Port netstack, hostinet, and netlink sockets to VFS2.

All three follow the same pattern: 1. Refactor VFS1 sockets into socketOpsCommon, so that most of the methods can be shared with VFS2. 2. Create a FileDescriptionImpl with the corresponding socket operations, rewriting the few that cannot be shared with VFS1. 3. Set up a VFS2 socket provider that creates a socket by setting up a dentry in the global Kernel.socketMount and connecting it with a new FileDescription. This mostly completes the work for porting sockets to VFS2, and many syscall tests can be enabled as a result. There are several networking-related syscall tests that are still not passing: 1. net gofer tests 2. socketpair gofer tests 2. sendfile tests (splice is not implemented in VFS2 yet) Updates #1478, #1484, #1485 PiperOrigin-RevId: 309457331
2026-05-22 17:12:49 -07:00 · 2020-05-01 12:54:41 -07:00
parent 8962b7840f
commit 82bae30cee
19 changed files with 1090 additions and 126 deletions
@@ -7,7 +7,9 @@ go_library(
    srcs = [
        "device.go",
        "netstack.go",
+        "netstack_vfs2.go",
        "provider.go",
+        "provider_vfs2.go",
        "save_restore.go",
        "stack.go",
    ],
@@ -25,6 +27,8 @@ go_library(
        "//pkg/sentry/device",
        "//pkg/sentry/fs",
        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fsimpl/kernfs",
+        "//pkg/sentry/fsimpl/sockfs",
        "//pkg/sentry/inet",
        "//pkg/sentry/kernel",
        "//pkg/sentry/kernel/auth",
@@ -32,6 +36,7 @@ go_library(
        "//pkg/sentry/socket",
        "//pkg/sentry/socket/netfilter",
        "//pkg/sentry/unimpl",
+        "//pkg/sentry/vfs",
        "//pkg/sync",
        "//pkg/syserr",
        "//pkg/syserror",
@@ -252,6 +252,8 @@ type commonEndpoint interface {
 	GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error)
 }

+// LINT.IfChange
+
 // SocketOperations encapsulates all the state needed to represent a network stack
 // endpoint in the kernel context.
 //
@@ -263,6 +265,14 @@ type SocketOperations struct {
 	fsutil.FileNoFsync              `state:"nosave"`
 	fsutil.FileNoMMap               `state:"nosave"`
 	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
+
+	socketOpsCommon
+}
+
+// socketOpsCommon contains the socket operations common to VFS1 and VFS2.
+//
+// +stateify savable
+type socketOpsCommon struct {
 	socket.SendReceiveTimeout
 	*waiter.Queue

@@ -314,11 +324,13 @@ func New(t *kernel.Task, family int, skType linux.SockType, protocol int, queue
 	dirent := socket.NewDirent(t, netstackDevice)
 	defer dirent.DecRef()
 	return fs.NewFile(t, dirent, fs.FileFlags{Read: true, Write: true, NonSeekable: true}, &SocketOperations{
-		Queue:    queue,
-		family:   family,
-		Endpoint: endpoint,
-		skType:   skType,
-		protocol: protocol,
+		socketOpsCommon: socketOpsCommon{
+			Queue:    queue,
+			family:   family,
+			Endpoint: endpoint,
+			skType:   skType,
+			protocol: protocol,
+		},
 	}), nil
 }

@@ -417,7 +429,7 @@ func AddressAndFamily(addr []byte) (tcpip.FullAddress, uint16, *syserr.Error) {
 	}
 }

-func (s *SocketOperations) isPacketBased() bool {
+func (s *socketOpsCommon) isPacketBased() bool {
 	return s.skType == linux.SOCK_DGRAM || s.skType == linux.SOCK_SEQPACKET || s.skType == linux.SOCK_RDM || s.skType == linux.SOCK_RAW
 }

@@ -425,7 +437,7 @@ func (s *SocketOperations) isPacketBased() bool {
 // empty. It assumes that the socket is locked.
 //
 // Precondition: s.readMu must be held.
-func (s *SocketOperations) fetchReadView() *syserr.Error {
+func (s *socketOpsCommon) fetchReadView() *syserr.Error {
 	if len(s.readView) > 0 {
 		return nil
 	}
@@ -446,7 +458,7 @@ func (s *SocketOperations) fetchReadView() *syserr.Error {
 }

 // Release implements fs.FileOperations.Release.
-func (s *SocketOperations) Release() {
+func (s *socketOpsCommon) Release() {
 	s.Endpoint.Close()
 }

