node_controller.go

package node

import (
	"context"
	"encoding/json"
	"fmt"
	"reflect"
	"time"

	"github.com/golang/glog"
	configv1 "github.com/openshift/api/config/v1"
	cligoinformersv1 "github.com/openshift/client-go/config/informers/externalversions/config/v1"
	cligolistersv1 "github.com/openshift/client-go/config/listers/config/v1"
	"github.com/openshift/library-go/pkg/operator/v1helpers"
	"github.com/openshift/machine-config-operator/internal"
	mcfgv1 "github.com/openshift/machine-config-operator/pkg/apis/machineconfiguration.openshift.io/v1"
	ctrlcommon "github.com/openshift/machine-config-operator/pkg/controller/common"
	daemonconsts "github.com/openshift/machine-config-operator/pkg/daemon/constants"
	mcfgclientset "github.com/openshift/machine-config-operator/pkg/generated/clientset/versioned"
	"github.com/openshift/machine-config-operator/pkg/generated/clientset/versioned/scheme"
	mcfginformersv1 "github.com/openshift/machine-config-operator/pkg/generated/informers/externalversions/machineconfiguration.openshift.io/v1"
	mcfglistersv1 "github.com/openshift/machine-config-operator/pkg/generated/listers/machineconfiguration.openshift.io/v1"
	goerrs "github.com/pkg/errors"
	corev1 "k8s.io/api/core/v1"
	"k8s.io/apimachinery/pkg/api/errors"
	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
	"k8s.io/apimachinery/pkg/labels"
	"k8s.io/apimachinery/pkg/types"
	intstrutil "k8s.io/apimachinery/pkg/util/intstr"
	utilruntime "k8s.io/apimachinery/pkg/util/runtime"
	"k8s.io/apimachinery/pkg/util/strategicpatch"
	"k8s.io/apimachinery/pkg/util/wait"
	coreinformersv1 "k8s.io/client-go/informers/core/v1"
	clientset "k8s.io/client-go/kubernetes"
	coreclientsetv1 "k8s.io/client-go/kubernetes/typed/core/v1"
	corelisterv1 "k8s.io/client-go/listers/core/v1"
	"k8s.io/client-go/tools/cache"
	"k8s.io/client-go/tools/record"
	clientretry "k8s.io/client-go/util/retry"
	"k8s.io/client-go/util/workqueue"
)

const (
	// WorkerLabel defines the label associated with worker node.
	WorkerLabel = "node-role.kubernetes.io/worker"

	// maxRetries is the number of times a machineconfig pool will be retried before it is dropped out of the queue.
	// With the current rate-limiter in use (5ms*2^(maxRetries-1)) the following numbers represent the times
	// a machineconfig pool is going to be requeued:
	//
	// 5ms, 10ms, 20ms, 40ms, 80ms, 160ms, 320ms, 640ms, 1.3s, 2.6s, 5.1s, 10.2s, 20.4s, 41s, 82s
	maxRetries = 15

	// updateDelay is a pause to deal with churn in MachineConfigs; see
	// https://github.com/openshift/machine-config-operator/issues/301
	updateDelay = 5 * time.Second

	// osLabel is used to identify which type of OS the node has
	osLabel = "kubernetes.io/os"

	// schedulerCRName that we're interested in watching.
	schedulerCRName = "cluster"

	// masterPoolName is the control plane MachineConfigPool name
	masterPoolName = "master"
)

var nodeUpdateBackoff = wait.Backoff{
	Steps:    5,
	Duration: 100 * time.Millisecond,
	Jitter:   1.0,
}

// Controller defines the node controller.
type Controller struct {
	client        mcfgclientset.Interface
	kubeClient    clientset.Interface
	eventRecorder record.EventRecorder

	syncHandler              func(mcp string) error
	enqueueMachineConfigPool func(*mcfgv1.MachineConfigPool)

	mcpLister  mcfglistersv1.MachineConfigPoolLister
	nodeLister corelisterv1.NodeLister

	mcpListerSynced  cache.InformerSynced
	nodeListerSynced cache.InformerSynced

	schedulerList         cligolistersv1.SchedulerLister
	schedulerListerSynced cache.InformerSynced

	queue workqueue.RateLimitingInterface
}

// New returns a new node controller.
func New(
	mcpInformer mcfginformersv1.MachineConfigPoolInformer,
	nodeInformer coreinformersv1.NodeInformer,
	schedulerInformer cligoinformersv1.SchedulerInformer,
	kubeClient clientset.Interface,
	mcfgClient mcfgclientset.Interface,
) *Controller {
	eventBroadcaster := record.NewBroadcaster()
	eventBroadcaster.StartLogging(glog.Infof)
	eventBroadcaster.StartRecordingToSink(&coreclientsetv1.EventSinkImpl{Interface: kubeClient.CoreV1().Events("")})

