Merge pull request kubernetes#72259 from bsalamat/fix_nominated_node

Fix a race in setting nominated node and the scheduling cycle after it.

Author: k8s-ci-robot

pkg/scheduler/core/generic_scheduler.go

@@ -362,7 +362,7 @@ func (g *genericScheduler) processPreemptionWithExtenders(
 // worth the complexity, especially because we generally expect to have a very
 // small number of nominated pods per node.
 func (g *genericScheduler) getLowerPriorityNominatedPods(pod *v1.Pod, nodeName string) []*v1.Pod {
-	pods := g.schedulingQueue.WaitingPodsForNode(nodeName)
+	pods := g.schedulingQueue.NominatedPodsForNode(nodeName)
 
 	if len(pods) == 0 {
 		return nil
@@ -509,7 +509,7 @@ func addNominatedPods(pod *v1.Pod, meta predicates.PredicateMetadata,
 		// This may happen only in tests.
 		return false, meta, nodeInfo
 	}
-	nominatedPods := queue.WaitingPodsForNode(nodeInfo.Node().Name)
+	nominatedPods := queue.NominatedPodsForNode(nodeInfo.Node().Name)
 	if nominatedPods == nil || len(nominatedPods) == 0 {
 		return false, meta, nodeInfo
 	}

pkg/scheduler/internal/queue/scheduling_queue.go

@@ -64,11 +64,14 @@ type SchedulingQueue interface {
 	MoveAllToActiveQueue()
 	AssignedPodAdded(pod *v1.Pod)
 	AssignedPodUpdated(pod *v1.Pod)
-	WaitingPodsForNode(nodeName string) []*v1.Pod
+	NominatedPodsForNode(nodeName string) []*v1.Pod
 	PendingPods() []*v1.Pod
 	// Close closes the SchedulingQueue so that the goroutine which is
 	// waiting to pop items can exit gracefully.
 	Close()
+	// UpdateNominatedPodForNode adds the given pod to the nominated pod map or
+	// updates it if it already exists.
+	UpdateNominatedPodForNode(pod *v1.Pod, nodeName string)
 	// DeleteNominatedPodIfExists deletes nominatedPod from internal cache
 	DeleteNominatedPodIfExists(pod *v1.Pod)
 	// NumUnschedulablePods returns the number of unschedulable pods exist in the SchedulingQueue.
@@ -152,9 +155,9 @@ func (f *FIFO) AssignedPodUpdated(pod *v1.Pod) {}
 // MoveAllToActiveQueue does nothing in FIFO as all pods are always in the active queue.
 func (f *FIFO) MoveAllToActiveQueue() {}
 
-// WaitingPodsForNode returns pods that are nominated to run on the given node,
+// NominatedPodsForNode returns pods that are nominated to run on the given node,
 // but FIFO does not support it.
-func (f *FIFO) WaitingPodsForNode(nodeName string) []*v1.Pod {
+func (f *FIFO) NominatedPodsForNode(nodeName string) []*v1.Pod {
 	return nil
 }
 
@@ -166,6 +169,9 @@ func (f *FIFO) Close() {
 // DeleteNominatedPodIfExists does nothing in FIFO.
 func (f *FIFO) DeleteNominatedPodIfExists(pod *v1.Pod) {}
 
