Merge pull request #123216 from giuseppe/runtime-handler-cri

KEP-127: require userns support from the CRI runtime before using it

pkg/kubelet/container/helpers.go

@@ -59,7 +59,7 @@ type RuntimeHelper interface {
 	GetExtraSupplementalGroupsForPod(pod *v1.Pod) []int64
 
 	// GetOrCreateUserNamespaceMappings returns the configuration for the sandbox user namespace
-	GetOrCreateUserNamespaceMappings(pod *v1.Pod) (*runtimeapi.UserNamespace, error)
+	GetOrCreateUserNamespaceMappings(pod *v1.Pod, runtimeHandler string) (*runtimeapi.UserNamespace, error)
 
 	// PrepareDynamicResources prepares resources for a pod.
 	PrepareDynamicResources(pod *v1.Pod) error

pkg/kubelet/container/runtime.go

@@ -526,6 +526,8 @@ const (
 type RuntimeStatus struct {
 	// Conditions is an array of current observed runtime conditions.
 	Conditions []RuntimeCondition
+	// Handlers is a map of current available handlers
+	Handlers map[string]RuntimeHandler
 }
 
 // GetRuntimeCondition gets a specified runtime condition from the runtime status.
@@ -542,10 +544,28 @@ func (r *RuntimeStatus) GetRuntimeCondition(t RuntimeConditionType) *RuntimeCond
 // String formats the runtime status into human readable string.
 func (r *RuntimeStatus) String() string {
 	var ss []string
+	var sh []string
 	for _, c := range r.Conditions {
 		ss = append(ss, c.String())
 	}
-	return fmt.Sprintf("Runtime Conditions: %s", strings.Join(ss, ", "))
+	for _, h := range r.Handlers {
+		sh = append(sh, h.String())
+	}
+	return fmt.Sprintf("Runtime Conditions: %s; Handlers: %s", strings.Join(ss, ", "), strings.Join(sh, ", "))
+}
+
+// RuntimeHandler contains condition information for the runtime handler.
+type RuntimeHandler struct {
+	// Name is the handler name.
+	Name string
+	// SupportsUserNamespaces is true if the handler has support for
+	// user namespaces.
+	SupportsUserNamespaces bool
+}
+
+// String formats the runtime handler into human readable string.
+func (h *RuntimeHandler) String() string {
+	return fmt.Sprintf("Name=%s SupportsUserNamespaces: %v", h.Name, h.SupportsUserNamespaces)
 }
 
 // RuntimeCondition contains condition information for the runtime.

pkg/kubelet/container/testing/fake_runtime_helper.go

@@ -68,7 +68,7 @@ func (f *FakeRuntimeHelper) GetExtraSupplementalGroupsForPod(pod *v1.Pod) []int6
 	return nil
 }
 
-func (f *FakeRuntimeHelper) GetOrCreateUserNamespaceMappings(pod *v1.Pod) (*runtimeapi.UserNamespace, error) {
+func (f *FakeRuntimeHelper) GetOrCreateUserNamespaceMappings(pod *v1.Pod, runtimeHandler string) (*runtimeapi.UserNamespace, error) {
 	return nil, nil
 }
 

pkg/kubelet/kubelet.go

@@ -1178,7 +1178,8 @@ type Kubelet struct {
 	updatePodCIDRMux sync.Mutex
 
 	// updateRuntimeMux is a lock on updating runtime, because this path is not thread-safe.
-	// This lock is used by Kubelet.updateRuntimeUp and Kubelet.fastNodeStatusUpdate functions and shouldn't be used anywhere else.
+	// This lock is used by Kubelet.updateRuntimeUp, Kubelet.fastNodeStatusUpdate and
+	// Kubelet.HandlerSupportsUserNamespaces functions and shouldn't be used anywhere else.
 	updateRuntimeMux sync.Mutex
 
 	// nodeLeaseController claims and renews the node lease for this Kubelet
@@ -2898,7 +2899,9 @@ func (kl *Kubelet) updateRuntimeUp() {
 		kl.runtimeState.setRuntimeState(fmt.Errorf("container runtime not ready: %v", runtimeReady))
 		return
 	}
+
 	kl.runtimeState.setRuntimeState(nil)
+	kl.runtimeState.setRuntimeHandlers(s.Handlers)
 	kl.oneTimeInitializer.Do(kl.initializeRuntimeDependentModules)
 	kl.runtimeState.setRuntimeSync(kl.clock.Now())
 }

pkg/kubelet/kubelet_getters.go

@@ -109,6 +109,20 @@ func (kl *Kubelet) ListPodsFromDisk() ([]types.UID, error) {
 	return kl.listPodsFromDisk()
 }
 
