Merge release-20200323.0-177-g12bde95 (automated)

10 files changed, 2540 insertions, 0 deletions

diff --git a/pkg/sentry/fsimpl/tmpfs/dentry_list.go b/pkg/sentry/fsimpl/tmpfs/dentry_list.go
new file mode 100755
index 000000000..9af621dd3
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/dentry_list.go
@@ -0,0 +1,193 @@
+package tmpfs
+
+// ElementMapper provides an identity mapping by default.
+//
+// This can be replaced to provide a struct that maps elements to linker
+// objects, if they are not the same. An ElementMapper is not typically
+// required if: Linker is left as is, Element is left as is, or Linker and
+// Element are the same type.
+type dentryElementMapper struct{}
+
+// linkerFor maps an Element to a Linker.
+//
+// This default implementation should be inlined.
+//
+//go:nosplit
+func (dentryElementMapper) linkerFor(elem *dentry) *dentry { return elem }
+
+// List is an intrusive list. Entries can be added to or removed from the list
+// in O(1) time and with no additional memory allocations.
+//
+// The zero value for List is an empty list ready to use.
+//
+// To iterate over a list (where l is a List):
+//      for e := l.Front(); e != nil; e = e.Next() {
+// 		// do something with e.
+//      }
+//
+// +stateify savable
+type dentryList struct {
+	head *dentry
+	tail *dentry
+}
+
+// Reset resets list l to the empty state.
+func (l *dentryList) Reset() {
+	l.head = nil
+	l.tail = nil
+}
+
+// Empty returns true iff the list is empty.
+func (l *dentryList) Empty() bool {
+	return l.head == nil
+}
+
+// Front returns the first element of list l or nil.
+func (l *dentryList) Front() *dentry {
+	return l.head
+}
+
+// Back returns the last element of list l or nil.
+func (l *dentryList) Back() *dentry {
+	return l.tail
+}
+
+// Len returns the number of elements in the list.
+//
+// NOTE: This is an O(n) operation.
+func (l *dentryList) Len() (count int) {
+	for e := l.Front(); e != nil; e = e.Next() {
+		count++
+	}
+	return count
+}
+
+// PushFront inserts the element e at the front of list l.
+func (l *dentryList) PushFront(e *dentry) {
+	linker := dentryElementMapper{}.linkerFor(e)
+	linker.SetNext(l.head)
+	linker.SetPrev(nil)
+	if l.head != nil {
+		dentryElementMapper{}.linkerFor(l.head).SetPrev(e)
+	} else {
+		l.tail = e
+	}
+
+	l.head = e
+}
+
+// PushBack inserts the element e at the back of list l.
+func (l *dentryList) PushBack(e *dentry) {
+	linker := dentryElementMapper{}.linkerFor(e)
+	linker.SetNext(nil)
+	linker.SetPrev(l.tail)
+	if l.tail != nil {
+		dentryElementMapper{}.linkerFor(l.tail).SetNext(e)
+	} else {
+		l.head = e
+	}
+
+	l.tail = e
+}
+
+// PushBackList inserts list m at the end of list l, emptying m.
+func (l *dentryList) PushBackList(m *dentryList) {
+	if l.head == nil {
+		l.head = m.head
+		l.tail = m.tail
+	} else if m.head != nil {
+		dentryElementMapper{}.linkerFor(l.tail).SetNext(m.head)
+		dentryElementMapper{}.linkerFor(m.head).SetPrev(l.tail)
+
+		l.tail = m.tail
+	}
+	m.head = nil
+	m.tail = nil
+}
+
+// InsertAfter inserts e after b.
+func (l *dentryList) InsertAfter(b, e *dentry) {
+	bLinker := dentryElementMapper{}.linkerFor(b)
+	eLinker := dentryElementMapper{}.linkerFor(e)
+
+	a := bLinker.Next()
+
+	eLinker.SetNext(a)
+	eLinker.SetPrev(b)
+	bLinker.SetNext(e)
+
+	if a != nil {
+		dentryElementMapper{}.linkerFor(a).SetPrev(e)
+	} else {
+		l.tail = e
+	}
+}
+
+// InsertBefore inserts e before a.
+func (l *dentryList) InsertBefore(a, e *dentry) {
+	aLinker := dentryElementMapper{}.linkerFor(a)
+	eLinker := dentryElementMapper{}.linkerFor(e)
+
+	b := aLinker.Prev()
+	eLinker.SetNext(a)
+	eLinker.SetPrev(b)
+	aLinker.SetPrev(e)
+
+	if b != nil {
+		dentryElementMapper{}.linkerFor(b).SetNext(e)
+	} else {
+		l.head = e
+	}
+}
+
+// Remove removes e from l.
+func (l *dentryList) Remove(e *dentry) {
+	linker := dentryElementMapper{}.linkerFor(e)
+	prev := linker.Prev()
+	next := linker.Next()
+
+	if prev != nil {
+		dentryElementMapper{}.linkerFor(prev).SetNext(next)
+	} else {
+		l.head = next
+	}
+
+	if next != nil {
+		dentryElementMapper{}.linkerFor(next).SetPrev(prev)
+	} else {
+		l.tail = prev
+	}
+
+	linker.SetNext(nil)
+	linker.SetPrev(nil)
+}
+
+// Entry is a default implementation of Linker. Users can add anonymous fields
+// of this type to their structs to make them automatically implement the
+// methods needed by List.
+//
+// +stateify savable
+type dentryEntry struct {
+	next *dentry
+	prev *dentry
+}
+
+// Next returns the entry that follows e in the list.
+func (e *dentryEntry) Next() *dentry {
+	return e.next
+}
+
+// Prev returns the entry that precedes e in the list.
+func (e *dentryEntry) Prev() *dentry {
+	return e.prev
+}
+
+// SetNext assigns 'entry' as the entry that follows e in the list.
+func (e *dentryEntry) SetNext(elem *dentry) {
+	e.next = elem
+}
+
+// SetPrev assigns 'entry' as the entry that precedes e in the list.
+func (e *dentryEntry) SetPrev(elem *dentry) {
+	e.prev = elem
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/device_file.go b/pkg/sentry/fsimpl/tmpfs/device_file.go
new file mode 100755
index 000000000..83bf885ee
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/device_file.go
@@ -0,0 +1,49 @@
+// Copyright 2020 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 tmpfs
+
+import (
+	"fmt"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+type deviceFile struct {
+	inode inode
+	kind  vfs.DeviceKind
+	major uint32
+	minor uint32
+}
+
+func (fs *filesystem) newDeviceFile(creds *auth.Credentials, mode linux.FileMode, kind vfs.DeviceKind, major, minor uint32) *inode {
+	file := &deviceFile{
+		kind:  kind,
+		major: major,
+		minor: minor,
+	}
+	switch kind {
+	case vfs.BlockDevice:
+		mode |= linux.S_IFBLK
+	case vfs.CharDevice:
+		mode |= linux.S_IFCHR
+	default:
+		panic(fmt.Sprintf("invalid DeviceKind: %v", kind))
+	}
+	file.inode.init(file, fs, creds, mode)
+	file.inode.nlink = 1 // from parent directory
+	return &file.inode
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go
new file mode 100755
index 000000000..45712c9b9
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/directory.go
@@ -0,0 +1,184 @@
+// Copyright 2019 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 tmpfs
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+type directory struct {
+	inode inode
+
+	// childList is a list containing (1) child Dentries and (2) fake Dentries
+	// (with inode == nil) that represent the iteration position of
+	// directoryFDs. childList is used to support directoryFD.IterDirents()
+	// efficiently. childList is protected by filesystem.mu.
+	childList dentryList
+}
+
+func (fs *filesystem) newDirectory(creds *auth.Credentials, mode linux.FileMode) *inode {
+	dir := &directory{}
+	dir.inode.init(dir, fs, creds, linux.S_IFDIR|mode)
+	dir.inode.nlink = 2 // from "." and parent directory or ".." for root
+	return &dir.inode
+}
+
+type directoryFD struct {
+	fileDescription
+	vfs.DirectoryFileDescriptionDefaultImpl
+
+	// Protected by filesystem.mu.
+	iter *dentry
+	off  int64
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *directoryFD) Release() {
+	if fd.iter != nil {
+		fs := fd.filesystem()
+		dir := fd.inode().impl.(*directory)
+		fs.mu.Lock()
+		dir.childList.Remove(fd.iter)
+		fs.mu.Unlock()
+		fd.iter = nil
+	}
+}
+
+// IterDirents implements vfs.FileDescriptionImpl.IterDirents.
+func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error {
+	fs := fd.filesystem()
+	vfsd := fd.vfsfd.VirtualDentry().Dentry()
+
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+
+	fd.inode().touchAtime(fd.vfsfd.Mount())
+
+	if fd.off == 0 {
+		if err := cb.Handle(vfs.Dirent{
+			Name:    ".",
+			Type:    linux.DT_DIR,
+			Ino:     vfsd.Impl().(*dentry).inode.ino,
+			NextOff: 1,
+		}); err != nil {
+			return err
+		}
+		fd.off++
+	}
+	if fd.off == 1 {
+		parentInode := vfsd.ParentOrSelf().Impl().(*dentry).inode
+		if err := cb.Handle(vfs.Dirent{
+			Name:    "..",
+			Type:    parentInode.direntType(),
+			Ino:     parentInode.ino,
+			NextOff: 2,
+		}); err != nil {
+			return err
+		}
+		fd.off++
+	}
+
+	dir := vfsd.Impl().(*dentry).inode.impl.(*directory)
+	var child *dentry
+	if fd.iter == nil {
+		// Start iteration at the beginning of dir.