	ctrl := &Controller{
		client:        mcfgClient,
		kubeClient:    kubeClient,
		eventRecorder: eventBroadcaster.NewRecorder(scheme.Scheme, corev1.EventSource{Component: "machineconfigcontroller-nodecontroller"}),
		queue:         workqueue.NewNamedRateLimitingQueue(workqueue.DefaultControllerRateLimiter(), "machineconfigcontroller-nodecontroller"),
	}

	mcpInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
		AddFunc:    ctrl.addMachineConfigPool,
		UpdateFunc: ctrl.updateMachineConfigPool,
		DeleteFunc: ctrl.deleteMachineConfigPool,
	})
	nodeInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
		AddFunc:    ctrl.addNode,
		UpdateFunc: ctrl.updateNode,
		DeleteFunc: ctrl.deleteNode,
	})
	schedulerInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
		AddFunc:    ctrl.checkMasterNodesOnAdd,
		UpdateFunc: ctrl.checkMasterNodesOnUpdate,
		DeleteFunc: ctrl.checkMasterNodesOnDelete,
	})
	ctrl.syncHandler = ctrl.syncMachineConfigPool
	ctrl.enqueueMachineConfigPool = ctrl.enqueueDefault

	ctrl.mcpLister = mcpInformer.Lister()
	ctrl.nodeLister = nodeInformer.Lister()
	ctrl.mcpListerSynced = mcpInformer.Informer().HasSynced
	ctrl.nodeListerSynced = nodeInformer.Informer().HasSynced

	ctrl.schedulerList = schedulerInformer.Lister()
	ctrl.schedulerListerSynced = schedulerInformer.Informer().HasSynced

	return ctrl
}

// Run executes the render controller.
func (ctrl *Controller) Run(workers int, stopCh <-chan struct{}) {
	defer utilruntime.HandleCrash()
	defer ctrl.queue.ShutDown()

	if !cache.WaitForCacheSync(stopCh, ctrl.mcpListerSynced, ctrl.nodeListerSynced, ctrl.schedulerListerSynced) {
		return
	}

	glog.Info("Starting MachineConfigController-NodeController")
	defer glog.Info("Shutting down MachineConfigController-NodeController")

	for i := 0; i < workers; i++ {
		go wait.Until(ctrl.worker, time.Second, stopCh)
	}

	<-stopCh
}

func (ctrl *Controller) getCurrentMasters() ([]*corev1.Node, error) {
	nodeList, err := ctrl.nodeLister.List(labels.SelectorFromSet(labels.Set{ctrlcommon.MasterLabel: ""}))
	if err != nil {
		return nil, fmt.Errorf("error while listing master nodes %v", err)
	}
	return nodeList, nil
}

// checkMasterNodesOnAdd makes the master nodes schedulable/unschedulable whenever scheduler config CR with name
// cluster is created
func (ctrl *Controller) checkMasterNodesOnAdd(obj interface{}) {
	ctrl.reconcileMasters()
}

// checkMasterNodesOnDelete makes the master nodes schedulable/unschedulable whenever scheduler config CR with name
// cluster is created
func (ctrl *Controller) checkMasterNodesOnDelete(obj interface{}) {
	scheduler := obj.(*configv1.Scheduler)
	if scheduler.Name != schedulerCRName {
		glog.V(4).Infof("We don't care about CRs other than cluster created for scheduler config")
		return
	}
	currentMasters, err := ctrl.getCurrentMasters()
	if err != nil {
		goerrs.Wrap(err, "Reconciling to make master nodes schedulable/unschedulable failed")
		return
	}
	// On deletion make all masters unschedulable to restore default behaviour
	errs := ctrl.makeMastersUnSchedulable(currentMasters)
	if len(errs) > 0 {
		err = v1helpers.NewMultiLineAggregate(errs)
		goerrs.Wrap(err, "Reconciling to make nodes schedulable/unschedulable failed")
		return
	}
	return
}

// checkMasterNodesonUpdate makes the master nodes schedulable/unschedulable whenever scheduler
// config CR with name cluster is updated
func (ctrl *Controller) checkMasterNodesOnUpdate(old, cur interface{}) {
	oldScheduler := old.(*configv1.Scheduler)
	curScheduler := cur.(*configv1.Scheduler)

	if oldScheduler.Name != schedulerCRName || curScheduler.Name != schedulerCRName {
		glog.V(4).Infof("We don't care about CRs other than cluster created for scheduler config")
		return
	}

	if reflect.DeepEqual(oldScheduler.Spec, curScheduler.Spec) {
		glog.V(4).Info("Scheduler config did not change")
		return
	}

	ctrl.reconcileMasters()
}

// makeMastersUnSchedulable makes all the masters in the cluster unschedulable
func (ctrl *Controller) makeMastersUnSchedulable(currentMasters []*corev1.Node) []error {
	var errs []error
	for _, node := range currentMasters {
		if err := ctrl.makeMasterNodeUnSchedulable(node); err != nil {
			errs = append(errs, fmt.Errorf("failed making node %v schedulable with error %v", node.Name, err))
		}
	}
	return errs

