateRangeCanonical(start, end)
+		} else if end > lowerTop && end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else if start < upperBottom {
+		if end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else {
+		w.iterateRangeCanonical(start, end)
+	}
+}
+
+// next returns the next address quantized by the given size.
+//
+//go:nosplit
+func lookupnext(start uintptr, size uintptr) uintptr {
+	start &= ^(size - 1)
+	start += size
+	return start
+}
+
+// iterateRangeCanonical walks a canonical range.
+//
+//go:nosplit
+func (w *lookupWalker) iterateRangeCanonical(start, end uintptr) {
+	for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ {
+		var (
+			pgdEntry   = &w.pageTables.root[pgdIndex]
+			pudEntries *PTEs
+		)
+		if !pgdEntry.Valid() {
+			if !w.visitor.requiresAlloc() {
+
+				start = lookupnext(start, pgdSize)
+				continue
+			}
+
+			pudEntries = w.pageTables.Allocator.NewPTEs()
+			pgdEntry.setPageTable(w.pageTables, pudEntries)
+		} else {
+			pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address())
+		}
+
+		clearPUDEntries := uint16(0)
+
+		for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ {
+			var (
+				pudEntry   = &pudEntries[pudIndex]
+				pmdEntries *PTEs
+			)
+			if !pudEntry.Valid() {
+				if !w.visitor.requiresAlloc() {
+
+					clearPUDEntries++
+					start = lookupnext(start, pudSize)
+					continue
+				}
+
+				if start&(pudSize-1) == 0 && end-start >= pudSize {
+					pudEntry.SetSuper()
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+					if pudEntry.Valid() {
+						start = lookupnext(start, pudSize)
+						continue
+					}
+				}
+
+				pmdEntries = w.pageTables.Allocator.NewPTEs()
+				pudEntry.setPageTable(w.pageTables, pmdEntries)
+
+			} else if pudEntry.IsSuper() {
+
+				if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < lookupnext(start, pudSize)) {
+
+					pmdEntries = w.pageTables.Allocator.NewPTEs()
+					for index := uint16(0); index < entriesPerPage; index++ {
+						pmdEntries[index].SetSuper()
+						pmdEntries[index].Set(
+							pudEntry.Address()+(pmdSize*uintptr(index)),
+							pudEntry.Opts())
+					}
+					pudEntry.setPageTable(w.pageTables, pmdEntries)
+				} else {
+
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+
+					if !pudEntry.Valid() {
+						clearPUDEntries++
+					}
+
+					start = lookupnext(start, pudSize)
+					continue
+				}
+			} else {
+				pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address())
+			}
+
+			clearPMDEntries := uint16(0)
+
+			for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ {
+				var (
+					pmdEntry   = &pmdEntries[pmdIndex]
+					pteEntries *PTEs
+				)
+				if !pmdEntry.Valid() {
+					if !w.visitor.requiresAlloc() {
+
+						clearPMDEntries++
+						start = lookupnext(start, pmdSize)
+						continue
+					}
+
+					if start&(pmdSize-1) == 0 && end-start >= pmdSize {
+						pmdEntry.SetSuper()
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+						if pmdEntry.Valid() {
+							start = lookupnext(start, pmdSize)
+							continue
+						}
+					}
+
+					pteEntries = w.pageTables.Allocator.NewPTEs()
+					pmdEntry.setPageTable(w.pageTables, pteEntries)
+
+				} else if pmdEntry.IsSuper() {
+
+					if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < lookupnext(start, pmdSize)) {
+
+						pteEntries = w.pageTables.Allocator.NewPTEs()
+						for index := uint16(0); index < entriesPerPage; index++ {
+							pteEntries[index].Set(
+								pmdEntry.Address()+(pteSize*uintptr(index)),
+								pmdEntry.Opts())
+						}
+						pmdEntry.setPageTable(w.pageTables, pteEntries)
+					} else {
+
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+
+						if !pmdEntry.Valid() {
+							clearPMDEntries++
+						}
+
+						start = lookupnext(start, pmdSize)
+						continue
+					}
+				} else {
+					pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address())
+				}
+
+				clearPTEEntries := uint16(0)
+
+				for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ {
+					var (
+						pteEntry = &pteEntries[pteIndex]
+					)
+					if !pteEntry.Valid() && !w.visitor.requiresAlloc() {
+						clearPTEEntries++
+						start += pteSize
+						continue
+					}
+
+					w.visitor.visit(uintptr(start), pteEntry, pteSize-1)
+					if !pteEntry.Valid() {
+						if w.visitor.requiresAlloc() {
+							panic("PTE not set after iteration with requiresAlloc!")
