Merge pull request #159 from razo7/more-info-e2e

Remediate Ready Worker Node in E2E

Author: openshift-merge-bot[bot]

test/e2e/far_e2e_test.go

@@ -38,7 +38,6 @@ const (
 
 	//TODO: try to minimize timeout
 	// eventually parameters
-	timeoutLogs        = "3m0s"
 	timeoutTaint       = "2s"   // Timeout for checking the FAR taint
 	timeoutReboot      = "6m0s" // fencing with fence_aws should be completed within 6 minutes
 	timeoutAfterReboot = "5s"   // Timeout for verifying steps  after the node has been rebooted
@@ -48,56 +47,71 @@ const (
 	skipOOSREnvVarName = "SKIP_OOST_REMEDIATION_VERIFICATION"
 )
 
-var remediationTimes []time.Duration
+var (
+	stopTesting      bool
+	remediationTimes []time.Duration
+)
 
 var _ = Describe("FAR E2e", func() {
 	var (
 		fenceAgent, nodeIdentifierPrefix string
 		testShareParam                   map[v1alpha1.ParameterName]string
 		testNodeParam                    map[v1alpha1.ParameterName]map[v1alpha1.NodeName]string
-		selectedNode                     *corev1.Node
-		nodeName                         string
-		pod                              *corev1.Pod
-		startTime, nodeBootTimeBefore    time.Time
 		err                              error
 	)
-	BeforeEach(func() {
-		// create FAR CR spec based on OCP platformn
-		clusterPlatform, err := e2eUtils.GetClusterInfo(configClient)
-		Expect(err).ToNot(HaveOccurred(), "can't identify the cluster platform")
-		log.Info("Begin e2e test", "Cluster name", string(clusterPlatform.Name), "PlatformType", string(clusterPlatform.Status.PlatformStatus.Type))
-
-		switch clusterPlatform.Status.PlatformStatus.Type {
-		case configv1.AWSPlatformType:
-			fenceAgent = fenceAgentAWS
-			nodeIdentifierPrefix = nodeIdentifierPrefixAWS
-			By("running fence_aws")
-		case configv1.BareMetalPlatformType:
-			fenceAgent = fenceAgentIPMI
-			nodeIdentifierPrefix = nodeIdentifierPrefixIPMI
-			By("running fence_ipmilan")
-		default:
-			Skip("FAR haven't been tested on this kind of cluster (non AWS or BareMetal)")
-		}
+	When("trying to identify cluster platform", func() {
+		It("should be AWS/BareMetal for finding the needed cluster and node parameters", func() {
+			// create FAR CR spec based on OCP platformn
+			clusterPlatform, err := e2eUtils.GetClusterInfo(configClient)
+			Expect(err).ToNot(HaveOccurred(), "can't identify the cluster platform")
+			log.Info("Getting Cluster Infromation", "Cluster name", string(clusterPlatform.Name), "PlatformType", string(clusterPlatform.Status.PlatformStatus.Type))
+
+			switch clusterPlatform.Status.PlatformStatus.Type {
+			case configv1.AWSPlatformType:
+				fenceAgent = fenceAgentAWS
+				nodeIdentifierPrefix = nodeIdentifierPrefixAWS
+				By("running fence_aws")
+			case configv1.BareMetalPlatformType:
+				fenceAgent = fenceAgentIPMI
+				nodeIdentifierPrefix = nodeIdentifierPrefixIPMI
+				By("running fence_ipmilan")
+			default:
+				stopTesting = true // Mark to stop subsequent tests
+				Fail("FAR haven't been tested on this kind of cluster (non AWS or BareMetal)")
+			}
 
-		testShareParam, err = buildSharedParameters(clusterPlatform, fenceAgentAction)
-		Expect(err).ToNot(HaveOccurred(), "can't get shared information")
-		testNodeParam, err = buildNodeParameters(clusterPlatform.Status.PlatformStatus.Type)
-		Expect(err).ToNot(HaveOccurred(), "can't get node information")
+			testShareParam, err = buildSharedParameters(clusterPlatform, fenceAgentAction)
+			Expect(err).ToNot(HaveOccurred(), "can't get shared information")
+			testNodeParam, err = buildNodeParameters(clusterPlatform.Status.PlatformStatus.Type)
+			Expect(err).ToNot(HaveOccurred(), "can't get node information")
+		})
 	})
 