+// UpdateNominatedPodForNode does nothing in FIFO.
+func (f *FIFO) UpdateNominatedPodForNode(pod *v1.Pod, nodeName string) {}
+
 // NumUnschedulablePods returns the number of unschedulable pods exist in the SchedulingQueue.
 func (f *FIFO) NumUnschedulablePods() int {
 	return 0
@@ -204,10 +210,9 @@ type PriorityQueue struct {
 	podBackoffQ *util.Heap
 	// unschedulableQ holds pods that have been tried and determined unschedulable.
 	unschedulableQ *UnschedulablePodsMap
-	// nominatedPods is a map keyed by a node name and the value is a list of
-	// pods which are nominated to run on the node. These are pods which can be in
-	// the activeQ or unschedulableQ.
-	nominatedPods map[string][]*v1.Pod
+	// nominatedPods is a structures that stores pods which are nominated to run
+	// on nodes.
+	nominatedPods *nominatedPodMap
 	// receivedMoveRequest is set to true whenever we receive a request to move a
 	// pod from the unschedulableQ to the activeQ, and is set to false, when we pop
 	// a pod from the activeQ. It indicates if we received a move request when a
@@ -257,7 +262,7 @@ func NewPriorityQueueWithClock(stop <-chan struct{}, clock util.Clock) *Priority
 		podBackoff:     util.CreatePodBackoffWithClock(1*time.Second, 10*time.Second, clock),
 		activeQ:        util.NewHeap(cache.MetaNamespaceKeyFunc, activeQComp),
 		unschedulableQ: newUnschedulablePodsMap(),
-		nominatedPods:  map[string][]*v1.Pod{},
+		nominatedPods:  newNominatedPodMap(),
 	}
 	pq.cond.L = &pq.lock
 	pq.podBackoffQ = util.NewHeap(cache.MetaNamespaceKeyFunc, pq.podsCompareBackoffCompleted)
@@ -272,49 +277,6 @@ func (p *PriorityQueue) run() {
 	go wait.Until(p.flushBackoffQCompleted, 1.0*time.Second, p.stop)
 }
 
-// addNominatedPodIfNeeded adds a pod to nominatedPods if it has a NominatedNodeName and it does not
-// already exist in the map. Adding an existing pod is not going to update the pod.
-func (p *PriorityQueue) addNominatedPodIfNeeded(pod *v1.Pod) {
-	nnn := NominatedNodeName(pod)
-	if len(nnn) <= 0 {
-		return
-	}
-	for _, np := range p.nominatedPods[nnn] {
-		if np.UID == pod.UID {
-			klog.V(4).Infof("Pod %v/%v already exists in the nominated map!", pod.Namespace, pod.Name)
-			return
-		}
-	}
-	p.nominatedPods[nnn] = append(p.nominatedPods[nnn], pod)
-}
-
-// deleteNominatedPodIfExists deletes a pod from the nominatedPods.
-// NOTE: this function assumes lock has been acquired in caller.
-func (p *PriorityQueue) deleteNominatedPodIfExists(pod *v1.Pod) {
-	nnn := NominatedNodeName(pod)
-	if len(nnn) <= 0 {
-		return
-	}
-	for i, np := range p.nominatedPods[nnn] {
-		if np.UID != pod.UID {
-			continue
-		}
-		p.nominatedPods[nnn] = append(p.nominatedPods[nnn][:i], p.nominatedPods[nnn][i+1:]...)
-		if len(p.nominatedPods[nnn]) == 0 {
-			delete(p.nominatedPods, nnn)
-		}
-		break
-	}
-}
-
-// updateNominatedPod updates a pod in the nominatedPods.
-func (p *PriorityQueue) updateNominatedPod(oldPod, newPod *v1.Pod) {
-	// Even if the nominated node name of the Pod is not changed, we must delete and add it again
-	// to ensure that its pointer is updated.
-	p.deleteNominatedPodIfExists(oldPod)
-	p.addNominatedPodIfNeeded(newPod)
-}
-
 // Add adds a pod to the active queue. It should be called only when a new pod
 // is added so there is no chance the pod is already in active/unschedulable/backoff queues
 func (p *PriorityQueue) Add(pod *v1.Pod) error {
@@ -326,14 +288,13 @@ func (p *PriorityQueue) Add(pod *v1.Pod) error {
 	}
 	if p.unschedulableQ.get(pod) != nil {
 		klog.Errorf("Error: pod %v/%v is already in the unschedulable queue.", pod.Namespace, pod.Name)
-		p.deleteNominatedPodIfExists(pod)
 		p.unschedulableQ.delete(pod)
 	}
 	// Delete pod from backoffQ if it is backing off
 	if err := p.podBackoffQ.Delete(pod); err == nil {
 		klog.Errorf("Error: pod %v/%v is already in the podBackoff queue.", pod.Namespace, pod.Name)
 	}
-	p.addNominatedPodIfNeeded(pod)
+	p.nominatedPods.add(pod, "")
 	p.cond.Broadcast()
 
 	return nil
@@ -357,7 +318,7 @@ func (p *PriorityQueue) AddIfNotPresent(pod *v1.Pod) error {
 	if err != nil {
 		klog.Errorf("Error adding pod %v/%v to the scheduling queue: %v", pod.Namespace, pod.Name, err)
 	} else {
-		p.addNominatedPodIfNeeded(pod)
+		p.nominatedPods.add(pod, "")
 		p.cond.Broadcast()
 	}
 	return err
@@ -420,7 +381,7 @@ func (p *PriorityQueue) AddUnschedulableIfNotPresent(pod *v1.Pod) error {
 	if !p.receivedMoveRequest && isPodUnschedulable(pod) {
 		p.backoffPod(pod)
 		p.unschedulableQ.addOrUpdate(pod)
-		p.addNominatedPodIfNeeded(pod)
+		p.nominatedPods.add(pod, "")
 		return nil
 	}
 