}

// makeMasterNodeUnSchedulable makes master node unschedulable by removing worker label and adding `NoSchedule`
// master taint to the master node
func (ctrl *Controller) makeMasterNodeUnSchedulable(node *corev1.Node) error {
	_, err := internal.UpdateNodeRetry(ctrl.kubeClient.CoreV1().Nodes(), ctrl.nodeLister, node.Name, func(node *corev1.Node) {
		// Remove worker label
		newLabels := node.Labels
		if _, hasWorkerLabel := newLabels[WorkerLabel]; hasWorkerLabel {
			delete(newLabels, WorkerLabel)
		}
		node.Labels = newLabels
		// Add master taint
		hasMasterTaint := false
		for _, taint := range node.Spec.Taints {
			if taint.Key == ctrlcommon.MasterLabel && taint.Effect == corev1.TaintEffectNoSchedule {
				hasMasterTaint = true
			}
		}
		if !hasMasterTaint {
			newTaints := node.Spec.Taints
			masterUnSchedulableTaint := corev1.Taint{Key: ctrlcommon.MasterLabel, Effect: corev1.TaintEffectNoSchedule}
			newTaints = append(newTaints, masterUnSchedulableTaint)
			node.Spec.Taints = newTaints
		}
	})
	if err != nil {
		return err
	}
	return nil
}

// makeMasterNodeSchedulable makes master node schedulable by removing NoSchedule master taint and
// adding worker label
func (ctrl *Controller) makeMasterNodeSchedulable(node *corev1.Node) error {
	_, err := internal.UpdateNodeRetry(ctrl.kubeClient.CoreV1().Nodes(), ctrl.nodeLister, node.Name, func(node *corev1.Node) {
		// Add worker label
		newLabels := node.Labels
		if _, hasWorkerLabel := newLabels[WorkerLabel]; !hasWorkerLabel {
			newLabels[WorkerLabel] = ""
		}
		node.Labels = newLabels
		// Remove master taint
		newTaints := []corev1.Taint{}
		for _, t := range node.Spec.Taints {
			if t.Key == ctrlcommon.MasterLabel && t.Effect == corev1.TaintEffectNoSchedule {
				continue
			}
			newTaints = append(newTaints, t)
		}
		node.Spec.Taints = newTaints
	})
	if err != nil {
		return err
	}
	return nil
}

func (ctrl *Controller) addMachineConfigPool(obj interface{}) {
	pool := obj.(*mcfgv1.MachineConfigPool)
	glog.V(4).Infof("Adding MachineConfigPool %s", pool.Name)
	ctrl.enqueueMachineConfigPool(pool)
}

func (ctrl *Controller) updateMachineConfigPool(old, cur interface{}) {
	oldPool := old.(*mcfgv1.MachineConfigPool)
	curPool := cur.(*mcfgv1.MachineConfigPool)

	glog.V(4).Infof("Updating MachineConfigPool %s", oldPool.Name)
	ctrl.enqueueMachineConfigPool(curPool)
}

func (ctrl *Controller) deleteMachineConfigPool(obj interface{}) {
	pool, ok := obj.(*mcfgv1.MachineConfigPool)
	if !ok {
		tombstone, ok := obj.(cache.DeletedFinalStateUnknown)
		if !ok {
			utilruntime.HandleError(fmt.Errorf("Couldn't get object from tombstone %#v", obj))
			return
		}
		pool, ok = tombstone.Obj.(*mcfgv1.MachineConfigPool)
		if !ok {
			utilruntime.HandleError(fmt.Errorf("Tombstone contained object that is not a MachineConfigPool %#v", obj))
			return
		}
	}
	glog.V(4).Infof("Deleting MachineConfigPool %s", pool.Name)
	// TODO(abhinavdahiya): handle deletes.
}

// Determine if masters are currently configured as schedulable
func (ctrl *Controller) getMastersSchedulable() (bool, error) {
	schedulerList, err := ctrl.schedulerList.List(labels.SelectorFromSet(nil))
	if err != nil {
		return false, fmt.Errorf("error while listing scheduler config %v", err)
	}
	for _, sched := range schedulerList {
		if sched.Name == schedulerCRName {
			return sched.Spec.MastersSchedulable, nil
		}
	}
	return false, nil
}

