bugfix(port-forward): Correctly handle known errors

These two errors can occur in the following scenarios:

ECONNRESET: the target process reset connection between CRI and itself.
see: #111825 for detail

EPIPE: the target process did not read the received data, causing the
buffer in the kernel to be full, resulting in the occurrence of Zero Window,
then closing the connection (FIN, RESET)
see: #74551 for detail

In both cases, we should RESET the httpStream.

Signed-off-by: wangxiang <scottwangsxll@gmail.com>

pkg/cri/streaming/portforward/httpstream.go

@@ -38,6 +38,7 @@ import (
 	"net/http"
 	"strconv"
 	"sync"
+	"syscall"
 	"time"
 
 	api "k8s.io/api/core/v1"
@@ -242,8 +243,16 @@ Loop:
 // portForward invokes the httpStreamHandler's forwarder.PortForward
 // function for the given stream pair.
 func (h *httpStreamHandler) portForward(p *httpStreamPair) {
-	defer p.dataStream.Close()
-	defer p.errorStream.Close()
+	resetStreams := false
+	defer func() {
+		if resetStreams {
+			p.dataStream.Reset()
+			p.errorStream.Reset()
+			return
+		}
+		p.dataStream.Close()
+		p.errorStream.Close()
+	}()
 
 	portString := p.dataStream.Headers().Get(api.PortHeader)
 	port, _ := strconv.ParseInt(portString, 10, 32)
@@ -252,11 +261,34 @@ func (h *httpStreamHandler) portForward(p *httpStreamPair) {
 	err := h.forwarder.PortForward(h.pod, h.uid, int32(port), p.dataStream)
 	klog.V(5).Infof("(conn=%p, request=%s) done invoking forwarder.PortForward for port %s", h.conn, p.requestID, portString)
 
-	if err != nil {
-		msg := fmt.Errorf("error forwarding port %d to pod %s, uid %v: %v", port, h.pod, h.uid, err)
-		utilruntime.HandleError(msg)
-		fmt.Fprint(p.errorStream, msg.Error())
+	// happy path, we have successfully completed forwarding task
+	if err == nil {
+		return
+	}
+
+	if errors.Is(err, syscall.EPIPE) || errors.Is(err, syscall.ECONNRESET) {
+		// In the process of forwarding, we encountered error types that can be handled:
+		//
+		// These two errors can occur in the following scenarios:
+		// ECONNRESET: the target process reset connection between CRI and itself.
+		// see: https://github.com/kubernetes/kubernetes/issues/111825 for detail
+		//
+		// EPIPE: the target process did not read the received data, causing the
+		// buffer in the kernel to be full, resulting in the occurrence of Zero Window,
+		// then closing the connection (FIN, RESET)
+		// see: https://github.com/kubernetes/kubernetes/issues/74551 for detail
+		//
+		// In both cases, we should RESET the httpStream.
+		klog.ErrorS(err, "forwarding port", "conn", h.conn, "request", p.requestID, "port", portString)
+		resetStreams = true
+		return
 	}
+
+	// We don't know how to deal with other types of errors,
+	// try to forward them to errStream, let our user know what happened
+	msg := fmt.Errorf("error forwarding port %d to pod %s, uid %v: %v", port, h.pod, h.uid, err)
+	utilruntime.HandleError(msg)
+	fmt.Fprint(p.errorStream, msg.Error())
 }
 
 // httpStreamPair represents the error and data streams for a port