monitor.go

package dockermonitor

import (
	"context"
	"errors"
	"fmt"
	"io"
	"os"
	"runtime"
	"strconv"
	"time"

	"github.com/dchest/siphash"
	"github.com/docker/docker/api/types"
	"github.com/docker/docker/api/types/events"
	"github.com/docker/docker/api/types/filters"
	dockerClient "github.com/docker/docker/client"
	"go.aporeto.io/trireme-lib/collector"
	"go.aporeto.io/trireme-lib/common"
	tevents "go.aporeto.io/trireme-lib/common"
	"go.aporeto.io/trireme-lib/monitor/config"
	"go.aporeto.io/trireme-lib/monitor/constants"
	"go.aporeto.io/trireme-lib/monitor/extractors"
	"go.aporeto.io/trireme-lib/monitor/registerer"
	"go.aporeto.io/trireme-lib/policy"
	"go.aporeto.io/trireme-lib/utils/cgnetcls"
	"go.aporeto.io/trireme-lib/utils/portspec"
	"go.uber.org/zap"
)

// DockerMonitor implements the connection to Docker and monitoring based on docker events.
type DockerMonitor struct {
	dockerClient               dockerClient.CommonAPIClient
	socketType                 string
	socketAddress              string
	metadataExtractor          extractors.DockerMetadataExtractor
	handlers                   map[Event]func(ctx context.Context, event *events.Message) error
	eventnotifications         []chan *events.Message
	stopprocessor              []chan bool
	numberOfQueues             int
	stoplistener               chan bool
	config                     *config.ProcessorConfig
	netcls                     cgnetcls.Cgroupnetcls
	killContainerOnPolicyError bool
	syncAtStart                bool
}

// New returns a new docker monitor.
func New() *DockerMonitor {
	return &DockerMonitor{}
}

// SetupConfig provides a configuration to implmentations. Every implementation
// can have its own config type.
func (d *DockerMonitor) SetupConfig(registerer registerer.Registerer, cfg interface{}) (err error) {

	defaultConfig := DefaultConfig()

	if cfg == nil {
		cfg = defaultConfig
	}

	dockerConfig, ok := cfg.(*Config)
	if !ok {
		return fmt.Errorf("Invalid configuration specified")
	}

	// Setup defaults
	dockerConfig = SetupDefaultConfig(dockerConfig)

	d.socketType = dockerConfig.SocketType
	d.socketAddress = dockerConfig.SocketAddress
	d.metadataExtractor = dockerConfig.EventMetadataExtractor
	d.syncAtStart = dockerConfig.SyncAtStart
	d.killContainerOnPolicyError = dockerConfig.KillContainerOnPolicyError
	d.handlers = make(map[Event]func(ctx context.Context, event *events.Message) error)
	d.stoplistener = make(chan bool)
	d.netcls = cgnetcls.NewDockerCgroupNetController()
	d.numberOfQueues = runtime.NumCPU() * 8
	d.eventnotifications = make([]chan *events.Message, d.numberOfQueues)
	d.stopprocessor = make([]chan bool, d.numberOfQueues)
	for i := 0; i < d.numberOfQueues; i++ {
		d.eventnotifications[i] = make(chan *events.Message, 1000)
		d.stopprocessor[i] = make(chan bool)
	}

	// Add handlers for the events that we know how to process
	d.addHandler(EventCreate, d.handleCreateEvent)
	d.addHandler(EventStart, d.handleStartEvent)
	d.addHandler(EventDie, d.handleDieEvent)
	d.addHandler(EventDestroy, d.handleDestroyEvent)
	d.addHandler(EventPause, d.handlePauseEvent)
	d.addHandler(EventUnpause, d.handleUnpauseEvent)

	return nil
}

// SetupHandlers sets up handlers for monitors to invoke for various events such as
// processing unit events and synchronization events. This will be called before Start()
// by the consumer of the monitor
func (d *DockerMonitor) SetupHandlers(c *config.ProcessorConfig) {

