transport: Optimize heap allocations (#8361)

Author: arjan-bal

internal/transport/controlbuf.go

@@ -40,6 +40,13 @@ var updateHeaderTblSize = func(e *hpack.Encoder, v uint32) {
 	e.SetMaxDynamicTableSizeLimit(v)
 }
 
+// itemNodePool is used to reduce heap allocations.
+var itemNodePool = sync.Pool{
+	New: func() any {
+		return &itemNode{}
+	},
+}
+
 type itemNode struct {
 	it   any
 	next *itemNode
@@ -51,7 +58,9 @@ type itemList struct {
 }
 
 func (il *itemList) enqueue(i any) {
-	n := &itemNode{it: i}
+	n := itemNodePool.Get().(*itemNode)
+	n.next = nil
+	n.it = i
 	if il.tail == nil {
 		il.head, il.tail = n, n
 		return
@@ -71,7 +80,9 @@ func (il *itemList) dequeue() any {
 		return nil
 	}
 	i := il.head.it
+	temp := il.head
 	il.head = il.head.next
+	itemNodePool.Put(temp)
 	if il.head == nil {
 		il.tail = nil
 	}

internal/transport/http2_server.go

@@ -132,6 +132,10 @@ type http2Server struct {
 	maxStreamID uint32 // max stream ID ever seen
 
 	logger *grpclog.PrefixLogger
+	// setResetPingStrikes is stored as a closure instead of making this a
+	// method on http2Server to avoid a heap allocation when converting a method
+	// to a closure for passing to frames objects.
+	setResetPingStrikes func()
 }
 
 // NewServerTransport creates a http2 transport with conn and configuration
@@ -266,6 +270,9 @@ func NewServerTransport(conn net.Conn, config *ServerConfig) (_ ServerTransport,
 		initialWindowSize: iwz,
 		bufferPool:        config.BufferPool,
 	}
+	t.setResetPingStrikes = func() {
+		atomic.StoreUint32(&t.resetPingStrikes, 1)
+	}
 	var czSecurity credentials.ChannelzSecurityValue
 	if au, ok := authInfo.(credentials.ChannelzSecurityInfo); ok {
 		czSecurity = au.GetSecurityValue()
@@ -1016,10 +1023,6 @@ func (t *http2Server) writeHeader(s *ServerStream, md metadata.MD) error {
 	return nil
 }
 
-func (t *http2Server) setResetPingStrikes() {
-	atomic.StoreUint32(&t.resetPingStrikes, 1)
-}
-
 func (t *http2Server) writeHeaderLocked(s *ServerStream) error {
 	// TODO(mmukhi): Benchmark if the performance gets better if count the metadata and other header fields
 	// first and create a slice of that exact size.