Improve darwin tun performance

internal/fdbased_darwin/endpoint.go

@@ -168,6 +168,7 @@ type endpoint struct {
 	mtu uint32
 
 	batchSize int
+	sendMsgX  bool
 }
 
 // Options specify the details about the fd-based endpoint to be created.
@@ -223,6 +224,9 @@ type Options struct {
 	// ProcessorsPerChannel is the number of goroutines used to handle packets
 	// from each FD.
 	ProcessorsPerChannel int
+
+	MultiPendingPackets bool
+	SendMsgX            bool
 }
 
 // New creates a new fd-based endpoint.
@@ -261,6 +265,12 @@ func New(opts *Options) (stack.LinkEndpoint, error) {
 	if opts.MaxSyscallHeaderBytes < 0 {
 		return nil, fmt.Errorf("opts.MaxSyscallHeaderBytes is negative")
 	}
+	var batchSize int
+	if opts.MultiPendingPackets {
+		batchSize = int((512*1024)/(opts.MTU)) + 1
+	} else {
+		batchSize = 1
+	}
 
 	e := &endpoint{
 		mtu:                   opts.MTU,
@@ -271,7 +281,8 @@ func New(opts *Options) (stack.LinkEndpoint, error) {
 		packetDispatchMode:    opts.PacketDispatchMode,
 		maxSyscallHeaderBytes: uintptr(opts.MaxSyscallHeaderBytes),
 		writevMaxIovs:         rawfile.MaxIovs,
-		batchSize:             int((512*1024)/(opts.MTU)) + 1,
+		batchSize:             batchSize,
+		sendMsgX:              opts.SendMsgX,
 	}
 	if e.maxSyscallHeaderBytes != 0 {
 		if max := int(e.maxSyscallHeaderBytes / rawfile.SizeofIovec); max < e.writevMaxIovs {
@@ -478,7 +489,7 @@ func (e *endpoint) writePacket(pkt *stack.PacketBuffer) tcpip.Error {
 
 func (e *endpoint) sendBatch(batchFDInfo fdInfo, pkts []*stack.PacketBuffer) (int, tcpip.Error) {
 	// Degrade to writePacket if underlying fd is not a socket.
-	if !batchFDInfo.isSocket {
+	if !batchFDInfo.isSocket || !e.sendMsgX {
 		var written int
 		var err tcpip.Error
 		for written < len(pkts) {

internal/fdbased_darwin/packet_dispatchers.go

@@ -116,7 +116,7 @@ func newRecvMMsgDispatcher(fd int, e *endpoint, opts *Options) (linkDispatcher,
 		return nil, err
 	}
 	var batchSize int
-	if opts.MTU < 49152 {
+	if opts.MultiPendingPackets {
 		batchSize = int((512*1024)/(opts.MTU)) + 1
 	} else {
 		batchSize = 1

stack_gvisor.go

@@ -40,6 +40,7 @@ type GVisor struct {
 
 type GVisorTun interface {
 	Tun
+	WritePacket(pkt *stack.PacketBuffer) (int, error)
 	NewEndpoint() (stack.LinkEndpoint, stack.NICOptions, error)
 }
 

stack_gvisor_filter.go

@@ -8,8 +8,6 @@ import (
 	"github.com/sagernet/gvisor/pkg/tcpip"
 	"github.com/sagernet/gvisor/pkg/tcpip/header"
 	"github.com/sagernet/gvisor/pkg/tcpip/stack"
-	"github.com/sagernet/sing/common/bufio"
-	N "github.com/sagernet/sing/common/network"
 )
 
 var _ stack.LinkEndpoint = (*LinkEndpointFilter)(nil)
@@ -17,7 +15,7 @@ var _ stack.LinkEndpoint = (*LinkEndpointFilter)(nil)
 type LinkEndpointFilter struct {
 	stack.LinkEndpoint
 	BroadcastAddress netip.Addr
-	Writer           N.VectorisedWriter
+	Writer           GVisorTun
 }
 