+		child = dir.childList.Front()
+		fd.iter = &dentry{}
+	} else {
+		// Continue iteration from where we left off.
+		child = fd.iter.Next()
+		dir.childList.Remove(fd.iter)
+	}
+	for child != nil {
+		// Skip other directoryFD iterators.
+		if child.inode != nil {
+			if err := cb.Handle(vfs.Dirent{
+				Name:    child.vfsd.Name(),
+				Type:    child.inode.direntType(),
+				Ino:     child.inode.ino,
+				NextOff: fd.off + 1,
+			}); err != nil {
+				dir.childList.InsertBefore(child, fd.iter)
+				return err
+			}
+			fd.off++
+		}
+		child = child.Next()
+	}
+	dir.childList.PushBack(fd.iter)
+	return nil
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+	fs := fd.filesystem()
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+
+	switch whence {
+	case linux.SEEK_SET:
+		// Use offset as given.
+	case linux.SEEK_CUR:
+		offset += fd.off
+	default:
+		return 0, syserror.EINVAL
+	}
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+
+	// If the offset isn't changing (e.g. due to lseek(0, SEEK_CUR)), don't
+	// seek even if doing so might reposition the iterator due to concurrent
+	// mutation of the directory. Compare fs/libfs.c:dcache_dir_lseek().
+	if fd.off == offset {
+		return offset, nil
+	}
+
+	fd.off = offset
+	// Compensate for "." and "..".
+	remChildren := int64(0)
+	if offset >= 2 {
+		remChildren = offset - 2
+	}
+
+	dir := fd.inode().impl.(*directory)
+
+	// Ensure that fd.iter exists and is not linked into dir.childList.
+	if fd.iter == nil {
+		fd.iter = &dentry{}
+	} else {
+		dir.childList.Remove(fd.iter)
+	}
+	// Insert fd.iter before the remChildren'th child, or at the end of the
+	// list if remChildren >= number of children.
+	child := dir.childList.Front()
+	for child != nil {
+		// Skip other directoryFD iterators.
+		if child.inode != nil {
+			if remChildren == 0 {
+				dir.childList.InsertBefore(child, fd.iter)
+				return offset, nil
+			}
+			remChildren--
+		}
+		child = child.Next()
+	}
+	dir.childList.PushBack(fd.iter)
+	return offset, nil
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go
new file mode 100755
index 000000000..660f5a29b
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go
@@ -0,0 +1,747 @@
+// Copyright 2019 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 tmpfs
+
+import (
+	"fmt"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fspath"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/socket/unix/transport"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// Sync implements vfs.FilesystemImpl.Sync.
+func (fs *filesystem) Sync(ctx context.Context) error {
+	// All filesystem state is in-memory.
+	return nil
+}
+
+// stepLocked resolves rp.Component() to an existing file, starting from the
+// given directory.
+//
+// stepLocked is loosely analogous to fs/namei.c:walk_component().
+//
+// Preconditions: filesystem.mu must be locked. !rp.Done().
+func stepLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) {
+	if !d.inode.isDir() {
+		return nil, syserror.ENOTDIR
+	}
+	if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil {
+		return nil, err
+	}
+afterSymlink:
+	if len(rp.Component()) > linux.NAME_MAX {
+		return nil, syserror.ENAMETOOLONG
+	}
+	nextVFSD, err := rp.ResolveComponent(&d.vfsd)
+	if err != nil {
+		return nil, err
+	}
+	if nextVFSD == nil {
+		// Since the Dentry tree is the sole source of truth for tmpfs, if it's
+		// not in the Dentry tree, it doesn't exist.
+		return nil, syserror.ENOENT
+	}
+	next := nextVFSD.Impl().(*dentry)
+	if symlink, ok := next.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() {
+		// TODO(gvisor.dev/issue/1197): Symlink traversals updates
+		// access time.
+		if err := rp.HandleSymlink(symlink.target); err != nil {
+			return nil, err
+		}
+		goto afterSymlink // don't check the current directory again
+	}
+	rp.Advance()
+	return next, nil
+}
+
+// walkParentDirLocked resolves all but the last path component of rp to an
+// existing directory, starting from the given directory (which is usually
+// rp.Start().Impl().(*dentry)). It does not check that the returned directory
+// is searchable by the provider of rp.
+//
+// walkParentDirLocked is loosely analogous to Linux's
+// fs/namei.c:path_parentat().
+//
+// Preconditions: filesystem.mu must be locked. !rp.Done().
+func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) {
+	for !rp.Final() {
+		next, err := stepLocked(rp, d)
+		if err != nil {
+			return nil, err
+		}
+		d = next
+	}
+	if !d.inode.isDir() {
+		return nil, syserror.ENOTDIR
+	}
+	return d, nil
+}
+
+// resolveLocked resolves rp to an existing file.
+//
+// resolveLocked is loosely analogous to Linux's fs/namei.c:path_lookupat().
+//
+// Preconditions: filesystem.mu must be locked.
+func resolveLocked(rp *vfs.ResolvingPath) (*dentry, error) {
+	d := rp.Start().Impl().(*dentry)
+	for !rp.Done() {
+		next, err := stepLocked(rp, d)
+		if err != nil {
+			return nil, err
+		}
+		d = next
+	}
+	if rp.MustBeDir() && !d.inode.isDir() {
+		return nil, syserror.ENOTDIR
+	}
+	return d, nil
+}
+
+// doCreateAt checks that creating a file at rp is permitted, then invokes
+// create to do so.
+//
+// doCreateAt is loosely analogous to a conjunction of Linux's
+// fs/namei.c:filename_create() and done_path_create().
+//
+// Preconditions: !rp.Done(). For the final path component in rp,
+// !rp.ShouldFollowSymlink().
+func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error {
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+	parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry))
+	if err != nil {
+		return err
+	}
+	if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+		return err
+	}
+	name := rp.Component()
+	if name == "." || name == ".." {
+		return syserror.EEXIST
+	}
+	if len(name) > linux.NAME_MAX {
+		return syserror.ENAMETOOLONG
+	}
+	// Call parent.vfsd.Child() instead of stepLocked() or rp.ResolveChild(),
+	// because if the child exists we want to return EEXIST immediately instead
+	// of attempting symlink/mount traversal.
+	if parent.vfsd.Child(name) != nil {
+		return syserror.EEXIST
+	}
+	if !dir && rp.MustBeDir() {
+		return syserror.ENOENT
+	}
+	// In memfs, the only way to cause a dentry to be disowned is by removing
+	// it from the filesystem, so this check is equivalent to checking if
+	// parent has been removed.
+	if parent.vfsd.IsDisowned() {
+		return syserror.ENOENT
+	}
+	mnt := rp.Mount()
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer mnt.EndWrite()
+	if err := create(parent, name); err != nil {
+		return err
+	}
+	parent.inode.touchCMtime()
+	return nil
+}
+
+// AccessAt implements vfs.Filesystem.Impl.AccessAt.
+func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return err
+	}
+	return d.inode.checkPermissions(creds, ats)
+}
+
+// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt.
+func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return nil, err
+	}
+	if opts.CheckSearchable {
+		if !d.inode.isDir() {
+			return nil, syserror.ENOTDIR
+		}
+		if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil {
+			return nil, err
+		}
+	}
+	d.IncRef()
+	return &d.vfsd, nil
+}
+
+// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt.
+func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry))
+	if err != nil {
+		return nil, err
+	}
+	d.IncRef()
+	return &d.vfsd, nil
+}
+
+// LinkAt implements vfs.FilesystemImpl.LinkAt.
+func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error {
+	return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error {
+		if rp.Mount() != vd.Mount() {
+			return syserror.EXDEV
+		}
+		d := vd.Dentry().Impl().(*dentry)
+		if d.inode.isDir() {
+			return syserror.EPERM
+		}
+		if d.inode.nlink == 0 {
+			return syserror.ENOENT
+		}
+		if d.inode.nlink == maxLinks {
+			return syserror.EMLINK
+		}
+		d.inode.incLinksLocked()
+		child := fs.newDentry(d.inode)
+		parent.vfsd.InsertChild(&child.vfsd, name)
+		parent.inode.impl.(*directory).childList.PushBack(child)
+		return nil
+	})
+}
+
+// MkdirAt implements vfs.FilesystemImpl.MkdirAt.
+func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error {
+	return fs.doCreateAt(rp, true /* dir */, func(parent *dentry, name string) error {
+		if parent.inode.nlink == maxLinks {
+			return syserror.EMLINK
+		}
+		parent.inode.incLinksLocked() // from child's ".."
+		child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode))
+		parent.vfsd.InsertChild(&child.vfsd, name)
+		parent.inode.impl.(*directory).childList.PushBack(child)
+		return nil
+	})
+}
+
+// MknodAt implements vfs.FilesystemImpl.MknodAt.
