Enhance subnet selection

charts/aws-vpc-cni/values.yaml

@@ -84,6 +84,7 @@ env:
   ENABLE_IPv4: "true"
   ENABLE_IPv6: "false"
   VPC_CNI_VERSION: "v1.16.4"
+  ENABLE_ADV_ENI_CFG: "true"
 
 # this flag enables you to use the match label that was present in the original daemonset deployed by EKS
 # You can then annotate and label the original aws-node resources and 'adopt' them into a helm release

config/master/aws-k8s-cni-cn.yaml

@@ -466,6 +466,8 @@ spec:
               value: "false"
             - name: VPC_CNI_VERSION
               value: "v1.16.4"
+            - name: ENABLE_ADV_ENI_CFG
+              value: "true"
             - name: WARM_ENI_TARGET
               value: "1"
             - name: WARM_PREFIX_TARGET

config/master/aws-k8s-cni-us-gov-east-1.yaml

@@ -466,6 +466,8 @@ spec:
               value: "false"
             - name: VPC_CNI_VERSION
               value: "v1.16.4"
+            - name: ENABLE_ADV_ENI_CFG
+              value: "true"
             - name: WARM_ENI_TARGET
               value: "1"
             - name: WARM_PREFIX_TARGET

config/master/aws-k8s-cni-us-gov-west-1.yaml

@@ -466,6 +466,8 @@ spec:
               value: "false"
             - name: VPC_CNI_VERSION
               value: "v1.16.4"
+            - name: ENABLE_ADV_ENI_CFG
+              value: "true"
             - name: WARM_ENI_TARGET
               value: "1"
             - name: WARM_PREFIX_TARGET

config/master/aws-k8s-cni.yaml

@@ -466,6 +466,8 @@ spec:
               value: "false"
             - name: VPC_CNI_VERSION
               value: "v1.16.4"
+            - name: ENABLE_ADV_ENI_CFG
+              value: "true"
             - name: WARM_ENI_TARGET
               value: "1"
             - name: WARM_PREFIX_TARGET

pkg/awsutils/awsutils.go

@@ -22,6 +22,7 @@ import (
 	"net"
 	"os"
 	"regexp"
+	"sort"
 	"strings"
 	"sync"
 	"time"
@@ -58,6 +59,7 @@ const (
 	eniClusterTagKey        = "cluster.k8s.amazonaws.com/name"
 	additionalEniTagsEnvVar = "ADDITIONAL_ENI_TAGS"
 	reservedTagKeyPrefix    = "k8s.amazonaws.com"
+	subnetSelectionTagKey   = "kubernetes.io/role/cni"
 	// UnknownInstanceType indicates that the instance type is not yet supported
 	UnknownInstanceType = "vpc ip resource(eni ip limit): unknown instance type"
 
@@ -85,7 +87,7 @@ var log = logger.Get()
 // APIs defines interfaces calls for adding/getting/deleting ENIs/secondary IPs. The APIs are not thread-safe.
 type APIs interface {
 	// AllocENI creates an ENI and attaches it to the instance
-	AllocENI(useCustomCfg bool, sg []*string, subnet string, numIPs int) (eni string, err error)
+	AllocENI(useCustomCfg bool, sg []*string, eniCfgSubnet string, numIPs int) (eni string, err error)
 
 	// FreeENI detaches ENI interface and deletes it
 	FreeENI(eniName string) error
@@ -200,10 +202,12 @@ type EC2InstanceMetadataCache struct {
 	primaryENImac    string
 	availabilityZone string
 	region           string
+	vpcID            string
 
 	unmanagedENIs          StringSet
 	useCustomNetworking    bool
 	multiCardENIs          StringSet
+	useAdvancedENIConfig   bool
 	enablePrefixDelegation bool
 
 	clusterName       string
@@ -347,7 +351,7 @@ func (i instrumentedIMDS) GetMetadataWithContext(ctx context.Context, p string)
 }
 
 // New creates an EC2InstanceMetadataCache
-func New(useCustomNetworking, disableLeakedENICleanup, v4Enabled, v6Enabled bool) (*EC2InstanceMetadataCache, error) {
+func New(useAdvancedENIConfig, useCustomNetworking, disableLeakedENICleanup, v4Enabled, v6Enabled bool) (*EC2InstanceMetadataCache, error) {
 	// ctx is passed to initWithEC2Metadata func to cancel spawned go-routines when tests are run
 	ctx := context.Background()
 
@@ -367,6 +371,8 @@ func New(useCustomNetworking, disableLeakedENICleanup, v4Enabled, v6Enabled bool
 	log.Debugf("Discovered region: %s", cache.region)
 	cache.useCustomNetworking = useCustomNetworking
 	log.Infof("Custom networking enabled %v", cache.useCustomNetworking)
+	cache.useAdvancedENIConfig = useAdvancedENIConfig
+	log.Infof("Advanced ENI Config enabled %v", cache.useAdvancedENIConfig)
 	cache.v4Enabled = v4Enabled
 	cache.v6Enabled = v6Enabled
 
