Merge branch 'master' into tcp-matchers-submit

12 files changed, 4071 insertions, 0 deletions

diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD
new file mode 100644
index 000000000..4ba76a1e8
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/BUILD
@@ -0,0 +1,55 @@
+load("//tools:defs.bzl", "go_library")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+
+licenses(["notice"])
+
+go_template_instance(
+    name = "dentry_list",
+    out = "dentry_list.go",
+    package = "gofer",
+    prefix = "dentry",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Element": "*dentry",
+        "Linker": "*dentry",
+    },
+)
+
+go_library(
+    name = "gofer",
+    srcs = [
+        "dentry_list.go",
+        "directory.go",
+        "filesystem.go",
+        "gofer.go",
+        "handle.go",
+        "handle_unsafe.go",
+        "p9file.go",
+        "pagemath.go",
+        "regular_file.go",
+        "special_file.go",
+        "symlink.go",
+        "time.go",
+    ],
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/context",
+        "//pkg/fd",
+        "//pkg/fspath",
+        "//pkg/log",
+        "//pkg/p9",
+        "//pkg/safemem",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/pgalloc",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/vfs",
+        "//pkg/syserror",
+        "//pkg/unet",
+        "//pkg/usermem",
+    ],
+)
diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go
new file mode 100644
index 000000000..baa2cdd8e
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/directory.go
@@ -0,0 +1,190 @@
+// 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 gofer
+
+import (
+	"sync"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/p9"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+func (d *dentry) isDir() bool {
+	return d.fileType() == linux.S_IFDIR
+}
+
+// Preconditions: d.dirMu must be locked. d.isDir(). fs.opts.interop !=
+// InteropModeShared.
+func (d *dentry) cacheNegativeChildLocked(name string) {
+	if d.negativeChildren == nil {
+		d.negativeChildren = make(map[string]struct{})
+	}
+	d.negativeChildren[name] = struct{}{}
+}
+
+type directoryFD struct {
+	fileDescription
+	vfs.DirectoryFileDescriptionDefaultImpl
+
+	mu      sync.Mutex
+	off     int64
+	dirents []vfs.Dirent
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *directoryFD) Release() {
+}
+
+// IterDirents implements vfs.FileDescriptionImpl.IterDirents.
+func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error {
+	fd.mu.Lock()
+	defer fd.mu.Unlock()
+
+	if fd.dirents == nil {
+		ds, err := fd.dentry().getDirents(ctx)
+		if err != nil {
+			return err
+		}
+		fd.dirents = ds
+	}
+
+	for fd.off < int64(len(fd.dirents)) {
+		if !cb.Handle(fd.dirents[fd.off]) {
+			return nil
+		}
+		fd.off++
+	}
+	return nil
+}
+
+// Preconditions: d.isDir(). There exists at least one directoryFD representing d.
+func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) {
+	// 9P2000.L's readdir does not specify behavior in the presence of
+	// concurrent mutation of an iterated directory, so implementations may
+	// duplicate or omit entries in this case, which violates POSIX semantics.
+	// Thus we read all directory entries while holding d.dirMu to exclude
+	// directory mutations. (Note that it is impossible for the client to
+	// exclude concurrent mutation from other remote filesystem users. Since
+	// there is no way to detect if the server has incorrectly omitted
+	// directory entries, we simply assume that the server is well-behaved
+	// under InteropModeShared.) This is inconsistent with Linux (which appears
+	// to assume that directory fids have the correct semantics, and translates
+	// struct file_operations::readdir calls directly to readdir RPCs), but is
+	// consistent with VFS1.
+
+	d.fs.renameMu.RLock()
+	defer d.fs.renameMu.RUnlock()
+	d.dirMu.Lock()
+	defer d.dirMu.Unlock()
+	if d.dirents != nil {
+		return d.dirents, nil
+	}
+
+	// It's not clear if 9P2000.L's readdir is expected to return "." and "..",
+	// so we generate them here.
+	parent := d.vfsd.ParentOrSelf().Impl().(*dentry)
+	dirents := []vfs.Dirent{
+		{
+			Name:    ".",
+			Type:    linux.DT_DIR,
+			Ino:     d.ino,
+			NextOff: 1,
+		},
+		{
+			Name:    "..",
+			Type:    uint8(atomic.LoadUint32(&parent.mode) >> 12),
+			Ino:     parent.ino,
+			NextOff: 2,
+		},
+	}
+	off := uint64(0)
+	const count = 64 * 1024 // for consistency with the vfs1 client
+	d.handleMu.RLock()
+	defer d.handleMu.RUnlock()
+	if !d.handleReadable {
+		// This should not be possible because a readable handle should have
+		// been opened when the calling directoryFD was opened.
+		panic("gofer.dentry.getDirents called without a readable handle")
+	}
+	for {
+		p9ds, err := d.handle.file.readdir(ctx, off, count)
+		if err != nil {
+			return nil, err
+		}
+		if len(p9ds) == 0 {
+			// Cache dirents for future directoryFDs if permitted.
+			if d.fs.opts.interop != InteropModeShared {
+				d.dirents = dirents
+			}
+			return dirents, nil
+		}
+		for _, p9d := range p9ds {
+			if p9d.Name == "." || p9d.Name == ".." {
+				continue
+			}
+			dirent := vfs.Dirent{
+				Name:    p9d.Name,
+				Ino:     p9d.QID.Path,
+				NextOff: int64(len(dirents) + 1),
+			}
+			// p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or
+			// DMSOCKET.
+			switch p9d.Type {
+			case p9.TypeSymlink:
+				dirent.Type = linux.DT_LNK
+			case p9.TypeDir:
+				dirent.Type = linux.DT_DIR
+			default:
+				dirent.Type = linux.DT_REG
+			}
+			dirents = append(dirents, dirent)
+		}
+		off = p9ds[len(p9ds)-1].Offset
+	}
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+	fd.mu.Lock()
+	defer fd.mu.Unlock()
+
+	switch whence {
+	case linux.SEEK_SET:
+		if offset < 0 {
+			return 0, syserror.EINVAL
+		}
+		if offset == 0 {
+			// Ensure that the next call to fd.IterDirents() calls
+			// fd.dentry().getDirents().
+			fd.dirents = nil
+		}
+		fd.off = offset
+		return fd.off, nil
+	case linux.SEEK_CUR:
+		offset += fd.off
+		if offset < 0 {
+			return 0, syserror.EINVAL
+		}
+		// Don't clear fd.dirents in this case, even if offset == 0.
+		fd.off = offset
+		return fd.off, nil
+	default:
+		return 0, syserror.EINVAL
+	}
+}
diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go
new file mode 100644
index 000000000..8eb61debf
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/filesystem.go
@@ -0,0 +1,1087 @@
+// 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 gofer
+
+import (
+	"sync"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fspath"
+	"gvisor.dev/gvisor/pkg/p9"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// Sync implements vfs.FilesystemImpl.Sync.
+func (fs *filesystem) Sync(ctx context.Context) error {
+	// Snapshot current dentries and special files.
+	fs.syncMu.Lock()
+	ds := make([]*dentry, 0, len(fs.dentries))
+	for d := range fs.dentries {
+		ds = append(ds, d)
+	}
+	sffds := make([]*specialFileFD, 0, len(fs.specialFileFDs))
+	for sffd := range fs.specialFileFDs {
+		sffds = append(sffds, sffd)
+	}
+	fs.syncMu.Unlock()
+
+	// Return the first error we encounter, but sync everything we can
+	// regardless.
+	var retErr error
+
+	// Sync regular files.
+	for _, d := range ds {
+		if !d.TryIncRef() {
+			continue
+		}
+		err := d.syncSharedHandle(ctx)
+		d.DecRef()
+		if err != nil && retErr == nil {
+			retErr = err
+		}
+	}
+
+	// Sync special files, which may be writable but do not use dentry shared
+	// handles (so they won't be synced by the above).
+	for _, sffd := range sffds {
+		if !sffd.vfsfd.TryIncRef() {
+			continue
+		}
+		err := sffd.Sync(ctx)
+		sffd.vfsfd.DecRef()
+		if err != nil && retErr == nil {
+			retErr = err
+		}
+	}
+
+	return retErr
+}
+
+// maxFilenameLen is the maximum length of a filename. This is dictated by 9P's
+// encoding of strings, which uses 2 bytes for the length prefix.
+const maxFilenameLen = (1 << 16) - 1
+
+// dentrySlicePool is a pool of *[]*dentry used to store dentries for which
+// dentry.checkCachingLocked() must be called. The pool holds pointers to
+// slices because Go lacks generics, so sync.Pool operates on interface{}, so
+// every call to (what should be) sync.Pool<[]*dentry>.Put() allocates a copy
+// of the slice header on the heap.
+var dentrySlicePool = sync.Pool{
+	New: func() interface{} {
+		ds := make([]*dentry, 0, 4) // arbitrary non-zero initial capacity
+		return &ds
+	},
+}
+
+func appendDentry(ds *[]*dentry, d *dentry) *[]*dentry {
+	if ds == nil {
+		ds = dentrySlicePool.Get().(*[]*dentry)
+	}
+	*ds = append(*ds, d)
+	return ds
+}
+
+// Preconditions: ds != nil.
+func putDentrySlice(ds *[]*dentry) {
+	// Allow dentries to be GC'd.
+	for i := range *ds {
+		(*ds)[i] = nil
+	}
+	*ds = (*ds)[:0]
+	dentrySlicePool.Put(ds)
+}
+
+// stepLocked resolves rp.Component() to an existing file, starting from the
+// given directory.
+//
+// Dentries which may become cached as a result of the traversal are appended
+// to *ds.
+//
+// Preconditions: fs.renameMu must be locked. d.dirMu must be locked.
+// !rp.Done(). If fs.opts.interop == InteropModeShared, then d's cached
+// metadata must be up to date.
+func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) {
+	if !d.isDir() {
+		return nil, syserror.ENOTDIR
+	}
+	if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil {
+		return nil, err
+	}
+afterSymlink:
+	name := rp.Component()
+	if name == "." {
+		rp.Advance()
+		return d, nil
+	}
+	if name == ".." {
+		parentVFSD, err := rp.ResolveParent(&d.vfsd)
+		if err != nil {
+			return nil, err
+		}
+		parent := parentVFSD.Impl().(*dentry)
+		if fs.opts.interop == InteropModeShared {
+			// We must assume that parentVFSD is correct, because if d has been
+			// moved elsewhere in the remote filesystem so that its parent has
+			// changed, we have no way of determining its new parent's location
+			// in the filesystem. Get updated metadata for parentVFSD.
+			_, attrMask, attr, err := parent.file.getAttr(ctx, dentryAttrMask())
+			if err != nil {
+				return nil, err
+			}
+			parent.updateFromP9Attrs(attrMask, &attr)
+		}
+		rp.Advance()
+		return parent, nil
+	}
+	childVFSD, err := rp.ResolveChild(&d.vfsd, name)
+	if err != nil {
+		return nil, err
+	}
+	// FIXME(jamieliu): Linux performs revalidation before mount lookup
+	// (fs/namei.c:lookup_fast() => __d_lookup_rcu(), d_revalidate(),
+	// __follow_mount_rcu()).
+	child, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, childVFSD, ds)
+	if err != nil {
+		return nil, err
+	}
+	if child == nil {
+		return nil, syserror.ENOENT
+	}
+	if child.isSymlink() && rp.ShouldFollowSymlink() {
+		target, err := child.readlink(ctx, rp.Mount())
+		if err != nil {
+			return nil, err
+		}
+		if err := rp.HandleSymlink(target); err != nil {
+			return nil, err
+		}
+		goto afterSymlink // don't check the current directory again
+	}
+	rp.Advance()
+	return child, nil
+}
+
+// revalidateChildLocked must be called after a call to parent.vfsd.Child(name)
+// or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be
+// nil) to verify that the returned child (or lack thereof) is correct. If no file
+// exists at name, revalidateChildLocked returns (nil, nil).
+//
+// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked.
+// parent.isDir(). name is not "." or "..".
+//
+// Postconditions: If revalidateChildLocked returns a non-nil dentry, its
+// cached metadata is up to date.
+func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, childVFSD *vfs.Dentry, ds **[]*dentry) (*dentry, error) {
+	if childVFSD != nil && fs.opts.interop != InteropModeShared {
+		// We have a cached dentry that is assumed to be correct.
+		return childVFSD.Impl().(*dentry), nil
+	}
+	// We either don't have a cached dentry or need to verify that it's still
+	// correct, either of which requires a remote lookup. Check if this name is
+	// valid before performing the lookup.
+	if len(name) > maxFilenameLen {
+		return nil, syserror.ENAMETOOLONG
+	}
+	// Check if we've already cached this lookup with a negative result.
+	if _, ok := parent.negativeChildren[name]; ok {
+		return nil, nil
+	}
+	// Perform the remote lookup.
+	qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name)
+	if err != nil && err != syserror.ENOENT {
+		return nil, err
+	}
+	if childVFSD != nil {
+		child := childVFSD.Impl().(*dentry)
+		if !file.isNil() && qid.Path == child.ino {
+			// The file at this path hasn't changed. Just update cached
+			// metadata.
+			file.close(ctx)
+			child.updateFromP9Attrs(attrMask, &attr)
+			return child, nil
+		}
+		// The file at this path has changed or no longer exists. Remove
+		// the stale dentry from the tree, and re-evaluate its caching
+		// status (i.e. if it has 0 references, drop it).
+		vfsObj.ForceDeleteDentry(childVFSD)
+		*ds = appendDentry(*ds, child)
+		childVFSD = nil
+	}
+	if file.isNil() {
+		// No file exists at this path now. Cache the negative lookup if
+		// allowed.
+		if fs.opts.interop != InteropModeShared {
+			parent.cacheNegativeChildLocked(name)
+		}
+		return nil, nil
+	}
+	// Create a new dentry representing the file.
+	child, err := fs.newDentry(ctx, file, qid, attrMask, &attr)
+	if err != nil {
+		file.close(ctx)
+		return nil, err
+	}
+	parent.IncRef() // reference held by child on its parent
+	parent.vfsd.InsertChild(&child.vfsd, name)
+	// For now, child has 0 references, so our caller should call
+	// child.checkCachingLocked().
+	*ds = appendDentry(*ds, child)
+	return child, 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.
+//
+// Preconditions: fs.renameMu must be locked. !rp.Done(). If fs.opts.interop ==
+// InteropModeShared, then d's cached metadata must be up to date.
+func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) {
+	for !rp.Final() {
+		d.dirMu.Lock()
+		next, err := fs.stepLocked(ctx, rp, d, ds)
+		d.dirMu.Unlock()
+		if err != nil {
+			return nil, err
+		}
+		d = next
+	}
+	if !d.isDir() {
+		return nil, syserror.ENOTDIR
+	}
+	return d, nil
+}
+
+// resolveLocked resolves rp to an existing file.
+//
+// Preconditions: fs.renameMu must be locked.
+func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) {
+	d := rp.Start().Impl().(*dentry)
+	if fs.opts.interop == InteropModeShared {
+		// Get updated metadata for rp.Start() as required by fs.stepLocked().
+		if err := d.updateFromGetattr(ctx); err != nil {
+			return nil, err
+		}
+	}
+	for !rp.Done() {
+		d.dirMu.Lock()
+		next, err := fs.stepLocked(ctx, rp, d, ds)
+		d.dirMu.Unlock()
+		if err != nil {
+			return nil, err
+		}
+		d = next
+	}
+	if rp.MustBeDir() && !d.isDir() {
+		return nil, syserror.ENOTDIR
+	}
+	return d, nil
+}
+
+// doCreateAt checks that creating a file at rp is permitted, then invokes
+// create to do so.
