Handle conntrack-related sysctls in supervisor agent setup

go.mod

@@ -33,6 +33,7 @@ replace (
 	github.com/prometheus/client_model => github.com/prometheus/client_model v0.0.0-20180712105110-5c3871d89910
 	github.com/prometheus/common => github.com/prometheus/common v0.0.0-20181126121408-4724e9255275
 	github.com/prometheus/procfs => github.com/prometheus/procfs v0.0.0-20181204211112-1dc9a6cbc91a
+	github.com/rancher/k3s/pkg/data => ./pkg/data
 	k8s.io/api => github.com/k3s-io/kubernetes/staging/src/k8s.io/api v1.18.19-k3s1
 	k8s.io/apiextensions-apiserver => github.com/k3s-io/kubernetes/staging/src/k8s.io/apiextensions-apiserver v1.18.19-k3s1
 	k8s.io/apimachinery => github.com/k3s-io/kubernetes/staging/src/k8s.io/apimachinery v1.18.19-k3s1
@@ -86,6 +87,7 @@ require (
 	github.com/go-bindata/go-bindata v3.1.2+incompatible
 	github.com/go-sql-driver/mysql v1.4.1
 	github.com/gogo/googleapis v1.3.0 // indirect
+	github.com/google/cadvisor v0.35.0
 	github.com/google/tcpproxy v0.0.0-20180808230851-dfa16c61dad2
 	github.com/gorilla/mux v1.7.3
 	github.com/gorilla/websocket v1.4.1
@@ -126,4 +128,5 @@ require (
 	k8s.io/cri-api v0.18.19
 	k8s.io/klog v1.0.0
 	k8s.io/kubernetes v1.18.19
+	k8s.io/utils v0.0.0-20200324210504-a9aa75ae1b89
 )

pkg/agent/run.go

@@ -2,7 +2,6 @@ package agent
 
 import (
 	"context"
-	"errors"
 	"io/ioutil"
 	"os"
 	"path/filepath"
@@ -12,6 +11,7 @@ import (
 	"k8s.io/apimachinery/pkg/labels"
 
 	systemd "github.com/coreos/go-systemd/daemon"
+	"github.com/pkg/errors"
 	"github.com/rancher/k3s/pkg/agent/config"
 	"github.com/rancher/k3s/pkg/agent/containerd"
 	"github.com/rancher/k3s/pkg/agent/flannel"
@@ -31,6 +31,9 @@ import (
 	"k8s.io/client-go/kubernetes"
 	v1 "k8s.io/client-go/kubernetes/typed/core/v1"
 	"k8s.io/client-go/tools/clientcmd"
+	app2 "k8s.io/kubernetes/cmd/kube-proxy/app"
+	kubeproxyconfig "k8s.io/kubernetes/pkg/proxy/apis/config"
+	utilpointer "k8s.io/utils/pointer"
 )
 
 const (
@@ -42,6 +45,12 @@ const (
 func run(ctx context.Context, cfg cmds.Agent, lb *loadbalancer.LoadBalancer) error {
 	nodeConfig := config.Get(ctx, cfg)
 
+	conntrackConfig, err := getConntrackConfig(nodeConfig)
+	if err != nil {
+		return errors.Wrap(err, "failed to validate kube-proxy conntrack configuration")
+	}
+	syssetup.Configure(conntrackConfig)
+
 	if !nodeConfig.NoFlannel {
 		if err := flannel.Prepare(ctx, nodeConfig); err != nil {
 			return err
@@ -86,6 +95,49 @@ func run(ctx context.Context, cfg cmds.Agent, lb *loadbalancer.LoadBalancer) err
 	return ctx.Err()
 }
 
