moby/libcontainerd/remote/client.go
Cory Snider 4bafaa00aa Refactor libcontainerd to minimize c8d RPCs
The containerd client is very chatty at the best of times. Because the
libcontained API is stateless and references containers and processes by
string ID for every method call, the implementation is essentially
forced to use the containerd client in a way which amplifies the number
of redundant RPCs invoked to perform any operation. The libcontainerd
remote implementation has to reload the containerd container, task
and/or process metadata for nearly every operation. This in turn
amplifies the number of context switches between dockerd and containerd
to perform any container operation or handle a containerd event,
increasing the load on the system which could otherwise be allocated to
workloads.

Overhaul the libcontainerd interface to reduce the impedance mismatch
with the containerd client so that the containerd client can be used
more efficiently. Split the API out into container, task and process
interfaces which the consumer is expected to retain so that
libcontainerd can retain state---especially the analogous containerd
client objects---without having to manage any state-store inside the
libcontainerd client.

Signed-off-by: Cory Snider <csnider@mirantis.com>
2022-08-24 14:59:08 -04:00

744 lines
21 KiB
Go

package remote // import "github.com/docker/docker/libcontainerd/remote"
import (
"context"
"encoding/json"
"io"
"os"
"path/filepath"
"reflect"
"runtime"
"strings"
"sync"
"syscall"
"time"
"github.com/containerd/containerd"
apievents "github.com/containerd/containerd/api/events"
"github.com/containerd/containerd/api/types"
"github.com/containerd/containerd/archive"
"github.com/containerd/containerd/cio"
"github.com/containerd/containerd/content"
containerderrors "github.com/containerd/containerd/errdefs"
"github.com/containerd/containerd/events"
"github.com/containerd/containerd/images"
v2runcoptions "github.com/containerd/containerd/runtime/v2/runc/options"
"github.com/containerd/typeurl"
"github.com/docker/docker/errdefs"
"github.com/docker/docker/libcontainerd/queue"
libcontainerdtypes "github.com/docker/docker/libcontainerd/types"
"github.com/docker/docker/pkg/ioutils"
v1 "github.com/opencontainers/image-spec/specs-go/v1"
specs "github.com/opencontainers/runtime-spec/specs-go"
"github.com/pkg/errors"
"github.com/sirupsen/logrus"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
)
// DockerContainerBundlePath is the label key pointing to the container's bundle path
const DockerContainerBundlePath = "com.docker/engine.bundle.path"
type client struct {
client *containerd.Client
stateDir string
logger *logrus.Entry
ns string
backend libcontainerdtypes.Backend
eventQ queue.Queue
}
type container struct {
client *client
c8dCtr containerd.Container
v2runcoptions *v2runcoptions.Options
}
type task struct {
containerd.Task
ctr *container
}
type process struct {
containerd.Process
}
// NewClient creates a new libcontainerd client from a containerd client
func NewClient(ctx context.Context, cli *containerd.Client, stateDir, ns string, b libcontainerdtypes.Backend) (libcontainerdtypes.Client, error) {
c := &client{
client: cli,
stateDir: stateDir,
logger: logrus.WithField("module", "libcontainerd").WithField("namespace", ns),
ns: ns,
backend: b,
}
go c.processEventStream(ctx, ns)
return c, nil
}
func (c *client) Version(ctx context.Context) (containerd.Version, error) {
return c.client.Version(ctx)
}
func (c *container) newTask(t containerd.Task) *task {
return &task{Task: t, ctr: c}
}
func (c *container) AttachTask(ctx context.Context, attachStdio libcontainerdtypes.StdioCallback) (_ libcontainerdtypes.Task, err error) {
var dio *cio.DirectIO
defer func() {
if err != nil && dio != nil {
dio.Cancel()
dio.Close()
}
}()
attachIO := func(fifos *cio.FIFOSet) (cio.IO, error) {
// dio must be assigned to the previously defined dio for the defer above
// to handle cleanup
dio, err = c.client.newDirectIO(ctx, fifos)
if err != nil {
return nil, err
}
return attachStdio(dio)
}
t, err := c.c8dCtr.Task(ctx, attachIO)
if err != nil {
return nil, errors.Wrap(wrapError(err), "error getting containerd task for container")
}
return c.newTask(t), nil
}
func (c *client) NewContainer(ctx context.Context, id string, ociSpec *specs.Spec, shim string, runtimeOptions interface{}, opts ...containerd.NewContainerOpts) (libcontainerdtypes.Container, error) {
bdir := c.bundleDir(id)
c.logger.WithField("bundle", bdir).WithField("root", ociSpec.Root.Path).Debug("bundle dir created")
newOpts := []containerd.NewContainerOpts{
containerd.WithSpec(ociSpec),
containerd.WithRuntime(shim, runtimeOptions),
WithBundle(bdir, ociSpec),
}
opts = append(opts, newOpts...)
