moby/daemon/daemon.go

1427 lines
47 KiB
Go
Raw Normal View History

// Package daemon exposes the functions that occur on the host server
// that the Docker daemon is running.
//
// In implementing the various functions of the daemon, there is often
// a method-specific struct for configuring the runtime behavior.
package daemon // import "github.com/docker/docker/daemon"
2013-01-19 00:13:39 +00:00
import (
"context"
2013-01-19 00:13:39 +00:00
"fmt"
"io/ioutil"
"math/rand"
"net"
"os"
"path"
"path/filepath"
"runtime"
Remove static errors from errors package. Moving all strings to the errors package wasn't a good idea after all. Our custom implementation of Go errors predates everything that's nice and good about working with errors in Go. Take as an example what we have to do to get an error message: ```go func GetErrorMessage(err error) string { switch err.(type) { case errcode.Error: e, _ := err.(errcode.Error) return e.Message case errcode.ErrorCode: ec, _ := err.(errcode.ErrorCode) return ec.Message() default: return err.Error() } } ``` This goes against every good practice for Go development. The language already provides a simple, intuitive and standard way to get error messages, that is calling the `Error()` method from an error. Reinventing the error interface is a mistake. Our custom implementation also makes very hard to reason about errors, another nice thing about Go. I found several (>10) error declarations that we don't use anywhere. This is a clear sign about how little we know about the errors we return. I also found several error usages where the number of arguments was different than the parameters declared in the error, another clear example of how difficult is to reason about errors. Moreover, our custom implementation didn't really make easier for people to return custom HTTP status code depending on the errors. Again, it's hard to reason about when to set custom codes and how. Take an example what we have to do to extract the message and status code from an error before returning a response from the API: ```go switch err.(type) { case errcode.ErrorCode: daError, _ := err.(errcode.ErrorCode) statusCode = daError.Descriptor().HTTPStatusCode errMsg = daError.Message() case errcode.Error: // For reference, if you're looking for a particular error // then you can do something like : // import ( derr "github.com/docker/docker/errors" ) // if daError.ErrorCode() == derr.ErrorCodeNoSuchContainer { ... } daError, _ := err.(errcode.Error) statusCode = daError.ErrorCode().Descriptor().HTTPStatusCode errMsg = daError.Message default: // This part of will be removed once we've // converted everything over to use the errcode package // FIXME: this is brittle and should not be necessary. // If we need to differentiate between different possible error types, // we should create appropriate error types with clearly defined meaning errStr := strings.ToLower(err.Error()) for keyword, status := range map[string]int{ "not found": http.StatusNotFound, "no such": http.StatusNotFound, "bad parameter": http.StatusBadRequest, "conflict": http.StatusConflict, "impossible": http.StatusNotAcceptable, "wrong login/password": http.StatusUnauthorized, "hasn't been activated": http.StatusForbidden, } { if strings.Contains(errStr, keyword) { statusCode = status break } } } ``` You can notice two things in that code: 1. We have to explain how errors work, because our implementation goes against how easy to use Go errors are. 2. At no moment we arrived to remove that `switch` statement that was the original reason to use our custom implementation. This change removes all our status errors from the errors package and puts them back in their specific contexts. IT puts the messages back with their contexts. That way, we know right away when errors used and how to generate their messages. It uses custom interfaces to reason about errors. Errors that need to response with a custom status code MUST implementent this simple interface: ```go type errorWithStatus interface { HTTPErrorStatusCode() int } ``` This interface is very straightforward to implement. It also preserves Go errors real behavior, getting the message is as simple as using the `Error()` method. I included helper functions to generate errors that use custom status code in `errors/errors.go`. By doing this, we remove the hard dependency we have eeverywhere to our custom errors package. Yes, you can use it as a helper to generate error, but it's still very easy to generate errors without it. Please, read this fantastic blog post about errors in Go: http://dave.cheney.net/2014/12/24/inspecting-errors Signed-off-by: David Calavera <david.calavera@gmail.com>
2016-02-25 15:53:35 +00:00
"strings"
"sync"
"time"
"google.golang.org/grpc"