+// getConntrackConfig uses the kube-proxy code to parse the user-provided kube-proxy-arg values, and
+// extract the conntrack settings so that K3s can set them itself. This allows us to soft-fail when
+// running K3s in Docker, where kube-proxy is no longer allowed to set conntrack sysctls on newer kernels.
+// When running rootless, we do not attempt to set conntrack sysctls - this behavior is copied from kubeadm.
+func getConntrackConfig(nodeConfig *daemonconfig.Node) (*kubeproxyconfig.KubeProxyConntrackConfiguration, error) {
+	ctConfig := &kubeproxyconfig.KubeProxyConntrackConfiguration{
+		MaxPerCore:            utilpointer.Int32Ptr(0),
+		Min:                   utilpointer.Int32Ptr(0),
+		TCPEstablishedTimeout: &metav1.Duration{},
+		TCPCloseWaitTimeout:   &metav1.Duration{},
+	}
+
+	if nodeConfig.AgentConfig.Rootless {
+		return ctConfig, nil
+	}
+
+	cmd := app2.NewProxyCommand()
+	if err := cmd.ParseFlags(daemonconfig.GetArgsList(map[string]string{}, nodeConfig.AgentConfig.ExtraKubeProxyArgs)); err != nil {
+		return nil, err
+	}
+	maxPerCore, err := cmd.Flags().GetInt32("conntrack-max-per-core")
+	if err != nil {
+		return nil, err
+	}
+	ctConfig.MaxPerCore = &maxPerCore
+	min, err := cmd.Flags().GetInt32("conntrack-min")
+	if err != nil {
+		return nil, err
+	}
+	ctConfig.Min = &min
+	establishedTimeout, err := cmd.Flags().GetDuration("conntrack-tcp-timeout-established")
+	if err != nil {
+		return nil, err
+	}
+	ctConfig.TCPEstablishedTimeout.Duration = establishedTimeout
+	closeWaitTimeout, err := cmd.Flags().GetDuration("conntrack-tcp-timeout-close-wait")
+	if err != nil {
+		return nil, err
+	}
+	ctConfig.TCPCloseWaitTimeout.Duration = closeWaitTimeout
+	return ctConfig, nil
+}
+
 func coreClient(cfg string) (kubernetes.Interface, error) {
 	restConfig, err := clientcmd.BuildConfigFromFlags("", cfg)
 	if err != nil {
@@ -99,7 +151,6 @@ func Run(ctx context.Context, cfg cmds.Agent) error {
 	if err := validate(); err != nil {
 		return err
 	}
-	syssetup.Configure()
 
 	if cfg.Rootless && !cfg.RootlessAlreadyUnshared {
 		if err := rootless.Rootless(cfg.DataDir); err != nil {

pkg/agent/syssetup/setup.go

@@ -6,8 +6,14 @@ import (
 	"io/ioutil"
 	"os"
 	"os/exec"
+	"runtime"
+	"time"
 
+	"github.com/google/cadvisor/machine"
+	"github.com/google/cadvisor/utils/sysfs"
 	"github.com/sirupsen/logrus"
+	kubeproxyconfig "k8s.io/kubernetes/pkg/proxy/apis/config"
+	"k8s.io/kubernetes/pkg/util/sysctl"
 )
 
 func loadKernelModule(moduleName string) {
@@ -16,18 +22,14 @@ func loadKernelModule(moduleName string) {
 		return
 	}
 
-	if err := exec.Command("modprobe", moduleName).Run(); err != nil {
-		logrus.Warnf("failed to start %s module", moduleName)
+	if err := exec.Command("modprobe", "--", moduleName).Run(); err != nil {
+		logrus.Warnf("Failed to load kernel module %v with modprobe", moduleName)
 	}
 }
 