@@ -633,7 +645,7 @@ func (s *SocketOperations) ReadFrom(ctx context.Context, _ *fs.File, r io.Reader
 }

 // Readiness returns a mask of ready events for socket s.
-func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+func (s *socketOpsCommon) Readiness(mask waiter.EventMask) waiter.EventMask {
 	r := s.Endpoint.Readiness(mask)

 	// Check our cached value iff the caller asked for readability and the
@@ -647,7 +659,7 @@ func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
 	return r
 }

-func (s *SocketOperations) checkFamily(family uint16, exact bool) *syserr.Error {
+func (s *socketOpsCommon) checkFamily(family uint16, exact bool) *syserr.Error {
 	if family == uint16(s.family) {
 		return nil
 	}
@@ -670,7 +682,7 @@ func (s *SocketOperations) checkFamily(family uint16, exact bool) *syserr.Error
 // represented by the empty string.
 //
 // TODO(gvisor.dev/issue/1556): remove this function.
-func (s *SocketOperations) mapFamily(addr tcpip.FullAddress, family uint16) tcpip.FullAddress {
+func (s *socketOpsCommon) mapFamily(addr tcpip.FullAddress, family uint16) tcpip.FullAddress {
 	if len(addr.Addr) == 0 && s.family == linux.AF_INET6 && family == linux.AF_INET {
 		addr.Addr = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\xff\xff\x00\x00\x00\x00"
 	}
@@ -679,7 +691,7 @@ func (s *SocketOperations) mapFamily(addr tcpip.FullAddress, family uint16) tcpi

 // Connect implements the linux syscall connect(2) for sockets backed by
 // tpcip.Endpoint.
-func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+func (s *socketOpsCommon) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
 	addr, family, err := AddressAndFamily(sockaddr)
 	if err != nil {
 		return err
@@ -725,7 +737,7 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo

 // Bind implements the linux syscall bind(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+func (s *socketOpsCommon) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
 	if len(sockaddr) < 2 {
 		return syserr.ErrInvalidArgument
 	}
@@ -771,13 +783,13 @@ func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {

 // Listen implements the linux syscall listen(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error {
+func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error {
 	return syserr.TranslateNetstackError(s.Endpoint.Listen(backlog))
 }

 // blockingAccept implements a blocking version of accept(2), that is, if no
 // connections are ready to be accept, it will block until one becomes ready.
-func (s *SocketOperations) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) {
+func (s *socketOpsCommon) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) {
 	// Register for notifications.
 	e, ch := waiter.NewChannelEntry(nil)
 	s.EventRegister(&e, waiter.EventIn)
@@ -863,7 +875,7 @@ func ConvertShutdown(how int) (tcpip.ShutdownFlags, *syserr.Error) {

 // Shutdown implements the linux syscall shutdown(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error {
+func (s *socketOpsCommon) Shutdown(t *kernel.Task, how int) *syserr.Error {
 	f, err := ConvertShutdown(how)
 	if err != nil {
 		return err
@@ -2258,7 +2270,7 @@ func ConvertAddress(family int, addr tcpip.FullAddress) (linux.SockAddr, uint32)

 // GetSockName implements the linux syscall getsockname(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) {
+func (s *socketOpsCommon) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) {
 	addr, err := s.Endpoint.GetLocalAddress()
 	if err != nil {
 		return nil, 0, syserr.TranslateNetstackError(err)
@@ -2270,7 +2282,7 @@ func (s *SocketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32,

 // GetPeerName implements the linux syscall getpeername(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) {
+func (s *socketOpsCommon) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) {
 	addr, err := s.Endpoint.GetRemoteAddress()
 	if err != nil {
 		return nil, 0, syserr.TranslateNetstackError(err)
@@ -2285,7 +2297,7 @@ func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32,
 // caller.
 //
 // Precondition: s.readMu must be locked.
-func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSequence, discard bool) (int, *syserr.Error) {
+func (s *socketOpsCommon) coalescingRead(ctx context.Context, dst usermem.IOSequence, discard bool) (int, *syserr.Error) {
 	var err *syserr.Error
 	var copied int

@@ -2337,7 +2349,7 @@ func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSeq
 	return 0, err
 }

