moby/daemon/cluster/executor/container/controller.go

package container // import "github.com/docker/docker/daemon/cluster/executor/container"

import (
	"context"
	"fmt"
	"os"
	"strconv"
	"strings"
	"time"

	"github.com/docker/docker/api/types"
	"github.com/docker/docker/api/types/events"
	executorpkg "github.com/docker/docker/daemon/cluster/executor"
	"github.com/docker/docker/libnetwork"
	"github.com/docker/go-connections/nat"
	gogotypes "github.com/gogo/protobuf/types"
	"github.com/moby/swarmkit/v2/agent/exec"
	"github.com/moby/swarmkit/v2/api"
	"github.com/moby/swarmkit/v2/log"
	"github.com/pkg/errors"
	"golang.org/x/time/rate"
)

const defaultGossipConvergeDelay = 2 * time.Second

// waitNodeAttachmentsTimeout defines the total period of time we should wait
// for node attachments to be ready before giving up on starting a task
const waitNodeAttachmentsTimeout = 30 * time.Second

// controller implements agent.Controller against docker's API.
//
// Most operations against docker's API are done through the container name,
// which is unique to the task.
type controller struct {
	task       *api.Task
	adapter    *containerAdapter
	closed     chan struct{}
	err        error
	pulled     chan struct{} // closed after pull
	cancelPull func()        // cancels pull context if not nil
	pullErr    error         // pull error, only read after pulled closed
}

var _ exec.Controller = &controller{}

// NewController returns a docker exec runner for the provided task.
func newController(b executorpkg.Backend, i executorpkg.ImageBackend, v executorpkg.VolumeBackend, task *api.Task, node *api.NodeDescription, dependencies exec.DependencyGetter) (*controller, error) {
	adapter, err := newContainerAdapter(b, i, v, task, node, dependencies)
	if err != nil {
		return nil, err
	}

	return &controller{
		task:    task,
		adapter: adapter,
		closed:  make(chan struct{}),
	}, nil
}

func (r *controller) Task() (*api.Task, error) {
	return r.task, nil
}

// ContainerStatus returns the container-specific status for the task.
func (r *controller) ContainerStatus(ctx context.Context) (*api.ContainerStatus, error) {
	ctnr, err := r.adapter.inspect(ctx)
	if err != nil {
		if isUnknownContainer(err) {
			return nil, nil
		}
		return nil, err
	}
	return parseContainerStatus(ctnr)
}

func (r *controller) PortStatus(ctx context.Context) (*api.PortStatus, error) {
	ctnr, err := r.adapter.inspect(ctx)
	if err != nil {
		if isUnknownContainer(err) {
			return nil, nil
		}

		return nil, err
	}

	return parsePortStatus(ctnr)
}

// Update tasks a recent task update and applies it to the container.
func (r *controller) Update(ctx context.Context, t *api.Task) error {
	// TODO(stevvooe): While assignment of tasks is idempotent, we do allow
	// updates of metadata, such as labelling, as well as any other properties
	// that make sense.
	return nil
}

// Prepare creates a container and ensures the image is pulled.
//
// If the container has already be created, exec.ErrTaskPrepared is returned.
func (r *controller) Prepare(ctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	// Before we create networks, we need to make sure that the node has all of
	// the network attachments that the task needs. This will block until that
	// is the case or the context has expired.
	// NOTE(dperny): Prepare doesn't time out on its own (that is, the context
	// passed in does not expire after any period of time), which means if the
	// node attachment never arrives (for example, if the network's IP address
	// space is exhausted), then the tasks on the node will park in PREPARING
	// forever (or until the node dies). To avoid this case, we create a new
	// context with a fixed deadline, and give up. In normal operation, a node
	// update with the node IP address should come in hot on the tail of the
	// task being assigned to the node, and this should exit on the order of
	// milliseconds, but to be extra conservative we'll give it 30 seconds to
	// time out before giving up.
	waitNodeAttachmentsContext, waitCancel := context.WithTimeout(ctx, waitNodeAttachmentsTimeout)
	defer waitCancel()
	if err := r.adapter.waitNodeAttachments(waitNodeAttachmentsContext); err != nil {
		return err
	}