-func enableSystemControl(file string) {
-	if err := ioutil.WriteFile(file, []byte("1"), 0640); err != nil {
-		logrus.Warnf("failed to write value 1 at %s: %v", file, err)
-	}
-}
-
-func Configure() {
+// Configure loads required kernel modules and sets sysctls required for other components to
+// function properly.
+func Configure(config *kubeproxyconfig.KubeProxyConntrackConfiguration) {
 	loadKernelModule("overlay")
 	loadKernelModule("nf_conntrack")
 	loadKernelModule("br_netfilter")
@@ -36,10 +38,67 @@ func Configure() {
 	// Kernel is inconsistent about how devconf is configured for
 	// new network namespaces between ipv4 and ipv6. Make sure to
 	// enable forwarding on all and default for both ipv4 and ipv8.
-	enableSystemControl("/proc/sys/net/ipv4/conf/all/forwarding")
-	enableSystemControl("/proc/sys/net/ipv4/conf/default/forwarding")
-	enableSystemControl("/proc/sys/net/ipv6/conf/all/forwarding")
-	enableSystemControl("/proc/sys/net/ipv6/conf/default/forwarding")
-	enableSystemControl("/proc/sys/net/bridge/bridge-nf-call-iptables")
-	enableSystemControl("/proc/sys/net/bridge/bridge-nf-call-ip6tables")
+	sysctls := map[string]int{
+		"net/ipv4/conf/all/forwarding":        1,
+		"net/ipv4/conf/default/forwarding":    1,
+		"net/ipv6/conf/all/forwarding":        1,
+		"net/ipv6/conf/default/forwarding":    1,
+		"net/bridge/bridge-nf-call-iptables":  1,
+		"net/bridge/bridge-nf-call-ip6tables": 1,
+	}
+
+	if conntrackMax := getConntrackMax(config); conntrackMax > 0 {
+		sysctls["net/netfilter/nf_conntrack_max"] = conntrackMax
+	}
+	if config.TCPEstablishedTimeout.Duration > 0 {
+		sysctls["net/netfilter/nf_conntrack_tcp_timeout_established"] = int(config.TCPEstablishedTimeout.Duration / time.Second)
+	}
+	if config.TCPCloseWaitTimeout.Duration > 0 {
+		sysctls["net/netfilter/nf_conntrack_tcp_timeout_close_wait"] = int(config.TCPCloseWaitTimeout.Duration / time.Second)
+	}
+
+	sys := sysctl.New()
+	for entry, value := range sysctls {
+		if val, _ := sys.GetSysctl(entry); val != value {
+			logrus.Infof("Set sysctl '%v' to %v", entry, value)
+			if err := sys.SetSysctl(entry, value); err != nil {
+				logrus.Errorf("Failed to set sysctl: %v", err)
+			}
+		}
+	}
+}
+
+// getConntrackMax is cribbed from kube-proxy, as recent kernels no longer allow non-init namespaces
+// to set conntrack-related sysctls.
+// ref: https://github.com/kubernetes/kubernetes/blob/v1.21.1/cmd/kube-proxy/app/server.go#L780
+// ref: https://github.com/kubernetes-sigs/kind/issues/2240
+func getConntrackMax(config *kubeproxyconfig.KubeProxyConntrackConfiguration) int {
+	if config.MaxPerCore != nil && *config.MaxPerCore > 0 {
+		floor := 0
+		if config.Min != nil {
+			floor = int(*config.Min)
+		}
+		scaled := int(*config.MaxPerCore) * detectNumCPU()
+		if scaled > floor {
+			logrus.Debugf("getConntrackMax: using scaled conntrack-max-per-core")
+			return scaled
+		}
+		logrus.Debugf("getConntrackMax: using conntrack-min")
+		return floor
+	}
+	return 0
+}
+
+// detectNumCPU is also cribbed from kube-proxy
+func detectNumCPU() int {
+	// try get numCPU from /sys firstly due to a known issue (https://github.com/kubernetes/kubernetes/issues/99225)
+	cpuinfo, err := ioutil.ReadFile("/proc/cpuinfo")
+	if err != nil {
+		return runtime.NumCPU()
+	}
+	_, numCPU, err := machine.GetTopology(sysfs.NewRealSysFs(), string(cpuinfo))
+	if err != nil || numCPU < 1 {
+		return runtime.NumCPU()
+	}
+	return numCPU
 }

pkg/daemons/agent/agent.go

@@ -40,6 +40,9 @@ func startKubeProxy(cfg *config.Agent) {
 		"healthz-bind-address":              "127.0.0.1",
 		"kubeconfig":                        cfg.KubeConfigKubeProxy,
 		"cluster-cidr":                      cfg.ClusterCIDR.String(),
+		"conntrack-max-per-core":            "0",
+		"conntrack-tcp-timeout-established": "0s",
+		"conntrack-tcp-timeout-close-wait":  "0s",
 	}
 	if cfg.NodeName != "" {
 		argsMap["hostname-override"] = cfg.NodeName

scripts/download

@@ -8,13 +8,19 @@ RUNC_VERSION=v1.0.0-rc94
 ROOT_VERSION=v0.8.1
 TRAEFIK_VERSION=1.81.0
 CHARTS_DIR=build/static/charts
+RUNC_DIR=build/src/github.com/opencontainers/runc
+DATA_DIR=build/data
 
+umask 022
+rm -rf ${CHARTS_DIR}
+rm -rf ${RUNC_DIR}
 mkdir -p ${CHARTS_DIR}
+mkdir -p ${DATA_DIR}
 
 curl --compressed -sfL https://github.com/rancher/k3s-root/releases/download/${ROOT_VERSION}/k3s-root-${ARCH}.tar | tar xf -
 
-git clone --depth=1 https://github.com/opencontainers/runc build/src/github.com/opencontainers/runc || true
-pushd build/src/github.com/opencontainers/runc
+git clone --depth=1 https://github.com/opencontainers/runc ${RUNC_DIR} || true
+pushd ${RUNC_DIR}
 git fetch --all --tags
 git checkout ${RUNC_VERSION} -b k3s
 popd