+//
+// Preconditions: !rp.Done(). For the final path component in rp,
+// !rp.ShouldFollowSymlink().
+func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	start := rp.Start().Impl().(*dentry)
+	if fs.opts.interop == InteropModeShared {
+		// Get updated metadata for start as required by
+		// fs.walkParentDirLocked().
+		if err := start.updateFromGetattr(ctx); err != nil {
+			return err
+		}
+	}
+	parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds)
+	if err != nil {
+		return err
+	}
+	if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil {
+		return err
+	}
+	if parent.isDeleted() {
+		return syserror.ENOENT
+	}
+	name := rp.Component()
+	if name == "." || name == ".." {
+		return syserror.EEXIST
+	}
+	if len(name) > maxFilenameLen {
+		return syserror.ENAMETOOLONG
+	}
+	if !dir && rp.MustBeDir() {
+		return syserror.ENOENT
+	}
+	mnt := rp.Mount()
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer mnt.EndWrite()
+	parent.dirMu.Lock()
+	defer parent.dirMu.Unlock()
+	if fs.opts.interop == InteropModeShared {
+		// The existence of a dentry at name would be inconclusive because the
+		// file it represents may have been deleted from the remote filesystem,
+		// so we would need to make an RPC to revalidate the dentry. Just
+		// attempt the file creation RPC instead. If a file does exist, the RPC
+		// will fail with EEXIST like we would have. If the RPC succeeds, and a
+		// stale dentry exists, the dentry will fail revalidation next time
+		// it's used.
+		return create(parent, name)
+	}
+	if parent.vfsd.Child(name) != nil {
+		return syserror.EEXIST
+	}
+	// No cached dentry exists; however, there might still be an existing file
+	// at name. As above, we attempt the file creation RPC anyway.
+	if err := create(parent, name); err != nil {
+		return err
+	}
+	parent.touchCMtime(ctx)
+	delete(parent.negativeChildren, name)
+	parent.dirents = nil
+	return nil
+}
+
+// Preconditions: !rp.Done().
+func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool) error {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	start := rp.Start().Impl().(*dentry)
+	if fs.opts.interop == InteropModeShared {
+		// Get updated metadata for start as required by
+		// fs.walkParentDirLocked().
+		if err := start.updateFromGetattr(ctx); err != nil {
+			return err
+		}
+	}
+	parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds)
+	if err != nil {
+		return err
+	}
+	if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil {
+		return err
+	}
+	if err := rp.Mount().CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer rp.Mount().EndWrite()
+
+	name := rp.Component()
+	if dir {
+		if name == "." {
+			return syserror.EINVAL
+		}
+		if name == ".." {
+			return syserror.ENOTEMPTY
+		}
+	} else {
+		if name == "." || name == ".." {
+			return syserror.EISDIR
+		}
+	}
+	vfsObj := rp.VirtualFilesystem()
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	parent.dirMu.Lock()
+	defer parent.dirMu.Unlock()
+	childVFSD := parent.vfsd.Child(name)
+	var child *dentry
+	// We only need a dentry representing the file at name if it can be a mount
+	// point. If childVFSD is nil, then it can't be a mount point. If childVFSD
+	// is non-nil but stale, the actual file can't be a mount point either; we
+	// detect this case by just speculatively calling PrepareDeleteDentry and
+	// only revalidating the dentry if that fails (indicating that the existing
+	// dentry is a mount point).
+	if childVFSD != nil {
+		child = childVFSD.Impl().(*dentry)
+		if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil {
+			child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, childVFSD, &ds)
+			if err != nil {
+				return err
+			}
+			if child != nil {
+				childVFSD = &child.vfsd
+				if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil {
+					return err
+				}
+			} else {
+				childVFSD = nil
+			}
+		}
+	} else if _, ok := parent.negativeChildren[name]; ok {
+		return syserror.ENOENT
+	}
+	flags := uint32(0)
+	if dir {
+		if child != nil && !child.isDir() {
+			return syserror.ENOTDIR
+		}
+		flags = linux.AT_REMOVEDIR
+	} else {
+		if child != nil && child.isDir() {
+			return syserror.EISDIR
+		}
+		if rp.MustBeDir() {
+			return syserror.ENOTDIR
+		}
+	}
+	err = parent.file.unlinkAt(ctx, name, flags)
+	if err != nil {
+		if childVFSD != nil {
+			vfsObj.AbortDeleteDentry(childVFSD)
+		}
+		return err
+	}
+	if fs.opts.interop != InteropModeShared {
+		parent.touchCMtime(ctx)
+		parent.cacheNegativeChildLocked(name)
+		parent.dirents = nil
+	}
+	if child != nil {
+		child.setDeleted()
+		vfsObj.CommitDeleteDentry(childVFSD)
+		ds = appendDentry(ds, child)
+	}
+	return nil
+}
+
+// renameMuRUnlockAndCheckCaching calls fs.renameMu.RUnlock(), then calls
+// dentry.checkCachingLocked on all dentries in *ds with fs.renameMu locked for
+// writing.
+//
+// ds is a pointer-to-pointer since defer evaluates its arguments immediately,
+// but dentry slices are allocated lazily, and it's much easier to say "defer
+// fs.renameMuRUnlockAndCheckCaching(&ds)" than "defer func() {
+// fs.renameMuRUnlockAndCheckCaching(ds) }()" to work around this.
+func (fs *filesystem) renameMuRUnlockAndCheckCaching(ds **[]*dentry) {
+	fs.renameMu.RUnlock()
+	if *ds == nil {
+		return
+	}
+	if len(**ds) != 0 {
+		fs.renameMu.Lock()
+		for _, d := range **ds {
+			d.checkCachingLocked()
+		}
+		fs.renameMu.Unlock()
+	}
+	putDentrySlice(*ds)
+}
+
+func (fs *filesystem) renameMuUnlockAndCheckCaching(ds **[]*dentry) {
+	if *ds == nil {
+		fs.renameMu.Unlock()
+		return
+	}
+	for _, d := range **ds {
+		d.checkCachingLocked()
+	}
+	fs.renameMu.Unlock()
+	putDentrySlice(*ds)
+}
+
+// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt.
+func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return nil, err
+	}
+	if opts.CheckSearchable {
+		if !d.isDir() {
+			return nil, syserror.ENOTDIR
+		}
+		if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); 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) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	start := rp.Start().Impl().(*dentry)
+	if fs.opts.interop == InteropModeShared {
+		// Get updated metadata for start as required by
+		// fs.walkParentDirLocked().
+		if err := start.updateFromGetattr(ctx); err != nil {
+			return nil, err
+		}
+	}
+	d, err := fs.walkParentDirLocked(ctx, rp, start, &ds)
+	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(ctx, rp, false /* dir */, func(parent *dentry, childName string) error {
+		if rp.Mount() != vd.Mount() {
+			return syserror.EXDEV
+		}
+		// 9P2000.L supports hard links, but we don't.
+		return syserror.EPERM
+	})
+}
+
+// MkdirAt implements vfs.FilesystemImpl.MkdirAt.
+func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error {
+	return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error {
+		creds := rp.Credentials()
+		_, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID))
+		return err
+	})
+}
+
+// MknodAt implements vfs.FilesystemImpl.MknodAt.
+func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error {
+	return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error {
+		creds := rp.Credentials()
+		_, err := parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID))
+		return err
+	})
+}
+
+// OpenAt implements vfs.FilesystemImpl.OpenAt.
+func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) {
+	// Reject O_TMPFILE, which is not supported; supporting it correctly in the
+	// presence of other remote filesystem users requires remote filesystem
+	// support, and it isn't clear that there's any way to implement this in
+	// 9P.
+	if opts.Flags&linux.O_TMPFILE != 0 {
+		return nil, syserror.EOPNOTSUPP
+	}
+	mayCreate := opts.Flags&linux.O_CREAT != 0
+	mustCreate := opts.Flags&(linux.O_CREAT|linux.O_EXCL) == (linux.O_CREAT | linux.O_EXCL)
+
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+
+	start := rp.Start().Impl().(*dentry)
+	if fs.opts.interop == InteropModeShared {
+		// Get updated metadata for start as required by fs.stepLocked().
+		if err := start.updateFromGetattr(ctx); err != nil {
+			return nil, err
+		}
+	}
+	if rp.Done() {
+		return start.openLocked(ctx, rp, opts.Flags)
+	}
+
+afterTrailingSymlink:
+	parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds)
+	if err != nil {
+		return nil, err
+	}
+	// Check for search permission in the parent directory.
+	if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil {
+		return nil, err
+	}
+	// Determine whether or not we need to create a file.
+	parent.dirMu.Lock()
+	child, err := fs.stepLocked(ctx, rp, parent, &ds)
+	if err == syserror.ENOENT && mayCreate {
+		fd, err := parent.createAndOpenChildLocked(ctx, rp, &opts)
+		parent.dirMu.Unlock()
+		return fd, err
+	}
+	if err != nil {
+		parent.dirMu.Unlock()
+		return nil, err
+	}
+	// Open existing child or follow symlink.
+	parent.dirMu.Unlock()
+	if mustCreate {
+		return nil, syserror.EEXIST
+	}
+	if child.isSymlink() && rp.ShouldFollowSymlink() {
+		target, err := child.readlink(ctx, rp.Mount())
+		if err != nil {
+			return nil, err
+		}
+		if err := rp.HandleSymlink(target); err != nil {
+			return nil, err
+		}
+		start = parent
+		goto afterTrailingSymlink
+	}
+	return child.openLocked(ctx, rp, opts.Flags)
+}
+
+// Preconditions: fs.renameMu must be locked.
+func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags uint32) (*vfs.FileDescription, error) {
+	ats := vfs.AccessTypesForOpenFlags(flags)
+	if err := d.checkPermissions(rp.Credentials(), ats, d.isDir()); err != nil {
+		return nil, err
+	}
+	mnt := rp.Mount()
+	filetype := d.fileType()
+	switch {
+	case filetype == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD:
+		if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0); err != nil {
+			return nil, err
+		}
+		fd := &regularFileFD{}
+		if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{
+			AllowDirectIO: true,
+		}); err != nil {
+			return nil, err
+		}
+		return &fd.vfsfd, nil
+	case filetype == linux.S_IFDIR:
+		// Can't open directories with O_CREAT.
+		if flags&linux.O_CREAT != 0 {
+			return nil, syserror.EISDIR
+		}
+		// Can't open directories writably.
+		if ats&vfs.MayWrite != 0 {
+			return nil, syserror.EISDIR
+		}
+		if flags&linux.O_DIRECT != 0 {
+			return nil, syserror.EINVAL
+		}
+		if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, false /* write */, false /* trunc */); err != nil {
+			return nil, err
+		}
+		fd := &directoryFD{}
+		if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+			return nil, err
+		}
+		return &fd.vfsfd, nil
+	case filetype == linux.S_IFLNK:
+		// Can't open symlinks without O_PATH (which is unimplemented).
+		return nil, syserror.ELOOP
+	default:
+		if flags&linux.O_DIRECT != 0 {
+			return nil, syserror.EINVAL
+		}
+		h, err := openHandle(ctx, d.file, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0)
+		if err != nil {
+			return nil, err
+		}
+		fd := &specialFileFD{
+			handle: h,
+		}
+		if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+			h.close(ctx)
+			return nil, err
+		}
+		return &fd.vfsfd, nil
+	}
+}
+
+// Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked.
+func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) {
+	if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil {
+		return nil, err
+	}
+	if d.isDeleted() {
+		return nil, syserror.ENOENT
+	}
+	mnt := rp.Mount()
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return nil, err
+	}
+	defer mnt.EndWrite()
+
+	// 9P2000.L's lcreate takes a fid representing the parent directory, and
+	// converts it into an open fid representing the created file, so we need
+	// to duplicate the directory fid first.
+	_, dirfile, err := d.file.walk(ctx, nil)
+	if err != nil {
+		return nil, err
+	}
+	creds := rp.Credentials()
+	name := rp.Component()
+	fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, (p9.OpenFlags)(opts.Flags), (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID))
+	if err != nil {
+		dirfile.close(ctx)
+		return nil, err
+	}
+	// Then we need to walk to the file we just created to get a non-open fid
+	// representing it, and to get its metadata. This must use d.file since, as
+	// explained above, dirfile was invalidated by dirfile.Create().
+	walkQID, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name)
+	if err != nil {
+		openFile.close(ctx)
+		if fdobj != nil {
+			fdobj.Close()
+		}
+		return nil, err
+	}
+	// Sanity-check that we walked to the file we created.
+	if createQID.Path != walkQID.Path {
+		// Probably due to concurrent remote filesystem mutation?
+		ctx.Warningf("gofer.dentry.createAndOpenChildLocked: created file has QID %v before walk, QID %v after (interop=%v)", createQID, walkQID, d.fs.opts.interop)
+		nonOpenFile.close(ctx)
+		openFile.close(ctx)
+		if fdobj != nil {
+			fdobj.Close()
+		}
+		return nil, syserror.EAGAIN
+	}
+
+	// Construct the new dentry.
+	child, err := d.fs.newDentry(ctx, nonOpenFile, createQID, attrMask, &attr)
+	if err != nil {
+		nonOpenFile.close(ctx)
+		openFile.close(ctx)
+		if fdobj != nil {
+			fdobj.Close()
+		}
+		return nil, err
+	}
+	// Incorporate the fid that was opened by lcreate.
+	useRegularFileFD := child.fileType() == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD
+	if useRegularFileFD {
+		child.handleMu.Lock()
+		child.handle.file = openFile
+		if fdobj != nil {
+			child.handle.fd = int32(fdobj.Release())
+		}
+		child.handleReadable = vfs.MayReadFileWithOpenFlags(opts.Flags)
+		child.handleWritable = vfs.MayWriteFileWithOpenFlags(opts.Flags)
+		child.handleMu.Unlock()
+	}
+	// Take a reference on the new dentry to be held by the new file
+	// description. (This reference also means that the new dentry is not
+	// eligible for caching yet, so we don't need to append to a dentry slice.)
+	child.refs = 1
+	// Insert the dentry into the tree.
+	d.IncRef() // reference held by child on its parent d
+	d.vfsd.InsertChild(&child.vfsd, name)
+	if d.fs.opts.interop != InteropModeShared {
+		d.touchCMtime(ctx)
+		delete(d.negativeChildren, name)
+		d.dirents = nil
+	}
+
+	// Finally, construct a file description representing the created file.
+	var childVFSFD *vfs.FileDescription
+	mnt.IncRef()
+	if useRegularFileFD {
+		fd := &regularFileFD{}
+		if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{
+			AllowDirectIO: true,
+		}); err != nil {
+			return nil, err
+		}
+		childVFSFD = &fd.vfsfd
+	} else {
+		fd := &specialFileFD{
+			handle: handle{
+				file: openFile,
+				fd:   -1,
+			},
+		}
+		if fdobj != nil {
+			fd.handle.fd = int32(fdobj.Release())
+		}
+		if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+			fd.handle.close(ctx)
+			return nil, err
+		}
+		childVFSFD = &fd.vfsfd
+	}
+	return childVFSFD, nil
+}
+
+// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt.
+func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return "", err
+	}
+	if !d.isSymlink() {
+		return "", syserror.EINVAL
+	}
+	return d.readlink(ctx, rp.Mount())
+}
+
+// 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 {
+		// Requires 9P support.
