// checkNodesReady waits up to nt for expect nodes accessed by c to be ready,
// returning an error if this doesn't happen in time. It returns the names of
// nodes it finds.
func checkNodesReady(c *client.Client, nt time.Duration, expect int) ([]string, error) {
// First, keep getting all of the nodes until we get the number we expect.
var nodeList *api.NodeList
var errLast error
start := time.Now()
found := wait.Poll(framework.Poll, nt, func() (bool, error) {
// A rolling-update (GCE/GKE implementation of restart) can complete before the apiserver
// knows about all of the nodes. Thus, we retry the list nodes call
// until we get the expected number of nodes.
nodeList, errLast = c.Nodes().List(api.ListOptions{
FieldSelector: fields.Set{"spec.unschedulable": "false"}.AsSelector()})
if errLast != nil {
return false, nil
}
if len(nodeList.Items) != expect {
errLast = fmt.Errorf("expected to find %d nodes but found only %d (%v elapsed)",
expect, len(nodeList.Items), time.Since(start))
framework.Logf("%v", errLast)
return false, nil
}
return true, nil
}) == nil
nodeNames := make([]string, len(nodeList.Items))
for i, n := range nodeList.Items {
nodeNames[i] = n.ObjectMeta.Name
}
if !found {
return nodeNames, fmt.Errorf("couldn't find %d nodes within %v; last error: %v",
expect, nt, errLast)
}
framework.Logf("Successfully found %d nodes", expect)
// Next, ensure in parallel that all the nodes are ready. We subtract the
// time we spent waiting above.
timeout := nt - time.Since(start)
result := make(chan bool, len(nodeList.Items))
for _, n := range nodeNames {
n := n
go func() { result <- framework.WaitForNodeToBeReady(c, n, timeout) }()
}
failed := false
// TODO(mbforbes): Change to `for range` syntax once we support only Go
// >= 1.4.
for i := range nodeList.Items {
_ = i
if !<-result {
failed = true
}
}
if failed {
return nodeNames, fmt.Errorf("at least one node failed to be ready")
}
return nodeNames, nil
}
// kubeletCommand performs `start`, `restart`, or `stop` on the kubelet running on the node of the target pod.
// Allowed kubeltOps are `kStart`, `kStop`, and `kRestart`
func kubeletCommand(kOp kubeletOpt, c clientset.Interface, pod *v1.Pod) {
nodeIP, err := framework.GetHostExternalAddress(c, pod)
Expect(err).NotTo(HaveOccurred())
nodeIP = nodeIP + ":22"
sshResult, err := framework.SSH("sudo /etc/init.d/kubelet "+string(kOp), nodeIP, framework.TestContext.Provider)
Expect(err).NotTo(HaveOccurred())
framework.LogSSHResult(sshResult)
// On restart, waiting for node NotReady prevents a race condition where the node takes a few moments to leave the
// Ready state which in turn short circuits WaitForNodeToBeReady()
if kOp == kStop || kOp == kRestart {
if ok := framework.WaitForNodeToBeNotReady(c, pod.Spec.NodeName, NodeStateTimeout); !ok {
framework.Failf("Node %s failed to enter NotReady state", pod.Spec.NodeName)
}
}
if kOp == kStart || kOp == kRestart {
if ok := framework.WaitForNodeToBeReady(c, pod.Spec.NodeName, NodeStateTimeout); !ok {
framework.Failf("Node %s failed to enter Ready state", pod.Spec.NodeName)
}
}
}
// rebootNode takes node name on provider through the following steps using c:
// - ensures the node is ready
// - ensures all pods on the node are running and ready
// - reboots the node (by executing rebootCmd over ssh)
// - ensures the node reaches some non-ready state
// - ensures the node becomes ready again
// - ensures all pods on the node become running and ready again
//
// It returns true through result only if all of the steps pass; at the first
// failed step, it will return false through result and not run the rest.
func rebootNode(c *client.Client, provider, name, rebootCmd string) bool {
// Setup
ns := api.NamespaceSystem
ps := framework.NewPodStore(c, ns, labels.Everything(), fields.OneTermEqualSelector(api.PodHostField, name))
defer ps.Stop()
// Get the node initially.
framework.Logf("Getting %s", name)
node, err := c.Nodes().Get(name)
if err != nil {
framework.Logf("Couldn't get node %s", name)
return false
}
// Node sanity check: ensure it is "ready".
if !framework.WaitForNodeToBeReady(c, name, framework.NodeReadyInitialTimeout) {
return false
}
// Get all the pods on the node that don't have liveness probe set.