// Determine if a given Node is a master
func (ctrl *Controller) isMaster(node *corev1.Node) bool {
	_, master := node.ObjectMeta.Labels[ctrlcommon.MasterLabel]
	return master
}

// isWindows checks if given node is a Windows node or a Linux node
func isWindows(node *corev1.Node) bool {
	windowsOsValue := "windows"
	if value, ok := node.ObjectMeta.Labels[osLabel]; ok {
		if value == windowsOsValue {
			return true
		}
		return false
	}
	// All the nodes should have a OS label populated by kubelet, if not just to maintain
	// backwards compatibility, we can returning true here.
	return false
}

// Given a master Node, ensure it reflects the current mastersSchedulable setting
func (ctrl *Controller) reconcileMaster(node *corev1.Node) {
	mastersSchedulable, err := ctrl.getMastersSchedulable()
	if err != nil {
		goerrs.Wrap(err, "Getting scheduler config failed")
		return
	}
	if mastersSchedulable {
		err = ctrl.makeMasterNodeSchedulable(node)
		if err != nil {
			goerrs.Wrap(err, "Failed making master Node schedulable")
			return
		}
	} else if !mastersSchedulable {
		err = ctrl.makeMasterNodeUnSchedulable(node)
		if err != nil {
			goerrs.Wrap(err, "Failed making master Node unschedulable")
			return
		}
	}
}

// Get a list of current masters and apply scheduler config to them
// TODO: Taint reconciliation should happen elsewhere, in a generic taint/label reconciler
func (ctrl *Controller) reconcileMasters() {
	currentMasters, err := ctrl.getCurrentMasters()
	if err != nil {
		goerrs.Wrap(err, "Reconciling to make master nodes schedulable/unschedulable failed")
		return
	}
	for _, node := range currentMasters {
		ctrl.reconcileMaster(node)
	}
}

func (ctrl *Controller) addNode(obj interface{}) {
	node := obj.(*corev1.Node)
	if node.DeletionTimestamp != nil {
		ctrl.deleteNode(node)
		return
	}

	if ctrl.isMaster(node) {
		ctrl.reconcileMaster(node)
	}

	pools, err := ctrl.getPoolsForNode(node)
	if err != nil {
		glog.Errorf("error finding pools for node %s: %v", node.Name, err)
		return
	}
	if pools == nil {
		return
	}
	glog.V(4).Infof("Node %s added", node.Name)
	for _, pool := range pools {
		ctrl.enqueueMachineConfigPool(pool)
	}
}

func (ctrl *Controller) logPool(pool *mcfgv1.MachineConfigPool, format string, args ...interface{}) {
	msg := fmt.Sprintf(format, args...)
	glog.Infof("Pool %s: %s", pool.Name, msg)
}

func (ctrl *Controller) logPoolNode(pool *mcfgv1.MachineConfigPool, node *corev1.Node, format string, args ...interface{}) {
	msg := fmt.Sprintf(format, args...)
	glog.Infof("Pool %s: node %s: %s", pool.Name, node.Name, msg)
}

func (ctrl *Controller) updateNode(old, cur interface{}) {
	oldNode := old.(*corev1.Node)
	curNode := cur.(*corev1.Node)

	if !isNodeManaged(curNode) {
		return
	}

	if ctrl.isMaster(curNode) {
		ctrl.reconcileMaster(curNode)
	}

	pool, err := ctrl.getPrimaryPoolForNode(curNode)
	if err != nil {
		glog.Errorf("error finding pool for node: %v", err)
		return
	}
	if pool == nil {
		return
	}
	glog.V(4).Infof("Node %s updated", curNode.Name)

	var changed bool
	oldReadyErr := checkNodeReady(oldNode)
	newReadyErr := checkNodeReady(curNode)

	oldReady := getErrorString(oldReadyErr)
	newReady := getErrorString(newReadyErr)

	if oldReady != newReady {
		changed = true
		if newReadyErr != nil {
			ctrl.logPoolNode(pool, curNode, "Reporting unready: %v", newReadyErr)
		} else {
			ctrl.logPoolNode(pool, curNode, "Reporting ready")
		}
	}

