Merge pull request #128083 from carlory/fix-126662-kubelet

kubelet: fix a bug where kubelet wrongly drops the QOSClass field of the Pod's s status when it rejects a Pod

pkg/kubelet/kubelet.go

@@ -42,6 +42,7 @@ import (
 	"k8s.io/client-go/informers"
 	"k8s.io/mount-utils"
 
+	v1qos "k8s.io/kubernetes/pkg/apis/core/v1/helper/qos"
 	utilfs "k8s.io/kubernetes/pkg/util/filesystem"
 	netutils "k8s.io/utils/net"
 
@@ -2292,9 +2293,10 @@ func (kl *Kubelet) deletePod(pod *v1.Pod) error {
 func (kl *Kubelet) rejectPod(pod *v1.Pod, reason, message string) {
 	kl.recorder.Eventf(pod, v1.EventTypeWarning, reason, message)
 	kl.statusManager.SetPodStatus(pod, v1.PodStatus{
-		Phase:   v1.PodFailed,
-		Reason:  reason,
-		Message: "Pod was rejected: " + message})
+		QOSClass: v1qos.GetPodQOS(pod), // keep it as is
+		Phase:    v1.PodFailed,
+		Reason:   reason,
+		Message:  "Pod was rejected: " + message})
 }
 
 // canAdmitPod determines if a pod can be admitted, and gives a reason if it

test/e2e/common/node/pod_admission.go

@@ -20,8 +20,10 @@ import (
 	"context"
 
 	"github.com/onsi/ginkgo/v2"
+	"github.com/onsi/gomega"
 
 	v1 "k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/labels"
 	"k8s.io/kubernetes/test/e2e/framework"
@@ -36,7 +38,7 @@ var _ = SIGDescribe("PodOSRejection", framework.WithNodeConformance(), func() {
 	f := framework.NewDefaultFramework("pod-os-rejection")
 	f.NamespacePodSecurityLevel = admissionapi.LevelBaseline
 	ginkgo.Context("Kubelet", func() {
-		ginkgo.It("should reject pod when the node OS doesn't match pod's OS", func(ctx context.Context) {
+		ginkgo.It("[LinuxOnly] should reject pod when the node OS doesn't match pod's OS", func(ctx context.Context) {
 			linuxNode, err := findLinuxNode(ctx, f)
 			framework.ExpectNoError(err)
 			pod := &v1.Pod{
@@ -65,6 +67,83 @@ var _ = SIGDescribe("PodOSRejection", framework.WithNodeConformance(), func() {
 	})
 })
 
+var _ = SIGDescribe("PodRejectionStatus", func() {
+	f := framework.NewDefaultFramework("pod-rejection-status")
+	f.NamespacePodSecurityLevel = admissionapi.LevelBaseline
+	ginkgo.Context("Kubelet", func() {
+		ginkgo.It("should reject pod when the node didn't have enough resource", func(ctx context.Context) {
+			node, err := e2enode.GetRandomReadySchedulableNode(ctx, f.ClientSet)
+			framework.ExpectNoError(err, "Failed to get a ready schedulable node")
+
+			// Create a pod that requests more CPU than the node has
+			pod := &v1.Pod{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "pod-out-of-cpu",
+					Namespace: f.Namespace.Name,
+				},
+				Spec: v1.PodSpec{
+					Containers: []v1.Container{
+						{
+							Name:  "pod-out-of-cpu",
+							Image: imageutils.GetPauseImageName(),
+							Resources: v1.ResourceRequirements{
+								Requests: v1.ResourceList{
+									v1.ResourceCPU: resource.MustParse("1000000000000"), // requests more CPU than any node has
+								},
+							},
+						},
+					},
+				},
+			}
+
+			pod = e2epod.NewPodClient(f).Create(ctx, pod)
+
+			// Wait for the scheduler to update the pod status
+			err = e2epod.WaitForPodNameUnschedulableInNamespace(ctx, f.ClientSet, pod.Name, pod.Namespace)
+			framework.ExpectNoError(err)
+
+			// Fetch the pod to get the latest status which should be last one observed by the scheduler
+			// before it rejected the pod
+			pod, err = f.ClientSet.CoreV1().Pods(pod.Namespace).Get(ctx, pod.Name, metav1.GetOptions{})
+			framework.ExpectNoError(err)
+
+			// force assign the Pod to a node in order to get rejection status later
+			binding := &v1.Binding{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      pod.Name,
+					Namespace: pod.Namespace,
+					UID:       pod.UID,
+				},
+				Target: v1.ObjectReference{
+					Kind: "Node",
+					Name: node.Name,
+				},
+			}
+			err = f.ClientSet.CoreV1().Pods(pod.Namespace).Bind(ctx, binding, metav1.CreateOptions{})
+			framework.ExpectNoError(err)
+
+			// kubelet has rejected the pod
+			err = e2epod.WaitForPodFailedReason(ctx, f.ClientSet, pod, "OutOfcpu", f.Timeouts.PodStartShort)
+			framework.ExpectNoError(err)
+
+			// fetch the reject Pod and compare the status
+			gotPod, err := f.ClientSet.CoreV1().Pods(pod.Namespace).Get(ctx, pod.Name, metav1.GetOptions{})
+			framework.ExpectNoError(err)
+
+			// This detects if there are any new fields in Status that were dropped by the pod rejection.
+			// These new fields either should be kept by kubelet's admission or added explicitly in the list of fields that are having a different value or must be cleared.
+			expectedStatus := pod.Status.DeepCopy()
+			expectedStatus.Phase = gotPod.Status.Phase
+			expectedStatus.Conditions = nil
+			expectedStatus.Message = gotPod.Status.Message
+			expectedStatus.Reason = gotPod.Status.Reason
+			expectedStatus.StartTime = gotPod.Status.StartTime
+			// expectedStatus.QOSClass keep it as is
+			gomega.Expect(gotPod.Status).To(gomega.Equal(*expectedStatus))
+		})
+	})
+})
+
 // findLinuxNode finds a Linux node that is Ready and Schedulable
 func findLinuxNode(ctx context.Context, f *framework.Framework) (v1.Node, error) {
 	selector := labels.Set{"kubernetes.io/os": "linux"}.AsSelector()