// Liveness probe may cause restart of a pod during node reboot, and the pod may not be running.
pods := ps.List()
podNames := []string{}
for _, p := range pods {
probe := false
for _, c := range p.Spec.Containers {
if c.LivenessProbe != nil {
probe = true
break
}
}
if !probe {
podNames = append(podNames, p.ObjectMeta.Name)
}
}
framework.Logf("Node %s has %d assigned pods with no liveness probes: %v", name, len(podNames), podNames)
// For each pod, we do a sanity check to ensure it's running / healthy
// or succeeded now, as that's what we'll be checking later.
if !framework.CheckPodsRunningReadyOrSucceeded(c, ns, podNames, framework.PodReadyBeforeTimeout) {
printStatusAndLogsForNotReadyPods(c, ns, podNames, pods)
return false
}
// Reboot the node.
if err = framework.IssueSSHCommand(rebootCmd, provider, node); err != nil {
framework.Logf("Error while issuing ssh command: %v", err)
return false
}
// Wait for some kind of "not ready" status.
if !framework.WaitForNodeToBeNotReady(c, name, rebootNodeNotReadyTimeout) {
return false
}
// Wait for some kind of "ready" status.
if !framework.WaitForNodeToBeReady(c, name, rebootNodeReadyAgainTimeout) {
return false
}
// Ensure all of the pods that we found on this node before the reboot are
// running / healthy, or succeeded.
if !framework.CheckPodsRunningReadyOrSucceeded(c, ns, podNames, rebootPodReadyAgainTimeout) {
newPods := ps.List()
printStatusAndLogsForNotReadyPods(c, ns, podNames, newPods)
return false
}
framework.Logf("Reboot successful on node %s", name)
return true
}
diskName, err := createPDWithRetry()
framework.ExpectNoError(err, "Error creating a pd")
host0Pod := testPDPod([]string{diskName}, host0Name, false, 1)
originalCount := len(nodes.Items)
containerName := "mycontainer"
nodeToDelete := &nodes.Items[0]
defer func() error {
By("Cleaning up PD-RW test env")
detachAndDeletePDs(diskName, []types.NodeName{host0Name})
nodeToDelete.ObjectMeta.SetResourceVersion("0")
// need to set the resource version or else the Create() fails
_, err := nodeClient.Create(nodeToDelete)
framework.ExpectNoError(err, "Unable to re-create the deleted node")
framework.ExpectNoError(WaitForGroupSize(framework.TestContext.CloudConfig.NodeInstanceGroup, int32(initialGroupSize)), "Unable to get the node group back to the original size")
framework.WaitForNodeToBeReady(f.ClientSet, nodeToDelete.Name, nodeStatusTimeout)
if len(nodes.Items) != originalCount {
return fmt.Errorf("The node count is not back to original count")
}
return nil
}()
By("submitting host0Pod to kubernetes")
_, err = podClient.Create(host0Pod)
framework.ExpectNoError(err, fmt.Sprintf("Failed to create host0pod: %v", err))
framework.ExpectNoError(f.WaitForPodRunningSlow(host0Pod.Name))
testFile := "/testpd1/tracker"
testFileContents := fmt.Sprintf("%v", mathrand.Int())
Expect(err).NotTo(HaveOccurred(), "Each pod should start running and responding")
By("choose a node with at least one pod - we will block some network traffic on this node")
label := labels.SelectorFromSet(labels.Set(map[string]string{"name": name}))
options := api.ListOptions{LabelSelector: label}
pods, err := c.Pods(ns).List(options) // list pods after all have been scheduled
Expect(err).NotTo(HaveOccurred())
nodeName := pods.Items[0].Spec.NodeName
node, err := c.Nodes().Get(nodeName)
Expect(err).NotTo(HaveOccurred())
By(fmt.Sprintf("block network traffic from node %s", node.Name))
performTemporaryNetworkFailure(c, ns, name, replicas, pods.Items[0].Name, node)
framework.Logf("Waiting %v for node %s to be ready once temporary network failure ends", resizeNodeReadyTimeout, node.Name)
if !framework.WaitForNodeToBeReady(c, node.Name, resizeNodeReadyTimeout) {
framework.Failf("Node %s did not become ready within %v", node.Name, resizeNodeReadyTimeout)
}
// sleep a bit, to allow Watch in NodeController to catch up.
time.Sleep(5 * time.Second)
By("verify whether new pods can be created on the re-attached node")
// increasing the RC size is not a valid way to test this
// since we have no guarantees the pod will be scheduled on our node.
additionalPod := "additionalpod"
err = newPodOnNode(c, ns, additionalPod, node.Name)
Expect(err).NotTo(HaveOccurred())
err = framework.VerifyPods(c, ns, additionalPod, true, 1)
Expect(err).NotTo(HaveOccurred())