Merge pull request #29700 from yujuhong/dshim3

pkg/kubelet/dockershim/convert.go

@@ -21,6 +21,7 @@ import (
 	"strings"
 
 	dockertypes "github.com/docker/engine-api/types"
+
 	runtimeApi "k8s.io/kubernetes/pkg/kubelet/api/v1alpha1/runtime"
 )
 
@@ -89,3 +90,25 @@ func toRuntimeAPIContainerState(state string) runtimeApi.ContainerState {
 		return runtimeApi.ContainerState_UNKNOWN
 	}
 }
+
+func toRuntimeAPISandboxState(state string) runtimeApi.PodSandBoxState {
+	// Parse the state string in dockertypes.Container. This could break when
+	// we upgrade docker.
+	switch {
+	case strings.HasPrefix(state, statusRunningPrefix):
+		return runtimeApi.PodSandBoxState_READY
+	default:
+		return runtimeApi.PodSandBoxState_NOTREADY
+	}
+}
+
+func toRuntimeAPISandbox(c *dockertypes.Container) *runtimeApi.PodSandbox {
+	state := toRuntimeAPISandboxState(c.Status)
+	return &runtimeApi.PodSandbox{
+		Id:        &c.ID,
+		Name:      &c.Names[0],
+		State:     &state,
+		CreatedAt: &c.Created, // TODO: Why do we need CreateAt timestamp for sandboxes?
+		Labels:    c.Labels,   // TODO: Need to disthinguish annotaions and labels.
+	}
+}

pkg/kubelet/dockershim/docker_container.go

@@ -35,15 +35,17 @@ func (ds *dockerService) ListContainers(filter *runtimeApi.ContainerFilter) ([]*
 	opts := dockertypes.ContainerListOptions{All: true}
 
 	opts.Filter = dockerfilters.NewArgs()
+	f := newDockerFilter(&opts.Filter)
+
 	if filter != nil {
 		if filter.Name != nil {
-			opts.Filter.Add("name", filter.GetName())
+			f.Add("name", filter.GetName())
 		}
 		if filter.Id != nil {
-			opts.Filter.Add("id", filter.GetId())
+			f.Add("id", filter.GetId())
 		}
 		if filter.State != nil {
-			opts.Filter.Add("status", toDockerContainerStatus(filter.GetState()))
+			f.Add("status", toDockerContainerStatus(filter.GetState()))
 		}
 		if filter.PodSandboxId != nil {
 			// TODO: implement this after sandbox functions are implemented.
@@ -51,9 +53,11 @@ func (ds *dockerService) ListContainers(filter *runtimeApi.ContainerFilter) ([]*
 
 		if filter.LabelSelector != nil {
 			for k, v := range filter.LabelSelector {
-				opts.Filter.Add("label", fmt.Sprintf("%s=%s", k, v))
+				f.AddLabel(k, v)
 			}
 		}
+		// Filter out sandbox containers.
+		f.AddLabel(containerTypeLabelKey, containerTypeLabelContainer)
 	}
 	containers, err := ds.client.ListContainers(opts)
 	if err != nil {
@@ -81,14 +85,9 @@ func (ds *dockerService) CreateContainer(podSandboxID string, config *runtimeApi
 	// Merge annotations and labels because docker supports only labels.
 	// TODO: add a prefix to annotations so that we can distinguish labels and
 	// annotations when reading back them from the docker container.
-	// TODO: should we apply docker-specific labels?
-	labels := config.GetLabels()
-	for k, v := range config.GetAnnotations() {
-		if _, ok := labels[k]; !ok {
-			// Only write to labels if the key doesn't exist.
-			labels[k] = v
-		}
-	}
+	labels := makeLabels(config.GetLabels(), config.GetAnnotations())
+	// Apply a the container type label.
+	labels[containerTypeLabelKey] = containerTypeLabelContainer
 
 	image := ""
 	if iSpec := config.GetImage(); iSpec != nil {
@@ -163,10 +162,7 @@ func (ds *dockerService) CreateContainer(podSandboxID string, config *runtimeApi
 		// Note: ShmSize is handled in kube_docker_client.go
 	}
 