-func (s *SocketOperations) fillCmsgInq(cmsg *socket.ControlMessages) {
+func (s *socketOpsCommon) fillCmsgInq(cmsg *socket.ControlMessages) {
 	if !s.sockOptInq {
 		return
 	}
@@ -2352,7 +2364,7 @@ func (s *SocketOperations) fillCmsgInq(cmsg *socket.ControlMessages) {
 // nonBlockingRead issues a non-blocking read.
 //
 // TODO(b/78348848): Support timestamps for stream sockets.
-func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSequence, peek, trunc, senderRequested bool) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) {
+func (s *socketOpsCommon) nonBlockingRead(ctx context.Context, dst usermem.IOSequence, peek, trunc, senderRequested bool) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) {
 	isPacket := s.isPacketBased()

 	// Fast path for regular reads from stream (e.g., TCP) endpoints. Note
@@ -2461,7 +2473,7 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe
 	return n, flags, addr, addrLen, cmsg, syserr.FromError(err)
 }

-func (s *SocketOperations) controlMessages() socket.ControlMessages {
+func (s *socketOpsCommon) controlMessages() socket.ControlMessages {
 	return socket.ControlMessages{
 		IP: tcpip.ControlMessages{
 			HasTimestamp:    s.readCM.HasTimestamp && s.sockOptTimestamp,
@@ -2480,7 +2492,7 @@ func (s *SocketOperations) controlMessages() socket.ControlMessages {
 // successfully writing packet data out to userspace.
 //
 // Precondition: s.readMu must be locked.
-func (s *SocketOperations) updateTimestamp() {
+func (s *socketOpsCommon) updateTimestamp() {
 	// Save the SIOCGSTAMP timestamp only if SO_TIMESTAMP is disabled.
 	if !s.sockOptTimestamp {
 		s.timestampValid = true
@@ -2490,7 +2502,7 @@ func (s *SocketOperations) updateTimestamp() {

 // RecvMsg implements the linux syscall recvmsg(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) {
+func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) {
 	trunc := flags&linux.MSG_TRUNC != 0
 	peek := flags&linux.MSG_PEEK != 0
 	dontWait := flags&linux.MSG_DONTWAIT != 0
@@ -2558,7 +2570,7 @@ func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags

 // SendMsg implements the linux syscall sendmsg(2) for sockets backed by
 // tcpip.Endpoint.
-func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) {
+func (s *socketOpsCommon) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) {
 	// Reject Unix control messages.
 	if !controlMessages.Unix.Empty() {
 		return 0, syserr.ErrInvalidArgument
@@ -2634,6 +2646,10 @@ func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []

 // Ioctl implements fs.FileOperations.Ioctl.
 func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	return s.socketOpsCommon.ioctl(ctx, io, args)
+}
+
+func (s *socketOpsCommon) ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
 	// SIOCGSTAMP is implemented by netstack rather than all commonEndpoint
 	// sockets.
 	// TODO(b/78348848): Add a commonEndpoint method to support SIOCGSTAMP.
@@ -2973,7 +2989,7 @@ func nicStateFlagsToLinux(f stack.NICStateFlags) uint32 {

 // State implements socket.Socket.State. State translates the internal state
 // returned by netstack to values defined by Linux.
-func (s *SocketOperations) State() uint32 {
+func (s *socketOpsCommon) State() uint32 {
 	if s.family != linux.AF_INET && s.family != linux.AF_INET6 {
 		// States not implemented for this socket's family.
 		return 0
@@ -3033,6 +3049,8 @@ func (s *SocketOperations) State() uint32 {
 }