+						}
+						clearPTEEntries++
+					}
+
+					start += pteSize
+					continue
+				}
+
+				if clearPTEEntries == entriesPerPage {
+					pmdEntry.Clear()
+					w.pageTables.Allocator.FreePTEs(pteEntries)
+					clearPMDEntries++
+				}
+			}
+
+			if clearPMDEntries == entriesPerPage {
+				pudEntry.Clear()
+				w.pageTables.Allocator.FreePTEs(pmdEntries)
+				clearPUDEntries++
+			}
+		}
+
+		if clearPUDEntries == entriesPerPage {
+			pgdEntry.Clear()
+			w.pageTables.Allocator.FreePTEs(pudEntries)
+		}
+	}
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/walker_map.go b/pkg/sentry/platform/ring0/pagetables/walker_map.go
new file mode 100755
index 000000000..61ee3c825
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/walker_map.go
@@ -0,0 +1,255 @@
+package pagetables
+
+// Walker walks page tables.
+type mapWalker struct {
+	// pageTables are the tables to walk.
+	pageTables *PageTables
+
+	// Visitor is the set of arguments.
+	visitor mapVisitor
+}
+
+// iterateRange iterates over all appropriate levels of page tables for the given range.
+//
+// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The
+// exception is super pages. If a valid super page (huge or jumbo) cannot be
+// installed, then the walk will continue to individual entries.
+//
+// This algorithm will attempt to maximize the use of super pages whenever
+// possible. Whether a super page is provided will be clear through the range
+// provided in the callback.
+//
+// Note that if requiresAlloc is true, then no gaps will be present. However,
+// if alloc is not set, then the iteration will likely be full of gaps.
+//
+// Note that this function should generally be avoided in favor of Map, Unmap,
+// etc. when not necessary.
+//
+// Precondition: start must be page-aligned.
+//
+// Precondition: start must be less than end.
+//
+// Precondition: If requiresAlloc is true, then start and end should not span
+// non-canonical ranges. If they do, a panic will result.
+//
+//go:nosplit
+func (w *mapWalker) iterateRange(start, end uintptr) {
+	if start%pteSize != 0 {
+		panic("unaligned start")
+	}
+	if end < start {
+		panic("start > end")
+	}
+	if start < lowerTop {
+		if end <= lowerTop {
+			w.iterateRangeCanonical(start, end)
+		} else if end > lowerTop && end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else if start < upperBottom {
+		if end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else {
+		w.iterateRangeCanonical(start, end)
+	}
+}
+
+// next returns the next address quantized by the given size.
+//
+//go:nosplit
+func mapnext(start uintptr, size uintptr) uintptr {
+	start &= ^(size - 1)
+	start += size
+	return start
+}
+
+// iterateRangeCanonical walks a canonical range.