	// could take a while for the cluster volumes to become available. set for
	// 5 minutes, I guess?
	// TODO(dperny): do this more intelligently. return a better error.
	waitClusterVolumesCtx, wcvcancel := context.WithTimeout(ctx, 5*time.Minute)
	defer wcvcancel()
	if err := r.adapter.waitClusterVolumes(waitClusterVolumesCtx); err != nil {
		return err
	}

	// Make sure all the networks that the task needs are created.
	if err := r.adapter.createNetworks(ctx); err != nil {
		return err
	}

	// Make sure all the volumes that the task needs are created.
	if err := r.adapter.createVolumes(ctx); err != nil {
		return err
	}

	if os.Getenv("DOCKER_SERVICE_PREFER_OFFLINE_IMAGE") != "1" {
		if r.pulled == nil {
			// Fork the pull to a different context to allow pull to continue
			// on re-entrant calls to Prepare. This ensures that Prepare can be
			// idempotent and not incur the extra cost of pulling when
			// cancelled on updates.
			var pctx context.Context

			r.pulled = make(chan struct{})
			pctx, r.cancelPull = context.WithCancel(context.Background()) // TODO(stevvooe): Bind a context to the entire controller.

			go func() {
				defer close(r.pulled)
				r.pullErr = r.adapter.pullImage(pctx) // protected by closing r.pulled
			}()
		}

		select {
		case <-ctx.Done():
			return ctx.Err()
		case <-r.pulled:
			if r.pullErr != nil {
				// NOTE(stevvooe): We always try to pull the image to make sure we have
				// the most up to date version. This will return an error, but we only
				// log it. If the image truly doesn't exist, the create below will
				// error out.
				//
				// This gives us some nice behavior where we use up to date versions of
				// mutable tags, but will still run if the old image is available but a
				// registry is down.
				//
				// If you don't want this behavior, lock down your image to an
				// immutable tag or digest.
				log.G(ctx).WithError(r.pullErr).Error("pulling image failed")
			}
		}
	}
	if err := r.adapter.create(ctx); err != nil {
		if isContainerCreateNameConflict(err) {
			if _, err := r.adapter.inspect(ctx); err != nil {
				return err
			}

			// container is already created. success!
			return exec.ErrTaskPrepared
		}

		return err
	}

	return nil
}

// Start the container. An error will be returned if the container is already started.
func (r *controller) Start(ctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	ctnr, err := r.adapter.inspect(ctx)
	if err != nil {
		return err
	}

	// Detect whether the container has *ever* been started. If so, we don't
	// issue the start.
	//
	// TODO(stevvooe): This is very racy. While reading inspect, another could
	// start the process and we could end up starting it twice.
	if ctnr.State.Status != "created" {
		return exec.ErrTaskStarted
	}

	var lnErr libnetwork.ErrNoSuchNetwork
	for {
		if err := r.adapter.start(ctx); err != nil {
			if errors.As(err, &lnErr) {
				// Retry network creation again if we
				// failed because some of the networks
				// were not found.
				if err := r.adapter.createNetworks(ctx); err != nil {
					return err
				}

				continue
			}

			return errors.Wrap(err, "starting container failed")
		}

		break
	}

	// no health check
	if ctnr.Config == nil || ctnr.Config.Healthcheck == nil || len(ctnr.Config.Healthcheck.Test) == 0 || ctnr.Config.Healthcheck.Test[0] == "NONE" {
		if err := r.adapter.activateServiceBinding(); err != nil {
			log.G(ctx).WithError(err).Errorf("failed to activate service binding for container %s which has no healthcheck config", r.adapter.container.name())
			return err
		}
		return nil
	}