ctr, err := c.client.NewContainer(ctx, id, opts...)
if err != nil {
if containerderrors.IsAlreadyExists(err) {
return nil, errors.WithStack(errdefs.Conflict(errors.New("id already in use")))
}
return nil, wrapError(err)
}
created := container{
client: c,
c8dCtr: ctr,
}
if x, ok := runtimeOptions.(*v2runcoptions.Options); ok {
created.v2runcoptions = x
}
return &created, nil
}
// Start create and start a task for the specified containerd id
func (c *container) Start(ctx context.Context, checkpointDir string, withStdin bool, attachStdio libcontainerdtypes.StdioCallback) (libcontainerdtypes.Task, error) {
var (
cp *types.Descriptor
t containerd.Task
rio cio.IO
stdinCloseSync = make(chan containerd.Process, 1)
)
if checkpointDir != "" {
// write checkpoint to the content store
tar := archive.Diff(ctx, "", checkpointDir)
cp, err := c.client.writeContent(ctx, images.MediaTypeContainerd1Checkpoint, checkpointDir, tar)
// remove the checkpoint when we're done
defer func() {
if cp != nil {
err := c.client.client.ContentStore().Delete(ctx, cp.Digest)
if err != nil {
c.client.logger.WithError(err).WithFields(logrus.Fields{
"ref": checkpointDir,
"digest": cp.Digest,
}).Warnf("failed to delete temporary checkpoint entry")
}
}
}()
if err := tar.Close(); err != nil {
return nil, errors.Wrap(err, "failed to close checkpoint tar stream")
}
if err != nil {
return nil, errors.Wrapf(err, "failed to upload checkpoint to containerd")
}
}
// Optimization: assume the relevant metadata has not changed in the
// moment since the container was created. Elide redundant RPC requests
// to refresh the metadata separately for spec and labels.
md, err := c.c8dCtr.Info(ctx, containerd.WithoutRefreshedMetadata)
if err != nil {
return nil, errors.Wrap(err, "failed to retrieve metadata")
}
bundle := md.Labels[DockerContainerBundlePath]
var spec specs.Spec
if err := json.Unmarshal(md.Spec.GetValue(), &spec); err != nil {
return nil, errors.Wrap(err, "failed to retrieve spec")
}
uid, gid := getSpecUser(&spec)
taskOpts := []containerd.NewTaskOpts{
func(_ context.Context, _ *containerd.Client, info *containerd.TaskInfo) error {
info.Checkpoint = cp
return nil
},
}
if runtime.GOOS != "windows" {
taskOpts = append(taskOpts, func(_ context.Context, _ *containerd.Client, info *containerd.TaskInfo) error {
if c.v2runcoptions != nil {
opts := *c.v2runcoptions
opts.IoUid = uint32(uid)
opts.IoGid = uint32(gid)
info.Options = &opts
}
return nil
})
} else {
taskOpts = append(taskOpts, withLogLevel(c.client.logger.Level))
}
t, err = c.c8dCtr.NewTask(ctx,
func(id string) (cio.IO, error) {
fifos := newFIFOSet(bundle, libcontainerdtypes.InitProcessName, withStdin, spec.Process.Terminal)
rio, err = c.createIO(fifos, libcontainerdtypes.InitProcessName, stdinCloseSync, attachStdio)
return rio, err
},
taskOpts...,
)
if err != nil {
close(stdinCloseSync)
if rio != nil {
rio.Cancel()
rio.Close()
}
return nil, errors.Wrap(wrapError(err), "failed to create task for container")
}
// Signal c.createIO that it can call CloseIO
stdinCloseSync <- t
if err := t.Start(ctx); err != nil {
if _, err := t.Delete(ctx); err != nil {
c.client.logger.WithError(err).WithField("container", c.c8dCtr.ID()).