	// Specifically log when a node has completed an update so the MCC logs are a useful central aggregate of state changes
	if oldNode.Annotations[daemonconsts.CurrentMachineConfigAnnotationKey] != oldNode.Annotations[daemonconsts.DesiredMachineConfigAnnotationKey] &&
		isNodeDone(curNode) {
		ctrl.logPoolNode(pool, curNode, "Completed update to %s", curNode.Annotations[daemonconsts.DesiredMachineConfigAnnotationKey])
		changed = true
	} else {
		annos := []string{
			daemonconsts.CurrentMachineConfigAnnotationKey,
			daemonconsts.DesiredMachineConfigAnnotationKey,
			daemonconsts.MachineConfigDaemonStateAnnotationKey,
		}
		for _, anno := range annos {
			newValue := curNode.Annotations[anno]
			if oldNode.Annotations[anno] != newValue {
				ctrl.logPoolNode(pool, curNode, "changed annotation %s = %s", anno, newValue)
				changed = true
				// For the control plane, emit events for these since they're important
				if pool.Name == masterPoolName {
					ctrl.eventRecorder.Eventf(pool, corev1.EventTypeNormal, "AnnotationChange", "Node %s now has %s=%s", curNode.Name, anno, newValue)
				}
			}
		}
		if !reflect.DeepEqual(oldNode.Labels, curNode.Labels) {
			ctrl.logPoolNode(pool, curNode, "changed labels")
			changed = true
		}
	}

	if !changed {
		return
	}

	pools, err := ctrl.getPoolsForNode(curNode)
	if err != nil {
		glog.Errorf("error finding pools for node: %v", err)
		return
	}
	if pools == nil {
		return
	}
	for _, pool := range pools {
		ctrl.enqueueMachineConfigPool(pool)
	}
}

func (ctrl *Controller) deleteNode(obj interface{}) {
	node, ok := obj.(*corev1.Node)

	if !ok {
		tombstone, ok := obj.(cache.DeletedFinalStateUnknown)
		if !ok {
			utilruntime.HandleError(fmt.Errorf("Couldn't get object from tombstone %#v", obj))
			return
		}
		node, ok = tombstone.Obj.(*corev1.Node)
		if !ok {
			utilruntime.HandleError(fmt.Errorf("Tombstone contained object that is not a Node %#v", obj))
			return
		}
	}

	pools, err := ctrl.getPoolsForNode(node)
	if err != nil {
		glog.Errorf("error finding pools for node: %v", err)
		return
	}
	if pools == nil {
		return
	}
	glog.V(4).Infof("Node %s delete", node.Name)
	for _, pool := range pools {
		ctrl.enqueueMachineConfigPool(pool)
	}
}

// getPoolsForNode chooses the MachineConfigPools that should be used for a given node.
// It disambiguates in the case where e.g. a node has both master/worker roles applied,
// and where a custom role may be used. It returns a slice of all the pools the node belongs to.
// It also ignores the Windows nodes.
func (ctrl *Controller) getPoolsForNode(node *corev1.Node) ([]*mcfgv1.MachineConfigPool, error) {
	if isWindows(node) {
		// This is not an error, is this a Windows Node and it won't be managed by MCO. We're explicitly logging
		// here at a high level to disambiguate this from other pools = nil  scenario
		glog.V(4).Infof("Node %v is a windows node so won't be managed by MCO", node.Name)
		return nil, nil
	}
	pl, err := ctrl.mcpLister.List(labels.Everything())
	if err != nil {
		return nil, err
	}

	var pools []*mcfgv1.MachineConfigPool
	for _, p := range pl {
		selector, err := metav1.LabelSelectorAsSelector(p.Spec.NodeSelector)
		if err != nil {
			return nil, fmt.Errorf("invalid label selector: %v", err)
		}

		// If a pool with a nil or empty selector creeps in, it should match nothing, not everything.
		if selector.Empty() || !selector.Matches(labels.Set(node.Labels)) {
			continue
		}

		pools = append(pools, p)
	}

	if len(pools) == 0 {
		// This is not an error, as there might be nodes in cluster that are not managed by machineconfigpool.
		return nil, nil
	}

	var master, worker *mcfgv1.MachineConfigPool
	var custom []*mcfgv1.MachineConfigPool
	for _, pool := range pools {
		if pool.Name == masterPoolName {
			master = pool
		} else if pool.Name == "worker" {
			worker = pool
		} else {
			custom = append(custom, pool)
		}
	}

	if len(custom) > 1 {
		return nil, fmt.Errorf("node %s belongs to %d custom roles, cannot proceed with this Node", node.Name, len(custom))
	} else if len(custom) == 1 {
		// We don't support making custom pools for masters
		if master != nil {
			return nil, fmt.Errorf("node %s has both master role and custom role %s", node.Name, custom[0].Name)
		}
		// One custom role, let's use its pool
		pls := []*mcfgv1.MachineConfigPool{custom[0]}
		if worker != nil {
			pls = append(pls, worker)
		}
		return pls, nil
	} else if master != nil {
		// In the case where a node is both master/worker, have it live under
		// the master pool. This occurs in CodeReadyContainers and general
		// "single node" deployments, which one may want to do for testing bare
		// metal, etc.
		return []*mcfgv1.MachineConfigPool{master}, nil
	}
	// Otherwise, it's a worker with no custom roles.
	return []*mcfgv1.MachineConfigPool{worker}, nil
}