 // Type implements socket.Socket.Type.
-func (s *SocketOperations) Type() (family int, skType linux.SockType, protocol int) {
+func (s *socketOpsCommon) Type() (family int, skType linux.SockType, protocol int) {
 	return s.family, s.skType, s.protocol
 }
+
+// LINT.ThenChange(./netstack_vfs2.go)
@@ -0,0 +1,327 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package netstack
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs"
+	"gvisor.dev/gvisor/pkg/sentry/inet"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/socket/netfilter"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserr"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/tcpip"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// SocketVFS2 encapsulates all the state needed to represent a network stack
+// endpoint in the kernel context.
+type SocketVFS2 struct {
+	vfsfd vfs.FileDescription
+	vfs.FileDescriptionDefaultImpl
+	vfs.DentryMetadataFileDescriptionImpl
+
+	socketOpsCommon
+}
+
+// NewVFS2 creates a new endpoint socket.
+func NewVFS2(t *kernel.Task, family int, skType linux.SockType, protocol int, queue *waiter.Queue, endpoint tcpip.Endpoint) (*vfs.FileDescription, *syserr.Error) {
+	if skType == linux.SOCK_STREAM {
+		if err := endpoint.SetSockOptBool(tcpip.DelayOption, true); err != nil {
+			return nil, syserr.TranslateNetstackError(err)
+		}
+	}
+
+	mnt := t.Kernel().SocketMount()
+	fs := mnt.Filesystem().Impl().(*kernfs.Filesystem)
+	d := sockfs.NewDentry(t.Credentials(), fs.NextIno())
+
+	s := &SocketVFS2{
+		socketOpsCommon: socketOpsCommon{
+			Queue:    queue,
+			family:   family,
+			Endpoint: endpoint,
+			skType:   skType,
+			protocol: protocol,
+		},
+	}
+	vfsfd := &s.vfsfd
+	if err := vfsfd.Init(s, linux.O_RDWR, mnt, d, &vfs.FileDescriptionOptions{
+		DenyPRead:         true,
+		DenyPWrite:        true,
+		UseDentryMetadata: true,
+	}); err != nil {
+		return nil, syserr.FromError(err)
+	}
+	return vfsfd, nil
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return s.socketOpsCommon.Readiness(mask)
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (s *SocketVFS2) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	s.socketOpsCommon.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (s *SocketVFS2) EventUnregister(e *waiter.Entry) {
+	s.socketOpsCommon.EventUnregister(e)
+}
+
+// PRead implements vfs.FileDescriptionImpl.
+func (s *SocketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	return 0, syserror.ESPIPE
+}
+
+// Read implements vfs.FileDescriptionImpl.
+func (s *SocketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	// All flags other than RWF_NOWAIT should be ignored.
+	// TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT.
+	if opts.Flags != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+	n, _, _, _, _, err := s.nonBlockingRead(ctx, dst, false, false, false)
+	if err == syserr.ErrWouldBlock {
+		return int64(n), syserror.ErrWouldBlock
+	}
+	if err != nil {
+		return 0, err.ToError()
+	}
+	return int64(n), nil
+}
+
+// PWrite implements vfs.FileDescriptionImpl.
+func (s *SocketVFS2) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+	return 0, syserror.ESPIPE
+}
+
+// Write implements vfs.FileDescriptionImpl.
+func (s *SocketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	// All flags other than RWF_NOWAIT should be ignored.
+	// TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT.
+	if opts.Flags != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	f := &ioSequencePayload{ctx: ctx, src: src}
+	n, resCh, err := s.Endpoint.Write(f, tcpip.WriteOptions{})
+	if err == tcpip.ErrWouldBlock {
+		return 0, syserror.ErrWouldBlock
+	}
+
+	if resCh != nil {
+		t := kernel.TaskFromContext(ctx)
+		if err := t.Block(resCh); err != nil {
+			return 0, syserr.FromError(err).ToError()
+		}
+
+		n, _, err = s.Endpoint.Write(f, tcpip.WriteOptions{})
+	}
+
+	if err != nil {
+		return 0, syserr.TranslateNetstackError(err).ToError()
+	}
+
+	if int64(n) < src.NumBytes() {
+		return int64(n), syserror.ErrWouldBlock
+	}
+
+	return int64(n), nil
+}
+
+// Accept implements the linux syscall accept(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) {
+	// Issue the accept request to get the new endpoint.
+	ep, wq, terr := s.Endpoint.Accept()
+	if terr != nil {
+		if terr != tcpip.ErrWouldBlock || !blocking {
+			return 0, nil, 0, syserr.TranslateNetstackError(terr)
+		}
+
+		var err *syserr.Error
+		ep, wq, err = s.blockingAccept(t)
