Merge pull request #73863 from gnufied/fix-attach-limit-delayed

Fix code to handle delayed binding volumes

pkg/scheduler/algorithm/predicates/csi_volume_predicate.go

@@ -20,6 +20,7 @@ import (
 	"fmt"
 
 	"k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/util/rand"
 	utilfeature "k8s.io/apiserver/pkg/util/feature"
 	"k8s.io/klog"
 	"k8s.io/kubernetes/pkg/features"
@@ -31,14 +32,18 @@ import (
 type CSIMaxVolumeLimitChecker struct {
 	pvInfo               PersistentVolumeInfo
 	pvcInfo              PersistentVolumeClaimInfo
+	scInfo               StorageClassInfo
+	randomVolumeIDPrefix string
 }
 
 // NewCSIMaxVolumeLimitPredicate returns a predicate for counting CSI volumes
 func NewCSIMaxVolumeLimitPredicate(
-	pvInfo PersistentVolumeInfo, pvcInfo PersistentVolumeClaimInfo) FitPredicate {
+	pvInfo PersistentVolumeInfo, pvcInfo PersistentVolumeClaimInfo, scInfo StorageClassInfo) FitPredicate {
 	c := &CSIMaxVolumeLimitChecker{
 		pvInfo:               pvInfo,
 		pvcInfo:              pvcInfo,
+		scInfo:               scInfo,
+		randomVolumeIDPrefix: rand.String(32),
 	}
 	return c.attachableLimitPredicate
 }
@@ -129,28 +134,70 @@ func (c *CSIMaxVolumeLimitChecker) filterAttachableVolumes(
 			continue
 		}
 
-		pvName := pvc.Spec.VolumeName
-		// TODO - the actual handling of unbound PVCs will be fixed by late binding design.
-		if pvName == "" {
-			klog.V(4).Infof("Persistent volume had no name for claim %s/%s", namespace, pvcName)
+		driverName, volumeHandle := c.getCSIDriver(pvc)
+		// if we can't find driver name or volume handle - we don't count this volume.
+		if driverName == "" || volumeHandle == "" {
 			continue
 		}
+		volumeLimitKey := volumeutil.GetCSIAttachLimitKey(driverName)
+		result[volumeHandle] = volumeLimitKey
+
+	}
+	return nil
+}
+
+func (c *CSIMaxVolumeLimitChecker) getCSIDriver(pvc *v1.PersistentVolumeClaim) (string, string) {
+	pvName := pvc.Spec.VolumeName
+	namespace := pvc.Namespace
+	pvcName := pvc.Name
+
+	placeHolderCSIDriver := ""
+	placeHolderHandle := ""
+	if pvName == "" {
+		klog.V(5).Infof("Persistent volume had no name for claim %s/%s", namespace, pvcName)
+		return c.getDriverNameFromSC(pvc)
+	}
 	pv, err := c.pvInfo.GetPersistentVolumeInfo(pvName)
 
 	if err != nil {
 		klog.V(4).Infof("Unable to look up PV info for PVC %s/%s and PV %s", namespace, pvcName, pvName)
-			continue
+		// If we can't fetch PV associated with PVC, may be it got deleted
+		// or PVC was prebound to a PVC that hasn't been created yet.
+		// fallback to using StorageClass for volume counting
+		return c.getDriverNameFromSC(pvc)
 	}
 
 	csiSource := pv.Spec.PersistentVolumeSource.CSI
 	if csiSource == nil {
-			klog.V(4).Infof("Not considering non-CSI volume %s/%s", namespace, pvcName)
-			continue
+		klog.V(5).Infof("Not considering non-CSI volume %s/%s", namespace, pvcName)
+		return placeHolderCSIDriver, placeHolderHandle
+	}
+	return csiSource.Driver, csiSource.VolumeHandle
 }
-		driverName := csiSource.Driver
-		volumeLimitKey := volumeutil.GetCSIAttachLimitKey(driverName)
-		result[csiSource.VolumeHandle] = volumeLimitKey
 
