net/tstun: restore tap mode functionality

It had bit-rotted likely during the transition to vector io in
76389d8baf. Tested on Ubuntu 24.04
by creating a netns and doing the DHCP dance to get an IP.

Updates #2589

Signed-off-by: Maisem Ali <maisem@tailscale.com>

cmd/k8s-operator/depaware.txt

@@ -310,7 +310,7 @@ tailscale.com/cmd/k8s-operator dependencies: (generated by github.com/tailscale/
         gvisor.dev/gvisor/pkg/tcpip/network/internal/ip              from gvisor.dev/gvisor/pkg/tcpip/network/ipv4+
         gvisor.dev/gvisor/pkg/tcpip/network/internal/multicast       from gvisor.dev/gvisor/pkg/tcpip/network/ipv4+
         gvisor.dev/gvisor/pkg/tcpip/network/ipv4                     from tailscale.com/net/tstun+
-        gvisor.dev/gvisor/pkg/tcpip/network/ipv6                     from tailscale.com/wgengine/netstack
+        gvisor.dev/gvisor/pkg/tcpip/network/ipv6                     from tailscale.com/wgengine/netstack+
         gvisor.dev/gvisor/pkg/tcpip/ports                            from gvisor.dev/gvisor/pkg/tcpip/stack+
         gvisor.dev/gvisor/pkg/tcpip/seqnum                           from gvisor.dev/gvisor/pkg/tcpip/header+
      💣 gvisor.dev/gvisor/pkg/tcpip/stack                            from gvisor.dev/gvisor/pkg/tcpip/adapters/gonet+

cmd/tailscaled/depaware.txt

@@ -221,7 +221,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         gvisor.dev/gvisor/pkg/tcpip/network/internal/ip              from gvisor.dev/gvisor/pkg/tcpip/network/ipv4+
         gvisor.dev/gvisor/pkg/tcpip/network/internal/multicast       from gvisor.dev/gvisor/pkg/tcpip/network/ipv4+
         gvisor.dev/gvisor/pkg/tcpip/network/ipv4                     from tailscale.com/net/tstun+
-        gvisor.dev/gvisor/pkg/tcpip/network/ipv6                     from tailscale.com/wgengine/netstack
+        gvisor.dev/gvisor/pkg/tcpip/network/ipv6                     from tailscale.com/wgengine/netstack+
         gvisor.dev/gvisor/pkg/tcpip/ports                            from gvisor.dev/gvisor/pkg/tcpip/stack+
         gvisor.dev/gvisor/pkg/tcpip/seqnum                           from gvisor.dev/gvisor/pkg/tcpip/header+
      💣 gvisor.dev/gvisor/pkg/tcpip/stack                            from gvisor.dev/gvisor/pkg/tcpip/adapters/gonet+

net/tstun/tap_linux.go

@@ -6,6 +6,7 @@
 package tstun
 
 import (
+	"bytes"
 	"fmt"
 	"net"
 	"net/netip"
@@ -20,10 +21,13 @@ import (
 	"gvisor.dev/gvisor/pkg/tcpip/checksum"
 	"gvisor.dev/gvisor/pkg/tcpip/header"
 	"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
+	"gvisor.dev/gvisor/pkg/tcpip/network/ipv6"
 	"gvisor.dev/gvisor/pkg/tcpip/transport/udp"
 	"tailscale.com/net/netaddr"
 	"tailscale.com/net/packet"
+	"tailscale.com/syncs"
 	"tailscale.com/types/ipproto"
+	"tailscale.com/types/logger"
 	"tailscale.com/util/multierr"
 )
 
@@ -35,13 +39,13 @@ var ourMAC = net.HardwareAddr{0x30, 0x2D, 0x66, 0xEC, 0x7A, 0x93}
 
 func init() { createTAP = createTAPLinux }
 
-func createTAPLinux(tapName, bridgeName string) (tun.Device, error) {
+func createTAPLinux(logf logger.Logf, tapName, bridgeName string) (tun.Device, error) {
 	fd, err := unix.Open("/dev/net/tun", unix.O_RDWR, 0)
 	if err != nil {
 		return nil, err
 	}
 