"github.com/containerd/containerd"
"github.com/containerd/containerd/defaults"
"github.com/containerd/containerd/pkg/dialer"
"github.com/containerd/containerd/remotes/docker"
"github.com/docker/distribution/reference"
"github.com/docker/docker/api/types"
containertypes "github.com/docker/docker/api/types/container"
"github.com/docker/docker/api/types/swarm"
"github.com/docker/docker/builder"
"github.com/docker/docker/container"
"github.com/docker/docker/daemon/config"
"github.com/docker/docker/daemon/discovery"
"github.com/docker/docker/daemon/events"
"github.com/docker/docker/daemon/exec"
"github.com/docker/docker/daemon/images"
"github.com/docker/docker/daemon/logger"
"github.com/docker/docker/daemon/network"
"github.com/docker/docker/errdefs"
"github.com/moby/buildkit/util/resolver"
"github.com/moby/buildkit/util/tracing"
"github.com/sirupsen/logrus"
// register graph drivers
_ "github.com/docker/docker/daemon/graphdriver/register"
"github.com/docker/docker/daemon/stats"
dmetadata "github.com/docker/docker/distribution/metadata"
"github.com/docker/docker/dockerversion"
"github.com/docker/docker/image"
"github.com/docker/docker/layer"
"github.com/docker/docker/libcontainerd"
"github.com/docker/docker/migrate/v1"
"github.com/docker/docker/pkg/idtools"
Fix race in attachable network attachment Attachable networks are networks created on the cluster which can then be attached to by non-swarm containers. These networks are lazily created on the node that wants to attach to that network. When no container is currently attached to one of these networks on a node, and then multiple containers which want that network are started concurrently, this can cause a race condition in the network attachment where essentially we try to attach the same network to the node twice. To easily reproduce this issue you must use a multi-node cluster with a worker node that has lots of CPUs (I used a 36 CPU node). Repro steps: 1. On manager, `docker network create -d overlay --attachable test` 2. On worker, `docker create --restart=always --network test busybox top`, many times... 200 is a good number (but not much more due to subnet size restrictions) 3. Restart the daemon When the daemon restarts, it will attempt to start all those containers simultaneously. Note that you could try to do this yourself over the API, but it's harder to trigger due to the added latency from going over the API. The error produced happens when the daemon tries to start the container upon allocating the network resources: ``` attaching to network failed, make sure your network options are correct and check manager logs: context deadline exceeded ``` What happens here is the worker makes a network attachment request to the manager. This is an async call which in the happy case would cause a task to be placed on the node, which the worker is waiting for to get the network configuration. In the case of this race, the error ocurrs on the manager like this: ``` task allocation failure" error="failed during network allocation for task n7bwwwbymj2o2h9asqkza8gom: failed to allocate network IP for task n7bwwwbymj2o2h9asqkza8gom network rj4szie2zfauqnpgh4eri1yue: could not find an available IP" module=node node.id=u3489c490fx1df8onlyfo1v6e ``` The task is not created and the worker times out waiting for the task. --- The mitigation for this is to make sure that only one attachment reuest is in flight for a given network at a time *when the network doesn't already exist on the node*. If the network already exists on the node there is no need for synchronization because the network is already allocated and on the node so there is no need to request it from the manager. This basically comes down to a race with `Find(network) || Create(network)` without any sort of syncronization. Signed-off-by: Brian Goff <cpuguy83@gmail.com>
2018-02-02 17:54:08 +00:00
"github.com/docker/docker/pkg/locker"
"github.com/docker/docker/pkg/plugingetter"
"github.com/docker/docker/pkg/sysinfo"
"github.com/docker/docker/pkg/system"
"github.com/docker/docker/pkg/truncindex"
"github.com/docker/docker/plugin"
pluginexec "github.com/docker/docker/plugin/executor/containerd"
refstore "github.com/docker/docker/reference"
"github.com/docker/docker/registry"
"github.com/docker/docker/runconfig"
volumesservice "github.com/docker/docker/volume/service"
"github.com/docker/libnetwork"
"github.com/docker/libnetwork/cluster"
nwconfig "github.com/docker/libnetwork/config"
"github.com/pkg/errors"
)
// ContainersNamespace is the name of the namespace used for users containers
const ContainersNamespace = "moby"
var (
errSystemNotSupported = errors.New("the Docker daemon is not supported on this platform")
)
2013-09-07 00:33:05 +00:00
// Daemon holds information about the Docker daemon.