	d.config = c
}

// Run will start the DockerPolicy Enforcement.
// It applies a policy to each Container already Up and Running.
// It listens to all ContainerEvents
func (d *DockerMonitor) Run(ctx context.Context) error {

	if err := d.config.IsComplete(); err != nil {
		return fmt.Errorf("docker: %s", err)
	}

	err := d.waitForDockerDaemon(ctx)
	if err != nil {
		zap.L().Error("Docker daemon is not running at startup - skipping container processing. periodic retries will be attempted",
			zap.Error(err),
			zap.Duration("retry interval", dockerRetryTimer),
		)
		return nil
	}

	return nil
}

func (d *DockerMonitor) initMonitor(ctx context.Context) error {
	if d.syncAtStart && d.config.Policy != nil {

		options := types.ContainerListOptions{All: true}
		containers, err := d.dockerClient.ContainerList(ctx, options)
		if err != nil {
			return fmt.Errorf("unable to get container list: %s", err)
		}

		// Starting the eventListener and wait to hear on channel for it to be ready.
		// Need to start before the resync process so that we don't loose any events.
		// They will be buffered. We don't want to start the listener before
		// getting the list from docker though, to avoid duplicates.
		listenerReady := make(chan struct{})
		go d.eventListener(ctx, listenerReady)
		<-listenerReady

		zap.L().Debug("Syncing all existing containers")
		// Syncing all Existing containers depending on MonitorSetting
		if err := d.resyncContainers(ctx, containers); err != nil {
			zap.L().Error("Unable to sync existing containers", zap.Error(err))
		}
	} else {
		// Starting the eventListener and wait to hear on channel for it to be ready.
		// We are not doing resync. We just start the listener.
		listenerReady := make(chan struct{})
		go d.eventListener(ctx, listenerReady)
		<-listenerReady
	}

	// Start processing the events
	go d.eventProcessors(ctx)
	return nil
}

// addHandler adds a callback handler for the given docker event.
// Interesting event names include 'start' and 'die'. For more on events see
// https://docs.docker.com/engine/reference/api/docker_remote_api/
// under the section 'Docker Events'.
func (d *DockerMonitor) addHandler(event Event, handler EventHandler) {
	d.handlers[event] = handler
}

// getHashKey returns key to loadbalance on. This ensures that all
// events from a pod/container fall onto the same queue.
func (d *DockerMonitor) getHashKey(r *events.Message) string {

	if isKubernetesContainer(r.Actor.Attributes) {
		return kubePodIdentifier(r.Actor.Attributes)
	}
	return r.ID
}

// sendRequestToQueue sends a request to a channel based on a hash function
func (d *DockerMonitor) sendRequestToQueue(r *events.Message) {

	key0 := uint64(256203161)
	key1 := uint64(982451653)

	key := d.getHashKey(r)
	h := siphash.Hash(key0, key1, []byte(key))

	d.eventnotifications[int(h%uint64(d.numberOfQueues))] <- r
}

// eventProcessor processes docker events. We are processing multiple
// queues in parallel so that we can activate containers as fast
// as possible.
func (d *DockerMonitor) eventProcessors(ctx context.Context) {

	for i := 0; i < d.numberOfQueues; i++ {
		go func(i int) {
			for {
				select {
				case event := <-d.eventnotifications[i]:
					if f, ok := d.handlers[Event(event.Action)]; ok {
						if err := f(ctx, event); err != nil {
							zap.L().Error("Unable to handle docker event",
								zap.String("action", event.Action),
								zap.Error(err),
							)
						}
						continue
					}
				case <-ctx.Done():
					return
				}
			}
		}(i)
	}
}

// eventListener listens to Docker events from the daemon and passes to
// to the processor through a buffered channel. This minimizes the chances
// that we will miss events because the processor is delayed
func (d *DockerMonitor) eventListener(ctx context.Context, listenerReady chan struct{}) {

	f := filters.NewArgs()
	f.Add("type", "container")
	options := types.EventsOptions{
		Filters: f,
	}

	messages, errs := d.dockerClient.Events(context.Background(), options)