-	dev, err := openDevice(fd, tapName, bridgeName)
+	dev, err := openDevice(logf, fd, tapName, bridgeName)
 	if err != nil {
 		unix.Close(fd)
 		return nil, err
@@ -50,7 +54,7 @@ func createTAPLinux(tapName, bridgeName string) (tun.Device, error) {
 	return dev, nil
 }
 
-func openDevice(fd int, tapName, bridgeName string) (tun.Device, error) {
+func openDevice(logf logger.Logf, fd int, tapName, bridgeName string) (tun.Device, error) {
 	ifr, err := unix.NewIfreq(tapName)
 	if err != nil {
 		return nil, err
@@ -71,7 +75,7 @@ func openDevice(fd int, tapName, bridgeName string) (tun.Device, error) {
 		}
 	}
 
-	return newTAPDevice(fd, tapName)
+	return newTAPDevice(logf, fd, tapName)
 }
 
 type etherType [2]byte
@@ -91,7 +95,7 @@ const (
 
 // handleTAPFrame handles receiving a raw TAP ethernet frame and reports whether
 // it's been handled (that is, whether it should NOT be passed to wireguard).
-func (t *Wrapper) handleTAPFrame(ethBuf []byte) bool {
+func (t *tapDevice) handleTAPFrame(ethBuf []byte) bool {
 
 	if len(ethBuf) < ethernetFrameSize {
 		// Corrupt. Ignore.
@@ -164,8 +168,7 @@ func (t *Wrapper) handleTAPFrame(ethBuf []byte) bool {
 			copy(res.HardwareAddressTarget(), req.HardwareAddressSender())
 			copy(res.ProtocolAddressTarget(), req.ProtocolAddressSender())
 
-			// TODO(raggi): reduce allocs!
+			n, err := t.WriteEthernet(buf)
-			n, err := t.tdev.Write([][]byte{buf}, 0)
 			if tapDebug {
 				t.logf("tap: wrote ARP reply %v, %v", n, err)
 			}
@@ -182,7 +185,7 @@ const routerIP = "100.70.145.1"    // must be in same netmask (currently hack at
 // handleDHCPRequest handles receiving a raw TAP ethernet frame and reports whether
 // it's been handled as a DHCP request. That is, it reports whether the frame should
 // be ignored by the caller and not passed on.
-func (t *Wrapper) handleDHCPRequest(ethBuf []byte) bool {
+func (t *tapDevice) handleDHCPRequest(ethBuf []byte) bool {
 	const udpHeader = 8
 	if len(ethBuf) < ethernetFrameSize+ipv4HeaderLen+udpHeader {
 		if tapDebug {
@@ -207,7 +210,7 @@ func (t *Wrapper) handleDHCPRequest(ethBuf []byte) bool {
 	if p.IPProto != ipproto.UDP || p.Src.Port() != 68 || p.Dst.Port() != 67 {
 		// Not a DHCP request.
 		if tapDebug {
-			t.logf("tap: DHCP wrong meta")
+			t.logf("tap: DHCP wrong meta: %+v", p)
 		}
 		return passOnPacket
 	}
@@ -250,8 +253,7 @@ func (t *Wrapper) handleDHCPRequest(ethBuf []byte) bool {
 			netip.AddrPortFrom(netaddr.IPv4(255, 255, 255, 255), 68), // dst
 		)
 
-		// TODO(raggi): reduce allocs!
+		n, err := t.WriteEthernet(pkt)
-		n, err := t.tdev.Write([][]byte{pkt}, 0)
 		if tapDebug {
 			t.logf("tap: wrote DHCP OFFER %v, %v", n, err)
 		}
@@ -278,8 +280,7 @@ func (t *Wrapper) handleDHCPRequest(ethBuf []byte) bool {
 			netip.AddrPortFrom(netaddr.IPv4(100, 100, 100, 100), 67), // src
 			netip.AddrPortFrom(netaddr.IPv4(255, 255, 255, 255), 68), // dst
 		)
-		// TODO(raggi): reduce allocs!
+		n, err := t.WriteEthernet(pkt)
-		n, err := t.tdev.Write([][]byte{pkt}, 0)
 		if tapDebug {
 			t.logf("tap: wrote DHCP ACK %v, %v", n, err)
 		}
@@ -291,6 +292,16 @@ func (t *Wrapper) handleDHCPRequest(ethBuf []byte) bool {
 	return consumePacket
 }
 