@@ -442,14 +448,22 @@ func (cache *EC2InstanceMetadataCache) initWithEC2Metadata(ctx context.Context)
 	}
 	log.Debugf("%s is the primary ENI of this instance", cache.primaryENI)
 
-	// retrieve sub-id
+	// retrieve subnet-id
 	cache.subnetID, err = cache.imds.GetSubnetID(ctx, mac)
 	if err != nil {
 		awsAPIErrInc("GetSubnetID", err)
 		return err
 	}
 	log.Debugf("Found subnet-id: %s ", cache.subnetID)
 
+	// retrieve vpc-id
+	cache.vpcID, err = cache.imds.GetVpcID(ctx, mac)
+	if err != nil {
+		awsAPIErrInc("GetVpcID", err)
+		return err
+	}
+	log.Debugf("Found vpc-id: %s ", cache.vpcID)
+
 	// We use the ctx here for testing, since we spawn go-routines above which will run forever.
 	select {
 	case <-ctx.Done():
@@ -730,8 +744,8 @@ func (cache *EC2InstanceMetadataCache) awsGetFreeDeviceNumber() (int, error) {
 
 // AllocENI creates an ENI and attaches it to the instance
 // returns: newly created ENI ID
-func (cache *EC2InstanceMetadataCache) AllocENI(useCustomCfg bool, sg []*string, subnet string, numIPs int) (string, error) {
-	eniID, err := cache.createENI(useCustomCfg, sg, subnet, numIPs)
+func (cache *EC2InstanceMetadataCache) AllocENI(useCustomCfg bool, sg []*string, eniCfgSubnet string, numIPs int) (string, error) {
+	eniID, err := cache.createENI(useCustomCfg, sg, eniCfgSubnet, numIPs)
 	if err != nil {
 		return "", errors.Wrap(err, "AllocENI: failed to create ENI")
 	}
@@ -803,7 +817,7 @@ func (cache *EC2InstanceMetadataCache) attachENI(eniID string) (string, error) {
 }
 
 // return ENI id, error
-func (cache *EC2InstanceMetadataCache) createENI(useCustomCfg bool, sg []*string, subnet string, numIPs int) (string, error) {
+func (cache *EC2InstanceMetadataCache) createENI(useCustomCfg bool, sg []*string, eniCfgSubnet string, numIPs int) (string, error) {
 	eniDescription := eniDescriptionPrefix + cache.instanceID
 	tags := map[string]string{
 		eniCreatedAtTagKey: time.Now().Format(time.RFC3339),
@@ -826,6 +840,8 @@ func (cache *EC2InstanceMetadataCache) createENI(useCustomCfg bool, sg []*string
 
 	log.Infof("Trying to allocate %d IP addresses on new ENI", needIPs)
 	log.Debugf("PD enabled - %t", cache.enablePrefixDelegation)
+
+	log.Info("Using same config as the primary interface for the new ENI")
 	input := &ec2.CreateNetworkInterfaceInput{}
 