Error("failed to delete task after fail start")
}
return nil, wrapError(err)
}
return c.newTask(t), nil
}
// Exec creates exec process.
//
// The containerd client calls Exec to register the exec config in the shim side.
// When the client calls Start, the shim will create stdin fifo if needs. But
// for the container main process, the stdin fifo will be created in Create not
// the Start call. stdinCloseSync channel should be closed after Start exec
// process.
func (t *task) Exec(ctx context.Context, processID string, spec *specs.Process, withStdin bool, attachStdio libcontainerdtypes.StdioCallback) (libcontainerdtypes.Process, error) {
var (
p containerd.Process
rio cio.IO
stdinCloseSync = make(chan containerd.Process, 1)
)
// Optimization: assume the DockerContainerBundlePath label has not been
// updated since the container metadata was last loaded/refreshed.
md, err := t.ctr.c8dCtr.Info(ctx, containerd.WithoutRefreshedMetadata)
if err != nil {
return nil, wrapError(err)
}
fifos := newFIFOSet(md.Labels[DockerContainerBundlePath], processID, withStdin, spec.Terminal)
defer func() {
if err != nil {
if rio != nil {
rio.Cancel()
rio.Close()
}
}
}()
p, err = t.Task.Exec(ctx, processID, spec, func(id string) (cio.IO, error) {
rio, err = t.ctr.createIO(fifos, processID, stdinCloseSync, attachStdio)
return rio, err
})
if err != nil {
close(stdinCloseSync)
if containerderrors.IsAlreadyExists(err) {
return nil, errors.WithStack(errdefs.Conflict(errors.New("id already in use")))
}
return nil, wrapError(err)
}
// Signal c.createIO that it can call CloseIO
//
// the stdin of exec process will be created after p.Start in containerd
defer func() { stdinCloseSync <- p }()
if err = p.Start(ctx); err != nil {
// use new context for cleanup because old one may be cancelled by user, but leave a timeout to make sure
// we are not waiting forever if containerd is unresponsive or to work around fifo cancelling issues in
// older containerd-shim
ctx, cancel := context.WithTimeout(context.Background(), 45*time.Second)
defer cancel()
p.Delete(ctx)
return nil, wrapError(err)
}
return process{p}, nil
}
func (t *task) Kill(ctx context.Context, signal syscall.Signal) error {
return wrapError(t.Task.Kill(ctx, signal))
}
func (p process) Kill(ctx context.Context, signal syscall.Signal) error {
return wrapError(p.Process.Kill(ctx, signal))
}
func (t *task) Pause(ctx context.Context) error {
return wrapError(t.Task.Pause(ctx))
}
func (t *task) Resume(ctx context.Context) error {
return wrapError(t.Task.Resume(ctx))
}
func (t *task) Stats(ctx context.Context) (*libcontainerdtypes.Stats, error) {
m, err := t.Metrics(ctx)
if err != nil {
return nil, err
}
v, err := typeurl.UnmarshalAny(m.Data)
if err != nil {
return nil, err
}
return libcontainerdtypes.InterfaceToStats(m.Timestamp, v), nil
}
func (t *task) Summary(ctx context.Context) ([]libcontainerdtypes.Summary, error) {
pis, err := t.Pids(ctx)
if err != nil {
return nil, err
}
var infos []libcontainerdtypes.Summary
for _, pi := range pis {
i, err := typeurl.UnmarshalAny(pi.Info)
if err != nil {
return nil, errors.Wrap(err, "unable to decode process details")
}
s, err := summaryFromInterface(i)
if err != nil {
return nil, err
}
infos = append(infos, *s)
}
return infos, nil
}
func (t *task) Delete(ctx context.Context) (*containerd.ExitStatus, error) {
s, err := t.Task.Delete(ctx)
return s, wrapError(err)
}
func (p process) Delete(ctx context.Context) (*containerd.ExitStatus, error) {
s, err := p.Process.Delete(ctx)
return s, wrapError(err)
}
func (c *container) Delete(ctx context.Context) error {
// Optimization: assume the DockerContainerBundlePath label has not been
// updated since the container metadata was last loaded/refreshed.