 func (w *LinkEndpointFilter) Attach(dispatcher stack.NetworkDispatcher) {
@@ -29,7 +27,7 @@ var _ stack.NetworkDispatcher = (*networkDispatcherFilter)(nil)
 type networkDispatcherFilter struct {
 	stack.NetworkDispatcher
 	broadcastAddress netip.Addr
-	writer           N.VectorisedWriter
+	writer           GVisorTun
 }
 
 func (w *networkDispatcherFilter) DeliverNetworkPacket(protocol tcpip.NetworkProtocolNumber, pkt *stack.PacketBuffer) {
@@ -49,7 +47,7 @@ func (w *networkDispatcherFilter) DeliverNetworkPacket(protocol tcpip.NetworkPro
 	}
 	destination := AddrFromAddress(network.DestinationAddress())
 	if destination == w.broadcastAddress || !destination.IsGlobalUnicast() {
-		_, _ = bufio.WriteVectorised(w.writer, pkt.AsSlices())
+		w.writer.WritePacket(pkt)
 		return
 	}
 	w.NetworkDispatcher.DeliverNetworkPacket(protocol, pkt)

stack_gvisor_tcp.go

@@ -13,7 +13,6 @@ import (
 	"github.com/sagernet/gvisor/pkg/tcpip/transport/tcp"
 	"github.com/sagernet/sing-tun/internal/gtcpip/checksum"
 	"github.com/sagernet/sing/common"
-	"github.com/sagernet/sing/common/bufio"
 	M "github.com/sagernet/sing/common/metadata"
 	N "github.com/sagernet/sing/common/network"
 )
@@ -56,7 +55,7 @@ func (f *TCPForwarder) HandlePacket(id stack.TransportEndpointID, pkt *stack.Pac
 			tcpHdr.SetChecksum(^checksum.Checksum(tcpHdr.Payload(), tcpHdr.CalculateChecksum(
 				header.PseudoHeaderChecksum(header.TCPProtocolNumber, ipHdr.SourceAddress(), ipHdr.DestinationAddress(), ipHdr.PayloadLength()),
 			)))
-			bufio.WriteVectorised(f.tun, pkt.AsSlices())
+			f.tun.WritePacket(pkt)
 			return true
 		}
 	}
@@ -70,7 +69,7 @@ func (f *TCPForwarder) HandlePacket(id stack.TransportEndpointID, pkt *stack.Pac
 			tcpHdr.SetChecksum(^checksum.Checksum(tcpHdr.Payload(), tcpHdr.CalculateChecksum(
 				header.PseudoHeaderChecksum(header.TCPProtocolNumber, ipHdr.SourceAddress(), ipHdr.DestinationAddress(), ipHdr.PayloadLength()),
 			)))
-			bufio.WriteVectorised(f.tun, pkt.AsSlices())
+			f.tun.WritePacket(pkt)
 			return true
 		}
 	}

stack_mixed.go

@@ -11,12 +11,12 @@ import (
 	"github.com/sagernet/gvisor/pkg/tcpip/transport/udp"
 	"github.com/sagernet/sing-tun/internal/gtcpip/header"
 	"github.com/sagernet/sing/common/buf"
-	"github.com/sagernet/sing/common/bufio"
 	E "github.com/sagernet/sing/common/exceptions"
 )
 
 type Mixed struct {
 	*System
+	tun      GVisorTun
 	stack    *stack.Stack
 	endpoint *channel.Endpoint
 }
@@ -30,6 +30,7 @@ func NewMixed(
 	}
 	return &Mixed{
 		System: system.(*System),
+		tun:    system.(*System).tun.(GVisorTun),
 	}, nil
 }
 