+		return syserror.EINVAL
+	}
+
+	var ds *[]*dentry
+	fs.renameMu.Lock()
+	defer fs.renameMuUnlockAndCheckCaching(&ds)
+	newParent, err := fs.walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry), &ds)
+	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 fs.opts.interop == InteropModeShared {
+		if err := oldParent.updateFromGetattr(ctx); err != nil {
+			return err
+		}
+	}
+	if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil {
+		return err
+	}
+	vfsObj := rp.VirtualFilesystem()
+	// We need a dentry representing the renamed file since, if it's a
+	// directory, we need to check for write permission on it.
+	oldParent.dirMu.Lock()
+	defer oldParent.dirMu.Unlock()
+	renamed, err := fs.revalidateChildLocked(ctx, vfsObj, oldParent, oldName, oldParent.vfsd.Child(oldName), &ds)
+	if err != nil {
+		return err
+	}
+	if renamed == nil {
+		return syserror.ENOENT
+	}
+	if renamed.isDir() {
+		if renamed == newParent || renamed.vfsd.IsAncestorOf(&newParent.vfsd) {
+			return syserror.EINVAL
+		}
+		if oldParent != newParent {
+			if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil {
+				return err
+			}
+		}
+	} else {
+		if opts.MustBeDir || rp.MustBeDir() {
+			return syserror.ENOTDIR
+		}
+	}
+
+	if oldParent != newParent {
+		if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil {
+			return err
+		}
+		newParent.dirMu.Lock()
+		defer newParent.dirMu.Unlock()
+	}
+	if newParent.isDeleted() {
+		return syserror.ENOENT
+	}
+	replacedVFSD := newParent.vfsd.Child(newName)
+	var replaced *dentry
+	// This is similar to unlinkAt, except:
+	//
+	// - We revalidate the replaced dentry unconditionally for simplicity.
+	//
+	// - If rp.MustBeDir(), then we need a dentry representing the replaced
+	// file regardless to confirm that it's a directory.
+	if replacedVFSD != nil || rp.MustBeDir() {
+		replaced, err = fs.revalidateChildLocked(ctx, vfsObj, newParent, newName, replacedVFSD, &ds)
+		if err != nil {
+			return err
+		}
+		if replaced != nil {
+			if replaced.isDir() {
+				if !renamed.isDir() {
+					return syserror.EISDIR
+				}
+			} else {
+				if rp.MustBeDir() || renamed.isDir() {
+					return syserror.ENOTDIR
+				}
+			}
+			replacedVFSD = &replaced.vfsd
+		} else {
+			replacedVFSD = nil
+		}
+	}
+
+	if oldParent == newParent && oldName == newName {
+		return nil
+	}
+	if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil {
+		return err
+	}
+	if err := renamed.file.rename(ctx, newParent.file, newName); err != nil {
+		vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD)
+		return err
+	}
+	if fs.opts.interop != InteropModeShared {
+		oldParent.cacheNegativeChildLocked(oldName)
+		oldParent.dirents = nil
+		delete(newParent.negativeChildren, newName)
+		newParent.dirents = nil
+	}
+	vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD)
+	return nil
+}
+
+// RmdirAt implements vfs.FilesystemImpl.RmdirAt.
+func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error {
+	return fs.unlinkAt(ctx, rp, true /* dir */)
+}
+
+// SetStatAt implements vfs.FilesystemImpl.SetStatAt.
+func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return err
+	}
+	return d.setStat(ctx, rp.Credentials(), &opts.Stat, rp.Mount())
+}
+
+// StatAt implements vfs.FilesystemImpl.StatAt.
+func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return linux.Statx{}, err
+	}
+	// Since walking updates metadata for all traversed dentries under
+	// InteropModeShared, including the returned one, we can return cached
+	// metadata here regardless of fs.opts.interop.
+	var stat linux.Statx
+	d.statTo(&stat)
+	return stat, nil
+}
+
+// StatFSAt implements vfs.FilesystemImpl.StatFSAt.
+func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return linux.Statfs{}, err
+	}
+	fsstat, err := d.file.statFS(ctx)
+	if err != nil {
+		return linux.Statfs{}, err
+	}
+	nameLen := uint64(fsstat.NameLength)
+	if nameLen > maxFilenameLen {
+		nameLen = maxFilenameLen
+	}
+	return linux.Statfs{
+		// This is primarily for distinguishing a gofer file system in
+		// tests. Testing is important, so instead of defining
+		// something completely random, use a standard value.
+		Type:            linux.V9FS_MAGIC,
+		BlockSize:       int64(fsstat.BlockSize),
+		Blocks:          fsstat.Blocks,
+		BlocksFree:      fsstat.BlocksFree,
+		BlocksAvailable: fsstat.BlocksAvailable,
+		Files:           fsstat.Files,
+		FilesFree:       fsstat.FilesFree,
+		NameLength:      nameLen,
+	}, nil
+}
+
+// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt.
+func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error {
+	return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error {
+		creds := rp.Credentials()
+		_, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID))
+		return err
+	})
+}
+
+// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt.
+func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error {
+	return fs.unlinkAt(ctx, rp, false /* dir */)
+}
+
+// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt.
+func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return nil, err
+	}
+	return d.listxattr(ctx)
+}
+
+// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt.
+func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return "", err
+	}
+	return d.getxattr(ctx, name)
+}
+
+// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt.
+func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return err
+	}
+	return d.setxattr(ctx, &opts)
+}
+
+// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt.
+func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error {
+	var ds *[]*dentry
+	fs.renameMu.RLock()
+	defer fs.renameMuRUnlockAndCheckCaching(&ds)
+	d, err := fs.resolveLocked(ctx, rp, &ds)
+	if err != nil {
+		return err
+	}
+	return d.removexattr(ctx, name)
+}
+
+// PrependPath implements vfs.FilesystemImpl.PrependPath.
+func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error {
+	fs.renameMu.RLock()
+	defer fs.renameMu.RUnlock()
+	return vfs.GenericPrependPath(vfsroot, vd, b)
+}
diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go
new file mode 100644
index 000000000..d0552bd99
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/gofer.go
@@ -0,0 +1,1147 @@
+// 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 gofer provides a filesystem implementation that is backed by a 9p
+// server, interchangably referred to as "gofers" throughout this package.
+//
+// Lock order:
+//   regularFileFD/directoryFD.mu
+//     filesystem.renameMu
+//       dentry.dirMu
+//         filesystem.syncMu
+//         dentry.metadataMu
+//           *** "memmap.Mappable locks" below this point
+//           dentry.mapsMu
+//             *** "memmap.Mappable locks taken by Translate" below this point
+//             dentry.handleMu
+//               dentry.dataMu
+//
+// Locking dentry.dirMu in multiple dentries requires holding
+// filesystem.renameMu for writing.
+package gofer
+
+import (
+	"fmt"
+	"strconv"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/p9"
+	"gvisor.dev/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
+	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/unet"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// FilesystemType implements vfs.FilesystemType.
+type FilesystemType struct{}
+
+// filesystem implements vfs.FilesystemImpl.
+type filesystem struct {
+	vfsfs vfs.Filesystem
+
+	// mfp is used to allocate memory that caches regular file contents. mfp is
+	// immutable.
+	mfp pgalloc.MemoryFileProvider
+
+	// Immutable options.
+	opts filesystemOptions
+
+	// client is the client used by this filesystem. client is immutable.
+	client *p9.Client
+
+	// uid and gid are the effective KUID and KGID of the filesystem's creator,
+	// and are used as the owner and group for files that don't specify one.
+	// uid and gid are immutable.
+	uid auth.KUID
+	gid auth.KGID
+
+	// renameMu serves two purposes:
+	//
+	// - It synchronizes path resolution with renaming initiated by this
+	// client.
+	//
+	// - It is held by path resolution to ensure that reachable dentries remain
+	// valid. A dentry is reachable by path resolution if it has a non-zero
+	// reference count (such that it is usable as vfs.ResolvingPath.Start() or
+	// is reachable from its children), or if it is a child dentry (such that
+	// it is reachable from its parent).
+	renameMu sync.RWMutex
+
+	// cachedDentries contains all dentries with 0 references. (Due to race
+	// conditions, it may also contain dentries with non-zero references.)
+	// cachedDentriesLen is the number of dentries in cachedDentries. These
+	// fields are protected by renameMu.
+	cachedDentries    dentryList
+	cachedDentriesLen uint64
+
+	// dentries contains all dentries in this filesystem. specialFileFDs
+	// contains all open specialFileFDs. These fields are protected by syncMu.
+	syncMu         sync.Mutex
+	dentries       map[*dentry]struct{}
+	specialFileFDs map[*specialFileFD]struct{}
+}
+
+type filesystemOptions struct {
+	// "Standard" 9P options.
+	fd      int
+	aname   string
+	interop InteropMode // derived from the "cache" mount option
+	msize   uint32
+	version string
+
+	// maxCachedDentries is the maximum number of dentries with 0 references
+	// retained by the client.
+	maxCachedDentries uint64
+
+	// If forcePageCache is true, host FDs may not be used for application
+	// memory mappings even if available; instead, the client must perform its
+	// own caching of regular file pages. This is primarily useful for testing.
+	forcePageCache bool
+
+	// If limitHostFDTranslation is true, apply maxFillRange() constraints to
+	// host FD mappings returned by dentry.(memmap.Mappable).Translate(). This
+	// makes memory accounting behavior more consistent between cases where
+	// host FDs are / are not available, but may increase the frequency of
+	// sentry-handled page faults on files for which a host FD is available.
+	limitHostFDTranslation bool
+
+	// If overlayfsStaleRead is true, O_RDONLY host FDs provided by the remote
+	// filesystem may not be coherent with writable host FDs opened later, so
+	// mappings of the former must be replaced by mappings of the latter. This
+	// is usually only the case when the remote filesystem is an overlayfs
+	// mount on Linux < 4.19.
+	overlayfsStaleRead bool
+
+	// If regularFilesUseSpecialFileFD is true, application FDs representing
+	// regular files will use distinct file handles for each FD, in the same
+	// way that application FDs representing "special files" such as sockets
+	// do. Note that this disables client caching and mmap for regular files.
+	regularFilesUseSpecialFileFD bool
+}
+
+// InteropMode controls the client's interaction with other remote filesystem
+// users.
+type InteropMode uint32
+
+const (
+	// InteropModeExclusive is appropriate when the filesystem client is the
+	// only user of the remote filesystem.
+	//
+	// - The client may cache arbitrary filesystem state (file data, metadata,
+	// filesystem structure, etc.).
+	//
+	// - Client changes to filesystem state may be sent to the remote
+	// filesystem asynchronously, except when server permission checks are
+	// necessary.
+	//
+	// - File timestamps are based on client clocks. This ensures that users of
+	// the client observe timestamps that are coherent with their own clocks
+	// and consistent with Linux's semantics. However, since it is not always
+	// possible for clients to set arbitrary atimes and mtimes, and never
+	// possible for clients to set arbitrary ctimes, file timestamp changes are
+	// stored in the client only and never sent to the remote filesystem.
+	InteropModeExclusive InteropMode = iota
+
+	// InteropModeWritethrough is appropriate when there are read-only users of
+	// the remote filesystem that expect to observe changes made by the
+	// filesystem client.
+	//
+	// - The client may cache arbitrary filesystem state.
+	//
+	// - Client changes to filesystem state must be sent to the remote
+	// filesystem synchronously.
+	//
+	// - File timestamps are based on client clocks. As a corollary, access
+	// timestamp changes from other remote filesystem users will not be visible
+	// to the client.
+	InteropModeWritethrough
+
+	// InteropModeShared is appropriate when there are users of the remote
+	// filesystem that may mutate its state other than the client.
+	//
+	// - The client must verify cached filesystem state before using it.
+	//
+	// - Client changes to filesystem state must be sent to the remote
+	// filesystem synchronously.
+	//
+	// - File timestamps are based on server clocks. This is necessary to
+	// ensure that timestamp changes are synchronized between remote filesystem
+	// users.
+	//
+	// Note that the correctness of InteropModeShared depends on the server
+	// correctly implementing 9P fids (i.e. each fid immutably represents a
+	// single filesystem object), even in the presence of remote filesystem
+	// mutations from other users. If this is violated, the behavior of the
+	// client is undefined.
+	InteropModeShared
+)
+
+// 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) {
+	mfp := pgalloc.MemoryFileProviderFromContext(ctx)
+	if mfp == nil {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: context does not provide a pgalloc.MemoryFileProvider")
+		return nil, nil, syserror.EINVAL
+	}
+
+	mopts := vfs.GenericParseMountOptions(opts.Data)
+	var fsopts filesystemOptions
+
+	// Check that the transport is "fd".
+	trans, ok := mopts["trans"]
+	if !ok {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: transport must be specified as 'trans=fd'")
+		return nil, nil, syserror.EINVAL
+	}
+	delete(mopts, "trans")
+	if trans != "fd" {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: unsupported transport: trans=%s", trans)
+		return nil, nil, syserror.EINVAL
+	}
+
+	// Check that read and write FDs are provided and identical.
+	rfdstr, ok := mopts["rfdno"]
+	if !ok {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD must be specified as 'rfdno=<file descriptor>")
+		return nil, nil, syserror.EINVAL
+	}
+	delete(mopts, "rfdno")
+	rfd, err := strconv.Atoi(rfdstr)
+	if err != nil {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid read FD: rfdno=%s", rfdstr)
+		return nil, nil, syserror.EINVAL
+	}
+	wfdstr, ok := mopts["wfdno"]
+	if !ok {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: write FD must be specified as 'wfdno=<file descriptor>")
+		return nil, nil, syserror.EINVAL
+	}
+	delete(mopts, "wfdno")
+	wfd, err := strconv.Atoi(wfdstr)
+	if err != nil {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid write FD: wfdno=%s", wfdstr)
+		return nil, nil, syserror.EINVAL
+	}
+	if rfd != wfd {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD (%d) and write FD (%d) must be equal", rfd, wfd)
+		return nil, nil, syserror.EINVAL
+	}
+	fsopts.fd = rfd
+
+	// Get the attach name.
+	fsopts.aname = "/"
+	if aname, ok := mopts["aname"]; ok {
+		delete(mopts, "aname")
+		fsopts.aname = aname
+	}
+
+	// Parse the cache policy. For historical reasons, this defaults to the
+	// least generally-applicable option, InteropModeExclusive.
+	fsopts.interop = InteropModeExclusive
+	if cache, ok := mopts["cache"]; ok {
+		delete(mopts, "cache")
+		switch cache {
+		case "fscache":
+			fsopts.interop = InteropModeExclusive
+		case "fscache_writethrough":
+			fsopts.interop = InteropModeWritethrough
+		case "none":
+			fsopts.regularFilesUseSpecialFileFD = true
+			fallthrough
+		case "remote_revalidating":
+			fsopts.interop = InteropModeShared
+		default:
+			ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid cache policy: cache=%s", cache)
+			return nil, nil, syserror.EINVAL
+		}
+	}
+
+	// Parse the 9P message size.
+	fsopts.msize = 1024 * 1024 // 1M, tested to give good enough performance up to 64M
+	if msizestr, ok := mopts["msize"]; ok {
+		delete(mopts, "msize")
+		msize, err := strconv.ParseUint(msizestr, 10, 32)
+		if err != nil {
+			ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid message size: msize=%s", msizestr)
+			return nil, nil, syserror.EINVAL
+		}
+		fsopts.msize = uint32(msize)
+	}
+
+	// Parse the 9P protocol version.
+	fsopts.version = p9.HighestVersionString()
+	if version, ok := mopts["version"]; ok {
+		delete(mopts, "version")
+		fsopts.version = version
+	}
+
+	// Parse the dentry cache limit.