// getPrimaryPoolForNode uses getPoolsForNode and returns the first one which is the one the node targets
func (ctrl *Controller) getPrimaryPoolForNode(node *corev1.Node) (*mcfgv1.MachineConfigPool, error) {
	pools, err := ctrl.getPoolsForNode(node)
	if err != nil {
		return nil, err
	}
	if pools == nil {
		return nil, nil
	}
	return pools[0], nil
}

func (ctrl *Controller) enqueue(pool *mcfgv1.MachineConfigPool) {
	key, err := cache.DeletionHandlingMetaNamespaceKeyFunc(pool)
	if err != nil {
		utilruntime.HandleError(fmt.Errorf("Couldn't get key for object %#v: %v", pool, err))
		return
	}

	ctrl.queue.Add(key)
}

func (ctrl *Controller) enqueueRateLimited(pool *mcfgv1.MachineConfigPool) {
	key, err := cache.DeletionHandlingMetaNamespaceKeyFunc(pool)
	if err != nil {
		utilruntime.HandleError(fmt.Errorf("Couldn't get key for object %#v: %v", pool, err))
		return
	}

	ctrl.queue.AddRateLimited(key)
}

// enqueueAfter will enqueue a pool after the provided amount of time.
func (ctrl *Controller) enqueueAfter(pool *mcfgv1.MachineConfigPool, after time.Duration) {
	key, err := cache.DeletionHandlingMetaNamespaceKeyFunc(pool)
	if err != nil {
		utilruntime.HandleError(fmt.Errorf("Couldn't get key for object %#v: %v", pool, err))
		return
	}

	ctrl.queue.AddAfter(key, after)
}

// enqueueDefault calls a default enqueue function
func (ctrl *Controller) enqueueDefault(pool *mcfgv1.MachineConfigPool) {
	ctrl.enqueueAfter(pool, updateDelay)
}

// worker runs a worker thread that just dequeues items, processes them, and marks them done.
// It enforces that the syncHandler is never invoked concurrently with the same key.
func (ctrl *Controller) worker() {
	for ctrl.processNextWorkItem() {
	}
}

func (ctrl *Controller) processNextWorkItem() bool {
	key, quit := ctrl.queue.Get()
	if quit {
		return false
	}
	defer ctrl.queue.Done(key)

	err := ctrl.syncHandler(key.(string))
	ctrl.handleErr(err, key)

	return true
}

func (ctrl *Controller) handleErr(err error, key interface{}) {
	if err == nil {
		ctrl.queue.Forget(key)
		return
	}

	if ctrl.queue.NumRequeues(key) < maxRetries {
		glog.V(2).Infof("Error syncing machineconfigpool %v: %v", key, err)
		ctrl.queue.AddRateLimited(key)
		return
	}

	utilruntime.HandleError(err)
	glog.V(2).Infof("Dropping machineconfigpool %q out of the queue: %v", key, err)
	ctrl.queue.Forget(key)
	ctrl.queue.AddAfter(key, 1*time.Minute)
}

// syncMachineConfigPool will sync the machineconfig pool with the given key.
// This function is not meant to be invoked concurrently with the same key.
func (ctrl *Controller) syncMachineConfigPool(key string) error {
	startTime := time.Now()
	glog.V(4).Infof("Started syncing machineconfigpool %q (%v)", key, startTime)
	defer func() {
		glog.V(4).Infof("Finished syncing machineconfigpool %q (%v)", key, time.Since(startTime))
	}()

	_, name, err := cache.SplitMetaNamespaceKey(key)
	if err != nil {
		return err
	}
	machineconfigpool, err := ctrl.mcpLister.Get(name)
	if errors.IsNotFound(err) {
		glog.V(2).Infof("MachineConfigPool %v has been deleted", key)
		return nil
	}
	if err != nil {
		return err
	}

	if machineconfigpool.Spec.Configuration.Name == "" {
		// Previously we spammed the logs about empty pools.
		// Let's just pause for a bit here to let the renderer
		// initialize them.
		glog.Infof("Pool %s is unconfigured, pausing %v for renderer to initialize", name, updateDelay)
		time.Sleep(updateDelay)
		return nil
	}