+func (c *CSIMaxVolumeLimitChecker) getDriverNameFromSC(pvc *v1.PersistentVolumeClaim) (string, string) {
+	namespace := pvc.Namespace
+	pvcName := pvc.Name
+	scName := pvc.Spec.StorageClassName
+
+	placeHolderCSIDriver := ""
+	placeHolderHandle := ""
+	if scName == nil {
+		// if StorageClass is not set or found, then PVC must be using immediate binding mode
+		// and hence it must be bound before scheduling. So it is safe to not count it.
+		klog.V(5).Infof("pvc %s/%s has no storageClass", namespace, pvcName)
+		return placeHolderCSIDriver, placeHolderHandle
 	}
-	return nil
+
+	storageClass, err := c.scInfo.GetStorageClassInfo(*scName)
+	if err != nil {
+		klog.V(5).Infof("no storage %s found for pvc %s/%s", *scName, namespace, pvcName)
+		return placeHolderCSIDriver, placeHolderHandle
+	}
+
+	// We use random prefix to avoid conflict with volume-ids. If PVC is bound in the middle
+	// predicate and there is another pod(on same node) that uses same volume then we will overcount
+	// the volume and consider both volumes as different.
+	volumeHandle := fmt.Sprintf("%s-%s/%s", c.randomVolumeIDPrefix, namespace, pvcName)
+	return storageClass.Provisioner, volumeHandle
 }

pkg/scheduler/algorithm/predicates/csi_volume_predicate_test.go

@@ -17,6 +17,7 @@ limitations under the License.
 package predicates
 
 import (
+	"fmt"
 	"reflect"
 	"testing"
 
@@ -35,7 +36,7 @@ func TestCSIVolumeCountPredicate(t *testing.T) {
 				{
 					VolumeSource: v1.VolumeSource{
 						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
-							ClaimName: "csi-ebs",
+							ClaimName: "csi-ebs-0",
 						},
 					},
 				},
@@ -77,38 +78,180 @@ func TestCSIVolumeCountPredicate(t *testing.T) {
 		},
 	}
 
+	pendingVolumePod := &v1.Pod{
+		Spec: v1.PodSpec{
+			Volumes: []v1.Volume{
+				{
+					VolumeSource: v1.VolumeSource{
+						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
+							ClaimName: "csi-4",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	// Different pod than pendingVolumePod, but using the same unbound PVC
+	unboundPVCPod2 := &v1.Pod{
+		Spec: v1.PodSpec{
+			Volumes: []v1.Volume{
+				{
+					VolumeSource: v1.VolumeSource{
+						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
+							ClaimName: "csi-4",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	missingPVPod := &v1.Pod{
+		Spec: v1.PodSpec{
+			Volumes: []v1.Volume{
+				{
+					VolumeSource: v1.VolumeSource{
+						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
+							ClaimName: "csi-6",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	noSCPVCPod := &v1.Pod{
+		Spec: v1.PodSpec{
+			Volumes: []v1.Volume{
+				{
+					VolumeSource: v1.VolumeSource{
+						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
+							ClaimName: "csi-5",
+						},
+					},
+				},
+			},
+		},
+	}
+	gceTwoVolPod := &v1.Pod{
+		Spec: v1.PodSpec{
+			Volumes: []v1.Volume{
+				{
+					VolumeSource: v1.VolumeSource{
+						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
+							ClaimName: "cs-gce-1",
+						},
+					},
+				},
+				{
+					VolumeSource: v1.VolumeSource{
+						PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
+							ClaimName: "csi-gce-2",
+						},
+					},
+				},
+			},
+		},
+	}
+
 	tests := []struct {
 		newPod       *v1.Pod
 		existingPods []*v1.Pod
 		filterName   string
 		maxVols      int
+		driverNames  []string
 		fits         bool
 		test         string
 	}{
 		{
 			newPod:       oneVolPod,
 			existingPods: []*v1.Pod{runningPod, twoVolPod},
-			filterName:   "csi-ebs",
+			filterName:   "csi",
 			maxVols:      4,
+			driverNames:  []string{"ebs"},
 			fits:         true,
 			test:         "fits when node capacity >= new pods CSI volume",
 		},
 		{
 			newPod:       oneVolPod,
 			existingPods: []*v1.Pod{runningPod, twoVolPod},
-			filterName:   "csi-ebs",
+			filterName:   "csi",
 			maxVols:      2,
+			driverNames:  []string{"ebs"},
 			fits:         false,
 			test:         "doesn't when node capacity <= pods CSI volume",
 		},
+		// should count pending PVCs
+		{
+			newPod:       oneVolPod,
+			existingPods: []*v1.Pod{pendingVolumePod, twoVolPod},
+			filterName:   "csi",
+			maxVols:      2,
+			driverNames:  []string{"ebs"},
+			fits:         false,
+			test:         "count pending PVCs towards capacity <= pods CSI volume",
+		},
+		// two same pending PVCs should be counted as 1
+		{
+			newPod:       oneVolPod,
+			existingPods: []*v1.Pod{pendingVolumePod, unboundPVCPod2, twoVolPod},
+			filterName:   "csi",
+			maxVols:      3,
+			driverNames:  []string{"ebs"},
+			fits:         true,
+			test:         "count multiple pending pvcs towards capacity >= pods CSI volume",
+		},
+		// should count PVCs with invalid PV name but valid SC
+		{
+			newPod:       oneVolPod,
+			existingPods: []*v1.Pod{missingPVPod, twoVolPod},
+			filterName:   "csi",
+			maxVols:      2,
+			driverNames:  []string{"ebs"},
+			fits:         false,
+			test:         "should count PVCs with invalid PV name but valid SC",
+		},
+		// don't count a volume which has storageclass missing
+		{
+			newPod:       oneVolPod,
+			existingPods: []*v1.Pod{runningPod, noSCPVCPod},
+			filterName:   "csi",
+			maxVols:      2,
+			driverNames:  []string{"ebs"},
+			fits:         true,
+			test:         "don't count pvcs with missing SC towards capacity",
+		},
+		// don't count multiple volume types
+		{
+			newPod:       oneVolPod,
+			existingPods: []*v1.Pod{gceTwoVolPod, twoVolPod},
+			filterName:   "csi",
+			maxVols:      2,
+			driverNames:  []string{"ebs", "gce"},
+			fits:         true,
+			test:         "don't count pvcs with different type towards capacity",
+		},
+		{
+			newPod:       gceTwoVolPod,
+			existingPods: []*v1.Pod{twoVolPod, runningPod},
+			filterName:   "csi",
+			maxVols:      2,
+			driverNames:  []string{"ebs", "gce"},
+			fits:         true,
+			test:         "don't count pvcs with different type towards capacity",
+		},
 	}
 