-	// TODO: Seccomp support. Need to figure out how to pass seccomp options
-	// through the runtime API (annotations?).See dockerManager.getSecurityOpts()
-	// for the details. Always set the default seccomp profile for now.
-	hc.SecurityOpt = []string{fmt.Sprintf("%s=%s", "seccomp", defaultSeccompProfile)}
+	hc.SecurityOpt = []string{getSeccompOpts()}
 	// TODO: Add or drop capabilities.
 
 	createConfig.HostConfig = hc

pkg/kubelet/dockershim/docker_sandbox.go

@@ -19,33 +19,243 @@ package dockershim
 import (
 	"fmt"
 
+	dockertypes "github.com/docker/engine-api/types"
+	dockercontainer "github.com/docker/engine-api/types/container"
+	dockerfilters "github.com/docker/engine-api/types/filters"
+
 	runtimeApi "k8s.io/kubernetes/pkg/kubelet/api/v1alpha1/runtime"
 )
 
+const (
+	defaultSandboxImage = "gcr.io/google_containers/pause-amd64:3.0"
+
+	// Various default sandbox resources requests/limits.
+	defaultSandboxCPUshares int64 = 2
+	defaultSandboxOOMScore  int   = -999
+
+	// Termination grace period
+	defaultSandboxGracePeriod int = 10
+)
+
 // CreatePodSandbox creates a pod-level sandbox.
 // The definition of PodSandbox is at https://github.com/kubernetes/kubernetes/pull/25899
+// For docker, PodSandbox is implemented by a container holding the network
+// namespace for the pod.
+// Note: docker doesn't use LogDirectory (yet).
 func (ds *dockerService) CreatePodSandbox(config *runtimeApi.PodSandboxConfig) (string, error) {
-	return "", fmt.Errorf("Not implemented")
+	// Step 1: Pull the image for the sandbox.
+	// TODO: How should we handle pulling custom pod infra container image
+	// (with credentials)?
+	image := defaultSandboxImage
+	if err := ds.client.PullImage(image, dockertypes.AuthConfig{}, dockertypes.ImagePullOptions{}); err != nil {
+		return "", fmt.Errorf("unable to pull image for the sandbox container: %v", err)
+	}
+
+	// Step 2: Create the sandbox container.
+	createConfig := makeSandboxDockerConfig(config, image)
+	createResp, err := ds.client.CreateContainer(*createConfig)
+	if err != nil || createResp == nil {
+		return "", fmt.Errorf("failed to create a sandbox for pod %q: %v", config.GetName(), err)
+	}
+
+	// Step 3: Start the sandbox container.
+	// Assume kubelet's garbage collector would remove the sandbox later, if
+	// startContainer failed.
+	err = ds.StartContainer(createResp.ID)
+	return createResp.ID, err
 }
 
 // StopPodSandbox stops the sandbox. If there are any running containers in the
 // sandbox, they should be force terminated.
 func (ds *dockerService) StopPodSandbox(podSandboxID string) error {
-	return fmt.Errorf("Not implemented")
+	return ds.client.StopContainer(podSandboxID, defaultSandboxGracePeriod)
+	// TODO: Stop all running containers in the sandbox.
 }
 
 // DeletePodSandbox deletes the sandbox. If there are running containers in the
 // sandbox, they should be forcibly deleted.
 func (ds *dockerService) DeletePodSandbox(podSandboxID string) error {
-	return fmt.Errorf("Not implemented")
+	return ds.client.RemoveContainer(podSandboxID, dockertypes.ContainerRemoveOptions{RemoveVolumes: true})
+	// TODO: remove all containers in the sandbox.
 }
 