@@ -430,14 +391,14 @@ func (p *PriorityQueue) AddUnschedulableIfNotPresent(pod *v1.Pod) error {
 		if err != nil {
 			klog.Errorf("Error adding pod %v to the backoff queue: %v", pod.Name, err)
 		} else {
-			p.addNominatedPodIfNeeded(pod)
+			p.nominatedPods.add(pod, "")
 		}
 		return err
 	}
 
 	err := p.activeQ.Add(pod)
 	if err == nil {
-		p.addNominatedPodIfNeeded(pod)
+		p.nominatedPods.add(pod, "")
 		p.cond.Broadcast()
 	}
 	return err
@@ -523,14 +484,14 @@ func (p *PriorityQueue) Update(oldPod, newPod *v1.Pod) error {
 	if oldPod != nil {
 		// If the pod is already in the active queue, just update it there.
 		if _, exists, _ := p.activeQ.Get(oldPod); exists {
-			p.updateNominatedPod(oldPod, newPod)
+			p.nominatedPods.update(oldPod, newPod)
 			err := p.activeQ.Update(newPod)
 			return err
 		}
 
 		// If the pod is in the backoff queue, update it there.
 		if _, exists, _ := p.podBackoffQ.Get(oldPod); exists {
-			p.updateNominatedPod(oldPod, newPod)
+			p.nominatedPods.update(oldPod, newPod)
 			p.podBackoffQ.Delete(newPod)
 			err := p.activeQ.Add(newPod)
 			if err == nil {
@@ -542,7 +503,7 @@ func (p *PriorityQueue) Update(oldPod, newPod *v1.Pod) error {
 
 	// If the pod is in the unschedulable queue, updating it may make it schedulable.
 	if usPod := p.unschedulableQ.get(newPod); usPod != nil {
-		p.updateNominatedPod(oldPod, newPod)
+		p.nominatedPods.update(oldPod, newPod)
 		if isPodUpdated(oldPod, newPod) {
 			// If the pod is updated reset backoff
 			p.clearPodBackoff(newPod)
@@ -560,7 +521,7 @@ func (p *PriorityQueue) Update(oldPod, newPod *v1.Pod) error {
 	// If pod is not in any of the two queue, we put it in the active queue.
 	err := p.activeQ.Add(newPod)
 	if err == nil {
-		p.addNominatedPodIfNeeded(newPod)
+		p.nominatedPods.add(newPod, "")
 		p.cond.Broadcast()
 	}
 	return err
@@ -571,7 +532,7 @@ func (p *PriorityQueue) Update(oldPod, newPod *v1.Pod) error {
 func (p *PriorityQueue) Delete(pod *v1.Pod) error {
 	p.lock.Lock()
 	defer p.lock.Unlock()
-	p.deleteNominatedPodIfExists(pod)
+	p.nominatedPods.delete(pod)
 	err := p.activeQ.Delete(pod)
 	if err != nil { // The item was probably not found in the activeQ.
 		p.clearPodBackoff(pod)
@@ -663,16 +624,13 @@ func (p *PriorityQueue) getUnschedulablePodsWithMatchingAffinityTerm(pod *v1.Pod
 	return podsToMove
 }
 
-// WaitingPodsForNode returns pods that are nominated to run on the given node,
+// NominatedPodsForNode returns pods that are nominated to run on the given node,
 // but they are waiting for other pods to be removed from the node before they
 // can be actually scheduled.
-func (p *PriorityQueue) WaitingPodsForNode(nodeName string) []*v1.Pod {
+func (p *PriorityQueue) NominatedPodsForNode(nodeName string) []*v1.Pod {
 	p.lock.RLock()
 	defer p.lock.RUnlock()
-	if list, ok := p.nominatedPods[nodeName]; ok {
-		return list
-	}
-	return nil
+	return p.nominatedPods.podsForNode(nodeName)
 }
 
