Merge pull request #110291 from danwinship/kep-3178-iptables-cleanup-kubelet

Implement KEP-3178 "iptables cleanup" in kubelet

pkg/features/kube_features.go

@@ -438,6 +438,13 @@ const (
 	// Allows Job controller to manage Pod completions per completion index.
 	IndexedJob featuregate.Feature = "IndexedJob"
 
+	// owner: @danwinship
+	// kep: http://kep.k8s.io/3178
+	// alpha: v1.25
+	//
+	// Causes kubelet to no longer create legacy IPTables rules
+	IPTablesOwnershipCleanup featuregate.Feature = "IPTablesOwnershipCleanup"
+
 	// owner: @ahg
 	// beta: v1.23
 	//
@@ -938,6 +945,8 @@ var defaultKubernetesFeatureGates = map[featuregate.Feature]featuregate.FeatureS
 
 	IndexedJob: {Default: true, PreRelease: featuregate.GA, LockToDefault: true}, // remove in 1.26
 
+	IPTablesOwnershipCleanup: {Default: false, PreRelease: featuregate.Alpha},
+
 	JobMutableNodeSchedulingDirectives: {Default: true, PreRelease: featuregate.Beta},
 
 	JobReadyPods: {Default: true, PreRelease: featuregate.Beta},

pkg/kubelet/kubelet_network_linux.go

@@ -24,10 +24,11 @@ import (
 	"time"
 
 	"k8s.io/apimachinery/pkg/util/wait"
+	utilfeature "k8s.io/apiserver/pkg/util/feature"
 	"k8s.io/klog/v2"
+	"k8s.io/kubernetes/pkg/features"
 	utiliptables "k8s.io/kubernetes/pkg/util/iptables"
 	utilexec "k8s.io/utils/exec"
-	utilnet "k8s.io/utils/net"
 )
 
 const (
@@ -51,73 +52,61 @@ const (
 
 func (kl *Kubelet) initNetworkUtil() {
 	exec := utilexec.New()
-	// TODO: @khenidak review when there is no IPv6 iptables exec  what should happen here (note: no error returned from this func)
+	iptClients := []utiliptables.Interface{
-	ipv6Primary := kl.nodeIPs != nil && utilnet.IsIPv6(kl.nodeIPs[0])
+		utiliptables.New(exec, utiliptables.ProtocolIPv4),
-
+		utiliptables.New(exec, utiliptables.ProtocolIPv6),
-	var iptClients []utiliptables.Interface
-	var protocols []utiliptables.Protocol
-
-	// assume 4,6
-	protocols = append(protocols, utiliptables.ProtocolIPv4)
-	iptClients = append(iptClients, utiliptables.New(exec, utiliptables.ProtocolIPv4))
-
-	protocols = append(protocols, utiliptables.ProtocolIPv6)
-	iptClients = append(iptClients, utiliptables.New(exec, utiliptables.ProtocolIPv6))
-
-	// and if they are not
-	if ipv6Primary {
-		protocols[0], protocols[1] = protocols[1], protocols[0]
-		iptClients[0], iptClients[1] = iptClients[1], iptClients[0]
 	}
 
 	for i := range iptClients {
 		iptClient := iptClients[i]
-		if kl.syncNetworkUtil(iptClient) {
+		if kl.syncIPTablesRules(iptClient) {
-			klog.InfoS("Initialized protocol iptables rules.", "protocol", protocols[i])
+			klog.InfoS("Initialized iptables rules.", "protocol", iptClient.Protocol())
 			go iptClient.Monitor(
 				utiliptables.Chain("KUBE-KUBELET-CANARY"),
 				[]utiliptables.Table{utiliptables.TableMangle, utiliptables.TableNAT, utiliptables.TableFilter},
-				func() { kl.syncNetworkUtil(iptClient) },
+				func() { kl.syncIPTablesRules(iptClient) },
 				1*time.Minute, wait.NeverStop,
 			)
 		} else {
-			klog.InfoS("Failed to initialize protocol iptables rules; some functionality may be missing.", "protocol", protocols[i])
+			klog.InfoS("Failed to initialize iptables rules; some functionality may be missing.", "protocol", iptClient.Protocol())
 		}
 	}
 }
 