+// HandlerSupportsUserNamespaces checks whether the specified handler supports
+// user namespaces.
+func (kl *Kubelet) HandlerSupportsUserNamespaces(rtHandler string) (bool, error) {
+	rtHandlers := kl.runtimeState.runtimeHandlers()
+	if rtHandlers == nil {
+		return false, fmt.Errorf("runtime handlers are not set")
+	}
+	h, found := rtHandlers[rtHandler]
+	if !found {
+		return false, fmt.Errorf("the handler %q is not known", rtHandler)
+	}
+	return h.SupportsUserNamespaces, nil
+}
+
 // getPodDir returns the full path to the per-pod directory for the pod with
 // the given UID.
 func (kl *Kubelet) getPodDir(podUID types.UID) string {

pkg/kubelet/kubelet_getters_test.go

@@ -21,6 +21,7 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+	kubecontainer "k8s.io/kubernetes/pkg/kubelet/container"
 )
 
 func TestKubeletDirs(t *testing.T) {
@@ -99,3 +100,32 @@ func TestKubeletDirs(t *testing.T) {
 	exp = filepath.Join(root, "pods/abc123/volume-subpaths")
 	assert.Equal(t, exp, got)
 }
+
+func TestHandlerSupportsUserNamespaces(t *testing.T) {
+	testKubelet := newTestKubelet(t, false /* controllerAttachDetachEnabled */)
+	defer testKubelet.Cleanup()
+	kubelet := testKubelet.kubelet
+
+	kubelet.runtimeState.setRuntimeHandlers(map[string]kubecontainer.RuntimeHandler{
+		"has-support": {
+			Name:                   "has-support",
+			SupportsUserNamespaces: true,
+		},
+		"has-no-support": {
+			Name:                   "has-support",
+			SupportsUserNamespaces: false,
+		},
+	})
+
+	got, err := kubelet.HandlerSupportsUserNamespaces("has-support")
+	assert.Equal(t, true, got)
+	assert.NoError(t, err)
+
+	got, err = kubelet.HandlerSupportsUserNamespaces("has-no-support")
+	assert.Equal(t, false, got)
+	assert.NoError(t, err)
+
+	got, err = kubelet.HandlerSupportsUserNamespaces("unknown")
+	assert.Equal(t, false, got)
+	assert.Error(t, err)
+}

pkg/kubelet/kubelet_pods.go

@@ -425,8 +425,8 @@ func truncatePodHostnameIfNeeded(podName, hostname string) (string, error) {
 }
 
 // GetOrCreateUserNamespaceMappings returns the configuration for the sandbox user namespace
-func (kl *Kubelet) GetOrCreateUserNamespaceMappings(pod *v1.Pod) (*runtimeapi.UserNamespace, error) {
-	return kl.usernsManager.GetOrCreateUserNamespaceMappings(pod)
+func (kl *Kubelet) GetOrCreateUserNamespaceMappings(pod *v1.Pod, runtimeHandler string) (*runtimeapi.UserNamespace, error) {
+	return kl.usernsManager.GetOrCreateUserNamespaceMappings(pod, runtimeHandler)
 }
 
 // GeneratePodHostNameAndDomain creates a hostname and domain name for a pod,

pkg/kubelet/kuberuntime/helpers.go

@@ -200,7 +200,7 @@ func parsePodUIDFromLogsDirectory(name string) types.UID {
 }
 
