Support VFS2 save/restore.

Inode number consistency checks are now skipped in save/restore tests for
reasons described in greatest detail in StatTest.StateDoesntChangeAfterRename.
They pass in VFS1 due to the bug described in new test case
SimpleStatTest.DifferentFilesHaveDifferentDeviceInodeNumberPairs.

Fixes #1663

PiperOrigin-RevId: 338776148

45 files changed, 1240 insertions, 425 deletions

diff --git a/pkg/sentry/control/state.go b/pkg/sentry/control/state.go
index 41feeffe3..d800f2c85 100644
--- a/pkg/sentry/control/state.go
+++ b/pkg/sentry/control/state.go
@@ -69,5 +69,5 @@ func (s *State) Save(o *SaveOpts, _ *struct{}) error {
 			s.Kernel.Kill(kernel.ExitStatus{})
 		},
 	}
-	return saveOpts.Save(s.Kernel, s.Watchdog)
+	return saveOpts.Save(s.Kernel.SupervisorContext(), s.Kernel, s.Watchdog)
 }
diff --git a/pkg/sentry/devices/tundev/tundev.go b/pkg/sentry/devices/tundev/tundev.go
index 655ea549b..ff5d49fbd 100644
--- a/pkg/sentry/devices/tundev/tundev.go
+++ b/pkg/sentry/devices/tundev/tundev.go
@@ -39,6 +39,8 @@ const (
 )
 
 // tunDevice implements vfs.Device for /dev/net/tun.
+//
+// +stateify savable
 type tunDevice struct{}
 
 // Open implements vfs.Device.Open.
@@ -53,6 +55,8 @@ func (tunDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, opt
 }
 
 // tunFD implements vfs.FileDescriptionImpl for /dev/net/tun.
+//
+// +stateify savable
 type tunFD struct {
 	vfsfd vfs.FileDescription
 	vfs.FileDescriptionDefaultImpl
diff --git a/pkg/sentry/fs/fsutil/host_file_mapper.go b/pkg/sentry/fs/fsutil/host_file_mapper.go
index 1390a9a7f..4468f5dd2 100644
--- a/pkg/sentry/fs/fsutil/host_file_mapper.go
+++ b/pkg/sentry/fs/fsutil/host_file_mapper.go
@@ -70,6 +70,13 @@ func (f *HostFileMapper) Init() {
 	f.mappings = make(map[uint64]mapping)
 }
 
+// IsInited returns true if f.Init() has been called. This is used when
+// restoring a checkpoint that contains a HostFileMapper that may or may not
+// have been initialized.
+func (f *HostFileMapper) IsInited() bool {
+	return f.refs != nil
+}
+
 // NewHostFileMapper returns an initialized HostFileMapper allocated on the
 // heap with no references or cached mappings.
 func NewHostFileMapper() *HostFileMapper {
diff --git a/pkg/sentry/fsimpl/devpts/line_discipline.go b/pkg/sentry/fsimpl/devpts/line_discipline.go
index e6b0e81cf..ae95fdd08 100644
--- a/pkg/sentry/fsimpl/devpts/line_discipline.go
+++ b/pkg/sentry/fsimpl/devpts/line_discipline.go
@@ -100,10 +100,10 @@ type lineDiscipline struct {
 	column int
 
 	// masterWaiter is used to wait on the master end of the TTY.
-	masterWaiter waiter.Queue `state:"zerovalue"`
+	masterWaiter waiter.Queue
 
 	// replicaWaiter is used to wait on the replica end of the TTY.
-	replicaWaiter waiter.Queue `state:"zerovalue"`
+	replicaWaiter waiter.Queue
 }
 
 func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline {
diff --git a/pkg/sentry/fsimpl/devtmpfs/BUILD b/pkg/sentry/fsimpl/devtmpfs/BUILD
index 01bbee5ad..e49a04c1b 100644
--- a/pkg/sentry/fsimpl/devtmpfs/BUILD
+++ b/pkg/sentry/fsimpl/devtmpfs/BUILD
@@ -4,7 +4,10 @@ licenses(["notice"])
 
 go_library(
     name = "devtmpfs",
-    srcs = ["devtmpfs.go"],
+    srcs = [
+        "devtmpfs.go",
+        "save_restore.go",
+    ],
     visibility = ["//pkg/sentry:internal"],
     deps = [
         "//pkg/abi/linux",
diff --git a/pkg/sentry/fsimpl/devtmpfs/save_restore.go b/pkg/sentry/fsimpl/devtmpfs/save_restore.go
new file mode 100644
index 000000000..28832d850
--- /dev/null
+++ b/pkg/sentry/fsimpl/devtmpfs/save_restore.go
@@ -0,0 +1,23 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devtmpfs
+
+// afterLoad is invoked by stateify.
+func (fst *FilesystemType) afterLoad() {
+	if fst.fs != nil {
+		// Ensure that we don't create another filesystem.
+		fst.initOnce.Do(func() {})
+	}
+}
diff --git a/pkg/sentry/fsimpl/eventfd/eventfd.go b/pkg/sentry/fsimpl/eventfd/eventfd.go
index 1c27ad700..5b29f2358 100644
--- a/pkg/sentry/fsimpl/eventfd/eventfd.go
+++ b/pkg/sentry/fsimpl/eventfd/eventfd.go
@@ -43,7 +43,7 @@ type EventFileDescription struct {
 
 	// queue is used to notify interested parties when the event object
 	// becomes readable or writable.
-	queue waiter.Queue `state:"zerovalue"`
+	queue waiter.Queue
 
 	// mu protects the fields below.
 	mu sync.Mutex `state:"nosave"`
diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD
index ad0afc41b..e3a090d95 100644
--- a/pkg/sentry/fsimpl/gofer/BUILD
+++ b/pkg/sentry/fsimpl/gofer/BUILD
@@ -38,6 +38,7 @@ go_library(
         "host_named_pipe.go",
         "p9file.go",
         "regular_file.go",
+        "save_restore.go",
         "socket.go",
         "special_file.go",
         "symlink.go",
@@ -70,6 +71,7 @@ go_library(
         "//pkg/sentry/socket/unix/transport",
         "//pkg/sentry/usage",
         "//pkg/sentry/vfs",
+        "//pkg/sync",
         "//pkg/syserr",
         "//pkg/syserror",
         "//pkg/unet",
diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go
index 18c884b59..c3af30f49 100644
--- a/pkg/sentry/fsimpl/gofer/directory.go
+++ b/pkg/sentry/fsimpl/gofer/directory.go
@@ -16,7 +16,6 @@ package gofer
 
 import (
 	"fmt"
-	"sync"
 	"sync/atomic"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
@@ -26,6 +25,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/kernel/pipe"
 	"gvisor.dev/gvisor/pkg/sentry/socket/unix/transport"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/usermem"
 )
@@ -92,7 +92,7 @@ func (d *dentry) createSyntheticChildLocked(opts *createSyntheticOpts) {
 	child := &dentry{
 		refs:      1, // held by d
 		fs:        d.fs,
-		ino:       d.fs.nextSyntheticIno(),
+		ino:       d.fs.nextIno(),
 		mode:      uint32(opts.mode),
 		uid:       uint32(opts.kuid),
 		gid:       uint32(opts.kgid),
@@ -235,7 +235,7 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) {
 				}
 				dirent := vfs.Dirent{
 					Name:    p9d.Name,
-					Ino:     uint64(inoFromPath(p9d.QID.Path)),
+					Ino:     d.fs.inoFromQIDPath(p9d.QID.Path),
 					NextOff: int64(len(dirents) + 1),
 				}
 				// p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or
diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go
index 94d96261b..baecb88c4 100644
--- a/pkg/sentry/fsimpl/gofer/filesystem.go
+++ b/pkg/sentry/fsimpl/gofer/filesystem.go
@@ -35,7 +35,7 @@ import (
 
 // Sync implements vfs.FilesystemImpl.Sync.
 func (fs *filesystem) Sync(ctx context.Context) error {
-	// Snapshot current syncable dentries and special files.
+	// Snapshot current syncable dentries and special file FDs.
 	fs.syncMu.Lock()
 	ds := make([]*dentry, 0, len(fs.syncableDentries))
 	for d := range fs.syncableDentries {
@@ -53,22 +53,28 @@ func (fs *filesystem) Sync(ctx context.Context) error {
 	// regardless.
 	var retErr error
 
-	// Sync regular files.
+	// Sync syncable dentries.
 	for _, d := range ds {
-		err := d.syncCachedFile(ctx)
+		err := d.syncCachedFile(ctx, true /* forFilesystemSync */)
 		d.DecRef(ctx)
-		if err != nil && retErr == nil {
-			retErr = err
+		if err != nil {
+			ctx.Infof("gofer.filesystem.Sync: dentry.syncCachedFile failed: %v", err)
+			if 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 {
-		err := sffd.Sync(ctx)
+		err := sffd.sync(ctx, true /* forFilesystemSync */)
 		sffd.vfsfd.DecRef(ctx)
-		if err != nil && retErr == nil {
-			retErr = err
+		if err != nil {
+			ctx.Infof("gofer.filesystem.Sync: specialFileFD.sync failed: %v", err)
+			if retErr == nil {
+				retErr = err
+			}
 		}
 	}
 
@@ -229,7 +235,7 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir
 		return nil, err
 	}
 	if child != nil {
-		if !file.isNil() && inoFromPath(qid.Path) == child.ino {
+		if !file.isNil() && qid.Path == child.qidPath {
 			// The file at this path hasn't changed. Just update cached metadata.
 			file.close(ctx)
 			child.updateFromP9AttrsLocked(attrMask, &attr)
@@ -1512,7 +1518,6 @@ func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath
 			d.IncRef()
 			return &endpoint{
 				dentry: d,
-				file:   d.file.file,
 				path:   opts.Addr,
 			}, nil
 		}
@@ -1591,7 +1596,3 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe
 	defer fs.renameMu.RUnlock()
 	return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b)
 }
-
-func (fs *filesystem) nextSyntheticIno() inodeNumber {
-	return inodeNumber(atomic.AddUint64(&fs.syntheticSeq, 1) | syntheticInoMask)
-}
diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go
index 8e179539c..f7c94cce1 100644
--- a/pkg/sentry/fsimpl/gofer/gofer.go
+++ b/pkg/sentry/fsimpl/gofer/gofer.go
@@ -26,6 +26,9 @@
 //             *** "memmap.Mappable locks taken by Translate" below this point
 //             dentry.handleMu
 //               dentry.dataMu
+//           filesystem.inoMu
+//   specialFileFD.mu
+//     specialFileFD.bufMu
 //
 // Locking dentry.dirMu in multiple dentries requires that either ancestor
 // dentries are locked before descendant dentries, or that filesystem.renameMu
@@ -36,7 +39,6 @@ import (
 	"fmt"
 	"strconv"
 	"strings"
-	"sync"
 	"sync/atomic"
 	"syscall"
 
@@ -53,6 +55,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
 	"gvisor.dev/gvisor/pkg/sentry/socket/unix/transport"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/unet"
 	"gvisor.dev/gvisor/pkg/usermem"
@@ -81,7 +84,7 @@ type filesystem struct {
 	iopts InternalFilesystemOptions
 
 	// client is the client used by this filesystem. client is immutable.
-	client *p9.Client `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
+	client *p9.Client `state:"nosave"`
 
 	// clock is a realtime clock used to set timestamps in file operations.
 	clock ktime.Clock
@@ -89,6 +92,9 @@ type filesystem struct {
 	// devMinor is the filesystem's minor device number. devMinor is immutable.
 	devMinor uint32
 
+	// root is the root dentry. root is immutable.
+	root *dentry
+
 	// renameMu serves two purposes:
 	//
 	// - It synchronizes path resolution with renaming initiated by this
@@ -108,34 +114,26 @@ type filesystem struct {
 	cachedDentries    dentryList
 	cachedDentriesLen uint64
 
-	// syncableDentries contains all dentries in this filesystem for which
-	// !dentry.file.isNil(). specialFileFDs contains all open specialFileFDs.
-	// These fields are protected by syncMu.
+	// syncableDentries contains all non-synthetic dentries. specialFileFDs
+	// contains all open specialFileFDs. These fields are protected by syncMu.
 	syncMu           sync.Mutex `state:"nosave"`
 	syncableDentries map[*dentry]struct{}
 	specialFileFDs   map[*specialFileFD]struct{}
 