type Daemon struct {
ID string
repository string
containers container.Store
containersReplica container.ViewDB
execCommands *exec.Store
imageService *images.ImageService
idIndex *truncindex.TruncIndex
configStore *config.Config
statsCollector *stats.Collector
defaultLogConfig containertypes.LogConfig
RegistryService registry.Service
EventsService *events.Events
netController libnetwork.NetworkController
volumes *volumesservice.VolumesService
discoveryWatcher discovery.Reloader
root string
seccompEnabled bool
apparmorEnabled bool
shutdown bool
idMapping *idtools.IdentityMapping
// TODO: move graphDrivers field to an InfoService
graphDrivers map[string]string // By operating system
PluginStore *plugin.Store // todo: remove
pluginManager *plugin.Manager
linkIndex *linkIndex
containerdCli *containerd.Client
containerd libcontainerd.Client
defaultIsolation containertypes.Isolation // Default isolation mode on Windows
clusterProvider cluster.Provider
cluster Cluster
genericResources []swarm.GenericResource
metricsPluginListener net.Listener
machineMemory uint64
seccompProfile []byte
seccompProfilePath string
diskUsageRunning int32
pruneRunning int32
hosts map[string]bool // hosts stores the addresses the daemon is listening on
startupDone chan struct{}
Fix race in attachable network attachment Attachable networks are networks created on the cluster which can then be attached to by non-swarm containers. These networks are lazily created on the node that wants to attach to that network. When no container is currently attached to one of these networks on a node, and then multiple containers which want that network are started concurrently, this can cause a race condition in the network attachment where essentially we try to attach the same network to the node twice. To easily reproduce this issue you must use a multi-node cluster with a worker node that has lots of CPUs (I used a 36 CPU node). Repro steps: 1. On manager, `docker network create -d overlay --attachable test` 2. On worker, `docker create --restart=always --network test busybox top`, many times... 200 is a good number (but not much more due to subnet size restrictions) 3. Restart the daemon When the daemon restarts, it will attempt to start all those containers simultaneously. Note that you could try to do this yourself over the API, but it's harder to trigger due to the added latency from going over the API. The error produced happens when the daemon tries to start the container upon allocating the network resources: ``` attaching to network failed, make sure your network options are correct and check manager logs: context deadline exceeded ``` What happens here is the worker makes a network attachment request to the manager. This is an async call which in the happy case would cause a task to be placed on the node, which the worker is waiting for to get the network configuration. In the case of this race, the error ocurrs on the manager like this: ``` task allocation failure" error="failed during network allocation for task n7bwwwbymj2o2h9asqkza8gom: failed to allocate network IP for task n7bwwwbymj2o2h9asqkza8gom network rj4szie2zfauqnpgh4eri1yue: could not find an available IP" module=node node.id=u3489c490fx1df8onlyfo1v6e ``` The task is not created and the worker times out waiting for the task. --- The mitigation for this is to make sure that only one attachment reuest is in flight for a given network at a time *when the network doesn't already exist on the node*. If the network already exists on the node there is no need for synchronization because the network is already allocated and on the node so there is no need to request it from the manager. This basically comes down to a race with `Find(network) || Create(network)` without any sort of syncronization. Signed-off-by: Brian Goff <cpuguy83@gmail.com>
2018-02-02 17:54:08 +00:00
attachmentStore network.AttachmentStore
attachableNetworkLock *locker.Locker
}
// StoreHosts stores the addresses the daemon is listening on
func (daemon *Daemon) StoreHosts(hosts []string) {
if daemon.hosts == nil {
daemon.hosts = make(map[string]bool)
}
for _, h := range hosts {
daemon.hosts[h] = true
}
}
// HasExperimental returns whether the experimental features of the daemon are enabled or not
func (daemon *Daemon) HasExperimental() bool {
return daemon.configStore != nil && daemon.configStore.Experimental
}
// Features returns the features map from configStore
func (daemon *Daemon) Features() *map[string]bool {
return &daemon.configStore.Features
}
// NewResolveOptionsFunc returns a call back function to resolve "registry-mirrors" and
// "insecure-registries" for buildkit
func (daemon *Daemon) NewResolveOptionsFunc() resolver.ResolveOptionsFunc {
return func(ref string) docker.ResolverOptions {
var (
registryKey = "docker.io"
mirrors = make([]string, len(daemon.configStore.Mirrors))
m = map[string]resolver.RegistryConf{}
)
// must trim "https://" or "http://" prefix
for i, v := range daemon.configStore.Mirrors {
v = strings.TrimPrefix(v, "https://")
v = strings.TrimPrefix(v, "http://")
mirrors[i] = v
}
// set "registry-mirrors"
m[registryKey] = resolver.RegistryConf{Mirrors: mirrors}
// set "insecure-registries"
for _, v := range daemon.configStore.InsecureRegistries {
v = strings.TrimPrefix(v, "http://")
m[v] = resolver.RegistryConf{
PlainHTTP: true,
}
}
def := docker.ResolverOptions{
Client: tracing.DefaultClient,
}
parsed, err := reference.ParseNormalizedNamed(ref)
if err != nil {
return def
}
host := reference.Domain(parsed)
c, ok := m[host]
if !ok {
return def
}
if len(c.Mirrors) > 0 {
def.Host = func(string) (string, error) {
return c.Mirrors[rand.Intn(len(c.Mirrors))], nil
}
}
def.PlainHTTP = c.PlainHTTP
return def
}
}
func (daemon *Daemon) restore() error {
containers := make(map[string]*container.Container)
logrus.Info("Loading containers: start.")