 	// runFARTests is a utility function to run FAR tests.
 	// It accepts a remediation strategy and a condition to determine if the tests should be skipped.
 	runFARTests := func(remediationStrategy v1alpha1.RemediationStrategyType, skipCondition func() bool) {
-		var availableWorkerNodes *corev1.NodeList
+		var (
+			availableWorkerNodes          *corev1.NodeList
+			selectedNode                  *corev1.Node
+			nodeName                      string
+			pod                           *corev1.Pod
+			startTime, nodeBootTimeBefore time.Time
+		)
 		BeforeEach(func() {
+			if stopTesting {
+				Skip("Skip testing due to unsupported platform")
+			}
 			if skipCondition() {
 				Skip("Skip this block due to unsupported condition")
 			}
 
 			if availableWorkerNodes == nil {
-				availableWorkerNodes = getAvailableWorkerNodes()
+				availableWorkerNodes = getReadyWorkerNodes()
 			}
+			if len(availableWorkerNodes.Items) < 1 {
+				Fail("There isn't an available (and Ready) worker node in the cluster")
+			}
+
 			selectedNode = pickRemediatedNode(availableWorkerNodes)
 			nodeName = selectedNode.Name
 			printNodeDetails(selectedNode, nodeIdentifierPrefix, testNodeParam)
@@ -107,7 +121,7 @@ var _ = Describe("FAR E2e", func() {
 			Expect(err).ToNot(HaveOccurred(), "failed to get boot time of the node")
 
 			// create tested pod which will be deleted by the far CR
-			pod = createTestedPod(nodeName, testContainerName)
+			pod = createTestedPod(nodeName)
 			DeferCleanup(cleanupTestedResources, pod)
 
 			// set the node as "unhealthy" by disabling kubelet
@@ -240,25 +254,31 @@ func buildNodeParameters(clusterPlatformType configv1.PlatformType) (map[v1alpha
 	return testNodeParam, nil
 }
 
-// getAvailableNodes a list of available worker nodes in the cluster
-func getAvailableWorkerNodes() *corev1.NodeList {
-	availableNodes := &corev1.NodeList{}
+// getReadyWorkerNodes returns a list of ready worker nodes in the cluster if any
+func getReadyWorkerNodes() *corev1.NodeList {
+	availableWorkerNodes := &corev1.NodeList{}
 	selector := labels.NewSelector()
-	requirement, _ := labels.NewRequirement(medik8sLabels.WorkerRole, selection.Exists, []string{})
+	requirement, err := labels.NewRequirement(medik8sLabels.WorkerRole, selection.Exists, []string{})
+	Expect(err).To(BeNil())
 	selector = selector.Add(*requirement)
-	Expect(k8sClient.List(context.Background(), availableNodes, &client.ListOptions{LabelSelector: selector})).ToNot(HaveOccurred())
-	if len(availableNodes.Items) < 1 {
-		Fail("No worker nodes found in the cluster")
+	Expect(k8sClient.List(context.Background(), availableWorkerNodes, &client.ListOptions{LabelSelector: selector})).ToNot(HaveOccurred())
+
+	// Filter nodes to only include those in "Ready" state
+	readyWorkerNodes := &corev1.NodeList{}
+	for _, node := range availableWorkerNodes.Items {
+		for _, condition := range node.Status.Conditions {
+			if condition.Type == corev1.NodeReady && condition.Status == corev1.ConditionTrue {
+				readyWorkerNodes.Items = append(readyWorkerNodes.Items, node)
+				break // "Ready" was found
+			}
+		}
 	}
-	return availableNodes
+	return readyWorkerNodes
 }
 
 // pickRemediatedNode randomly returns a next remediated node from the current available nodes,
 // and then the node is removed from the list of available nodes
 func pickRemediatedNode(availableNodes *corev1.NodeList) *corev1.Node {
-	if len(availableNodes.Items) < 1 {
-		Fail("No available node found for remediation")
-	}
 	// Generate a random seed based on the current time
 	r := rand.New(rand.NewSource(time.Now().UnixNano()))
 	// Randomly select a worker node
@@ -270,14 +290,14 @@ func pickRemediatedNode(availableNodes *corev1.NodeList) *corev1.Node {
 }
 
 // createTestedPod creates tested pod which will be deleted by the far CR
-func createTestedPod(nodeName, containerName string) *corev1.Pod {
+func createTestedPod(nodeName string) *corev1.Pod {
 	pod := e2eUtils.GetPod(nodeName, testContainerName)
 	pod.Name = testPodName
 	pod.Namespace = testNsName
 	pod.Spec.Tolerations = []corev1.Toleration{
 		{
 			Key:      v1alpha1.FARNoExecuteTaintKey,
-			Operator: corev1.TolerationOpEqual,
+			Operator: corev1.TolerationOpExists,
 			Effect:   corev1.TaintEffectNoExecute,
 		},
 	}
@@ -303,7 +323,7 @@ func createFAR(nodeName string, agent string, sharedParameters map[v1alpha1.Para
 			SharedParameters:    sharedParameters,
 			NodeParameters:      nodeParameters,
 			RemediationStrategy: strategy,
-			RetryCount:          5,
+			RetryCount:          10,
 			RetryInterval:       metav1.Duration{Duration: 20 * time.Second},
 			Timeout:             metav1.Duration{Duration: 60 * time.Second},
 		},

test/e2e/utils/cluster.go

@@ -48,6 +48,23 @@ func GetSecretData(clientSet *kubernetes.Clientset, secretName, secretNamespace,
 	return string(secret.Data[secretData1]), string(secret.Data[secretData2]), nil
 }
 