-	// syntheticSeq stores a counter to used to generate unique inodeNumber for
-	// synthetic dentries.
-	syntheticSeq uint64
-}
-
-// inodeNumber represents inode number reported in Dirent.Ino. For regular
-// dentries, it comes from QID.Path from the 9P server. Synthetic dentries
-// have have their inodeNumber generated sequentially, with the MSB reserved to
-// prevent conflicts with regular dentries.
-//
-// +stateify savable
-type inodeNumber uint64
+	// inoByQIDPath maps previously-observed QID.Paths to inode numbers
+	// assigned to those paths. inoByQIDPath is not preserved across
+	// checkpoint/restore because QIDs may be reused between different gofer
+	// processes, so QIDs may be repeated for different files across
+	// checkpoint/restore. inoByQIDPath is protected by inoMu.
+	inoMu        sync.Mutex        `state:"nosave"`
+	inoByQIDPath map[uint64]uint64 `state:"nosave"`
 
-// Reserve MSB for synthetic mounts.
-const syntheticInoMask = uint64(1) << 63
+	// lastIno is the last inode number assigned to a file. lastIno is accessed
+	// using atomic memory operations.
+	lastIno uint64
 
-func inoFromPath(path uint64) inodeNumber {
-	if path&syntheticInoMask != 0 {
-		log.Warningf("Dropping MSB from ino, collision is possible. Original: %d, new: %d", path, path&^syntheticInoMask)
-	}
-	return inodeNumber(path &^ syntheticInoMask)
+	// savedDentryRW records open read/write handles during save/restore.
+	savedDentryRW map[*dentry]savedDentryRW
 }
 
 // +stateify savable
@@ -247,6 +245,10 @@ const (
 //
 // +stateify savable
 type InternalFilesystemOptions struct {
+	// If UniqueID is non-empty, it is an opaque string used to reassociate the
+	// filesystem with a new server FD during restoration from checkpoint.
+	UniqueID string
+
 	// If LeakConnection is true, do not close the connection to the server
 	// when the Filesystem is released. This is necessary for deployments in
 	// which servers can handle only a single client and report failure if that
@@ -286,46 +288,11 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 	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)
+	fd, err := getFDFromMountOptionsMap(ctx, mopts)
 	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
+		return nil, nil, err
 	}
-	fsopts.fd = rfd
+	fsopts.fd = fd
 
 	// Get the attach name.
 	fsopts.aname = "/"
@@ -441,57 +408,44 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 	}
 	// If !ok, iopts being the zero value is correct.
 
-	// Establish a connection with the server.
-	conn, err := unet.NewSocket(fsopts.fd)
+	// Construct the filesystem object.
+	devMinor, err := vfsObj.GetAnonBlockDevMinor()
 	if err != nil {
 		return nil, nil, err
 	}
+	fs := &filesystem{
+		mfp:              mfp,
+		opts:             fsopts,
+		iopts:            iopts,
+		clock:            ktime.RealtimeClockFromContext(ctx),
+		devMinor:         devMinor,
+		syncableDentries: make(map[*dentry]struct{}),
+		specialFileFDs:   make(map[*specialFileFD]struct{}),
+		inoByQIDPath:     make(map[uint64]uint64),
+	}
+	fs.vfsfs.Init(vfsObj, &fstype, fs)
 
-	// 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()
+	// Connect to the server.
+	if err := fs.dial(ctx); err != nil {
 		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)
+	attached, err := fs.client.Attach(fsopts.aname)
 	ctx.UninterruptibleSleepFinish(false)
 	if err != nil {
-		client.Close()
+		fs.vfsfs.DecRef(ctx)
 		return nil, nil, err
 	}
 	attachFile := p9file{attached}
 	qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask())
 	if err != nil {
 		attachFile.close(ctx)
-		client.Close()
+		fs.vfsfs.DecRef(ctx)
 		return nil, nil, err
 	}
 
-	// Construct the filesystem object.
-	devMinor, err := vfsObj.GetAnonBlockDevMinor()
-	if err != nil {
-		attachFile.close(ctx)
-		client.Close()
-		return nil, nil, err
-	}
-	fs := &filesystem{
-		mfp:              mfp,
-		opts:             fsopts,
-		iopts:            iopts,
-		client:           client,
-		clock:            ktime.RealtimeClockFromContext(ctx),
-		devMinor:         devMinor,
-		syncableDentries: make(map[*dentry]struct{}),
-		specialFileFDs:   make(map[*specialFileFD]struct{}),
-	}
-	fs.vfsfs.Init(vfsObj, &fstype, fs)
-
 	// Construct the root dentry.
 	root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr)
 	if err != nil {
@@ -504,10 +458,72 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 	// being "cached" and subsequently evicted. Its resources will still be
 	// cleaned up by fs.Release().
 	root.refs = 2
+	fs.root = root
 
 	return &fs.vfsfs, &root.vfsd, nil
 }
 
+func getFDFromMountOptionsMap(ctx context.Context, mopts map[string]string) (int, error) {
+	// Check that the transport is "fd".
+	trans, ok := mopts["trans"]
+	if !ok || trans != "fd" {
+		ctx.Warningf("gofer.getFDFromMountOptionsMap: transport must be specified as 'trans=fd'")
+		return -1, syserror.EINVAL
+	}
+	delete(mopts, "trans")
+
+	// Check that read and write FDs are provided and identical.
+	rfdstr, ok := mopts["rfdno"]
+	if !ok {
+		ctx.Warningf("gofer.getFDFromMountOptionsMap: read FD must be specified as 'rfdno=<file descriptor>'")
+		return -1, syserror.EINVAL
+	}
+	delete(mopts, "rfdno")
+	rfd, err := strconv.Atoi(rfdstr)
+	if err != nil {
+		ctx.Warningf("gofer.getFDFromMountOptionsMap: invalid read FD: rfdno=%s", rfdstr)
+		return -1, syserror.EINVAL
+	}
+	wfdstr, ok := mopts["wfdno"]
+	if !ok {
+		ctx.Warningf("gofer.getFDFromMountOptionsMap: write FD must be specified as 'wfdno=<file descriptor>'")
+		return -1, syserror.EINVAL
+	}
+	delete(mopts, "wfdno")
+	wfd, err := strconv.Atoi(wfdstr)
+	if err != nil {
+		ctx.Warningf("gofer.getFDFromMountOptionsMap: invalid write FD: wfdno=%s", wfdstr)
+		return -1, syserror.EINVAL
+	}
+	if rfd != wfd {
+		ctx.Warningf("gofer.getFDFromMountOptionsMap: read FD (%d) and write FD (%d) must be equal", rfd, wfd)
+		return -1, syserror.EINVAL
+	}
+	return rfd, nil
+}
+
+// Preconditions: fs.client == nil.
+func (fs *filesystem) dial(ctx context.Context) error {
+	// Establish a connection with the server.
+	conn, err := unet.NewSocket(fs.opts.fd)
+	if err != nil {
+		return err
+	}
+
+	// Perform version negotiation with the server.
+	ctx.UninterruptibleSleepStart(false)
+	client, err := p9.NewClient(conn, fs.opts.msize, fs.opts.version)
+	ctx.UninterruptibleSleepFinish(false)
+	if err != nil {
+		conn.Close()
+		return err
+	}
+	// Ownership of conn has been transferred to client.
+
+	fs.client = client
+	return nil
+}
+
 // Release implements vfs.FilesystemImpl.Release.
 func (fs *filesystem) Release(ctx context.Context) {
 	mf := fs.mfp.MemoryFile()
@@ -574,12 +590,15 @@ type dentry struct {
 	// filesystem.renameMu.
 	name string
 
+	// qidPath is the p9.QID.Path for this file. qidPath is immutable.
+	qidPath uint64
+
 	// file is the unopened p9.File that backs this dentry. file is immutable.
 	//
 	// If file.isNil(), this dentry represents a synthetic file, i.e. a file
 	// that does not exist on the remote filesystem. As of this writing, the
 	// only files that can be synthetic are sockets, pipes, and directories.
-	file p9file `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
+	file p9file `state:"nosave"`
 
 	// If deleted is non-zero, the file represented by this dentry has been
 	// deleted. deleted is accessed using atomic memory operations.
@@ -623,12 +642,12 @@ type dentry struct {
 	// To mutate:
 	//   - Lock metadataMu and use atomic operations to update because we might
 	//     have atomic readers that don't hold the lock.
-	metadataMu sync.Mutex  `state:"nosave"`
-	ino        inodeNumber // 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
+	metadataMu sync.Mutex `state:"nosave"`
+	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
@@ -679,9 +698,9 @@ type dentry struct {
 	// (isNil() == false), it may be mutated with handleMu locked, but cannot
 	// be closed until the dentry is destroyed.
 	handleMu  sync.RWMutex `state:"nosave"`
-	readFile  p9file       `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
-	writeFile p9file       `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
-	hostFD    int32
+	readFile  p9file       `state:"nosave"`
+	writeFile p9file       `state:"nosave"`
+	hostFD    int32        `state:"nosave"`
 
 	dataMu sync.RWMutex `state:"nosave"`
 
@@ -758,8 +777,9 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma
 
 	d := &dentry{
 		fs:        fs,
+		qidPath:   qid.Path,
 		file:      file,
-		ino:       inoFromPath(qid.Path),
+		ino:       fs.inoFromQIDPath(qid.Path),
 		mode:      uint32(attr.Mode),
 		uid:       uint32(fs.opts.dfltuid),
 		gid:       uint32(fs.opts.dfltgid),
@@ -802,6 +822,21 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma
 	return d, nil
 }
 
+func (fs *filesystem) inoFromQIDPath(qidPath uint64) uint64 {
+	fs.inoMu.Lock()
+	defer fs.inoMu.Unlock()
+	if ino, ok := fs.inoByQIDPath[qidPath]; ok {
+		return ino
+	}
+	ino := fs.nextIno()
+	fs.inoByQIDPath[qidPath] = ino
+	return ino
+}
+
+func (fs *filesystem) nextIno() uint64 {
+	return atomic.AddUint64(&fs.lastIno, 1)
+}
+
 func (d *dentry) isSynthetic() bool {
 	return d.file.isNil()
 }
@@ -853,7 +888,7 @@ func (d *dentry) updateFromP9AttrsLocked(mask p9.AttrMask, attr *p9.Attr) {
 	}
 }
 
-// Preconditions: !d.isSynthetic()
+// Preconditions: !d.isSynthetic().
 func (d *dentry) updateFromGetattr(ctx context.Context) error {
 	// Use d.readFile or d.writeFile, which represent 9P fids that have been
 	// opened, in preference to d.file, which represents a 9P fid that has not.
@@ -1269,33 +1304,40 @@ func (d *dentry) checkCachingLocked(ctx context.Context) {
 	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.
-		if atomic.LoadInt64(&victim.refs) == 0 {
-			if victim.parent != nil {
-				victim.parent.dirMu.Lock()
-				if !victim.vfsd.IsDead() {
-					// Note that victim can't be a mount point (in any mount
-					// namespace), since VFS holds references on mount points.
-					d.fs.vfsfs.VirtualFilesystem().InvalidateDentry(ctx, &victim.vfsd)
-					delete(victim.parent.children, victim.name)
-					// We're only deleting the dentry, not the file it
-					// represents, so we don't need to update
-					// victimParent.dirents etc.
-				}
-				victim.parent.dirMu.Unlock()
-			}
-			victim.destroyLocked(ctx)
-		}
+		d.fs.evictCachedDentryLocked(ctx)
 		// Whether or not victim was destroyed, we brought fs.cachedDentriesLen
 		// back down to fs.opts.maxCachedDentries, so we don't loop.
 	}
 }
 
+// Preconditions:
+// * fs.renameMu must be locked for writing; it may be temporarily unlocked.
+// * fs.cachedDentriesLen != 0.
+func (fs *filesystem) evictCachedDentryLocked(ctx context.Context) {
+	victim := fs.cachedDentries.Back()
+	fs.cachedDentries.Remove(victim)
+	fs.cachedDentriesLen--
+	victim.cached = false
+	// victim.refs may have become non-zero from an earlier path resolution
+	// since it was inserted into fs.cachedDentries.
+	if atomic.LoadInt64(&victim.refs) == 0 {
+		if victim.parent != nil {
+			victim.parent.dirMu.Lock()
+			if !victim.vfsd.IsDead() {
+				// Note that victim can't be a mount point (in any mount
+				// namespace), since VFS holds references on mount points.
+				fs.vfsfs.VirtualFilesystem().InvalidateDentry(ctx, &victim.vfsd)
+				delete(victim.parent.children, victim.name)
+				// We're only deleting the dentry, not the file it
+				// represents, so we don't need to update
+				// victimParent.dirents etc.
+			}
+			victim.parent.dirMu.Unlock()
+		}
+		victim.destroyLocked(ctx)
+	}
+}
+
 // destroyLocked destroys the dentry.
 //
 // Preconditions:
@@ -1623,6 +1665,33 @@ func (d *dentry) syncRemoteFileLocked(ctx context.Context) error {
 	return nil
 }
 