dir, err := ioutil.ReadDir(daemon.repository)
2013-01-19 00:13:39 +00:00
if err != nil {
return err
}
2013-12-18 18:43:42 +00:00
for _, v := range dir {
id := v.Name()
container, err := daemon.load(id)
2013-01-19 00:13:39 +00:00
if err != nil {
logrus.Errorf("Failed to load container %v: %v", id, err)
2013-01-19 00:13:39 +00:00
continue
}
if !system.IsOSSupported(container.OS) {
logrus.Errorf("Failed to load container %v: %s (%q)", id, system.ErrNotSupportedOperatingSystem, container.OS)
continue
}
// Ignore the container if it does not support the current driver being used by the graph
currentDriverForContainerOS := daemon.graphDrivers[container.OS]
if (container.Driver == "" && currentDriverForContainerOS == "aufs") || container.Driver == currentDriverForContainerOS {
rwlayer, err := daemon.imageService.GetLayerByID(container.ID, container.OS)
if err != nil {
logrus.Errorf("Failed to load container mount %v: %v", id, err)
continue
}
container.RWLayer = rwlayer
logrus.Debugf("Loaded container %v, isRunning: %v", container.ID, container.IsRunning())
containers[container.ID] = container
} else {
logrus.Debugf("Cannot load container %s because it was created with another graph driver.", container.ID)
}
}
removeContainers := make(map[string]*container.Container)
restartContainers := make(map[*container.Container]chan struct{})
activeSandboxes := make(map[string]interface{})
for id, c := range containers {
if err := daemon.registerName(c); err != nil {
logrus.Errorf("Failed to register container name %s: %s", c.ID, err)
delete(containers, id)
continue
}
if err := daemon.Register(c); err != nil {
logrus.Errorf("Failed to register container %s: %s", c.ID, err)
delete(containers, id)
continue
}
// The LogConfig.Type is empty if the container was created before docker 1.12 with default log driver.
// We should rewrite it to use the daemon defaults.
// Fixes https://github.com/docker/docker/issues/22536
if c.HostConfig.LogConfig.Type == "" {
if err := daemon.mergeAndVerifyLogConfig(&c.HostConfig.LogConfig); err != nil {
logrus.Errorf("Failed to verify log config for container %s: %q", c.ID, err)
continue
}
}
}
var (
wg sync.WaitGroup
mapLock sync.Mutex
)
for _, c := range containers {
wg.Add(1)
go func(c *container.Container) {
defer wg.Done()
daemon.backportMountSpec(c)
if err := daemon.checkpointAndSave(c); err != nil {
logrus.WithError(err).WithField("container", c.ID).Error("error saving backported mountspec to disk")
}
daemon.setStateCounter(c)
logrus.WithFields(logrus.Fields{
"container": c.ID,
"running": c.IsRunning(),
"paused": c.IsPaused(),
}).Debug("restoring container")
var (
err error
alive bool
ec uint32
exitedAt time.Time
)
alive, _, err = daemon.containerd.Restore(context.Background(), c.ID, c.InitializeStdio)
if err != nil && !errdefs.IsNotFound(err) {
logrus.Errorf("Failed to restore container %s with containerd: %s", c.ID, err)
return
}
if !alive {
ec, exitedAt, err = daemon.containerd.DeleteTask(context.Background(), c.ID)
if err != nil && !errdefs.IsNotFound(err) {
logrus.WithError(err).Errorf("Failed to delete container %s from containerd", c.ID)
return
}
} else if !daemon.configStore.LiveRestoreEnabled {
if err := daemon.kill(c, c.StopSignal()); err != nil && !errdefs.IsNotFound(err) {
logrus.WithError(err).WithField("container", c.ID).Error("error shutting down container")
return
}
}
if c.IsRunning() || c.IsPaused() {
c.RestartManager().Cancel() // manually start containers because some need to wait for swarm networking
if c.IsPaused() && alive {
s, err := daemon.containerd.Status(context.Background(), c.ID)
if err != nil {
logrus.WithError(err).WithField("container", c.ID).
