Merge pull request #21641 from bprashanth/promisc

cluster/gce/config-default.sh

@@ -116,8 +116,8 @@ OPENCONTRAIL_TAG="${OPENCONTRAIL_TAG:-R2.20}"
 OPENCONTRAIL_KUBERNETES_TAG="${OPENCONTRAIL_KUBERNETES_TAG:-master}"
 OPENCONTRAIL_PUBLIC_SUBNET="${OPENCONTRAIL_PUBLIC_SUBNET:-10.1.0.0/16}"
 
-# Should the kubelet configure hairpin mode on the bridge?
-HAIRPIN_MODE="${HAIRPIN_MODE:-true}" # true, false
+# How should the kubelet configure hairpin mode?
+HAIRPIN_MODE="${HAIRPIN_MODE:-promiscuous-bridge}" # promiscuous-bridge, hairpin-veth, none
 
 # Optional: if set to true, kube-up will configure the cluster to run e2e tests.
 E2E_STORAGE_TEST_ENVIRONMENT=${KUBE_E2E_STORAGE_TEST_ENVIRONMENT:-false}

cluster/gce/config-test.sh

@@ -134,8 +134,8 @@ OPENCONTRAIL_TAG="${OPENCONTRAIL_TAG:-R2.20}"
 OPENCONTRAIL_KUBERNETES_TAG="${OPENCONTRAIL_KUBERNETES_TAG:-master}"
 OPENCONTRAIL_PUBLIC_SUBNET="${OPENCONTRAIL_PUBLIC_SUBNET:-10.1.0.0/16}"
 
-# Should the kubelet configure hairpin mode on the bridge?
-HAIRPIN_MODE="${HAIRPIN_MODE:-true}" # true, false
+# How should the kubelet configure hairpin mode?
+HAIRPIN_MODE="${HAIRPIN_MODE:-promiscuous-bridge}" # promiscuous-bridge, hairpin-veth, none
 
 # Optional: if set to true, kube-up will configure the cluster to run e2e tests.
 E2E_STORAGE_TEST_ENVIRONMENT=${KUBE_E2E_STORAGE_TEST_ENVIRONMENT:-false}

cluster/saltbase/salt/kubelet/default

@@ -107,12 +107,12 @@
 {% set system_container = "" -%}
 {% set kubelet_container = "" -%}
 {% set runtime_container = "" -%}
-{% if grains['os_family'] == 'Debian' -%}  
+{% if grains['os_family'] == 'Debian' -%}
   {% if pillar.get('is_systemd') %}
     {% set cgroup_root = "--cgroup-root=docker" -%}
   {% else %}
     {% set cgroup_root = "--cgroup-root=/" -%}
-    {% set system_container = "--system-cgroups=/system" -%}    
+    {% set system_container = "--system-cgroups=/system" -%}
     {% set runtime_container = "--runtime-cgroups=/docker-daemon" -%}
     {% set kubelet_container= "--kubelet-cgroups=/kubelet" -%}
   {% endif %}
@@ -151,12 +151,19 @@
   {% endif -%}
 {% endif -%}
 
-{% set configure_hairpin_mode = "--configure-hairpin-mode=true" -%}
-{% if pillar.get('hairpin_mode', '').lower() == 'false' %}
-  {% set configure_hairpin_mode = "--configure-hairpin-mode=false" %}
+# Don't pipe the --hairpin-mode flag by default. This allows the kubelet to pick
+# an appropriate value.
+{% set hairpin_mode = "" -%}
+# The master cannot see Services because it doesn't run kube-proxy, so we don't
+# need to make its container bridge promiscuous. We also don't want to set
+# the hairpin-veth flag on the master because it increases the chances of
+# running into the kernel bug described in #20096.
+{% if grains['roles'][0] == 'kubernetes-master' -%}
+  {% set hairpin_mode = "--hairpin-mode=none" -%}
+{% elif pillar['hairpin_mode'] is defined and pillar['hairpin_mode'] in ['promiscuous-bridge', 'hairpin-veth', 'none'] -%}
+  {% set hairpin_mode = "--hairpin-mode=" + pillar['hairpin_mode'] -%}
 {% endif -%}
 