md, err := c.c8dCtr.Info(ctx, containerd.WithoutRefreshedMetadata)
if err != nil {
return err
}
bundle := md.Labels[DockerContainerBundlePath]
if err := c.c8dCtr.Delete(ctx); err != nil {
return wrapError(err)
}
if os.Getenv("LIBCONTAINERD_NOCLEAN") != "1" {
if err := os.RemoveAll(bundle); err != nil {
c.client.logger.WithContext(ctx).WithError(err).WithFields(logrus.Fields{
"container": c.c8dCtr.ID(),
"bundle": bundle,
}).Error("failed to remove state dir")
}
}
return nil
}
func (t *task) ForceDelete(ctx context.Context) error {
_, err := t.Task.Delete(ctx, containerd.WithProcessKill)
return wrapError(err)
}
func (t *task) Status(ctx context.Context) (containerd.Status, error) {
s, err := t.Task.Status(ctx)
return s, wrapError(err)
}
func (p process) Status(ctx context.Context) (containerd.Status, error) {
s, err := p.Process.Status(ctx)
return s, wrapError(err)
}
func (c *container) getCheckpointOptions(exit bool) containerd.CheckpointTaskOpts {
return func(r *containerd.CheckpointTaskInfo) error {
if r.Options == nil && c.v2runcoptions != nil {
r.Options = &v2runcoptions.CheckpointOptions{}
}
switch opts := r.Options.(type) {
case *v2runcoptions.CheckpointOptions:
opts.Exit = exit
}
return nil
}
}
func (t *task) CreateCheckpoint(ctx context.Context, checkpointDir string, exit bool) error {
img, err := t.Task.Checkpoint(ctx, t.ctr.getCheckpointOptions(exit))
if err != nil {
return wrapError(err)
}
// Whatever happens, delete the checkpoint from containerd
defer func() {
err := t.ctr.client.client.ImageService().Delete(ctx, img.Name())
if err != nil {
t.ctr.client.logger.WithError(err).WithField("digest", img.Target().Digest).
Warnf("failed to delete checkpoint image")
}
}()
b, err := content.ReadBlob(ctx, t.ctr.client.client.ContentStore(), img.Target())
if err != nil {
return errdefs.System(errors.Wrapf(err, "failed to retrieve checkpoint data"))
}
var index v1.Index
if err := json.Unmarshal(b, &index); err != nil {
return errdefs.System(errors.Wrapf(err, "failed to decode checkpoint data"))
}
var cpDesc *v1.Descriptor
for _, m := range index.Manifests {
m := m
if m.MediaType == images.MediaTypeContainerd1Checkpoint {
cpDesc = &m //nolint:gosec
break
}
}
if cpDesc == nil {
return errdefs.System(errors.Wrapf(err, "invalid checkpoint"))
}
rat, err := t.ctr.client.client.ContentStore().ReaderAt(ctx, *cpDesc)
if err != nil {
return errdefs.System(errors.Wrapf(err, "failed to get checkpoint reader"))
}
defer rat.Close()
_, err = archive.Apply(ctx, checkpointDir, content.NewReader(rat))
if err != nil {
return errdefs.System(errors.Wrapf(err, "failed to read checkpoint reader"))
}
return err
}
// LoadContainer loads the containerd container.