Errorf("Failed to get container status")
} else {
logrus.WithField("container", c.ID).WithField("state", s).
Info("restored container paused")
switch s {
case libcontainerd.StatusPaused, libcontainerd.StatusPausing:
// nothing to do
case libcontainerd.StatusStopped:
alive = false
case libcontainerd.StatusUnknown:
logrus.WithField("container", c.ID).
Error("Unknown status for container during restore")
default:
// running
c.Lock()
c.Paused = false
daemon.setStateCounter(c)
if err := c.CheckpointTo(daemon.containersReplica); err != nil {
logrus.WithError(err).WithField("container", c.ID).
Error("Failed to update stopped container state")
}
c.Unlock()
}
}
}
if !alive {
c.Lock()
c.SetStopped(&container.ExitStatus{ExitCode: int(ec), ExitedAt: exitedAt})
daemon.Cleanup(c)
if err := c.CheckpointTo(daemon.containersReplica); err != nil {
logrus.Errorf("Failed to update stopped container %s state: %v", c.ID, err)
}
c.Unlock()
}
// we call Mount and then Unmount to get BaseFs of the container
if err := daemon.Mount(c); err != nil {
// The mount is unlikely to fail. However, in case mount fails
// the container should be allowed to restore here. Some functionalities
// (like docker exec -u user) might be missing but container is able to be
// stopped/restarted/removed.
// See #29365 for related information.
// The error is only logged here.
logrus.Warnf("Failed to mount container on getting BaseFs path %v: %v", c.ID, err)
} else {
if err := daemon.Unmount(c); err != nil {
logrus.Warnf("Failed to umount container on getting BaseFs path %v: %v", c.ID, err)
}
}
c.ResetRestartManager(false)
if !c.HostConfig.NetworkMode.IsContainer() && c.IsRunning() {
options, err := daemon.buildSandboxOptions(c)
if err != nil {
logrus.Warnf("Failed build sandbox option to restore container %s: %v", c.ID, err)
}
mapLock.Lock()
activeSandboxes[c.NetworkSettings.SandboxID] = options
mapLock.Unlock()
}
}
// get list of containers we need to restart
// Do not autostart containers which
// has endpoints in a swarm scope
// network yet since the cluster is
// not initialized yet. We will start
// it after the cluster is
// initialized.
if daemon.configStore.AutoRestart && c.ShouldRestart() && !c.NetworkSettings.HasSwarmEndpoint && c.HasBeenStartedBefore {
mapLock.Lock()
restartContainers[c] = make(chan struct{})
mapLock.Unlock()
} else if c.HostConfig != nil && c.HostConfig.AutoRemove {
mapLock.Lock()
removeContainers[c.ID] = c
mapLock.Unlock()
}
c.Lock()
if c.RemovalInProgress {
// We probably crashed in the middle of a removal, reset
// the flag.
//
// We DO NOT remove the container here as we do not
// know if the user had requested for either the
// associated volumes, network links or both to also
// be removed. So we put the container in the "dead"
// state and leave further processing up to them.