+// getNodeRoleFromMachine return node role "master/control-plane" or "worker" from machine label if present, otherwise "unknown"
+func getNodeRoleFromMachine(nodeLabels map[string]string) string {
+	machineLabelPrefixRole := "machine.openshift.io/cluster-api-machine-"
+	// look for machine.openshift.io/cluster-api-machine-role or machine.openshift.io/cluster-api-machine-type label
+	for _, labelKey := range []string{machineLabelPrefixRole + "role", machineLabelPrefixRole + "type"} {
+		if labelVal, isFound := nodeLabels[labelKey]; isFound {
+			if labelVal == "worker" {
+				return "worker"
+			}
+			if labelVal == "master" {
+				return "master/control-plane"
+			}
+		}
+	}
+	return "unknown"
+}
+
 // GetAWSNodeInfoList returns a list of the node names and their identification, e.g., AWS instance ID
 func GetAWSNodeInfoList(machineClient *machineclient.Clientset) (map[v1alpha1.NodeName]string, error) {
 	//  oc get machine -n openshift-machine-api MACHINE_NAME -o jsonpath='{.spec.providerID}'
@@ -62,24 +79,28 @@ func GetAWSNodeInfoList(machineClient *machineclient.Clientset) (map[v1alpha1.No
 	}
 
 	var missNodeMachineErr error
+	missNodeMachineNames := ""
 	// creates map for nodeName and AWS instance ID
 	for _, machine := range machineList.Items {
 		if machine.Status.NodeRef == nil || machine.Spec.ProviderID == nil {
 			if missNodeMachineErr != nil {
-				missNodeMachineErr = fmt.Errorf("machine %s is not associated with any node or it't provider ID is missing\n%w", machine.Spec.Name, missNodeMachineErr)
+				missNodeMachineNames += ", " + machine.ObjectMeta.GetName()
+				missNodeMachineErr = fmt.Errorf("machines %s are not associated with any node or there provider ID is missing", missNodeMachineNames)
 			} else {
-				missNodeMachineErr = fmt.Errorf("machine %s is not associated with any node or it't provider ID is missing", machine.Spec.Name)
+				missNodeMachineNames = machine.ObjectMeta.GetName()
+				missNodeMachineErr = fmt.Errorf("machine %s is not associated with any node or it's provider ID is missing", machine.ObjectMeta.GetName())
 			}
-			continue
+		} else {
+			nodeName := v1alpha1.NodeName(machine.Status.NodeRef.Name)
+			nodeRole := getNodeRoleFromMachine(machine.Labels)
+			providerID := *machine.Spec.ProviderID
+
+			// Get the instance ID from the provider ID aws:///us-east-1b/i-082ac37ab919a82c2 -> i-082ac37ab919a82c2
+			splitedProviderID := strings.Split(providerID, "/i-")
+			instanceID := "i-" + splitedProviderID[1]
+			nodeList[nodeName] = instanceID
+			fmt.Printf("node: %s, Role: %s, Instance ID: %s \n", nodeName, nodeRole, instanceID)
 		}
-		nodeName := v1alpha1.NodeName(machine.Status.NodeRef.Name)
-		providerID := *machine.Spec.ProviderID
-
-		// Get the instance ID from the provider ID aws:///us-east-1b/i-082ac37ab919a82c2 -> i-082ac37ab919a82c2
-		splitedProviderID := strings.Split(providerID, "/i-")
-		instanceID := "i-" + splitedProviderID[1]
-		nodeList[nodeName] = instanceID
-		fmt.Printf("node: %s Instance ID: %s \n", nodeName, instanceID)
 	}
 	return nodeList, missNodeMachineErr
 }

test/e2e/utils/command.go

@@ -193,7 +193,7 @@ func GetPod(nodeName, containerName string) *corev1.Pod {
 			Tolerations: []corev1.Toleration{
 				{
 					Key:      v1alpha1.FARNoExecuteTaintKey,
-					Operator: corev1.TolerationOpEqual,
+					Operator: corev1.TolerationOpExists,
 					Effect:   corev1.TaintEffectNoExecute,
 				},
 				{