KUBE-1815: pods with wildcard tolerations are ignored during drain

Author: cirisked

internal/k8s/kubernetes.go

@@ -457,7 +457,7 @@ func PartitionPodsForEviction(pods []*v1.Pod, castNamespace string, skipDeletedT
 			continue
 		}
 
-		if IsDaemonSetPod(p) || IsStaticPod(p) {
+		if IsDaemonSetPod(p) || IsStaticPod(p) || HasWildcardToleration(p) {
 			nonEvictable = append(nonEvictable, p)
 			continue
 		}
@@ -493,7 +493,19 @@ func IsControlledBy(p *v1.Pod, kind string) bool {
 }
 
 func IsNonEvictible(p *v1.Pod) bool {
-	return IsDaemonSetPod(p) || IsStaticPod(p)
+	return IsDaemonSetPod(p) || IsStaticPod(p) || HasWildcardToleration(p)
+}
+
+// HasWildcardToleration returns true if the pod has a toleration that matches all taints
+// (i.e. key is empty and operator is Exists). Such pods would be rescheduled back onto
+// a cordoned node since they tolerate the node.kubernetes.io/unschedulable taint.
+func HasWildcardToleration(p *v1.Pod) bool {
+	for _, t := range p.Spec.Tolerations {
+		if t.Key == "" && t.Operator == v1.TolerationOpExists {
+			return true
+		}
+	}
+	return false
 }
 
 // PatchNode patches a node with the given change function.

internal/k8s/kubernetes_test.go

@@ -1087,6 +1087,40 @@ func TestPartitionPodsForEviction(t *testing.T) {
 			skipDeletedTimeoutSecs: 60,
 			wantEvictableLen:       0,
 		},
+		{
+			name: "pod with wildcard toleration is non-evictable",
+			pods: []v1.Pod{
+				{
+					ObjectMeta: metav1.ObjectMeta{Name: "wildcard-pod", Namespace: "default"},
+					Spec: v1.PodSpec{
+						Tolerations: []v1.Toleration{
+							{Operator: v1.TolerationOpExists},
+						},
+					},
+					Status: v1.PodStatus{Phase: v1.PodRunning},
+				},
+			},
+			castNamespace:            testCastNamespace,
+			wantNonEvictableLen:      1,
+			wantNonEvictablePodNames: []string{"wildcard-pod"},
+		},
+		{
+			name: "pod with specific key toleration is evictable",
+			pods: []v1.Pod{
+				{
+					ObjectMeta: metav1.ObjectMeta{Name: "specific-toleration-pod", Namespace: "default"},
+					Spec: v1.PodSpec{
+						Tolerations: []v1.Toleration{
+							{Key: "node.kubernetes.io/unschedulable", Operator: v1.TolerationOpExists},
+						},
+					},
+					Status: v1.PodStatus{Phase: v1.PodRunning},
+				},
+			},
+			castNamespace:         testCastNamespace,
+			wantEvictableLen:      1,
+			wantEvictablePodNames: []string{"specific-toleration-pod"},
+		},
 		{
 			name: "mixed pods are partitioned correctly",
 			pods: []v1.Pod{