+	fsopts.maxCachedDentries = 1000
+	if str, ok := mopts["dentry_cache_limit"]; ok {
+		delete(mopts, "dentry_cache_limit")
+		maxCachedDentries, err := strconv.ParseUint(str, 10, 64)
+		if err != nil {
+			ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid dentry cache limit: dentry_cache_limit=%s", str)
+			return nil, nil, syserror.EINVAL
+		}
+		fsopts.maxCachedDentries = maxCachedDentries
+	}
+
+	// Handle simple flags.
+	if _, ok := mopts["force_page_cache"]; ok {
+		delete(mopts, "force_page_cache")
+		fsopts.forcePageCache = true
+	}
+	if _, ok := mopts["limit_host_fd_translation"]; ok {
+		delete(mopts, "limit_host_fd_translation")
+		fsopts.limitHostFDTranslation = true
+	}
+	if _, ok := mopts["overlayfs_stale_read"]; ok {
+		delete(mopts, "overlayfs_stale_read")
+		fsopts.overlayfsStaleRead = true
+	}
+	// fsopts.regularFilesUseSpecialFileFD can only be enabled by specifying
+	// "cache=none".
+
+	// Check for unparsed options.
+	if len(mopts) != 0 {
+		ctx.Warningf("gofer.FilesystemType.GetFilesystem: unknown options: %v", mopts)
+		return nil, nil, syserror.EINVAL
+	}
+
+	// Establish a connection with the server.
+	conn, err := unet.NewSocket(fsopts.fd)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	// Perform version negotiation with the server.
+	ctx.UninterruptibleSleepStart(false)
+	client, err := p9.NewClient(conn, fsopts.msize, fsopts.version)
+	ctx.UninterruptibleSleepFinish(false)
+	if err != nil {
+		conn.Close()
+		return nil, nil, err
+	}
+	// Ownership of conn has been transferred to client.
+
+	// Perform attach to obtain the filesystem root.
+	ctx.UninterruptibleSleepStart(false)
+	attached, err := client.Attach(fsopts.aname)
+	ctx.UninterruptibleSleepFinish(false)
+	if err != nil {
+		client.Close()
+		return nil, nil, err
+	}
+	attachFile := p9file{attached}
+	qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask())
+	if err != nil {
+		attachFile.close(ctx)
+		client.Close()
+		return nil, nil, err
+	}
+
+	// Construct the filesystem object.
+	fs := &filesystem{
+		mfp:            mfp,
+		opts:           fsopts,
+		uid:            creds.EffectiveKUID,
+		gid:            creds.EffectiveKGID,
+		client:         client,
+		dentries:       make(map[*dentry]struct{}),
+		specialFileFDs: make(map[*specialFileFD]struct{}),
+	}
+	fs.vfsfs.Init(vfsObj, fs)
+
+	// Construct the root dentry.
+	root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr)
+	if err != nil {
+		attachFile.close(ctx)
+		fs.vfsfs.DecRef()
+		return nil, nil, err
+	}
+	// Set the root's reference count to 2. One reference is returned to the
+	// caller, and the other is deliberately leaked to prevent the root from
+	// being "cached" and subsequently evicted. Its resources will still be
+	// cleaned up by fs.Release().
+	root.refs = 2
+
+	return &fs.vfsfs, &root.vfsd, nil
+}
+
+// Release implements vfs.FilesystemImpl.Release.
+func (fs *filesystem) Release() {
+	ctx := context.Background()
+	mf := fs.mfp.MemoryFile()
+
+	fs.syncMu.Lock()
+	for d := range fs.dentries {
+		d.handleMu.Lock()
+		d.dataMu.Lock()
+		if d.handleWritable {
+			// Write dirty cached data to the remote file.
+			if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt); err != nil {
+				log.Warningf("gofer.filesystem.Release: failed to flush dentry: %v", err)
+			}
+			// TODO(jamieliu): Do we need to flushf/fsync d?
+		}
+		// Discard cached pages.
+		d.cache.DropAll(mf)
+		d.dirty.RemoveAll()
+		d.dataMu.Unlock()
+		// Close the host fd if one exists.
+		if d.handle.fd >= 0 {
+			syscall.Close(int(d.handle.fd))
+			d.handle.fd = -1
+		}
+		d.handleMu.Unlock()
+	}
+	// There can't be any specialFileFDs still using fs, since each such
+	// FileDescription would hold a reference on a Mount holding a reference on
+	// fs.
+	fs.syncMu.Unlock()
+
+	// Close the connection to the server. This implicitly clunks all fids.
+	fs.client.Close()
+}
+
+// dentry implements vfs.DentryImpl.
+type dentry struct {
+	vfsd vfs.Dentry
+
+	// refs is the reference count. Each dentry holds a reference on its
+	// parent, even if disowned. refs is accessed using atomic memory
+	// operations.
+	refs int64
+
+	// fs is the owning filesystem. fs is immutable.
+	fs *filesystem
+
+	// We don't support hard links, so each dentry maps 1:1 to an inode.
+
+	// file is the unopened p9.File that backs this dentry. file is immutable.
+	file p9file
+
+	// If deleted is non-zero, the file represented by this dentry has been
+	// deleted. deleted is accessed using atomic memory operations.
+	deleted uint32
+
+	// If cached is true, dentryEntry links dentry into
+	// filesystem.cachedDentries. cached and dentryEntry are protected by
+	// filesystem.renameMu.
+	cached bool
+	dentryEntry
+
+	dirMu sync.Mutex
+
+	// If this dentry represents a directory, and InteropModeShared is not in
+	// effect, negativeChildren is a set of child names in this directory that
+	// are known not to exist. negativeChildren is protected by dirMu.
+	negativeChildren map[string]struct{}
+
+	// If this dentry represents a directory, InteropModeShared is not in
+	// effect, and dirents is not nil, it is a cache of all entries in the
+	// directory, in the order they were returned by the server. dirents is
+	// protected by dirMu.
+	dirents []vfs.Dirent
+
+	// Cached metadata; protected by metadataMu and accessed using atomic
+	// memory operations unless otherwise specified.
+	metadataMu sync.Mutex
+	ino        uint64 // immutable
+	mode       uint32 // type is immutable, perms are mutable
+	uid        uint32 // auth.KUID, but stored as raw uint32 for sync/atomic
+	gid        uint32 // auth.KGID, but ...
+	blockSize  uint32 // 0 if unknown
+	// Timestamps, all nsecs from the Unix epoch.
+	atime int64
+	mtime int64
+	ctime int64
+	btime int64
+	// File size, protected by both metadataMu and dataMu (i.e. both must be
+	// locked to mutate it).
+	size uint64
+
+	mapsMu sync.Mutex
+
+	// If this dentry represents a regular file, mappings tracks mappings of
+	// the file into memmap.MappingSpaces. mappings is protected by mapsMu.
+	mappings memmap.MappingSet
+
+	// If this dentry represents a regular file or directory:
+	//
+	// - handle is the I/O handle used by all regularFileFDs/directoryFDs
+	// representing this dentry.
+	//
+	// - handleReadable is true if handle is readable.
+	//
+	// - handleWritable is true if handle is writable.
+	//
+	// Invariants:
+	//
+	// - If handleReadable == handleWritable == false, then handle.file == nil
+	// (i.e. there is no open handle). Conversely, if handleReadable ||
+	// handleWritable == true, then handle.file != nil (i.e. there is an open
+	// handle).
+	//
+	// - handleReadable and handleWritable cannot transition from true to false
+	// (i.e. handles may not be downgraded).
+	//
+	// These fields are protected by handleMu.
+	handleMu       sync.RWMutex
+	handle         handle
+	handleReadable bool
+	handleWritable bool
+
+	dataMu sync.RWMutex
+
+	// If this dentry represents a regular file that is client-cached, cache
+	// maps offsets into the cached file to offsets into
+	// filesystem.mfp.MemoryFile() that store the file's data. cache is
+	// protected by dataMu.
+	cache fsutil.FileRangeSet
+
+	// If this dentry represents a regular file that is client-cached, dirty
+	// tracks dirty segments in cache. dirty is protected by dataMu.
+	dirty fsutil.DirtySet
+
+	// pf implements platform.File for mappings of handle.fd.
+	pf dentryPlatformFile
+
+	// If this dentry represents a symbolic link, InteropModeShared is not in
+	// effect, and haveTarget is true, target is the symlink target. haveTarget
+	// and target are protected by dataMu.
+	haveTarget bool
+	target     string
+}
+
+// dentryAttrMask returns a p9.AttrMask enabling all attributes used by the
+// gofer client.
+func dentryAttrMask() p9.AttrMask {
+	return p9.AttrMask{
+		Mode:  true,
+		UID:   true,
+		GID:   true,
+		ATime: true,
+		MTime: true,
+		CTime: true,
+		Size:  true,
+		BTime: true,
+	}
+}
+
+// newDentry creates a new dentry representing the given file. The dentry
+// initially has no references, but is not cached; it is the caller's
+// responsibility to set the dentry's reference count and/or call
+// dentry.checkCachingLocked() as appropriate.
+func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, mask p9.AttrMask, attr *p9.Attr) (*dentry, error) {
+	if !mask.Mode {
+		ctx.Warningf("can't create gofer.dentry without file type")
+		return nil, syserror.EIO
+	}
+	if attr.Mode.FileType() == p9.ModeRegular && !mask.Size {
+		ctx.Warningf("can't create regular file gofer.dentry without file size")
+		return nil, syserror.EIO
+	}
+
+	d := &dentry{
+		fs:        fs,
+		file:      file,
+		ino:       qid.Path,
+		mode:      uint32(attr.Mode),
+		uid:       uint32(fs.uid),
+		gid:       uint32(fs.gid),
+		blockSize: usermem.PageSize,
+		handle: handle{
+			fd: -1,
+		},
+	}
+	d.pf.dentry = d
+	if mask.UID {
+		d.uid = uint32(attr.UID)
+	}
+	if mask.GID {
+		d.gid = uint32(attr.GID)
+	}
+	if mask.Size {
+		d.size = attr.Size
+	}
+	if attr.BlockSize != 0 {
+		d.blockSize = uint32(attr.BlockSize)
+	}
+	if mask.ATime {
+		d.atime = dentryTimestampFromP9(attr.ATimeSeconds, attr.ATimeNanoSeconds)
+	}
+	if mask.MTime {
+		d.mtime = dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds)
+	}
+	if mask.CTime {
+		d.ctime = dentryTimestampFromP9(attr.CTimeSeconds, attr.CTimeNanoSeconds)
+	}
+	if mask.BTime {
+		d.btime = dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds)
+	}
+	d.vfsd.Init(d)
+
+	fs.syncMu.Lock()
+	fs.dentries[d] = struct{}{}
+	fs.syncMu.Unlock()
+	return d, nil
+}
+
+// updateFromP9Attrs is called to update d's metadata after an update from the
+// remote filesystem.
+func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) {
+	d.metadataMu.Lock()
+	if mask.Mode {
+		if got, want := uint32(attr.Mode.FileType()), d.fileType(); got != want {
+			d.metadataMu.Unlock()
+			panic(fmt.Sprintf("gofer.dentry file type changed from %#o to %#o", want, got))
+		}
+		atomic.StoreUint32(&d.mode, uint32(attr.Mode))
+	}
+	if mask.UID {
+		atomic.StoreUint32(&d.uid, uint32(attr.UID))
+	}
+	if mask.GID {
+		atomic.StoreUint32(&d.gid, uint32(attr.GID))
+	}
+	// There is no P9_GETATTR_* bit for I/O block size.
+	if attr.BlockSize != 0 {
+		atomic.StoreUint32(&d.blockSize, uint32(attr.BlockSize))
+	}
+	if mask.ATime {
+		atomic.StoreInt64(&d.atime, dentryTimestampFromP9(attr.ATimeSeconds, attr.ATimeNanoSeconds))
+	}
+	if mask.MTime {
+		atomic.StoreInt64(&d.mtime, dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds))
+	}
+	if mask.CTime {
+		atomic.StoreInt64(&d.ctime, dentryTimestampFromP9(attr.CTimeSeconds, attr.CTimeNanoSeconds))
+	}
+	if mask.BTime {
+		atomic.StoreInt64(&d.btime, dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds))
+	}
+	if mask.Size {
+		d.dataMu.Lock()
+		atomic.StoreUint64(&d.size, attr.Size)
+		d.dataMu.Unlock()
+	}
+	d.metadataMu.Unlock()
+}
+
+func (d *dentry) updateFromGetattr(ctx context.Context) error {
+	// Use d.handle.file, which represents a 9P fid that has been opened, in
+	// preference to d.file, which represents a 9P fid that has not. This may
+	// be significantly more efficient in some implementations.
+	var (
+		file            p9file
+		handleMuRLocked bool
+	)
+	d.handleMu.RLock()
+	if !d.handle.file.isNil() {
+		file = d.handle.file
+		handleMuRLocked = true
+	} else {
+		file = d.file
+		d.handleMu.RUnlock()
+	}
+	_, attrMask, attr, err := file.getAttr(ctx, dentryAttrMask())
+	if handleMuRLocked {
+		d.handleMu.RUnlock()
+	}
+	if err != nil {
+		return err
+	}
+	d.updateFromP9Attrs(attrMask, &attr)
+	return nil
+}
+
+func (d *dentry) fileType() uint32 {
+	return atomic.LoadUint32(&d.mode) & linux.S_IFMT
+}
+
+func (d *dentry) statTo(stat *linux.Statx) {
+	stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME
+	stat.Blksize = atomic.LoadUint32(&d.blockSize)
+	stat.Nlink = 1
+	if d.isDir() {
+		stat.Nlink = 2
+	}
+	stat.UID = atomic.LoadUint32(&d.uid)
+	stat.GID = atomic.LoadUint32(&d.gid)
+	stat.Mode = uint16(atomic.LoadUint32(&d.mode))
+	stat.Ino = d.ino
+	stat.Size = atomic.LoadUint64(&d.size)
+	// This is consistent with regularFileFD.Seek(), which treats regular files
+	// as having no holes.
+	stat.Blocks = (stat.Size + 511) / 512
+	stat.Atime = statxTimestampFromDentry(atomic.LoadInt64(&d.atime))
+	stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime))
+	stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime))
+	stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime))
+	// TODO(jamieliu): device number
+}
+
+func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mnt *vfs.Mount) 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_SIZE) != 0 {
+		return syserror.EPERM
+	}
+	if err := vfs.CheckSetStat(creds, stat, uint16(atomic.LoadUint32(&d.mode))&^linux.S_IFMT, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil {
+		return err
+	}
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return err
+	}
+	defer mnt.EndWrite()
+	setLocalAtime := false
+	setLocalMtime := false
+	if d.fs.opts.interop != InteropModeShared {
+		// Timestamp updates will be handled locally.
+		setLocalAtime = stat.Mask&linux.STATX_ATIME != 0
+		setLocalMtime = stat.Mask&linux.STATX_MTIME != 0
+		stat.Mask &^= linux.STATX_ATIME | linux.STATX_MTIME
+		if !setLocalMtime && (stat.Mask&linux.STATX_SIZE != 0) {
+			// Truncate updates mtime.