-// syncNetworkUtil ensures the network utility are present on host.
+// syncIPTablesRules ensures the KUBE-IPTABLES-HINT chain exists, and the martian packet
-// Network util includes:
+// protection rule is installed. If the IPTablesOwnershipCleanup feature gate is disabled
-//  1. In nat table, KUBE-MARK-DROP rule to mark connections for dropping
+// it will also synchronize additional deprecated iptables rules.
-//     Marked connection will be drop on INPUT/OUTPUT Chain in filter table
+func (kl *Kubelet) syncIPTablesRules(iptClient utiliptables.Interface) bool {
-//  2. In nat table, KUBE-MARK-MASQ rule to mark connections for SNAT
+	// Create hint chain so other components can see whether we are using iptables-legacy
-//     Marked connection will get SNAT on POSTROUTING Chain in nat table
+	// or iptables-nft.
-func (kl *Kubelet) syncNetworkUtil(iptClient utiliptables.Interface) bool {
+	if _, err := iptClient.EnsureChain(utiliptables.TableMangle, KubeIPTablesHintChain); err != nil {
-	// Setup KUBE-MARK-DROP rules
+		klog.ErrorS(err, "Failed to ensure that iptables hint chain exists")
-	dropMark := getIPTablesMark(kl.iptablesDropBit)
-	if _, err := iptClient.EnsureChain(utiliptables.TableNAT, KubeMarkDropChain); err != nil {
-		klog.ErrorS(err, "Failed to ensure that nat chain exists KUBE-MARK-DROP chain")
-		return false
-	}
-	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableNAT, KubeMarkDropChain, "-j", "MARK", "--or-mark", dropMark); err != nil {
-		klog.ErrorS(err, "Failed to ensure marking rule for KUBE-MARK-DROP chain")
-		return false
-	}
-	if _, err := iptClient.EnsureChain(utiliptables.TableFilter, KubeFirewallChain); err != nil {
-		klog.ErrorS(err, "Failed to ensure that filter table exists KUBE-FIREWALL chain")
-		return false
-	}
-	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableFilter, KubeFirewallChain,
-		"-m", "comment", "--comment", "kubernetes firewall for dropping marked packets",
-		"-m", "mark", "--mark", fmt.Sprintf("%s/%s", dropMark, dropMark),
-		"-j", "DROP"); err != nil {
-		klog.ErrorS(err, "Failed to ensure rule to drop packet marked by the KUBE-MARK-DROP in KUBE-FIREWALL chain")
 		return false
 	}
 
-	// drop all non-local packets to localhost if they're not part of an existing
-	// forwarded connection. See #90259
 	if !iptClient.IsIPv6() { // ipv6 doesn't have this issue
+		// Set up the KUBE-FIREWALL chain and martian packet protection rule.
+		// (See below.)
+		if _, err := iptClient.EnsureChain(utiliptables.TableFilter, KubeFirewallChain); err != nil {
+			klog.ErrorS(err, "Failed to ensure that filter table KUBE-FIREWALL chain exists")
+			return false
+		}
+
+		if _, err := iptClient.EnsureRule(utiliptables.Prepend, utiliptables.TableFilter, utiliptables.ChainOutput, "-j", string(KubeFirewallChain)); err != nil {
+			klog.ErrorS(err, "Failed to ensure that OUTPUT chain jumps to KUBE-FIREWALL")
+			return false
+		}
+		if _, err := iptClient.EnsureRule(utiliptables.Prepend, utiliptables.TableFilter, utiliptables.ChainInput, "-j", string(KubeFirewallChain)); err != nil {
+			klog.ErrorS(err, "Failed to ensure that INPUT chain jumps to KUBE-FIREWALL")
+			return false
+		}
+
+		// Kube-proxy's use of `route_localnet` to enable NodePorts on localhost
+		// creates a security hole (https://issue.k8s.io/90259) which this
+		// iptables rule mitigates. This rule should have been added to
+		// kube-proxy, but it mistakenly ended up in kubelet instead, and we are
+		// keeping it in kubelet for now in case other third-party components
+		// depend on it.
 		if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableFilter, KubeFirewallChain,
 			"-m", "comment", "--comment", "block incoming localnet connections",
 			"--dst", "127.0.0.0/8",
@@ -130,32 +119,61 @@ func (kl *Kubelet) syncNetworkUtil(iptClient utiliptables.Interface) bool {
 		}
 	}
 