+//
+//go:nosplit
+func (w *mapWalker) iterateRangeCanonical(start, end uintptr) {
+	for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ {
+		var (
+			pgdEntry   = &w.pageTables.root[pgdIndex]
+			pudEntries *PTEs
+		)
+		if !pgdEntry.Valid() {
+			if !w.visitor.requiresAlloc() {
+
+				start = mapnext(start, pgdSize)
+				continue
+			}
+
+			pudEntries = w.pageTables.Allocator.NewPTEs()
+			pgdEntry.setPageTable(w.pageTables, pudEntries)
+		} else {
+			pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address())
+		}
+
+		clearPUDEntries := uint16(0)
+
+		for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ {
+			var (
+				pudEntry   = &pudEntries[pudIndex]
+				pmdEntries *PTEs
+			)
+			if !pudEntry.Valid() {
+				if !w.visitor.requiresAlloc() {
+
+					clearPUDEntries++
+					start = mapnext(start, pudSize)
+					continue
+				}
+
+				if start&(pudSize-1) == 0 && end-start >= pudSize {
+					pudEntry.SetSuper()
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+					if pudEntry.Valid() {
+						start = mapnext(start, pudSize)
+						continue
+					}
+				}
+
+				pmdEntries = w.pageTables.Allocator.NewPTEs()
+				pudEntry.setPageTable(w.pageTables, pmdEntries)
+
+			} else if pudEntry.IsSuper() {
+
+				if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < mapnext(start, pudSize)) {
+
+					pmdEntries = w.pageTables.Allocator.NewPTEs()
+					for index := uint16(0); index < entriesPerPage; index++ {
+						pmdEntries[index].SetSuper()
+						pmdEntries[index].Set(
+							pudEntry.Address()+(pmdSize*uintptr(index)),
+							pudEntry.Opts())
+					}
+					pudEntry.setPageTable(w.pageTables, pmdEntries)
+				} else {
+
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+
+					if !pudEntry.Valid() {
+						clearPUDEntries++
+					}
+
+					start = mapnext(start, pudSize)
+					continue
+				}
+			} else {
+				pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address())
+			}
+
+			clearPMDEntries := uint16(0)
+
+			for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ {
+				var (
+					pmdEntry   = &pmdEntries[pmdIndex]
+					pteEntries *PTEs
+				)
+				if !pmdEntry.Valid() {
+					if !w.visitor.requiresAlloc() {
+
+						clearPMDEntries++
+						start = mapnext(start, pmdSize)
+						continue
+					}
+
+					if start&(pmdSize-1) == 0 && end-start >= pmdSize {
+						pmdEntry.SetSuper()
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+						if pmdEntry.Valid() {
+							start = mapnext(start, pmdSize)
+							continue
+						}
+					}
+
+					pteEntries = w.pageTables.Allocator.NewPTEs()
+					pmdEntry.setPageTable(w.pageTables, pteEntries)
+
+				} else if pmdEntry.IsSuper() {
+
+					if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < mapnext(start, pmdSize)) {
+
+						pteEntries = w.pageTables.Allocator.NewPTEs()
+						for index := uint16(0); index < entriesPerPage; index++ {
+							pteEntries[index].Set(
+								pmdEntry.Address()+(pteSize*uintptr(index)),
+								pmdEntry.Opts())
+						}
+						pmdEntry.setPageTable(w.pageTables, pteEntries)
+					} else {
+
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+
+						if !pmdEntry.Valid() {
+							clearPMDEntries++
+						}
+
+						start = mapnext(start, pmdSize)
+						continue
+					}
+				} else {
+					pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address())
+				}
+
+				clearPTEEntries := uint16(0)
+
+				for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ {
+					var (
+						pteEntry = &pteEntries[pteIndex]
+					)
+					if !pteEntry.Valid() && !w.visitor.requiresAlloc() {
+						clearPTEEntries++
+						start += pteSize
+						continue
+					}
+
+					w.visitor.visit(uintptr(start), pteEntry, pteSize-1)
+					if !pteEntry.Valid() {
+						if w.visitor.requiresAlloc() {
+							panic("PTE not set after iteration with requiresAlloc!")