+			setLocalMtime = true
+			stat.Mtime.Nsec = linux.UTIME_NOW
+		}
+	}
+	d.metadataMu.Lock()
+	defer d.metadataMu.Unlock()
+	if stat.Mask != 0 {
+		if err := d.file.setAttr(ctx, p9.SetAttrMask{
+			Permissions:        stat.Mask&linux.STATX_MODE != 0,
+			UID:                stat.Mask&linux.STATX_UID != 0,
+			GID:                stat.Mask&linux.STATX_GID != 0,
+			Size:               stat.Mask&linux.STATX_SIZE != 0,
+			ATime:              stat.Mask&linux.STATX_ATIME != 0,
+			MTime:              stat.Mask&linux.STATX_MTIME != 0,
+			ATimeNotSystemTime: stat.Atime.Nsec != linux.UTIME_NOW,
+			MTimeNotSystemTime: stat.Mtime.Nsec != linux.UTIME_NOW,
+		}, p9.SetAttr{
+			Permissions:      p9.FileMode(stat.Mode),
+			UID:              p9.UID(stat.UID),
+			GID:              p9.GID(stat.GID),
+			Size:             stat.Size,
+			ATimeSeconds:     uint64(stat.Atime.Sec),
+			ATimeNanoSeconds: uint64(stat.Atime.Nsec),
+			MTimeSeconds:     uint64(stat.Mtime.Sec),
+			MTimeNanoSeconds: uint64(stat.Mtime.Nsec),
+		}); err != nil {
+			return err
+		}
+	}
+	if d.fs.opts.interop == InteropModeShared {
+		// There's no point to updating d's metadata in this case since it'll
+		// be overwritten by revalidation before the next time it's used
+		// anyway. (InteropModeShared inhibits client caching of regular file
+		// data, so there's no cache to truncate either.)
+		return nil
+	}
+	now, haveNow := nowFromContext(ctx)
+	if !haveNow {
+		ctx.Warningf("gofer.dentry.setStat: current time not available")
+	}
+	if stat.Mask&linux.STATX_MODE != 0 {
+		atomic.StoreUint32(&d.mode, d.fileType()|uint32(stat.Mode))
+	}
+	if stat.Mask&linux.STATX_UID != 0 {
+		atomic.StoreUint32(&d.uid, stat.UID)
+	}
+	if stat.Mask&linux.STATX_GID != 0 {
+		atomic.StoreUint32(&d.gid, stat.GID)
+	}
+	if setLocalAtime {
+		if stat.Atime.Nsec == linux.UTIME_NOW {
+			if haveNow {
+				atomic.StoreInt64(&d.atime, now)
+			}
+		} else {
+			atomic.StoreInt64(&d.atime, dentryTimestampFromStatx(stat.Atime))
+		}
+	}
+	if setLocalMtime {
+		if stat.Mtime.Nsec == linux.UTIME_NOW {
+			if haveNow {
+				atomic.StoreInt64(&d.mtime, now)
+			}
+		} else {
+			atomic.StoreInt64(&d.mtime, dentryTimestampFromStatx(stat.Mtime))
+		}
+	}
+	if haveNow {
+		atomic.StoreInt64(&d.ctime, now)
+	}
+	if stat.Mask&linux.STATX_SIZE != 0 {
+		d.dataMu.Lock()
+		oldSize := d.size
+		d.size = stat.Size
+		// d.dataMu must be unlocked to lock d.mapsMu and invalidate mappings
+		// below. This allows concurrent calls to Read/Translate/etc. These
+		// functions synchronize with truncation by refusing to use cache
+		// contents beyond the new d.size. (We are still holding d.metadataMu,
+		// so we can't race with Write or another truncate.)
+		d.dataMu.Unlock()
+		if d.size < oldSize {
+			oldpgend := pageRoundUp(oldSize)
+			newpgend := pageRoundUp(d.size)
+			if oldpgend != newpgend {
+				d.mapsMu.Lock()
+				d.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{
+					// Compare Linux's mm/truncate.c:truncate_setsize() =>
+					// truncate_pagecache() =>
+					// mm/memory.c:unmap_mapping_range(evencows=1).
+					InvalidatePrivate: true,
+				})
+				d.mapsMu.Unlock()
+			}
+			// We are now guaranteed that there are no translations of
+			// truncated pages, and can remove them from the cache. Since
+			// truncated pages have been removed from the remote file, they
+			// should be dropped without being written back.
+			d.dataMu.Lock()
+			d.cache.Truncate(d.size, d.fs.mfp.MemoryFile())
+			d.dirty.KeepClean(memmap.MappableRange{d.size, oldpgend})
+			d.dataMu.Unlock()
+		}
+	}
+	return nil
+}
+
+func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error {
+	return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&d.mode))&0777, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid)))
+}
+
+// IncRef implements vfs.DentryImpl.IncRef.
+func (d *dentry) IncRef() {
+	// d.refs may be 0 if d.fs.renameMu is locked, which serializes against
+	// d.checkCachingLocked().
+	atomic.AddInt64(&d.refs, 1)
+}
+
+// TryIncRef implements vfs.DentryImpl.TryIncRef.
+func (d *dentry) TryIncRef() bool {
+	for {
+		refs := atomic.LoadInt64(&d.refs)
+		if refs == 0 {
+			return false
+		}
+		if atomic.CompareAndSwapInt64(&d.refs, refs, refs+1) {
+			return true
+		}
+	}
+}
+
+// DecRef implements vfs.DentryImpl.DecRef.
+func (d *dentry) DecRef() {
+	if refs := atomic.AddInt64(&d.refs, -1); refs == 0 {
+		d.fs.renameMu.Lock()
+		d.checkCachingLocked()
+		d.fs.renameMu.Unlock()
+	} else if refs < 0 {
+		panic("gofer.dentry.DecRef() called without holding a reference")
+	}
+}
+
+// checkCachingLocked should be called after d's reference count becomes 0 or it
+// becomes disowned.
+//
+// Preconditions: d.fs.renameMu must be locked for writing.
+func (d *dentry) checkCachingLocked() {
+	// Dentries with a non-zero reference count must be retained. (The only way
+	// to obtain a reference on a dentry with zero references is via path
+	// resolution, which requires renameMu, so if d.refs is zero then it will
+	// remain zero while we hold renameMu for writing.)
+	if atomic.LoadInt64(&d.refs) != 0 {
+		if d.cached {
+			d.fs.cachedDentries.Remove(d)
+			d.fs.cachedDentriesLen--
+			d.cached = false
+		}
+		return
+	}
+	// Non-child dentries with zero references are no longer reachable by path
+	// resolution and should be dropped immediately.
+	if d.vfsd.Parent() == nil || d.vfsd.IsDisowned() {
+		if d.cached {
+			d.fs.cachedDentries.Remove(d)
+			d.fs.cachedDentriesLen--
+			d.cached = false
+		}
+		d.destroyLocked()
+		return
+	}
+	// If d is already cached, just move it to the front of the LRU.
+	if d.cached {
+		d.fs.cachedDentries.Remove(d)
+		d.fs.cachedDentries.PushFront(d)
+		return
+	}
+	// Cache the dentry, then evict the least recently used cached dentry if
+	// the cache becomes over-full.
+	d.fs.cachedDentries.PushFront(d)
+	d.fs.cachedDentriesLen++
+	d.cached = true
+	if d.fs.cachedDentriesLen > d.fs.opts.maxCachedDentries {
+		victim := d.fs.cachedDentries.Back()
+		d.fs.cachedDentries.Remove(victim)
+		d.fs.cachedDentriesLen--
+		victim.cached = false
+		// victim.refs may have become non-zero from an earlier path
+		// resolution since it was inserted into fs.cachedDentries; see
+		// dentry.incRefLocked(). Either way, we brought
+		// fs.cachedDentriesLen back down to fs.opts.maxCachedDentries, so
+		// we don't loop.
+		if atomic.LoadInt64(&victim.refs) == 0 {
+			if victimParentVFSD := victim.vfsd.Parent(); victimParentVFSD != nil {
+				victimParent := victimParentVFSD.Impl().(*dentry)
+				victimParent.dirMu.Lock()
+				if !victim.vfsd.IsDisowned() {
+					// victim can't be a mount point (in any mount
+					// namespace), since VFS holds references on mount
+					// points.
+					d.fs.vfsfs.VirtualFilesystem().ForceDeleteDentry(&victim.vfsd)
+					// We're only deleting the dentry, not the file it
+					// represents, so we don't need to update
+					// victimParent.dirents etc.
+				}
+				victimParent.dirMu.Unlock()
+			}
+			victim.destroyLocked()
+		}
+	}
+}
+
+// Preconditions: d.fs.renameMu must be locked for writing. d.refs == 0. d is
+// not a child dentry.
+func (d *dentry) destroyLocked() {
+	ctx := context.Background()
+	d.handleMu.Lock()
+	if !d.handle.file.isNil() {
+		mf := d.fs.mfp.MemoryFile()
+		d.dataMu.Lock()
+		// Write dirty pages back to the remote filesystem.
+		if d.handleWritable {
+			if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil {
+				log.Warningf("gofer.dentry.DecRef: failed to write dirty data back: %v", err)
+			}
+		}
+		// Discard cached data.
+		d.cache.DropAll(mf)
+		d.dirty.RemoveAll()
+		d.dataMu.Unlock()
+		// Clunk open fids and close open host FDs.
+		d.handle.close(ctx)
+	}
+	d.handleMu.Unlock()
+	d.file.close(ctx)
+	// Remove d from the set of all dentries.
+	d.fs.syncMu.Lock()
+	delete(d.fs.dentries, d)
+	d.fs.syncMu.Unlock()
+	// Drop the reference held by d on its parent.
+	if parentVFSD := d.vfsd.Parent(); parentVFSD != nil {
+		parent := parentVFSD.Impl().(*dentry)
+		// This is parent.DecRef() without recursive locking of d.fs.renameMu.
+		if refs := atomic.AddInt64(&parent.refs, -1); refs == 0 {
+			parent.checkCachingLocked()
+		} else if refs < 0 {
+			panic("gofer.dentry.DecRef() called without holding a reference")
+		}
+	}
+}
+
+func (d *dentry) isDeleted() bool {
+	return atomic.LoadUint32(&d.deleted) != 0
+}
+
+func (d *dentry) setDeleted() {
+	atomic.StoreUint32(&d.deleted, 1)
+}
+
+func (d *dentry) listxattr(ctx context.Context) ([]string, error) {
+	return nil, syserror.ENOTSUP
+}
+
+func (d *dentry) getxattr(ctx context.Context, name string) (string, error) {
+	// TODO(jamieliu): add vfs.GetxattrOptions.Size
+	return d.file.getXattr(ctx, name, linux.XATTR_SIZE_MAX)
+}
+
+func (d *dentry) setxattr(ctx context.Context, opts *vfs.SetxattrOptions) error {
+	return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags)
+}
+
+func (d *dentry) removexattr(ctx context.Context, name string) error {
+	return syserror.ENOTSUP
+}
+
+// Preconditions: d.isRegularFile() || d.isDirectory().
+func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool) error {
+	// O_TRUNC unconditionally requires us to obtain a new handle (opened with
+	// O_TRUNC).
+	if !trunc {
+		d.handleMu.RLock()
+		if (!read || d.handleReadable) && (!write || d.handleWritable) {
+			// The current handle is sufficient.
+			d.handleMu.RUnlock()
+			return nil
+		}
+		d.handleMu.RUnlock()
+	}
+
+	haveOldFD := false
+	d.handleMu.Lock()
+	if (read && !d.handleReadable) || (write && !d.handleWritable) || trunc {
+		// Get a new handle.
+		wantReadable := d.handleReadable || read
+		wantWritable := d.handleWritable || write
+		h, err := openHandle(ctx, d.file, wantReadable, wantWritable, trunc)
+		if err != nil {
+			d.handleMu.Unlock()
+			return err
+		}
+		if !d.handle.file.isNil() {
+			// Check that old and new handles are compatible: If the old handle
+			// includes a host file descriptor but the new one does not, or
+			// vice versa, old and new memory mappings may be incoherent.
+			haveOldFD = d.handle.fd >= 0
+			haveNewFD := h.fd >= 0
+			if haveOldFD != haveNewFD {
+				d.handleMu.Unlock()
+				ctx.Warningf("gofer.dentry.ensureSharedHandle: can't change host FD availability from %v to %v across dentry handle upgrade", haveOldFD, haveNewFD)
+				h.close(ctx)
+				return syserror.EIO
+			}
+			if haveOldFD {
+				// We may have raced with callers of d.pf.FD() that are now
+				// using the old file descriptor, preventing us from safely
+				// closing it. We could handle this by invalidating existing
+				// memmap.Translations, but this is expensive. Instead, use
+				// dup2() to make the old file descriptor refer to the new file
+				// description, then close the new file descriptor (which is no
+				// longer needed). Racing callers may use the old or new file
+				// description, but this doesn't matter since they refer to the
+				// same file (unless d.fs.opts.overlayfsStaleRead is true,
+				// which we handle separately).
+				if err := syscall.Dup2(int(h.fd), int(d.handle.fd)); err != nil {
+					d.handleMu.Unlock()
+					ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err)
+					h.close(ctx)
+					return err
+				}
+				syscall.Close(int(h.fd))
+				h.fd = d.handle.fd
+				if d.fs.opts.overlayfsStaleRead {
+					// Replace sentry mappings of the old FD with mappings of
+					// the new FD, since the two are not necessarily coherent.
+					if err := d.pf.hostFileMapper.RegenerateMappings(int(h.fd)); err != nil {
+						d.handleMu.Unlock()
+						ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to replace sentry mappings of old FD with mappings of new FD: %v", err)
+						h.close(ctx)
+						return err
+					}
+				}
+				// Clunk the old fid before making the new handle visible (by
+				// unlocking d.handleMu).
+				d.handle.file.close(ctx)
+			}
+		}
+		// Switch to the new handle.
+		d.handle = h
+		d.handleReadable = wantReadable
+		d.handleWritable = wantWritable
+	}
+	d.handleMu.Unlock()
+
+	if d.fs.opts.overlayfsStaleRead && haveOldFD {
+		// Invalidate application mappings that may be using the old FD; they
+		// will be replaced with mappings using the new FD after future calls
+		// to d.Translate(). This requires holding d.mapsMu, which precedes
+		// d.handleMu in the lock order.
+		d.mapsMu.Lock()
+		d.mappings.InvalidateAll(memmap.InvalidateOpts{})
+		d.mapsMu.Unlock()
+	}
+
+	return nil
+}
+
+// fileDescription is embedded by gofer 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) dentry() *dentry {
+	return fd.vfsfd.Dentry().Impl().(*dentry)
+}
+
+// Stat implements vfs.FileDescriptionImpl.Stat.
+func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
+	d := fd.dentry()
+	if d.fs.opts.interop == InteropModeShared && opts.Mask&(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE|linux.STATX_BLOCKS|linux.STATX_BTIME) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC {
+		// TODO(jamieliu): Use specialFileFD.handle.file for the getattr if
+		// available?
+		if err := d.updateFromGetattr(ctx); err != nil {
+			return linux.Statx{}, err
+		}
+	}
+	var stat linux.Statx
+	d.statTo(&stat)
+	return stat, nil
+}
+
+// SetStat implements vfs.FileDescriptionImpl.SetStat.
+func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
+	return fd.dentry().setStat(ctx, auth.CredentialsFromContext(ctx), &opts.Stat, fd.vfsfd.Mount())
+}
+
+// Listxattr implements vfs.FileDescriptionImpl.Listxattr.
+func (fd *fileDescription) Listxattr(ctx context.Context) ([]string, error) {
+	return fd.dentry().listxattr(ctx)
+}
+
+// Getxattr implements vfs.FileDescriptionImpl.Getxattr.
+func (fd *fileDescription) Getxattr(ctx context.Context, name string) (string, error) {
+	return fd.dentry().getxattr(ctx, name)
+}
+
+// Setxattr implements vfs.FileDescriptionImpl.Setxattr.
+func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error {
+	return fd.dentry().setxattr(ctx, &opts)
+}
+
+// Removexattr implements vfs.FileDescriptionImpl.Removexattr.