 	defer utilfeaturetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, features.AttachVolumeLimit, true)()
 	expectedFailureReasons := []PredicateFailureReason{ErrMaxVolumeCountExceeded}
 	// running attachable predicate tests with feature gate and limit present on nodes
 	for _, test := range tests {
-		node := getNodeWithPodAndVolumeLimits(test.existingPods, int64(test.maxVols), test.filterName)
-		pred := NewCSIMaxVolumeLimitPredicate(getFakeCSIPVInfo("csi-ebs", "csi-ebs"), getFakeCSIPVCInfo("csi-ebs"))
+		node := getNodeWithPodAndVolumeLimits(test.existingPods, int64(test.maxVols), test.driverNames...)
+		pred := NewCSIMaxVolumeLimitPredicate(getFakeCSIPVInfo(test.filterName, test.driverNames...),
+			getFakeCSIPVCInfo(test.filterName, "csi-sc", test.driverNames...),
+			getFakeCSIStorageClassInfo("csi-sc", test.driverNames[0]))
+
 		fits, reasons, err := pred(test.newPod, GetPredicateMetadata(test.newPod, nil), node)
 		if err != nil {
 			t.Errorf("Using allocatable [%s]%s: unexpected error: %v", test.filterName, test.test, err)
@@ -122,57 +265,63 @@ func TestCSIVolumeCountPredicate(t *testing.T) {
 	}
 }
 
-func getFakeCSIPVInfo(volumeName, driverName string) FakePersistentVolumeInfo {
-	return FakePersistentVolumeInfo{
-		{
-			ObjectMeta: metav1.ObjectMeta{Name: volumeName},
+func getFakeCSIPVInfo(volumeName string, driverNames ...string) FakePersistentVolumeInfo {
+	pvInfos := FakePersistentVolumeInfo{}
+	for _, driver := range driverNames {
+		for j := 0; j < 4; j++ {
+			volumeHandle := fmt.Sprintf("%s-%s-%d", volumeName, driver, j)
+			pv := v1.PersistentVolume{
+				ObjectMeta: metav1.ObjectMeta{Name: volumeHandle},
 				Spec: v1.PersistentVolumeSpec{
 					PersistentVolumeSource: v1.PersistentVolumeSource{
 						CSI: &v1.CSIPersistentVolumeSource{
-						Driver:       driverName,
-						VolumeHandle: volumeName,
-					},
-				},
-			},
-		},
-		{
-			ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-2"},
-			Spec: v1.PersistentVolumeSpec{
-				PersistentVolumeSource: v1.PersistentVolumeSource{
-					CSI: &v1.CSIPersistentVolumeSource{
-						Driver:       driverName,
-						VolumeHandle: volumeName + "-2",
-					},
-				},
-			},
-		},
-		{
-			ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-3"},
-			Spec: v1.PersistentVolumeSpec{
-				PersistentVolumeSource: v1.PersistentVolumeSource{
-					CSI: &v1.CSIPersistentVolumeSource{
-						Driver:       driverName,
-						VolumeHandle: volumeName + "-3",
-					},
+							Driver:       driver,
+							VolumeHandle: volumeHandle,
 						},
 					},
 				},
 			}
+			pvInfos = append(pvInfos, pv)
+		}
+
+	}
+	return pvInfos
+}
+
+func getFakeCSIPVCInfo(volumeName, scName string, driverNames ...string) FakePersistentVolumeClaimInfo {
+	pvcInfos := FakePersistentVolumeClaimInfo{}
+	for _, driver := range driverNames {
+		for j := 0; j < 4; j++ {
+			v := fmt.Sprintf("%s-%s-%d", volumeName, driver, j)
+			pvc := v1.PersistentVolumeClaim{
+				ObjectMeta: metav1.ObjectMeta{Name: v},
+				Spec:       v1.PersistentVolumeClaimSpec{VolumeName: v},
+			}
+			pvcInfos = append(pvcInfos, pvc)
+		}
 	}
 