+func (d *dentry) syncCachedFile(ctx context.Context, forFilesystemSync bool) error {
+	d.handleMu.RLock()
+	defer d.handleMu.RUnlock()
+	h := d.writeHandleLocked()
+	if h.isOpen() {
+		// Write back dirty pages to the remote file.
+		d.dataMu.Lock()
+		err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), h.writeFromBlocksAt)
+		d.dataMu.Unlock()
+		if err != nil {
+			return err
+		}
+	}
+	if err := d.syncRemoteFileLocked(ctx); err != nil {
+		if !forFilesystemSync {
+			return err
+		}
+		// Only return err if we can reasonably have expected sync to succeed
+		// (d is a regular file and was opened for writing).
+		if d.isRegularFile() && h.isOpen() {
+			return err
+		}
+		ctx.Debugf("gofer.dentry.syncCachedFile: syncing non-writable or non-regular-file dentry failed: %v", err)
+	}
+	return nil
+}
+
 // incLinks increments link count.
 func (d *dentry) incLinks() {
 	if atomic.LoadUint32(&d.nlink) == 0 {
@@ -1650,7 +1719,7 @@ type fileDescription struct {
 	vfs.FileDescriptionDefaultImpl
 	vfs.LockFD
 
-	lockLogging sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
+	lockLogging sync.Once `state:"nosave"`
 }
 
 func (fd *fileDescription) filesystem() *filesystem {
diff --git a/pkg/sentry/fsimpl/gofer/gofer_test.go b/pkg/sentry/fsimpl/gofer/gofer_test.go
index bfe75dfe4..76f08e252 100644
--- a/pkg/sentry/fsimpl/gofer/gofer_test.go
+++ b/pkg/sentry/fsimpl/gofer/gofer_test.go
@@ -26,12 +26,13 @@ import (
 func TestDestroyIdempotent(t *testing.T) {
 	ctx := contexttest.Context(t)
 	fs := filesystem{
-		mfp:              pgalloc.MemoryFileProviderFromContext(ctx),
-		syncableDentries: make(map[*dentry]struct{}),
+		mfp: pgalloc.MemoryFileProviderFromContext(ctx),
 		opts: filesystemOptions{
 			// Test relies on no dentry being held in the cache.
 			maxCachedDentries: 0,
 		},
+		syncableDentries: make(map[*dentry]struct{}),
+		inoByQIDPath:     make(map[uint64]uint64),
 	}
 
 	attr := &p9.Attr{
diff --git a/pkg/sentry/fsimpl/gofer/host_named_pipe.go b/pkg/sentry/fsimpl/gofer/host_named_pipe.go
index 7294de7d6..c7bf10007 100644
--- a/pkg/sentry/fsimpl/gofer/host_named_pipe.go
+++ b/pkg/sentry/fsimpl/gofer/host_named_pipe.go
@@ -51,8 +51,24 @@ func blockUntilNonblockingPipeHasWriter(ctx context.Context, fd int32) error {
 		if ok {
 			return nil
 		}
-		if err := sleepBetweenNamedPipeOpenChecks(ctx); err != nil {
-			return err
+		if sleepErr := sleepBetweenNamedPipeOpenChecks(ctx); sleepErr != nil {
+			// Another application thread may have opened this pipe for
+			// writing, succeeded because we previously opened the pipe for
+			// reading, and subsequently interrupted us for checkpointing (e.g.
+			// this occurs in mknod tests under cooperative save/restore). In
+			// this case, our open has to succeed for the checkpoint to include
+			// a readable FD for the pipe, which is in turn necessary to
+			// restore the other thread's writable FD for the same pipe
+			// (otherwise it will get ENXIO). So we have to check
+			// nonblockingPipeHasWriter() once last time.
+			ok, err := nonblockingPipeHasWriter(fd)
+			if err != nil {
+				return err
+			}
+			if ok {
+				return nil
+			}
+			return sleepErr
 		}
 	}
 }
diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go
index f8b19bae7..dc8a890cb 100644
--- a/pkg/sentry/fsimpl/gofer/regular_file.go
+++ b/pkg/sentry/fsimpl/gofer/regular_file.go
@@ -18,7 +18,6 @@ import (
 	"fmt"
 	"io"
 	"math"
-	"sync"
 	"sync/atomic"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
@@ -31,6 +30,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
 	"gvisor.dev/gvisor/pkg/sentry/usage"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/usermem"
 )
@@ -624,23 +624,7 @@ func regularFileSeekLocked(ctx context.Context, d *dentry, fdOffset, offset int6
 
 // Sync implements vfs.FileDescriptionImpl.Sync.
 func (fd *regularFileFD) Sync(ctx context.Context) error {
-	return fd.dentry().syncCachedFile(ctx)
-}
-
-func (d *dentry) syncCachedFile(ctx context.Context) error {
-	d.handleMu.RLock()
-	defer d.handleMu.RUnlock()
-
-	if h := d.writeHandleLocked(); h.isOpen() {
-		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(), h.writeFromBlocksAt)
-		d.dataMu.Unlock()
-		if err != nil {
-			return err
-		}
-	}
-	return d.syncRemoteFileLocked(ctx)
+	return fd.dentry().syncCachedFile(ctx, false /* lowSyncExpectations */)
 }
 
 // ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap.
@@ -913,7 +897,7 @@ type dentryPlatformFile struct {
 	hostFileMapper fsutil.HostFileMapper
 
 	// hostFileMapperInitOnce is used to lazily initialize hostFileMapper.
-	hostFileMapperInitOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
+	hostFileMapperInitOnce sync.Once `state:"nosave"`
 }
 
 // IncRef implements memmap.File.IncRef.
diff --git a/pkg/sentry/fsimpl/gofer/save_restore.go b/pkg/sentry/fsimpl/gofer/save_restore.go
new file mode 100644
index 000000000..e995619a6
--- /dev/null
+++ b/pkg/sentry/fsimpl/gofer/save_restore.go
@@ -0,0 +1,327 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"fmt"
+	"io"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fdnotifier"
+	"gvisor.dev/gvisor/pkg/p9"
+	"gvisor.dev/gvisor/pkg/safemem"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+type saveRestoreContextID int
+
+const (
+	// CtxRestoreServerFDMap is a Context.Value key for a map[string]int
+	// mapping filesystem unique IDs (cf. InternalFilesystemOptions.UniqueID)
+	// to host FDs.
+	CtxRestoreServerFDMap saveRestoreContextID = iota
+)
+
+// +stateify savable
+type savedDentryRW struct {
+	read  bool
+	write bool
+}
+
+// PreprareSave implements vfs.FilesystemImplSaveRestoreExtension.PrepareSave.
+func (fs *filesystem) PrepareSave(ctx context.Context) error {
+	if len(fs.iopts.UniqueID) == 0 {
+		return fmt.Errorf("gofer.filesystem with no UniqueID cannot be saved")
+	}
+
+	// Purge cached dentries, which may not be reopenable after restore due to
+	// permission changes.
+	fs.renameMu.Lock()
+	for fs.cachedDentriesLen != 0 {
+		fs.evictCachedDentryLocked(ctx)
+	}
+	fs.renameMu.Unlock()
+
+	// Buffer pipe data so that it's available for reading after restore. (This
+	// is a legacy VFS1 feature.)
+	fs.syncMu.Lock()
+	for sffd := range fs.specialFileFDs {
+		if sffd.dentry().fileType() == linux.S_IFIFO && sffd.vfsfd.IsReadable() {
+			if err := sffd.savePipeData(ctx); err != nil {
+				fs.syncMu.Unlock()
+				return err
+			}
+		}
+	}
+	fs.syncMu.Unlock()
+
+	// Flush local state to the remote filesystem.
+	if err := fs.Sync(ctx); err != nil {
+		return err
+	}
+
+	fs.savedDentryRW = make(map[*dentry]savedDentryRW)
+	return fs.root.prepareSaveRecursive(ctx)
+}
+
+// Preconditions:
+// * fd represents a pipe.
+// * fd is readable.
+func (fd *specialFileFD) savePipeData(ctx context.Context) error {
+	fd.bufMu.Lock()
+	defer fd.bufMu.Unlock()
+	var buf [usermem.PageSize]byte
+	for {
+		n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:])), ^uint64(0))
+		if n != 0 {
+			fd.buf = append(fd.buf, buf[:n]...)
+		}
+		if err != nil {
+			if err == io.EOF || err == syserror.EAGAIN {
+				break
+			}
+			return err
+		}
+	}
+	if len(fd.buf) != 0 {
+		atomic.StoreUint32(&fd.haveBuf, 1)
+	}
+	return nil
+}
+
+func (d *dentry) prepareSaveRecursive(ctx context.Context) error {
+	if d.isRegularFile() && !d.cachedMetadataAuthoritative() {
+		// Get updated metadata for d in case we need to perform metadata
+		// validation during restore.
+		if err := d.updateFromGetattr(ctx); err != nil {
+			return err
+		}
+	}
+	if !d.readFile.isNil() || !d.writeFile.isNil() {
+		d.fs.savedDentryRW[d] = savedDentryRW{
+			read:  !d.readFile.isNil(),
+			write: !d.writeFile.isNil(),
+		}
+	}
+	d.dirMu.Lock()
+	defer d.dirMu.Unlock()
+	for _, child := range d.children {
+		if child != nil {
+			if err := child.prepareSaveRecursive(ctx); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// beforeSave is invoked by stateify.
+func (d *dentry) beforeSave() {
+	if d.vfsd.IsDead() {
+		panic(fmt.Sprintf("gofer.dentry(%q).beforeSave: deleted and invalidated dentries can't be restored", genericDebugPathname(d)))
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (d *dentry) afterLoad() {
+	d.hostFD = -1
+}
+
+// afterLoad is invoked by stateify.
+func (d *dentryPlatformFile) afterLoad() {
+	if d.hostFileMapper.IsInited() {
+		// Ensure that we don't call d.hostFileMapper.Init() again.
+		d.hostFileMapperInitOnce.Do(func() {})
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (fd *specialFileFD) afterLoad() {
+	fd.handle.fd = -1
+}
+
+// CompleteRestore implements
+// vfs.FilesystemImplSaveRestoreExtension.CompleteRestore.
+func (fs *filesystem) CompleteRestore(ctx context.Context, opts vfs.CompleteRestoreOptions) error {
+	fdmapv := ctx.Value(CtxRestoreServerFDMap)
+	if fdmapv == nil {
+		return fmt.Errorf("no server FD map available")
+	}
+	fdmap := fdmapv.(map[string]int)
+	fd, ok := fdmap[fs.iopts.UniqueID]
+	if !ok {
+		return fmt.Errorf("no server FD available for filesystem with unique ID %q", fs.iopts.UniqueID)
+	}
+	fs.opts.fd = fd
+	if err := fs.dial(ctx); err != nil {
+		return err
+	}
+	fs.inoByQIDPath = make(map[uint64]uint64)
+
+	// Restore the filesystem root.
+	ctx.UninterruptibleSleepStart(false)
+	attached, err := fs.client.Attach(fs.opts.aname)
+	ctx.UninterruptibleSleepFinish(false)
+	if err != nil {
+		return err
+	}
+	attachFile := p9file{attached}
+	qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask())
+	if err != nil {
+		return err
+	}
+	if err := fs.root.restoreFile(ctx, attachFile, qid, attrMask, &attr, &opts); err != nil {
+		return err
+	}
+
+	// Restore remaining dentries.
+	if err := fs.root.restoreDescendantsRecursive(ctx, &opts); err != nil {
+		return err
+	}
+
+	// Re-open handles for specialFileFDs. Unlike the initial open
+	// (dentry.openSpecialFile()), pipes are always opened without blocking;
+	// non-readable pipe FDs are opened last to ensure that they don't get
+	// ENXIO if another specialFileFD represents the read end of the same pipe.
+	// This is consistent with VFS1.
+	haveWriteOnlyPipes := false
+	for fd := range fs.specialFileFDs {
+		if fd.dentry().fileType() == linux.S_IFIFO && !fd.vfsfd.IsReadable() {
+			haveWriteOnlyPipes = true
+			continue
+		}
+		if err := fd.completeRestore(ctx); err != nil {
+			return err
+		}
+	}
+	if haveWriteOnlyPipes {
+		for fd := range fs.specialFileFDs {
+			if fd.dentry().fileType() == linux.S_IFIFO && !fd.vfsfd.IsReadable() {
+				if err := fd.completeRestore(ctx); err != nil {
+					return err
+				}
+			}
+		}
+	}
+
+	// Discard state only required during restore.
+	fs.savedDentryRW = nil
+
+	return nil
+}
+
+func (d *dentry) restoreFile(ctx context.Context, file p9file, qid p9.QID, attrMask p9.AttrMask, attr *p9.Attr, opts *vfs.CompleteRestoreOptions) error {
+	d.file = file
+
+	// Gofers do not preserve QID across checkpoint/restore, so:
+	//
+	// - We must assume that the remote filesystem did not change in a way that
+	// would invalidate dentries, since we can't revalidate dentries by
+	// checking QIDs.
+	//
+	// - We need to associate the new QID.Path with the existing d.ino.
+	d.qidPath = qid.Path
+	d.fs.inoMu.Lock()
+	d.fs.inoByQIDPath[qid.Path] = d.ino
+	d.fs.inoMu.Unlock()
+
+	// Check metadata stability before updating metadata.
+	d.metadataMu.Lock()
+	defer d.metadataMu.Unlock()
+	if d.isRegularFile() {
+		if opts.ValidateFileSizes {
+			if !attrMask.Size {
+				return fmt.Errorf("gofer.dentry(%q).restoreFile: file size validation failed: file size not available", genericDebugPathname(d))
+			}
+			if d.size != attr.Size {
+				return fmt.Errorf("gofer.dentry(%q).restoreFile: file size validation failed: size changed from %d to %d", genericDebugPathname(d), d.size, attr.Size)
+			}
+		}
+		if opts.ValidateFileModificationTimestamps {
+			if !attrMask.MTime {
+				return fmt.Errorf("gofer.dentry(%q).restoreFile: mtime validation failed: mtime not available", genericDebugPathname(d))
+			}
+			if want := dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds); d.mtime != want {
+				return fmt.Errorf("gofer.dentry(%q).restoreFile: mtime validation failed: mtime changed from %+v to %+v", genericDebugPathname(d), linux.NsecToStatxTimestamp(d.mtime), linux.NsecToStatxTimestamp(want))
+			}
+		}
+	}
+	if !d.cachedMetadataAuthoritative() {
+		d.updateFromP9AttrsLocked(attrMask, attr)
+	}
+
+	if rw, ok := d.fs.savedDentryRW[d]; ok {
+		if err := d.ensureSharedHandle(ctx, rw.read, rw.write, false /* trunc */); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// Preconditions: d is not synthetic.
+func (d *dentry) restoreDescendantsRecursive(ctx context.Context, opts *vfs.CompleteRestoreOptions) error {
+	for _, child := range d.children {
+		if child == nil {
+			continue
+		}
+		if _, ok := d.fs.syncableDentries[child]; !ok {
+			// child is synthetic.
+			continue
+		}
+		if err := child.restoreRecursive(ctx, opts); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// Preconditions: d is not synthetic (but note that since this function
+// restores d.file, d.file.isNil() is always true at this point, so this can
+// only be detected by checking filesystem.syncableDentries). d.parent has been
+// restored.
+func (d *dentry) restoreRecursive(ctx context.Context, opts *vfs.CompleteRestoreOptions) error {
+	qid, file, attrMask, attr, err := d.parent.file.walkGetAttrOne(ctx, d.name)
+	if err != nil {
+		return err
+	}
+	if err := d.restoreFile(ctx, file, qid, attrMask, &attr, opts); err != nil {
+		return err
+	}
+	return d.restoreDescendantsRecursive(ctx, opts)
+}
+
+func (fd *specialFileFD) completeRestore(ctx context.Context) error {
+	d := fd.dentry()
+	h, err := openHandle(ctx, d.file, fd.vfsfd.IsReadable(), fd.vfsfd.IsWritable(), false /* trunc */)
+	if err != nil {
+		return err
+	}
+	fd.handle = h
+
+	ftype := d.fileType()
+	fd.haveQueue = (ftype == linux.S_IFIFO || ftype == linux.S_IFSOCK) && fd.handle.fd >= 0
+	if fd.haveQueue {
+		if err := fdnotifier.AddFD(fd.handle.fd, &fd.queue); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
diff --git a/pkg/sentry/fsimpl/gofer/socket.go b/pkg/sentry/fsimpl/gofer/socket.go
index 326b940a7..a21199eac 100644
--- a/pkg/sentry/fsimpl/gofer/socket.go
+++ b/pkg/sentry/fsimpl/gofer/socket.go
@@ -42,9 +42,6 @@ type endpoint struct {
 	// dentry is the filesystem dentry which produced this endpoint.
 	dentry *dentry
 