@@ -77,7 +78,7 @@ func (m *Mixed) tunLoop() {
 			return
 		}
 	}
-	if darwinTUN, isDarwinTUN := m.tun.(DarwinTUN); isDarwinTUN && m.mtu < 49152 {
+	if darwinTUN, isDarwinTUN := m.tun.(DarwinTUN); isDarwinTUN && m.multiPendingPackets {
 		m.batchLoopDarwin(darwinTUN)
 		return
 	}
@@ -265,11 +266,11 @@ func (m *Mixed) processIPv6(ipHdr header.IPv6) (writeBack bool, err error) {
 
 func (m *Mixed) packetLoop() {
 	for {
-		packet := m.endpoint.ReadContext(m.ctx)
+		pkt := m.endpoint.ReadContext(m.ctx)
-		if packet == nil {
+		if pkt == nil {
 			break
 		}
-		bufio.WriteVectorised(m.tun, packet.AsSlices())
+		m.tun.WritePacket(pkt)
-		packet.DecRef()
+		pkt.DecRef()
 	}
 }

stack_system.go

@@ -49,6 +49,7 @@ type System struct {
 	interfaceFinder      control.InterfaceFinder
 	frontHeadroom        int
 	txChecksumOffload    bool
+	multiPendingPackets  bool
 }
 
 type Session struct {
@@ -74,6 +75,7 @@ func NewSystem(options StackOptions) (Stack, error) {
 		broadcastAddr:        BroadcastAddr(options.TunOptions.Inet4Address),
 		bindInterface:        options.ForwarderBindInterface,
 		interfaceFinder:      options.InterfaceFinder,
+		multiPendingPackets:  options.TunOptions.EXP_MultiPendingPackets,
 	}
 	if len(options.TunOptions.Inet4Address) > 0 {
 		if !HasNextAddress(options.TunOptions.Inet4Address[0], 1) {
@@ -174,7 +176,7 @@ func (s *System) tunLoop() {
 			return
 		}
 	}
-	if darwinTUN, isDarwinTUN := s.tun.(DarwinTUN); isDarwinTUN && s.mtu < 49152 {
+	if darwinTUN, isDarwinTUN := s.tun.(DarwinTUN); isDarwinTUN && s.multiPendingPackets {
 		s.batchLoopDarwin(darwinTUN)
 		return
 	}
@@ -320,6 +322,13 @@ func (s *System) acceptLoop(listener net.Listener) {
 		if err != nil {
 			return
 		}
+		err = acceptConn(conn)
+		if err != nil {
+			s.logger.Error("set buffer for conn: ", err)
+			_ = conn.Close()
+			listener.Close()
+			return
+		}
 		connPort := M.SocksaddrFromNet(conn.RemoteAddr()).Port
 		session := s.tcpNat.LookupBack(connPort)
 		if session == nil {

stack_system_unix.go

@@ -0,0 +1,26 @@
+//go:build !windows
+
+package tun
+
+import (
+	"net"
+
+	"github.com/sagernet/sing/common/control"
+
+	"golang.org/x/sys/unix"
+)
+
+func acceptConn(conn net.Conn) error {
+	return control.Conn(conn.(*net.TCPConn), func(fd uintptr) error {
+		const bufferSize = 1024 * 1024
+		oErr := unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, unix.SO_RCVBUF, bufferSize)
+		if oErr != nil {
+			return oErr
+		}
+		oErr = unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, unix.SO_SNDBUF, bufferSize)
+		if oErr != nil {
+			return oErr
+		}
+		return nil
+	})
+}

stack_system_windows.go

@@ -2,6 +2,7 @@ package tun
 
 import (
 	"errors"
+	"net"
 	"os"
 	"path/filepath"
 
@@ -31,3 +32,7 @@ func fixWindowsFirewall() error {
 func retryableListenError(err error) bool {
 	return errors.Is(err, windows.WSAEADDRNOTAVAIL)
 }
+
+func acceptConn(conn net.Conn) error {
+	return nil
+}

tun.go

@@ -25,7 +25,6 @@ type Handler interface {
 
 type Tun interface {
 	io.ReadWriter
-	N.VectorisedWriter
 	Name() (string, error)
 	Start() error
 	Close() error
@@ -97,6 +96,13 @@ type Options struct {
 
 	// For library usages.
 	EXP_DisableDNSHijack bool
+
+	// For gvisor stack, it should be enabled when MTU is less than 32768; otherwise it should be less than or equal to 8192.
+	// The above condition is just an estimate and not exact, calculated on M4 pro.
+	EXP_MultiPendingPackets bool
+
+	// Will cause the darwin network to die, do not use.
+	EXP_SendMsgX bool
 }
 
 func (o *Options) Inet4GatewayAddr() netip.Addr {

tun_darwin.go

@@ -14,9 +14,7 @@ import (
 	"github.com/sagernet/sing-tun/internal/stopfd_darwin"
 	"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/shell"
 
