Improve vectorised writer

common/buf/multi.go

@@ -16,6 +16,14 @@ func ToSliceMulti(buffers []*Buffer) [][]byte {
 	})
 }
 
+func CopyMulti(toBuffer []byte, buffers []*Buffer) int {
+	var n int
+	for _, buffer := range buffers {
+		n += copy(toBuffer[n:], buffer.Bytes())
+	}
+	return n
+}
+
 func ReleaseMulti(buffers []*Buffer) {
 	for _, buffer := range buffers {
 		buffer.Release()

common/bufio/chunk.go

@@ -49,11 +49,29 @@ func (c *ChunkReader) Read(p []byte) (n int, err error) {
 	c.cache.FullReset()
 	err = c.upstream.ReadBuffer(c.cache)
 	if err != nil {
+		c.cache.Release()
+		c.cache = nil
 		return
 	}
 	return c.cache.Read(p)
 }
 
+func (c *ChunkReader) ReadChunk() (*buf.Buffer, error) {
+	if c.cache == nil {
+		c.cache = buf.NewSize(c.maxChunkSize)
+	} else if !c.cache.IsEmpty() {
+		return c.cache, nil
+	}
+	c.cache.FullReset()
+	err := c.upstream.ReadBuffer(c.cache)
+	if err != nil {
+		c.cache.Release()
+		c.cache = nil
+		return nil, err
+	}
+	return c.cache, nil
+}
+
 func (c *ChunkReader) MTU() int {
 	return c.maxChunkSize
 }

common/bufio/vectorised.go

@@ -3,7 +3,6 @@ package bufio
 import (
 	"io"
 	"net"
-	"sync"
 	"syscall"
 
 	"github.com/sagernet/sing/common"
@@ -16,7 +15,7 @@ func NewVectorisedWriter(writer io.Writer) N.VectorisedWriter {
 	if vectorisedWriter, ok := CreateVectorisedWriter(writer); ok {
 		return vectorisedWriter
 	}
-	return &SerialVectorisedWriter{upstream: writer}
+	return &BufferedVectorisedWriter{upstream: writer}
 }
 
 func CreateVectorisedWriter(writer any) (N.VectorisedWriter, bool) {
@@ -57,26 +56,30 @@ func CreateVectorisedPacketWriter(writer any) (N.VectorisedPacketWriter, bool) {
 	return nil, false
 }
 
-var _ N.VectorisedWriter = (*SerialVectorisedWriter)(nil)
+var _ N.VectorisedWriter = (*BufferedVectorisedWriter)(nil)
 
-type SerialVectorisedWriter struct {
+type BufferedVectorisedWriter struct {
 	upstream io.Writer
-	access   sync.Mutex
 }
 
-func (w *SerialVectorisedWriter) WriteVectorised(buffers []*buf.Buffer) error {
-	w.access.Lock()
-	defer w.access.Unlock()
-	for _, buffer := range buffers {
-		_, err := w.upstream.Write(buffer.Bytes())
-		if err != nil {
-			return err
-		}
+func (w *BufferedVectorisedWriter) WriteVectorised(buffers []*buf.Buffer) error {
+	defer buf.ReleaseMulti(buffers)
+	bufferLen := buf.LenMulti(buffers)
+	var bufferBytes []byte
+	if bufferLen > 65535 {
+		bufferBytes = make([]byte, bufferLen)
+	} else {
+		_buffer := buf.StackNewSize(bufferLen)
+		defer common.KeepAlive(_buffer)
+		buffer := common.Dup(_buffer)
+		defer buffer.Release()
+		bufferBytes = buffer.FreeBytes()
 	}
-	return nil
+	buf.CopyMulti(bufferBytes, buffers)
+	return common.Error(w.upstream.Write(bufferBytes))
 }
 
-func (w *SerialVectorisedWriter) Upstream() any {
+func (w *BufferedVectorisedWriter) Upstream() any {
 	return w.upstream
 }
 

go.sum

@@ -1,4 +1,2 @@
-golang.org/x/sys v0.4.0 h1:Zr2JFtRQNX3BCZ8YtxRE9hNJYC8J6I1MVbMg6owUp18=
-golang.org/x/sys v0.4.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0 h1:MUK/U/4lj1t1oPg0HfuXDN/Z1wv31ZJ/YcPiGccS4DU=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=