func (c *client) LoadContainer(ctx context.Context, id string) (libcontainerdtypes.Container, error) {
ctr, err := c.client.LoadContainer(ctx, id)
if err != nil {
if containerderrors.IsNotFound(err) {
return nil, errors.WithStack(errdefs.NotFound(errors.New("no such container")))
}
return nil, wrapError(err)
}
return &container{client: c, c8dCtr: ctr}, nil
}
func (c *container) Task(ctx context.Context) (libcontainerdtypes.Task, error) {
t, err := c.c8dCtr.Task(ctx, nil)
if err != nil {
return nil, wrapError(err)
}
return c.newTask(t), nil
}
// createIO creates the io to be used by a process
// This needs to get a pointer to interface as upon closure the process may not have yet been registered
func (c *container) createIO(fifos *cio.FIFOSet, processID string, stdinCloseSync chan containerd.Process, attachStdio libcontainerdtypes.StdioCallback) (cio.IO, error) {
var (
io *cio.DirectIO
err error
)
io, err = c.client.newDirectIO(context.Background(), fifos)
if err != nil {
return nil, err
}
if io.Stdin != nil {
var (
err error
stdinOnce sync.Once
)
pipe := io.Stdin
io.Stdin = ioutils.NewWriteCloserWrapper(pipe, func() error {
stdinOnce.Do(func() {
err = pipe.Close()
// Do the rest in a new routine to avoid a deadlock if the
// Exec/Start call failed.
go func() {
p, ok := <-stdinCloseSync
if !ok {
return
}
err = p.CloseIO(context.Background(), containerd.WithStdinCloser)
if err != nil && strings.Contains(err.Error(), "transport is closing") {
err = nil
}
}()
})
return err
})
}
rio, err := attachStdio(io)
if err != nil {
io.Cancel()
io.Close()
}
return rio, err
}
func (c *client) processEvent(ctx context.Context, et libcontainerdtypes.EventType, ei libcontainerdtypes.EventInfo) {
c.eventQ.Append(ei.ContainerID, func() {
err := c.backend.ProcessEvent(ei.ContainerID, et, ei)
if err != nil {
c.logger.WithContext(ctx).WithError(err).WithFields(logrus.Fields{
"container": ei.ContainerID,
"event": et,
"event-info": ei,
}).Error("failed to process event")
}
})
}
func (c *client) waitServe(ctx context.Context) bool {
t := 100 * time.Millisecond
delay := time.NewTimer(t)
if !delay.Stop() {
<-delay.C
}
defer delay.Stop()
// `IsServing` will actually block until the service is ready.
// However it can return early, so we'll loop with a delay to handle it.
for {
serving, err := c.client.IsServing(ctx)
if err != nil {
if errors.Is(err, context.DeadlineExceeded) || errors.Is(err, context.Canceled) {
return false
}
logrus.WithError(err).Warn("Error while testing if containerd API is ready")
}
if serving {
return true
}
delay.Reset(t)
select {
case <-ctx.Done():
return false
case <-delay.C:
}
}
}
func (c *client) processEventStream(ctx context.Context, ns string) {
var (
err error
ev *events.Envelope
et libcontainerdtypes.EventType
ei libcontainerdtypes.EventInfo
)
// Create a new context specifically for this subscription.
// The context must be cancelled to cancel the subscription.