 	"golang.org/x/net/route"
@@ -33,15 +31,16 @@ type NativeTun struct {
 	batchSize           int
 	iovecs              []iovecBuffer
 	iovecsOutput        []iovecBuffer
+	iovecsOutputDefault []unix.Iovec
 	msgHdrs             []rawfile.MsgHdrX
 	msgHdrsOutput       []rawfile.MsgHdrX
 	buffers             []*buf.Buffer
 	stopFd              stopfd.StopFD
-	tunWriter     N.VectorisedWriter
 	options             Options
 	inet4Address        [4]byte
 	inet6Address        [16]byte
 	routeSet            bool
+	writeMsgX           bool
 }
 
 type iovecBuffer struct {
@@ -111,14 +110,14 @@ func New(options Options) (Tun, error) {
 			unix.Close(tunFd)
 			return nil, err
 		}
-		err = configure(tunFd, options.MTU, batchSize)
+		err = configure(tunFd, options.EXP_MultiPendingPackets, batchSize)
 		if err != nil {
 			unix.Close(tunFd)
 			return nil, err
 		}
 	} else {
 		tunFd = options.FileDescriptor
-		err := configure(tunFd, options.MTU, batchSize)
+		err := configure(tunFd, options.EXP_MultiPendingPackets, batchSize)
 		if err != nil {
 			return nil, err
 		}
@@ -133,6 +132,7 @@ func New(options Options) (Tun, error) {
 		msgHdrs:       make([]rawfile.MsgHdrX, batchSize),
 		msgHdrsOutput: make([]rawfile.MsgHdrX, batchSize),
 		stopFd:        common.Must1(stopfd.New()),
+		writeMsgX:     options.EXP_SendMsgX,
 	}
 	for i := 0; i < batchSize; i++ {
 		nativeTun.iovecs[i] = newIovecBuffer(int(options.MTU))
@@ -144,11 +144,6 @@ func New(options Options) (Tun, error) {
 	if len(options.Inet6Address) > 0 {
 		nativeTun.inet6Address = options.Inet6Address[0].Addr().As16()
 	}
-	var ok bool
-	nativeTun.tunWriter, ok = bufio.CreateVectorisedWriter(nativeTun.tunFile)
-	if !ok {
-		panic("create vectorised writer")
-	}
 	return nativeTun, nil
 }
 
@@ -182,17 +177,6 @@ func init() {
 	packetHeaderVec6.SetLen(4)
 }
 
-func (t *NativeTun) WriteVectorised(buffers []*buf.Buffer) error {
-	var packetHeader []byte
-	switch header.IPVersion(buffers[0].Bytes()) {
-	case header.IPv4Version:
-		packetHeader = packetHeader4[:]
-	case header.IPv6Version:
-		packetHeader = packetHeader6[:]
-	}
-	return t.tunWriter.WriteVectorised(append([]*buf.Buffer{buf.As(packetHeader)}, buffers...))
-}
-
 const utunControlName = "com.apple.net.utun_control"
 
 const (
@@ -332,12 +316,12 @@ func create(tunFd int, ifIndex int, name string, options Options) error {
 	return nil
 }
 
-func configure(tunFd int, tunMTU uint32, batchSize int) error {
+func configure(tunFd int, multiPendingPackets bool, batchSize int) error {
 	err := unix.SetNonblock(tunFd, true)
 	if err != nil {
 		return os.NewSyscallError("SetNonblock", err)
 	}
-	if tunMTU < 49152 {
+	if multiPendingPackets {
 		const UTUN_OPT_MAX_PENDING_PACKETS = 16
 		err = unix.SetsockoptInt(tunFd, 2, UTUN_OPT_MAX_PENDING_PACKETS, batchSize)
 		if err != nil {
@@ -347,10 +331,6 @@ func configure(tunFd int, tunMTU uint32, batchSize int) error {
 	return nil
 }
 