 // PendingPods returns all the pending pods in the queue. This function is
@@ -702,10 +660,20 @@ func (p *PriorityQueue) Close() {
 	p.cond.Broadcast()
 }
 
-// DeleteNominatedPodIfExists deletes pod from internal cache if it's a nominatedPod
+// DeleteNominatedPodIfExists deletes pod nominatedPods.
 func (p *PriorityQueue) DeleteNominatedPodIfExists(pod *v1.Pod) {
 	p.lock.Lock()
-	p.deleteNominatedPodIfExists(pod)
+	p.nominatedPods.delete(pod)
+	p.lock.Unlock()
+}
+
+// UpdateNominatedPodForNode adds a pod to the nominated pods of the given node.
+// This is called during the preemption process after a node is nominated to run
+// the pod. We update the structure before sending a request to update the pod
+// object to avoid races with the following scheduling cycles.
+func (p *PriorityQueue) UpdateNominatedPodForNode(pod *v1.Pod, nodeName string) {
+	p.lock.Lock()
+	p.nominatedPods.add(pod, nodeName)
 	p.lock.Unlock()
 }
 
@@ -762,3 +730,77 @@ func newUnschedulablePodsMap() *UnschedulablePodsMap {
 		keyFunc: util.GetPodFullName,
 	}
 }
+
+// nominatedPodMap is a structure that stores pods nominated to run on nodes.
+// It exists because nominatedNodeName of pod objects stored in the structure
+// may be different than what scheduler has here. We should be able to find pods
+// by their UID and update/delete them.
+type nominatedPodMap struct {
+	// nominatedPods is a map keyed by a node name and the value is a list of
+	// pods which are nominated to run on the node. These are pods which can be in
+	// the activeQ or unschedulableQ.
+	nominatedPods map[string][]*v1.Pod
+	// nominatedPodToNode is map keyed by a Pod UID to the node name where it is
+	// nominated.
+	nominatedPodToNode map[ktypes.UID]string
+}
+
+func (npm *nominatedPodMap) add(p *v1.Pod, nodeName string) {
+	// always delete the pod if it already exist, to ensure we never store more than
+	// one instance of the pod.
+	npm.delete(p)
+
+	nnn := nodeName
+	if len(nnn) == 0 {
+		nnn = NominatedNodeName(p)
+		if len(nnn) == 0 {
+			return
+		}
+	}
+	npm.nominatedPodToNode[p.UID] = nnn
+	for _, np := range npm.nominatedPods[nnn] {
+		if np.UID == p.UID {
+			klog.V(4).Infof("Pod %v/%v already exists in the nominated map!", p.Namespace, p.Name)
+			return
+		}
+	}
+	npm.nominatedPods[nnn] = append(npm.nominatedPods[nnn], p)
+}
+
+func (npm *nominatedPodMap) delete(p *v1.Pod) {
+	nnn, ok := npm.nominatedPodToNode[p.UID]
+	if !ok {
+		return
+	}
+	for i, np := range npm.nominatedPods[nnn] {
+		if np.UID == p.UID {
+			npm.nominatedPods[nnn] = append(npm.nominatedPods[nnn][:i], npm.nominatedPods[nnn][i+1:]...)
+			if len(npm.nominatedPods[nnn]) == 0 {
+				delete(npm.nominatedPods, nnn)
+			}
+			break
+		}
+	}
+	delete(npm.nominatedPodToNode, p.UID)
+}
+
+func (npm *nominatedPodMap) update(oldPod, newPod *v1.Pod) {
+	// We update irrespective of the nominatedNodeName changed or not, to ensure
+	// that pod pointer is updated.
+	npm.delete(oldPod)
+	npm.add(newPod, "")
+}
+
+func (npm *nominatedPodMap) podsForNode(nodeName string) []*v1.Pod {
+	if list, ok := npm.nominatedPods[nodeName]; ok {
+		return list
+	}
+	return nil
+}
+
+func newNominatedPodMap() *nominatedPodMap {
+	return &nominatedPodMap{
+		nominatedPods:      make(map[string][]*v1.Pod),
+		nominatedPodToNode: make(map[ktypes.UID]string),
+	}
+}