+func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error {
+	return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error {
+		var childInode *inode
+		switch opts.Mode.FileType() {
+		case 0, linux.S_IFREG:
+			childInode = fs.newRegularFile(rp.Credentials(), opts.Mode)
+		case linux.S_IFIFO:
+			childInode = fs.newNamedPipe(rp.Credentials(), opts.Mode)
+		case linux.S_IFBLK:
+			childInode = fs.newDeviceFile(rp.Credentials(), opts.Mode, vfs.BlockDevice, opts.DevMajor, opts.DevMinor)
+		case linux.S_IFCHR:
+			childInode = fs.newDeviceFile(rp.Credentials(), opts.Mode, vfs.CharDevice, opts.DevMajor, opts.DevMinor)
+		case linux.S_IFSOCK:
+			childInode = fs.newSocketFile(rp.Credentials(), opts.Mode, opts.Endpoint)
+		default:
+			return syserror.EINVAL
+		}
+		child := fs.newDentry(childInode)
+		parent.vfsd.InsertChild(&child.vfsd, name)
+		parent.inode.impl.(*directory).childList.PushBack(child)
+		return nil
+	})
+}
+
+// OpenAt implements vfs.FilesystemImpl.OpenAt.
+func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) {
+	if opts.Flags&linux.O_TMPFILE != 0 {
+		// Not yet supported.
+		return nil, syserror.EOPNOTSUPP
+	}
+
+	// Handle O_CREAT and !O_CREAT separately, since in the latter case we
+	// don't need fs.mu for writing.
+	if opts.Flags&linux.O_CREAT == 0 {
+		fs.mu.RLock()
+		defer fs.mu.RUnlock()
+		d, err := resolveLocked(rp)
+		if err != nil {
+			return nil, err
+		}
+		return d.open(ctx, rp, &opts, false /* afterCreate */)
+	}
+
+	mustCreate := opts.Flags&linux.O_EXCL != 0
+	start := rp.Start().Impl().(*dentry)
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+	if rp.Done() {
+		// Reject attempts to open directories with O_CREAT.
+		if rp.MustBeDir() {
+			return nil, syserror.EISDIR
+		}
+		if mustCreate {
+			return nil, syserror.EEXIST
+		}
+		return start.open(ctx, rp, &opts, false /* afterCreate */)
+	}
+afterTrailingSymlink:
+	parent, err := walkParentDirLocked(rp, start)
+	if err != nil {
+		return nil, err
+	}
+	// Check for search permission in the parent directory.
+	if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil {
+		return nil, err
+	}
+	// Reject attempts to open directories with O_CREAT.
+	if rp.MustBeDir() {
+		return nil, syserror.EISDIR
+	}
+	name := rp.Component()
+	if name == "." || name == ".." {
+		return nil, syserror.EISDIR
+	}
+	// Determine whether or not we need to create a file.
+	child, err := stepLocked(rp, parent)
+	if err == syserror.ENOENT {
+		// Already checked for searchability above; now check for writability.
+		if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil {
+			return nil, err
+		}
+		if err := rp.Mount().CheckBeginWrite(); err != nil {
+			return nil, err
+		}
+		defer rp.Mount().EndWrite()
+		// Create and open the child.
+		child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode))
+		parent.vfsd.InsertChild(&child.vfsd, name)
+		parent.inode.impl.(*directory).childList.PushBack(child)
+		fd, err := child.open(ctx, rp, &opts, true)
+		if err != nil {
+			return nil, err
+		}
+		parent.inode.touchCMtime()
+		return fd, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	// Do we need to resolve a trailing symlink?
+	if !rp.Done() {
+		start = parent
+		goto afterTrailingSymlink
+	}
+	// Open existing file.
+	if mustCreate {
+		return nil, syserror.EEXIST
+	}
+	return child.open(ctx, rp, &opts, false)
+}
+
+func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, afterCreate bool) (*vfs.FileDescription, error) {
+	ats := vfs.AccessTypesForOpenFlags(opts)
+	if !afterCreate {
+		if err := d.inode.checkPermissions(rp.Credentials(), ats); err != nil {
+			return nil, err
+		}
+	}
+	switch impl := d.inode.impl.(type) {
+	case *regularFile:
+		var fd regularFileFD
+		if err := fd.vfsfd.Init(&fd, opts.Flags, rp.Mount(), &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+			return nil, err
+		}
+		if opts.Flags&linux.O_TRUNC != 0 {
+			if _, err := impl.truncate(0); err != nil {
+				return nil, err
+			}
+		}
+		return &fd.vfsfd, nil
+	case *directory:
+		// Can't open directories writably.
+		if ats&vfs.MayWrite != 0 {
+			return nil, syserror.EISDIR
+		}
+		var fd directoryFD
+		if err := fd.vfsfd.Init(&fd, opts.Flags, rp.Mount(), &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+			return nil, err
+		}
+		return &fd.vfsfd, nil
+	case *symlink:
+		// Can't open symlinks without O_PATH (which is unimplemented).
+		return nil, syserror.ELOOP
+	case *namedPipe:
+		return impl.pipe.Open(ctx, rp.Mount(), &d.vfsd, opts.Flags)
+	case *deviceFile:
+		return rp.VirtualFilesystem().OpenDeviceSpecialFile(ctx, rp.Mount(), &d.vfsd, impl.kind, impl.major, impl.minor, opts)
+	case *socketFile:
+		return nil, syserror.ENXIO
+	default:
+		panic(fmt.Sprintf("unknown inode type: %T", d.inode.impl))
+	}
+}
+
+// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt.
+func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return "", err
+	}
+	symlink, ok := d.inode.impl.(*symlink)
+	if !ok {
+		return "", syserror.EINVAL
+	}
+	symlink.inode.touchAtime(rp.Mount())
+	return symlink.target, nil
+}
+
+// RenameAt implements vfs.FilesystemImpl.RenameAt.
+func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error {
+	if opts.Flags != 0 {
+		// TODO(b/145974740): Support renameat2 flags.
+		return syserror.EINVAL
+	}
+
+	// Resolve newParent first to verify that it's on this Mount.
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+	newParent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry))
+	if err != nil {
+		return err
+	}
+	newName := rp.Component()
+	if newName == "." || newName == ".." {
+		return syserror.EBUSY
+	}
+	mnt := rp.Mount()
+	if mnt != oldParentVD.Mount() {
+		return syserror.EXDEV
+	}
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer mnt.EndWrite()
+
+	oldParent := oldParentVD.Dentry().Impl().(*dentry)
+	if err := oldParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+		return err
+	}
+	// Call vfs.Dentry.Child() instead of stepLocked() or rp.ResolveChild(),
+	// because if the existing child is a symlink or mount point then we want
+	// to rename over it rather than follow it.
+	renamedVFSD := oldParent.vfsd.Child(oldName)
+	if renamedVFSD == nil {
+		return syserror.ENOENT
+	}
+	renamed := renamedVFSD.Impl().(*dentry)
+	if renamed.inode.isDir() {
+		if renamed == newParent || renamedVFSD.IsAncestorOf(&newParent.vfsd) {
+			return syserror.EINVAL
+		}
+		if oldParent != newParent {
+			// Writability is needed to change renamed's "..".
+			if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil {
+				return err
+			}
+		}
+	} else {
+		if opts.MustBeDir || rp.MustBeDir() {
+			return syserror.ENOTDIR
+		}
+	}
+
+	if err := newParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+		return err
+	}
+	replacedVFSD := newParent.vfsd.Child(newName)
+	var replaced *dentry
+	if replacedVFSD != nil {
+		replaced = replacedVFSD.Impl().(*dentry)
+		if replaced.inode.isDir() {
+			if !renamed.inode.isDir() {
+				return syserror.EISDIR
+			}
+			if replaced.vfsd.HasChildren() {
+				return syserror.ENOTEMPTY
+			}
+		} else {
+			if rp.MustBeDir() {
+				return syserror.ENOTDIR
+			}
+			if renamed.inode.isDir() {
+				return syserror.ENOTDIR
+			}
+		}
+	} else {
+		if renamed.inode.isDir() && newParent.inode.nlink == maxLinks {
+			return syserror.EMLINK
+		}
+	}
+	if newParent.vfsd.IsDisowned() {
+		return syserror.ENOENT
+	}
+
+	// Linux places this check before some of those above; we do it here for
+	// simplicity, under the assumption that applications are not intentionally
+	// doing noop renames expecting them to succeed where non-noop renames
+	// would fail.
+	if renamedVFSD == replacedVFSD {
+		return nil
+	}
+	vfsObj := rp.VirtualFilesystem()
+	oldParentDir := oldParent.inode.impl.(*directory)
+	newParentDir := newParent.inode.impl.(*directory)
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
+	if err := vfsObj.PrepareRenameDentry(mntns, renamedVFSD, replacedVFSD); err != nil {
+		return err
+	}
+	if replaced != nil {
+		newParentDir.childList.Remove(replaced)
+		if replaced.inode.isDir() {
+			newParent.inode.decLinksLocked() // from replaced's ".."
+		}
+		replaced.inode.decLinksLocked()
+	}
+	oldParentDir.childList.Remove(renamed)
+	newParentDir.childList.PushBack(renamed)
+	if renamed.inode.isDir() {
+		oldParent.inode.decLinksLocked()
+		newParent.inode.incLinksLocked()
+	}
+	oldParent.inode.touchCMtime()
+	newParent.inode.touchCMtime()
+	renamed.inode.touchCtime()
+	// TODO(gvisor.dev/issue/1197): Update timestamps and parent directory
+	// sizes.