 	if cache.enablePrefixDelegation {
@@ -845,33 +861,117 @@ func (cache *EC2InstanceMetadataCache) createENI(useCustomCfg bool, sg []*string
 			SecondaryPrivateIpAddressCount: aws.Int64(int64(needIPs)),
 		}
 	}
-	if useCustomCfg {
-		log.Info("Using a custom network config for the new ENI")
-		if len(sg) != 0 {
-			input.Groups = sg
+
+	var err error
+	var networkInterfaceID string
+	if cache.useCustomNetworking {
+		input = createENIUsingCustomCfg(sg, eniCfgSubnet, input)
+		log.Infof("Creating ENI with security groups: %v in subnet: %s", aws.StringValueSlice(input.Groups), aws.StringValue(input.SubnetId))
+
+		networkInterfaceID, err = cache.tryCreateNetworkInterface(input)
+		if err == nil {
+			return networkInterfaceID, nil
+		}
+	} else {
+		if cache.useAdvancedENIConfig {
+			subnetResult, vpcErr := cache.getVpcSubnets()
+			if vpcErr != nil {
+				return "", errors.Wrap(vpcErr, "failed to get subnets in vpc")
+			}
+
+			for _, subnet := range subnetResult {
+				if *subnet.SubnetId != cache.subnetID {
+					if !validTag(subnet) {
+						continue
+					}
+				}
+				log.Infof("Creating ENI with security groups: %v in subnet: %s", aws.StringValueSlice(input.Groups), aws.StringValue(input.SubnetId))
+
+				input.SubnetId = subnet.SubnetId
+				networkInterfaceID, err = cache.tryCreateNetworkInterface(input)
+				if err == nil {
+					return networkInterfaceID, nil
+				}
+			}
 		} else {
-			log.Warnf("No custom networking security group found, will use the node's primary ENI's SG: %v", aws.StringValueSlice(input.Groups))
+			networkInterfaceID, err = cache.tryCreateNetworkInterface(input)
+			if err == nil {
+				return networkInterfaceID, nil
+			}
+		}
+	}
+	return "", errors.Wrap(err, "failed to create network interface")
+}
+
+func (cache *EC2InstanceMetadataCache) getVpcSubnets() ([]*ec2.Subnet, error) {
+	describeSubnetInput := &ec2.DescribeSubnetsInput{
+		Filters: []*ec2.Filter{
+			{
+				Name:   aws.String("vpc-id"),
+				Values: []*string{aws.String(cache.vpcID)},
+			},
+			{
+				Name:   aws.String("availability-zone"),
+				Values: []*string{aws.String(cache.availabilityZone)},
+			},
+		},
+	}
+
+	start := time.Now()
+	subnetResult, err := cache.ec2SVC.DescribeSubnetsWithContext(context.Background(), describeSubnetInput)
+	prometheusmetrics.Ec2ApiReq.WithLabelValues("DescribeSubnets").Inc()
+	prometheusmetrics.AwsAPILatency.WithLabelValues("DescribeSubnets", fmt.Sprint(err != nil), awsReqStatus(err)).Observe(msSince(start))
+	if err != nil {
+		checkAPIErrorAndBroadcastEvent(err, "ec2:DescribeSubnets")
+		awsAPIErrInc("DescribeSubnets", err)
+		prometheusmetrics.Ec2ApiErr.WithLabelValues("DescribeSubnets").Inc()
+		return nil, errors.Wrap(err, "AllocENI: unable to describe subnets")
+	}
+
+	// Sort the subnet by available IP address counter (desc order) before determining subnet to use
+	sort.SliceStable(subnetResult.Subnets, func(i, j int) bool {
+		return *subnetResult.Subnets[j].AvailableIpAddressCount < *subnetResult.Subnets[i].AvailableIpAddressCount
+	})
+
+	return subnetResult.Subnets, nil
+}
+
+func validTag(subnet *ec2.Subnet) bool {
+	for _, tag := range subnet.Tags {
+		if *tag.Key == subnetSelectionTagKey {
+			return true
 		}
-		input.SubnetId = aws.String(subnet)
+	}
+	return false
+}
+
+func createENIUsingCustomCfg(sg []*string, eniCfgSubnet string, input *ec2.CreateNetworkInterfaceInput) *ec2.CreateNetworkInterfaceInput {
+	log.Info("Using a custom network config for the new ENI")
+
+	if len(sg) != 0 {
+		input.Groups = sg
 	} else {
-		log.Info("Using same config as the primary interface for the new ENI")
+		log.Warnf("No custom networking security group found, will use the node's primary ENI's SG: %v", aws.StringValueSlice(input.Groups))
 	}
+	input.SubnetId = aws.String(eniCfgSubnet)
 
-	log.Infof("Creating ENI with security groups: %v in subnet: %s", aws.StringValueSlice(input.Groups), aws.StringValue(input.SubnetId))
+	return input
+}
 
+func (cache *EC2InstanceMetadataCache) tryCreateNetworkInterface(input *ec2.CreateNetworkInterfaceInput) (string, error) {
 	start := time.Now()
 	result, err := cache.ec2SVC.CreateNetworkInterfaceWithContext(context.Background(), input)
 	prometheusmetrics.Ec2ApiReq.WithLabelValues("CreateNetworkInterface").Inc()
 	prometheusmetrics.AwsAPILatency.WithLabelValues("CreateNetworkInterface", fmt.Sprint(err != nil), awsReqStatus(err)).Observe(msSince(start))
-	if err != nil {
-		checkAPIErrorAndBroadcastEvent(err, "ec2:CreateNetworkInterface")
-		awsAPIErrInc("CreateNetworkInterface", err)
-		prometheusmetrics.Ec2ApiErr.WithLabelValues("CreateNetworkInterface").Inc()
-		log.Errorf("Failed to CreateNetworkInterface %v", err)
-		return "", errors.Wrap(err, "failed to create network interface")
-	}
-	log.Infof("Created a new ENI: %s", aws.StringValue(result.NetworkInterface.NetworkInterfaceId))
-	return aws.StringValue(result.NetworkInterface.NetworkInterfaceId), nil
+	if err == nil {
+		log.Infof("Created a new ENI: %s", aws.StringValue(result.NetworkInterface.NetworkInterfaceId))
+		return aws.StringValue(result.NetworkInterface.NetworkInterfaceId), nil
+	}
+	checkAPIErrorAndBroadcastEvent(err, "ec2:CreateNetworkInterface")
+	awsAPIErrInc("CreateNetworkInterface", err)
+	prometheusmetrics.Ec2ApiErr.WithLabelValues("CreateNetworkInterface").Inc()
+	log.Errorf("Failed to CreateNetworkInterface %v for subnet %s", err, *input.SubnetId)
+	return "", err
 }
 
 // buildENITags computes the desired AWS Tags for eni