kubelet: new feature gate: SupportNoneCgroupDriver

cmd/kubeadm/app/phases/kubelet/flags.go

@@ -80,6 +80,8 @@ func buildKubeletArgMap(opts kubeletFlagsOpts) map[string]string {
 		if err != nil {
 			klog.Warningf("cannot automatically assign a '--cgroup-driver' value when starting the Kubelet: %v\n", err)
 		} else {
+			// NOTE: As of Docker v19.03.0-beta5, rootless dockerd uses "cgroupfs" driver but it is substantially "none" driver.
+			// User would need to override the kubelet configuration manually.
 			kubeletFlags["cgroup-driver"] = driver
 		}
 		if opts.pauseImage != "" {

cmd/kubelet/app/options/options.go

@@ -525,7 +525,7 @@ func AddKubeletConfigFlags(mainfs *pflag.FlagSet, c *kubeletconfig.KubeletConfig
 	fs.StringVar(&c.SystemCgroups, "system-cgroups", c.SystemCgroups, "Optional absolute name of cgroups in which to place all non-kernel processes that are not already inside a cgroup under `/`. Empty for no container. Rolling back the flag requires a reboot.")
 
 	fs.BoolVar(&c.CgroupsPerQOS, "cgroups-per-qos", c.CgroupsPerQOS, "Enable creation of QoS cgroup hierarchy, if true top level QoS and pod cgroups are created.")
-	fs.StringVar(&c.CgroupDriver, "cgroup-driver", c.CgroupDriver, "Driver that the kubelet uses to manipulate cgroups on the host.  Possible values: 'cgroupfs', 'systemd'")
+	fs.StringVar(&c.CgroupDriver, "cgroup-driver", c.CgroupDriver, "Driver that the kubelet uses to manipulate cgroups on the host.  Possible values: 'cgroupfs', 'systemd', 'none'")
 	fs.StringVar(&c.CgroupRoot, "cgroup-root", c.CgroupRoot, "Optional root cgroup to use for pods. This is handled by the container runtime on a best effort basis. Default: '', which means use the container runtime default.")
 	fs.StringVar(&c.CPUManagerPolicy, "cpu-manager-policy", c.CPUManagerPolicy, "CPU Manager policy to use. Possible values: 'none', 'static'. Default: 'none'")
 	fs.DurationVar(&c.CPUManagerReconcilePeriod.Duration, "cpu-manager-reconcile-period", c.CPUManagerReconcilePeriod.Duration, "<Warning: Alpha feature> CPU Manager reconciliation period. Examples: '10s', or '1m'. If not supplied, defaults to `NodeStatusUpdateFrequency`")

cmd/kubelet/app/server.go

@@ -621,7 +621,9 @@ func run(s *options.KubeletServer, kubeDeps *kubelet.Dependencies, stopCh <-chan
 	}
 
 	var cgroupRoots []string
-
+	if s.CgroupDriver == "none" {
+		cgroupRoots = []string{"/"}
+	} else {
 		cgroupRoots = append(cgroupRoots, cm.NodeAllocatableRoot(s.CgroupRoot, s.CgroupDriver))
 		kubeletCgroup, err := cm.GetKubeletContainer(s.KubeletCgroups)
 		if err != nil {
@@ -644,6 +646,7 @@ func run(s *options.KubeletServer, kubeDeps *kubelet.Dependencies, stopCh <-chan
 			// SystemCgroups is optional, so ignore if it isn't specified
 			cgroupRoots = append(cgroupRoots, s.SystemCgroups)
 		}
+	}
 
 	if kubeDeps.CAdvisorInterface == nil {
 		imageFsInfoProvider := cadvisor.NewImageFsInfoProvider(s.ContainerRuntime, s.RemoteRuntimeEndpoint)

pkg/features/kube_features.go

@@ -479,6 +479,18 @@ const (
 	//
 	// Enable support for specifying an existing PVC as a DataSource
 	VolumePVCDataSource featuregate.Feature = "VolumePVCDataSource"
+
+	// owner: @AkihiroSuda
+	// alpha: v1.XX
+	//
+	// Enable support for "none" cgroup driver.
+	//
+	// The "none" driver is expected to be used in "rootless" mode until OCI/CRI runtime get
+	// support for cgroup2 (unified) mode with nsdelegate.
+	//
+	// Even after cgroup2 gets supported in the ecosystem, the "none" driver will remain
+	// because nested containers might not always get support for cgroup2 (via systemd).
+	SupportNoneCgroupDriver featuregate.Feature = "SupportNoneCgroupDriver"
 )
 
 func init() {
@@ -559,6 +571,7 @@ var defaultKubernetesFeatureGates = map[featuregate.Feature]featuregate.FeatureS
 	LocalStorageCapacityIsolationFSQuotaMonitoring: {Default: false, PreRelease: featuregate.Alpha},
 	NonPreemptingPriority:                          {Default: false, PreRelease: featuregate.Alpha},
 	VolumePVCDataSource:                            {Default: false, PreRelease: featuregate.Alpha},
+	SupportNoneCgroupDriver:                        {Default: false, PreRelease: featuregate.Alpha},
 