+	vfsObj.CommitRenameReplaceDentry(renamedVFSD, &newParent.vfsd, newName, replacedVFSD)
+	return nil
+}
+
+// RmdirAt implements vfs.FilesystemImpl.RmdirAt.
+func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error {
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+	parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry))
+	if err != nil {
+		return err
+	}
+	if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+		return err
+	}
+	name := rp.Component()
+	if name == "." {
+		return syserror.EINVAL
+	}
+	if name == ".." {
+		return syserror.ENOTEMPTY
+	}
+	childVFSD := parent.vfsd.Child(name)
+	if childVFSD == nil {
+		return syserror.ENOENT
+	}
+	child := childVFSD.Impl().(*dentry)
+	if !child.inode.isDir() {
+		return syserror.ENOTDIR
+	}
+	if childVFSD.HasChildren() {
+		return syserror.ENOTEMPTY
+	}
+	mnt := rp.Mount()
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer mnt.EndWrite()
+	vfsObj := rp.VirtualFilesystem()
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
+	if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil {
+		return err
+	}
+	parent.inode.impl.(*directory).childList.Remove(child)
+	parent.inode.decLinksLocked() // from child's ".."
+	child.inode.decLinksLocked()
+	vfsObj.CommitDeleteDentry(childVFSD)
+	parent.inode.touchCMtime()
+	return nil
+}
+
+// SetStatAt implements vfs.FilesystemImpl.SetStatAt.
+func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return err
+	}
+	return d.inode.setStat(ctx, rp.Credentials(), &opts.Stat)
+}
+
+// StatAt implements vfs.FilesystemImpl.StatAt.
+func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return linux.Statx{}, err
+	}
+	var stat linux.Statx
+	d.inode.statTo(&stat)
+	return stat, nil
+}
+
+// StatFSAt implements vfs.FilesystemImpl.StatFSAt.
+func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	_, err := resolveLocked(rp)
+	if err != nil {
+		return linux.Statfs{}, err
+	}
+	// TODO(gvisor.dev/issue/1197): Actually implement statfs.
+	return linux.Statfs{}, syserror.ENOSYS
+}
+
+// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt.
+func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error {
+	return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error {
+		child := fs.newDentry(fs.newSymlink(rp.Credentials(), target))
+		parent.vfsd.InsertChild(&child.vfsd, name)
+		parent.inode.impl.(*directory).childList.PushBack(child)
+		return nil
+	})
+}
+
+// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt.
+func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error {
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+	parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry))
+	if err != nil {
+		return err
+	}
+	if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+		return err
+	}
+	name := rp.Component()
+	if name == "." || name == ".." {
+		return syserror.EISDIR
+	}
+	childVFSD := parent.vfsd.Child(name)
+	if childVFSD == nil {
+		return syserror.ENOENT
+	}
+	child := childVFSD.Impl().(*dentry)
+	if child.inode.isDir() {
+		return syserror.EISDIR
+	}
+	if rp.MustBeDir() {
+		return syserror.ENOTDIR
+	}
+	mnt := rp.Mount()
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer mnt.EndWrite()
+	vfsObj := rp.VirtualFilesystem()
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
+	if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil {
+		return err
+	}
+	parent.inode.impl.(*directory).childList.Remove(child)
+	child.inode.decLinksLocked()
+	vfsObj.CommitDeleteDentry(childVFSD)
+	parent.inode.touchCMtime()
+	return nil
+}
+
+// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt.
+func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath) (transport.BoundEndpoint, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return nil, err
+	}
+	switch impl := d.inode.impl.(type) {
+	case *socketFile:
+		return impl.ep, nil
+	default:
+		return nil, syserror.ECONNREFUSED
+	}
+}
+
+// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt.
+func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return nil, err
+	}
+	return d.inode.listxattr(size)
+}
+
+// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt.
+func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return "", err
+	}
+	return d.inode.getxattr(rp.Credentials(), &opts)
+}
+
+// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt.
+func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return err
+	}
+	return d.inode.setxattr(rp.Credentials(), &opts)
+}
+
+// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt.
+func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	d, err := resolveLocked(rp)
+	if err != nil {
+		return err
+	}
+	return d.inode.removexattr(rp.Credentials(), name)
+}
+
+// PrependPath implements vfs.FilesystemImpl.PrependPath.
+func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error {
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+	return vfs.GenericPrependPath(vfsroot, vd, b)
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go
new file mode 100755
index 000000000..8d77b3fa8
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go
@@ -0,0 +1,38 @@
+// Copyright 2019 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 tmpfs
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/pipe"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+type namedPipe struct {
+	inode inode
+
+	pipe *pipe.VFSPipe
+}
+
+// Preconditions:
+//   * fs.mu must be locked.
+//   * rp.Mount().CheckBeginWrite() has been called successfully.
+func (fs *filesystem) newNamedPipe(creds *auth.Credentials, mode linux.FileMode) *inode {
+	file := &namedPipe{pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize)}
+	file.inode.init(file, fs, creds, linux.S_IFIFO|mode)
+	file.inode.nlink = 1 // Only the parent has a link.
+	return &file.inode
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go
new file mode 100755
index 000000000..57e5e28ec
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go
@@ -0,0 +1,579 @@
+// Copyright 2019 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 tmpfs
+
+import (
+	"fmt"
+	"io"
+	"math"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/safemem"
+	"gvisor.dev/gvisor/pkg/sentry/fs"
+	"gvisor.dev/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.dev/gvisor/pkg/sentry/fs/lock"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
+	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
+	"gvisor.dev/gvisor/pkg/sentry/usage"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// regularFile is a regular (=S_IFREG) tmpfs file.
+type regularFile struct {
+	inode inode
+
+	// memFile is a platform.File used to allocate pages to this regularFile.
+	memFile *pgalloc.MemoryFile
+
+	// mapsMu protects mappings.
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// mappings tracks mappings of the file into memmap.MappingSpaces.
+	//
+	// Protected by mapsMu.
+	mappings memmap.MappingSet
+
+	// writableMappingPages tracks how many pages of virtual memory are mapped
+	// as potentially writable from this file. If a page has multiple mappings,
+	// each mapping is counted separately.
+	//
+	// This counter is susceptible to overflow as we can potentially count
+	// mappings from many VMAs. We count pages rather than bytes to slightly
+	// mitigate this.
+	//
+	// Protected by mapsMu.
+	writableMappingPages uint64
+
+	// dataMu protects the fields below.
+	dataMu sync.RWMutex
+
+	// data maps offsets into the file to offsets into memFile that store
+	// the file's data.
+	//
+	// Protected by dataMu.
+	data fsutil.FileRangeSet
+
+	// seals represents file seals on this inode.
+	//
+	// Protected by dataMu.
+	seals uint32
+
+	// size is the size of data.
+	//
+	// Protected by both dataMu and inode.mu; reading it requires holding
+	// either mutex, while writing requires holding both AND using atomics.
+	// Readers that do not require consistency (like Stat) may read the
+	// value atomically without holding either lock.
+	size uint64
+}
+
+func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode {
+	file := &regularFile{
+		memFile: fs.memFile,
+	}
+	file.inode.init(file, fs, creds, linux.S_IFREG|mode)
+	file.inode.nlink = 1 // from parent directory
+	return &file.inode
+}
+
+// truncate grows or shrinks the file to the given size. It returns true if the
+// file size was updated.
+func (rf *regularFile) truncate(newSize uint64) (bool, error) {
+	rf.inode.mu.Lock()
+	defer rf.inode.mu.Unlock()
+	return rf.truncateLocked(newSize)
+}
+
+// Preconditions: rf.inode.mu must be held.
+func (rf *regularFile) truncateLocked(newSize uint64) (bool, error) {
+	oldSize := rf.size
+	if newSize == oldSize {
+		// Nothing to do.
+		return false, nil
+	}
+
+	// Need to hold inode.mu and dataMu while modifying size.
+	rf.dataMu.Lock()
+	if newSize > oldSize {
+		// Can we grow the file?
+		if rf.seals&linux.F_SEAL_GROW != 0 {
+			rf.dataMu.Unlock()
+			return false, syserror.EPERM
+		}
+		// We only need to update the file size.
+		atomic.StoreUint64(&rf.size, newSize)
+		rf.dataMu.Unlock()
+		return true, nil
+	}
+
+	// We are shrinking the file. First check if this is allowed.
+	if rf.seals&linux.F_SEAL_SHRINK != 0 {
+		rf.dataMu.Unlock()
+		return false, syserror.EPERM
+	}
+
+	// Update the file size.
+	atomic.StoreUint64(&rf.size, newSize)
+	rf.dataMu.Unlock()
+
+	// Invalidate past translations of truncated pages.
+	oldpgend := fs.OffsetPageEnd(int64(oldSize))
+	newpgend := fs.OffsetPageEnd(int64(newSize))
+	if newpgend < oldpgend {
+		rf.mapsMu.Lock()
+		rf.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{
+			// Compare Linux's mm/shmem.c:shmem_setattr() =>
+			// mm/memory.c:unmap_mapping_range(evencows=1).
+			InvalidatePrivate: true,
+		})
+		rf.mapsMu.Unlock()
+	}
+
+	// We are now guaranteed that there are no translations of truncated pages,
+	// and can remove them.