-	// file is the p9 file that contains a single unopened fid.
-	file p9.File `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
-
 	// path is the sentry path where this endpoint is bound.
 	path string
 }
@@ -116,7 +113,7 @@ func (e *endpoint) UnidirectionalConnect(ctx context.Context) (transport.Connect
 }
 
 func (e *endpoint) newConnectedEndpoint(ctx context.Context, flags p9.ConnectFlags, queue *waiter.Queue) (*host.SCMConnectedEndpoint, *syserr.Error) {
-	hostFile, err := e.file.Connect(flags)
+	hostFile, err := e.dentry.file.connect(ctx, flags)
 	if err != nil {
 		return nil, syserr.ErrConnectionRefused
 	}
@@ -131,7 +128,7 @@ func (e *endpoint) newConnectedEndpoint(ctx context.Context, flags p9.ConnectFla
 
 	c, serr := host.NewSCMEndpoint(ctx, hostFD, queue, e.path)
 	if serr != nil {
-		log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v flags %+v: %v", e.file, flags, serr)
+		log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v flags %+v: %v", e.dentry.file, flags, serr)
 		return nil, serr
 	}
 	return c, nil
diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go
index 71581736c..625400c0b 100644
--- a/pkg/sentry/fsimpl/gofer/special_file.go
+++ b/pkg/sentry/fsimpl/gofer/special_file.go
@@ -15,7 +15,6 @@
 package gofer
 
 import (
-	"sync"
 	"sync/atomic"
 	"syscall"
 
@@ -25,6 +24,7 @@ import (
 	"gvisor.dev/gvisor/pkg/p9"
 	"gvisor.dev/gvisor/pkg/safemem"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/usermem"
 	"gvisor.dev/gvisor/pkg/waiter"
@@ -40,7 +40,7 @@ type specialFileFD struct {
 	fileDescription
 
 	// handle is used for file I/O. handle is immutable.
-	handle handle `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
+	handle handle `state:"nosave"`
 
 	// isRegularFile is true if this FD represents a regular file which is only
 	// possible when filesystemOptions.regularFilesUseSpecialFileFD is in
@@ -54,12 +54,20 @@ type specialFileFD struct {
 
 	// haveQueue is true if this file description represents a file for which
 	// queue may send I/O readiness events. haveQueue is immutable.
-	haveQueue bool
+	haveQueue bool `state:"nosave"`
 	queue     waiter.Queue
 
 	// If seekable is true, off is the file offset. off is protected by mu.
 	mu  sync.Mutex `state:"nosave"`
 	off int64
+
+	// If haveBuf is non-zero, this FD represents a pipe, and buf contains data
+	// read from the pipe from previous calls to specialFileFD.savePipeData().
+	// haveBuf and buf are protected by bufMu. haveBuf is accessed using atomic
+	// memory operations.
+	bufMu   sync.Mutex `state:"nosave"`
+	haveBuf uint32
+	buf     []byte
 }
 
 func newSpecialFileFD(h handle, mnt *vfs.Mount, d *dentry, locks *vfs.FileLocks, flags uint32) (*specialFileFD, error) {
@@ -87,6 +95,9 @@ func newSpecialFileFD(h handle, mnt *vfs.Mount, d *dentry, locks *vfs.FileLocks,
 		}
 		return nil, err
 	}
+	d.fs.syncMu.Lock()
+	d.fs.specialFileFDs[fd] = struct{}{}
+	d.fs.syncMu.Unlock()
 	return fd, nil
 }
 
@@ -161,26 +172,51 @@ func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs
 		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.cachedMetadataAuthoritative() {
 		d.touchAtime(fd.vfsfd.Mount())
 	}
+
+	bufN := int64(0)
+	if atomic.LoadUint32(&fd.haveBuf) != 0 {
+		var err error
+		fd.bufMu.Lock()
+		if len(fd.buf) != 0 {
+			var n int
+			n, err = dst.CopyOut(ctx, fd.buf)
+			dst = dst.DropFirst(n)
+			fd.buf = fd.buf[n:]
+			if len(fd.buf) == 0 {
+				atomic.StoreUint32(&fd.haveBuf, 0)
+				fd.buf = nil
+			}
+			bufN = int64(n)
+			if offset >= 0 {
+				offset += bufN
+			}
+		}
+		fd.bufMu.Unlock()
+		if err != nil {
+			return bufN, err
+		}
+	}
+
+	// Going through dst.CopyOutFrom() would hold 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.
 	buf := make([]byte, dst.NumBytes())
 	n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset))
 	if err == syserror.EAGAIN {
 		err = syserror.ErrWouldBlock
 	}
 	if n == 0 {
-		return 0, err
+		return bufN, err
 	}
 	if cp, cperr := dst.CopyOut(ctx, buf[:n]); cperr != nil {
-		return int64(cp), cperr
+		return bufN + int64(cp), cperr
 	}
-	return int64(n), err
+	return bufN + int64(n), err
 }
 
 // Read implements vfs.FileDescriptionImpl.Read.
@@ -217,16 +253,16 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off
 	}
 
 	d := fd.dentry()
-	// If the regular file fd was opened with O_APPEND, make sure the file size
-	// is updated. There is a possible race here if size is modified externally
-	// after metadata cache is updated.
-	if fd.isRegularFile && fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 && !d.cachedMetadataAuthoritative() {
-		if err := d.updateFromGetattr(ctx); err != nil {
-			return 0, offset, err
+	if fd.isRegularFile {
+		// If the regular file fd was opened with O_APPEND, make sure the file
+		// size is updated. There is a possible race here if size is modified
+		// externally after metadata cache is updated.
+		if fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 && !d.cachedMetadataAuthoritative() {
+			if err := d.updateFromGetattr(ctx); err != nil {
+				return 0, offset, err
+			}
 		}
-	}
 
-	if fd.isRegularFile {
 		// We need to hold the metadataMu *while* writing to a regular file.
 		d.metadataMu.Lock()
 		defer d.metadataMu.Unlock()
@@ -306,13 +342,31 @@ func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) (
 
 // Sync implements vfs.FileDescriptionImpl.Sync.
 func (fd *specialFileFD) Sync(ctx context.Context) error {
-	// If we have a host FD, fsyncing it is likely to be faster than an fsync
-	// RPC.
-	if fd.handle.fd >= 0 {
-		ctx.UninterruptibleSleepStart(false)
-		err := syscall.Fsync(int(fd.handle.fd))
-		ctx.UninterruptibleSleepFinish(false)
-		return err
+	return fd.sync(ctx, false /* forFilesystemSync */)
+}
+
+func (fd *specialFileFD) sync(ctx context.Context, forFilesystemSync bool) error {
+	err := func() error {
+		// If we have a host FD, fsyncing it is likely to be faster than an fsync
+		// RPC.
+		if fd.handle.fd >= 0 {
+			ctx.UninterruptibleSleepStart(false)
+			err := syscall.Fsync(int(fd.handle.fd))
+			ctx.UninterruptibleSleepFinish(false)
+			return err
+		}
+		return fd.handle.file.fsync(ctx)
+	}()
+	if err != nil {
+		if !forFilesystemSync {
+			return err
+		}
+		// Only return err if we can reasonably have expected sync to succeed
+		// (fd represents a regular file that was opened for writing).
+		if fd.isRegularFile && fd.vfsfd.IsWritable() {
+			return err
+		}
+		ctx.Debugf("gofer.specialFileFD.sync: syncing non-writable or non-regular-file FD failed: %v", err)
 	}
-	return fd.handle.file.fsync(ctx)
+	return nil
 }
diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD
index 8a3e5c81e..dc0f86061 100644
--- a/pkg/sentry/fsimpl/host/BUILD
+++ b/pkg/sentry/fsimpl/host/BUILD
@@ -34,6 +34,7 @@ go_library(
         "inode_refs.go",
         "ioctl_unsafe.go",
         "mmap.go",
+        "save_restore.go",
         "socket.go",
         "socket_iovec.go",
         "socket_unsafe.go",
diff --git a/pkg/sentry/fsimpl/host/control.go b/pkg/sentry/fsimpl/host/control.go
index 0135e4428..13ef48cb5 100644
--- a/pkg/sentry/fsimpl/host/control.go
+++ b/pkg/sentry/fsimpl/host/control.go
@@ -79,7 +79,7 @@ func fdsToFiles(ctx context.Context, fds []int) []*vfs.FileDescription {
 		}
 
