Fix GSO write

stack_system.go

@@ -186,13 +186,13 @@ func (s *System) wintunLoop(winTun WinTun) {
 	}
 }
 
-func (m *System) batchLoop(linuxTUN BatchTUN, batchSize int) {
-	frontHeadroom := m.tun.FrontHeadroom()
+func (s *System) batchLoop(linuxTUN BatchTUN, batchSize int) {
+	frontHeadroom := s.tun.FrontHeadroom()
 	packetBuffers := make([][]byte, batchSize)
 	readBuffers := make([][]byte, batchSize)
 	packetSizes := make([]int, batchSize)
 	for i := range packetBuffers {
-		packetBuffers[i] = make([]byte, m.mtu+frontHeadroom+PacketOffset)
+		packetBuffers[i] = make([]byte, s.mtu+frontHeadroom+PacketOffset)
 		readBuffers[i] = packetBuffers[i][frontHeadroom:]
 	}
 	for {
@@ -201,7 +201,7 @@ func (m *System) batchLoop(linuxTUN BatchTUN, batchSize int) {
 			if E.IsClosed(err) {
 				return
 			}
-			m.logger.Error(E.Cause(err, "batch read packet"))
+			s.logger.Error(E.Cause(err, "batch read packet"))
 		}
 		if n == 0 {
 			continue
@@ -214,7 +214,7 @@ func (m *System) batchLoop(linuxTUN BatchTUN, batchSize int) {
 			packetBuffer := packetBuffers[i]
 			rawPacket := packetBuffer[:frontHeadroom+packetSize]
 			packet := packetBuffer[frontHeadroom+PacketOffset : frontHeadroom+packetSize]
-			m.processPacket(rawPacket, packet)
+			s.processPacket(rawPacket, packet)
 		}
 	}
 }

tun.go

@@ -23,6 +23,7 @@ type Handler interface {
 
 type Tun interface {
 	io.ReadWriter
+	N.VectorisedWriter
 	N.FrontHeadroom
 	Close() error
 }
@@ -33,6 +34,7 @@ type WinTun interface {
 }
 
 type BatchTUN interface {
+	Tun
 	BatchSize() int
 	BatchRead(buffers [][]byte, readN []int) (n int, err error)
 }

tun_linux.go

@@ -1,7 +1,6 @@
 package tun
 
 import (
-	"io"
 	"math/rand"
 	"net"
 	"net/netip"
@@ -13,7 +12,10 @@ import (
 
 	"github.com/sagernet/netlink"
 	"github.com/sagernet/sing/common"
+	"github.com/sagernet/sing/common/buf"
+	"github.com/sagernet/sing/common/bufio"
 	E "github.com/sagernet/sing/common/exceptions"
+	N "github.com/sagernet/sing/common/network"
 	"github.com/sagernet/sing/common/rw"
 	"github.com/sagernet/sing/common/shell"
 	"github.com/sagernet/sing/common/x/list"
@@ -26,16 +28,16 @@ var _ BatchTUN = (*NativeTun)(nil)
 type NativeTun struct {
 	tunFd             int
 	tunFile           *os.File
+	tunWriter         N.VectorisedWriter
 	interfaceCallback *list.Element[DefaultInterfaceUpdateCallback]
 	options           Options
 	ruleIndex6        []int
 	gsoEnabled        bool
 	gsoBuffer         []byte
-	tcp4GROTable      *tcpGROTable
-	tcp6GROTable      *tcpGROTable
 }
 
 func New(options Options) (Tun, error) {
+	var nativeTun *NativeTun
 	if options.FileDescriptor == 0 {
 		tunFd, err := open(options.Name, options.GSO)
 		if err != nil {
@@ -45,26 +47,28 @@ func New(options Options) (Tun, error) {
 		if err != nil {
 			return nil, E.Errors(err, unix.Close(tunFd))
 		}
-		nativeTun := &NativeTun{
+		nativeTun = &NativeTun{
 			tunFd:   tunFd,
 			tunFile: os.NewFile(uintptr(tunFd), "tun"),
 			options: options,
 		}
-		runtime.SetFinalizer(nativeTun.tunFile, nil)
 		err = nativeTun.configure(tunLink)
 		if err != nil {
 			return nil, E.Errors(err, unix.Close(tunFd))
 		}
-		return nativeTun, nil
 	} else {
-		nativeTun := &NativeTun{
+		nativeTun = &NativeTun{
 			tunFd:   options.FileDescriptor,
 			tunFile: os.NewFile(uintptr(options.FileDescriptor), "tun"),
 			options: options,
 		}
-		runtime.SetFinalizer(nativeTun.tunFile, nil)
-		return nativeTun, nil
 	}
+	var ok bool
+	nativeTun.tunWriter, ok = bufio.CreateVectorisedWriter(nativeTun.tunFile)
+	if !ok {
+		panic("create vectorised writer")
+	}
+	return nativeTun, nil
 }
 
 func (t *NativeTun) FrontHeadroom() int {
@@ -74,14 +78,6 @@ func (t *NativeTun) FrontHeadroom() int {
 	return 0
 }
 
-func (t *NativeTun) UpstreamWriter() io.Writer {
-	return t.tunFile
-}
-
-func (t *NativeTun) WriterReplaceable() bool {
-	return !t.gsoEnabled
-}
-
 func (t *NativeTun) Read(p []byte) (n int, err error) {
 	if t.gsoEnabled {
 		n, err = t.tunFile.Read(t.gsoBuffer)
@@ -105,22 +101,19 @@ func (t *NativeTun) Read(p []byte) (n int, err error) {
 
 func (t *NativeTun) Write(p []byte) (n int, err error) {
 	if t.gsoEnabled {
-		defer func() {
-			t.tcp4GROTable.reset()
-			t.tcp6GROTable.reset()
-		}()
-		var toWrite []int
-		err = handleGRO([][]byte{p}, virtioNetHdrLen, t.tcp4GROTable, t.tcp6GROTable, &toWrite)
-		if err != nil {
-			return
-		}
-		if len(toWrite) == 0 {
-			return
-		}
+		common.ClearArray(p[:virtioNetHdrLen])
 	}
 	return t.tunFile.Write(p)
 }
 
+func (t *NativeTun) WriteVectorised(buffers []*buf.Buffer) error {
+	if t.gsoEnabled {
+		return t.WriteVectorised(append([]*buf.Buffer{buf.As(rw.ZeroBytes[:virtioNetHdrLen])}, buffers...))
+	} else {
+		return t.WriteVectorised(buffers)
+	}
+}
+
 func (t *NativeTun) BatchSize() int {
 	if !t.gsoEnabled {
 		return 1
@@ -233,8 +226,6 @@ func (t *NativeTun) configure(tunLink netlink.Link) error {
 		}
 		t.gsoEnabled = true
 		t.gsoBuffer = make([]byte, virtioNetHdrLen+int(t.options.GSOMaxSize))
-		t.tcp4GROTable = newTCPGROTable()
-		t.tcp6GROTable = newTCPGROTable()
 	}
 
 	err = netlink.LinkSetUp(tunLink)