logrus.Debugf("Resetting RemovalInProgress flag from %v", c.ID)
c.RemovalInProgress = false
c.Dead = true
if err := c.CheckpointTo(daemon.containersReplica); err != nil {
logrus.Errorf("Failed to update RemovalInProgress container %s state: %v", c.ID, err)
}
}
c.Unlock()
}(c)
}
wg.Wait()
daemon.netController, err = daemon.initNetworkController(daemon.configStore, activeSandboxes)
if err != nil {
return fmt.Errorf("Error initializing network controller: %v", err)
}
// Now that all the containers are registered, register the links
for _, c := range containers {
if err := daemon.registerLinks(c, c.HostConfig); err != nil {
logrus.Errorf("failed to register link for container %s: %v", c.ID, err)
}
}
group := sync.WaitGroup{}
for c, notifier := range restartContainers {
group.Add(1)
go func(c *container.Container, chNotify chan struct{}) {
defer group.Done()
logrus.Debugf("Starting container %s", c.ID)
// ignore errors here as this is a best effort to wait for children to be
// running before we try to start the container
children := daemon.children(c)
timeout := time.After(5 * time.Second)
for _, child := range children {
if notifier, exists := restartContainers[child]; exists {
select {
case <-notifier:
case <-timeout:
}
}
}
// Make sure networks are available before starting
daemon.waitForNetworks(c)
if err := daemon.containerStart(c, "", "", true); err != nil {
logrus.Errorf("Failed to start container %s: %s", c.ID, err)
}
close(chNotify)
}(c, notifier)
}
group.Wait()
removeGroup := sync.WaitGroup{}
for id := range removeContainers {
removeGroup.Add(1)
go func(cid string) {
if err := daemon.ContainerRm(cid, &types.ContainerRmConfig{ForceRemove: true, RemoveVolume: true}); err != nil {
logrus.Errorf("Failed to remove container %s: %s", cid, err)
}
removeGroup.Done()
}(id)
}
removeGroup.Wait()
// any containers that were started above would already have had this done,
// however we need to now prepare the mountpoints for the rest of the containers as well.
// This shouldn't cause any issue running on the containers that already had this run.
// This must be run after any containers with a restart policy so that containerized plugins
// can have a chance to be running before we try to initialize them.
for _, c := range containers {
// if the container has restart policy, do not
// prepare the mountpoints since it has been done on restarting.
// This is to speed up the daemon start when a restart container
// has a volume and the volume driver is not available.
if _, ok := restartContainers[c]; ok {
continue
} else if _, ok := removeContainers[c.ID]; ok {
// container is automatically removed, skip it.
continue
}
group.Add(1)
go func(c *container.Container) {
defer group.Done()
if err := daemon.prepareMountPoints(c); err != nil {
logrus.Error(err)
}
}(c)
}
group.Wait()
logrus.Info("Loading containers: done.")
2013-01-19 00:13:39 +00:00
return nil
}
// RestartSwarmContainers restarts any autostart container which has a
// swarm endpoint.
func (daemon *Daemon) RestartSwarmContainers() {
group := sync.WaitGroup{}
for _, c := range daemon.List() {
if !c.IsRunning() && !c.IsPaused() {
// Autostart all the containers which has a
// swarm endpoint now that the cluster is
// initialized.
if daemon.configStore.AutoRestart && c.ShouldRestart() && c.NetworkSettings.HasSwarmEndpoint && c.HasBeenStartedBefore {
group.Add(1)
go func(c *container.Container) {
defer group.Done()
if err := daemon.containerStart(c, "", "", true); err != nil {
logrus.Error(err)
}
}(c)
}
}
}
group.Wait()
}
// waitForNetworks is used during daemon initialization when starting up containers
// It ensures that all of a container's networks are available before the daemon tries to start the container.
// In practice it just makes sure the discovery service is available for containers which use a network that require discovery.
func (daemon *Daemon) waitForNetworks(c *container.Container) {
if daemon.discoveryWatcher == nil {
return
}
// Make sure if the container has a network that requires discovery that the discovery service is available before starting
for netName := range c.NetworkSettings.Networks {
// If we get `ErrNoSuchNetwork` here, we can assume that it is due to discovery not being ready
// Most likely this is because the K/V store used for discovery is in a container and needs to be started
if _, err := daemon.netController.NetworkByName(netName); err != nil {
if _, ok := err.(libnetwork.ErrNoSuchNetwork); !ok {
continue
}
// use a longish timeout here due to some slowdowns in libnetwork if the k/v store is on anything other than --net=host
// FIXME: why is this slow???
logrus.Debugf("Container %s waiting for network to be ready", c.Name)
select {
case <-daemon.discoveryWatcher.ReadyCh():
case <-time.After(60 * time.Second):
}
return
}
}
}
func (daemon *Daemon) children(c *container.Container) map[string]*container.Container {
return daemon.linkIndex.children(c)
}
// parents returns the names of the parent containers of the container
// with the given name.