+func writeEthernetFrame(buf []byte, srcMAC, dstMAC net.HardwareAddr, proto tcpip.NetworkProtocolNumber) {
+	// Ethernet header
+	eth := header.Ethernet(buf)
+	eth.Encode(&header.EthernetFields{
+		SrcAddr: tcpip.LinkAddress(srcMAC),
+		DstAddr: tcpip.LinkAddress(dstMAC),
+		Type:    proto,
+	})
+}
+
 func packLayer2UDP(payload []byte, srcMAC, dstMAC net.HardwareAddr, src, dst netip.AddrPort) []byte {
 	buf := make([]byte, header.EthernetMinimumSize+header.UDPMinimumSize+header.IPv4MinimumSize+len(payload))
 	payloadStart := len(buf) - len(payload)
@@ -300,12 +311,7 @@ func packLayer2UDP(payload []byte, srcMAC, dstMAC net.HardwareAddr, src, dst net
 	dstB := dst.Addr().As4()
 	dstIP := tcpip.AddrFromSlice(dstB[:])
 	// Ethernet header
-	eth := header.Ethernet(buf)
+	writeEthernetFrame(buf, srcMAC, dstMAC, ipv4.ProtocolNumber)
-	eth.Encode(&header.EthernetFields{
-		SrcAddr: tcpip.LinkAddress(srcMAC),
-		DstAddr: tcpip.LinkAddress(dstMAC),
-		Type:    ipv4.ProtocolNumber,
-	})
 	// IP header
 	ipbuf := buf[header.EthernetMinimumSize:]
 	ip := header.IPv4(ipbuf)
@@ -342,17 +348,18 @@ func run(prog string, args ...string) error {
 	return nil
 }
 
-func (t *Wrapper) destMAC() [6]byte {
+func (t *tapDevice) destMAC() [6]byte {
 	return t.destMACAtomic.Load()
 }
 
-func newTAPDevice(fd int, tapName string) (tun.Device, error) {
+func newTAPDevice(logf logger.Logf, fd int, tapName string) (tun.Device, error) {
 	err := unix.SetNonblock(fd, true)
 	if err != nil {
 		return nil, err
 	}
 	file := os.NewFile(uintptr(fd), "/dev/tap")
 	d := &tapDevice{
+		logf:   logf,
 		file:   file,
 		events: make(chan tun.Event),
 		name:   tapName,
@@ -360,20 +367,14 @@ func newTAPDevice(fd int, tapName string) (tun.Device, error) {
 	return d, nil
 }
 
-var (
-	_ setWrapperer = &tapDevice{}
-)
-
 type tapDevice struct {
 	file      *os.File
+	logf      func(format string, args ...any)
 	events    chan tun.Event
 	name      string
-	wrapper   *Wrapper
 	closeOnce sync.Once
-}
 
-func (t *tapDevice) setWrapper(wrapper *Wrapper) {
+	destMACAtomic syncs.AtomicValue[[6]byte]
-	t.wrapper = wrapper
 }
 
 func (t *tapDevice) File() *os.File {
@@ -384,36 +385,63 @@ func (t *tapDevice) Name() (string, error) {
 	return t.name, nil
 }
 
+// Read reads an IP packet from the TAP device. It strips the ethernet frame header.
 func (t *tapDevice) Read(buffs [][]byte, sizes []int, offset int) (int, error) {
+	n, err := t.ReadEthernet(buffs, sizes, offset)
+	if err != nil || n == 0 {
+		return n, err
+	}
+	// Strip the ethernet frame header.
+	copy(buffs[0][offset:], buffs[0][offset+ethernetFrameSize:offset+sizes[0]])
+	sizes[0] -= ethernetFrameSize
+	return 1, nil
+}
+
+// ReadEthernet reads a raw ethernet frame from the TAP device.
+func (t *tapDevice) ReadEthernet(buffs [][]byte, sizes []int, offset int) (int, error) {
 	n, err := t.file.Read(buffs[0][offset:])
 	if err != nil {
 		return 0, err
 	}
+	if t.handleTAPFrame(buffs[0][offset : offset+n]) {
+		return 0, nil
+	}
 	sizes[0] = n
 	return 1, nil
 }
 