 		// Create the file backed by hostFD.
-		file, err := ImportFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fd, false /* isTTY */)
+		file, err := NewFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fd, &NewFDOptions{})
 		if err != nil {
 			ctx.Warningf("Error creating file from host FD: %v", err)
 			break
diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go
index 698e913fe..eeed0f97d 100644
--- a/pkg/sentry/fsimpl/host/host.go
+++ b/pkg/sentry/fsimpl/host/host.go
@@ -19,6 +19,7 @@ package host
 import (
 	"fmt"
 	"math"
+	"sync/atomic"
 	"syscall"
 
 	"golang.org/x/sys/unix"
@@ -40,34 +41,106 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
-func newInode(fs *filesystem, hostFD int, fileType linux.FileMode, isTTY bool) (*inode, error) {
-	// Determine if hostFD is seekable. If not, this syscall will return ESPIPE
-	// (see fs/read_write.c:llseek), e.g. for pipes, sockets, and some character
-	// devices.
+// inode implements kernfs.Inode.
+//
+// +stateify savable
+type inode struct {
+	kernfs.InodeNoStatFS
+	kernfs.InodeNotDirectory
+	kernfs.InodeNotSymlink
+	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
+
+	locks vfs.FileLocks
+
+	// When the reference count reaches zero, the host fd is closed.
+	inodeRefs
+
+	// hostFD contains the host fd that this file was originally created from,
+	// which must be available at time of restore.
+	//
+	// This field is initialized at creation time and is immutable.
+	hostFD int
+
+	// ino is an inode number unique within this filesystem.
+	//
+	// This field is initialized at creation time and is immutable.
+	ino uint64
+
+	// ftype is the file's type (a linux.S_IFMT mask).
+	//
+	// This field is initialized at creation time and is immutable.
+	ftype uint16
+
+	// mayBlock is true if hostFD is non-blocking, and operations on it may
+	// return EAGAIN or EWOULDBLOCK instead of blocking.
+	//
+	// This field is initialized at creation time and is immutable.
+	mayBlock bool
+
+	// seekable is false if lseek(hostFD) returns ESPIPE. We assume that file
+	// offsets are meaningful iff seekable is true.
+	//
+	// This field is initialized at creation time and is immutable.
+	seekable bool
+
+	// isTTY is true if this file represents a TTY.
+	//
+	// This field is initialized at creation time and is immutable.
+	isTTY bool
+
+	// savable is true if hostFD may be saved/restored by its numeric value.
+	//
+	// This field is initialized at creation time and is immutable.
+	savable bool
+
+	// Event queue for blocking operations.
+	queue waiter.Queue
+
+	// mapsMu protects mappings.
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// If this file is mmappable, mappings tracks mappings of hostFD into
+	// memmap.MappingSpaces.
+	mappings memmap.MappingSet
+
+	// pf implements platform.File for mappings of hostFD.
+	pf inodePlatformFile
+
+	// If haveBuf is non-zero, hostFD represents a pipe, and buf contains data
+	// read from the pipe from previous calls to inode.beforeSave(). haveBuf
+	// and buf are protected by bufMu. haveBuf is accessed using atomic memory
+	// operations.
+	bufMu   sync.Mutex `state:"nosave"`
+	haveBuf uint32
+	buf     []byte
+}
+
+func newInode(ctx context.Context, fs *filesystem, hostFD int, savable bool, fileType linux.FileMode, isTTY bool) (*inode, error) {
+	// Determine if hostFD is seekable.
 	_, err := unix.Seek(hostFD, 0, linux.SEEK_CUR)
 	seekable := err != syserror.ESPIPE
+	// We expect regular files to be seekable, as this is required for them to
+	// be memory-mappable.
+	if !seekable && fileType == syscall.S_IFREG {
+		ctx.Infof("host.newInode: host FD %d is a non-seekable regular file", hostFD)
+		return nil, syserror.ESPIPE
+	}
 
 	i := &inode{
-		hostFD:     hostFD,
-		ino:        fs.NextIno(),
-		isTTY:      isTTY,
-		wouldBlock: wouldBlock(uint32(fileType)),
-		seekable:   seekable,
-		// NOTE(b/38213152): Technically, some obscure char devices can be memory
-		// mapped, but we only allow regular files.
-		canMap: fileType == linux.S_IFREG,
+		hostFD:   hostFD,
+		ino:      fs.NextIno(),
+		ftype:    uint16(fileType),
+		mayBlock: fileType != syscall.S_IFREG && fileType != syscall.S_IFDIR,
+		seekable: seekable,
+		isTTY:    isTTY,
+		savable:  savable,
 	}
 	i.pf.inode = i
 	i.EnableLeakCheck()
 
-	// Non-seekable files can't be memory mapped, assert this.
-	if !i.seekable && i.canMap {
-		panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped")
-	}
-
-	// If the hostFD would block, we must set it to non-blocking and handle
-	// blocking behavior in the sentry.
-	if i.wouldBlock {
+	// If the hostFD can return EWOULDBLOCK when set to non-blocking, do so and
+	// handle blocking behavior in the sentry.
+	if i.mayBlock {
 		if err := syscall.SetNonblock(i.hostFD, true); err != nil {
 			return nil, err
 		}
@@ -80,6 +153,11 @@ func newInode(fs *filesystem, hostFD int, fileType linux.FileMode, isTTY bool) (
 
 // NewFDOptions contains options to NewFD.
 type NewFDOptions struct {
+	// If Savable is true, the host file descriptor may be saved/restored by
+	// numeric value; the sandbox API requires a corresponding host FD with the
+	// same numeric value to be provieded at time of restore.
+	Savable bool
+
 	// If IsTTY is true, the file descriptor is a TTY.
 	IsTTY bool
 
@@ -114,7 +192,7 @@ func NewFD(ctx context.Context, mnt *vfs.Mount, hostFD int, opts *NewFDOptions)
 	}
 
 	d := &kernfs.Dentry{}
-	i, err := newInode(fs, hostFD, linux.FileMode(s.Mode).FileType(), opts.IsTTY)
+	i, err := newInode(ctx, fs, hostFD, opts.Savable, linux.FileMode(s.Mode).FileType(), opts.IsTTY)
 	if err != nil {
 		return nil, err
 	}
@@ -132,7 +210,8 @@ func NewFD(ctx context.Context, mnt *vfs.Mount, hostFD int, opts *NewFDOptions)
 // ImportFD sets up and returns a vfs.FileDescription from a donated fd.
 func ImportFD(ctx context.Context, mnt *vfs.Mount, hostFD int, isTTY bool) (*vfs.FileDescription, error) {
 	return NewFD(ctx, mnt, hostFD, &NewFDOptions{
-		IsTTY: isTTY,
+		Savable: true,
+		IsTTY:   isTTY,
 	})
 }
 
@@ -191,68 +270,6 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe
 	return vfs.PrependPathSyntheticError{}
 }
 
-// inode implements kernfs.Inode.
-//
-// +stateify savable
-type inode struct {
-	kernfs.InodeNoStatFS
-	kernfs.InodeNotDirectory
-	kernfs.InodeNotSymlink
-	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
-
-	locks vfs.FileLocks
-
-	// When the reference count reaches zero, the host fd is closed.
-	inodeRefs
-
-	// hostFD contains the host fd that this file was originally created from,
-	// which must be available at time of restore.
-	//
-	// This field is initialized at creation time and is immutable.
-	hostFD int
-
-	// ino is an inode number unique within this filesystem.
-	//
-	// This field is initialized at creation time and is immutable.
-	ino uint64
-
-	// isTTY is true if this file represents a TTY.
-	//
-	// This field is initialized at creation time and is immutable.
-	isTTY bool
-
-	// seekable is false if the host fd points to a file representing a stream,
-	// e.g. a socket or a pipe. Such files are not seekable and can return
-	// EWOULDBLOCK for I/O operations.
-	//
-	// This field is initialized at creation time and is immutable.
-	seekable bool
-
-	// wouldBlock is true if the host FD would return EWOULDBLOCK for
-	// operations that would block.
-	//
-	// This field is initialized at creation time and is immutable.
-	wouldBlock bool
-
-	// Event queue for blocking operations.
-	queue waiter.Queue
-
-	// canMap specifies whether we allow the file to be memory mapped.
-	//
-	// This field is initialized at creation time and is immutable.
-	canMap bool
-
-	// mapsMu protects mappings.
-	mapsMu sync.Mutex `state:"nosave"`
-
-	// If canMap is true, mappings tracks mappings of hostFD into
-	// memmap.MappingSpaces.
-	mappings memmap.MappingSet
-
-	// pf implements platform.File for mappings of hostFD.
-	pf inodePlatformFile
-}
-
 // CheckPermissions implements kernfs.Inode.CheckPermissions.
 func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error {
 	var s syscall.Stat_t
@@ -448,7 +465,7 @@ func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Cre
 // DecRef implements kernfs.Inode.DecRef.
 func (i *inode) DecRef(ctx context.Context) {
 	i.inodeRefs.DecRef(func() {
-		if i.wouldBlock {
+		if i.mayBlock {
 			fdnotifier.RemoveFD(int32(i.hostFD))
 		}
 		if err := unix.Close(i.hostFD); err != nil {
@@ -567,6 +584,13 @@ func (f *fileDescription) Allocate(ctx context.Context, mode, offset, length uin
 
 // PRead implements vfs.FileDescriptionImpl.PRead.
 func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	// Check that flags are supported.
+	//
+	// TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
+	if opts.Flags&^linux.RWF_HIPRI != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
 	i := f.inode
 	if !i.seekable {
 		return 0, syserror.ESPIPE
@@ -577,19 +601,31 @@ func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, off
 
 // Read implements vfs.FileDescriptionImpl.Read.
 func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	// Check that flags are supported.
+	//
+	// TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
+	if opts.Flags&^linux.RWF_HIPRI != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
 	i := f.inode
 	if !i.seekable {
+		bufN, err := i.readFromBuf(ctx, &dst)
+		if err != nil {
+			return bufN, err
+		}
 		n, err := readFromHostFD(ctx, i.hostFD, dst, -1, opts.Flags)
+		total := bufN + n
 		if isBlockError(err) {
 			// If we got any data at all, return it as a "completed" partial read
 			// rather than retrying until complete.
-			if n != 0 {
+			if total != 0 {
 				err = nil
 			} else {
 				err = syserror.ErrWouldBlock
 			}
 		}
-		return n, err
+		return total, err
 	}
 
 	f.offsetMu.Lock()
@@ -599,13 +635,26 @@ func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts
 	return n, err
 }
 
-func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags uint32) (int64, error) {
-	// Check that flags are supported.
-	//
-	// TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
-	if flags&^linux.RWF_HIPRI != 0 {
-		return 0, syserror.EOPNOTSUPP
+func (i *inode) readFromBuf(ctx context.Context, dst *usermem.IOSequence) (int64, error) {
+	if atomic.LoadUint32(&i.haveBuf) == 0 {
+		return 0, nil
 	}
+	i.bufMu.Lock()
+	defer i.bufMu.Unlock()
+	if len(i.buf) == 0 {
+		return 0, nil
+	}
+	n, err := dst.CopyOut(ctx, i.buf)
+	*dst = dst.DropFirst(n)
+	i.buf = i.buf[n:]
+	if len(i.buf) == 0 {
+		atomic.StoreUint32(&i.haveBuf, 0)
+		i.buf = nil
+	}
+	return int64(n), err
+}
+
+func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags uint32) (int64, error) {
 	reader := hostfd.GetReadWriterAt(int32(hostFD), offset, flags)
 	n, err := dst.CopyOutFrom(ctx, reader)
 	hostfd.PutReadWriterAt(reader)
@@ -735,14 +784,16 @@ func (f *fileDescription) Seek(_ context.Context, offset int64, whence int32) (i
 }
 
 // Sync implements vfs.FileDescriptionImpl.Sync.
-func (f *fileDescription) Sync(context.Context) error {
+func (f *fileDescription) Sync(ctx context.Context) error {
 	// TODO(gvisor.dev/issue/1897): Currently, we always sync everything.
 	return unix.Fsync(f.inode.hostFD)
 }
 
 // ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap.
 func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error {
-	if !f.inode.canMap {
+	// NOTE(b/38213152): Technically, some obscure char devices can be memory
+	// mapped, but we only allow regular files.
+	if f.inode.ftype != syscall.S_IFREG {
 		return syserror.ENODEV
 	}
 	i := f.inode
@@ -753,13 +804,17 @@ func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts
 // EventRegister implements waiter.Waitable.EventRegister.
 func (f *fileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
 	f.inode.queue.EventRegister(e, mask)
-	fdnotifier.UpdateFD(int32(f.inode.hostFD))
+	if f.inode.mayBlock {
+		fdnotifier.UpdateFD(int32(f.inode.hostFD))
+	}
 }
 
 // EventUnregister implements waiter.Waitable.EventUnregister.
 func (f *fileDescription) EventUnregister(e *waiter.Entry) {
 	f.inode.queue.EventUnregister(e)
-	fdnotifier.UpdateFD(int32(f.inode.hostFD))
+	if f.inode.mayBlock {
+		fdnotifier.UpdateFD(int32(f.inode.hostFD))
+	}
 }
 
 // Readiness uses the poll() syscall to check the status of the underlying FD.
diff --git a/pkg/sentry/fsimpl/host/mmap.go b/pkg/sentry/fsimpl/host/mmap.go
index b51a17bed..3d7eb2f96 100644
--- a/pkg/sentry/fsimpl/host/mmap.go
+++ b/pkg/sentry/fsimpl/host/mmap.go
@@ -43,7 +43,7 @@ type inodePlatformFile struct {
 	fileMapper fsutil.HostFileMapper
 
 	// fileMapperInitOnce is used to lazily initialize fileMapper.
-	fileMapperInitOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported.
+	fileMapperInitOnce sync.Once `state:"nosave"`
 }
 
 // IncRef implements memmap.File.IncRef.
diff --git a/pkg/sentry/fsimpl/host/save_restore.go b/pkg/sentry/fsimpl/host/save_restore.go
new file mode 100644
index 000000000..7e32a8863
--- /dev/null
+++ b/pkg/sentry/fsimpl/host/save_restore.go
@@ -0,0 +1,78 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"fmt"
+	"io"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/fdnotifier"
+	"gvisor.dev/gvisor/pkg/safemem"
+	"gvisor.dev/gvisor/pkg/sentry/hostfd"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// beforeSave is invoked by stateify.
+func (i *inode) beforeSave() {
+	if !i.savable {
+		panic("host.inode is not savable")
+	}
+	if i.ftype == syscall.S_IFIFO {
+		// If this pipe FD is readable, drain it so that bytes in the pipe can
+		// be read after restore. (This is a legacy VFS1 feature.) We don't
+		// know if the pipe FD is readable, so just try reading and tolerate
+		// EBADF from the read.
+		i.bufMu.Lock()
+		defer i.bufMu.Unlock()
+		var buf [usermem.PageSize]byte
+		for {
+			n, err := hostfd.Preadv2(int32(i.hostFD), safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:])), -1 /* offset */, 0 /* flags */)
+			if n != 0 {
+				i.buf = append(i.buf, buf[:n]...)
+			}
+			if err != nil {
+				if err == io.EOF || err == syscall.EAGAIN || err == syscall.EBADF {
+					break
+				}
+				panic(fmt.Errorf("host.inode.beforeSave: buffering from pipe failed: %v", err))
+			}
+		}
+		if len(i.buf) != 0 {
+			atomic.StoreUint32(&i.haveBuf, 1)
+		}
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (i *inode) afterLoad() {
+	if i.mayBlock {
+		if err := syscall.SetNonblock(i.hostFD, true); err != nil {
+			panic(fmt.Sprintf("host.inode.afterLoad: failed to set host FD %d non-blocking: %v", i.hostFD, err))
+		}
+		if err := fdnotifier.AddFD(int32(i.hostFD), &i.queue); err != nil {
+			panic(fmt.Sprintf("host.inode.afterLoad: fdnotifier.AddFD(%d) failed: %v", i.hostFD, err))
+		}
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (i *inodePlatformFile) afterLoad() {
+	if i.fileMapper.IsInited() {
+		// Ensure that we don't call i.fileMapper.Init() again.
+		i.fileMapperInitOnce.Do(func() {})
+	}
+}
diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go
index 412bdb2eb..b2f43a119 100644
--- a/pkg/sentry/fsimpl/host/util.go
+++ b/pkg/sentry/fsimpl/host/util.go
@@ -43,12 +43,6 @@ func timespecToStatxTimestamp(ts unix.Timespec) linux.StatxTimestamp {
 	return linux.StatxTimestamp{Sec: int64(ts.Sec), Nsec: uint32(ts.Nsec)}
 }
 
-// wouldBlock returns true for file types that can return EWOULDBLOCK
-// for blocking operations, e.g. pipes, character devices, and sockets.
-func wouldBlock(fileType uint32) bool {
-	return fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK
-}
-
 // isBlockError checks if an error is EAGAIN or EWOULDBLOCK.
 // If so, they can be transformed into syserror.ErrWouldBlock.
 func isBlockError(err error) bool {
diff --git a/pkg/sentry/fsimpl/overlay/overlay.go b/pkg/sentry/fsimpl/overlay/overlay.go
index 4c5de8d32..f28411b5b 100644
--- a/pkg/sentry/fsimpl/overlay/overlay.go
+++ b/pkg/sentry/fsimpl/overlay/overlay.go
@@ -458,9 +458,9 @@ type dentry struct {
 	//
 	// - isMappable is non-zero iff wrappedMappable is non-nil. isMappable is
 	// accessed using atomic memory operations.
-	mapsMu          sync.Mutex
+	mapsMu          sync.Mutex `state:"nosave"`
 	lowerMappings   memmap.MappingSet
-	dataMu          sync.RWMutex
+	dataMu          sync.RWMutex `state:"nosave"`
 	wrappedMappable memmap.Mappable
 	isMappable      uint32
 
diff --git a/pkg/sentry/fsimpl/tmpfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD
index 85850cf34..fe520b6fd 100644
--- a/pkg/sentry/fsimpl/tmpfs/BUILD
+++ b/pkg/sentry/fsimpl/tmpfs/BUILD
@@ -48,6 +48,7 @@ go_library(
         "inode_refs.go",
         "named_pipe.go",
         "regular_file.go",
+        "save_restore.go",
         "socket_file.go",
         "symlink.go",
         "tmpfs.go",
diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go
index ce4e3eda7..98680fde9 100644
--- a/pkg/sentry/fsimpl/tmpfs/regular_file.go
+++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go
@@ -42,7 +42,7 @@ type regularFile struct {
 	inode inode
 
 	// memFile is a platform.File used to allocate pages to this regularFile.
-	memFile *pgalloc.MemoryFile
+	memFile *pgalloc.MemoryFile `state:"nosave"`
 
 	// memoryUsageKind is the memory accounting category under which pages backing
 	// this regularFile's contents are accounted.
@@ -92,7 +92,7 @@ type regularFile struct {
 
 func (fs *filesystem) newRegularFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode {
 	file := &regularFile{
-		memFile:         fs.memFile,
+		memFile:         fs.mfp.MemoryFile(),
 		memoryUsageKind: usage.Tmpfs,
 		seals:           linux.F_SEAL_SEAL,
 	}
diff --git a/pkg/sentry/fsimpl/tmpfs/save_restore.go b/pkg/sentry/fsimpl/tmpfs/save_restore.go
new file mode 100644
index 000000000..b27f75cc2
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/save_restore.go
@@ -0,0 +1,20 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tmpfs
+
+// afterLoad is called by stateify.
+func (rf *regularFile) afterLoad() {
+	rf.memFile = rf.inode.fs.mfp.MemoryFile()
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
index e2a0aac69..4ce859d57 100644
--- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
@@ -61,8 +61,9 @@ type FilesystemType struct{}
 type filesystem struct {
 	vfsfs vfs.Filesystem
 
-	// memFile is used to allocate pages to for regular files.
-	memFile *pgalloc.MemoryFile
+	// mfp is used to allocate memory that stores regular file contents. mfp is
+	// immutable.
+	mfp pgalloc.MemoryFileProvider
 
 	// clock is a realtime clock used to set timestamps in file operations.
 	clock time.Clock
@@ -106,8 +107,8 @@ type FilesystemOpts struct {
 
 // GetFilesystem implements vfs.FilesystemType.GetFilesystem.
 func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, _ string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
-	memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx)
-	if memFileProvider == nil {
+	mfp := pgalloc.MemoryFileProviderFromContext(ctx)
+	if mfp == nil {
 		panic("MemoryFileProviderFromContext returned nil")
 	}
 
@@ -181,7 +182,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 	}
 	clock := time.RealtimeClockFromContext(ctx)
 	fs := filesystem{
-		memFile:  memFileProvider.MemoryFile(),
+		mfp:      mfp,
 		clock:    clock,
 		devMinor: devMinor,
 	}
diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go
index 81a998966..7aba31587 100644
--- a/pkg/sentry/kernel/fd_table.go
+++ b/pkg/sentry/kernel/fd_table.go
@@ -240,6 +240,10 @@ func (f *FDTable) String() string {
 
 		case fileVFS2 != nil:
 			vfsObj := fileVFS2.Mount().Filesystem().VirtualFilesystem()
+			vd := fileVFS2.VirtualDentry()
+			if vd.Dentry() == nil {
+				panic(fmt.Sprintf("fd %d (type %T) has nil dentry: %#v", fd, fileVFS2.Impl(), fileVFS2))
+			}
 			name, err := vfsObj.PathnameWithDeleted(ctx, vfs.VirtualDentry{}, fileVFS2.VirtualDentry())
 			if err != nil {
 				fmt.Fprintf(&buf, "<err: %v>\n", err)
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index 0eb2bf7bd..9b2be44d4 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -430,9 +430,8 @@ func (k *Kernel) Init(args InitKernelArgs) error {
 // SaveTo saves the state of k to w.
 //
 // Preconditions: The kernel must be paused throughout the call to SaveTo.
-func (k *Kernel) SaveTo(w wire.Writer) error {
+func (k *Kernel) SaveTo(ctx context.Context, w wire.Writer) error {
 	saveStart := time.Now()
-	ctx := k.SupervisorContext()
 
 	// Do not allow other Kernel methods to affect it while it's being saved.
 	k.extMu.Lock()
@@ -446,38 +445,55 @@ func (k *Kernel) SaveTo(w wire.Writer) error {
 	k.mf.StartEvictions()
 	k.mf.WaitForEvictions()
 
-	// Flush write operations on open files so data reaches backing storage.
-	// This must come after MemoryFile eviction since eviction may cause file
-	// writes.
-	if err := k.tasks.flushWritesToFiles(ctx); err != nil {
-		return err
-	}
+	if VFS2Enabled {
+		// Discard unsavable mappings, such as those for host file descriptors.
+		if err := k.invalidateUnsavableMappings(ctx); err != nil {
+			return fmt.Errorf("failed to invalidate unsavable mappings: %v", err)
+		}
 
-	// Remove all epoll waiter objects from underlying wait queues.
-	// NOTE: for programs to resume execution in future snapshot scenarios,
-	// we will need to re-establish these waiter objects after saving.
-	k.tasks.unregisterEpollWaiters(ctx)
+		// Prepare filesystems for saving. This must be done after
+		// invalidateUnsavableMappings(), since dropping memory mappings may
+		// affect filesystem state (e.g. page cache reference counts).
+		if err := k.vfs.PrepareSave(ctx); err != nil {
+			return err
+		}
+	} else {
+		// Flush cached file writes to backing storage. This must come after
+		// MemoryFile eviction since eviction may cause file writes.
+		if err := k.flushWritesToFiles(ctx); err != nil {
+			return err
+		}
 
-	// Clear the dirent cache before saving because Dirents must be Loaded in a
-	// particular order (parents before children), and Loading dirents from a cache
-	// breaks that order.
-	if err := k.flushMountSourceRefs(ctx); err != nil {
-		return err
-	}
+		// Remove all epoll waiter objects from underlying wait queues.
+		// NOTE: for programs to resume execution in future snapshot scenarios,
+		// we will need to re-establish these waiter objects after saving.
+		k.tasks.unregisterEpollWaiters(ctx)
 
-	// Ensure that all inode and mount release operations have completed.
-	fs.AsyncBarrier()
+		// Clear the dirent cache before saving because Dirents must be Loaded in a
+		// particular order (parents before children), and Loading dirents from a cache
+		// breaks that order.
+		if err := k.flushMountSourceRefs(ctx); err != nil {
+			return err
+		}
 
-	// Once all fs work has completed (flushed references have all been released),
-	// reset mount mappings. This allows individual mounts to save how inodes map
-	// to filesystem resources. Without this, fs.Inodes cannot be restored.
-	fs.SaveInodeMappings()
+		// Ensure that all inode and mount release operations have completed.
+		fs.AsyncBarrier()
 