+						}
+						clearPTEEntries++
+					}
+
+					start += pteSize
+					continue
+				}
+
+				if clearPTEEntries == entriesPerPage {
+					pmdEntry.Clear()
+					w.pageTables.Allocator.FreePTEs(pteEntries)
+					clearPMDEntries++
+				}
+			}
+
+			if clearPMDEntries == entriesPerPage {
+				pudEntry.Clear()
+				w.pageTables.Allocator.FreePTEs(pmdEntries)
+				clearPUDEntries++
+			}
+		}
+
+		if clearPUDEntries == entriesPerPage {
+			pgdEntry.Clear()
+			w.pageTables.Allocator.FreePTEs(pudEntries)
+		}
+	}
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/walker_unmap.go b/pkg/sentry/platform/ring0/pagetables/walker_unmap.go
new file mode 100755
index 000000000..be2aa0ce4
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/walker_unmap.go
@@ -0,0 +1,255 @@
+package pagetables
+
+// Walker walks page tables.
+type unmapWalker struct {
+	// pageTables are the tables to walk.
+	pageTables *PageTables
+
+	// Visitor is the set of arguments.
+	visitor unmapVisitor
+}
+
+// iterateRange iterates over all appropriate levels of page tables for the given range.
+//
+// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The
+// exception is super pages. If a valid super page (huge or jumbo) cannot be
+// installed, then the walk will continue to individual entries.
+//
+// This algorithm will attempt to maximize the use of super pages whenever
+// possible. Whether a super page is provided will be clear through the range
+// provided in the callback.
+//
+// Note that if requiresAlloc is true, then no gaps will be present. However,
+// if alloc is not set, then the iteration will likely be full of gaps.
+//
+// Note that this function should generally be avoided in favor of Map, Unmap,
+// etc. when not necessary.
+//
+// Precondition: start must be page-aligned.
+//
+// Precondition: start must be less than end.
+//
+// Precondition: If requiresAlloc is true, then start and end should not span
+// non-canonical ranges. If they do, a panic will result.
+//
+//go:nosplit
+func (w *unmapWalker) iterateRange(start, end uintptr) {
+	if start%pteSize != 0 {
+		panic("unaligned start")
+	}
+	if end < start {
+		panic("start > end")
+	}
+	if start < lowerTop {
+		if end <= lowerTop {
+			w.iterateRangeCanonical(start, end)
+		} else if end > lowerTop && end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(start, lowerTop)
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else if start < upperBottom {
+		if end <= upperBottom {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+		} else {
+			if w.visitor.requiresAlloc() {
+				panic("alloc spans non-canonical range")
+			}
+			w.iterateRangeCanonical(upperBottom, end)
+		}
+	} else {
+		w.iterateRangeCanonical(start, end)
+	}
+}
+
+// next returns the next address quantized by the given size.
+//
+//go:nosplit
+func unmapnext(start uintptr, size uintptr) uintptr {
+	start &= ^(size - 1)
+	start += size
+	return start
+}
+
+// iterateRangeCanonical walks a canonical range.