 	// inherited features from generic apiserver, relisted here to get a conflict if it is changed
 	// unintentionally on either side:

pkg/kubelet/apis/config/types.go

@@ -189,7 +189,7 @@ type KubeletConfiguration struct {
 	// And all Burstable and BestEffort pods are brought up under their
 	// specific top level QoS cgroup.
 	CgroupsPerQOS bool
-	// driver that the kubelet uses to manipulate cgroups on the host (cgroupfs or systemd)
+	// driver that the kubelet uses to manipulate cgroups on the host (cgroupfs, systemd, none)
 	CgroupDriver string
 	// CPUManagerPolicy is the name of the policy to use.
 	// Requires the CPUManager feature gate to be enabled.

pkg/kubelet/cm/cgroup_manager_linux.go

@@ -45,6 +45,9 @@ const (
 	libcontainerCgroupfs libcontainerCgroupManagerType = "cgroupfs"
 	// libcontainerSystemd means use libcontainer with systemd
 	libcontainerSystemd libcontainerCgroupManagerType = "systemd"
+	// noneDriver is the name of the "NOP" driver, which is used when
+	// cgroup is not accessible
+	noneDriver = "none"
 	// systemdSuffix is the cgroup name suffix for systemd
 	systemdSuffix string = ".slice"
 )
@@ -191,7 +194,15 @@ type cgroupManagerImpl struct {
 var _ CgroupManager = &cgroupManagerImpl{}
 
 // NewCgroupManager is a factory method that returns a CgroupManager
-func NewCgroupManager(cs *CgroupSubsystems, cgroupDriver string) CgroupManager {
+func NewCgroupManager(cs *CgroupSubsystems, cgroupDriver string) (CgroupManager, error) {
+	if cgroupDriver == noneDriver {
+		if !utilfeature.DefaultFeatureGate.Enabled(kubefeatures.SupportNoneCgroupDriver) {
+			return nil, fmt.Errorf("cgroup driver %q requires SupportNoneCgroupDriver feature gate", cgroupDriver)
+		}
+		cm := &noneCgroupManager{}
+		cm.init()
+		return cm, nil
+	}
 	managerType := libcontainerCgroupfs
 	if cgroupDriver == string(libcontainerSystemd) {
 		managerType = libcontainerSystemd
@@ -199,7 +210,7 @@ func NewCgroupManager(cs *CgroupSubsystems, cgroupDriver string) CgroupManager {
 	return &cgroupManagerImpl{
 		subsystems: cs,
 		adapter:    newLibcontainerAdapter(managerType),
-	}
+	}, nil
 }
 
 // Name converts the cgroup to the driver specific value in cgroupfs form.
@@ -591,3 +602,57 @@ func (m *cgroupManagerImpl) GetResourceStats(name CgroupName) (*ResourceStats, e
 	}
 	return toResourceStats(stats), nil
 }
+
+type noneCgroupManager struct {
+	names map[string]struct{}
+}
+
+func (m *noneCgroupManager) init() {
+	m.names = make(map[string]struct{})
+}
+
+func (m *noneCgroupManager) Create(c *CgroupConfig) error {
+	name := m.Name(c.Name)
+	m.names[name] = struct{}{}
+	return nil
+}
+
+func (m *noneCgroupManager) Destroy(c *CgroupConfig) error {
+	name := m.Name(c.Name)
+	delete(m.names, name)
+	return nil
+}
+
+func (m *noneCgroupManager) Update(c *CgroupConfig) error {
+	name := m.Name(c.Name)
+	m.names[name] = struct{}{}
+	return nil
+}
+
+func (m *noneCgroupManager) Exists(cgname CgroupName) bool {
+	name := m.Name(cgname)
+	_, ok := m.names[name]
+	return ok
+}
+
+func (m *noneCgroupManager) Name(cgname CgroupName) string {
+	return cgname.ToCgroupfs()
+}
+
+func (m *noneCgroupManager) CgroupName(name string) CgroupName {
+	return ParseCgroupfsToCgroupName(name)
+}
+
+func (m *noneCgroupManager) Pids(_ CgroupName) []int {
+	return nil
+}
+
+func (m *noneCgroupManager) ReduceCPULimits(cgroupName CgroupName) error {
+	return nil
+}
+
+func (m *noneCgroupManager) GetResourceStats(name CgroupName) (*ResourceStats, error) {
+	return &ResourceStats{
+		MemoryStats: &MemoryStats{},
+	}, nil
+}

pkg/kubelet/cm/cgroup_manager_unsupported.go

@@ -30,8 +30,8 @@ type CgroupSubsystems struct {
 	MountPoints map[string]string
 }
 