func (daemon *Daemon) parents(c *container.Container) map[string]*container.Container {
return daemon.linkIndex.parents(c)
}
func (daemon *Daemon) registerLink(parent, child *container.Container, alias string) error {
fullName := path.Join(parent.Name, alias)
if err := daemon.containersReplica.ReserveName(fullName, child.ID); err != nil {
if err == container.ErrNameReserved {
logrus.Warnf("error registering link for %s, to %s, as alias %s, ignoring: %v", parent.ID, child.ID, alias, err)
return nil
}
return err
}
daemon.linkIndex.link(parent, child, fullName)
return nil
}
// DaemonJoinsCluster informs the daemon has joined the cluster and provides
// the handler to query the cluster component
func (daemon *Daemon) DaemonJoinsCluster(clusterProvider cluster.Provider) {
daemon.setClusterProvider(clusterProvider)
}
// DaemonLeavesCluster informs the daemon has left the cluster
func (daemon *Daemon) DaemonLeavesCluster() {
// Daemon is in charge of removing the attachable networks with
// connected containers when the node leaves the swarm
daemon.clearAttachableNetworks()
// We no longer need the cluster provider, stop it now so that
// the network agent will stop listening to cluster events.
daemon.setClusterProvider(nil)
// Wait for the networking cluster agent to stop
daemon.netController.AgentStopWait()
// Daemon is in charge of removing the ingress network when the
// node leaves the swarm. Wait for job to be done or timeout.
// This is called also on graceful daemon shutdown. We need to
// wait, because the ingress release has to happen before the
// network controller is stopped.
if done, err := daemon.ReleaseIngress(); err == nil {
select {
case <-done:
case <-time.After(5 * time.Second):
logrus.Warn("timeout while waiting for ingress network removal")
}
} else {
logrus.Warnf("failed to initiate ingress network removal: %v", err)
}
daemon.attachmentStore.ClearAttachments()
}
// setClusterProvider sets a component for querying the current cluster state.
func (daemon *Daemon) setClusterProvider(clusterProvider cluster.Provider) {
daemon.clusterProvider = clusterProvider
daemon.netController.SetClusterProvider(clusterProvider)
Fix race in attachable network attachment Attachable networks are networks created on the cluster which can then be attached to by non-swarm containers. These networks are lazily created on the node that wants to attach to that network. When no container is currently attached to one of these networks on a node, and then multiple containers which want that network are started concurrently, this can cause a race condition in the network attachment where essentially we try to attach the same network to the node twice. To easily reproduce this issue you must use a multi-node cluster with a worker node that has lots of CPUs (I used a 36 CPU node). Repro steps: 1. On manager, `docker network create -d overlay --attachable test` 2. On worker, `docker create --restart=always --network test busybox top`, many times... 200 is a good number (but not much more due to subnet size restrictions) 3. Restart the daemon When the daemon restarts, it will attempt to start all those containers simultaneously. Note that you could try to do this yourself over the API, but it's harder to trigger due to the added latency from going over the API. The error produced happens when the daemon tries to start the container upon allocating the network resources: ``` attaching to network failed, make sure your network options are correct and check manager logs: context deadline exceeded ``` What happens here is the worker makes a network attachment request to the manager. This is an async call which in the happy case would cause a task to be placed on the node, which the worker is waiting for to get the network configuration. In the case of this race, the error ocurrs on the manager like this: ``` task allocation failure" error="failed during network allocation for task n7bwwwbymj2o2h9asqkza8gom: failed to allocate network IP for task n7bwwwbymj2o2h9asqkza8gom network rj4szie2zfauqnpgh4eri1yue: could not find an available IP" module=node node.id=u3489c490fx1df8onlyfo1v6e ``` The task is not created and the worker times out waiting for the task. --- The mitigation for this is to make sure that only one attachment reuest is in flight for a given network at a time *when the network doesn't already exist on the node*. If the network already exists on the node there is no need for synchronization because the network is already allocated and on the node so there is no need to request it from the manager. This basically comes down to a race with `Find(network) || Create(network)` without any sort of syncronization. Signed-off-by: Brian Goff <cpuguy83@gmail.com>
2018-02-02 17:54:08 +00:00
daemon.attachableNetworkLock = locker.New()
}
// IsSwarmCompatible verifies if the current daemon
// configuration is compatible with the swarm mode
func (daemon *Daemon) IsSwarmCompatible() error {
if daemon.configStore == nil {
return nil
}
return daemon.configStore.IsSwarmCompatible()
}
// NewDaemon sets up everything for the daemon to be able to service
// requests from the webserver.