+		if err != nil {
+			return 0, nil, 0, err
+		}
+	}
+
+	ns, err := NewVFS2(t, s.family, s.skType, s.protocol, wq, ep)
+	if err != nil {
+		return 0, nil, 0, err
+	}
+	defer ns.DecRef()
+
+	if err := ns.SetStatusFlags(t, t.Credentials(), uint32(flags&linux.SOCK_NONBLOCK)); err != nil {
+		return 0, nil, 0, syserr.FromError(err)
+	}
+
+	var addr linux.SockAddr
+	var addrLen uint32
+	if peerRequested {
+		// Get address of the peer and write it to peer slice.
+		var err *syserr.Error
+		addr, addrLen, err = ns.Impl().(*SocketVFS2).GetPeerName(t)
+		if err != nil {
+			return 0, nil, 0, err
+		}
+	}
+
+	fd, e := t.NewFDFromVFS2(0, ns, kernel.FDFlags{
+		CloseOnExec: flags&linux.SOCK_CLOEXEC != 0,
+	})
+
+	t.Kernel().RecordSocketVFS2(ns)
+
+	return fd, addr, addrLen, syserr.FromError(e)
+}
+
+// Ioctl implements vfs.FileDescriptionImpl.
+func (s *SocketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	return s.socketOpsCommon.ioctl(ctx, uio, args)
+}
+
+// GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) {
+	// TODO(b/78348848): Unlike other socket options, SO_TIMESTAMP is
+	// implemented specifically for netstack.SocketVFS2 rather than
+	// commonEndpoint. commonEndpoint should be extended to support socket
+	// options where the implementation is not shared, as unix sockets need
+	// their own support for SO_TIMESTAMP.
+	if level == linux.SOL_SOCKET && name == linux.SO_TIMESTAMP {
+		if outLen < sizeOfInt32 {
+			return nil, syserr.ErrInvalidArgument
+		}
+		val := int32(0)
+		s.readMu.Lock()
+		defer s.readMu.Unlock()
+		if s.sockOptTimestamp {
+			val = 1
+		}
+		return val, nil
+	}
+	if level == linux.SOL_TCP && name == linux.TCP_INQ {
+		if outLen < sizeOfInt32 {
+			return nil, syserr.ErrInvalidArgument
+		}
+		val := int32(0)
+		s.readMu.Lock()
+		defer s.readMu.Unlock()
+		if s.sockOptInq {
+			val = 1
+		}
+		return val, nil
+	}
+
+	if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP {
+		switch name {
+		case linux.IPT_SO_GET_INFO:
+			if outLen < linux.SizeOfIPTGetinfo {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			stack := inet.StackFromContext(t)
+			if stack == nil {
+				return nil, syserr.ErrNoDevice
+			}
+			info, err := netfilter.GetInfo(t, stack.(*Stack).Stack, outPtr)
+			if err != nil {
+				return nil, err
+			}
+			return info, nil
+
+		case linux.IPT_SO_GET_ENTRIES:
+			if outLen < linux.SizeOfIPTGetEntries {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			stack := inet.StackFromContext(t)
+			if stack == nil {
+				return nil, syserr.ErrNoDevice
+			}
+			entries, err := netfilter.GetEntries(t, stack.(*Stack).Stack, outPtr, outLen)
+			if err != nil {
+				return nil, err
+			}
+			return entries, nil
+
+		}
+	}
+
+	return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outLen)
+}
+
+// SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketVFS2) SetSockOpt(t *kernel.Task, level int, name int, optVal []byte) *syserr.Error {
+	// TODO(b/78348848): Unlike other socket options, SO_TIMESTAMP is
+	// implemented specifically for netstack.SocketVFS2 rather than
+	// commonEndpoint. commonEndpoint should be extended to support socket
+	// options where the implementation is not shared, as unix sockets need
+	// their own support for SO_TIMESTAMP.
+	if level == linux.SOL_SOCKET && name == linux.SO_TIMESTAMP {
+		if len(optVal) < sizeOfInt32 {
+			return syserr.ErrInvalidArgument
+		}
+		s.readMu.Lock()
+		defer s.readMu.Unlock()
+		s.sockOptTimestamp = usermem.ByteOrder.Uint32(optVal) != 0
+		return nil
+	}
+	if level == linux.SOL_TCP && name == linux.TCP_INQ {
+		if len(optVal) < sizeOfInt32 {
+			return syserr.ErrInvalidArgument
+		}
+		s.readMu.Lock()
+		defer s.readMu.Unlock()
+		s.sockOptInq = usermem.ByteOrder.Uint32(optVal) != 0
+		return nil
+	}
+
+	if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP {
+		switch name {
+		case linux.IPT_SO_SET_REPLACE:
+			if len(optVal) < linux.SizeOfIPTReplace {
+				return syserr.ErrInvalidArgument
+			}
+
+			stack := inet.StackFromContext(t)
+			if stack == nil {
+				return syserr.ErrNoDevice
+			}
+			// Stack must be a netstack stack.
+			return netfilter.SetEntries(stack.(*Stack).Stack, optVal)
+
+		case linux.IPT_SO_SET_ADD_COUNTERS:
+			// TODO(gvisor.dev/issue/170): Counter support.
+			return nil
+		}
+	}
+
+	return SetSockOpt(t, s, s.Endpoint, level, name, optVal)
+}
@@ -33,6 +33,8 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )

+// LINT.IfChange
+
 // provider is an inet socket provider.
 type provider struct {
 	family   int
@@ -167,6 +169,8 @@ func packetSocket(t *kernel.Task, epStack *Stack, stype linux.SockType, protocol
 	return New(t, linux.AF_PACKET, stype, protocol, wq, ep)
 }

+// LINT.ThenChange(./provider_vfs2.go)
+
 // Pair just returns nil sockets (not supported).
 func (*provider) Pair(*kernel.Task, linux.SockType, int) (*fs.File, *fs.File, *syserr.Error) {
 	return nil, nil, nil
@@ -0,0 +1,141 @@
+// 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 netstack
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/socket"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserr"
+	"gvisor.dev/gvisor/pkg/tcpip"
+	"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
+	"gvisor.dev/gvisor/pkg/tcpip/network/ipv6"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// providerVFS2 is an inet socket provider.
+type providerVFS2 struct {
+	family   int
+	netProto tcpip.NetworkProtocolNumber
+}
+
+// Socket creates a new socket object for the AF_INET, AF_INET6, or AF_PACKET
+// family.
+func (p *providerVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) {
+	// Fail right away if we don't have a stack.
+	stack := t.NetworkContext()
+	if stack == nil {
+		// Don't propagate an error here. Instead, allow the socket
+		// code to continue searching for another provider.
+		return nil, nil
+	}
+	eps, ok := stack.(*Stack)
+	if !ok {
+		return nil, nil
+	}
+
+	// Packet sockets are handled separately, since they are neither INET
+	// nor INET6 specific.
+	if p.family == linux.AF_PACKET {
+		return packetSocketVFS2(t, eps, stype, protocol)
+	}
+
+	// Figure out the transport protocol.
+	transProto, associated, err := getTransportProtocol(t, stype, protocol)
+	if err != nil {
+		return nil, err
+	}
+
+	// Create the endpoint.
+	var ep tcpip.Endpoint
+	var e *tcpip.Error
+	wq := &waiter.Queue{}
+	if stype == linux.SOCK_RAW {
+		ep, e = eps.Stack.NewRawEndpoint(transProto, p.netProto, wq, associated)
+	} else {
+		ep, e = eps.Stack.NewEndpoint(transProto, p.netProto, wq)
+
+		// Assign task to PacketOwner interface to get the UID and GID for
+		// iptables owner matching.
+		if e == nil {
+			ep.SetOwner(t)
+		}
+	}
+	if e != nil {
+		return nil, syserr.TranslateNetstackError(e)
+	}
+
+	return NewVFS2(t, p.family, stype, int(transProto), wq, ep)
+}
+
+func packetSocketVFS2(t *kernel.Task, epStack *Stack, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) {
+	// Packet sockets require CAP_NET_RAW.
+	creds := auth.CredentialsFromContext(t)
+	if !creds.HasCapability(linux.CAP_NET_RAW) {
+		return nil, syserr.ErrNotPermitted
+	}
+
+	// "cooked" packets don't contain link layer information.
+	var cooked bool
+	switch stype {
+	case linux.SOCK_DGRAM:
+		cooked = true
+	case linux.SOCK_RAW:
+		cooked = false
+	default:
+		return nil, syserr.ErrProtocolNotSupported
+	}
+
+	// protocol is passed in network byte order, but netstack wants it in
+	// host order.
+	netProto := tcpip.NetworkProtocolNumber(ntohs(uint16(protocol)))
+
+	wq := &waiter.Queue{}
+	ep, err := epStack.Stack.NewPacketEndpoint(cooked, netProto, wq)
+	if err != nil {
+		return nil, syserr.TranslateNetstackError(err)
+	}
+
+	return NewVFS2(t, linux.AF_PACKET, stype, protocol, wq, ep)
+}
+
+// Pair just returns nil sockets (not supported).
+func (*providerVFS2) Pair(*kernel.Task, linux.SockType, int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) {
+	return nil, nil, nil
+}
+
+// init registers socket providers for AF_INET, AF_INET6, and AF_PACKET.
+func init() {
+	// Providers backed by netstack.
+	p := []providerVFS2{
+		{
+			family:   linux.AF_INET,
+			netProto: ipv4.ProtocolNumber,
+		},
+		{
+			family:   linux.AF_INET6,
+			netProto: ipv6.ProtocolNumber,
+		},
+		{
+			family: linux.AF_PACKET,
+		},
+	}
+
+	for i := range p {
+		socket.RegisterProviderVFS2(p[i].family, &p[i])
+	}
+}