// In cases where we have to restart event stream processing,
// we'll need the original context b/c this one will be cancelled
subCtx, cancel := context.WithCancel(ctx)
defer cancel()
// Filter on both namespace *and* topic. To create an "and" filter,
// this must be a single, comma-separated string
eventStream, errC := c.client.EventService().Subscribe(subCtx, "namespace=="+ns+",topic~=|^/tasks/|")
c.logger.Debug("processing event stream")
for {
select {
case err = <-errC:
if err != nil {
errStatus, ok := status.FromError(err)
if !ok || errStatus.Code() != codes.Canceled {
c.logger.WithError(err).Error("Failed to get event")
c.logger.Info("Waiting for containerd to be ready to restart event processing")
if c.waitServe(ctx) {
go c.processEventStream(ctx, ns)
return
}
}
c.logger.WithError(ctx.Err()).Info("stopping event stream following graceful shutdown")
}
return
case ev = <-eventStream:
if ev.Event == nil {
c.logger.WithField("event", ev).Warn("invalid event")
continue
}
v, err := typeurl.UnmarshalAny(ev.Event)
if err != nil {
c.logger.WithError(err).WithField("event", ev).Warn("failed to unmarshal event")
continue
}
c.logger.WithField("topic", ev.Topic).Debug("event")
switch t := v.(type) {
case *apievents.TaskCreate:
et = libcontainerdtypes.EventCreate
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
ProcessID: t.ContainerID,
Pid: t.Pid,
}
case *apievents.TaskStart:
et = libcontainerdtypes.EventStart
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
ProcessID: t.ContainerID,
Pid: t.Pid,
}
case *apievents.TaskExit:
et = libcontainerdtypes.EventExit
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
ProcessID: t.ID,
Pid: t.Pid,
ExitCode: t.ExitStatus,
ExitedAt: t.ExitedAt,
}
case *apievents.TaskOOM:
et = libcontainerdtypes.EventOOM
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
}
case *apievents.TaskExecAdded:
et = libcontainerdtypes.EventExecAdded
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
ProcessID: t.ExecID,
}
case *apievents.TaskExecStarted:
et = libcontainerdtypes.EventExecStarted
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
ProcessID: t.ExecID,
Pid: t.Pid,
}
case *apievents.TaskPaused:
et = libcontainerdtypes.EventPaused
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
}
case *apievents.TaskResumed:
et = libcontainerdtypes.EventResumed
ei = libcontainerdtypes.EventInfo{
ContainerID: t.ContainerID,
}
case *apievents.TaskDelete:
c.logger.WithFields(logrus.Fields{
"topic": ev.Topic,
"type": reflect.TypeOf(t),
"container": t.ContainerID},
).Info("ignoring event")
continue
default:
c.logger.WithFields(logrus.Fields{
"topic": ev.Topic,
"type": reflect.TypeOf(t)},
).Info("ignoring event")
continue
}
c.processEvent(ctx, et, ei)
}
}
}
func (c *client) writeContent(ctx context.Context, mediaType, ref string, r io.Reader) (*types.Descriptor, error) {
writer, err := c.client.ContentStore().Writer(ctx, content.WithRef(ref))
if err != nil {
return nil, err
}
defer writer.Close()
size, err := io.Copy(writer, r)
if err != nil {
return nil, err
}
labels := map[string]string{
"containerd.io/gc.root": time.Now().UTC().Format(time.RFC3339),
}
if err := writer.Commit(ctx, 0, "", content.WithLabels(labels)); err != nil {
return nil, err
}
return &types.Descriptor{
MediaType: mediaType,
Digest: writer.Digest(),
Size_: size,
}, nil
}
func (c *client) bundleDir(id string) string {
return filepath.Join(c.stateDir, id)
}
func wrapError(err error) error {
switch {
case err == nil:
return nil
case containerderrors.IsNotFound(err):
return errdefs.NotFound(err)
}
msg := err.Error()
for _, s := range []string{"container does not exist", "not found", "no such container"} {
if strings.Contains(msg, s) {
return errdefs.NotFound(err)
}
}
return err
}