-// PodSandboxStatus returns the Status of the PodSandbox.
+// PodSandboxStatus returns the status of the PodSandbox.
 func (ds *dockerService) PodSandboxStatus(podSandboxID string) (*runtimeApi.PodSandboxStatus, error) {
-	return nil, fmt.Errorf("Not implemented")
+	// Inspect the container.
+	r, err := ds.client.InspectContainer(podSandboxID)
+	if err != nil {
+		return nil, err
+	}
+
+	// Parse the timstamps.
+	createdAt, _, _, err := getContainerTimestamps(r)
+	if err != nil {
+		return nil, fmt.Errorf("failed to parse timestamp for container %q: %v", podSandboxID, err)
+	}
+	ct := createdAt.Unix()
+
+	// Translate container to sandbox state.
+	state := runtimeApi.PodSandBoxState_NOTREADY
+	if r.State.Running {
+		state = runtimeApi.PodSandBoxState_READY
+	}
+
+	// TODO: We can't really get the IP address from the network plugin, which
+	// is handled by kubelet as of now. Should we amend the interface? How is
+	// this handled in the new remote runtime integration?
+	// See DockerManager.determineContainerIP() for more details.
+	// For now, just assume that there is no network plugin.
+	// Related issue: https://github.com/kubernetes/kubernetes/issues/28667
+	var IP string
+	if r.NetworkSettings != nil {
+		IP = r.NetworkSettings.IPAddress
+		// Fall back to IPv6 address if no IPv4 address is present
+		if IP == "" {
+			IP = r.NetworkSettings.GlobalIPv6Address
+		}
+	}
+	network := &runtimeApi.PodSandboxNetworkStatus{Ip: &IP}
+	netNS := getNetworkNamespace(r)
+
+	return &runtimeApi.PodSandboxStatus{
+		Id:        &r.ID,
+		Name:      &r.Name,
+		State:     &state,
+		CreatedAt: &ct,
+		// TODO: We write annotations as labels on the docker containers. All
+		// these annotations will be read back as labels. Need to fix this.
+		// Also filter out labels only relevant to this shim.
+		Labels:  r.Config.Labels,
+		Network: network,
+		Linux:   &runtimeApi.LinuxPodSandboxStatus{Namespaces: &runtimeApi.Namespace{Network: &netNS}},
+	}, nil
 }
 
