Merge pull request #12933 from k8s-infra-cherrypick-robot/cherry-pick-12923-to-release-1.11

[release-1.11] 🌱 CAPD: recreate container if we re-enter reconciliation and it exists but is not running

Author: k8s-ci-robot

test/infrastructure/docker/exp/internal/controllers/dockermachinepool_controller_phases.go

@@ -97,6 +97,14 @@ func createDockerContainer(ctx context.Context, name string, cluster *clusterv1.
 		}
 	}
 
+	// If re-entering the reconcile loop and reaching this point, the container is expected to be running. If it is not, delete it so we can try to create it again.
+	if externalMachine.Exists() && !externalMachine.IsRunning() {
+		// This deletes the machine and results in re-creating it below.
+		if err := externalMachine.Delete(ctx); err != nil {
+			return errors.Wrap(err, "Failed to delete not running DockerMachine")
+		}
+	}
+
 	log.Info("Creating container for machinePool", "name", name, "MachinePool", klog.KObj(machinePool))
 	if err := externalMachine.Create(ctx, dockerMachinePool.Spec.Template.CustomImage, constants.WorkerNodeRoleValue, machinePool.Spec.Template.Spec.Version, labels, dockerMachinePool.Spec.Template.ExtraMounts); err != nil {
 		return errors.Wrapf(err, "failed to create docker machine with name %s", name)

test/infrastructure/docker/internal/controllers/backends/docker/dockermachine_backend.go

@@ -193,6 +193,14 @@ func (r *MachineBackendReconciler) ReconcileNormal(ctx context.Context, cluster
 		role = constants.ControlPlaneNodeRoleValue
 	}
 
+	// If re-entering the reconcile loop and reaching this point, the container is expected to be running. If it is not, delete it so we can try to create it again.
+	if externalMachine.Exists() && !externalMachine.IsRunning() {
+		// This deletes the machine and results in re-creating it below.
+		if err := externalMachine.Delete(ctx); err != nil {
+			return ctrl.Result{}, errors.Wrap(err, "Failed to delete not running DockerMachine")
+		}
+	}
+
 	// Create the machine if not existing yet
 	if !externalMachine.Exists() {
 		// NOTE: FailureDomains don't mean much in CAPD since it's all local, but we are setting a label on

test/infrastructure/docker/internal/docker/machine.go

@@ -158,6 +158,15 @@ func (m *Machine) Exists() bool {
 	return m.container != nil
 }
 
+// IsRunning returns true if the container for this machine is running.
+func (m *Machine) IsRunning() bool {
+	if !m.Exists() {
+		return false
+	}
+
+	return m.container.IsRunning()
+}
+
 // Name returns the name of the machine.
 func (m *Machine) Name() string {
 	return m.machine
@@ -540,6 +549,8 @@ func (m *Machine) Delete(ctx context.Context) error {
 		if err := m.container.Delete(ctx); err != nil {
 			return err
 		}
+
+		m.container = nil
 	}
 	return nil
 }