 // toKubeRuntimeStatus converts the runtimeapi.RuntimeStatus to kubecontainer.RuntimeStatus.
-func toKubeRuntimeStatus(status *runtimeapi.RuntimeStatus) *kubecontainer.RuntimeStatus {
+func toKubeRuntimeStatus(status *runtimeapi.RuntimeStatus, handlers []*runtimeapi.RuntimeHandler) *kubecontainer.RuntimeStatus {
 	conditions := []kubecontainer.RuntimeCondition{}
 	for _, c := range status.GetConditions() {
 		conditions = append(conditions, kubecontainer.RuntimeCondition{
@@ -210,7 +210,18 @@ func toKubeRuntimeStatus(status *runtimeapi.RuntimeStatus) *kubecontainer.Runtim
 			Message: c.Message,
 		})
 	}
-	return &kubecontainer.RuntimeStatus{Conditions: conditions}
+	retHandlers := make(map[string]kubecontainer.RuntimeHandler)
+	for _, h := range handlers {
+		supportsUserns := false
+		if h.Features != nil {
+			supportsUserns = h.Features.UserNamespaces
+		}
+		retHandlers[h.Name] = kubecontainer.RuntimeHandler{
+			Name:                   h.Name,
+			SupportsUserNamespaces: supportsUserns,
+		}
+	}
+	return &kubecontainer.RuntimeStatus{Conditions: conditions, Handlers: retHandlers}
 }
 
 func fieldSeccompProfile(scmp *v1.SeccompProfile, profileRootPath string, fallbackToRuntimeDefault bool) (*runtimeapi.SecurityProfile, error) {

pkg/kubelet/kuberuntime/kuberuntime_manager.go

@@ -353,7 +353,7 @@ func (m *kubeGenericRuntimeManager) Status(ctx context.Context) (*kubecontainer.
 	if resp.GetStatus() == nil {
 		return nil, errors.New("runtime status is nil")
 	}
-	return toKubeRuntimeStatus(resp.GetStatus()), nil
+	return toKubeRuntimeStatus(resp.GetStatus(), resp.GetRuntimeHandlers()), nil
 }
 
 // GetPods returns a list of containers grouped by pods. The boolean parameter

pkg/kubelet/kuberuntime/util/util.go

@@ -100,7 +100,11 @@ func PidNamespaceForPod(pod *v1.Pod) runtimeapi.NamespaceMode {
 // namespacesForPod returns the runtimeapi.NamespaceOption for a given pod.
 // An empty or nil pod can be used to get the namespace defaults for v1.Pod.
 func NamespacesForPod(pod *v1.Pod, runtimeHelper kubecontainer.RuntimeHelper) (*runtimeapi.NamespaceOption, error) {
-	userNs, err := runtimeHelper.GetOrCreateUserNamespaceMappings(pod)
+	runtimeHandler := ""
+	if pod != nil && pod.Spec.RuntimeClassName != nil {
+		runtimeHandler = *pod.Spec.RuntimeClassName
+	}
+	userNs, err := runtimeHelper.GetOrCreateUserNamespaceMappings(pod, runtimeHandler)
 	if err != nil {
 		return nil, err
 	}

pkg/kubelet/runtime.go

@@ -23,6 +23,7 @@ import (
 	"time"
 
 	utilerrors "k8s.io/apimachinery/pkg/util/errors"
+	kubecontainer "k8s.io/kubernetes/pkg/kubelet/container"
 )
 
 type runtimeState struct {
@@ -34,6 +35,7 @@ type runtimeState struct {
 	storageError             error
 	cidr                     string
 	healthChecks             []*healthCheck
+	rtHandlers               map[string]kubecontainer.RuntimeHandler
 }
 
 // A health check function should be efficient and not rely on external
@@ -69,6 +71,18 @@ func (s *runtimeState) setRuntimeState(err error) {
 	s.runtimeError = err
 }
 
+func (s *runtimeState) setRuntimeHandlers(rtHandlers map[string]kubecontainer.RuntimeHandler) {
+	s.Lock()
+	defer s.Unlock()
+	s.rtHandlers = rtHandlers
+}
+
+func (s *runtimeState) runtimeHandlers() map[string]kubecontainer.RuntimeHandler {
+	s.RLock()
+	defer s.RUnlock()
+	return s.rtHandlers
+}
+
 func (s *runtimeState) setStorageState(err error) {
 	s.Lock()
 	defer s.Unlock()

pkg/kubelet/userns/userns_manager.go

@@ -49,6 +49,7 @@ const maxPods = 1024
 const mapReInitializeThreshold = 1000
 
 type userNsPodsManager interface {
+	HandlerSupportsUserNamespaces(runtimeHandler string) (bool, error)
 	GetPodDir(podUID types.UID) string
 	ListPodsFromDisk() ([]types.UID, error)
 }
@@ -379,20 +380,41 @@ func (m *UsernsManager) createUserNs(pod *v1.Pod) (userNs userNamespace, err err
 }
 