	// Once the buffered event channel was returned by Docker we return the ready status.
	listenerReady <- struct{}{}

	for {
		select {
		case message := <-messages:
			zap.L().Debug("Got message from docker client",
				zap.String("action", message.Action),
				zap.String("ID", message.ID),
			)
			d.sendRequestToQueue(&message)

		case err := <-errs:
			if err != nil && err != io.EOF {
				zap.L().Warn("Received docker event error",
					zap.Error(err),
				)
			}
		case <-ctx.Done():
			return
		}
	}
}

// Resync resyncs all the existing containers on the Host, using the
// same process as when a container is initially spawn up
func (d *DockerMonitor) Resync(ctx context.Context) error {

	if !d.syncAtStart || d.config.Policy == nil {
		zap.L().Debug("No synchronization of containers performed")
		return nil
	}

	zap.L().Debug("Syncing all existing containers")

	options := types.ContainerListOptions{All: true}
	containers, err := d.dockerClient.ContainerList(ctx, options)
	if err != nil {
		return fmt.Errorf("unable to get container list: %s", err)
	}

	return d.resyncContainers(ctx, containers)
}

func (d *DockerMonitor) resyncContainers(ctx context.Context, containers []types.Container) error {

	// resync containers that share host network first.
	if err := d.resyncContainersByOrder(ctx, containers, true); err != nil {
		zap.L().Error("Unable to sync container", zap.Error(err))
	}

	// resync remaining containers.
	if err := d.resyncContainersByOrder(ctx, containers, false); err != nil {
		zap.L().Error("Unable to sync container", zap.Error(err))
	}

	return nil
}

//container.HostConfig.NetworkMode == constants.DockerHostMode
func (d *DockerMonitor) resyncContainersByOrder(ctx context.Context, containers []types.Container, syncHost bool) error {
	for _, c := range containers {
		container, err := d.dockerClient.ContainerInspect(ctx, c.ID)
		if err != nil {
			continue
		}

		if (syncHost && container.HostConfig.NetworkMode != constants.DockerHostMode) ||
			(!syncHost && container.HostConfig.NetworkMode == constants.DockerHostMode) {
			continue
		}

		puID, _ := puIDFromDockerID(container.ID)

		runtime, err := d.extractMetadata(&container)
		if err != nil {
			continue
		}

		event := common.EventStop
		if container.State.Running {
			if !container.State.Paused {
				event = common.EventStart
			} else {
				event = common.EventPause
			}
		}

		// If it is a host container, we need to activate it as a Linux process. We will
		// override the options that the metadata extractor provided.
		if container.HostConfig.NetworkMode == constants.DockerHostMode {
			options := hostModeOptions(&container)
			options.PolicyExtensions = runtime.Options().PolicyExtensions
			runtime.SetOptions(*options)
			runtime.SetPUType(common.LinuxProcessPU)
		}

		runtime.SetOptions(runtime.Options())

		if err := d.config.Policy.HandlePUEvent(ctx, puID, event, runtime); err != nil {
			zap.L().Error("Unable to sync existing Container",
				zap.String("dockerID", c.ID),
				zap.Error(err),
			)
		}

		// if the container has hostnet set to true or is linked
		// to container with hostnet set to true, program the cgroup.
		if isHostNetworkContainer(runtime) {
			if err = d.setupHostMode(puID, runtime, &container); err != nil {
				return fmt.Errorf("unable to setup host mode for container %s: %s", puID, err)
			}
		}

	}

	return nil
}

// setupHostMode sets up the net_cls cgroup for the host mode
func (d *DockerMonitor) setupHostMode(puID string, runtimeInfo policy.RuntimeReader, dockerInfo *types.ContainerJSON) (err error) {

	pausePUID := puID
	if dockerInfo.HostConfig.NetworkMode == constants.DockerHostMode {
		if err = d.netcls.Creategroup(puID); err != nil {
			return err
		}