-func NewCgroupManager(_ interface{}) CgroupManager {
-	return &unsupportedCgroupManager{}
+func NewCgroupManager(_ interface{}) (CgroupManager, error) {
+	return &unsupportedCgroupManager{}, nil
 }
 
 func (m *unsupportedCgroupManager) Name(_ CgroupName) string {

pkg/kubelet/cm/container_manager_linux.go

@@ -253,9 +253,15 @@ func NewContainerManager(mountUtil mount.Interface, cadvisorInterface cadvisor.I
 
 	// Turn CgroupRoot from a string (in cgroupfs path format) to internal CgroupName
 	cgroupRoot := ParseCgroupfsToCgroupName(nodeConfig.CgroupRoot)
-	cgroupManager := NewCgroupManager(subsystems, nodeConfig.CgroupDriver)
+	cgroupManager, err := NewCgroupManager(subsystems, nodeConfig.CgroupDriver)
+	if err != nil {
+		return nil, err
+	}
 	// Check if Cgroup-root actually exists on the node
 	if nodeConfig.CgroupsPerQOS {
+		if nodeConfig.CgroupDriver == noneDriver {
+			return nil, fmt.Errorf("invalid configuration: cgroups-per-qos is not supported for %s cgroup driver", nodeConfig.CgroupDriver)
+		}
 		// this does default to / when enabled, but this tests against regressions.
 		if nodeConfig.CgroupRoot == "" {
 			return nil, fmt.Errorf("invalid configuration: cgroups-per-qos was specified and cgroup-root was not specified. To enable the QoS cgroup hierarchy you need to specify a valid cgroup-root")
@@ -265,7 +271,7 @@ func NewContainerManager(mountUtil mount.Interface, cadvisorInterface cadvisor.I
 		// of note, we always use the cgroupfs driver when performing this check since
 		// the input is provided in that format.
 		// this is important because we do not want any name conversion to occur.
-		if !cgroupManager.Exists(cgroupRoot) {
+		if !cgroupManager.Exists(cgroupRoot) && nodeConfig.CgroupDriver != noneDriver {
 			return nil, fmt.Errorf("invalid configuration: cgroup-root %q doesn't exist", cgroupRoot)
 		}
 		klog.Infof("container manager verified user specified cgroup-root exists: %v", cgroupRoot)

pkg/kubelet/cm/pod_container_manager_linux_test.go

@@ -100,8 +100,12 @@ func TestIsCgroupPod(t *testing.T) {
 		},
 	}
 	for _, cgroupDriver := range []string{"cgroupfs", "systemd"} {
+		cm, err := NewCgroupManager(nil, cgroupDriver)
+		if err != nil {
+			t.Fatal(err)
+		}
 		pcm := &podContainerManagerImpl{
-			cgroupManager:     NewCgroupManager(nil, cgroupDriver),
+			cgroupManager:     cm,
 			enforceCPULimits:  true,
 			qosContainersInfo: qosContainersInfo,
 		}

pkg/kubelet/dockershim/docker_service.go

@@ -265,7 +265,8 @@ func NewDockerService(config *ClientConfig, podSandboxImage string, streamingCon
 	} else {
 		cgroupDriver = dockerInfo.CgroupDriver
 	}
-	if len(kubeCgroupDriver) != 0 && kubeCgroupDriver != cgroupDriver {
+	// NOTE: As of Docker v19.03.0-beta5, rootless dockerd uses "cgroupfs" driver but it is substantially "none" driver.
+	if len(kubeCgroupDriver) != 0 && kubeCgroupDriver != "none" && kubeCgroupDriver != cgroupDriver {
 		return nil, fmt.Errorf("misconfiguration: kubelet cgroup driver: %q is different from docker cgroup driver: %q", kubeCgroupDriver, cgroupDriver)
 	}
 	klog.Infof("Setting cgroupDriver to %s", cgroupDriver)

test/e2e_node/node_container_manager_test.go

@@ -159,7 +159,10 @@ func runTest(f *framework.Framework) error {
 	}
 
 	// Create a cgroup manager object for manipulating cgroups.
-	cgroupManager := cm.NewCgroupManager(subsystems, oldCfg.CgroupDriver)
+	cgroupManager, err := cm.NewCgroupManager(subsystems, oldCfg.CgroupDriver)
+	if err != nil {
+		return nil
+	}
 
 	defer destroyTemporaryCgroupsForReservation(cgroupManager)
 	defer func() {