-	if _, err := iptClient.EnsureRule(utiliptables.Prepend, utiliptables.TableFilter, utiliptables.ChainOutput, "-j", string(KubeFirewallChain)); err != nil {
+	if !utilfeature.DefaultFeatureGate.Enabled(features.IPTablesOwnershipCleanup) {
-		klog.ErrorS(err, "Failed to ensure that filter table  from OUTPUT chain jumps to KUBE-FIREWALL chain")
+		ok := kl.syncIPTablesRulesDeprecated(iptClient)
+		if !ok {
+			return false
+		}
+	}
+
+	return true
+}
+
+// syncIPTablesRulesDeprecated ensures deprecated iptables rules are present:
+//  1. In nat table, KUBE-MARK-DROP rule to mark connections for dropping
+//     Marked connection will be drop on INPUT/OUTPUT Chain in filter table
+//  2. In nat table, KUBE-MARK-MASQ rule to mark connections for SNAT
+//     Marked connection will get SNAT on POSTROUTING Chain in nat table
+func (kl *Kubelet) syncIPTablesRulesDeprecated(iptClient utiliptables.Interface) bool {
+	// Setup KUBE-MARK-DROP rules
+	dropMark := getIPTablesMark(kl.iptablesDropBit)
+	if _, err := iptClient.EnsureChain(utiliptables.TableNAT, KubeMarkDropChain); err != nil {
+		klog.ErrorS(err, "Failed to ensure that KUBE-MARK-DROP chain exists")
 		return false
 	}
-	if _, err := iptClient.EnsureRule(utiliptables.Prepend, utiliptables.TableFilter, utiliptables.ChainInput, "-j", string(KubeFirewallChain)); err != nil {
+	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableNAT, KubeMarkDropChain, "-j", "MARK", "--or-mark", dropMark); err != nil {
-		klog.ErrorS(err, "Failed to ensure that filter table INPUT chain jumps to KUBE-FIREWALL chain")
+		klog.ErrorS(err, "Failed to ensure that KUBE-MARK-DROP rule exists")
+		return false
+	}
+	if _, err := iptClient.EnsureChain(utiliptables.TableFilter, KubeFirewallChain); err != nil {
+		klog.ErrorS(err, "Failed to ensure that KUBE-FIREWALL chain exists")
+		return false
+	}
+	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableFilter, KubeFirewallChain,
+		"-m", "comment", "--comment", "kubernetes firewall for dropping marked packets",
+		"-m", "mark", "--mark", fmt.Sprintf("%s/%s", dropMark, dropMark),
+		"-j", "DROP"); err != nil {
+		klog.ErrorS(err, "Failed to ensure that KUBE-FIREWALL rule exists")
 		return false
 	}
 