 // ListPodSandbox returns a list of Sandbox.
 func (ds *dockerService) ListPodSandbox(filter *runtimeApi.PodSandboxFilter) ([]*runtimeApi.PodSandbox, error) {
-	return nil, fmt.Errorf("Not implemented")
+	// By default, list all containers whether they are running or not.
+	opts := dockertypes.ContainerListOptions{All: true}
+	filterOutReadySandboxes := false
+
+	opts.Filter = dockerfilters.NewArgs()
+	f := newDockerFilter(&opts.Filter)
+	if filter != nil {
+		if filter.Name != nil {
+			f.Add("name", filter.GetName())
+		}
+		if filter.Id != nil {
+			f.Add("id", filter.GetId())
+		}
+		if filter.State != nil {
+			if filter.GetState() == runtimeApi.PodSandBoxState_READY {
+				// Only list running containers.
+				opts.All = false
+			} else {
+				// runtimeApi.PodSandBoxState_NOTREADY can mean the
+				// container is in any of the non-running state (e.g., created,
+				// exited). We can't tell docker to filter out running
+				// containers directly, so we'll need to filter them out
+				// ourselves after getting the results.
+				filterOutReadySandboxes = true
+			}
+		}
+
+		if filter.LabelSelector != nil {
+			for k, v := range filter.LabelSelector {
+				f.AddLabel(k, v)
+			}
+		}
+		// Filter out sandbox containers.
+		f.AddLabel(containerTypeLabelKey, containerTypeLabelSandbox)
+	}
+	containers, err := ds.client.ListContainers(opts)
+	if err != nil {
+		return nil, err
+	}
+
+	// Convert docker containers to runtime api sandboxes.
+	result := []*runtimeApi.PodSandbox{}
+	for _, c := range containers {
+		s := toRuntimeAPISandbox(&c)
+		if filterOutReadySandboxes && s.GetState() == runtimeApi.PodSandBoxState_READY {
+			continue
+		}
+		result = append(result, s)
+	}
+	return result, nil
+}
+
+func makeSandboxDockerConfig(c *runtimeApi.PodSandboxConfig, image string) *dockertypes.ContainerCreateConfig {
+	// Merge annotations and labels because docker supports only labels.
+	labels := makeLabels(c.GetLabels(), c.GetAnnotations())
+	// Apply a label to distinguish sandboxes from regular containers.
+	labels[containerTypeLabelKey] = containerTypeLabelSandbox
+
+	hc := &dockercontainer.HostConfig{}
+	createConfig := &dockertypes.ContainerCreateConfig{
+		Name: c.GetName(),
+		Config: &dockercontainer.Config{
+			Hostname: c.GetHostname(),
+			// TODO: Handle environment variables.
+			Image:  image,
+			Labels: labels,
+		},
+		HostConfig: hc,
+	}
+
+	// Apply linux-specific options.
+	if lc := c.GetLinux(); lc != nil {
+		// Apply Cgroup options.
+		// TODO: Check if this works with per-pod cgroups.
+		hc.CgroupParent = lc.GetCgroupParent()
+
+		// Apply namespace options.
+		hc.NetworkMode, hc.UTSMode, hc.PidMode = "", "", ""
+		nsOpts := lc.GetNamespaceOptions()
+		if nsOpts != nil {
+			if nsOpts.GetHostNetwork() {
+				hc.NetworkMode = namespaceModeHost
+			} else {
+				// Assume kubelet uses either the cni or the kubenet plugin.
+				// TODO: support docker networking.
+				hc.NetworkMode = "none"
+			}
+			if nsOpts.GetHostIpc() {
+				hc.IpcMode = namespaceModeHost
+			}
+			if nsOpts.GetHostPid() {
+				hc.PidMode = namespaceModeHost
+			}
+		}
+	}
+	// Set port mappings.
+	exposedPorts, portBindings := makePortsAndBindings(c.GetPortMappings())
+	createConfig.Config.ExposedPorts = exposedPorts
+	hc.PortBindings = portBindings
+
+	// Set DNS options.
+	if dnsOpts := c.GetDnsOptions(); dnsOpts != nil {
+		hc.DNS = dnsOpts.GetServers()
+		hc.DNSSearch = dnsOpts.GetSearches()
+	}
+
+	// Apply resource options.
+	setSandboxResources(c.GetResources(), hc)
+
+	// Set security options.
+	hc.SecurityOpt = []string{getSeccompOpts()}
+
+	return createConfig
+}
+
+func setSandboxResources(_ *runtimeApi.PodSandboxResources, hc *dockercontainer.HostConfig) {
+	// Ignore the resource requests and limits for now and just use the docker
+	// defaults.
+	// TODO: apply resource limits based on the configuration.
+	hc.Resources = dockercontainer.Resources{
+		MemorySwap: -1, // Always disable memory swap.
+		CPUShares:  defaultSandboxCPUshares,
+		// Use docker's default cpu quota/period.
+	}
+	hc.OomScoreAdj = defaultSandboxOOMScore
 }

pkg/kubelet/dockershim/docker_sandbox_test.go

+/*
+Copyright 2016 The Kubernetes Authors.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package dockershim
+
+import (
+	"testing"
+
+	dockertypes "github.com/docker/engine-api/types"
+
+	runtimeApi "k8s.io/kubernetes/pkg/kubelet/api/v1alpha1/runtime"
+)
+
+func TestCreateSandbox(t *testing.T) {
+	ds, fakeDocker := newTestDockerSevice()
+	name := "FOO"
+	// We don't really want to test k8s name format and parsing,
+	// but FakeDockerClient parses the name internally during AssertCreated().
+	// TODO: fix this.
+	fullName := "k8s_" + name + ".abcde_foo_new_12345678_0"
+	config := &runtimeApi.PodSandboxConfig{Name: &fullName}
+	id, err := ds.CreatePodSandbox(config)
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if err := fakeDocker.AssertCreated([]string{name}); err != nil {
+		t.Errorf("%v", err)
+	}
+	if err := fakeDocker.AssertStarted([]string{id}); err != nil {
+		t.Errorf("%v", err)
+	}
+
+	// List running containers and verify that there is one (and only one)
+	// running container that we just created.
+	containers, err := fakeDocker.ListContainers(dockertypes.ContainerListOptions{All: false})
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if len(containers) != 1 {
+		t.Errorf("More than one running containers: %+v", containers)
+	}
+	if containers[0].ID != id {
+		t.Errorf("Expected id %q, got %v", id, containers[0].ID)
+	}
+}