-
 {% set kubelet_port = "" -%}
 {% if pillar['kubelet_port'] is defined -%}
   {% set kubelet_port="--port=" + pillar['kubelet_port'] %}
@@ -173,4 +180,4 @@
 {% endif -%}
 
 # test_args has to be kept at the end, so they'll overwrite any prior configuration
-DAEMON_ARGS="{{daemon_args}} {{api_servers_with_port}} {{debugging_handlers}} {{hostname_override}} {{cloud_provider}} {{config}} {{manifest_url}} --allow-privileged={{pillar['allow_privileged']}} {{log_level}} {{cluster_dns}} {{cluster_domain}} {{docker_root}} {{kubelet_root}} {{configure_cbr0}} {{non_masquerade_cidr}} {{cgroup_root}} {{system_container}} {{pod_cidr}} {{ master_kubelet_args }} {{cpu_cfs_quota}} {{network_plugin}} {{kubelet_port}} {{experimental_flannel_overlay}} {{ reconcile_cidr_args }} {{ configure_hairpin_mode }} {{enable_custom_metrics}} {{runtime_container}} {{kubelet_container}} {{test_args}}"
+DAEMON_ARGS="{{daemon_args}} {{api_servers_with_port}} {{debugging_handlers}} {{hostname_override}} {{cloud_provider}} {{config}} {{manifest_url}} --allow-privileged={{pillar['allow_privileged']}} {{log_level}} {{cluster_dns}} {{cluster_domain}} {{docker_root}} {{kubelet_root}} {{configure_cbr0}} {{non_masquerade_cidr}} {{cgroup_root}} {{system_container}} {{pod_cidr}} {{ master_kubelet_args }} {{cpu_cfs_quota}} {{network_plugin}} {{kubelet_port}} {{experimental_flannel_overlay}} {{ reconcile_cidr_args }} {{ hairpin_mode }} {{enable_custom_metrics}} {{runtime_container}} {{kubelet_container}} {{test_args}}"

cmd/kubelet/app/options/options.go

@@ -35,7 +35,6 @@ import (
 const (
 	defaultRootDir             = "/var/lib/kubelet"
 	experimentalFlannelOverlay = false
-	hairpinMode                = true
 )
 
 // KubeletServer encapsulates all of the parameters necessary for starting up
@@ -126,7 +125,7 @@ func NewKubeletServer() *KubeletServer {
 			KubeAPIBurst:                   10,
 			ExperimentalFlannelOverlay:     experimentalFlannelOverlay,
 			OutOfDiskTransitionFrequency:   unversioned.Duration{5 * time.Minute},
-			HairpinMode:                    hairpinMode,
+			HairpinMode:                    componentconfig.PromiscuousBridge,
 		},
 	}
 }
@@ -208,7 +207,7 @@ func (s *KubeletServer) AddFlags(fs *pflag.FlagSet) {
 	fs.StringVar(&s.RktPath, "rkt-path", s.RktPath, "Path of rkt binary. Leave empty to use the first rkt in $PATH.  Only used if --container-runtime='rkt'")
 	fs.StringVar(&s.RktStage1Image, "rkt-stage1-image", s.RktStage1Image, "image to use as stage1. Local paths and http/https URLs are supported. If empty, the 'stage1.aci' in the same directory as '--rkt-path' will be used")
 	fs.BoolVar(&s.ConfigureCBR0, "configure-cbr0", s.ConfigureCBR0, "If true, kubelet will configure cbr0 based on Node.Spec.PodCIDR.")
-	fs.BoolVar(&s.HairpinMode, "configure-hairpin-mode", s.HairpinMode, "If true, kubelet will set the hairpin mode flag on container interfaces. This allows endpoints of a Service to loadbalance back to themselves if they should try to access their own Service.")
+	fs.StringVar(&s.HairpinMode, "hairpin-mode", s.HairpinMode, "How should the kubelet setup hairpin NAT. This allows endpoints of a Service to loadbalance back to themselves if they should try to access their own Service. Valid values are \"promiscuous-bridge\", \"hairpin-veth\" and \"none\".")
 	fs.IntVar(&s.MaxPods, "max-pods", s.MaxPods, "Number of Pods that can run on this Kubelet.")
 	fs.StringVar(&s.DockerExecHandlerName, "docker-exec-handler", s.DockerExecHandlerName, "Handler to use when executing a command in a container. Valid values are 'native' and 'nsenter'. Defaults to 'native'.")
 	fs.StringVar(&s.NonMasqueradeCIDR, "non-masquerade-cidr", s.NonMasqueradeCIDR, "Traffic to IPs outside this range will use IP masquerade.")