	// wait for container to be healthy
	eventq := r.adapter.events(ctx)

	var healthErr error
	for {
		select {
		case event := <-eventq:
			if !r.matchevent(event) {
				continue
			}

			switch event.Action {
			case events.ActionDie: // exit on terminal events
				ctnr, err := r.adapter.inspect(ctx)
				if err != nil {
					return errors.Wrap(err, "die event received")
				} else if ctnr.State.ExitCode != 0 {
					return &exitError{code: ctnr.State.ExitCode, cause: healthErr}
				}

				return nil
			case events.ActionDestroy:
				// If we get here, something has gone wrong but we want to exit
				// and report anyways.
				return ErrContainerDestroyed
			case events.ActionHealthStatusUnhealthy:
				// in this case, we stop the container and report unhealthy status
				if err := r.Shutdown(ctx); err != nil {
					return errors.Wrap(err, "unhealthy container shutdown failed")
				}
				// set health check error, and wait for container to fully exit ("die" event)
				healthErr = ErrContainerUnhealthy
			case events.ActionHealthStatusHealthy:
				if err := r.adapter.activateServiceBinding(); err != nil {
					log.G(ctx).WithError(err).Errorf("failed to activate service binding for container %s after healthy event", r.adapter.container.name())
					return err
				}
				return nil
			}
		case <-ctx.Done():
			return ctx.Err()
		case <-r.closed:
			return r.err
		}
	}
}

// Wait on the container to exit.
func (r *controller) Wait(pctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	ctx, cancel := context.WithCancel(pctx)
	defer cancel()

	healthErr := make(chan error, 1)
	go func() {
		ectx, cancel := context.WithCancel(ctx) // cancel event context on first event
		defer cancel()
		if err := r.checkHealth(ectx); err == ErrContainerUnhealthy {
			healthErr <- ErrContainerUnhealthy
			if err := r.Shutdown(ectx); err != nil {
				log.G(ectx).WithError(err).Debug("shutdown failed on unhealthy")
			}
		}
	}()

	waitC, err := r.adapter.wait(ctx)
	if err != nil {
		return err
	}

	if status := <-waitC; status.ExitCode() != 0 {
		exitErr := &exitError{
			code: status.ExitCode(),
		}

		// Set the cause if it is knowable.
		select {
		case e := <-healthErr:
			exitErr.cause = e
		default:
			if status.Err() != nil {
				exitErr.cause = status.Err()
			}
		}

		return exitErr
	}

	return nil
}

func (r *controller) hasServiceBinding() bool {
	if r.task == nil {
		return false
	}

	// service is attached to a network besides the default bridge
	for _, na := range r.task.Networks {
		if na.Network == nil ||
			na.Network.DriverState == nil ||
			na.Network.DriverState.Name == "bridge" && na.Network.Spec.Annotations.Name == "bridge" {
			continue
		}
		return true
	}

	return false
}

// Shutdown the container cleanly.
func (r *controller) Shutdown(ctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	if r.cancelPull != nil {
		r.cancelPull()
	}

	if r.hasServiceBinding() {
		// remove container from service binding
		if err := r.adapter.deactivateServiceBinding(); err != nil {
			log.G(ctx).WithError(err).Warningf("failed to deactivate service binding for container %s", r.adapter.container.name())
			// Don't return an error here, because failure to deactivate
			// the service binding is expected if the container was never
			// started.
		}

		// add a delay for gossip converge
		// TODO(dongluochen): this delay should be configurable to fit different cluster size and network delay.
		time.Sleep(defaultGossipConvergeDelay)
	}

	if err := r.adapter.shutdown(ctx); err != nil {
		if !(isUnknownContainer(err) || isStoppedContainer(err)) {
			return err
		}
	}

	// Try removing networks referenced in this task in case this
	// task is the last one referencing it
	if err := r.adapter.removeNetworks(ctx); err != nil {
		if !isUnknownContainer(err) {
			return err
		}
	}

	return nil
}

// Terminate the container, with force.
func (r *controller) Terminate(ctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	if r.cancelPull != nil {
		r.cancelPull()
	}

	if err := r.adapter.terminate(ctx); err != nil {
		if isUnknownContainer(err) {
			return nil
		}

		return err
	}

	return nil
}

// Remove the container and its resources.
func (r *controller) Remove(ctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	if r.cancelPull != nil {
		r.cancelPull()
	}

