[Windows] Bugfix for MTU configuration on nonexistent interface

Configure MTU in a separate goroutine after the host interface is
created.

Signed-off-by: wenyingd <wenyingd@vmware.com>

pkg/agent/cniserver/interface_configuration_linux.go

@@ -341,6 +341,7 @@ func (ic *ifConfigurator) configureContainerLink(
 	brSriovVFDeviceID string,
 	podSriovVFDeviceID string,
 	result *current.Result,
+	containerAccess *containerAccessArbitrator,
 ) error {
 	if brSriovVFDeviceID != "" {
 		if !ic.isOvsHardwareOffloadEnabled {

pkg/agent/cniserver/interface_configuration_windows.go

@@ -30,6 +30,7 @@ import (
 	cnitypes "github.com/containernetworking/cni/pkg/types"
 	"github.com/containernetworking/cni/pkg/types/current"
 	"k8s.io/apimachinery/pkg/util/sets"
+	"k8s.io/apimachinery/pkg/util/wait"
 	"k8s.io/klog/v2"
 
 	"antrea.io/antrea/pkg/agent/interfacestore"
@@ -43,6 +44,8 @@ const (
 	notFoundHNSEndpoint = "The endpoint was not found"
 )
 
+type postInterfaceCreateHook func() error
+
 type ifConfigurator struct {
 	hnsNetwork *hcsshim.HNSNetwork
 	epCache    *sync.Map
@@ -106,6 +109,7 @@ func (ic *ifConfigurator) configureContainerLink(
 	brSriovVFDeviceID string,
 	podSriovVFDeviceID string,
 	result *current.Result,
+	containerAccess *containerAccessArbitrator,
 ) error {
 	if brSriovVFDeviceID != "" {
 		return fmt.Errorf("OVS hardware offload is not supported on windows")
@@ -152,9 +156,19 @@ func (ic *ifConfigurator) configureContainerLink(
 	ifaceIdx := 1
 	containerIP.Interface = &ifaceIdx
 
-	ifaceName := fmt.Sprintf("vEthernet (%s)", endpoint.Name)
-	if err := util.SetInterfaceMTU(ifaceName, mtu); err != nil {
-		return fmt.Errorf("failed to configure MTU on container interface '%s': %v", ifaceName, err)
+	// MTU is configured only when the infrastructure container is created.
+	if containerID == infraContainerID {
+		// Configure MTU in another separate goroutine to ensure it is executed after the host interface is created.
+		// The reasons include, 1) for containerd runtime, the interface is created by containerd after the CNI
+		// CmdAdd request is returned; 2) for docker runtime, the interface is created after hcsshim.HotAttachEndpoint,
+		// and the hcsshim call is not synchronized from the observation.
+		return ic.addPostInterfaceCreateHook(infraContainerID, epName, containerAccess, func() error {
+			ifaceName := fmt.Sprintf("%s (%s)", util.ContainerVNICPrefix, epName)
+			if err := util.SetInterfaceMTU(ifaceName, mtu); err != nil {
+				return fmt.Errorf("failed to configure MTU on container interface '%s': %v", ifaceName, err)
+			}
+			return nil
+		})
 	}
 	return nil
 }
@@ -504,3 +518,42 @@ func (ic *ifConfigurator) getInterceptedInterfaces(
 func (ic *ifConfigurator) getOVSInterfaceType() int {
 	return internalOVSInterfaceType
 }
+
+func (ic *ifConfigurator) addPostInterfaceCreateHook(containerID, endpointName string, containerAccess *containerAccessArbitrator, hook postInterfaceCreateHook) error {
+	if containerAccess == nil {
+		return fmt.Errorf("container lock cannot be null")
+	}
+	expectedEP, ok := ic.getEndpoint(endpointName)
+	if !ok {
+		return fmt.Errorf("failed to find HNSEndpoint %s", endpointName)
+	}
+	go func() {
+		ifaceName := fmt.Sprintf("vEthernet (%s)", endpointName)
+		err := wait.PollImmediate(time.Second, 60*time.Second, func() (done bool, err error) {
+			containerAccess.lockContainer(containerID)
+			defer containerAccess.unlockContainer(containerID)
+			currentEP, ok := ic.getEndpoint(endpointName)
+			if !ok {
+				klog.InfoS("HNSEndpoint doesn't exist in cache, exit current goroutine", "HNSEndpoint Name", endpointName)
+				return true, nil
+			}
+			if currentEP.Id != expectedEP.Id {
+				klog.InfoS("Detected HNSEndpoint change, exit current goroutine", "HNSEndpoint Name", endpointName)
+				return true, nil
+			}
+			if !util.HostInterfaceExists(ifaceName) {
+				klog.InfoS("Waiting for interface to be created", "interface name", ifaceName)
+				return false, nil
+			}
+			if err := hook(); err != nil {
+				return false, err
+			}
+			return true, nil
+		})
+
+		if err != nil {
+			klog.Errorf("Failed to configure host interface %s: %v", ifaceName, err)
+		}
+	}()
+	return nil
+}