 // GetOrCreateUserNamespaceMappings returns the configuration for the sandbox user namespace
-func (m *UsernsManager) GetOrCreateUserNamespaceMappings(pod *v1.Pod) (*runtimeapi.UserNamespace, error) {
-	if !utilfeature.DefaultFeatureGate.Enabled(features.UserNamespacesSupport) {
+func (m *UsernsManager) GetOrCreateUserNamespaceMappings(pod *v1.Pod, runtimeHandler string) (*runtimeapi.UserNamespace, error) {
+	featureEnabled := utilfeature.DefaultFeatureGate.Enabled(features.UserNamespacesSupport)
+
+	if pod == nil || pod.Spec.HostUsers == nil {
+		// if the feature is enabled, specify to use the node mode...
+		if featureEnabled {
+			return &runtimeapi.UserNamespace{
+				Mode: runtimeapi.NamespaceMode_NODE,
+			}, nil
+		}
+		// ...otherwise don't even specify it
 		return nil, nil
 	}
-
-	m.lock.Lock()
-	defer m.lock.Unlock()
-
-	if pod.Spec.HostUsers == nil || *pod.Spec.HostUsers {
+	// pod.Spec.HostUsers is set to true/false
+	if !featureEnabled {
+		return nil, fmt.Errorf("the feature gate %q is disabled: can't set spec.HostUsers", features.UserNamespacesSupport)
+	}
+	if *pod.Spec.HostUsers {
 		return &runtimeapi.UserNamespace{
 			Mode: runtimeapi.NamespaceMode_NODE,
 		}, nil
 	}
 
+	// From here onwards, hostUsers=false and the feature gate is enabled.
+
+	// if the pod requested a user namespace and the runtime doesn't support user namespaces then return an error.
+	if handlerSupportsUserns, err := m.kl.HandlerSupportsUserNamespaces(runtimeHandler); err != nil {
+		return nil, err
+	} else if !handlerSupportsUserns {
+		return nil, fmt.Errorf("RuntimeClass handler %q does not support user namespaces", runtimeHandler)
+	}
+
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
 	content, err := m.readMappingsFromFile(pod.UID)
 	if err != nil && err != utilstore.ErrKeyNotFound {
 		return nil, err

pkg/kubelet/userns/userns_manager_disabled_test.go

@@ -21,6 +21,7 @@ import (
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	v1 "k8s.io/api/core/v1"
 	utilfeature "k8s.io/apiserver/pkg/util/feature"
 	featuregatetesting "k8s.io/component-base/featuregate/testing"
 	pkgfeatures "k8s.io/kubernetes/pkg/features"
@@ -49,13 +50,61 @@ func TestReleaseDisabled(t *testing.T) {
 func TestGetOrCreateUserNamespaceMappingsDisabled(t *testing.T) {
 	defer featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, pkgfeatures.UserNamespacesSupport, false)()
 
-	testUserNsPodsManager := &testUserNsPodsManager{}
-	m, err := MakeUserNsManager(testUserNsPodsManager)
-	require.NoError(t, err)
+	trueVal := true
+	falseVal := false
 
-	userns, err := m.GetOrCreateUserNamespaceMappings(nil)
-	assert.NoError(t, err)
-	assert.Nil(t, userns)
+	tests := []struct {
+		name    string
+		pod     *v1.Pod
+		success bool
+	}{
+		{
+			name:    "pod is nil",
+			pod:     nil,
+			success: true,
+		},
+		{
+			name: "hostUsers is nil",
+			pod: &v1.Pod{
+				Spec: v1.PodSpec{
+					HostUsers: nil,
+				},
+			},
+			success: true,
+		},
+		{
+			name: "hostUsers is true",
+			pod: &v1.Pod{
+				Spec: v1.PodSpec{
+					HostUsers: &trueVal,
+				},
+			},
+		},
+		{
+			name: "hostUsers is false",
+			pod: &v1.Pod{
+				Spec: v1.PodSpec{
+					HostUsers: &falseVal,
+				},
+			},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			testUserNsPodsManager := &testUserNsPodsManager{}
+			m, err := MakeUserNsManager(testUserNsPodsManager)
+			require.NoError(t, err)
+
+			userns, err := m.GetOrCreateUserNamespaceMappings(test.pod, "")
+			assert.Nil(t, userns)
+			if test.success {
+				assert.NoError(t, err)
+			} else {
+				assert.Error(t, err)
+			}
+		})
+	}
 }
 