	// It may be necessary to shut down the task before removing it.
	if err := r.Shutdown(ctx); err != nil {
		if isUnknownContainer(err) {
			return nil
		}
		// This may fail if the task was already shut down.
		log.G(ctx).WithError(err).Debug("shutdown failed on removal")
	}

	if err := r.adapter.remove(ctx); err != nil {
		if isUnknownContainer(err) {
			return nil
		}

		return err
	}
	return nil
}

// waitReady waits for a container to be "ready".
// Ready means it's past the started state.
func (r *controller) waitReady(pctx context.Context) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	ctx, cancel := context.WithCancel(pctx)
	defer cancel()

	eventq := r.adapter.events(ctx)

	ctnr, err := r.adapter.inspect(ctx)
	if err != nil {
		if !isUnknownContainer(err) {
			return errors.Wrap(err, "inspect container failed")
		}
	} else {
		switch ctnr.State.Status {
		case "running", "exited", "dead":
			return nil
		}
	}

	for {
		select {
		case event := <-eventq:
			if !r.matchevent(event) {
				continue
			}

			switch event.Action {
			case "start":
				return nil
			}
		case <-ctx.Done():
			return ctx.Err()
		case <-r.closed:
			return r.err
		}
	}
}

func (r *controller) Logs(ctx context.Context, publisher exec.LogPublisher, options api.LogSubscriptionOptions) error {
	if err := r.checkClosed(); err != nil {
		return err
	}

	// if we're following, wait for this container to be ready. there is a
	// problem here: if the container will never be ready (for example, it has
	// been totally deleted) then this will wait forever. however, this doesn't
	// actually cause any UI issues, and shouldn't be a problem. the stuck wait
	// will go away when the follow (context) is canceled.
	if options.Follow {
		if err := r.waitReady(ctx); err != nil {
			return errors.Wrap(err, "container not ready for logs")
		}
	}
	// if we're not following, we're not gonna wait for the container to be
	// ready. just call logs. if the container isn't ready, the call will fail
	// and return an error. no big deal, we don't care, we only want the logs
	// we can get RIGHT NOW with no follow

	logsContext, cancel := context.WithCancel(ctx)
	msgs, err := r.adapter.logs(logsContext, options)
	defer cancel()
	if err != nil {
		return errors.Wrap(err, "failed getting container logs")
	}

	var (
		// use a rate limiter to keep things under control but also provides some
		// ability coalesce messages.
		// this will implement a "token bucket" of size 10 MB, initially full and refilled
		// at rate 10 MB tokens per second.
		limiter = rate.NewLimiter(10<<20, 10<<20) // 10 MB/s
		msgctx  = api.LogContext{
			NodeID:    r.task.NodeID,
			ServiceID: r.task.ServiceID,
			TaskID:    r.task.ID,
		}
	)

	for {
		msg, ok := <-msgs
		if !ok {
			// we're done here, no more messages
			return nil
		}

		if msg.Err != nil {
			// the deferred cancel closes the adapter's log stream
			return msg.Err
		}

		// wait here for the limiter to catch up
		if err := limiter.WaitN(ctx, len(msg.Line)); err != nil {
			return errors.Wrap(err, "failed rate limiter")
		}
		tsp, err := gogotypes.TimestampProto(msg.Timestamp)
		if err != nil {
			return errors.Wrap(err, "failed to convert timestamp")
		}
		var stream api.LogStream
		if msg.Source == "stdout" {
			stream = api.LogStreamStdout
		} else if msg.Source == "stderr" {
			stream = api.LogStreamStderr
		}

		// parse the details out of the Attrs map
		var attrs []api.LogAttr
		if len(msg.Attrs) != 0 {
			attrs = make([]api.LogAttr, 0, len(msg.Attrs))
			for _, attr := range msg.Attrs {
				attrs = append(attrs, api.LogAttr{Key: attr.Key, Value: attr.Value})
			}
		}

		if err := publisher.Publish(ctx, api.LogMessage{
			Context:   msgctx,
			Timestamp: tsp,
			Stream:    stream,
			Attrs:     attrs,
			Data:      msg.Line,
		}); err != nil {
			return errors.Wrap(err, "failed to publish log message")
		}
	}
}