-func (t *NativeTun) BatchSize() int {
-	return t.batchSize
-}
-
 func (t *NativeTun) BatchRead() ([]*buf.Buffer, error) {
 	for i := 0; i < t.batchSize; i++ {
 		iovecs := t.iovecs[i].nextIovecs()
@@ -384,6 +364,17 @@ func (t *NativeTun) BatchRead() ([]*buf.Buffer, error) {
 }
 
 func (t *NativeTun) BatchWrite(buffers []*buf.Buffer) error {
+	if !t.writeMsgX {
+		for i, buffer := range buffers {
+			t.iovecsOutput[i].nextIovecsOutput(buffer)
+		}
+		for i := range buffers {
+			errno := rawfile.NonBlockingWriteIovec(t.tunFd, t.iovecsOutput[i].iovecs)
+			if errno != 0 {
+				return errno
+			}
+		}
+	} else {
 		for i, buffer := range buffers {
 			iovecs := t.iovecsOutput[i].nextIovecsOutput(buffer)
 			t.msgHdrsOutput[i] = rawfile.MsgHdrX{}
@@ -394,6 +385,7 @@ func (t *NativeTun) BatchWrite(buffers []*buf.Buffer) error {
 		if errno != 0 {
 			return errno
 		}
+	}
 	return nil
 }
 

tun_darwin_gvisor.go

@@ -6,16 +6,43 @@ import (
 	"github.com/sagernet/gvisor/pkg/tcpip/link/qdisc/fifo"
 	"github.com/sagernet/gvisor/pkg/tcpip/stack"
 	"github.com/sagernet/sing-tun/internal/fdbased_darwin"
+	"github.com/sagernet/sing-tun/internal/rawfile_darwin"
+
+	"golang.org/x/sys/unix"
 )
 
 var _ GVisorTun = (*NativeTun)(nil)
 
+func (t *NativeTun) WritePacket(pkt *stack.PacketBuffer) (int, error) {
+	iovecs := t.iovecsOutputDefault
+	var dataLen int
+	for _, packetSlice := range pkt.AsSlices() {
+		dataLen += len(packetSlice)
+		iovec := unix.Iovec{
+			Base: &packetSlice[0],
+		}
+		iovec.SetLen(len(packetSlice))
+		iovecs = append(iovecs, iovec)
+	}
+	if cap(iovecs) > cap(t.iovecsOutputDefault) {
+		t.iovecsOutputDefault = iovecs[:0]
+	}
+	errno := rawfile.NonBlockingWriteIovec(t.tunFd, iovecs)
+	if errno == 0 {
+		return dataLen, nil
+	} else {
+		return 0, errno
+	}
+}
+
 func (t *NativeTun) NewEndpoint() (stack.LinkEndpoint, stack.NICOptions, error) {
 	ep, err := fdbased.New(&fdbased.Options{
 		FDs:                 []int{t.tunFd},
 		MTU:                 t.options.MTU,
 		RXChecksumOffload:   true,
 		PacketDispatchMode:  fdbased.RecvMMsg,
+		MultiPendingPackets: t.options.EXP_MultiPendingPackets,
+		SendMsgX:            t.options.EXP_SendMsgX,
 	})
 	if err != nil {
 		return nil, stack.NICOptions{}, err

tun_linux.go

@@ -18,10 +18,8 @@ import (
 	"github.com/sagernet/sing-tun/internal/gtcpip/header"
 	"github.com/sagernet/sing/common"
 	"github.com/sagernet/sing/common/buf"
-	"github.com/sagernet/sing/common/bufio"
 	"github.com/sagernet/sing/common/control"
 	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"
@@ -34,7 +32,7 @@ var _ LinuxTUN = (*NativeTun)(nil)
 type NativeTun struct {
 	tunFd               int
 	tunFile             *os.File
-	tunWriter         N.VectorisedWriter
+	iovecsOutputDefault []unix.Iovec
 	interfaceCallback   *list.Element[DefaultInterfaceUpdateCallback]
 	options             Options
 	ruleIndex6          []int
@@ -77,11 +75,6 @@ func New(options Options) (Tun, error) {
 			options: options,
 		}
 	}
-	var ok bool
-	nativeTun.tunWriter, ok = bufio.CreateVectorisedWriter(nativeTun.tunFile)
-	if !ok {
-		panic("create vectorised writer")
-	}
 	return nativeTun, nil
 }
 