pkg/agent/cniserver/pod_configuration.go

@@ -204,7 +204,7 @@ func (pc *podConfigurator) configureInterfaces(
 	createOVSPort bool,
 	containerAccess *containerAccessArbitrator,
 ) error {
-	err := pc.ifConfigurator.configureContainerLink(podName, podNameSpace, containerID, containerNetNS, containerIFDev, mtu, sriovVFDeviceID, "", result)
+	err := pc.ifConfigurator.configureContainerLink(podName, podNameSpace, containerID, containerNetNS, containerIFDev, mtu, sriovVFDeviceID, "", result, containerAccess)
 	if err != nil {
 		return err
 	}

pkg/agent/cniserver/pod_configuration_windows.go

@@ -19,11 +19,9 @@ package cniserver
 
 import (
 	"fmt"
-	"time"
 
 	"github.com/containernetworking/cni/pkg/types/current"
 	"k8s.io/apimachinery/pkg/util/sets"
-	"k8s.io/apimachinery/pkg/util/wait"
 	"k8s.io/klog/v2"
 
 	"antrea.io/antrea/pkg/agent/interfacestore"
@@ -34,43 +32,10 @@ import (
 // in another goroutine. The function is for Containerd runtime. The host interface is created after
 // CNI call completes.
 func (pc *podConfigurator) connectInterfaceToOVSAsync(ifConfig *interfacestore.InterfaceConfig, containerAccess *containerAccessArbitrator) error {
-	if containerAccess == nil {
-		return fmt.Errorf("container lock cannot be null")
-	}
 	ovsPortName := ifConfig.InterfaceName
-	expectedEp, ok := pc.ifConfigurator.getEndpoint(ovsPortName)
-	if !ok {
-		return fmt.Errorf("failed to find HNSEndpoint %s", ovsPortName)
-	}
-	hostIfAlias := fmt.Sprintf("%s (%s)", util.ContainerVNICPrefix, ovsPortName)
-	containerID := ifConfig.ContainerID
-	go func() {
-		klog.Infof("Waiting for interface %s to be created", hostIfAlias)
-		err := wait.PollImmediate(time.Second, 60*time.Second, func() (bool, error) {
-			containerAccess.lockContainer(containerID)
-			defer containerAccess.unlockContainer(containerID)
-			curEp, ok := pc.ifConfigurator.getEndpoint(ovsPortName)
-			if !ok {
-				return true, fmt.Errorf("failed to find HNSEndpoint %s", ovsPortName)
-			}
-			if curEp.Id != expectedEp.Id {
-				klog.Warningf("Detected HNSEndpoint change for port %s, exiting current goroutine", ovsPortName)
-				return true, nil
-			}
-			if !util.HostInterfaceExists(hostIfAlias) {
-				klog.Infof("Waiting for interface %s to be created", hostIfAlias)
-				return false, nil
-			}
-			if err := pc.connectInterfaceToOVSCommon(ovsPortName, ifConfig); err != nil {
-				return true, fmt.Errorf("failed to connect to OVS for container %s: %v", containerID, err)
-			}
-			return true, nil
-		})
-		if err != nil {
-			klog.Errorf("Failed to create OVS port for container %s: %v", containerID, err)
-		}
-	}()
-	return nil
+	return pc.ifConfigurator.addPostInterfaceCreateHook(ifConfig.ContainerID, ovsPortName, containerAccess, func() error {
+		return pc.connectInterfaceToOVSCommon(ovsPortName, ifConfig)
+	})
 }
 
 // connectInterfaceToOVS connects an existing interface to OVS br-int.

pkg/agent/cniserver/sriov.go

@@ -125,7 +125,7 @@ func (pc *podConfigurator) configureSecondaryInterface(
 		return fmt.Errorf("error getting the Pod SR-IOV VF device ID")
 	}
 
-	err := pc.ifConfigurator.configureContainerLink(podName, podNameSpace, containerID, containerNetNS, containerIFDev, mtu, "", podSriovVFDeviceID, result)
+	err := pc.ifConfigurator.configureContainerLink(podName, podNameSpace, containerID, containerNetNS, containerIFDev, mtu, "", podSriovVFDeviceID, result, nil)
 	if err != nil {
 		return err
 	}