+	rf.dataMu.Lock()
+	rf.data.Truncate(newSize, rf.memFile)
+	rf.dataMu.Unlock()
+	return true, nil
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (rf *regularFile) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error {
+	rf.mapsMu.Lock()
+	defer rf.mapsMu.Unlock()
+	rf.dataMu.RLock()
+	defer rf.dataMu.RUnlock()
+
+	// Reject writable mapping if F_SEAL_WRITE is set.
+	if rf.seals&linux.F_SEAL_WRITE != 0 && writable {
+		return syserror.EPERM
+	}
+
+	rf.mappings.AddMapping(ms, ar, offset, writable)
+	if writable {
+		pagesBefore := rf.writableMappingPages
+
+		// ar is guaranteed to be page aligned per memmap.Mappable.
+		rf.writableMappingPages += uint64(ar.Length() / usermem.PageSize)
+
+		if rf.writableMappingPages < pagesBefore {
+			panic(fmt.Sprintf("Overflow while mapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages))
+		}
+	}
+
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (rf *regularFile) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) {
+	rf.mapsMu.Lock()
+	defer rf.mapsMu.Unlock()
+
+	rf.mappings.RemoveMapping(ms, ar, offset, writable)
+
+	if writable {
+		pagesBefore := rf.writableMappingPages
+
+		// ar is guaranteed to be page aligned per memmap.Mappable.
+		rf.writableMappingPages -= uint64(ar.Length() / usermem.PageSize)
+
+		if rf.writableMappingPages > pagesBefore {
+			panic(fmt.Sprintf("Underflow while unmapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages))
+		}
+	}
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (rf *regularFile) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error {
+	return rf.AddMapping(ctx, ms, dstAR, offset, writable)
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (rf *regularFile) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	rf.dataMu.Lock()
+	defer rf.dataMu.Unlock()
+
+	// Constrain translations to f.attr.Size (rounded up) to prevent
+	// translation to pages that may be concurrently truncated.
+	pgend := fs.OffsetPageEnd(int64(rf.size))
+	var beyondEOF bool
+	if required.End > pgend {
+		if required.Start >= pgend {
+			return nil, &memmap.BusError{io.EOF}
+		}
+		beyondEOF = true
+		required.End = pgend
+	}
+	if optional.End > pgend {
+		optional.End = pgend
+	}
+
+	cerr := rf.data.Fill(ctx, required, optional, rf.memFile, usage.Tmpfs, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) {
+		// Newly-allocated pages are zeroed, so we don't need to do anything.
+		return dsts.NumBytes(), nil
+	})
+
+	var ts []memmap.Translation
+	var translatedEnd uint64
+	for seg := rf.data.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() {
+		segMR := seg.Range().Intersect(optional)
+		ts = append(ts, memmap.Translation{
+			Source: segMR,
+			File:   rf.memFile,
+			Offset: seg.FileRangeOf(segMR).Start,
+			Perms:  usermem.AnyAccess,
+		})
+		translatedEnd = segMR.End
+	}
+
+	// Don't return the error returned by f.data.Fill if it occurred outside of
+	// required.
+	if translatedEnd < required.End && cerr != nil {
+		return ts, &memmap.BusError{cerr}
+	}
+	if beyondEOF {
+		return ts, &memmap.BusError{io.EOF}
+	}
+	return ts, nil
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (*regularFile) InvalidateUnsavable(context.Context) error {
+	return nil
+}
+
+type regularFileFD struct {
+	fileDescription
+
+	// off is the file offset. off is accessed using atomic memory operations.
+	// offMu serializes operations that may mutate off.
+	off   int64
+	offMu sync.Mutex
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *regularFileFD) Release() {
+	// noop
+}
+
+// PRead implements vfs.FileDescriptionImpl.PRead.
+func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+	f := fd.inode().impl.(*regularFile)
+	rw := getRegularFileReadWriter(f, offset)
+	n, err := dst.CopyOutFrom(ctx, rw)
+	putRegularFileReadWriter(rw)
+	fd.inode().touchAtime(fd.vfsfd.Mount())
+	return n, err
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	fd.offMu.Lock()
+	n, err := fd.PRead(ctx, dst, fd.off, opts)
+	fd.off += n
+	fd.offMu.Unlock()
+	return n, err
+}
+
+// PWrite implements vfs.FileDescriptionImpl.PWrite.
+func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	srclen := src.NumBytes()
+	if srclen == 0 {
+		return 0, nil
+	}
+	f := fd.inode().impl.(*regularFile)
+	if end := offset + srclen; end < offset {
+		// Overflow.
+		return 0, syserror.EFBIG
+	}
+
+	var err error
+	srclen, err = vfs.CheckLimit(ctx, offset, srclen)
+	if err != nil {
+		return 0, err
+	}
+	src = src.TakeFirst64(srclen)
+
+	f.inode.mu.Lock()
+	rw := getRegularFileReadWriter(f, offset)
+	n, err := src.CopyInTo(ctx, rw)
+	fd.inode().touchCMtimeLocked()
+	f.inode.mu.Unlock()
+	putRegularFileReadWriter(rw)
+	return n, err
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	fd.offMu.Lock()
+	n, err := fd.PWrite(ctx, src, fd.off, opts)
+	fd.off += n
+	fd.offMu.Unlock()
+	return n, err
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+	fd.offMu.Lock()
+	defer fd.offMu.Unlock()
+	switch whence {
+	case linux.SEEK_SET:
+		// use offset as specified
+	case linux.SEEK_CUR:
+		offset += fd.off
+	case linux.SEEK_END:
+		offset += int64(atomic.LoadUint64(&fd.inode().impl.(*regularFile).size))
+	default:
+		return 0, syserror.EINVAL
+	}
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	fd.off = offset
+	return offset, nil
+}
+
+// Sync implements vfs.FileDescriptionImpl.Sync.
+func (fd *regularFileFD) Sync(ctx context.Context) error {
+	return nil
+}
+
+// LockBSD implements vfs.FileDescriptionImpl.LockBSD.
+func (fd *regularFileFD) LockBSD(ctx context.Context, uid lock.UniqueID, t lock.LockType, block lock.Blocker) error {
+	return fd.inode().lockBSD(uid, t, block)
+}
+
+// UnlockBSD implements vfs.FileDescriptionImpl.UnlockBSD.
+func (fd *regularFileFD) UnlockBSD(ctx context.Context, uid lock.UniqueID) error {
+	fd.inode().unlockBSD(uid)
+	return nil
+}
+
+// LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX.
+func (fd *regularFileFD) LockPOSIX(ctx context.Context, uid lock.UniqueID, t lock.LockType, rng lock.LockRange, block lock.Blocker) error {
+	return fd.inode().lockPOSIX(uid, t, rng, block)
+}
+
+// UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX.
+func (fd *regularFileFD) UnlockPOSIX(ctx context.Context, uid lock.UniqueID, rng lock.LockRange) error {
+	fd.inode().unlockPOSIX(uid, rng)
+	return nil
+}
+
+// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap.
+func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error {
+	file := fd.inode().impl.(*regularFile)
+	return vfs.GenericConfigureMMap(&fd.vfsfd, file, opts)
+}
+
+// regularFileReadWriter implements safemem.Reader and Safemem.Writer.
+type regularFileReadWriter struct {
+	file *regularFile
+
+	// Offset into the file to read/write at. Note that this may be
+	// different from the FD offset if PRead/PWrite is used.
+	off uint64
+}
+
+var regularFileReadWriterPool = sync.Pool{
+	New: func() interface{} {
+		return &regularFileReadWriter{}
+	},
+}
+
+func getRegularFileReadWriter(file *regularFile, offset int64) *regularFileReadWriter {
+	rw := regularFileReadWriterPool.Get().(*regularFileReadWriter)
+	rw.file = file
+	rw.off = uint64(offset)
+	return rw
+}
+
+func putRegularFileReadWriter(rw *regularFileReadWriter) {
+	rw.file = nil
+	regularFileReadWriterPool.Put(rw)
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	rw.file.dataMu.RLock()
+	defer rw.file.dataMu.RUnlock()
+	size := rw.file.size
+
+	// Compute the range to read (limited by file size and overflow-checked).
+	if rw.off >= size {
+		return 0, io.EOF
+	}
+	end := size
+	if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end {
+		end = rend
+	}
+
+	var done uint64
+	seg, gap := rw.file.data.Find(uint64(rw.off))
+	for rw.off < end {
+		mr := memmap.MappableRange{uint64(rw.off), uint64(end)}
+		switch {
+		case seg.Ok():
+			// Get internal mappings.
+			ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read)
+			if err != nil {
+				return done, err
+			}
+
+			// Copy from internal mappings.
+			n, err := safemem.CopySeq(dsts, ims)
+			done += n
+			rw.off += uint64(n)
+			dsts = dsts.DropFirst64(n)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			// Tmpfs holes are zero-filled.
+			gapmr := gap.Range().Intersect(mr)
+			dst := dsts.TakeFirst64(gapmr.Length())
+			n, err := safemem.ZeroSeq(dst)
+			done += n
+			rw.off += uint64(n)
+			dsts = dsts.DropFirst64(n)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{}
+		}
+	}
+	return done, nil
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+//
+// Preconditions: inode.mu must be held.
+func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	// Hold dataMu so we can modify size.