+// WriteEthernet writes a raw ethernet frame to the TAP device.
+func (t *tapDevice) WriteEthernet(buf []byte) (int, error) {
+	return t.file.Write(buf)
+}
+
+// ethBufPool holds a pool of bytes.Buffers for use in [tapDevice.Write].
+var ethBufPool = syncs.Pool[*bytes.Buffer]{New: func() *bytes.Buffer { return new(bytes.Buffer) }}
+
+// Write writes a raw IP packet to the TAP device. It adds the ethernet frame header.
 func (t *tapDevice) Write(buffs [][]byte, offset int) (int, error) {
 	errs := make([]error, 0)
 	wrote := 0
+	m := t.destMAC()
+	dstMac := net.HardwareAddr(m[:])
+	buf := ethBufPool.Get()
+	defer ethBufPool.Put(buf)
 	for _, buff := range buffs {
-		if offset < ethernetFrameSize {
+		buf.Reset()
-			errs = append(errs, fmt.Errorf("[unexpected] weird offset %d for TAP write", offset))
+		buf.Grow(header.EthernetMinimumSize + len(buff) - offset)
-			return 0, multierr.New(errs...)
+
-		}
+		var ebuf [14]byte
-		eth := buff[offset-ethernetFrameSize:]
+		switch buff[offset] >> 4 {
-		dst := t.wrapper.destMAC()
+		case 4:
-		copy(eth[:6], dst[:])
+			writeEthernetFrame(ebuf[:], ourMAC, dstMac, ipv4.ProtocolNumber)
-		copy(eth[6:12], ourMAC[:])
+		case 6:
-		et := etherTypeIPv4
+			writeEthernetFrame(ebuf[:], ourMAC, dstMac, ipv6.ProtocolNumber)
-		if buff[offset]>>4 == 6 {
+		default:
-			et = etherTypeIPv6
+			continue
 		}
-		eth[12], eth[13] = et[0], et[1]
+		buf.Write(ebuf[:])
-		if tapDebug {
+		buf.Write(buff[offset:])
-			t.wrapper.logf("tap: tapWrite off=%v % x", offset, buff)
+		_, err := t.WriteEthernet(buf.Bytes())
-		}
-		_, err := t.file.Write(buff[offset-ethernetFrameSize:])
 		if err != nil {
 			errs = append(errs, err)
 		} else {
@@ -428,8 +456,7 @@ func (t *tapDevice) MTU() (int, error) {
 	if err != nil {
 		return 0, err
 	}
-	err = unix.IoctlIfreq(int(t.file.Fd()), unix.SIOCGIFMTU, ifr)
+	if err := unix.IoctlIfreq(int(t.file.Fd()), unix.SIOCGIFMTU, ifr); err != nil {
-	if err != nil {
 		return 0, err
 	}
 	return int(ifr.Uint32()), nil

net/tstun/tap_unsupported.go

@@ -1,8 +0,0 @@
-// Copyright (c) Tailscale Inc & AUTHORS
-// SPDX-License-Identifier: BSD-3-Clause
-
-//go:build !linux || ts_omit_tap
-
-package tstun
-
-func (*Wrapper) handleTAPFrame([]byte) bool { panic("unreachable") }

net/tstun/tun.go

@@ -18,7 +18,7 @@ import (
 )
 
 // createTAP is non-nil on Linux.
-var createTAP func(tapName, bridgeName string) (tun.Device, error)
+var createTAP func(logf logger.Logf, tapName, bridgeName string) (tun.Device, error)
 
