Add systemd cgroup controller support

pkg/agent/containerd/config_linux.go

@@ -45,18 +45,22 @@ func setupContainerdConfig(ctx context.Context, cfg *config.Node) error {
 	}
 
 	isRunningInUserNS := userns.RunningInUserNS()
-	_, _, hasCFS, hasPIDs := cgroups.CheckCgroups()
+	_, _, controllers := cgroups.CheckCgroups()
 	// "/sys/fs/cgroup" is namespaced
 	cgroupfsWritable := unix.Access("/sys/fs/cgroup", unix.W_OK) == nil
-	disableCgroup := isRunningInUserNS && (!hasCFS || !hasPIDs || !cgroupfsWritable)
+	disableCgroup := isRunningInUserNS && (!controllers["cpu"] || !controllers["pids"] || !cgroupfsWritable)
 	if disableCgroup {
 		logrus.Warn("cgroup v2 controllers are not delegated for rootless. Disabling cgroup.")
 	}
 
+	systemdCgroup := controllers["cpuset"] && os.Getenv("NOTIFY_SOCKET") != ""
+	cfg.AgentConfig.Systemd = systemdCgroup
+
 	var containerdTemplate string
 	containerdConfig := templates.ContainerdConfig{
 		NodeConfig:            cfg,
 		DisableCgroup:         disableCgroup,
+		SystemdCgroup:         systemdCgroup,
 		IsRunningInUserNS:     isRunningInUserNS,
 		PrivateRegistryConfig: privRegistries.Registry,
 		ExtraRuntimes:         findNvidiaContainerRuntimes(os.DirFS(string(os.PathSeparator))),

pkg/agent/containerd/config_windows.go

@@ -45,6 +45,7 @@ func setupContainerdConfig(ctx context.Context, cfg *config.Node) error {
 	containerdConfig := templates.ContainerdConfig{
 		NodeConfig:            cfg,
 		DisableCgroup:         true,
+		SystemdCgroup:         false,
 		IsRunningInUserNS:     false,
 		PrivateRegistryConfig: privRegistries.Registry,
 	}

pkg/agent/templates/templates.go

@@ -14,6 +14,7 @@ type ContainerdRuntimeConfig struct {
 type ContainerdConfig struct {
 	NodeConfig            *config.Node
 	DisableCgroup         bool
+	SystemdCgroup         bool
 	IsRunningInUserNS     bool
 	PrivateRegistryConfig *registries.Registry
 	ExtraRuntimes         map[string]ContainerdRuntimeConfig

pkg/agent/templates/templates_linux.go

@@ -81,6 +81,9 @@ enable_keychain = true
 [plugins.cri.containerd.runtimes.runc]
   runtime_type = "io.containerd.runc.v2"
 
+[plugins.cri.containerd.runtimes.runc.options]
+	SystemdCgroup = {{ .SystemdCgroup }}
+
 {{ if .PrivateRegistryConfig }}
 {{ if .PrivateRegistryConfig.Mirrors }}
 [plugins.cri.registry.mirrors]{{end}}

pkg/cgroups/cgroups_linux.go

@@ -65,34 +65,30 @@ func validateCgroupsV2() error {
 	return nil
 }
 
-func CheckCgroups() (kubeletRoot, runtimeRoot string, hasCFS, hasPIDs bool) {
+func CheckCgroups() (kubeletRoot, runtimeRoot string, controllers map[string]bool) {
 	cgroupsModeV2 := cgroups.Mode() == cgroups.Unified
+	controllers = make(map[string]bool)
 
 	// For Unified (v2) cgroups we can directly check to see what controllers are mounted
 	// under the unified hierarchy.
 	if cgroupsModeV2 {
 		m, err := cgroupsv2.LoadManager("/sys/fs/cgroup", "/")
 		if err != nil {
-			return "", "", false, false
+			return
 		}
-		controllers, err := m.Controllers()
+		enabledControllers, err := m.Controllers()
 		if err != nil {
-			return "", "", false, false
+			return
 		}
 		// Intentionally using an expressionless switch to match the logic below
-		for _, controller := range controllers {
-			switch {
-			case controller == "cpu":
-				hasCFS = true
-			case controller == "pids":
-				hasPIDs = true
-			}
+		for _, controller := range enabledControllers {
+			controllers[controller] = true
 		}
 	}
 
 	f, err := os.Open("/proc/self/cgroup")
 	if err != nil {
-		return "", "", false, false
+		return
 	}
 	defer f.Close()
 
@@ -102,10 +98,10 @@ func CheckCgroups() (kubeletRoot, runtimeRoot string, hasCFS, hasPIDs bool) {
 		if len(parts) < 3 {
 			continue
 		}
-		controllers := strings.Split(parts[1], ",")
+		enabledControllers := strings.Split(parts[1], ",")
 		// For v1 or hybrid, controller can be a single value {"blkio"}, or a comounted set {"cpu","cpuacct"}
-		// For v2, controllers = {""} (only contains a single empty string)
-		for _, controller := range controllers {
+		// For v2, controllers = {""} (only contains a single empty string) so this section is not used.
+		for _, controller := range enabledControllers {
 			switch {
 			case controller == "name=systemd" || cgroupsModeV2:
 				// If we detect that we are running under a `.scope` unit with systemd
@@ -128,10 +124,10 @@ func CheckCgroups() (kubeletRoot, runtimeRoot string, hasCFS, hasPIDs bool) {
 				// can fail if we use the comma-separated name. Instead, we check for the controller using the symlink.
 				p := filepath.Join("/sys/fs/cgroup", controller, parts[2], "cpu.cfs_period_us")
 				if _, err := os.Stat(p); err == nil {
-					hasCFS = true
+					controllers[controller] = true
 				}
-			case controller == "pids":
-				hasPIDs = true
+			default:
+				controllers[controller] = true
 			}
 		}
 	}
@@ -146,7 +142,7 @@ func CheckCgroups() (kubeletRoot, runtimeRoot string, hasCFS, hasPIDs bool) {
 		// a host PID scenario but we don't support this.
 		g, err := os.Open("/proc/1/cgroup")
 		if err != nil {
-			return "", "", false, false
+			return
 		}
 		defer g.Close()
 		scan = bufio.NewScanner(g)
@@ -170,5 +166,5 @@ func CheckCgroups() (kubeletRoot, runtimeRoot string, hasCFS, hasPIDs bool) {
 			}
 		}
 	}
-	return kubeletRoot, runtimeRoot, hasCFS, hasPIDs
+	return
 }