	// Deep-copy otherwise we are mutating our cache.
	// TODO: Deep-copy only when needed.
	pool := machineconfigpool.DeepCopy()
	everything := metav1.LabelSelector{}

	if reflect.DeepEqual(pool.Spec.NodeSelector, &everything) {
		ctrl.eventRecorder.Eventf(pool, corev1.EventTypeWarning, "SelectingAll", "This machineconfigpool is selecting all nodes. A non-empty selector is required.")
		return nil
	}

	if pool.DeletionTimestamp != nil {
		return ctrl.syncStatusOnly(pool)
	}

	if pool.Spec.Paused {
		return ctrl.syncStatusOnly(pool)
	}

	nodes, err := ctrl.getNodesForPool(pool)
	if err != nil {
		if syncErr := ctrl.syncStatusOnly(pool); syncErr != nil {
			return goerrs.Wrapf(err, "error getting nodes for pool %q, sync error: %v", pool.Name, syncErr)
		}
		return err
	}

	maxunavail, err := maxUnavailable(pool, nodes)
	if err != nil {
		if syncErr := ctrl.syncStatusOnly(pool); syncErr != nil {
			return goerrs.Wrapf(err, "error getting max unavailable count for pool %q, sync error: %v", pool.Name, syncErr)
		}
		return err
	}

	candidates, capacity := getAllCandidateMachines(pool, nodes, maxunavail)
	if len(candidates) > 0 {
		ctrl.logPool(pool, "%d candidate nodes for update, capacity: %d", len(candidates), capacity)
		if err := ctrl.updateCandidateMachines(pool, candidates, capacity); err != nil {
			if syncErr := ctrl.syncStatusOnly(pool); syncErr != nil {
				return goerrs.Wrapf(err, "error setting desired machine config annotation for pool %q, sync error: %v", pool.Name, syncErr)
			}
			return err
		}
	}
	return ctrl.syncStatusOnly(pool)
}

func (ctrl *Controller) getNodesForPool(pool *mcfgv1.MachineConfigPool) ([]*corev1.Node, error) {
	selector, err := metav1.LabelSelectorAsSelector(pool.Spec.NodeSelector)
	if err != nil {
		return nil, fmt.Errorf("invalid label selector: %v", err)
	}

	initialNodes, err := ctrl.nodeLister.List(selector)
	if err != nil {
		return nil, err
	}

	nodes := []*corev1.Node{}
	for _, n := range initialNodes {
		p, err := ctrl.getPrimaryPoolForNode(n)
		if err != nil {
			glog.Warningf("can't get pool for node %q: %v", n.Name, err)
			continue
		}
		if p == nil {
			continue
		}
		if p.Name != pool.Name {
			continue
		}
		nodes = append(nodes, n)
	}
	return nodes, nil
}

func (ctrl *Controller) setDesiredMachineConfigAnnotation(nodeName, currentConfig string) error {
	return clientretry.RetryOnConflict(nodeUpdateBackoff, func() error {
		oldNode, err := ctrl.kubeClient.CoreV1().Nodes().Get(context.TODO(), nodeName, metav1.GetOptions{})
		if err != nil {
			return err
		}
		oldData, err := json.Marshal(oldNode)
		if err != nil {
			return err
		}

		newNode := oldNode.DeepCopy()
		if newNode.Annotations == nil {
			newNode.Annotations = map[string]string{}
		}
		if newNode.Annotations[daemonconsts.DesiredMachineConfigAnnotationKey] == currentConfig {
			return nil
		}
		newNode.Annotations[daemonconsts.DesiredMachineConfigAnnotationKey] = currentConfig
		newData, err := json.Marshal(newNode)
		if err != nil {
			return err
		}

		patchBytes, err := strategicpatch.CreateTwoWayMergePatch(oldData, newData, corev1.Node{})
		if err != nil {
			return fmt.Errorf("failed to create patch for node %q: %v", nodeName, err)
		}
		_, err = ctrl.kubeClient.CoreV1().Nodes().Patch(context.TODO(), nodeName, types.StrategicMergePatchType, patchBytes, metav1.PatchOptions{})
		return err
	})
}

// getAllCandidateMachines returns all possible nodes which can be updated to the target config, along with a maximum
// capacity.  It is the reponsibility of the caller to choose a subset of the nodes given the capacity.
func getAllCandidateMachines(pool *mcfgv1.MachineConfigPool, nodesInPool []*corev1.Node, maxUnavailable int) ([]*corev1.Node, uint) {
	targetConfig := pool.Spec.Configuration.Name

	unavail := getUnavailableMachines(nodesInPool)
	// If we're at capacity, there's nothing to do.
	if len(unavail) >= maxUnavailable {
		return nil, 0
	}
	capacity := maxUnavailable - len(unavail)
	failingThisConfig := 0
	// We only look at nodes which aren't already targeting our desired config
	var nodes []*corev1.Node
	for _, node := range nodesInPool {
		if node.Annotations[daemonconsts.DesiredMachineConfigAnnotationKey] == targetConfig {
			if isNodeMCDFailing(node) {
				failingThisConfig++
			}
			continue
		}

		nodes = append(nodes, node)
	}

	// Nodes which are failing to target this config also count against
	// availability - it might be a transient issue, and if the issue
	// clears we don't want multiple to update at once.
	if failingThisConfig >= capacity {
		return nil, 0
	}
	capacity -= failingThisConfig

	return nodes, uint(capacity)
}

// getCandidateMachines returns the maximum subset of nodes which can be updated to the target config given availability constraints.
func getCandidateMachines(pool *mcfgv1.MachineConfigPool, nodesInPool []*corev1.Node, maxUnavailable int) []*corev1.Node {
	nodes, capacity := getAllCandidateMachines(pool, nodesInPool, maxUnavailable)
	if uint(len(nodes)) < capacity {
		return nodes
	}
	return nodes[:capacity]
}