 	// Setup KUBE-MARK-MASQ rules
 	masqueradeMark := getIPTablesMark(kl.iptablesMasqueradeBit)
 	if _, err := iptClient.EnsureChain(utiliptables.TableNAT, KubeMarkMasqChain); err != nil {
-		klog.ErrorS(err, "Failed to ensure that nat table exists KUBE-MARK-MASQ chain")
+		klog.ErrorS(err, "Failed to ensure that KUBE-MARK-MASQ chain exists")
 		return false
 	}
 	if _, err := iptClient.EnsureChain(utiliptables.TableNAT, KubePostroutingChain); err != nil {
-		klog.ErrorS(err, "Failed to ensure that nat table exists kube POSTROUTING chain")
+		klog.ErrorS(err, "Failed to ensure that KUBE-POSTROUTING chain exists")
 		return false
 	}
 	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableNAT, KubeMarkMasqChain, "-j", "MARK", "--or-mark", masqueradeMark); err != nil {
-		klog.ErrorS(err, "Failed to ensure marking rule for KUBE-MARK-MASQ chain")
+		klog.ErrorS(err, "Failed to ensure that KUBE-MARK-MASQ rule exists")
 		return false
 	}
 	if _, err := iptClient.EnsureRule(utiliptables.Prepend, utiliptables.TableNAT, utiliptables.ChainPostrouting,
 		"-m", "comment", "--comment", "kubernetes postrouting rules", "-j", string(KubePostroutingChain)); err != nil {
-		klog.ErrorS(err, "Failed to ensure that nat table from POSTROUTING chain jumps to KUBE-POSTROUTING chain")
+		klog.ErrorS(err, "Failed to ensure that POSTROUTING chain jumps to KUBE-POSTROUTING")
 		return false
 	}
 
@@ -165,7 +183,7 @@ func (kl *Kubelet) syncNetworkUtil(iptClient utiliptables.Interface) bool {
 	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableNAT, KubePostroutingChain,
 		"-m", "mark", "!", "--mark", fmt.Sprintf("%s/%s", masqueradeMark, masqueradeMark),
 		"-j", "RETURN"); err != nil {
-		klog.ErrorS(err, "Failed to ensure filtering rule for KUBE-POSTROUTING chain")
+		klog.ErrorS(err, "Failed to ensure first masquerading rule exists")
 		return false
 	}
 	// Clear the mark to avoid re-masquerading if the packet re-traverses the network stack.
@@ -173,7 +191,7 @@ func (kl *Kubelet) syncNetworkUtil(iptClient utiliptables.Interface) bool {
 	// to Sprintf another bitmask).
 	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableNAT, KubePostroutingChain,
 		"-j", "MARK", "--xor-mark", masqueradeMark); err != nil {
-		klog.ErrorS(err, "Failed to ensure unmarking rule for KUBE-POSTROUTING chain")
+		klog.ErrorS(err, "Failed to ensure second masquerading rule exists")
 		return false
 	}
 	masqRule := []string{
@@ -184,14 +202,7 @@ func (kl *Kubelet) syncNetworkUtil(iptClient utiliptables.Interface) bool {
 		masqRule = append(masqRule, "--random-fully")
 	}
 	if _, err := iptClient.EnsureRule(utiliptables.Append, utiliptables.TableNAT, KubePostroutingChain, masqRule...); err != nil {
-		klog.ErrorS(err, "Failed to ensure SNAT rule for packets marked by KUBE-MARK-MASQ chain in nat table KUBE-POSTROUTING chain")
+		klog.ErrorS(err, "Failed to ensure third masquerading rule exists")
-		return false
-	}
-
-	// Create hint chain so other components can see whether we are using iptables-legacy
-	// or iptables-nft.
-	if _, err := iptClient.EnsureChain(utiliptables.TableMangle, KubeIPTablesHintChain); err != nil {
-		klog.ErrorS(err, "Failed to ensure that iptables hint chain exists")
 		return false
 	}
 

test/e2e/network/networking.go

@@ -618,13 +618,13 @@ var _ = common.SIGDescribe("Networking", func() {
 
 		ginkgo.By("verifying that kubelet rules are eventually recreated")
 		err = utilwait.PollImmediate(framework.Poll, framework.RestartNodeReadyAgainTimeout, func() (bool, error) {
-			result, err = e2essh.SSH("sudo iptables-save -t nat", host, framework.TestContext.Provider)
+			result, err = e2essh.SSH("sudo iptables-save -t mangle", host, framework.TestContext.Provider)
 			if err != nil || result.Code != 0 {
 				e2essh.LogResult(result)
 				return false, err
 			}
 
-			if strings.Contains(result.Stdout, "\n-A KUBE-MARK-DROP ") {
+			if strings.Contains(result.Stdout, "\n:KUBE-IPTABLES-HINT") {
 				return true, nil
 			}
 			return false, nil