diff --git a/session.go b/session.go
index e3dfea5..19b008c 100644
--- a/session.go
+++ b/session.go
@@ -1,7 +1,6 @@
 package smux
 
 import (
-	"bytes"
 	"io"
 	"sync"
 	"sync/atomic"
@@ -31,9 +30,8 @@ type Session struct {
 	bucket     int32      // token bucket
 	bucketCond *sync.Cond // used for waiting for tokens
 
-	streamBuffers map[uint32]*bytes.Buffer // stream input buffer
-	streams       map[uint32]*Stream       // all streams in this session
-	streamLock    sync.Mutex               // locks streams && frameQueues
+	streams    map[uint32]*Stream // all streams in this session
+	streamLock sync.Mutex         // locks streams && frameQueues
 
 	die            chan struct{} // flag session has died
 	dieLock        sync.Mutex
@@ -49,7 +47,6 @@ func newSession(config *Config, conn io.ReadWriteCloser, client bool) *Session {
 	s.conn = conn
 	s.config = config
 	s.streams = make(map[uint32]*Stream)
-	s.streamBuffers = make(map[uint32]*bytes.Buffer)
 	s.chAccepts = make(chan *Stream, defaultAcceptBacklog)
 	s.chClosedStream = make(chan uint32, defaultCloseWait)
 	s.bucket = int32(config.MaxReceiveBuffer)
@@ -73,11 +70,9 @@ func (s *Session) OpenStream() (*Stream, error) {
 
 	sid := atomic.AddUint32(&s.nextStreamID, 2)
 	stream := newStream(sid, s.config.MaxFrameSize, s)
-	streamBuffer := new(bytes.Buffer)
 
 	s.streamLock.Lock()
 	s.streams[sid] = stream
-	s.streamBuffers[sid] = streamBuffer
 	s.streamLock.Unlock()
 
 	s.writeFrame(newFrame(cmdSYN, sid))
@@ -146,18 +141,11 @@ func (s *Session) streamClosed(sid uint32) {
 	}()
 }
 
-// nonblocking read from session pool, for streams
-func (s *Session) nioread(sid uint32, p []byte) (n int) {
-	s.streamLock.Lock()
-	if streamBuffer, ok := s.streamBuffers[sid]; ok {
-		n, _ = streamBuffer.Read(p)
-		if n > 0 {
-			atomic.AddInt32(&s.bucket, int32(n))
-			s.bucketCond.Signal()
-		}
+// returnTokens
+func (s *Session) returnTokens(n int) {
+	if atomic.AddInt32(&s.bucket, int32(n)) > 0 {
+		s.bucketCond.Signal()
 	}
-	s.streamLock.Unlock()
-	return
 }
 
 // session read a frame from underlying connection
@@ -189,12 +177,12 @@ func (s *Session) monitor() {
 		select {
 		case sid := <-s.chClosedStream:
 			s.streamLock.Lock()
-			delete(s.streams, sid)
-			if n := s.streamBuffers[sid].Len(); n > 0 { // return remaining tokens to the bucket
-				atomic.AddInt32(&s.bucket, int32(n))
-				s.bucketCond.Signal()
+			if n := s.streams[sid].recycleTokens(); n > 0 { // return remaining tokens to the bucket
+				if atomic.AddInt32(&s.bucket, int32(n)) > 0 {
+					s.bucketCond.Signal()
+				}
 			}
-			delete(s.streamBuffers, sid)
+			delete(s.streams, sid)
 			s.streamLock.Unlock()
 		case <-s.die:
 			return
@@ -225,9 +213,7 @@ func (s *Session) recvLoop() {
 				s.streamLock.Lock()
 				if _, ok := s.streams[f.sid]; !ok {
 					stream := newStream(f.sid, s.config.MaxFrameSize, s)
-					streamBuffer := new(bytes.Buffer)
 					s.streams[f.sid] = stream
-					s.streamBuffers[f.sid] = streamBuffer
 					go func() { s.chAccepts <- stream }()
 				} else { // stream exists, RST the peer
 					go s.writeFrame(newFrame(cmdRST, f.sid))
@@ -244,7 +230,7 @@ func (s *Session) recvLoop() {
 				s.streamLock.Lock()
 				if stream, ok := s.streams[f.sid]; ok {
 					atomic.AddInt32(&s.bucket, -int32(len(f.data)))
-					s.streamBuffers[f.sid].Write(f.data)
+					stream.pushBytes(f.data)
 					stream.notifyReadEvent()
 				} else { // stream is absent
 					go s.writeFrame(newFrame(cmdRST, f.sid))
diff --git a/stream.go b/stream.go
index 17d6717..9a21155 100644
--- a/stream.go
+++ b/stream.go
@@ -13,6 +13,8 @@ type Stream struct {
 	id          uint32
 	rstflag     int32
 	sess        *Session
+	buffer      bytes.Buffer
+	bufferLock  sync.Mutex
 	frameSize   int
 	chReadEvent chan struct{} // notify a read event
 	die         chan struct{} // flag the stream has closed
@@ -39,8 +41,13 @@ READ:
 	default:
 	}
 
-	if n = s.sess.nioread(s.id, b); n > 0 {
-		return n, err
+	s.bufferLock.Lock()
+	n, err = s.buffer.Read(b)
+	s.bufferLock.Unlock()
+
+	if n > 0 {
+		s.sess.returnTokens(n)
+		return n, nil
 	} else if atomic.LoadInt32(&s.rstflag) == 1 {
 		s.Close()
 		return 0, errors.New(errBrokenPipe)
@@ -106,6 +113,22 @@ func (s *Stream) sessionClose() error {
 	return nil
 }
 
+// pushBytes a slice into buffer
+func (s *Stream) pushBytes(p []byte) {
+	s.bufferLock.Lock()
+	s.buffer.Write(p)
+	s.bufferLock.Unlock()
+}
+
+// recycleTokens transform remaining bytes to tokens(will truncate buffer)
+func (s *Stream) recycleTokens() (n int) {
+	s.bufferLock.Lock()
+	n = s.buffer.Len()
+	s.buffer.Reset()
+	s.bufferLock.Unlock()
+	return
+}
+
 // split large byte buffer into smaller frames
 func (s *Stream) split(bts []byte, cmd byte, sid uint32) []Frame {
 	var frames []Frame