+func (fd *fileDescription) Removexattr(ctx context.Context, name string) error {
+	return fd.dentry().removexattr(ctx, name)
+}
diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go
new file mode 100644
index 000000000..cfe66f797
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/handle.go
@@ -0,0 +1,135 @@
+// 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 gofer
+
+import (
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/p9"
+	"gvisor.dev/gvisor/pkg/safemem"
+)
+
+// handle represents a remote "open file descriptor", consisting of an opened
+// fid (p9.File) and optionally a host file descriptor.
+type handle struct {
+	file p9file
+	fd   int32 // -1 if unavailable
+}
+
+// Preconditions: read || write.
+func openHandle(ctx context.Context, file p9file, read, write, trunc bool) (handle, error) {
+	_, newfile, err := file.walk(ctx, nil)
+	if err != nil {
+		return handle{fd: -1}, err
+	}
+	var flags p9.OpenFlags
+	switch {
+	case read && !write:
+		flags = p9.ReadOnly
+	case !read && write:
+		flags = p9.WriteOnly
+	case read && write:
+		flags = p9.ReadWrite
+	}
+	if trunc {
+		flags |= p9.OpenTruncate
+	}
+	fdobj, _, _, err := newfile.open(ctx, flags)
+	if err != nil {
+		newfile.close(ctx)
+		return handle{fd: -1}, err
+	}
+	fd := int32(-1)
+	if fdobj != nil {
+		fd = int32(fdobj.Release())
+	}
+	return handle{
+		file: newfile,
+		fd:   fd,
+	}, nil
+}
+
+func (h *handle) close(ctx context.Context) {
+	h.file.close(ctx)
+	h.file = p9file{}
+	if h.fd >= 0 {
+		syscall.Close(int(h.fd))
+		h.fd = -1
+	}
+}
+
+func (h *handle) readToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) {
+	if dsts.IsEmpty() {
+		return 0, nil
+	}
+	if h.fd >= 0 {
+		ctx.UninterruptibleSleepStart(false)
+		n, err := hostPreadv(h.fd, dsts, int64(offset))
+		ctx.UninterruptibleSleepFinish(false)
+		return n, err
+	}
+	if dsts.NumBlocks() == 1 && !dsts.Head().NeedSafecopy() {
+		n, err := h.file.readAt(ctx, dsts.Head().ToSlice(), offset)
+		return uint64(n), err
+	}
+	// Buffer the read since p9.File.ReadAt() takes []byte.
+	buf := make([]byte, dsts.NumBytes())
+	n, err := h.file.readAt(ctx, buf, offset)
+	if n == 0 {
+		return 0, err
+	}
+	if cp, cperr := safemem.CopySeq(dsts, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:n]))); cperr != nil {
+		return cp, cperr
+	}
+	return uint64(n), err
+}
+
+func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) {
+	if srcs.IsEmpty() {
+		return 0, nil
+	}
+	if h.fd >= 0 {
+		ctx.UninterruptibleSleepStart(false)
+		n, err := hostPwritev(h.fd, srcs, int64(offset))
+		ctx.UninterruptibleSleepFinish(false)
+		return n, err
+	}
+	if srcs.NumBlocks() == 1 && !srcs.Head().NeedSafecopy() {
+		n, err := h.file.writeAt(ctx, srcs.Head().ToSlice(), offset)
+		return uint64(n), err
+	}
+	// Buffer the write since p9.File.WriteAt() takes []byte.
+	buf := make([]byte, srcs.NumBytes())
+	cp, cperr := safemem.CopySeq(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), srcs)
+	if cp == 0 {
+		return 0, cperr
+	}
+	n, err := h.file.writeAt(ctx, buf[:cp], offset)
+	if err != nil {
+		return uint64(n), err
+	}
+	return cp, cperr
+}
+
+func (h *handle) sync(ctx context.Context) error {
+	if h.fd >= 0 {
+		ctx.UninterruptibleSleepStart(false)
+		err := syscall.Fsync(int(h.fd))
+		ctx.UninterruptibleSleepFinish(false)
+		return err
+	}
+	return h.file.fsync(ctx)
+}
diff --git a/pkg/sentry/fsimpl/gofer/handle_unsafe.go b/pkg/sentry/fsimpl/gofer/handle_unsafe.go
new file mode 100644
index 000000000..19560ab26
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/handle_unsafe.go
@@ -0,0 +1,66 @@
+// 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 gofer
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.dev/gvisor/pkg/safemem"
+)
+
+// Preconditions: !dsts.IsEmpty().
+func hostPreadv(fd int32, dsts safemem.BlockSeq, off int64) (uint64, error) {
+	// No buffering is necessary regardless of safecopy; host syscalls will
+	// return EFAULT if appropriate, instead of raising SIGBUS.
+	if dsts.NumBlocks() == 1 {
+		// Use pread() instead of preadv() to avoid iovec allocation and
+		// copying.
+		dst := dsts.Head()
+		n, _, e := syscall.Syscall6(syscall.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(off), 0, 0)
+		if e != 0 {
+			return 0, e
+		}
+		return uint64(n), nil
+	}
+	iovs := safemem.IovecsFromBlockSeq(dsts)
+	n, _, e := syscall.Syscall6(syscall.SYS_PREADV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0)
+	if e != 0 {
+		return 0, e
+	}
+	return uint64(n), nil
+}
+
+// Preconditions: !srcs.IsEmpty().
+func hostPwritev(fd int32, srcs safemem.BlockSeq, off int64) (uint64, error) {
+	// No buffering is necessary regardless of safecopy; host syscalls will
+	// return EFAULT if appropriate, instead of raising SIGBUS.
+	if srcs.NumBlocks() == 1 {
+		// Use pwrite() instead of pwritev() to avoid iovec allocation and
+		// copying.
+		src := srcs.Head()
+		n, _, e := syscall.Syscall6(syscall.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(off), 0, 0)
+		if e != 0 {
+			return 0, e
+		}
+		return uint64(n), nil
+	}
+	iovs := safemem.IovecsFromBlockSeq(srcs)
+	n, _, e := syscall.Syscall6(syscall.SYS_PWRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0)
+	if e != 0 {
+		return 0, e
+	}
+	return uint64(n), nil
+}
diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go
new file mode 100644
index 000000000..755ac2985
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/p9file.go
@@ -0,0 +1,219 @@
+// 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 gofer
+
+import (
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fd"
+	"gvisor.dev/gvisor/pkg/p9"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// p9file is a wrapper around p9.File that provides methods that are
+// Context-aware.
+type p9file struct {
+	file p9.File
+}
+
+func (f p9file) isNil() bool {
+	return f.file == nil
+}
+
+func (f p9file) walk(ctx context.Context, names []string) ([]p9.QID, p9file, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qids, newfile, err := f.file.Walk(names)
+	ctx.UninterruptibleSleepFinish(false)
+	return qids, p9file{newfile}, err
+}
+
+func (f p9file) walkGetAttr(ctx context.Context, names []string) ([]p9.QID, p9file, p9.AttrMask, p9.Attr, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qids, newfile, attrMask, attr, err := f.file.WalkGetAttr(names)
+	ctx.UninterruptibleSleepFinish(false)
+	return qids, p9file{newfile}, attrMask, attr, err
+}
+
+// walkGetAttrOne is a wrapper around p9.File.WalkGetAttr that takes a single
+// path component and returns a single qid.
+func (f p9file) walkGetAttrOne(ctx context.Context, name string) (p9.QID, p9file, p9.AttrMask, p9.Attr, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qids, newfile, attrMask, attr, err := f.file.WalkGetAttr([]string{name})
+	ctx.UninterruptibleSleepFinish(false)
+	if err != nil {
+		return p9.QID{}, p9file{}, p9.AttrMask{}, p9.Attr{}, err
+	}
+	if len(qids) != 1 {
+		ctx.Warningf("p9.File.WalkGetAttr returned %d qids (%v), wanted 1", len(qids), qids)
+		if newfile != nil {
+			p9file{newfile}.close(ctx)
+		}
+		return p9.QID{}, p9file{}, p9.AttrMask{}, p9.Attr{}, syserror.EIO
+	}
+	return qids[0], p9file{newfile}, attrMask, attr, nil
+}
+
+func (f p9file) statFS(ctx context.Context) (p9.FSStat, error) {
+	ctx.UninterruptibleSleepStart(false)
+	fsstat, err := f.file.StatFS()
+	ctx.UninterruptibleSleepFinish(false)
+	return fsstat, err
+}
+
+func (f p9file) getAttr(ctx context.Context, req p9.AttrMask) (p9.QID, p9.AttrMask, p9.Attr, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qid, attrMask, attr, err := f.file.GetAttr(req)
+	ctx.UninterruptibleSleepFinish(false)
+	return qid, attrMask, attr, err
+}
+
+func (f p9file) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAttr) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.SetAttr(valid, attr)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) getXattr(ctx context.Context, name string, size uint64) (string, error) {
+	ctx.UninterruptibleSleepStart(false)
+	val, err := f.file.GetXattr(name, size)
+	ctx.UninterruptibleSleepFinish(false)
+	return val, err
+}
+
+func (f p9file) setXattr(ctx context.Context, name, value string, flags uint32) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.SetXattr(name, value, flags)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.Allocate(mode, offset, length)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) close(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.Close()
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) open(ctx context.Context, flags p9.OpenFlags) (*fd.FD, p9.QID, uint32, error) {
+	ctx.UninterruptibleSleepStart(false)
+	fdobj, qid, iounit, err := f.file.Open(flags)
+	ctx.UninterruptibleSleepFinish(false)
+	return fdobj, qid, iounit, err
+}
+
+func (f p9file) readAt(ctx context.Context, p []byte, offset uint64) (int, error) {
+	ctx.UninterruptibleSleepStart(false)
+	n, err := f.file.ReadAt(p, offset)
+	ctx.UninterruptibleSleepFinish(false)
+	return n, err
+}
+
+func (f p9file) writeAt(ctx context.Context, p []byte, offset uint64) (int, error) {
+	ctx.UninterruptibleSleepStart(false)
+	n, err := f.file.WriteAt(p, offset)
+	ctx.UninterruptibleSleepFinish(false)
+	return n, err
+}
+
+func (f p9file) fsync(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.FSync()
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) create(ctx context.Context, name string, flags p9.OpenFlags, permissions p9.FileMode, uid p9.UID, gid p9.GID) (*fd.FD, p9file, p9.QID, uint32, error) {
+	ctx.UninterruptibleSleepStart(false)
+	fdobj, newfile, qid, iounit, err := f.file.Create(name, flags, permissions, uid, gid)
+	ctx.UninterruptibleSleepFinish(false)
+	return fdobj, p9file{newfile}, qid, iounit, err
+}
+
+func (f p9file) mkdir(ctx context.Context, name string, permissions p9.FileMode, uid p9.UID, gid p9.GID) (p9.QID, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qid, err := f.file.Mkdir(name, permissions, uid, gid)
+	ctx.UninterruptibleSleepFinish(false)
+	return qid, err
+}
+
+func (f p9file) symlink(ctx context.Context, oldName string, newName string, uid p9.UID, gid p9.GID) (p9.QID, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qid, err := f.file.Symlink(oldName, newName, uid, gid)
+	ctx.UninterruptibleSleepFinish(false)
+	return qid, err
+}
+
+func (f p9file) link(ctx context.Context, target p9file, newName string) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.Link(target.file, newName)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) mknod(ctx context.Context, name string, mode p9.FileMode, major uint32, minor uint32, uid p9.UID, gid p9.GID) (p9.QID, error) {
+	ctx.UninterruptibleSleepStart(false)
+	qid, err := f.file.Mknod(name, mode, major, minor, uid, gid)
+	ctx.UninterruptibleSleepFinish(false)
+	return qid, err
+}
+
+func (f p9file) rename(ctx context.Context, newDir p9file, newName string) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.Rename(newDir.file, newName)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) unlinkAt(ctx context.Context, name string, flags uint32) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.UnlinkAt(name, flags)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) readdir(ctx context.Context, offset uint64, count uint32) ([]p9.Dirent, error) {
+	ctx.UninterruptibleSleepStart(false)
+	dirents, err := f.file.Readdir(offset, count)
+	ctx.UninterruptibleSleepFinish(false)
+	return dirents, err
+}
+
+func (f p9file) readlink(ctx context.Context) (string, error) {
+	ctx.UninterruptibleSleepStart(false)
+	target, err := f.file.Readlink()
+	ctx.UninterruptibleSleepFinish(false)
+	return target, err
+}
+
+func (f p9file) flush(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := f.file.Flush()
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
+func (f p9file) connect(ctx context.Context, flags p9.ConnectFlags) (*fd.FD, error) {
+	ctx.UninterruptibleSleepStart(false)
+	fdobj, err := f.file.Connect(flags)
+	ctx.UninterruptibleSleepFinish(false)
+	return fdobj, err
+}
diff --git a/pkg/sentry/fsimpl/gofer/pagemath.go b/pkg/sentry/fsimpl/gofer/pagemath.go
new file mode 100644
index 000000000..847cb0784
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/pagemath.go
@@ -0,0 +1,31 @@
+// 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 gofer
+
+import (
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// This are equivalent to usermem.Addr.RoundDown/Up, but without the
+// potentially truncating conversion to usermem.Addr. This is necessary because
+// there is no way to define generic "PageRoundDown/Up" functions in Go.
+
+func pageRoundDown(x uint64) uint64 {
+	return x &^ (usermem.PageSize - 1)
+}
+
+func pageRoundUp(x uint64) uint64 {
+	return pageRoundDown(x + usermem.PageSize - 1)
+}
diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go
new file mode 100644
index 000000000..8e11e06b3
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/regular_file.go
@@ -0,0 +1,860 @@
+// 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 gofer
+
+import (
+	"fmt"
+	"io"
+	"math"
+	"sync"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/safemem"
+	"gvisor.dev/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
+	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
+	"gvisor.dev/gvisor/pkg/sentry/platform"
+	"gvisor.dev/gvisor/pkg/sentry/usage"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+func (d *dentry) isRegularFile() bool {
+	return d.fileType() == linux.S_IFREG
+}
+
+type regularFileFD struct {
+	fileDescription
+
+	// off is the file offset. off is protected by mu.
+	mu  sync.Mutex
+	off int64
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *regularFileFD) Release() {
+}
+
+// OnClose implements vfs.FileDescriptionImpl.OnClose.
+func (fd *regularFileFD) OnClose(ctx context.Context) error {
+	if !fd.vfsfd.IsWritable() {
+		return nil
+	}
+	// Skip flushing if writes may be buffered by the client, since (as with
+	// the VFS1 client) we don't flush buffered writes on close anyway.
+	d := fd.dentry()
+	if d.fs.opts.interop == InteropModeExclusive {
+		return nil
+	}
+	d.handleMu.RLock()
+	defer d.handleMu.RUnlock()
+	return d.handle.file.flush(ctx)
+}
+
+// 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 opts.Flags != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	// Check for reading at EOF before calling into MM (but not under
+	// InteropModeShared, which makes d.size unreliable).
+	d := fd.dentry()
+	if d.fs.opts.interop != InteropModeShared && uint64(offset) >= atomic.LoadUint64(&d.size) {
+		return 0, io.EOF
+	}
+
+	if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 {
+		// Lock d.metadataMu for the rest of the read to prevent d.size from
+		// changing.
+		d.metadataMu.Lock()
+		defer d.metadataMu.Unlock()
+		// Write dirty cached pages that will be touched by the read back to
+		// the remote file.
+		if err := d.writeback(ctx, offset, dst.NumBytes()); err != nil {
+			return 0, err
+		}
+	}
+
+	rw := getDentryReadWriter(ctx, d, offset)
+	if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 {
+		// Require the read to go to the remote file.