 func TestCleanupOrphanedPodUsernsAllocationsDisabled(t *testing.T) {

pkg/kubelet/userns/userns_manager_test.go

@@ -17,6 +17,7 @@ limitations under the License.
 package userns
 
 import (
+	"errors"
 	"fmt"
 	"os"
 	"testing"
@@ -36,6 +37,7 @@ import (
 type testUserNsPodsManager struct {
 	podDir  string
 	podList []types.UID
+	userns  bool
 }
 
 func (m *testUserNsPodsManager) GetPodDir(podUID types.UID) string {
@@ -52,6 +54,13 @@ func (m *testUserNsPodsManager) ListPodsFromDisk() ([]types.UID, error) {
 	return m.podList, nil
 }
 
+func (m *testUserNsPodsManager) HandlerSupportsUserNamespaces(runtimeHandler string) (bool, error) {
+	if runtimeHandler == "error" {
+		return false, errors.New("unknown runtime")
+	}
+	return m.userns, nil
+}
+
 func TestUserNsManagerAllocate(t *testing.T) {
 	defer featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, pkgfeatures.UserNamespacesSupport, true)()
 
@@ -192,10 +201,12 @@ func TestGetOrCreateUserNamespaceMappings(t *testing.T) {
 	falseVal := false
 
 	cases := []struct {
-		name    string
-		pod     *v1.Pod
-		expMode runtimeapi.NamespaceMode
-		success bool
+		name           string
+		pod            *v1.Pod
+		expMode        runtimeapi.NamespaceMode
+		runtimeUserns  bool
+		runtimeHandler string
+		success        bool
 	}{
 		{
 			name:    "no user namespace",
@@ -203,6 +214,12 @@ func TestGetOrCreateUserNamespaceMappings(t *testing.T) {
 			expMode: runtimeapi.NamespaceMode_NODE,
 			success: true,
 		},
+		{
+			name:    "nil pod",
+			pod:     nil,
+			expMode: runtimeapi.NamespaceMode_NODE,
+			success: true,
+		},
 		{
 			name: "opt-in to host user namespace",
 			pod: &v1.Pod{
@@ -220,19 +237,42 @@ func TestGetOrCreateUserNamespaceMappings(t *testing.T) {
 					HostUsers: &falseVal,
 				},
 			},
-			expMode: runtimeapi.NamespaceMode_POD,
-			success: true,
+			expMode:       runtimeapi.NamespaceMode_POD,
+			runtimeUserns: true,
+			success:       true,
+		},
+		{
+			name: "user namespace, but no runtime support",
+			pod: &v1.Pod{
+				Spec: v1.PodSpec{
+					HostUsers: &falseVal,
+				},
+			},
+			runtimeUserns: false,
+		},
+		{
+			name: "user namespace, but runtime returns error",
+			pod: &v1.Pod{
+				Spec: v1.PodSpec{
+					HostUsers: &falseVal,
+				},
+			},
+			// This handler name makes the fake runtime return an error.
+			runtimeHandler: "error",
 		},
 	}
 
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
 			// These tests will create the userns file, so use an existing podDir.
-			testUserNsPodsManager := &testUserNsPodsManager{podDir: t.TempDir()}
+			testUserNsPodsManager := &testUserNsPodsManager{
+				podDir: t.TempDir(),
+				userns: tc.runtimeUserns,
+			}
 			m, err := MakeUserNsManager(testUserNsPodsManager)
 			assert.NoError(t, err)
 
-			userns, err := m.GetOrCreateUserNamespaceMappings(tc.pod)
+			userns, err := m.GetOrCreateUserNamespaceMappings(tc.pod, tc.runtimeHandler)
 			if (tc.success && err != nil) || (!tc.success && err == nil) {
 				t.Errorf("expected success: %v but got error: %v", tc.success, err)
 			}