+//
+//go:nosplit
+func (w *unmapWalker) iterateRangeCanonical(start, end uintptr) {
+	for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ {
+		var (
+			pgdEntry   = &w.pageTables.root[pgdIndex]
+			pudEntries *PTEs
+		)
+		if !pgdEntry.Valid() {
+			if !w.visitor.requiresAlloc() {
+
+				start = unmapnext(start, pgdSize)
+				continue
+			}
+
+			pudEntries = w.pageTables.Allocator.NewPTEs()
+			pgdEntry.setPageTable(w.pageTables, pudEntries)
+		} else {
+			pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address())
+		}
+
+		clearPUDEntries := uint16(0)
+
+		for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ {
+			var (
+				pudEntry   = &pudEntries[pudIndex]
+				pmdEntries *PTEs
+			)
+			if !pudEntry.Valid() {
+				if !w.visitor.requiresAlloc() {
+
+					clearPUDEntries++
+					start = unmapnext(start, pudSize)
+					continue
+				}
+
+				if start&(pudSize-1) == 0 && end-start >= pudSize {
+					pudEntry.SetSuper()
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+					if pudEntry.Valid() {
+						start = unmapnext(start, pudSize)
+						continue
+					}
+				}
+
+				pmdEntries = w.pageTables.Allocator.NewPTEs()
+				pudEntry.setPageTable(w.pageTables, pmdEntries)
+
+			} else if pudEntry.IsSuper() {
+
+				if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < unmapnext(start, pudSize)) {
+
+					pmdEntries = w.pageTables.Allocator.NewPTEs()
+					for index := uint16(0); index < entriesPerPage; index++ {
+						pmdEntries[index].SetSuper()
+						pmdEntries[index].Set(
+							pudEntry.Address()+(pmdSize*uintptr(index)),
+							pudEntry.Opts())
+					}
+					pudEntry.setPageTable(w.pageTables, pmdEntries)
+				} else {
+
+					w.visitor.visit(uintptr(start), pudEntry, pudSize-1)
+
+					if !pudEntry.Valid() {
+						clearPUDEntries++
+					}
+
+					start = unmapnext(start, pudSize)
+					continue
+				}
+			} else {
+				pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address())
+			}
+
+			clearPMDEntries := uint16(0)
+
+			for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ {
+				var (
+					pmdEntry   = &pmdEntries[pmdIndex]
+					pteEntries *PTEs
+				)
+				if !pmdEntry.Valid() {
+					if !w.visitor.requiresAlloc() {
+
+						clearPMDEntries++
+						start = unmapnext(start, pmdSize)
+						continue
+					}
+
+					if start&(pmdSize-1) == 0 && end-start >= pmdSize {
+						pmdEntry.SetSuper()
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+						if pmdEntry.Valid() {
+							start = unmapnext(start, pmdSize)
+							continue
+						}
+					}
+
+					pteEntries = w.pageTables.Allocator.NewPTEs()
+					pmdEntry.setPageTable(w.pageTables, pteEntries)
+
+				} else if pmdEntry.IsSuper() {
+
+					if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < unmapnext(start, pmdSize)) {
+
+						pteEntries = w.pageTables.Allocator.NewPTEs()
+						for index := uint16(0); index < entriesPerPage; index++ {
+							pteEntries[index].Set(
+								pmdEntry.Address()+(pteSize*uintptr(index)),
+								pmdEntry.Opts())
+						}
+						pmdEntry.setPageTable(w.pageTables, pteEntries)
+					} else {
+
+						w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1)
+
+						if !pmdEntry.Valid() {
+							clearPMDEntries++
+						}
+
+						start = unmapnext(start, pmdSize)
+						continue
+					}
+				} else {
+					pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address())
+				}
+
+				clearPTEEntries := uint16(0)
+
+				for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ {
+					var (
+						pteEntry = &pteEntries[pteIndex]
+					)
+					if !pteEntry.Valid() && !w.visitor.requiresAlloc() {
+						clearPTEEntries++
+						start += pteSize
+						continue
+					}
+
+					w.visitor.visit(uintptr(start), pteEntry, pteSize-1)
+					if !pteEntry.Valid() {
+						if w.visitor.requiresAlloc() {
+							panic("PTE not set after iteration with requiresAlloc!")
+						}
+						clearPTEEntries++
+					}
+
+					start += pteSize
+					continue
+				}
+
+				if clearPTEEntries == entriesPerPage {
+					pmdEntry.Clear()
+					w.pageTables.Allocator.FreePTEs(pteEntries)
+					clearPMDEntries++
+				}
+			}
+
+			if clearPMDEntries == entriesPerPage {
+				pudEntry.Clear()
+				w.pageTables.Allocator.FreePTEs(pmdEntries)
+				clearPUDEntries++
+			}
+		}
+
+		if clearPUDEntries == entriesPerPage {
+			pgdEntry.Clear()
+			w.pageTables.Allocator.FreePTEs(pudEntries)
+		}
+	}
+}