+	rw.file.dataMu.Lock()
+	defer rw.file.dataMu.Unlock()
+
+	// Compute the range to write (overflow-checked).
+	end := rw.off + srcs.NumBytes()
+	if end <= rw.off {
+		end = math.MaxInt64
+	}
+
+	// Check if seals prevent either file growth or all writes.
+	switch {
+	case rw.file.seals&linux.F_SEAL_WRITE != 0: // Write sealed
+		return 0, syserror.EPERM
+	case end > rw.file.size && rw.file.seals&linux.F_SEAL_GROW != 0: // Grow sealed
+		// When growth is sealed, Linux effectively allows writes which would
+		// normally grow the file to partially succeed up to the current EOF,
+		// rounded down to the page boundary before the EOF.
+		//
+		// This happens because writes (and thus the growth check) for tmpfs
+		// files proceed page-by-page on Linux, and the final write to the page
+		// containing EOF fails, resulting in a partial write up to the start of
+		// that page.
+		//
+		// To emulate this behaviour, artifically truncate the write to the
+		// start of the page containing the current EOF.
+		//
+		// See Linux, mm/filemap.c:generic_perform_write() and
+		// mm/shmem.c:shmem_write_begin().
+		if pgstart := uint64(usermem.Addr(rw.file.size).RoundDown()); end > pgstart {
+			end = pgstart
+		}
+		if end <= rw.off {
+			// Truncation would result in no data being written.
+			return 0, syserror.EPERM
+		}
+	}
+
+	// Page-aligned mr for when we need to allocate memory. RoundUp can't
+	// overflow since end is an int64.
+	pgstartaddr := usermem.Addr(rw.off).RoundDown()
+	pgendaddr, _ := usermem.Addr(end).RoundUp()
+	pgMR := memmap.MappableRange{uint64(pgstartaddr), uint64(pgendaddr)}
+
+	var (
+		done   uint64
+		retErr error
+	)
+	seg, gap := rw.file.data.Find(uint64(rw.off))
+	for rw.off < end {
+		mr := memmap.MappableRange{uint64(rw.off), uint64(end)}
+		switch {
+		case seg.Ok():
+			// Get internal mappings.
+			ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Write)
+			if err != nil {
+				retErr = err
+				goto exitLoop
+			}
+
+			// Copy to internal mappings.
+			n, err := safemem.CopySeq(ims, srcs)
+			done += n
+			rw.off += uint64(n)
+			srcs = srcs.DropFirst64(n)
+			if err != nil {
+				retErr = err
+				goto exitLoop
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			// Allocate memory for the write.
+			gapMR := gap.Range().Intersect(pgMR)
+			fr, err := rw.file.memFile.Allocate(gapMR.Length(), usage.Tmpfs)
+			if err != nil {
+				retErr = err
+				goto exitLoop
+			}
+
+			// Write to that memory as usual.
+			seg, gap = rw.file.data.Insert(gap, gapMR, fr.Start), fsutil.FileRangeGapIterator{}
+		}
+	}
+exitLoop:
+	// If the write ends beyond the file's previous size, it causes the
+	// file to grow.
+	if rw.off > rw.file.size {
+		rw.file.size = rw.off
+	}
+
+	return done, retErr
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/socket_file.go b/pkg/sentry/fsimpl/tmpfs/socket_file.go
new file mode 100755
index 000000000..25c2321af
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/socket_file.go
@@ -0,0 +1,34 @@
+// Copyright 2019 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 tmpfs
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/socket/unix/transport"
+)
+
+// socketFile is a socket (=S_IFSOCK) tmpfs file.
+type socketFile struct {
+	inode inode
+	ep    transport.BoundEndpoint
+}
+
+func (fs *filesystem) newSocketFile(creds *auth.Credentials, mode linux.FileMode, ep transport.BoundEndpoint) *inode {
+	file := &socketFile{ep: ep}
+	file.inode.init(file, fs, creds, mode)
+	file.inode.nlink = 1 // from parent directory
+	return &file.inode
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go
new file mode 100755
index 000000000..47e075ed4
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/symlink.go
@@ -0,0 +1,37 @@
+// Copyright 2019 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 tmpfs
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+)
+
+type symlink struct {
+	inode  inode
+	target string // immutable
+}
+
+func (fs *filesystem) newSymlink(creds *auth.Credentials, target string) *inode {
+	link := &symlink{
+		target: target,
+	}
+	link.inode.init(link, fs, creds, linux.S_IFLNK|0777)
+	link.inode.nlink = 1 // from parent directory
+	return &link.inode
+}
+
+// O_PATH is unimplemented, so there's no way to get a FileDescription
+// representing a symlink yet.
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
new file mode 100755
index 000000000..a59b24d45
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
@@ -0,0 +1,641 @@
+// Copyright 2019 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 tmpfs provides a filesystem implementation that behaves like tmpfs:
+// the Dentry tree is the sole source of truth for the state of the filesystem.
+//
+// Lock order:
+//
+// filesystem.mu
+//   inode.mu
+//     regularFileFD.offMu
+//       regularFile.mapsMu
+//         regularFile.dataMu
+package tmpfs
+
+import (
+	"fmt"
+	"math"
+	"strings"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/time"
+	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/sentry/vfs/lock"
+	"gvisor.dev/gvisor/pkg/sentry/vfs/memxattr"
+	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// Name is the default filesystem name.
+const Name = "tmpfs"
+
+// FilesystemType implements vfs.FilesystemType.
+type FilesystemType struct{}
+
+// filesystem implements vfs.FilesystemImpl.
+type filesystem struct {
+	vfsfs vfs.Filesystem
+
+	// memFile is used to allocate pages to for regular files.
+	memFile *pgalloc.MemoryFile
+
+	// clock is a realtime clock used to set timestamps in file operations.
+	clock time.Clock
+
+	// mu serializes changes to the Dentry tree.
+	mu sync.RWMutex
+
+	nextInoMinusOne uint64 // accessed using atomic memory operations
+}
+
+// Name implements vfs.FilesystemType.Name.
+func (FilesystemType) Name() string {
+	return Name
+}
+
+// FilesystemOpts is used to pass configuration data to tmpfs.
+type FilesystemOpts struct {
+	// RootFileType is the FileType of the filesystem root. Valid values
+	// are: S_IFDIR, S_IFREG, and S_IFLNK. Defaults to S_IFDIR.
+	RootFileType uint16
+
+	// RootSymlinkTarget is the target of the root symlink. Only valid if
+	// RootFileType == S_IFLNK.
+	RootSymlinkTarget string
+
+	// FilesystemType allows setting a different FilesystemType for this
+	// tmpfs filesystem. This allows tmpfs to "impersonate" other
+	// filesystems, like ramdiskfs and cgroupfs.
+	FilesystemType vfs.FilesystemType
+}
+
+// GetFilesystem implements vfs.FilesystemType.GetFilesystem.
+func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
+	memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx)
+	if memFileProvider == nil {
+		panic("MemoryFileProviderFromContext returned nil")
+	}
+	clock := time.RealtimeClockFromContext(ctx)
+	fs := filesystem{
+		memFile: memFileProvider.MemoryFile(),
+		clock:   clock,
+	}
+
+	rootFileType := uint16(linux.S_IFDIR)
+	newFSType := vfs.FilesystemType(&fstype)
+	tmpfsOpts, ok := opts.InternalData.(FilesystemOpts)
+	if ok {
+		if tmpfsOpts.RootFileType != 0 {
+			rootFileType = tmpfsOpts.RootFileType
+		}
+		if tmpfsOpts.FilesystemType != nil {
+			newFSType = tmpfsOpts.FilesystemType
+		}
+	}
+
+	fs.vfsfs.Init(vfsObj, newFSType, &fs)
+
+	var root *inode
+	switch rootFileType {
+	case linux.S_IFREG:
+		root = fs.newRegularFile(creds, 0777)
+	case linux.S_IFLNK:
+		root = fs.newSymlink(creds, tmpfsOpts.RootSymlinkTarget)
+	case linux.S_IFDIR:
+		root = fs.newDirectory(creds, 01777)
+	default:
+		return nil, nil, fmt.Errorf("invalid tmpfs root file type: %#o", rootFileType)
+	}
+	return &fs.vfsfs, &fs.newDentry(root).vfsd, nil
+}
+
+// Release implements vfs.FilesystemImpl.Release.
+func (fs *filesystem) Release() {
+}
+
+// dentry implements vfs.DentryImpl.
+type dentry struct {
+	vfsd vfs.Dentry
+
+	// inode is the inode represented by this dentry. Multiple Dentries may
+	// share a single non-directory inode (with hard links). inode is
+	// immutable.
+	inode *inode
+
+	// tmpfs doesn't count references on dentries; because the dentry tree is
+	// the sole source of truth, it is by definition always consistent with the
+	// state of the filesystem. However, it does count references on inodes,
+	// because inode resources are released when all references are dropped.
+	// (tmpfs doesn't really have resources to release, but we implement
+	// reference counting because tmpfs regular files will.)
+
+	// dentryEntry (ugh) links dentries into their parent directory.childList.
+	dentryEntry
+}
+
+func (fs *filesystem) newDentry(inode *inode) *dentry {
+	d := &dentry{
+		inode: inode,
+	}
+	d.vfsd.Init(d)
+	return d
+}
+
+// IncRef implements vfs.DentryImpl.IncRef.