// getCurrentEtcdLeader is not yet implemented
func (ctrl *Controller) getCurrentEtcdLeader(candidates []*corev1.Node) (*corev1.Node, error) {
	return nil, nil
}

// filterControlPlaneCandidateNodes adjusts the candidates and capacity specifically
// for the control plane, e.g. based on which node is the etcd leader at the time.
// nolint:unparam
func (ctrl *Controller) filterControlPlaneCandidateNodes(pool *mcfgv1.MachineConfigPool, candidates []*corev1.Node, capacity uint) ([]*corev1.Node, uint, error) {
	if len(candidates) <= 1 {
		return candidates, capacity, nil
	}
	etcdLeader, err := ctrl.getCurrentEtcdLeader(candidates)
	if err != nil {
		glog.Warningf("Failed to find current etcd leader (continuing anyways): %v", err)
	}
	var newCandidates []*corev1.Node
	for _, node := range candidates {
		if node == etcdLeader {
			// For now make this an event so we know it's working, even though it's more of a non-event
			ctrl.eventRecorder.Eventf(pool, corev1.EventTypeNormal, "DeferringEtcdLeaderUpdate", "Deferring update of etcd leader %s", node.Name)
			glog.Infof("Deferring update of etcd leader: %s", node.Name)
			continue
		}
		newCandidates = append(newCandidates, node)
	}
	return newCandidates, capacity, nil
}

// updateCandidateMachines sets the desiredConfig annotation the candidate machines
func (ctrl *Controller) updateCandidateMachines(pool *mcfgv1.MachineConfigPool, candidates []*corev1.Node, capacity uint) error {
	if pool.Name == masterPoolName {
		var err error
		candidates, capacity, err = ctrl.filterControlPlaneCandidateNodes(pool, candidates, capacity)
		if err != nil {
			return err
		}
		// In practice right now these counts will be 1 but let's stay general to support 5 etcd nodes in the future
		ctrl.logPool(pool, "filtered to %d candidate nodes for update, capacity: %d", len(candidates), capacity)
	}
	if capacity < uint(len(candidates)) {
		// Arbitrarily pick the first N candidates; no attempt at sorting.
		// Perhaps later we allow admins to weight somehow, or do something more intelligent.
		candidates = candidates[:capacity]
	}
	targetConfig := pool.Spec.Configuration.Name
	for _, node := range candidates {
		ctrl.logPool(pool, "Setting node %s target to %s", node.Name, targetConfig)
		if err := ctrl.setDesiredMachineConfigAnnotation(node.Name, targetConfig); err != nil {
			return goerrs.Wrapf(err, "setting desired config for node %s", node.Name)
		}
	}
	if len(candidates) == 1 {
		candidate := candidates[0]
		ctrl.eventRecorder.Eventf(pool, corev1.EventTypeNormal, "SetDesiredConfig", "Targeted node %s to config %s", candidate.Name, targetConfig)
	} else {
		ctrl.eventRecorder.Eventf(pool, corev1.EventTypeNormal, "SetDesiredConfig", "Set target for %d nodes to config %s", targetConfig)
	}
	return nil
}

func maxUnavailable(pool *mcfgv1.MachineConfigPool, nodes []*corev1.Node) (int, error) {
	intOrPercent := intstrutil.FromInt(1)
	if pool.Spec.MaxUnavailable != nil {
		intOrPercent = *pool.Spec.MaxUnavailable
	}
	maxunavail, err := intstrutil.GetValueFromIntOrPercent(&intOrPercent, len(nodes), false)
	if err != nil {
		return 0, err
	}
	if maxunavail == 0 {
		maxunavail = 1
	}
	if pool.Name == masterPoolName {
		// calculate the fault tolerance dynamically for the master pool
		// to avoid risking losing etcd quorum.
		tolerance := len(nodes) - ((len(nodes) / 2) + 1)
		if maxunavail > tolerance {
			glog.Warningf("Refusing to honor master pool maxUnavailable %d to prevent losing etcd quorum, using %d instead", maxunavail, tolerance)
			return tolerance, nil
		}
	}
	return maxunavail, nil
}

// getErrorString returns error string if not nil and empty string if error is nil
func getErrorString(err error) string {
	if err != nil {
		return err.Error()
	}
	return ""
}