Reduce allocations in WebSocket compression

Signed-off-by: Neil Twigg <neil@nats.io>

server/client.go

@@ -346,20 +346,23 @@ func nbPoolGet(sz int) []byte {
 	}
 }
 
-func nbPoolPut(b []byte) {
+func nbPoolPut(in []byte) {
-	switch cap(b) {
+	ca := cap(in)
-	case nbPoolSizeSmall:
+	for in = in[:ca]; ca >= nbPoolSizeSmall; ca = cap(in) {
-		b := (*[nbPoolSizeSmall]byte)(b[0:nbPoolSizeSmall])
+		switch {
-		nbPoolSmall.Put(b)
+		case ca >= nbPoolSizeLarge:
-	case nbPoolSizeMedium:
+			b := (*[nbPoolSizeLarge]byte)(in[0:nbPoolSizeLarge:nbPoolSizeLarge])
-		b := (*[nbPoolSizeMedium]byte)(b[0:nbPoolSizeMedium])
+			nbPoolLarge.Put(b)
-		nbPoolMedium.Put(b)
+			in = in[nbPoolSizeLarge:]
-	case nbPoolSizeLarge:
+		case ca >= nbPoolSizeMedium:
-		b := (*[nbPoolSizeLarge]byte)(b[0:nbPoolSizeLarge])
+			b := (*[nbPoolSizeMedium]byte)(in[0:nbPoolSizeMedium:nbPoolSizeMedium])
-		nbPoolLarge.Put(b)
+			nbPoolMedium.Put(b)
-	default:
+			in = in[nbPoolSizeMedium:]
-		// Ignore frames that are the wrong size, this might happen
+		case ca >= nbPoolSizeSmall:
-		// with WebSocket/MQTT messages as they are framed
+			b := (*[nbPoolSizeSmall]byte)(in[0:nbPoolSizeSmall:nbPoolSizeSmall])
+			nbPoolSmall.Put(b)
+			in = in[nbPoolSizeSmall:]
+		}
 	}
 }
 

server/websocket.go

@@ -115,6 +115,7 @@ type websocket struct {
 	nocompfrag bool // No fragment for compressed frames
 	maskread   bool
 	maskwrite  bool
+	compressor *flate.Writer
 	cookieJwt  string
 	clientIP   string
 }
@@ -1295,7 +1296,13 @@ func (c *client) wsCollapsePtoNB() (net.Buffers, int64) {
 			mfs = 0
 		}
 		buf := bytes.NewBuffer(nbPoolGet(usz))
-		cp, _ := flate.NewWriter(buf, flate.BestSpeed)
+		cp := c.ws.compressor
+		if cp == nil {
+			c.ws.compressor, _ = flate.NewWriter(buf, flate.BestSpeed)
+			cp = c.ws.compressor
+		} else {
+			cp.Reset(buf)
+		}
 		var csz int
 		for _, b := range nb {
 			cp.Write(b)
@@ -1323,9 +1330,7 @@ func (c *client) wsCollapsePtoNB() (net.Buffers, int64) {
 				if mask {
 					wsMaskBuf(key, p[:lp])
 				}
-				new := nbPoolGet(wsFrameSizeForBrowsers)
+				bufs = append(bufs, fh[:n], p[:lp])
-				lp = copy(new[:wsFrameSizeForBrowsers], p[:lp])
-				bufs = append(bufs, fh[:n], new[:lp])
 				csz += n + lp
 				p = p[lp:]
 			}
@@ -1335,16 +1340,9 @@ func (c *client) wsCollapsePtoNB() (net.Buffers, int64) {
 			if mask {
 				wsMaskBuf(key, p)
 			}
-			bufs = append(bufs, h)
+			bufs = append(bufs, h, p)
-			for len(p) > 0 {
-				new := nbPoolGet(len(p))
-				n := copy(new[:cap(new)], p)
-				bufs = append(bufs, new[:n])
-				p = p[n:]
-			}
 			csz = len(h) + ol
 		}
-		nbPoolPut(b) // No longer needed as we copied from above.
 		// Add to pb the compressed data size (including headers), but
 		// remove the original uncompressed data size that was added
 		// during the queueing.
@@ -1355,14 +1353,15 @@ func (c *client) wsCollapsePtoNB() (net.Buffers, int64) {
 		if mfs > 0 {
 			// We are limiting the frame size.
 			startFrame := func() int {
-				bufs = append(bufs, nbPoolGet(wsMaxFrameHeaderSize)[:wsMaxFrameHeaderSize])
+				bufs = append(bufs, nbPoolGet(wsMaxFrameHeaderSize))
 				return len(bufs) - 1
 			}
 			endFrame := func(idx, size int) {
+				bufs[idx] = bufs[idx][:wsMaxFrameHeaderSize]
 				n, key := wsFillFrameHeader(bufs[idx], mask, wsFirstFrame, wsFinalFrame, wsUncompressedFrame, wsBinaryMessage, size)
+				bufs[idx] = bufs[idx][:n]
 				c.out.pb += int64(n)
 				c.ws.fs += int64(n + size)
-				bufs[idx] = bufs[idx][:n]
 				if mask {
 					wsMaskBufs(key, bufs[idx+1:])
 				}
@@ -1388,10 +1387,8 @@ func (c *client) wsCollapsePtoNB() (net.Buffers, int64) {
 					if endStart {
 						fhIdx = startFrame()
 					}
-					new := nbPoolGet(total)
+					bufs = append(bufs, b[:total])
-					n := copy(new[:cap(new)], b[:total])
+					b = b[total:]
-					bufs = append(bufs, new[:n])
-					b = b[n:]
 				}
 			}
 			if total > 0 {