cmd/kubelet/app/server.go

@@ -759,7 +759,7 @@ type KubeletConfig struct {
 	ExperimentalFlannelOverlay bool
 	NodeIP                     net.IP
 	ContainerRuntimeOptions    []kubecontainer.Option
-	HairpinMode                bool
+	HairpinMode                string
 }
 
 func CreateAndInitKubelet(kc *KubeletConfig) (k KubeletBootstrap, pc *config.PodConfig, err error) {

docs/admin/kubelet.md

@@ -77,7 +77,6 @@ kubelet
       --cluster-domain="": Domain for this cluster.  If set, kubelet will configure all containers to search this domain in addition to the host's search domains
       --config="": Path to the config file or directory of files
       --configure-cbr0[=false]: If true, kubelet will configure cbr0 based on Node.Spec.PodCIDR.
-      --configure-hairpin-mode[=true]: If true, kubelet will set the hairpin mode flag on container interfaces. This allows endpoints of a Service to loadbalance back to themselves if they should try to access their own Service.
       --container-runtime="docker": The container runtime to use. Possible values: 'docker', 'rkt'. Default: 'docker'.
       --containerized[=false]: Experimental support for running kubelet in a container.  Intended for testing. [default=false]
       --cpu-cfs-quota[=true]: Enable CPU CFS quota enforcement for containers that specify CPU limits
@@ -91,6 +90,7 @@ kubelet
       --experimental-flannel-overlay[=false]: Experimental support for starting the kubelet with the default overlay network (flannel). Assumes flanneld is already running in client mode. [default=false]
       --file-check-frequency=20s: Duration between checking config files for new data
       --google-json-key="": The Google Cloud Platform Service Account JSON Key to use for authentication.
+      --hairpin-mode="promiscuous-bridge": How should the kubelet setup hairpin NAT. This allows endpoints of a Service to loadbalance back to themselves if they should try to access their own Service. Valid values are "promiscuous-bridge", "hairpin-veth" and "none".
       --healthz-bind-address=127.0.0.1: The IP address for the healthz server to serve on, defaulting to 127.0.0.1 (set to 0.0.0.0 for all interfaces)
       --healthz-port=10248: The port of the localhost healthz endpoint
       --host-ipc-sources="*": Comma-separated list of sources from which the Kubelet allows pods to use the host ipc namespace. [default="*"]
@@ -152,7 +152,7 @@ kubelet
       --volume-stats-agg-period=1m0s: Specifies interval for kubelet to calculate and cache the volume disk usage for all pods and volumes.  To disable volume calculations, set to 0.  Default: '1m'
 ```
 
-###### Auto generated by spf13/cobra on 17-Feb-2016
+###### Auto generated by spf13/cobra on 18-Feb-2016
 
 
 <!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

hack/jenkins/job-configs/kubernetes-soak.yaml

@@ -96,7 +96,6 @@
             branch: 'master'
             provider-env: '{gce-provider-env}'
             job-env: |
-                export HAIRPIN_MODE="false"
                 export PROJECT="k8s-jkns-gce-soak"
         - 'gce-2':
             deploy-description: Clone of kubernetes-soak-weekly-deploy-gce.
@@ -104,6 +103,7 @@
             branch: 'master'
             provider-env: '{gce-provider-env}'
             job-env: |
+                export HAIRPIN_MODE="hairpin-veth"
                 export PROJECT="k8s-jkns-gce-soak-2"
         - 'gce-1.1':
             deploy-description: |

hack/verify-flags/exceptions.txt