+		rw.direct = true
+	}
+	n, err := dst.CopyOutFrom(ctx, rw)
+	putDentryReadWriter(rw)
+	if d.fs.opts.interop != InteropModeShared {
+		// Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed().
+		d.touchAtime(ctx, 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.mu.Lock()
+	n, err := fd.PRead(ctx, dst, fd.off, opts)
+	fd.off += n
+	fd.mu.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
+	}
+	if opts.Flags != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	d := fd.dentry()
+	d.metadataMu.Lock()
+	defer d.metadataMu.Unlock()
+	if d.fs.opts.interop != InteropModeShared {
+		// Compare Linux's mm/filemap.c:__generic_file_write_iter() =>
+		// file_update_time(). This is d.touchCMtime(), but without locking
+		// d.metadataMu (recursively).
+		if now, ok := nowFromContext(ctx); ok {
+			atomic.StoreInt64(&d.mtime, now)
+			atomic.StoreInt64(&d.ctime, now)
+		}
+	}
+	if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 {
+		// Write dirty cached pages that will be touched by the write back to
+		// the remote file.
+		if err := d.writeback(ctx, offset, src.NumBytes()); err != nil {
+			return 0, err
+		}
+		// Remove touched pages from the cache.
+		pgstart := pageRoundDown(uint64(offset))
+		pgend := pageRoundUp(uint64(offset + src.NumBytes()))
+		if pgend < pgstart {
+			return 0, syserror.EINVAL
+		}
+		mr := memmap.MappableRange{pgstart, pgend}
+		var freed []platform.FileRange
+		d.dataMu.Lock()
+		cseg := d.cache.LowerBoundSegment(mr.Start)
+		for cseg.Ok() && cseg.Start() < mr.End {
+			cseg = d.cache.Isolate(cseg, mr)
+			freed = append(freed, platform.FileRange{cseg.Value(), cseg.Value() + cseg.Range().Length()})
+			cseg = d.cache.Remove(cseg).NextSegment()
+		}
+		d.dataMu.Unlock()
+		// Invalidate mappings of removed pages.
+		d.mapsMu.Lock()
+		d.mappings.Invalidate(mr, memmap.InvalidateOpts{})
+		d.mapsMu.Unlock()
+		// Finally free pages removed from the cache.
+		mf := d.fs.mfp.MemoryFile()
+		for _, freedFR := range freed {
+			mf.DecRef(freedFR)
+		}
+	}
+	rw := getDentryReadWriter(ctx, d, offset)
+	if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 {
+		// Require the write to go to the remote file.
+		rw.direct = true
+	}
+	n, err := src.CopyInTo(ctx, rw)
+	putDentryReadWriter(rw)
+	if n != 0 && fd.vfsfd.StatusFlags()&(linux.O_DSYNC|linux.O_SYNC) != 0 {
+		// Write dirty cached pages touched by the write back to the remote
+		// file.
+		if err := d.writeback(ctx, offset, src.NumBytes()); err != nil {
+			return 0, err
+		}
+		// Request the remote filesystem to sync the remote file.
+		if err := d.handle.file.fsync(ctx); err != nil {
+			return 0, err
+		}
+	}
+	return n, err
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	fd.mu.Lock()
+	n, err := fd.PWrite(ctx, src, fd.off, opts)
+	fd.off += n
+	fd.mu.Unlock()
+	return n, err
+}
+
+type dentryReadWriter struct {
+	ctx    context.Context
+	d      *dentry
+	off    uint64
+	direct bool
+}
+
+var dentryReadWriterPool = sync.Pool{
+	New: func() interface{} {
+		return &dentryReadWriter{}
+	},
+}
+
+func getDentryReadWriter(ctx context.Context, d *dentry, offset int64) *dentryReadWriter {
+	rw := dentryReadWriterPool.Get().(*dentryReadWriter)
+	rw.ctx = ctx
+	rw.d = d
+	rw.off = uint64(offset)
+	rw.direct = false
+	return rw
+}
+
+func putDentryReadWriter(rw *dentryReadWriter) {
+	rw.ctx = nil
+	rw.d = nil
+	dentryReadWriterPool.Put(rw)
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *dentryReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	if dsts.IsEmpty() {
+		return 0, nil
+	}
+
+	// If we have a mmappable host FD (which must be used here to ensure
+	// coherence with memory-mapped I/O), or if InteropModeShared is in effect
+	// (which prevents us from caching file contents and makes dentry.size
+	// unreliable), or if the file was opened O_DIRECT, read directly from
+	// dentry.handle without locking dentry.dataMu.
+	rw.d.handleMu.RLock()
+	if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct {
+		n, err := rw.d.handle.readToBlocksAt(rw.ctx, dsts, rw.off)
+		rw.d.handleMu.RUnlock()
+		rw.off += n
+		return n, err
+	}
+
+	// Otherwise read from/through the cache.
+	mf := rw.d.fs.mfp.MemoryFile()
+	fillCache := mf.ShouldCacheEvictable()
+	var dataMuUnlock func()
+	if fillCache {
+		rw.d.dataMu.Lock()
+		dataMuUnlock = rw.d.dataMu.Unlock
+	} else {
+		rw.d.dataMu.RLock()
+		dataMuUnlock = rw.d.dataMu.RUnlock
+	}
+
+	// Compute the range to read (limited by file size and overflow-checked).
+	if rw.off >= rw.d.size {
+		dataMuUnlock()
+		rw.d.handleMu.RUnlock()
+		return 0, io.EOF
+	}
+	end := rw.d.size
+	if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end {
+		end = rend
+	}
+
+	var done uint64
+	seg, gap := rw.d.cache.Find(rw.off)
+	for rw.off < end {
+		mr := memmap.MappableRange{rw.off, end}
+		switch {
+		case seg.Ok():
+			// Get internal mappings from the cache.
+			ims, err := mf.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read)
+			if err != nil {
+				dataMuUnlock()
+				rw.d.handleMu.RUnlock()
+				return done, err
+			}
+
+			// Copy from internal mappings.
+			n, err := safemem.CopySeq(dsts, ims)
+			done += n
+			rw.off += n
+			dsts = dsts.DropFirst64(n)
+			if err != nil {
+				dataMuUnlock()
+				rw.d.handleMu.RUnlock()
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			gapMR := gap.Range().Intersect(mr)
+			if fillCache {
+				// Read into the cache, then re-enter the loop to read from the
+				// cache.
+				reqMR := memmap.MappableRange{
+					Start: pageRoundDown(gapMR.Start),
+					End:   pageRoundUp(gapMR.End),
+				}
+				optMR := gap.Range()
+				err := rw.d.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), mf, usage.PageCache, rw.d.handle.readToBlocksAt)
+				mf.MarkEvictable(rw.d, pgalloc.EvictableRange{optMR.Start, optMR.End})
+				seg, gap = rw.d.cache.Find(rw.off)
+				if !seg.Ok() {
+					dataMuUnlock()
+					rw.d.handleMu.RUnlock()
+					return done, err
+				}
+				// err might have occurred in part of gap.Range() outside
+				// gapMR. Forget about it for now; if the error matters and
+				// persists, we'll run into it again in a later iteration of
+				// this loop.
+			} else {
+				// Read directly from the file.
+				gapDsts := dsts.TakeFirst64(gapMR.Length())
+				n, err := rw.d.handle.readToBlocksAt(rw.ctx, gapDsts, gapMR.Start)
+				done += n
+				rw.off += n
+				dsts = dsts.DropFirst64(n)
+				// Partial reads are fine. But we must stop reading.
+				if n != gapDsts.NumBytes() || err != nil {
+					dataMuUnlock()
+					rw.d.handleMu.RUnlock()
+					return done, err
+				}
+
+				// Continue.
+				seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{}
+			}
+		}
+	}
+	dataMuUnlock()
+	rw.d.handleMu.RUnlock()
+	return done, nil
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+//
+// Preconditions: rw.d.metadataMu must be locked.
+func (rw *dentryReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	if srcs.IsEmpty() {
+		return 0, nil
+	}
+
+	// If we have a mmappable host FD (which must be used here to ensure
+	// coherence with memory-mapped I/O), or if InteropModeShared is in effect
+	// (which prevents us from caching file contents), or if the file was
+	// opened with O_DIRECT, write directly to dentry.handle without locking
+	// dentry.dataMu.
+	rw.d.handleMu.RLock()
+	if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct {
+		n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, srcs, rw.off)
+		rw.d.handleMu.RUnlock()
+		rw.off += n
+		return n, err
+	}
+
+	// Otherwise write to/through the cache.
+	mf := rw.d.fs.mfp.MemoryFile()
+	rw.d.dataMu.Lock()
+
+	// Compute the range to write (overflow-checked).
+	start := rw.off
+	end := rw.off + srcs.NumBytes()
+	if end <= rw.off {
+		end = math.MaxInt64
+	}
+
+	var (
+		done   uint64
+		retErr error
+	)
+	seg, gap := rw.d.cache.Find(rw.off)
+	for rw.off < end {
+		mr := memmap.MappableRange{rw.off, end}
+		switch {
+		case seg.Ok():
+			// Get internal mappings from the cache.
+			segMR := seg.Range().Intersect(mr)
+			ims, err := mf.MapInternal(seg.FileRangeOf(segMR), usermem.Write)
+			if err != nil {
+				retErr = err
+				goto exitLoop
+			}
+
+			// Copy to internal mappings.
+			n, err := safemem.CopySeq(ims, srcs)
+			done += n
+			rw.off += n
+			srcs = srcs.DropFirst64(n)
+			rw.d.dirty.MarkDirty(segMR)
+			if err != nil {
+				retErr = err
+				goto exitLoop
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			// Write directly to the file. At present, we never fill the cache
+			// when writing, since doing so can convert small writes into
+			// inefficient read-modify-write cycles, and we have no mechanism
+			// for detecting or avoiding this.
+			gapMR := gap.Range().Intersect(mr)
+			gapSrcs := srcs.TakeFirst64(gapMR.Length())
+			n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, gapSrcs, gapMR.Start)
+			done += n
+			rw.off += n
+			srcs = srcs.DropFirst64(n)
+			// Partial writes are fine. But we must stop writing.
+			if n != gapSrcs.NumBytes() || err != nil {
+				retErr = err
+				goto exitLoop
+			}
+
+			// Continue.
+			seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{}
+		}
+	}
+exitLoop:
+	if rw.off > rw.d.size {
+		atomic.StoreUint64(&rw.d.size, rw.off)
+		// The remote file's size will implicitly be extended to the correct
+		// value when we write back to it.
+	}
+	// If InteropModeWritethrough is in effect, flush written data back to the
+	// remote filesystem.
+	if rw.d.fs.opts.interop == InteropModeWritethrough && done != 0 {
+		if err := fsutil.SyncDirty(rw.ctx, memmap.MappableRange{
+			Start: start,
+			End:   rw.off,
+		}, &rw.d.cache, &rw.d.dirty, rw.d.size, mf, rw.d.handle.writeFromBlocksAt); err != nil {
+			// We have no idea how many bytes were actually flushed.
+			rw.off = start
+			done = 0
+			retErr = err
+		}
+	}
+	rw.d.dataMu.Unlock()
+	rw.d.handleMu.RUnlock()
+	return done, retErr
+}
+
+func (d *dentry) writeback(ctx context.Context, offset, size int64) error {
+	if size == 0 {
+		return nil
+	}
+	d.handleMu.RLock()
+	defer d.handleMu.RUnlock()
+	d.dataMu.Lock()
+	defer d.dataMu.Unlock()
+	// Compute the range of valid bytes (overflow-checked).
+	if uint64(offset) >= d.size {
+		return nil
+	}
+	end := int64(d.size)
+	if rend := offset + size; rend > offset && rend < end {
+		end = rend
+	}
+	return fsutil.SyncDirty(ctx, memmap.MappableRange{
+		Start: uint64(offset),
+		End:   uint64(end),
+	}, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt)
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+	fd.mu.Lock()
+	defer fd.mu.Unlock()
+	switch whence {
+	case linux.SEEK_SET:
+		// Use offset as specified.
+	case linux.SEEK_CUR:
+		offset += fd.off
+	case linux.SEEK_END, linux.SEEK_DATA, linux.SEEK_HOLE:
+		// Ensure file size is up to date.
+		d := fd.dentry()
+		if fd.filesystem().opts.interop == InteropModeShared {
+			if err := d.updateFromGetattr(ctx); err != nil {
+				return 0, err
+			}
+		}
+		size := int64(atomic.LoadUint64(&d.size))
+		// For SEEK_DATA and SEEK_HOLE, treat the file as a single contiguous
+		// block of data.
+		switch whence {
+		case linux.SEEK_END:
+			offset += size
+		case linux.SEEK_DATA:
+			if offset > size {
+				return 0, syserror.ENXIO
+			}
+			// Use offset as specified.
+		case linux.SEEK_HOLE:
+			if offset > size {
+				return 0, syserror.ENXIO
+			}
+			offset = 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 fd.dentry().syncSharedHandle(ctx)
+}
+
+func (d *dentry) syncSharedHandle(ctx context.Context) error {
+	d.handleMu.RLock()
+	if !d.handleWritable {
+		d.handleMu.RUnlock()
+		return nil
+	}
+	d.dataMu.Lock()
+	// Write dirty cached data to the remote file.
+	err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt)
+	d.dataMu.Unlock()
+	if err == nil {
+		// Sync the remote file.
+		err = d.handle.sync(ctx)
+	}
+	d.handleMu.RUnlock()
+	return err
+}
+
+// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap.
+func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error {
+	d := fd.dentry()
+	switch d.fs.opts.interop {
+	case InteropModeExclusive:
+		// Any mapping is fine.
+	case InteropModeWritethrough:
+		// Shared writable mappings require a host FD, since otherwise we can't
+		// synchronously flush memory-mapped writes to the remote file.
+		if opts.Private || !opts.MaxPerms.Write {
+			break
+		}
+		fallthrough
+	case InteropModeShared:
+		// All mappings require a host FD to be coherent with other filesystem
+		// users.
+		if d.fs.opts.forcePageCache {
+			// Whether or not we have a host FD, we're not allowed to use it.
+			return syserror.ENODEV
+		}
+		d.handleMu.RLock()
+		haveFD := d.handle.fd >= 0
+		d.handleMu.RUnlock()
+		if !haveFD {
+			return syserror.ENODEV
+		}
+	default:
+		panic(fmt.Sprintf("unknown InteropMode %v", d.fs.opts.interop))
+	}
+	return vfs.GenericConfigureMMap(&fd.vfsfd, d, opts)
+}
+
+func (d *dentry) mayCachePages() bool {
+	if d.fs.opts.interop == InteropModeShared {
+		return false
+	}
+	if d.fs.opts.forcePageCache {
+		return true
+	}
+	d.handleMu.RLock()
+	haveFD := d.handle.fd >= 0
+	d.handleMu.RUnlock()
+	return haveFD
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (d *dentry) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error {
+	d.mapsMu.Lock()
+	mapped := d.mappings.AddMapping(ms, ar, offset, writable)
+	// Do this unconditionally since whether we have a host FD can change
+	// across save/restore.
+	for _, r := range mapped {
+		d.pf.hostFileMapper.IncRefOn(r)
+	}
+	if d.mayCachePages() {
+		// d.Evict() will refuse to evict memory-mapped pages, so tell the
+		// MemoryFile to not bother trying.