 // New returns a tun.Device for the requested device name, along with
 // the OS-dependent name that was allocated to the device.
@@ -42,7 +42,7 @@ func New(logf logger.Logf, tunName string) (tun.Device, string, error) {
 		default:
 			return nil, "", errors.New("bogus tap argument")
 		}
-		dev, err = createTAP(tapName, bridgeName)
+		dev, err = createTAP(logf, tapName, bridgeName)
 	} else {
 		dev, err = tun.CreateTUN(tunName, int(DefaultTUNMTU()))
 	}

net/tstun/wrap.go

@@ -109,9 +109,7 @@ type Wrapper struct {
 	lastActivityAtomic mono.Time // time of last send or receive
 
 	destIPActivity syncs.AtomicValue[map[netip.Addr]func()]
-	//lint:ignore U1000 used in tap_linux.go
+	discoKey       syncs.AtomicValue[key.DiscoPublic]
-	destMACAtomic syncs.AtomicValue[[6]byte]
-	discoKey      syncs.AtomicValue[key.DiscoPublic]
 
 	// timeNow, if non-nil, will be used to obtain the current time.
 	timeNow func() time.Time
@@ -257,12 +255,6 @@ type tunVectorReadResult struct {
 	dataOffset int
 }
 
-type setWrapperer interface {
-	// setWrapper enables the underlying TUN/TAP to have access to the Wrapper.
-	// It MUST be called only once during initialization, other usage is unsafe.
-	setWrapper(*Wrapper)
-}
-
 // Start unblocks any Wrapper.Read calls that have already started
 // and makes the Wrapper functional.
 //
@@ -313,10 +305,6 @@ func wrap(logf logger.Logf, tdev tun.Device, isTAP bool, m *usermetric.Registry)
 	w.bufferConsumed <- struct{}{}
 	w.noteActivity()
 
-	if sw, ok := w.tdev.(setWrapperer); ok {
-		sw.setWrapper(w)
-	}
-
 	return w
 }
 
@@ -459,12 +447,18 @@ const ethernetFrameSize = 14 // 2 six byte MACs, 2 bytes ethertype
 func (t *Wrapper) pollVector() {
 	sizes := make([]int, len(t.vectorBuffer))
 	readOffset := PacketStartOffset
+	reader := t.tdev.Read
 	if t.isTAP {
-		readOffset = PacketStartOffset - ethernetFrameSize
+		type tapReader interface {
+			ReadEthernet(buffs [][]byte, sizes []int, offset int) (int, error)
+		}
+		if r, ok := t.tdev.(tapReader); ok {
+			readOffset = PacketStartOffset - ethernetFrameSize
+			reader = r.ReadEthernet
+		}
 	}
 
 	for range t.bufferConsumed {
-	DoRead:
 		for i := range t.vectorBuffer {
 			t.vectorBuffer[i] = t.vectorBuffer[i][:cap(t.vectorBuffer[i])]
 		}
@@ -474,7 +468,7 @@ func (t *Wrapper) pollVector() {
 			if t.isClosed() {
 				return
 			}
-			n, err = t.tdev.Read(t.vectorBuffer[:], sizes, readOffset)
+			n, err = reader(t.vectorBuffer[:], sizes, readOffset)
 			if t.isTAP && tapDebug {
 				s := fmt.Sprintf("% x", t.vectorBuffer[0][:])
 				for strings.HasSuffix(s, " 00") {
@@ -486,21 +480,6 @@ func (t *Wrapper) pollVector() {
 		for i := range sizes[:n] {
 			t.vectorBuffer[i] = t.vectorBuffer[i][:readOffset+sizes[i]]
 		}
-		if t.isTAP {
-			if err == nil {
-				ethernetFrame := t.vectorBuffer[0][readOffset:]
-				if t.handleTAPFrame(ethernetFrame) {
-					goto DoRead
-				}
-			}
-			// Fall through. We got an IP packet.
-			if sizes[0] >= ethernetFrameSize {
-				t.vectorBuffer[0] = t.vectorBuffer[0][:readOffset+sizes[0]-ethernetFrameSize]
-			}
-			if tapDebug {
-				t.logf("tap regular frame: %x", t.vectorBuffer[0][PacketStartOffset:PacketStartOffset+sizes[0]])
-			}
-		}
 		t.sendVectorOutbound(tunVectorReadResult{
 			data:       t.vectorBuffer[:n],
 			dataOffset: PacketStartOffset,