@@ -108,3 +108,4 @@ test/e2e/host_path.go:			fmt.Sprintf("--retry_time=%d", retryDuration),
 test/images/mount-tester/mt.go:	flag.BoolVar(&breakOnExpectedContent, "break_on_expected_content", true, "Break out of loop on expected content, (use with --file_content_in_loop flag only)")
 test/images/mount-tester/mt.go:	flag.IntVar(&retryDuration, "retry_time", 180, "Retry time during the loop")
 test/images/mount-tester/mt.go:	flag.StringVar(&readFileContentInLoopPath, "file_content_in_loop", "", "Path to read the file content in loop from")
+pkg/kubelet/network/hairpin/hairpin.go:	hairpinModeRelativePath = "hairpin_mode"

hack/verify-flags/known-flags.txt

@@ -390,5 +390,5 @@ leader-elect-lease-duration
 leader-elect-renew-deadline
 leader-elect-retry-period
 watch-cache-sizes
-configure-hairpin-mode
 lock-file
+hairpin-mode

pkg/apis/componentconfig/types.generated.go

@@ -1094,6 +1094,32 @@ func (x *ProxyMode) CodecDecodeSelf(d *codec1978.Decoder) {
 	}
 }
 
+func (x HairpinMode) CodecEncodeSelf(e *codec1978.Encoder) {
+	var h codecSelfer1234
+	z, r := codec1978.GenHelperEncoder(e)
+	_, _, _ = h, z, r
+	yym1 := z.EncBinary()
+	_ = yym1
+	if false {
+	} else if z.HasExtensions() && z.EncExt(x) {
+	} else {
+		r.EncodeString(codecSelferC_UTF81234, string(x))
+	}
+}
+
+func (x *HairpinMode) CodecDecodeSelf(d *codec1978.Decoder) {
+	var h codecSelfer1234
+	z, r := codec1978.GenHelperDecoder(d)
+	_, _, _ = h, z, r
+	yym1 := z.DecBinary()
+	_ = yym1
+	if false {
+	} else if z.HasExtensions() && z.DecExt(x) {
+	} else {
+		*((*string)(x)) = r.DecodeString()
+	}
+}
+
 func (x *KubeletConfiguration) CodecEncodeSelf(e *codec1978.Encoder) {
 	var h codecSelfer1234
 	z, r := codec1978.GenHelperEncoder(e)
@@ -2354,17 +2380,17 @@ func (x *KubeletConfiguration) CodecEncodeSelf(e *codec1978.Encoder) {
 				_ = yym194
 				if false {
 				} else {
-					r.EncodeBool(bool(x.HairpinMode))
+					r.EncodeString(codecSelferC_UTF81234, string(x.HairpinMode))
 				}
 			} else {
 				z.EncSendContainerState(codecSelfer_containerMapKey1234)
-				r.EncodeString(codecSelferC_UTF81234, string("configureHairpinMode"))
+				r.EncodeString(codecSelferC_UTF81234, string("hairpinMode"))
 				z.EncSendContainerState(codecSelfer_containerMapValue1234)
 				yym195 := z.EncBinary()
 				_ = yym195
 				if false {
 				} else {
-					r.EncodeBool(bool(x.HairpinMode))
+					r.EncodeString(codecSelferC_UTF81234, string(x.HairpinMode))
 				}
 			}
 			if yyr2 || yy2arr2 {
@@ -3218,11 +3244,11 @@ func (x *KubeletConfiguration) codecDecodeSelfFromMap(l int, d *codec1978.Decode
 			} else {
 				x.ConfigureCBR0 = bool(r.DecodeBool())
 			}
-		case "configureHairpinMode":
+		case "hairpinMode":
 			if r.TryDecodeAsNil() {
-				x.HairpinMode = false
+				x.HairpinMode = ""
 			} else {
-				x.HairpinMode = bool(r.DecodeBool())
+				x.HairpinMode = string(r.DecodeString())
 			}
 		case "maxPods":
 			if r.TryDecodeAsNil() {
@@ -4373,9 +4399,9 @@ func (x *KubeletConfiguration) codecDecodeSelfFromArray(l int, d *codec1978.Deco
 	}
 	z.DecSendContainerState(codecSelfer_containerArrayElem1234)
 	if r.TryDecodeAsNil() {
-		x.HairpinMode = false
+		x.HairpinMode = ""
 	} else {
-		x.HairpinMode = bool(r.DecodeBool())
+		x.HairpinMode = string(r.DecodeString())
 	}
 	yyj91++
 	if yyhl91 {

pkg/apis/componentconfig/types.go

@@ -78,6 +78,24 @@ const (
 	ProxyModeIPTables  ProxyMode = "iptables"
 )
 
+// HairpinMode denotes how the kubelet should configure networking to handle
