Handle conntrack-related sysctls in supervisor agent setup

Signed-off-by: Brad Davidson <brad.davidson@rancher.com>

go.mod

@@ -26,6 +26,7 @@ replace (
 	// LOOK TO scripts/download FOR THE VERSION OF runc THAT WE ARE BUILDING/SHIPPING
 	github.com/opencontainers/runc => github.com/opencontainers/runc v1.0.0-rc93.0.20210414171415-3397a09ee932
 	github.com/opencontainers/runtime-spec => github.com/opencontainers/runtime-spec v1.0.3-0.20210316141917-a8c4a9ee0f6b
+	github.com/rancher/k3s/pkg/data => ./pkg/data
 	go.etcd.io/etcd => github.com/k3s-io/etcd v0.5.0-alpha.5.0.20201208200253-50621aee4aea
 	golang.org/x/crypto => golang.org/x/crypto v0.0.0-20210220033148-5ea612d1eb83
 	golang.org/x/net => golang.org/x/net v0.0.0-20210224082022-3d97a244fca7
@@ -78,6 +79,7 @@ require (
 	github.com/go-bindata/go-bindata v3.1.2+incompatible
 	github.com/go-sql-driver/mysql v1.4.1
 	github.com/golangplus/testing v1.0.0 // indirect
+	github.com/google/cadvisor v0.39.0
 	github.com/google/tcpproxy v0.0.0-20180808230851-dfa16c61dad2
 	github.com/google/uuid v1.2.0
 	github.com/gorilla/mux v1.8.0

pkg/agent/run.go

@@ -38,7 +38,10 @@ import (
 	v1 "k8s.io/client-go/kubernetes/typed/core/v1"
 	"k8s.io/client-go/tools/clientcmd"
 	"k8s.io/controller-manager/app"
+	app2 "k8s.io/kubernetes/cmd/kube-proxy/app"
+	kubeproxyconfig "k8s.io/kubernetes/pkg/proxy/apis/config"
 	utilsnet "k8s.io/utils/net"
+	utilpointer "k8s.io/utils/pointer"
 )
 
 const (
@@ -86,7 +89,12 @@ func run(ctx context.Context, cfg cmds.Agent, proxy proxy.Proxy) error {
 		return errors.Wrap(err, "failed to validate node-ip")
 	}
 
-	syssetup.Configure(dualCluster || dualService || dualNode)
+	enableIPv6 := dualCluster || dualService || dualNode
+	conntrackConfig, err := getConntrackConfig(nodeConfig)
+	if err != nil {
+		return errors.Wrap(err, "failed to validate kube-proxy conntrack configuration")
+	}
+	syssetup.Configure(enableIPv6, conntrackConfig)
 
 	if err := setupCriCtlConfig(cfg, nodeConfig); err != nil {
 		return err
@@ -138,6 +146,49 @@ func run(ctx context.Context, cfg cmds.Agent, proxy proxy.Proxy) error {
 	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 {

pkg/agent/syssetup/setup.go

@@ -3,11 +3,16 @@
 package syssetup
 
 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 +21,14 @@ func loadKernelModule(moduleName string) {
 		return
 	}
 
-	if err := exec.Command("modprobe", moduleName).Run(); err != nil {
-		logrus.Warn("Failed to start " + moduleName + " module")
+	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 "+file+": %v", err)
-	}
-}
-
-func Configure(enableIPv6 bool) {
+// Configure loads required kernel modules and sets sysctls required for other components to
+// function properly.
+func Configure(enableIPv6 bool, config *kubeproxyconfig.KubeProxyConntrackConfiguration) {
 	loadKernelModule("overlay")
 	loadKernelModule("nf_conntrack")
 	loadKernelModule("br_netfilter")
@@ -39,12 +40,66 @@ func Configure(enableIPv6 bool) {
 	// 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 ipv6.
-	enableSystemControl("/proc/sys/net/ipv4/conf/all/forwarding")
-	enableSystemControl("/proc/sys/net/ipv4/conf/default/forwarding")
-	enableSystemControl("/proc/sys/net/bridge/bridge-nf-call-iptables")
+	sysctls := map[string]int{
+		"net/ipv4/conf/all/forwarding":       1,
+		"net/ipv4/conf/default/forwarding":   1,
+		"net/bridge/bridge-nf-call-iptables": 1,
+	}
+
 	if enableIPv6 {
-		enableSystemControl("/proc/sys/net/ipv6/conf/all/forwarding")
-		enableSystemControl("/proc/sys/net/ipv6/conf/default/forwarding")
-		enableSystemControl("/proc/sys/net/bridge/bridge-nf-call-ip6tables")
+		sysctls["net/ipv6/conf/all/forwarding"] = 1
+		sysctls["net/ipv6/conf/default/forwarding"] = 1
+		sysctls["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)
+	_, numCPU, err := machine.GetTopology(sysfs.NewRealSysFs())
+	if err != nil || numCPU < 1 {
+		return runtime.NumCPU()
+	}
+	return numCPU
+}

pkg/daemons/agent/agent.go

@@ -49,6 +49,9 @@ func startKubeProxy(cfg *config.Agent) error {
 		"healthz-bind-address":              "127.0.0.1",
 		"kubeconfig":                        cfg.KubeConfigKubeProxy,
 		"cluster-cidr":                      util.JoinIPNets(cfg.ClusterCIDRs),
+		"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

@@ -8,17 +8,20 @@ RUNC_VERSION=v1.0.0-rc94
 ROOT_VERSION=v0.8.1
 TRAEFIK_VERSION=9.18.2  # appVersion: 2.4.8
 CHARTS_DIR=build/static/charts
+RUNC_DIR=build/src/github.com/opencontainers/runc
 DATA_DIR=build/data
 export TZ=UTC
 
 umask 022
+rm -rf ${CHARTS_DIR}
+rm -rf ${RUNC_DIR}
 mkdir -p ${CHARTS_DIR}
 mkdir -p ${DATA_DIR}
 
 curl --compressed -sfL https://github.com/k3s-io/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

vendor/modules.txt

@@ -554,6 +554,7 @@ github.com/golang/snappy
 # github.com/google/btree v1.0.0
 github.com/google/btree
 # github.com/google/cadvisor v0.39.0
+## explicit
 github.com/google/cadvisor/accelerators
 github.com/google/cadvisor/cache/memory
 github.com/google/cadvisor/collector
@@ -3195,6 +3196,7 @@ sigs.k8s.io/yaml
 # github.com/matryer/moq => github.com/rancher/moq v0.0.0-20190404221404-ee5226d43009
 # github.com/opencontainers/runc => github.com/opencontainers/runc v1.0.0-rc93.0.20210414171415-3397a09ee932
 # github.com/opencontainers/runtime-spec => github.com/opencontainers/runtime-spec v1.0.3-0.20210316141917-a8c4a9ee0f6b
+# github.com/rancher/k3s/pkg/data => ./pkg/data
 # go.etcd.io/etcd => github.com/k3s-io/etcd v0.5.0-alpha.5.0.20201208200253-50621aee4aea
 # golang.org/x/crypto => golang.org/x/crypto v0.0.0-20210220033148-5ea612d1eb83
 # golang.org/x/net => golang.org/x/net v0.0.0-20210224082022-3d97a244fca7