-	// Discard unsavable mappings, such as those for host file descriptors.
-	// This must be done after waiting for "asynchronous fs work", which
-	// includes async I/O that may touch application memory.
-	if err := k.invalidateUnsavableMappings(ctx); err != nil {
-		return fmt.Errorf("failed to invalidate unsavable mappings: %v", err)
+		// Once all fs work has completed (flushed references have all been released),
+		// reset mount mappings. This allows individual mounts to save how inodes map
+		// to filesystem resources. Without this, fs.Inodes cannot be restored.
+		fs.SaveInodeMappings()
+
+		// Discard unsavable mappings, such as those for host file descriptors.
+		// This must be done after waiting for "asynchronous fs work", which
+		// includes async I/O that may touch application memory.
+		//
+		// TODO(gvisor.dev/issue/1624): This rationale is believed to be
+		// obsolete since AIO callbacks are now waited-for by Kernel.Pause(),
+		// but this order is conservatively retained for VFS1.
+		if err := k.invalidateUnsavableMappings(ctx); err != nil {
+			return fmt.Errorf("failed to invalidate unsavable mappings: %v", err)
+		}
 	}
 
 	// Save the CPUID FeatureSet before the rest of the kernel so we can
@@ -486,14 +502,14 @@ func (k *Kernel) SaveTo(w wire.Writer) error {
 	//
 	// N.B. This will also be saved along with the full kernel save below.
 	cpuidStart := time.Now()
-	if _, err := state.Save(k.SupervisorContext(), w, k.FeatureSet()); err != nil {
+	if _, err := state.Save(ctx, w, k.FeatureSet()); err != nil {
 		return err
 	}
 	log.Infof("CPUID save took [%s].", time.Since(cpuidStart))
 
 	// Save the kernel state.
 	kernelStart := time.Now()
-	stats, err := state.Save(k.SupervisorContext(), w, k)
+	stats, err := state.Save(ctx, w, k)
 	if err != nil {
 		return err
 	}
@@ -502,7 +518,7 @@ func (k *Kernel) SaveTo(w wire.Writer) error {
 
 	// Save the memory file's state.
 	memoryStart := time.Now()
-	if err := k.mf.SaveTo(k.SupervisorContext(), w); err != nil {
+	if err := k.mf.SaveTo(ctx, w); err != nil {
 		return err
 	}
 	log.Infof("Memory save took [%s].", time.Since(memoryStart))
@@ -514,11 +530,9 @@ func (k *Kernel) SaveTo(w wire.Writer) error {
 
 // flushMountSourceRefs flushes the MountSources for all mounted filesystems
 // and open FDs.
+//
+// Preconditions: !VFS2Enabled.
 func (k *Kernel) flushMountSourceRefs(ctx context.Context) error {
-	if VFS2Enabled {
-		return nil // Not relevant.
-	}
-
 	// Flush all mount sources for currently mounted filesystems in each task.
 	flushed := make(map[*fs.MountNamespace]struct{})
 	k.tasks.mu.RLock()
@@ -561,13 +575,9 @@ func (ts *TaskSet) forEachFDPaused(ctx context.Context, f func(*fs.File, *vfs.Fi
 	return err
 }
 
-func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error {
-	// TODO(gvisor.dev/issue/1663): Add save support for VFS2.
-	if VFS2Enabled {
-		return nil
-	}
-
-	return ts.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error {
+// Preconditions: !VFS2Enabled.
+func (k *Kernel) flushWritesToFiles(ctx context.Context) error {
+	return k.tasks.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error {
 		if flags := file.Flags(); !flags.Write {
 			return nil
 		}
@@ -589,37 +599,8 @@ func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error {
 	})
 }
 
-// Preconditions: The kernel must be paused.
-func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error {
-	invalidated := make(map[*mm.MemoryManager]struct{})
-	k.tasks.mu.RLock()
-	defer k.tasks.mu.RUnlock()
-	for t := range k.tasks.Root.tids {
-		// We can skip locking Task.mu here since the kernel is paused.
-		if mm := t.tc.MemoryManager; mm != nil {
-			if _, ok := invalidated[mm]; !ok {
-				if err := mm.InvalidateUnsavable(ctx); err != nil {
-					return err
-				}
-				invalidated[mm] = struct{}{}
-			}
-		}
-		// I really wish we just had a sync.Map of all MMs...
-		if r, ok := t.runState.(*runSyscallAfterExecStop); ok {
-			if err := r.tc.MemoryManager.InvalidateUnsavable(ctx); err != nil {
-				return err
-			}
-		}
-	}
-	return nil
-}
-
+// Preconditions: !VFS2Enabled.
 func (ts *TaskSet) unregisterEpollWaiters(ctx context.Context) {
-	// TODO(gvisor.dev/issue/1663): Add save support for VFS2.
-	if VFS2Enabled {
-		return
-	}
-
 	ts.mu.RLock()
 	defer ts.mu.RUnlock()
 
@@ -644,8 +625,33 @@ func (ts *TaskSet) unregisterEpollWaiters(ctx context.Context) {
 	}
 }
 
+// Preconditions: The kernel must be paused.
+func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error {
+	invalidated := make(map[*mm.MemoryManager]struct{})
+	k.tasks.mu.RLock()
+	defer k.tasks.mu.RUnlock()
+	for t := range k.tasks.Root.tids {
+		// We can skip locking Task.mu here since the kernel is paused.
+		if mm := t.tc.MemoryManager; mm != nil {
+			if _, ok := invalidated[mm]; !ok {
+				if err := mm.InvalidateUnsavable(ctx); err != nil {
+					return err
+				}
+				invalidated[mm] = struct{}{}
+			}
+		}
+		// I really wish we just had a sync.Map of all MMs...
+		if r, ok := t.runState.(*runSyscallAfterExecStop); ok {
+			if err := r.tc.MemoryManager.InvalidateUnsavable(ctx); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
 // LoadFrom returns a new Kernel loaded from args.
-func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clocks) error {
+func (k *Kernel) LoadFrom(ctx context.Context, r wire.Reader, net inet.Stack, clocks sentrytime.Clocks, vfsOpts *vfs.CompleteRestoreOptions) error {
 	loadStart := time.Now()
 
 	initAppCores := k.applicationCores
@@ -656,7 +662,7 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock
 	// don't need to explicitly install it in the Kernel.
 	cpuidStart := time.Now()
 	var features cpuid.FeatureSet
-	if _, err := state.Load(k.SupervisorContext(), r, &features); err != nil {
+	if _, err := state.Load(ctx, r, &features); err != nil {
 		return err
 	}
 	log.Infof("CPUID load took [%s].", time.Since(cpuidStart))
@@ -671,7 +677,7 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock
 
 	// Load the kernel state.
 	kernelStart := time.Now()
-	stats, err := state.Load(k.SupervisorContext(), r, k)
+	stats, err := state.Load(ctx, r, k)
 	if err != nil {
 		return err
 	}
@@ -684,7 +690,7 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock
 
 	// Load the memory file's state.
 	memoryStart := time.Now()
-	if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil {
+	if err := k.mf.LoadFrom(ctx, r); err != nil {
 		return err
 	}
 	log.Infof("Memory load took [%s].", time.Since(memoryStart))
@@ -696,11 +702,17 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock
 		net.Resume()
 	}
 
-	// Ensure that all pending asynchronous work is complete:
-	//   - namedpipe opening
-	//   - inode file opening
-	if err := fs.AsyncErrorBarrier(); err != nil {
-		return err
+	if VFS2Enabled {
+		if err := k.vfs.CompleteRestore(ctx, vfsOpts); err != nil {
+			return err
+		}
+	} else {
+		// Ensure that all pending asynchronous work is complete:
+		//   - namedpipe opening
+		//   - inode file opening
+		if err := fs.AsyncErrorBarrier(); err != nil {
+			return err
+		}
 	}
 
 	tcpip.AsyncLoading.Wait()
diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go
index 1a152142b..d96bf253b 100644
--- a/pkg/sentry/kernel/pipe/vfs.go
+++ b/pkg/sentry/kernel/pipe/vfs.go
@@ -33,6 +33,8 @@ import (
 
 // VFSPipe represents the actual pipe, analagous to an inode. VFSPipes should
 // not be copied.
+//
+// +stateify savable
 type VFSPipe struct {
 	// mu protects the fields below.
 	mu sync.Mutex `state:"nosave"`
@@ -164,6 +166,8 @@ func (vp *VFSPipe) newFD(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, l
 // VFSPipeFD implements vfs.FileDescriptionImpl for pipes. It also implements
 // non-atomic usermem.IO methods, allowing it to be passed as usermem.IO to
 // other FileDescriptions for splice(2) and tee(2).
+//
+// +stateify savable
 type VFSPipeFD struct {
 	vfsfd vfs.FileDescription
 	vfs.FileDescriptionDefaultImpl
diff --git a/pkg/sentry/socket/control/control_vfs2.go b/pkg/sentry/socket/control/control_vfs2.go
index d9621968c..5047b6261 100644
--- a/pkg/sentry/socket/control/control_vfs2.go
+++ b/pkg/sentry/socket/control/control_vfs2.go
@@ -24,6 +24,8 @@ import (
 )
 
 // SCMRightsVFS2 represents a SCM_RIGHTS socket control message.
+//
+// +stateify savable
 type SCMRightsVFS2 interface {
 	transport.RightsControlMessage
 
@@ -34,9 +36,11 @@ type SCMRightsVFS2 interface {
 	Files(ctx context.Context, max int) (rf RightsFilesVFS2, truncated bool)
 }
 
-// RightsFiles represents a SCM_RIGHTS socket control message. A reference is
+// RightsFilesVFS2 represents a SCM_RIGHTS socket control message. A reference is
 // maintained for each vfs.FileDescription and is release either when an FD is created or
 // when the Release method is called.
+//
+// +stateify savable
 type RightsFilesVFS2 []*vfs.FileDescription
 
 // NewSCMRightsVFS2 creates a new SCM_RIGHTS socket control message
diff --git a/pkg/sentry/socket/hostinet/socket_vfs2.go b/pkg/sentry/socket/hostinet/socket_vfs2.go
index ce48a9903..9a2cac40b 100644
--- a/pkg/sentry/socket/hostinet/socket_vfs2.go
+++ b/pkg/sentry/socket/hostinet/socket_vfs2.go
@@ -33,6 +33,7 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
+// +stateify savable
 type socketVFS2 struct {
 	vfsfd vfs.FileDescription
 	vfs.FileDescriptionDefaultImpl
diff --git a/pkg/sentry/socket/netlink/socket_vfs2.go b/pkg/sentry/socket/netlink/socket_vfs2.go
index c83b23242..461d524e5 100644
--- a/pkg/sentry/socket/netlink/socket_vfs2.go
+++ b/pkg/sentry/socket/netlink/socket_vfs2.go
@@ -37,6 +37,8 @@ import (
 // to/from the kernel.
 //
 // SocketVFS2 implements socket.SocketVFS2 and transport.Credentialer.
+//
+// +stateify savable
 type SocketVFS2 struct {
 	vfsfd vfs.FileDescription
 	vfs.FileDescriptionDefaultImpl
diff --git a/pkg/sentry/socket/netstack/netstack_vfs2.go b/pkg/sentry/socket/netstack/netstack_vfs2.go
index adffba3c6..b0d9e4d9e 100644
--- a/pkg/sentry/socket/netstack/netstack_vfs2.go
+++ b/pkg/sentry/socket/netstack/netstack_vfs2.go
@@ -35,6 +35,8 @@ import (
 
 // SocketVFS2 encapsulates all the state needed to represent a network stack
 // endpoint in the kernel context.
+//
+// +stateify savable
 type SocketVFS2 struct {
 	vfsfd vfs.FileDescription
 	vfs.FileDescriptionDefaultImpl
diff --git a/pkg/sentry/state/BUILD b/pkg/sentry/state/BUILD
index 0ea4aab8b..563d60578 100644
--- a/pkg/sentry/state/BUILD
+++ b/pkg/sentry/state/BUILD
@@ -12,10 +12,12 @@ go_library(
     visibility = ["//pkg/sentry:internal"],
     deps = [
         "//pkg/abi/linux",
+        "//pkg/context",
         "//pkg/log",
         "//pkg/sentry/inet",
         "//pkg/sentry/kernel",
         "//pkg/sentry/time",
+        "//pkg/sentry/vfs",
         "//pkg/sentry/watchdog",
         "//pkg/state/statefile",
         "//pkg/syserror",
diff --git a/pkg/sentry/state/state.go b/pkg/sentry/state/state.go
index 245d2c5cf..167754537 100644
--- a/pkg/sentry/state/state.go
+++ b/pkg/sentry/state/state.go
@@ -19,10 +19,12 @@ import (
 	"fmt"
 	"io"
 