		// Clean the cgroup on exit, if we have failed t activate.
		defer func() {
			if err != nil {
				if derr := d.netcls.DeleteCgroup(puID); derr != nil {
					zap.L().Warn("Failed to clean cgroup",
						zap.String("puID", puID),
						zap.Error(derr),
						zap.Error(err),
					)
				}
			}
		}()

		markval := runtimeInfo.Options().CgroupMark
		if markval == "" {
			return errors.New("mark value not found")
		}

		mark, _ := strconv.ParseUint(markval, 10, 32)
		if err := d.netcls.AssignMark(puID, mark); err != nil {
			return err
		}
	} else {
		// Add the container pid that is linked to hostnet to
		// the cgroup of the parent container.

		pausePUID = getPausePUID(policyExtensions(runtimeInfo))
	}

	return d.netcls.AddProcess(pausePUID, dockerInfo.State.Pid)
}

func (d *DockerMonitor) retrieveDockerInfo(ctx context.Context, event *events.Message) (*types.ContainerJSON, error) {

	info, err := d.dockerClient.ContainerInspect(ctx, event.ID)
	if err != nil {
		// If we see errors, we will kill the container for security reasons if DockerMonitor was configured to do so.
		if d.killContainerOnPolicyError {
			timeout := 0 * time.Second
			if err1 := d.dockerClient.ContainerStop(ctx, event.ID, &timeout); err1 != nil {
				zap.L().Warn("Unable to stop illegal container",
					zap.String("dockerID", event.ID),
					zap.Error(err1),
				)
			}
			d.config.Collector.CollectContainerEvent(&collector.ContainerRecord{
				ContextID: event.ID,
				IPAddress: nil,
				Tags:      nil,
				Event:     collector.ContainerFailed,
			})
			return nil, fmt.Errorf("unable to read container information: container %s killed: %s", event.ID, err)
		}
		return nil, fmt.Errorf("unable to read container information: container %s kept alive per policy: %s", event.ID, err)
	}
	return &info, nil
}

// ExtractMetadata generates the RuntimeInfo based on Docker primitive
func (d *DockerMonitor) extractMetadata(dockerInfo *types.ContainerJSON) (*policy.PURuntime, error) {

	if dockerInfo == nil {
		return nil, errors.New("docker info is empty")
	}

	if d.metadataExtractor != nil {
		return d.metadataExtractor(dockerInfo)
	}

	return extractors.DefaultMetadataExtractor(dockerInfo)
}

// handleCreateEvent generates a create event type. We extract the metadata
// and start the policy resolution at the create event. No need to wait
// for the start event.
func (d *DockerMonitor) handleCreateEvent(ctx context.Context, event *events.Message) error {

	puID, err := puIDFromDockerID(event.ID)
	if err != nil {
		return err
	}

	container, err := d.retrieveDockerInfo(ctx, event)
	if err != nil {
		return err
	}

	runtime, err := d.extractMetadata(container)
	if err != nil {
		return err
	}

	// If it is a host container, we need to activate it as a Linux process. We will
	// override the options that the metadata extractor provided. We will maintain
	// any policy extensions in the object.
	if container.HostConfig.NetworkMode == constants.DockerHostMode {
		options := hostModeOptions(container)
		options.PolicyExtensions = runtime.Options().PolicyExtensions
		runtime.SetOptions(*options)
		runtime.SetPUType(common.LinuxProcessPU)
	}

	runtime.SetOptions(runtime.Options())

	return d.config.Policy.HandlePUEvent(ctx, puID, tevents.EventCreate, runtime)
}

// handleStartEvent will notify the policy engine immediately about the event in order
// to start the implementation of the functions. At this point we know the process ID
// that is needed for the remote enforcers.
func (d *DockerMonitor) handleStartEvent(ctx context.Context, event *events.Message) error {