// Close the runner and clean up any ephemeral resources.
func (r *controller) Close() error {
	select {
	case <-r.closed:
		return r.err
	default:
		if r.cancelPull != nil {
			r.cancelPull()
		}

		r.err = exec.ErrControllerClosed
		close(r.closed)
	}
	return nil
}

func (r *controller) matchevent(event events.Message) bool {
	if event.Type != events.ContainerEventType {
		return false
	}
	// we can't filter using id since it will have huge chances to introduce a deadlock. see #33377.
	return event.Actor.Attributes["name"] == r.adapter.container.name()
}

func (r *controller) checkClosed() error {
	select {
	case <-r.closed:
		return r.err
	default:
		return nil
	}
}

func parseContainerStatus(ctnr types.ContainerJSON) (*api.ContainerStatus, error) {
	status := &api.ContainerStatus{
		ContainerID: ctnr.ID,
		PID:         int32(ctnr.State.Pid),
		ExitCode:    int32(ctnr.State.ExitCode),
	}

	return status, nil
}

func parsePortStatus(ctnr types.ContainerJSON) (*api.PortStatus, error) {
	status := &api.PortStatus{}

	if ctnr.NetworkSettings != nil && len(ctnr.NetworkSettings.Ports) > 0 {
		exposedPorts, err := parsePortMap(ctnr.NetworkSettings.Ports)
		if err != nil {
			return nil, err
		}
		status.Ports = exposedPorts
	}

	return status, nil
}

func parsePortMap(portMap nat.PortMap) ([]*api.PortConfig, error) {
	exposedPorts := make([]*api.PortConfig, 0, len(portMap))

	for portProtocol, mapping := range portMap {
		p, proto, ok := strings.Cut(string(portProtocol), "/")
		if !ok {
			return nil, fmt.Errorf("invalid port mapping: %s", portProtocol)
		}

		port, err := strconv.ParseUint(p, 10, 16)
		if err != nil {
			return nil, err
		}

		var protocol api.PortConfig_Protocol
		switch strings.ToLower(proto) {
		case "tcp":
			protocol = api.ProtocolTCP
		case "udp":
			protocol = api.ProtocolUDP
		case "sctp":
			protocol = api.ProtocolSCTP
		default:
			return nil, fmt.Errorf("invalid protocol: %s", proto)
		}

		for _, binding := range mapping {
			hostPort, err := strconv.ParseUint(binding.HostPort, 10, 16)
			if err != nil {
				return nil, err
			}

			// TODO(aluzzardi): We're losing the port `name` here since
			// there's no way to retrieve it back from the Engine.
			exposedPorts = append(exposedPorts, &api.PortConfig{
				PublishMode:   api.PublishModeHost,
				Protocol:      protocol,
				TargetPort:    uint32(port),
				PublishedPort: uint32(hostPort),
			})
		}
	}

	return exposedPorts, nil
}

type exitError struct {
	code  int
	cause error
}

func (e *exitError) Error() string {
	if e.cause != nil {
		return fmt.Sprintf("task: non-zero exit (%v): %v", e.code, e.cause)
	}

	return fmt.Sprintf("task: non-zero exit (%v)", e.code)
}

func (e *exitError) ExitCode() int {
	return e.code
}

func (e *exitError) Cause() error {
	return e.cause
}

// checkHealth blocks until unhealthy container is detected or ctx exits
func (r *controller) checkHealth(ctx context.Context) error {
	eventq := r.adapter.events(ctx)

	for {
		select {
		case <-ctx.Done():
			return nil
		case <-r.closed:
			return nil
		case event := <-eventq:
			if !r.matchevent(event) {
				continue
			}

			switch event.Action {
			case events.ActionHealthStatusUnhealthy:
				return ErrContainerUnhealthy
			}
		}
	}
}