+	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/log"
 	"gvisor.dev/gvisor/pkg/sentry/inet"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/time"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sentry/watchdog"
 	"gvisor.dev/gvisor/pkg/state/statefile"
 	"gvisor.dev/gvisor/pkg/syserror"
@@ -57,7 +59,7 @@ type SaveOpts struct {
 }
 
 // Save saves the system state.
-func (opts SaveOpts) Save(k *kernel.Kernel, w *watchdog.Watchdog) error {
+func (opts SaveOpts) Save(ctx context.Context, k *kernel.Kernel, w *watchdog.Watchdog) error {
 	log.Infof("Sandbox save started, pausing all tasks.")
 	k.Pause()
 	k.ReceiveTaskStates()
@@ -81,7 +83,7 @@ func (opts SaveOpts) Save(k *kernel.Kernel, w *watchdog.Watchdog) error {
 		err = ErrStateFile{err}
 	} else {
 		// Save the kernel.
-		err = k.SaveTo(wc)
+		err = k.SaveTo(ctx, wc)
 
 		// ENOSPC is a state file error. This error can only come from
 		// writing the state file, and not from fs.FileOperations.Fsync
@@ -108,7 +110,7 @@ type LoadOpts struct {
 }
 
 // Load loads the given kernel, setting the provided platform and stack.
-func (opts LoadOpts) Load(k *kernel.Kernel, n inet.Stack, clocks time.Clocks) error {
+func (opts LoadOpts) Load(ctx context.Context, k *kernel.Kernel, n inet.Stack, clocks time.Clocks, vfsOpts *vfs.CompleteRestoreOptions) error {
 	// Open the file.
 	r, m, err := statefile.NewReader(opts.Source, opts.Key)
 	if err != nil {
@@ -118,5 +120,5 @@ func (opts LoadOpts) Load(k *kernel.Kernel, n inet.Stack, clocks time.Clocks) er
 	previousMetadata = m
 
 	// Restore the Kernel object graph.
-	return k.LoadFrom(r, n, clocks)
+	return k.LoadFrom(ctx, r, n, clocks, vfsOpts)
 }
diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD
index 996af7332..440c9307c 100644
--- a/pkg/sentry/vfs/BUILD
+++ b/pkg/sentry/vfs/BUILD
@@ -87,6 +87,7 @@ go_library(
         "pathname.go",
         "permissions.go",
         "resolving_path.go",
+        "save_restore.go",
         "vfs.go",
     ],
     visibility = ["//pkg/sentry:internal"],
diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go
index 8f36c3e3b..a98aac52b 100644
--- a/pkg/sentry/vfs/epoll.go
+++ b/pkg/sentry/vfs/epoll.go
@@ -74,7 +74,7 @@ type epollInterestKey struct {
 // +stateify savable
 type epollInterest struct {
 	// epoll is the owning EpollInstance. epoll is immutable.
-	epoll *EpollInstance
+	epoll *EpollInstance `state:"wait"`
 
 	// key is the file to which this epollInterest applies. key is immutable.
 	key epollInterestKey
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go
index 183957ad8..546e445aa 100644
--- a/pkg/sentry/vfs/file_description.go
+++ b/pkg/sentry/vfs/file_description.go
@@ -183,7 +183,6 @@ func (fd *FileDescription) DecRef(ctx context.Context) {
 		}
 		fd.vd.DecRef(ctx)
 		fd.flagsMu.Lock()
-		// TODO(gvisor.dev/issue/1663): We may need to unregister during save, as we do in VFS1.
 		if fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil {
 			fd.asyncHandler.Unregister(fd)
 		}
diff --git a/pkg/sentry/vfs/genericfstree/genericfstree.go b/pkg/sentry/vfs/genericfstree/genericfstree.go
index 2d27d9d35..ba6e6ed49 100644
--- a/pkg/sentry/vfs/genericfstree/genericfstree.go
+++ b/pkg/sentry/vfs/genericfstree/genericfstree.go
@@ -71,7 +71,7 @@ func PrependPath(vfsroot vfs.VirtualDentry, mnt *vfs.Mount, d *Dentry, b *fspath
 		if mnt == vfsroot.Mount() && &d.vfsd == vfsroot.Dentry() {
 			return vfs.PrependPathAtVFSRootError{}
 		}
-		if &d.vfsd == mnt.Root() {
+		if mnt != nil && &d.vfsd == mnt.Root() {
 			return nil
 		}
 		if d.parent == nil {
@@ -81,3 +81,12 @@ func PrependPath(vfsroot vfs.VirtualDentry, mnt *vfs.Mount, d *Dentry, b *fspath
 		d = d.parent
 	}
 }
+
+// DebugPathname returns a pathname to d relative to its filesystem root.
+// DebugPathname does not correspond to any Linux function; it's used to
+// generate dentry pathnames for debugging.
+func DebugPathname(d *Dentry) string {
+	var b fspath.Builder
+	_ = PrependPath(vfs.VirtualDentry{}, nil, d, &b)
+	return b.String()
+}
diff --git a/pkg/sentry/vfs/lock.go b/pkg/sentry/vfs/lock.go
index 55783d4eb..1ff202f2a 100644
--- a/pkg/sentry/vfs/lock.go
+++ b/pkg/sentry/vfs/lock.go
@@ -12,11 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// Package lock provides POSIX and BSD style file locking for VFS2 file
-// implementations.
-//
-// The actual implementations can be found in the lock package under
-// sentry/fs/lock.
 package vfs
 
 import (
diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go
index b7d122d22..cb48c37a1 100644
--- a/pkg/sentry/vfs/mount_unsafe.go
+++ b/pkg/sentry/vfs/mount_unsafe.go
@@ -98,7 +98,6 @@ type mountTable struct {
 	// length and cap in separate uint32s) for ~free.
 	size uint64
 
-	// FIXME(gvisor.dev/issue/1663): Slots need to be saved.
 	slots unsafe.Pointer `state:"nosave"` // []mountSlot; never nil after Init
 }
 
@@ -212,6 +211,26 @@ loop:
 	}
 }
 
+// Range calls f on each Mount in mt. If f returns false, Range stops iteration
+// and returns immediately.
+func (mt *mountTable) Range(f func(*Mount) bool) {
+	tcap := uintptr(1) << (mt.size & mtSizeOrderMask)
+	slotPtr := mt.slots
+	last := unsafe.Pointer(uintptr(mt.slots) + ((tcap - 1) * mountSlotBytes))
+	for {
+		slot := (*mountSlot)(slotPtr)
+		if slot.value != nil {
+			if !f((*Mount)(slot.value)) {
+				return
+			}
+		}
+		if slotPtr == last {
+			return
+		}
+		slotPtr = unsafe.Pointer(uintptr(slotPtr) + mountSlotBytes)
+	}
+}
+
 // Insert inserts the given mount into mt.
 //
 // Preconditions: mt must not already contain a Mount with the same mount point
diff --git a/pkg/sentry/vfs/save_restore.go b/pkg/sentry/vfs/save_restore.go
new file mode 100644
index 000000000..7aa073510
--- /dev/null
+++ b/pkg/sentry/vfs/save_restore.go
@@ -0,0 +1,117 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package vfs
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/context"
+)
+
+// FilesystemImplSaveRestoreExtension is an optional extension to
+// FilesystemImpl.
+type FilesystemImplSaveRestoreExtension interface {
+	// PrepareSave prepares this filesystem for serialization.
+	PrepareSave(ctx context.Context) error
+
+	// CompleteRestore completes restoration from checkpoint for this
+	// filesystem after deserialization.
+	CompleteRestore(ctx context.Context, opts CompleteRestoreOptions) error
+}
+
+// PrepareSave prepares all filesystems for serialization.
+func (vfs *VirtualFilesystem) PrepareSave(ctx context.Context) error {
+	failures := 0
+	for fs := range vfs.getFilesystems() {
+		if ext, ok := fs.impl.(FilesystemImplSaveRestoreExtension); ok {
+			if err := ext.PrepareSave(ctx); err != nil {
+				ctx.Warningf("%T.PrepareSave failed: %v", fs.impl, err)
+				failures++
+			}
+		}
+		fs.DecRef(ctx)
+	}
+	if failures != 0 {
+		return fmt.Errorf("%d filesystems failed to prepare for serialization", failures)
+	}
+	return nil
+}
+
+// CompleteRestore completes restoration from checkpoint for all filesystems
+// after deserialization.
+func (vfs *VirtualFilesystem) CompleteRestore(ctx context.Context, opts *CompleteRestoreOptions) error {
+	failures := 0
+	for fs := range vfs.getFilesystems() {
+		if ext, ok := fs.impl.(FilesystemImplSaveRestoreExtension); ok {
+			if err := ext.CompleteRestore(ctx, *opts); err != nil {
+				ctx.Warningf("%T.CompleteRestore failed: %v", fs.impl, err)
+				failures++
+			}
+		}
+		fs.DecRef(ctx)
+	}
+	if failures != 0 {
+		return fmt.Errorf("%d filesystems failed to complete restore after deserialization", failures)
+	}
+	return nil
+}
+
+// CompleteRestoreOptions contains options to
+// VirtualFilesystem.CompleteRestore() and
+// FilesystemImplSaveRestoreExtension.CompleteRestore().
+type CompleteRestoreOptions struct {
+	// If ValidateFileSizes is true, filesystem implementations backed by
+	// remote filesystems should verify that file sizes have not changed
+	// between checkpoint and restore.
+	ValidateFileSizes bool
+
+	// If ValidateFileModificationTimestamps is true, filesystem
+	// implementations backed by remote filesystems should validate that file
+	// mtimes have not changed between checkpoint and restore.
+	ValidateFileModificationTimestamps bool
+}
+
+// saveMounts is called by stateify.
+func (vfs *VirtualFilesystem) saveMounts() []*Mount {
+	if atomic.LoadPointer(&vfs.mounts.slots) == nil {
+		// vfs.Init() was never called.
+		return nil
+	}
+	var mounts []*Mount
+	vfs.mounts.Range(func(mount *Mount) bool {
+		mounts = append(mounts, mount)
+		return true
+	})
+	return mounts
+}
+
+// loadMounts is called by stateify.
+func (vfs *VirtualFilesystem) loadMounts(mounts []*Mount) {
+	if mounts == nil {
+		return
+	}
+	vfs.mounts.Init()
+	for _, mount := range mounts {
+		vfs.mounts.Insert(mount)
+	}
+}
+
+// afterLoad is called by stateify.
+func (epi *epollInterest) afterLoad() {
+	// Mark all epollInterests as ready after restore so that the next call to
+	// EpollInstance.ReadEvents() rechecks their readiness.
+	epi.Callback(nil)
+}
diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go
index 38d2701d2..48d6252f7 100644
--- a/pkg/sentry/vfs/vfs.go
+++ b/pkg/sentry/vfs/vfs.go
@@ -71,7 +71,7 @@ type VirtualFilesystem struct {
 	// points.
 	//
 	// mounts is analogous to Linux's mount_hashtable.
-	mounts mountTable
+	mounts mountTable `state:".([]*Mount)"`
 
 	// mountpoints maps mount points to mounts at those points in all
 	// namespaces. mountpoints is protected by mountMu.
@@ -780,23 +780,27 @@ func (vfs *VirtualFilesystem) RemoveXattrAt(ctx context.Context, creds *auth.Cre
 
 // SyncAllFilesystems has the semantics of Linux's sync(2).
 func (vfs *VirtualFilesystem) SyncAllFilesystems(ctx context.Context) error {
+	var retErr error
+	for fs := range vfs.getFilesystems() {
+		if err := fs.impl.Sync(ctx); err != nil && retErr == nil {
+			retErr = err
+		}
+		fs.DecRef(ctx)
+	}
+	return retErr
+}
+
+func (vfs *VirtualFilesystem) getFilesystems() map[*Filesystem]struct{} {
 	fss := make(map[*Filesystem]struct{})
 	vfs.filesystemsMu.Lock()
+	defer vfs.filesystemsMu.Unlock()
 	for fs := range vfs.filesystems {
 		if !fs.TryIncRef() {
 			continue
 		}
 		fss[fs] = struct{}{}
 	}
-	vfs.filesystemsMu.Unlock()
-	var retErr error
-	for fs := range fss {
-		if err := fs.impl.Sync(ctx); err != nil && retErr == nil {
-			retErr = err
-		}
-		fs.DecRef(ctx)
-	}
-	return retErr
+	return fss
 }
 
 // MkdirAllAt recursively creates non-existent directories on the given path