+// hairpin packets.
+type HairpinMode string
+
+// Enum settings for different ways to handle hairpin packets.
+const (
+	// Set the hairpin flag on the veth of containers in the respective
+	// container runtime.
+	HairpinVeth = "hairpin-veth"
+	// Make the container bridge promiscuous. This will force it to accept
+	// hairpin packets, even if the flag isn't set on ports of the bridge.
+	PromiscuousBridge = "promiscuous-bridge"
+	// Neither of the above. If the kubelet is started in this hairpin mode
+	// and kube-proxy is running in iptables mode, hairpin packets will be
+	// dropped by the container bridge.
+	HairpinNone = "none"
+)
+
 // TODO: curate the ordering and structure of this config object
 type KubeletConfiguration struct {
 	// config is the path to the config file or directory of files
@@ -252,11 +270,16 @@ type KubeletConfiguration struct {
 	// configureCBR0 enables the kublet to configure cbr0 based on
 	// Node.Spec.PodCIDR.
 	ConfigureCBR0 bool `json:"configureCbr0"`
-	// Should the kubelet set the hairpin flag on veth interfaces for containers
-	// it creates? Setting this flag allows endpoints in a Service to
-	// loadbalance back to themselves if they should try to access their own
-	// Service.
-	HairpinMode bool `json:"configureHairpinMode"`
+	// How should the kubelet configure the container bridge for hairpin packets.
+	// Setting this flag allows endpoints in a Service to loadbalance back to
+	// themselves if they should try to access their own Service. Values:
+	//   "promiscuous-bridge": make the container bridge promiscuous.
+	//   "hairpin-veth":       set the hairpin flag on container veth interfaces.
+	//   "none":               do nothing.
+	// Setting --configure-cbr0 to false implies that to achieve hairpin NAT
+	// one must set --hairpin-mode=veth-flag, because bridge assumes the
+	// existence of a container bridge named cbr0.
+	HairpinMode string `json:"hairpinMode"`
 	// maxPods is the number of pods that can run on this Kubelet.
 	MaxPods int `json:"maxPods"`
 	// dockerExecHandlerName is the handler to use when executing a command

pkg/kubelet/container_bridge.go

@@ -59,7 +59,7 @@ func createCBR0(wantCIDR *net.IPNet) error {
 	return nil
 }
 
-func ensureCbr0(wantCIDR *net.IPNet) error {
+func ensureCbr0(wantCIDR *net.IPNet, promiscuous bool) error {
 	exists, err := cbr0Exists()
 	if err != nil {
 		return err
@@ -80,7 +80,20 @@ func ensureCbr0(wantCIDR *net.IPNet) error {
 			glog.Error(err)
 			return err
 		}
-		return createCBR0(wantCIDR)
+		if err := createCBR0(wantCIDR); err != nil {
+			glog.Error(err)
+			return err
+		}
+	}
+	// Put the container bridge into promiscuous mode to force it to accept hairpin packets.
+	// TODO: Remove this once the kernel bug (#20096) is fixed.
+	if promiscuous {
+		// Checking if the bridge is in promiscuous mode is as expensive and more brittle than
+		// simply setting the flag everytime.
+		if err := exec.Command("ip", "link", "set", "cbr0", "promisc", "on").Run(); err != nil {
+			glog.Error(err)
+			return err
+		}
 	}
 	return nil
 }

pkg/kubelet/kubelet.go

@@ -40,6 +40,7 @@ import (
 	"k8s.io/kubernetes/pkg/api/resource"
 	"k8s.io/kubernetes/pkg/api/unversioned"
 	"k8s.io/kubernetes/pkg/api/validation"
+	"k8s.io/kubernetes/pkg/apis/componentconfig"
 	"k8s.io/kubernetes/pkg/client/cache"
 	clientset "k8s.io/kubernetes/pkg/client/clientset_generated/internalclientset"
 	"k8s.io/kubernetes/pkg/client/record"