	container, err := d.retrieveDockerInfo(ctx, event)
	if err != nil {
		return err
	}

	if !container.State.Running {
		return nil
	}

	puID, err := puIDFromDockerID(container.ID)
	if err != nil {
		return err
	}

	runtime, err := d.extractMetadata(container)
	if err != nil {
		return err
	}

	// If it is a host container, we need to activate it as a Linux process. We will
	// override the options that the metadata extractor provided.
	if container.HostConfig.NetworkMode == constants.DockerHostMode {
		options := hostModeOptions(container)
		options.PolicyExtensions = runtime.Options().PolicyExtensions
		runtime.SetOptions(*options)
		runtime.SetPUType(common.LinuxProcessPU)
	}

	runtime.SetOptions(runtime.Options())

	if err = d.config.Policy.HandlePUEvent(ctx, puID, tevents.EventStart, runtime); err != nil {
		if d.killContainerOnPolicyError {
			timeout := 0 * time.Second
			if err1 := d.dockerClient.ContainerStop(ctx, event.ID, &timeout); err1 != nil {
				zap.L().Warn("Unable to stop illegal container",
					zap.String("dockerID", event.ID),
					zap.Error(err1),
				)
			}
			d.config.Collector.CollectContainerEvent(&collector.ContainerRecord{
				ContextID: event.ID,
				IPAddress: nil,
				Tags:      nil,
				Event:     collector.ContainerFailed,
			})
			return fmt.Errorf("unable to start container because of policy: container %s killed: %s", event.ID, err)
		}
		return fmt.Errorf("unable to set policy: container %s kept alive per policy: %s", puID, err)
	}

	// if the container has hostnet set to true or is linked
	// to container with hostnet set to true, program the cgroup.
	if isHostNetworkContainer(runtime) {
		if err = d.setupHostMode(puID, runtime, container); err != nil {
			return fmt.Errorf("unable to setup host mode for container %s: %s", puID, err)
		}
	}
	return nil
}

//handleDie event is called when a container dies. It generates a "Stop" event.
func (d *DockerMonitor) handleDieEvent(ctx context.Context, event *events.Message) error {

	puID, err := puIDFromDockerID(event.ID)
	if err != nil {
		return err
	}

	runtime := policy.NewPURuntimeWithDefaults()
	runtime.SetOptions(runtime.Options())

	return d.config.Policy.HandlePUEvent(ctx, puID, tevents.EventStop, runtime)
}

// handleDestroyEvent handles destroy events from Docker. It generated a "Destroy event"
func (d *DockerMonitor) handleDestroyEvent(ctx context.Context, event *events.Message) error {

	puID, err := puIDFromDockerID(event.ID)
	if err != nil {
		return err
	}
	runtime := policy.NewPURuntimeWithDefaults()
	runtime.SetOptions(runtime.Options())

	err = d.config.Policy.HandlePUEvent(ctx, puID, tevents.EventDestroy, runtime)
	if err != nil {
		zap.L().Error("Failed to handle delete event",
			zap.Error(err),
		)
	}

	if err := d.netcls.DeleteCgroup(puID); err != nil {
		zap.L().Warn("Failed to clean netcls group",
			zap.String("puID", puID),
			zap.Error(err),
		)
	}

	return nil
}

// handlePauseEvent generates a create event type.
func (d *DockerMonitor) handlePauseEvent(ctx context.Context, event *events.Message) error {
	zap.L().Info("UnPause Event for nativeID", zap.String("ID", event.ID))

	puID, err := puIDFromDockerID(event.ID)
	if err != nil {
		return err
	}

	runtime := policy.NewPURuntimeWithDefaults()
	runtime.SetOptions(runtime.Options())

	return d.config.Policy.HandlePUEvent(ctx, puID, tevents.EventPause, runtime)
}

// handleCreateEvent generates a create event type.
func (d *DockerMonitor) handleUnpauseEvent(ctx context.Context, event *events.Message) error {