-func getFakeCSIPVCInfo(volumeName string) FakePersistentVolumeClaimInfo {
-	return FakePersistentVolumeClaimInfo{
+	pvcInfos = append(pvcInfos, v1.PersistentVolumeClaim{
+		ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-4"},
+		Spec:       v1.PersistentVolumeClaimSpec{StorageClassName: &scName},
+	})
+	pvcInfos = append(pvcInfos, v1.PersistentVolumeClaim{
+		ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-5"},
+		Spec:       v1.PersistentVolumeClaimSpec{},
+	})
+	// a pvc with missing PV but available storageclass.
+	pvcInfos = append(pvcInfos, v1.PersistentVolumeClaim{
+		ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-6"},
+		Spec:       v1.PersistentVolumeClaimSpec{StorageClassName: &scName, VolumeName: "missing-in-action"},
+	})
+	return pvcInfos
+}
+
+func getFakeCSIStorageClassInfo(scName, provisionerName string) FakeStorageClassInfo {
+	return FakeStorageClassInfo{
 		{
-			ObjectMeta: metav1.ObjectMeta{Name: volumeName},
-			Spec:       v1.PersistentVolumeClaimSpec{VolumeName: volumeName},
-		},
-		{
-			ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-2"},
-			Spec:       v1.PersistentVolumeClaimSpec{VolumeName: volumeName + "-2"},
-		},
-		{
-			ObjectMeta: metav1.ObjectMeta{Name: volumeName + "-3"},
-			Spec:       v1.PersistentVolumeClaimSpec{VolumeName: volumeName + "-3"},
+			ObjectMeta:  metav1.ObjectMeta{Name: scName},
+			Provisioner: provisionerName,
 		},
 	}
 }

pkg/scheduler/algorithm/predicates/max_attachable_volume_predicate_test.go

@@ -937,16 +937,17 @@ func TestMaxVolumeFuncM4(t *testing.T) {
 	}
 }
 
-func getNodeWithPodAndVolumeLimits(pods []*v1.Pod, limit int64, filter string) *schedulernodeinfo.NodeInfo {
+func getNodeWithPodAndVolumeLimits(pods []*v1.Pod, limit int64, driverNames ...string) *schedulernodeinfo.NodeInfo {
 	nodeInfo := schedulernodeinfo.NewNodeInfo(pods...)
 	node := &v1.Node{
 		ObjectMeta: metav1.ObjectMeta{Name: "node-for-max-pd-test-1"},
 		Status: v1.NodeStatus{
-			Allocatable: v1.ResourceList{
-				getVolumeLimitKey(filter): *resource.NewQuantity(limit, resource.DecimalSI),
-			},
+			Allocatable: v1.ResourceList{},
 		},
 	}
+	for _, driver := range driverNames {
+		node.Status.Allocatable[getVolumeLimitKey(driver)] = *resource.NewQuantity(limit, resource.DecimalSI)
+	}
 	nodeInfo.SetNode(node)
 	return nodeInfo
 }

pkg/scheduler/algorithmprovider/defaults/register_predicates.go

@@ -81,7 +81,7 @@ func init() {
 	factory.RegisterFitPredicateFactory(
 		predicates.MaxCSIVolumeCountPred,
 		func(args factory.PluginFactoryArgs) predicates.FitPredicate {
-			return predicates.NewCSIMaxVolumeLimitPredicate(args.PVInfo, args.PVCInfo)
+			return predicates.NewCSIMaxVolumeLimitPredicate(args.PVInfo, args.PVCInfo, args.StorageClassInfo)
 		},
 	)
 	factory.RegisterFitPredicateFactory(