+func (d *dentry) IncRef() {
+	d.inode.incRef()
+}
+
+// TryIncRef implements vfs.DentryImpl.TryIncRef.
+func (d *dentry) TryIncRef() bool {
+	return d.inode.tryIncRef()
+}
+
+// DecRef implements vfs.DentryImpl.DecRef.
+func (d *dentry) DecRef() {
+	d.inode.decRef()
+}
+
+// inode represents a filesystem object.
+type inode struct {
+	// clock is a realtime clock used to set timestamps in file operations.
+	clock time.Clock
+
+	// refs is a reference count. refs is accessed using atomic memory
+	// operations.
+	//
+	// A reference is held on all inodes that are reachable in the filesystem
+	// tree. For non-directories (which may have multiple hard links), this
+	// means that a reference is dropped when nlink reaches 0. For directories,
+	// nlink never reaches 0 due to the "." entry; instead,
+	// filesystem.RmdirAt() drops the reference.
+	refs int64
+
+	// xattrs implements extended attributes.
+	//
+	// TODO(b/148380782): Support xattrs other than user.*
+	xattrs memxattr.SimpleExtendedAttributes
+
+	// Inode metadata. Writing multiple fields atomically requires holding
+	// mu, othewise atomic operations can be used.
+	mu    sync.Mutex
+	mode  uint32 // file type and mode
+	nlink uint32 // protected by filesystem.mu instead of inode.mu
+	uid   uint32 // auth.KUID, but stored as raw uint32 for sync/atomic
+	gid   uint32 // auth.KGID, but ...
+	ino   uint64 // immutable
+
+	// Linux's tmpfs has no concept of btime.
+	atime int64 // nanoseconds
+	ctime int64 // nanoseconds
+	mtime int64 // nanoseconds
+
+	// Only meaningful for device special files.
+	rdevMajor uint32
+	rdevMinor uint32
+
+	// Advisory file locks, which lock at the inode level.
+	locks lock.FileLocks
+
+	impl interface{} // immutable
+}
+
+const maxLinks = math.MaxUint32
+
+func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, mode linux.FileMode) {
+	if mode.FileType() == 0 {
+		panic("file type is required in FileMode")
+	}
+	i.clock = fs.clock
+	i.refs = 1
+	i.mode = uint32(mode)
+	i.uid = uint32(creds.EffectiveKUID)
+	i.gid = uint32(creds.EffectiveKGID)
+	i.ino = atomic.AddUint64(&fs.nextInoMinusOne, 1)
+	// Tmpfs creation sets atime, ctime, and mtime to current time.
+	now := i.clock.Now().Nanoseconds()
+	i.atime = now
+	i.ctime = now
+	i.mtime = now
+	// i.nlink initialized by caller
+	i.impl = impl
+}
+
+// incLinksLocked increments i's link count.
+//
+// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0.
+// i.nlink < maxLinks.
+func (i *inode) incLinksLocked() {
+	if i.nlink == 0 {
+		panic("tmpfs.inode.incLinksLocked() called with no existing links")
+	}
+	if i.nlink == maxLinks {
+		panic("memfs.inode.incLinksLocked() called with maximum link count")
+	}
+	atomic.AddUint32(&i.nlink, 1)
+}
+
+// decLinksLocked decrements i's link count.
+//
+// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0.
+func (i *inode) decLinksLocked() {
+	if i.nlink == 0 {
+		panic("tmpfs.inode.decLinksLocked() called with no existing links")
+	}
+	atomic.AddUint32(&i.nlink, ^uint32(0))
+}
+
+func (i *inode) incRef() {
+	if atomic.AddInt64(&i.refs, 1) <= 1 {
+		panic("tmpfs.inode.incRef() called without holding a reference")
+	}
+}
+
+func (i *inode) tryIncRef() bool {
+	for {
+		refs := atomic.LoadInt64(&i.refs)
+		if refs == 0 {
+			return false
+		}
+		if atomic.CompareAndSwapInt64(&i.refs, refs, refs+1) {
+			return true
+		}
+	}
+}
+
+func (i *inode) decRef() {
+	if refs := atomic.AddInt64(&i.refs, -1); refs == 0 {
+		if regFile, ok := i.impl.(*regularFile); ok {
+			// Hold inode.mu and regFile.dataMu while mutating
+			// size.
+			i.mu.Lock()
+			regFile.dataMu.Lock()
+			regFile.data.DropAll(regFile.memFile)
+			atomic.StoreUint64(&regFile.size, 0)
+			regFile.dataMu.Unlock()
+			i.mu.Unlock()
+		}
+	} else if refs < 0 {
+		panic("tmpfs.inode.decRef() called without holding a reference")
+	}
+}
+
+func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error {
+	mode := linux.FileMode(atomic.LoadUint32(&i.mode))
+	return vfs.GenericCheckPermissions(creds, ats, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid)))
+}
+
+// Go won't inline this function, and returning linux.Statx (which is quite
+// big) means spending a lot of time in runtime.duffcopy(), so instead it's an
+// output parameter.
+//
+// Note that Linux does not guarantee to return consistent data (in the case of
+// a concurrent modification), so we do not require holding inode.mu.
+func (i *inode) statTo(stat *linux.Statx) {
+	stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK |
+		linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_ATIME |
+		linux.STATX_BTIME | linux.STATX_CTIME | linux.STATX_MTIME
+	stat.Blksize = 1 // usermem.PageSize in tmpfs
+	stat.Nlink = atomic.LoadUint32(&i.nlink)
+	stat.UID = atomic.LoadUint32(&i.uid)
+	stat.GID = atomic.LoadUint32(&i.gid)
+	stat.Mode = uint16(atomic.LoadUint32(&i.mode))
+	stat.Ino = i.ino
+	// Linux's tmpfs has no concept of btime, so zero-value is returned.
+	stat.Atime = linux.NsecToStatxTimestamp(i.atime)
+	stat.Ctime = linux.NsecToStatxTimestamp(i.ctime)
+	stat.Mtime = linux.NsecToStatxTimestamp(i.mtime)
+	// TODO(gvisor.dev/issue/1197): Device number.
+	switch impl := i.impl.(type) {
+	case *regularFile:
+		stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS
+		stat.Size = uint64(atomic.LoadUint64(&impl.size))
+		// In tmpfs, this will be FileRangeSet.Span() / 512 (but also cached in
+		// a uint64 accessed using atomic memory operations to avoid taking
+		// locks).
+		stat.Blocks = allocatedBlocksForSize(stat.Size)
+	case *symlink:
+		stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS
+		stat.Size = uint64(len(impl.target))
+		stat.Blocks = allocatedBlocksForSize(stat.Size)
+	case *deviceFile:
+		stat.RdevMajor = impl.major
+		stat.RdevMinor = impl.minor
+	case *socketFile, *directory, *namedPipe:
+		// Nothing to do.
+	default:
+		panic(fmt.Sprintf("unknown inode type: %T", i.impl))
+	}
+}
+
+func (i *inode) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx) error {
+	if stat.Mask == 0 {
+		return nil
+	}
+	if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE) != 0 {
+		return syserror.EPERM
+	}
+	mode := linux.FileMode(atomic.LoadUint32(&i.mode))
+	if err := vfs.CheckSetStat(ctx, creds, stat, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))); err != nil {
+		return err
+	}
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	var (
+		needsMtimeBump bool
+		needsCtimeBump bool
+	)
+	mask := stat.Mask
+	if mask&linux.STATX_MODE != 0 {
+		ft := atomic.LoadUint32(&i.mode) & linux.S_IFMT
+		atomic.StoreUint32(&i.mode, ft|uint32(stat.Mode&^linux.S_IFMT))
+		needsCtimeBump = true
+	}
+	if mask&linux.STATX_UID != 0 {
+		atomic.StoreUint32(&i.uid, stat.UID)
+		needsCtimeBump = true
+	}
+	if mask&linux.STATX_GID != 0 {
+		atomic.StoreUint32(&i.gid, stat.GID)
+		needsCtimeBump = true
+	}
+	if mask&linux.STATX_SIZE != 0 {
+		switch impl := i.impl.(type) {
+		case *regularFile:
+			updated, err := impl.truncateLocked(stat.Size)
+			if err != nil {
+				return err
+			}
+			if updated {
+				needsMtimeBump = true
+				needsCtimeBump = true
+			}
+		case *directory:
+			return syserror.EISDIR
+		default:
+			return syserror.EINVAL
+		}
+	}
+	now := i.clock.Now().Nanoseconds()
+	if mask&linux.STATX_ATIME != 0 {
+		if stat.Atime.Nsec == linux.UTIME_NOW {
+			atomic.StoreInt64(&i.atime, now)
+		} else {
+			atomic.StoreInt64(&i.atime, stat.Atime.ToNsecCapped())
+		}
+		needsCtimeBump = true
+	}
+	if mask&linux.STATX_MTIME != 0 {
+		if stat.Mtime.Nsec == linux.UTIME_NOW {
+			atomic.StoreInt64(&i.mtime, now)
+		} else {
+			atomic.StoreInt64(&i.mtime, stat.Mtime.ToNsecCapped())
+		}
+		needsCtimeBump = true
+		// Ignore the mtime bump, since we just set it ourselves.