	puID, err := puIDFromDockerID(event.ID)
	if err != nil {
		return err
	}

	runtime := policy.NewPURuntimeWithDefaults()
	runtime.SetOptions(runtime.Options())

	return d.config.Policy.HandlePUEvent(ctx, puID, tevents.EventUnpause, runtime)
}

func puIDFromDockerID(dockerID string) (string, error) {

	if dockerID == "" {
		return "", errors.New("unable to generate context id: empty docker id")
	}

	if len(dockerID) < 12 {
		return "", fmt.Errorf("unable to generate context id: dockerid smaller than 12 characters: %s", dockerID)
	}

	return dockerID[:12], nil
}

func initDockerClient(socketType string, socketAddress string) (*dockerClient.Client, error) {

	var socket string

	switch socketType {
	case "tcp":
		socket = "https://" + socketAddress
	case "unix":
		// Sanity check that this path exists
		if _, oserr := os.Stat(socketAddress); os.IsNotExist(oserr) {
			return nil, oserr
		}
		socket = "unix://" + socketAddress
	default:
		return nil, fmt.Errorf("bad socket type: %s", socketType)
	}

	defaultHeaders := map[string]string{"User-Agent": "engine-api-dockerClient-1.0"}

	dc, err := dockerClient.NewClient(socket, DockerClientVersion, nil, defaultHeaders)
	if err != nil {
		return nil, fmt.Errorf("unable to create docker client: %s", err)
	}

	return dc, nil
}

func (d *DockerMonitor) setupDockerDaemon() (err error) {

	if d.dockerClient == nil {
		// Initialize client
		dockerClient, err := initDockerClient(d.socketType, d.socketAddress)
		if err != nil {
			// Reset this here since the interface = nil check will fail later this is partly initialized.
			// cheaper than doing reflect and check later
			return err
		}
		d.dockerClient = dockerClient
	}

	ctx, cancel := context.WithTimeout(context.Background(), dockerPingTimeout)
	defer cancel()

	_, err = d.dockerClient.Ping(ctx)
	return err
}

// waitForDockerDaemon is a blocking call which will try to bring up docker, if not return err
// with timeout
func (d *DockerMonitor) waitForDockerDaemon(ctx context.Context) (err error) {

	done := make(chan bool)
	zap.L().Info("Trying to initialize docker monitor")
	go func(gctx context.Context) {

		for {
			errg := d.setupDockerDaemon()
			if errg == nil {
				if err := d.initMonitor(gctx); err != nil {
					zap.L().Error("Unable to init monitor", zap.Error(err))
				}
				break
			}

			select {
			case <-gctx.Done():
				return
			case <-time.After(dockerRetryTimer):
				continue
			}

		}
		done <- true
	}(ctx)

	select {
	case <-ctx.Done():
		return nil
	case <-time.After(dockerInitializationWait):
		return fmt.Errorf("Unable to connect to docker daemon")
	case <-done:
		zap.L().Info("Started Docker Monitor")
	}

	return nil
}

// hostModeOptions creates the default options for a host-mode container. The
// container must be activated as a Linux Process.
func hostModeOptions(dockerInfo *types.ContainerJSON) *policy.OptionsType {

	options := policy.OptionsType{
		CgroupName:        strconv.Itoa(dockerInfo.State.Pid),
		CgroupMark:        strconv.FormatUint(cgnetcls.MarkVal(), 10),
		ConvertedDockerPU: true,
		AutoPort:          true,
	}

	for p := range dockerInfo.Config.ExposedPorts {
		if p.Proto() == "tcp" {
			s, err := portspec.NewPortSpecFromString(p.Port(), nil)
			if err != nil {
				continue
			}

			options.Services = append(options.Services, common.Service{
				Protocol: uint8(6),
				Ports:    s,
			})
		}
	}

	return &options
}