pkg/cgroups/cgroups_windows.go

+//go:build windows
 // +build windows
 
 package cgroups
@@ -6,6 +7,6 @@ func Validate() error {
 	return nil
 }
 
-func CheckCgroups() (kubeletRoot, runtimeRoot string, hasCFS, hasPIDs bool) {
-	return "", "", false, false
+func CheckCgroups() (kubeletRoot, runtimeRoot string, controllers map[string]bool) {
+	return
 }

pkg/daemons/agent/agent_linux.go

@@ -18,17 +18,15 @@ import (
 	"k8s.io/kubernetes/pkg/kubeapiserver/authorizer/modes"
 )
 
-func createRootlessConfig(argsMap map[string]string, hasCFS, hasPIDs bool) {
+func createRootlessConfig(argsMap map[string]string, controllers map[string]bool) {
 	argsMap["feature-gates=KubeletInUserNamespace"] = "true"
 	// "/sys/fs/cgroup" is namespaced
 	cgroupfsWritable := unix.Access("/sys/fs/cgroup", unix.W_OK) == nil
-	if hasCFS && hasPIDs && cgroupfsWritable {
+	if controllers["cpu"] && controllers["pids"] && cgroupfsWritable {
 		logrus.Info("cgroup v2 controllers are delegated for rootless.")
-		// cgroupfs v2, delegated for rootless by systemd
-		argsMap["cgroup-driver"] = "cgroupfs"
-	} else {
-		logrus.Fatal("delegated cgroup v2 controllers are required for rootless.")
+		return
 	}
+	logrus.Fatal("delegated cgroup v2 controllers are required for rootless.")
 }
 
 func checkRuntimeEndpoint(cfg *config.Agent, argsMap map[string]string) {
@@ -67,14 +65,13 @@ func kubeletArgs(cfg *config.Agent) map[string]string {
 		bindAddress = "::1"
 	}
 	argsMap := map[string]string{
-		"healthz-bind-address":     bindAddress,
-		"read-only-port":           "0",
-		"cluster-domain":           cfg.ClusterDomain,
-		"kubeconfig":               cfg.KubeConfigKubelet,
-		"eviction-hard":            "imagefs.available<5%,nodefs.available<5%",
-		"eviction-minimum-reclaim": "imagefs.available=10%,nodefs.available=10%",
-		"fail-swap-on":             "false",
-		//"cgroup-root": "/k3s",
+		"healthz-bind-address":         bindAddress,
+		"read-only-port":               "0",
+		"cluster-domain":               cfg.ClusterDomain,
+		"kubeconfig":                   cfg.KubeConfigKubelet,
+		"eviction-hard":                "imagefs.available<5%,nodefs.available<5%",
+		"eviction-minimum-reclaim":     "imagefs.available=10%,nodefs.available=10%",
+		"fail-swap-on":                 "false",
 		"cgroup-driver":                "cgroupfs",
 		"authentication-token-webhook": "true",
 		"anonymous-auth":               "false",
@@ -138,13 +135,13 @@ func kubeletArgs(cfg *config.Agent) map[string]string {
 	if err != nil || defaultIP.String() != cfg.NodeIP {
 		argsMap["node-ip"] = cfg.NodeIP
 	}
-	kubeletRoot, runtimeRoot, hasCFS, hasPIDs := cgroups.CheckCgroups()
-	if !hasCFS {
-		logrus.Warn("Disabling CPU quotas due to missing cpu.cfs_period_us")
+	kubeletRoot, runtimeRoot, controllers := cgroups.CheckCgroups()
+	if !controllers["cpu"] {
+		logrus.Warn("Disabling CPU quotas due to missing cpu controller or cpu.cfs_period_us")
 		argsMap["cpu-cfs-quota"] = "false"
 	}
-	if !hasPIDs {
-		logrus.Fatal("PIDS cgroup support not found")
+	if !controllers["pids"] {
+		logrus.Fatal("pids cgroup controller not found")
 	}
 	if kubeletRoot != "" {
 		argsMap["kubelet-cgroups"] = kubeletRoot
@@ -172,7 +169,11 @@ func kubeletArgs(cfg *config.Agent) map[string]string {
 	}
 
 	if cfg.Rootless {
-		createRootlessConfig(argsMap, hasCFS, hasCFS)
+		createRootlessConfig(argsMap, controllers)
+	}
+
+	if cfg.Systemd {
+		argsMap["cgroup-driver"] = "systemd"
 	}
 
 	if cfg.ProtectKernelDefaults {

pkg/daemons/config/types.go

@@ -90,6 +90,7 @@ type Agent struct {
 	ExtraKubeProxyArgs      []string
 	PauseImage              string
 	Snapshotter             string
+	Systemd                 bool
 	CNIPlugin               bool
 	NodeTaints              []string
 	NodeLabels              []string