pkg/kubelet/dockershim/docker_service.go

@@ -36,7 +36,17 @@ const (
 	// '{{.HostConfig.NetworkMode}}'.
 	namespaceModeHost = "host"
 
+	dockerNetNSFmt = "/proc/%v/ns/net"
+
 	defaultSeccompProfile = "unconfined"
+
+	// Internal docker labels used to identify whether a container is a sandbox
+	// or a regular container.
+	// TODO: This is not backward compatible with older containers. We will
+	// need to add filtering based on names.
+	containerTypeLabelKey       = "io.kubernetes.docker.type"
+	containerTypeLabelSandbox   = "podsandbox"
+	containerTypeLabelContainer = "container"
 )
 
 func NewDockerSevice(client dockertools.DockerInterface) DockerLegacyService {

pkg/kubelet/dockershim/docker_service_test.go

@@ -20,8 +20,7 @@ import (
 	"k8s.io/kubernetes/pkg/kubelet/dockertools"
 )
 
-func newFakeDockerSevice() *dockerService {
-	return &dockerService{
-		client: dockertools.NewFakeDockerClient(),
-	}
+func newTestDockerSevice() (*dockerService, *dockertools.FakeDockerClient) {
+	c := dockertools.NewFakeDockerClient()
+	return &dockerService{client: c}, c
 }

pkg/kubelet/dockershim/helpers.go

@@ -18,8 +18,15 @@ package dockershim
 
 import (
 	"fmt"
+	"strconv"
+	"strings"
 
+	dockertypes "github.com/docker/engine-api/types"
+	dockerfilters "github.com/docker/engine-api/types/filters"
 	dockerapiversion "github.com/docker/engine-api/types/versions"
+	dockernat "github.com/docker/go-connections/nat"
+	"github.com/golang/glog"
+
 	runtimeApi "k8s.io/kubernetes/pkg/kubelet/api/v1alpha1/runtime"
 )
 
@@ -50,6 +57,23 @@ func generateEnvList(envs []*runtimeApi.KeyValue) (result []string) {
 	return
 }
 