@@ -402,20 +395,6 @@ func (t *NativeTun) Write(p []byte) (n int, err error) {
 	return t.tunFile.Write(p)
 }
 
-func (t *NativeTun) WriteVectorised(buffers []*buf.Buffer) error {
-	if t.vnetHdr {
-		n := buf.LenMulti(buffers)
-		buffer := buf.NewSize(virtioNetHdrLen + n)
-		buffer.Truncate(virtioNetHdrLen)
-		buf.CopyMulti(buffer.Extend(n), buffers)
-		_, err := t.tunFile.Write(buffer.Bytes())
-		buffer.Release()
-		return err
-	} else {
-		return t.tunWriter.WriteVectorised(buffers)
-	}
-}
-
 func (t *NativeTun) FrontHeadroom() int {
 	if t.vnetHdr {
 		return virtioNetHdrLen

tun_linux_gvisor.go

@@ -3,8 +3,11 @@
 package tun
 
 import (
+	"github.com/sagernet/gvisor/pkg/rawfile"
 	"github.com/sagernet/gvisor/pkg/tcpip/link/fdbased"
 	"github.com/sagernet/gvisor/pkg/tcpip/stack"
+
+	"golang.org/x/sys/unix"
 )
 
 func init() {
@@ -13,6 +16,28 @@ func init() {
 
 var _ GVisorTun = (*NativeTun)(nil)
 
+func (t *NativeTun) WritePacket(pkt *stack.PacketBuffer) (int, error) {
+	iovecs := t.iovecsOutputDefault
+	var dataLen int
+	for _, packetSlice := range pkt.AsSlices() {
+		dataLen += len(packetSlice)
+		iovec := unix.Iovec{
+			Base: &packetSlice[0],
+		}
+		iovec.SetLen(len(packetSlice))
+		iovecs = append(iovecs, iovec)
+	}
+	if cap(iovecs) > cap(t.iovecsOutputDefault) {
+		t.iovecsOutputDefault = iovecs[:0]
+	}
+	errno := rawfile.NonBlockingWriteIovec(t.tunFd, iovecs)
+	if errno == 0 {
+		return dataLen, nil
+	} else {
+		return 0, errno
+	}
+}
+
 func (t *NativeTun) NewEndpoint() (stack.LinkEndpoint, stack.NICOptions, error) {
 	if t.vnetHdr {
 		ep, err := fdbased.New(&fdbased.Options{

tun_windows.go

@@ -17,7 +17,6 @@ import (
 	"github.com/sagernet/sing-tun/internal/wintun"
 	"github.com/sagernet/sing/common"
 	"github.com/sagernet/sing/common/atomic"
-	"github.com/sagernet/sing/common/buf"
 	E "github.com/sagernet/sing/common/exceptions"
 	"github.com/sagernet/sing/common/windnsapi"
 
@@ -517,11 +516,6 @@ func (t *NativeTun) write(packetElementList [][]byte) (n int, err error) {
 	return 0, fmt.Errorf("write failed: %w", err)
 }
 
-func (t *NativeTun) WriteVectorised(buffers []*buf.Buffer) error {
-	defer buf.ReleaseMulti(buffers)
-	return common.Error(t.write(buf.ToSliceMulti(buffers)))
-}
-
 func (t *NativeTun) Close() error {
 	var err error
 	t.closeOnce.Do(func() {

tun_windows_gvisor.go

@@ -11,6 +11,10 @@ import (
 
 var _ GVisorTun = (*NativeTun)(nil)
 
+func (t *NativeTun) WritePacket(pkt *stack.PacketBuffer) (int, error) {
+	return t.write(pkt.AsSlices())
+}
+
 func (t *NativeTun) NewEndpoint() (stack.LinkEndpoint, stack.NICOptions, error) {
 	return &WintunEndpoint{tun: t}, stack.NICOptions{}, nil
 }