@@ -204,7 +205,7 @@ func NewMainKubelet(
 	enableCustomMetrics bool,
 	volumeStatsAggPeriod time.Duration,
 	containerRuntimeOptions []kubecontainer.Option,
-	hairpinMode bool,
+	hairpinMode string,
 ) (*Kubelet, error) {
 	if rootDirectory == "" {
 		return nil, fmt.Errorf("invalid root directory %q", rootDirectory)
@@ -326,6 +327,7 @@ func NewMainKubelet(
 		outOfDiskTransitionFrequency: outOfDiskTransitionFrequency,
 		reservation:                  reservation,
 		enableCustomMetrics:          enableCustomMetrics,
+		hairpinMode:                  componentconfig.HairpinMode(hairpinMode),
 	}
 	// TODO: Factor out "StatsProvider" from Kubelet so we don't have a cyclic dependency
 	klet.resourceAnalyzer = stats.NewResourceAnalyzer(klet, volumeStatsAggPeriod)
@@ -357,10 +359,41 @@ func NewMainKubelet(
 
 	klet.podCache = kubecontainer.NewCache()
 
+	// The hairpin mode setting doesn't matter if:
+	// - We're not using a bridge network. This is hard to check because we might
+	//   be using a plugin. It matters if --configure-cbr0=true, and we currently
+	//   don't pipe it down to any plugins.
+	// - It's set to hairpin-veth for a container runtime that doesn't know how
+	//   to set the hairpin flag on the veth's of containers. Currently the
+	//   docker runtime is the only one that understands this.
+	// - It's set to "none" or an unrecognized string.
+	switch hairpinMode {
+	case componentconfig.PromiscuousBridge:
+		if !configureCBR0 {
+			glog.Warningf("Hairpin mode set to %v but configureCBR0 is false", hairpinMode)
+			break
+		}
+		fallthrough
+	case componentconfig.HairpinVeth:
+		if containerRuntime != "docker" {
+			glog.Warningf("Hairpin mode set to %v but container runtime is %v", hairpinMode, containerRuntime)
+			break
+		}
+		fallthrough
+	case componentconfig.HairpinNone:
+		if configureCBR0 {
+			glog.Warningf("Hairpin mode set to %q and configureCBR0 is true, this might result in loss of hairpin packets.", hairpinMode)
+			break
+		}
+		glog.Infof("Hairpin mode set to %q", hairpinMode)
+	default:
+		glog.Infof("Unrecognized hairpin mode setting %q, setting it to %v", hairpinMode, componentconfig.HairpinNone)
+		hairpinMode = componentconfig.HairpinNone
+	}
+
 	// Initialize the runtime.
 	switch containerRuntime {
 	case "docker":
-		glog.Infof("Hairpin mode set to %v", hairpinMode)
 		// Only supported one for now, continue.
 		klet.containerRuntime = dockertools.NewDockerManager(
 			dockerClient,
@@ -383,7 +416,7 @@ func NewMainKubelet(
 			imageBackOff,
 			serializeImagePulls,
 			enableCustomMetrics,
-			hairpinMode,
+			hairpinMode == componentconfig.HairpinVeth,
 			containerRuntimeOptions...,
 		)
 	case "rkt":
@@ -683,6 +716,11 @@ type Kubelet struct {
 
 	// support gathering custom metrics.
 	enableCustomMetrics bool
+
+	// How the Kubelet should setup hairpin NAT. Can take the values: "promiscuous-bridge"
+	// (make cbr0 promiscuous), "hairpin-veth" (set the hairpin flag on veth interfaces)
+	// or "none" (do nothing).
+	hairpinMode componentconfig.HairpinMode
 }
 
 // Validate given node IP belongs to the current host
@@ -2633,7 +2671,7 @@ func (kl *Kubelet) reconcileCBR0(podCIDR string) error {
 	}
 	// Set cbr0 interface address to first address in IPNet
 	cidr.IP.To4()[3] += 1
-	if err := ensureCbr0(cidr); err != nil {
+	if err := ensureCbr0(cidr, kl.hairpinMode == componentconfig.PromiscuousBridge); err != nil {
 		return err
 	}
 	if kl.shaper == nil {