+// Merge annotations and labels because docker supports only labels.
+// TODO: Need to be able to distinguish annotations from labels; otherwise, we
+// couldn't restore the information when reading the labels back from docker.
+func makeLabels(labels, annotations map[string]string) map[string]string {
+	merged := make(map[string]string)
+	for k, v := range labels {
+		merged[k] = v
+	}
+	for k, v := range annotations {
+		if _, ok := merged[k]; !ok {
+			// Don't overwrite the key if it already exists.
+			merged[k] = v
+		}
+	}
+	return merged
+}
+
 // generateMountBindings converts the mount list to a list of strings that
 // can be understood by docker.
 // Each element in the string is in the form of:
@@ -76,3 +100,78 @@ func generateMountBindings(mounts []*runtimeApi.Mount) (result []string) {
 	}
 	return
 }
+
+func makePortsAndBindings(pm []*runtimeApi.PortMapping) (map[dockernat.Port]struct{}, map[dockernat.Port][]dockernat.PortBinding) {
+	exposedPorts := map[dockernat.Port]struct{}{}
+	portBindings := map[dockernat.Port][]dockernat.PortBinding{}
+	for _, port := range pm {
+		exteriorPort := port.GetHostPort()
+		if exteriorPort == 0 {
+			// No need to do port binding when HostPort is not specified
+			continue
+		}
+		interiorPort := port.GetContainerPort()
+		// Some of this port stuff is under-documented voodoo.
+		// See http://stackoverflow.com/questions/20428302/binding-a-port-to-a-host-interface-using-the-rest-api
+		var protocol string
+		switch strings.ToUpper(string(port.GetProtocol())) {
+		case "UDP":
+			protocol = "/udp"
+		case "TCP":
+			protocol = "/tcp"
+		default:
+			glog.Warningf("Unknown protocol %q: defaulting to TCP", port.Protocol)
+			protocol = "/tcp"
+		}
+
+		dockerPort := dockernat.Port(strconv.Itoa(int(interiorPort)) + protocol)
+		exposedPorts[dockerPort] = struct{}{}
+
+		hostBinding := dockernat.PortBinding{
+			HostPort: strconv.Itoa(int(exteriorPort)),
+			HostIP:   port.GetHostIp(),
+		}
+
+		// Allow multiple host ports bind to same docker port
+		if existedBindings, ok := portBindings[dockerPort]; ok {
+			// If a docker port already map to a host port, just append the host ports
+			portBindings[dockerPort] = append(existedBindings, hostBinding)
+		} else {
+			// Otherwise, it's fresh new port binding
+			portBindings[dockerPort] = []dockernat.PortBinding{
+				hostBinding,
+			}
+		}
+	}
+	return exposedPorts, portBindings
+}
+
+// TODO: Seccomp support. Need to figure out how to pass seccomp options
+// through the runtime API (annotations?).See dockerManager.getSecurityOpts()
+// for the details. Always set the default seccomp profile for now.
+// Also need to support syntax for different docker versions.
+func getSeccompOpts() string {
+	return fmt.Sprintf("%s=%s", "seccomp", defaultSeccompProfile)
+}
+
+func getNetworkNamespace(c *dockertypes.ContainerJSON) string {
+	return fmt.Sprintf(dockerNetNSFmt, c.State.Pid)
+}
+
+// dockerFilter wraps around dockerfilters.Args and provides methods to modify
+// the filter easily.
+type dockerFilter struct {
+	f *dockerfilters.Args
+}
+
+func newDockerFilter(args *dockerfilters.Args) *dockerFilter {
+	return &dockerFilter{f: args}
+}
+
+func (f *dockerFilter) Add(key, value string) {
+	f.Add(key, value)
+}
+
+func (f *dockerFilter) AddLabel(key, value string) {
+	f.Add("label", fmt.Sprintf("%s=%s", key, value))
+}

pkg/kubelet/dockertools/fake_docker_client.go

@@ -51,6 +51,7 @@ type FakeDockerClient struct {
 
 	// Created, Stopped and Removed all container docker ID
 	Created         []string
+	Started         []string
 	Stopped         []string
 	Removed         []string
 	VersionInfo     dockertypes.Version
@@ -229,6 +230,17 @@ func (f *FakeDockerClient) AssertCreated(created []string) error {
 	return nil
 }
 
+func (f *FakeDockerClient) AssertStarted(started []string) error {
+	f.Lock()
+	defer f.Unlock()
+	sort.StringSlice(started).Sort()
+	sort.StringSlice(f.Started).Sort()
+	if !reflect.DeepEqual(started, f.Started) {
+		return fmt.Errorf("expected %#v, got %#v", started, f.Started)
+	}
+	return nil
+}
+
 func (f *FakeDockerClient) AssertStopped(stopped []string) error {
 	f.Lock()
 	defer f.Unlock()
@@ -339,6 +351,7 @@ func (f *FakeDockerClient) StartContainer(id string) error {
 	if err := f.popError("start"); err != nil {
 		return err
 	}
+	f.Started = append(f.Started, id)
 	container, ok := f.ContainerMap[id]
 	if !ok {
 		container = convertFakeContainer(&FakeContainer{ID: id, Name: id})