+		needsMtimeBump = false
+	}
+	if mask&linux.STATX_CTIME != 0 {
+		if stat.Ctime.Nsec == linux.UTIME_NOW {
+			atomic.StoreInt64(&i.ctime, now)
+		} else {
+			atomic.StoreInt64(&i.ctime, stat.Ctime.ToNsecCapped())
+		}
+		// Ignore the ctime bump, since we just set it ourselves.
+		needsCtimeBump = false
+	}
+	if needsMtimeBump {
+		atomic.StoreInt64(&i.mtime, now)
+	}
+	if needsCtimeBump {
+		atomic.StoreInt64(&i.ctime, now)
+	}
+
+	return nil
+}
+
+// TODO(gvisor.dev/issue/1480): support file locking for file types other than regular.
+func (i *inode) lockBSD(uid fslock.UniqueID, t fslock.LockType, block fslock.Blocker) error {
+	switch i.impl.(type) {
+	case *regularFile:
+		return i.locks.LockBSD(uid, t, block)
+	}
+	return syserror.EBADF
+}
+
+// TODO(gvisor.dev/issue/1480): support file locking for file types other than regular.
+func (i *inode) unlockBSD(uid fslock.UniqueID) error {
+	switch i.impl.(type) {
+	case *regularFile:
+		i.locks.UnlockBSD(uid)
+		return nil
+	}
+	return syserror.EBADF
+}
+
+// TODO(gvisor.dev/issue/1480): support file locking for file types other than regular.
+func (i *inode) lockPOSIX(uid fslock.UniqueID, t fslock.LockType, rng fslock.LockRange, block fslock.Blocker) error {
+	switch i.impl.(type) {
+	case *regularFile:
+		return i.locks.LockPOSIX(uid, t, rng, block)
+	}
+	return syserror.EBADF
+}
+
+// TODO(gvisor.dev/issue/1480): support file locking for file types other than regular.
+func (i *inode) unlockPOSIX(uid fslock.UniqueID, rng fslock.LockRange) error {
+	switch i.impl.(type) {
+	case *regularFile:
+		i.locks.UnlockPOSIX(uid, rng)
+		return nil
+	}
+	return syserror.EBADF
+}
+
+// allocatedBlocksForSize returns the number of 512B blocks needed to
+// accommodate the given size in bytes, as appropriate for struct
+// stat::st_blocks and struct statx::stx_blocks. (Note that this 512B block
+// size is independent of the "preferred block size for I/O", struct
+// stat::st_blksize and struct statx::stx_blksize.)
+func allocatedBlocksForSize(size uint64) uint64 {
+	return (size + 511) / 512
+}
+
+func (i *inode) direntType() uint8 {
+	switch impl := i.impl.(type) {
+	case *regularFile:
+		return linux.DT_REG
+	case *directory:
+		return linux.DT_DIR
+	case *symlink:
+		return linux.DT_LNK
+	case *socketFile:
+		return linux.DT_SOCK
+	case *deviceFile:
+		switch impl.kind {
+		case vfs.BlockDevice:
+			return linux.DT_BLK
+		case vfs.CharDevice:
+			return linux.DT_CHR
+		default:
+			panic(fmt.Sprintf("unknown vfs.DeviceKind: %v", impl.kind))
+		}
+	default:
+		panic(fmt.Sprintf("unknown inode type: %T", i.impl))
+	}
+}
+
+func (i *inode) isDir() bool {
+	return linux.FileMode(i.mode).FileType() == linux.S_IFDIR
+}
+
+func (i *inode) touchAtime(mnt *vfs.Mount) {
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return
+	}
+	now := i.clock.Now().Nanoseconds()
+	i.mu.Lock()
+	atomic.StoreInt64(&i.atime, now)
+	i.mu.Unlock()
+	mnt.EndWrite()
+}
+
+// Preconditions: The caller has called vfs.Mount.CheckBeginWrite().
+func (i *inode) touchCtime() {
+	now := i.clock.Now().Nanoseconds()
+	i.mu.Lock()
+	atomic.StoreInt64(&i.ctime, now)
+	i.mu.Unlock()
+}
+
+// Preconditions: The caller has called vfs.Mount.CheckBeginWrite().
+func (i *inode) touchCMtime() {
+	now := i.clock.Now().Nanoseconds()
+	i.mu.Lock()
+	atomic.StoreInt64(&i.mtime, now)
+	atomic.StoreInt64(&i.ctime, now)
+	i.mu.Unlock()
+}
+
+// Preconditions: The caller has called vfs.Mount.CheckBeginWrite() and holds
+// inode.mu.
+func (i *inode) touchCMtimeLocked() {
+	now := i.clock.Now().Nanoseconds()
+	atomic.StoreInt64(&i.mtime, now)
+	atomic.StoreInt64(&i.ctime, now)
+}
+
+func (i *inode) listxattr(size uint64) ([]string, error) {
+	return i.xattrs.Listxattr(size)
+}
+
+func (i *inode) getxattr(creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) {
+	if err := i.checkPermissions(creds, vfs.MayRead); err != nil {
+		return "", err
+	}
+	if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) {
+		return "", syserror.EOPNOTSUPP
+	}
+	if !i.userXattrSupported() {
+		return "", syserror.ENODATA
+	}
+	return i.xattrs.Getxattr(opts)
+}
+
+func (i *inode) setxattr(creds *auth.Credentials, opts *vfs.SetxattrOptions) error {
+	if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+		return err
+	}
+	if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) {
+		return syserror.EOPNOTSUPP
+	}
+	if !i.userXattrSupported() {
+		return syserror.EPERM
+	}
+	return i.xattrs.Setxattr(opts)
+}
+
+func (i *inode) removexattr(creds *auth.Credentials, name string) error {
+	if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+		return err
+	}
+	if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) {
+		return syserror.EOPNOTSUPP
+	}
+	if !i.userXattrSupported() {
+		return syserror.EPERM
+	}
+	return i.xattrs.Removexattr(name)
+}
+
+// Extended attributes in the user.* namespace are only supported for regular
+// files and directories.
+func (i *inode) userXattrSupported() bool {
+	filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode)
+	return filetype == linux.S_IFREG || filetype == linux.S_IFDIR
+}
+
+// fileDescription is embedded by tmpfs implementations of
+// vfs.FileDescriptionImpl.
+type fileDescription struct {
+	vfsfd vfs.FileDescription
+	vfs.FileDescriptionDefaultImpl
+}
+
+func (fd *fileDescription) filesystem() *filesystem {
+	return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem)
+}
+
+func (fd *fileDescription) inode() *inode {
+	return fd.vfsfd.Dentry().Impl().(*dentry).inode
+}
+
+// Stat implements vfs.FileDescriptionImpl.Stat.
+func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
+	var stat linux.Statx
+	fd.inode().statTo(&stat)
+	return stat, nil
+}
+
+// SetStat implements vfs.FileDescriptionImpl.SetStat.
+func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
+	creds := auth.CredentialsFromContext(ctx)
+	return fd.inode().setStat(ctx, creds, &opts.Stat)
+}
+
+// Listxattr implements vfs.FileDescriptionImpl.Listxattr.
+func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) {
+	return fd.inode().listxattr(size)
+}
+
+// Getxattr implements vfs.FileDescriptionImpl.Getxattr.
+func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) {
+	return fd.inode().getxattr(auth.CredentialsFromContext(ctx), &opts)
+}
+
+// Setxattr implements vfs.FileDescriptionImpl.Setxattr.
+func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error {
+	return fd.inode().setxattr(auth.CredentialsFromContext(ctx), &opts)
+}
+
+// Removexattr implements vfs.FileDescriptionImpl.Removexattr.
+func (fd *fileDescription) Removexattr(ctx context.Context, name string) error {
+	return fd.inode().removexattr(auth.CredentialsFromContext(ctx), name)
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs_state_autogen.go b/pkg/sentry/fsimpl/tmpfs/tmpfs_state_autogen.go
new file mode 100755
index 000000000..10d8b49a6
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs_state_autogen.go
@@ -0,0 +1,38 @@
+// automatically generated by stateify.
+
+package tmpfs
+
+import (
+	"gvisor.dev/gvisor/pkg/state"
+)
+
+func (x *dentryList) beforeSave() {}
+func (x *dentryList) save(m state.Map) {
+	x.beforeSave()
+	m.Save("head", &x.head)
+	m.Save("tail", &x.tail)
+}
+
+func (x *dentryList) afterLoad() {}
+func (x *dentryList) load(m state.Map) {
+	m.Load("head", &x.head)
+	m.Load("tail", &x.tail)
+}
+
+func (x *dentryEntry) beforeSave() {}
+func (x *dentryEntry) save(m state.Map) {
+	x.beforeSave()
+	m.Save("next", &x.next)
+	m.Save("prev", &x.prev)
+}
+
+func (x *dentryEntry) afterLoad() {}
+func (x *dentryEntry) load(m state.Map) {
+	m.Load("next", &x.next)
+	m.Load("prev", &x.prev)
+}
+
+func init() {
+	state.Register("pkg/sentry/fsimpl/tmpfs.dentryList", (*dentryList)(nil), state.Fns{Save: (*dentryList).save, Load: (*dentryList).load})
+	state.Register("pkg/sentry/fsimpl/tmpfs.dentryEntry", (*dentryEntry)(nil), state.Fns{Save: (*dentryEntry).save, Load: (*dentryEntry).load})
+}