Only evict via live migration when only a single active pod is runnin…

…g for a VMI

Signed-off-by: David Vossel <[email protected]>

pkg/virt-controller/watch/application.go

@@ -499,6 +499,7 @@ func (vca *VirtControllerApp) initEvacuationController() {
 		vca.vmiInformer,
 		vca.migrationInformer,
 		vca.nodeInformer,
+		vca.kvPodInformer,
 		recorder,
 		vca.clientSet,
 		vca.clusterConfig,

pkg/virt-controller/watch/drain/evacuation/evacuation.go

@@ -34,6 +34,7 @@ type EvacuationController struct {
 	clientset             kubecli.KubevirtClient
 	Queue                 workqueue.RateLimitingInterface
 	vmiInformer           cache.SharedIndexInformer
+	vmiPodInformer        cache.SharedIndexInformer
 	migrationInformer     cache.SharedIndexInformer
 	recorder              record.EventRecorder
 	migrationExpectations *controller.UIDTrackingControllerExpectations
@@ -45,6 +46,7 @@ func NewEvacuationController(
 	vmiInformer cache.SharedIndexInformer,
 	migrationInformer cache.SharedIndexInformer,
 	nodeInformer cache.SharedIndexInformer,
+	vmiPodInformer cache.SharedIndexInformer,
 	recorder record.EventRecorder,
 	clientset kubecli.KubevirtClient,
 	clusterConfig *virtconfig.ClusterConfig,
@@ -55,6 +57,7 @@ func NewEvacuationController(
 		vmiInformer:           vmiInformer,
 		migrationInformer:     migrationInformer,
 		nodeInformer:          nodeInformer,
+		vmiPodInformer:        vmiPodInformer,
 		recorder:              recorder,
 		clientset:             clientset,
 		migrationExpectations: controller.NewUIDTrackingControllerExpectations(controller.NewControllerExpectations()),
@@ -463,13 +466,41 @@ func (c *EvacuationController) listVMIsOnNode(nodeName string) ([]*virtv1.Virtua
 	return vmis, nil
 }
 
+func (c *EvacuationController) numRunningPods(vmi *virtv1.VirtualMachineInstance) int {
+
+	objs, err := c.vmiPodInformer.GetIndexer().ByIndex(cache.NamespaceIndex, vmi.Namespace)
+	if err != nil {
+		return 0
+	}
+
+	running := 0
+	for _, obj := range objs {
+		pod := obj.(*k8sv1.Pod)
+
+		if pod.Status.Phase == k8sv1.PodSucceeded || pod.Status.Phase == k8sv1.PodFailed {
+			// not interested in terminated pods
+			continue
+		} else if !controller.IsControlledBy(pod, vmi) {
+			// not interested pods not associated with the vmi
+			continue
+		}
+		running++
+	}
+
+	return running
+}
+
 func (c *EvacuationController) filterRunningNonMigratingVMIs(vmis []*virtv1.VirtualMachineInstance, migrations []*virtv1.VirtualMachineInstanceMigration) (migrateable []*virtv1.VirtualMachineInstance, nonMigrateable []*virtv1.VirtualMachineInstance) {
 	lookup := map[string]bool{}
 	for _, migration := range migrations {
 		lookup[migration.Namespace+"/"+migration.Spec.VMIName] = true
 	}
 
 	for _, vmi := range vmis {
+		// vmi is shutting down
+		if vmi.IsFinal() || vmi.DeletionTimestamp != nil {
+			continue
+		}
 
 		// does not want to migrate
 		if vmi.Spec.EvictionStrategy == nil || *vmi.Spec.EvictionStrategy != virtv1.EvictionStrategyLiveMigrate {
@@ -480,12 +511,25 @@ func (c *EvacuationController) filterRunningNonMigratingVMIs(vmis []*virtv1.Virt
 			nonMigrateable = append(nonMigrateable, vmi)
 			continue
 		}
-		if exists := lookup[vmi.Namespace+"/"+vmi.Name]; !exists &&
-			!vmi.IsFinal() && vmi.DeletionTimestamp == nil {
-			// no migration exists,
-			// the vmi is running,
-			migrateable = append(migrateable, vmi)
+
+		hasMigration := lookup[vmi.Namespace+"/"+vmi.Name]
+		// already migrating
+		if hasMigration {
+			continue
+		}
+
+		if c.numRunningPods(vmi) > 1 {
+			// waiting on target/source pods from a previous migration to terminate
+			//
+			// We only want to create a migration when num pods == 1 or else we run the
+			// risk of invalidating our pdb which prevents the VMI from being evicted
+			continue
 		}
+
+		// no migration exists,
+		// the vmi is running,
+		// only one pod is currently active for vmi
+		migrateable = append(migrateable, vmi)
 	}
 	return migrateable, nonMigrateable
 }