+		mf := d.fs.mfp.MemoryFile()
+		for _, r := range mapped {
+			mf.MarkUnevictable(d, pgalloc.EvictableRange{r.Start, r.End})
+		}
+	}
+	d.mapsMu.Unlock()
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (d *dentry) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) {
+	d.mapsMu.Lock()
+	unmapped := d.mappings.RemoveMapping(ms, ar, offset, writable)
+	for _, r := range unmapped {
+		d.pf.hostFileMapper.DecRefOn(r)
+	}
+	if d.mayCachePages() {
+		// Pages that are no longer referenced by any application memory
+		// mappings are now considered unused; allow MemoryFile to evict them
+		// when necessary.
+		mf := d.fs.mfp.MemoryFile()
+		d.dataMu.Lock()
+		for _, r := range unmapped {
+			// Since these pages are no longer mapped, they are no longer
+			// concurrently dirtyable by a writable memory mapping.
+			d.dirty.AllowClean(r)
+			mf.MarkEvictable(d, pgalloc.EvictableRange{r.Start, r.End})
+		}
+		d.dataMu.Unlock()
+	}
+	d.mapsMu.Unlock()
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (d *dentry) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error {
+	return d.AddMapping(ctx, ms, dstAR, offset, writable)
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (d *dentry) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	d.handleMu.RLock()
+	if d.handle.fd >= 0 && !d.fs.opts.forcePageCache {
+		d.handleMu.RUnlock()
+		mr := optional
+		if d.fs.opts.limitHostFDTranslation {
+			mr = maxFillRange(required, optional)
+		}
+		return []memmap.Translation{
+			{
+				Source: mr,
+				File:   &d.pf,
+				Offset: mr.Start,
+				Perms:  usermem.AnyAccess,
+			},
+		}, nil
+	}
+
+	d.dataMu.Lock()
+
+	// Constrain translations to d.size (rounded up) to prevent translation to
+	// pages that may be concurrently truncated.
+	pgend := pageRoundUp(d.size)
+	var beyondEOF bool
+	if required.End > pgend {
+		if required.Start >= pgend {
+			d.dataMu.Unlock()
+			d.handleMu.RUnlock()
+			return nil, &memmap.BusError{io.EOF}
+		}
+		beyondEOF = true
+		required.End = pgend
+	}
+	if optional.End > pgend {
+		optional.End = pgend
+	}
+
+	mf := d.fs.mfp.MemoryFile()
+	cerr := d.cache.Fill(ctx, required, maxFillRange(required, optional), mf, usage.PageCache, d.handle.readToBlocksAt)
+
+	var ts []memmap.Translation
+	var translatedEnd uint64
+	for seg := d.cache.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() {
+		segMR := seg.Range().Intersect(optional)
+		// TODO(jamieliu): Make Translations writable even if writability is
+		// not required if already kept-dirty by another writable translation.
+		perms := usermem.AccessType{
+			Read:    true,
+			Execute: true,
+		}
+		if at.Write {
+			// From this point forward, this memory can be dirtied through the
+			// mapping at any time.
+			d.dirty.KeepDirty(segMR)
+			perms.Write = true
+		}
+		ts = append(ts, memmap.Translation{
+			Source: segMR,
+			File:   mf,
+			Offset: seg.FileRangeOf(segMR).Start,
+			Perms:  perms,
+		})
+		translatedEnd = segMR.End
+	}
+
+	d.dataMu.Unlock()
+	d.handleMu.RUnlock()
+
+	// Don't return the error returned by c.cache.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
+}
+
+func maxFillRange(required, optional memmap.MappableRange) memmap.MappableRange {
+	const maxReadahead = 64 << 10 // 64 KB, chosen arbitrarily
+	if required.Length() >= maxReadahead {
+		return required
+	}
+	if optional.Length() <= maxReadahead {
+		return optional
+	}
+	optional.Start = required.Start
+	if optional.Length() <= maxReadahead {
+		return optional
+	}
+	optional.End = optional.Start + maxReadahead
+	return optional
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (d *dentry) InvalidateUnsavable(ctx context.Context) error {
+	// Whether we have a host fd (and consequently what platform.File is
+	// mapped) can change across save/restore, so invalidate all translations
+	// unconditionally.
+	d.mapsMu.Lock()
+	defer d.mapsMu.Unlock()
+	d.mappings.InvalidateAll(memmap.InvalidateOpts{})
+
+	// Write the cache's contents back to the remote file so that if we have a
+	// host fd after restore, the remote file's contents are coherent.
+	mf := d.fs.mfp.MemoryFile()
+	d.dataMu.Lock()
+	defer d.dataMu.Unlock()
+	if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil {
+		return err
+	}
+
+	// Discard the cache so that it's not stored in saved state. This is safe
+	// because per InvalidateUnsavable invariants, no new translations can have
+	// been returned after we invalidated all existing translations above.
+	d.cache.DropAll(mf)
+	d.dirty.RemoveAll()
+
+	return nil
+}
+
+// Evict implements pgalloc.EvictableMemoryUser.Evict.
+func (d *dentry) Evict(ctx context.Context, er pgalloc.EvictableRange) {
+	d.mapsMu.Lock()
+	defer d.mapsMu.Unlock()
+	d.dataMu.Lock()
+	defer d.dataMu.Unlock()
+
+	mr := memmap.MappableRange{er.Start, er.End}
+	mf := d.fs.mfp.MemoryFile()
+	// Only allow pages that are no longer memory-mapped to be evicted.
+	for mgap := d.mappings.LowerBoundGap(mr.Start); mgap.Ok() && mgap.Start() < mr.End; mgap = mgap.NextGap() {
+		mgapMR := mgap.Range().Intersect(mr)
+		if mgapMR.Length() == 0 {
+			continue
+		}
+		if err := fsutil.SyncDirty(ctx, mgapMR, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil {
+			log.Warningf("Failed to writeback cached data %v: %v", mgapMR, err)
+		}
+		d.cache.Drop(mgapMR, mf)
+		d.dirty.KeepClean(mgapMR)
+	}
+}
+
+// dentryPlatformFile implements platform.File. It exists solely because dentry
+// cannot implement both vfs.DentryImpl.IncRef and platform.File.IncRef.
+//
+// dentryPlatformFile is only used when a host FD representing the remote file
+// is available (i.e. dentry.handle.fd >= 0), and that FD is used for
+// application memory mappings (i.e. !filesystem.opts.forcePageCache).
+type dentryPlatformFile struct {
+	*dentry
+
+	// fdRefs counts references on platform.File offsets. fdRefs is protected
+	// by dentry.dataMu.
+	fdRefs fsutil.FrameRefSet
+
+	// If this dentry represents a regular file, and handle.fd >= 0,
+	// hostFileMapper caches mappings of handle.fd.
+	hostFileMapper fsutil.HostFileMapper
+}
+
+// IncRef implements platform.File.IncRef.
+func (d *dentryPlatformFile) IncRef(fr platform.FileRange) {
+	d.dataMu.Lock()
+	seg, gap := d.fdRefs.Find(fr.Start)
+	for {
+		switch {
+		case seg.Ok() && seg.Start() < fr.End:
+			seg = d.fdRefs.Isolate(seg, fr)
+			seg.SetValue(seg.Value() + 1)
+			seg, gap = seg.NextNonEmpty()
+		case gap.Ok() && gap.Start() < fr.End:
+			newRange := gap.Range().Intersect(fr)
+			usage.MemoryAccounting.Inc(newRange.Length(), usage.Mapped)
+			seg, gap = d.fdRefs.InsertWithoutMerging(gap, newRange, 1).NextNonEmpty()
+		default:
+			d.fdRefs.MergeAdjacent(fr)
+			d.dataMu.Unlock()
+			return
+		}
+	}
+}
+
+// DecRef implements platform.File.DecRef.
+func (d *dentryPlatformFile) DecRef(fr platform.FileRange) {
+	d.dataMu.Lock()
+	seg := d.fdRefs.FindSegment(fr.Start)
+
+	for seg.Ok() && seg.Start() < fr.End {
+		seg = d.fdRefs.Isolate(seg, fr)
+		if old := seg.Value(); old == 1 {
+			usage.MemoryAccounting.Dec(seg.Range().Length(), usage.Mapped)
+			seg = d.fdRefs.Remove(seg).NextSegment()
+		} else {
+			seg.SetValue(old - 1)
+			seg = seg.NextSegment()
+		}
+	}
+	d.fdRefs.MergeAdjacent(fr)
+	d.dataMu.Unlock()
+
+}
+
+// MapInternal implements platform.File.MapInternal.
+func (d *dentryPlatformFile) MapInternal(fr platform.FileRange, at usermem.AccessType) (safemem.BlockSeq, error) {
+	d.handleMu.RLock()
+	bs, err := d.hostFileMapper.MapInternal(fr, int(d.handle.fd), at.Write)
+	d.handleMu.RUnlock()
+	return bs, err
+}
+
+// FD implements platform.File.FD.
+func (d *dentryPlatformFile) FD() int {
+	d.handleMu.RLock()
+	fd := d.handle.fd
+	d.handleMu.RUnlock()
+	return int(fd)
+}
diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go
new file mode 100644
index 000000000..08c691c47
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/special_file.go
@@ -0,0 +1,159 @@
+// 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 gofer
+
+import (
+	"sync"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/safemem"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// specialFileFD implements vfs.FileDescriptionImpl for files other than
+// regular files, directories, and symlinks: pipes, sockets, etc. It is also
+// used for regular files when filesystemOptions.specialRegularFiles is in
+// effect. specialFileFD differs from regularFileFD by using per-FD handles
+// instead of shared per-dentry handles, and never buffering I/O.
+type specialFileFD struct {
+	fileDescription
+
+	// handle is immutable.
+	handle handle
+
+	// off is the file offset. off is protected by mu. (POSIX 2.9.7 only
+	// requires operations using the file offset to be atomic for regular files
+	// and symlinks; however, since specialFileFD may be used for regular
+	// files, we apply this atomicity unconditionally.)
+	mu  sync.Mutex
+	off int64
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *specialFileFD) Release() {
+	fd.handle.close(context.Background())
+	fs := fd.vfsfd.Mount().Filesystem().Impl().(*filesystem)
+	fs.syncMu.Lock()
+	delete(fs.specialFileFDs, fd)
+	fs.syncMu.Unlock()
+}
+
+// OnClose implements vfs.FileDescriptionImpl.OnClose.
+func (fd *specialFileFD) OnClose(ctx context.Context) error {
+	if !fd.vfsfd.IsWritable() {
+		return nil
+	}
+	return fd.handle.file.flush(ctx)
+}
+
+// PRead implements vfs.FileDescriptionImpl.PRead.
+func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	if opts.Flags != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	// Going through dst.CopyOutFrom() holds MM locks around file operations of
+	// unknown duration. For regularFileFD, doing so is necessary to support
+	// mmap due to lock ordering; MM locks precede dentry.dataMu. That doesn't
+	// hold here since specialFileFD doesn't client-cache data. Just buffer the
+	// read instead.
+	if d := fd.dentry(); d.fs.opts.interop != InteropModeShared {
+		d.touchAtime(ctx, fd.vfsfd.Mount())
+	}
+	buf := make([]byte, dst.NumBytes())
+	n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset))
+	if n == 0 {
+		return 0, err
+	}
+	if cp, cperr := dst.CopyOut(ctx, buf[:n]); cperr != nil {
+		return int64(cp), cperr
+	}
+	return int64(n), err
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (fd *specialFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	fd.mu.Lock()
+	n, err := fd.PRead(ctx, dst, fd.off, opts)
+	fd.off += n
+	fd.mu.Unlock()
+	return n, err
+}
+
+// PWrite implements vfs.FileDescriptionImpl.PWrite.
+func (fd *specialFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	if opts.Flags != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	// Do a buffered write. See rationale in PRead.
+	if d := fd.dentry(); d.fs.opts.interop != InteropModeShared {
+		d.touchCMtime(ctx)
+	}
+	buf := make([]byte, src.NumBytes())
+	// Don't do partial writes if we get a partial read from src.
+	if _, err := src.CopyIn(ctx, buf); err != nil {
+		return 0, err
+	}
+	n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset))
+	return int64(n), err
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (fd *specialFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	fd.mu.Lock()
+	n, err := fd.PWrite(ctx, src, fd.off, opts)
+	fd.off += n
+	fd.mu.Unlock()
+	return n, err
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+	fd.mu.Lock()
+	defer fd.mu.Unlock()
+	switch whence {
+	case linux.SEEK_SET:
+		// Use offset as given.
+	case linux.SEEK_CUR:
+		offset += fd.off
+	default:
+		// SEEK_END, SEEK_DATA, and SEEK_HOLE aren't supported since it's not
+		// clear that file size is even meaningful for these files.
+		return 0, syserror.EINVAL
+	}
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	fd.off = offset
+	return offset, nil
+}
+
+// Sync implements vfs.FileDescriptionImpl.Sync.
+func (fd *specialFileFD) Sync(ctx context.Context) error {
+	if !fd.vfsfd.IsWritable() {
+		return nil
+	}
+	return fd.handle.sync(ctx)
+}
diff --git a/pkg/sentry/fsimpl/gofer/symlink.go b/pkg/sentry/fsimpl/gofer/symlink.go
new file mode 100644
index 000000000..adf43be60
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/symlink.go
@@ -0,0 +1,47 @@
+// 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 gofer
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+func (d *dentry) isSymlink() bool {
+	return d.fileType() == linux.S_IFLNK
+}
+
+// Precondition: d.isSymlink().
+func (d *dentry) readlink(ctx context.Context, mnt *vfs.Mount) (string, error) {
+	if d.fs.opts.interop != InteropModeShared {
+		d.touchAtime(ctx, mnt)
+		d.dataMu.Lock()
+		if d.haveTarget {
+			target := d.target
+			d.dataMu.Unlock()
+			return target, nil
+		}
+	}
+	target, err := d.file.readlink(ctx)
+	if d.fs.opts.interop != InteropModeShared {
+		if err == nil {
+			d.haveTarget = true
+			d.target = target
+		}
+		d.dataMu.Unlock()
+	}
+	return target, err
+}
diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go
new file mode 100644
index 000000000..7598ec6a8
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/time.go
@@ -0,0 +1,75 @@
+// 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 gofer
+
+import (
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+func dentryTimestampFromP9(s, ns uint64) int64 {
+	return int64(s*1e9 + ns)
+}
+
+func dentryTimestampFromStatx(ts linux.StatxTimestamp) int64 {
+	return ts.Sec*1e9 + int64(ts.Nsec)
+}
+
+func statxTimestampFromDentry(ns int64) linux.StatxTimestamp {
+	return linux.StatxTimestamp{
+		Sec:  ns / 1e9,
+		Nsec: uint32(ns % 1e9),
+	}
+}
+
+func nowFromContext(ctx context.Context) (int64, bool) {
+	if clock := ktime.RealtimeClockFromContext(ctx); clock != nil {
+		return clock.Now().Nanoseconds(), true
+	}
+	return 0, false
+}
+
+// Preconditions: fs.interop != InteropModeShared.
+func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) {
+	if err := mnt.CheckBeginWrite(); err != nil {
+		return
+	}
+	now, ok := nowFromContext(ctx)
+	if !ok {
+		mnt.EndWrite()
+		return
+	}
+	d.metadataMu.Lock()
+	atomic.StoreInt64(&d.atime, now)
+	d.metadataMu.Unlock()
+	mnt.EndWrite()
+}
+
+// Preconditions: fs.interop != InteropModeShared. The caller has successfully
+// called vfs.Mount.CheckBeginWrite().
+func (d *dentry) touchCMtime(ctx context.Context) {
+	now, ok := nowFromContext(ctx)
+	if !ok {
+		return
+	}
+	d.metadataMu.Lock()
+	atomic.StoreInt64(&d.mtime, now)
+	atomic.StoreInt64(&d.ctime, now)
+	d.metadataMu.Unlock()
+}