func NewDaemon(ctx context.Context, config *config.Config, pluginStore *plugin.Store) (daemon *Daemon, err error) {
setDefaultMtu(config)
registryService, err := registry.NewService(config.ServiceOptions)
if err != nil {
return nil, err
}
// Ensure that we have a correct root key limit for launching containers.
if err := ModifyRootKeyLimit(); err != nil {
logrus.Warnf("unable to modify root key limit, number of containers could be limited by this quota: %v", err)
}
// Ensure we have compatible and valid configuration options
if err := verifyDaemonSettings(config); err != nil {
return nil, err
}
// Do we have a disabled network?
config.DisableBridge = isBridgeNetworkDisabled(config)
// Setup the resolv.conf
setupResolvConf(config)
// Verify the platform is supported as a daemon
if !platformSupported {
return nil, errSystemNotSupported
}
// Validate platform-specific requirements
if err := checkSystem(); err != nil {
return nil, err
}
idMapping, err := setupRemappedRoot(config)
if err != nil {
return nil, err
}
rootIDs := idMapping.RootPair()
if err := setupDaemonProcess(config); err != nil {
return nil, err
}
// set up the tmpDir to use a canonical path
tmp, err := prepareTempDir(config.Root, rootIDs)
if err != nil {
return nil, fmt.Errorf("Unable to get the TempDir under %s: %s", config.Root, err)
}
realTmp, err := getRealPath(tmp)
if err != nil {
return nil, fmt.Errorf("Unable to get the full path to the TempDir (%s): %s", tmp, err)
}
if runtime.GOOS == "windows" {
if _, err := os.Stat(realTmp); err != nil && os.IsNotExist(err) {
if err := system.MkdirAll(realTmp, 0700, ""); err != nil {
return nil, fmt.Errorf("Unable to create the TempDir (%s): %s", realTmp, err)
}
}
os.Setenv("TEMP", realTmp)
os.Setenv("TMP", realTmp)
} else {
os.Setenv("TMPDIR", realTmp)
}
d := &Daemon{
configStore: config,
PluginStore: pluginStore,
startupDone: make(chan struct{}),
}
// Ensure the daemon is properly shutdown if there is a failure during
// initialization
defer func() {
if err != nil {
if err := d.Shutdown(); err != nil {
logrus.Error(err)
}
}
}()
if err := d.setGenericResources(config); err != nil {
return nil, err
}
// set up SIGUSR1 handler on Unix-like systems, or a Win32 global event
// on Windows to dump Go routine stacks
stackDumpDir := config.Root
if execRoot := config.GetExecRoot(); execRoot != "" {
stackDumpDir = execRoot
}
d.setupDumpStackTrap(stackDumpDir)
if err := d.setupSeccompProfile(); err != nil {
return nil, err
}
// Set the default isolation mode (only applicable on Windows)
if err := d.setDefaultIsolation(); err != nil {
return nil, fmt.Errorf("error setting default isolation mode: %v", err)
}
if err := configureMaxThreads(config); err != nil {
logrus.Warnf("Failed to configure golang's threads limit: %v", err)
}
if err := ensureDefaultAppArmorProfile(); err != nil {
logrus.Errorf(err.Error())
}
daemonRepo := filepath.Join(config.Root, "containers")
Simplify/fix MkdirAll usage This subtle bug keeps lurking in because error checking for `Mkdir()` and `MkdirAll()` is slightly different wrt to `EEXIST`/`IsExist`: - for `Mkdir()`, `IsExist` error should (usually) be ignored (unless you want to make sure directory was not there before) as it means "the destination directory was already there" - for `MkdirAll()`, `IsExist` error should NEVER be ignored. Mostly, this commit just removes ignoring the IsExist error, as it should not be ignored. Also, there are a couple of cases then IsExist is handled as "directory already exist" which is wrong. As a result, some code that never worked as intended is now removed. NOTE that `idtools.MkdirAndChown()` behaves like `os.MkdirAll()` rather than `os.Mkdir()` -- so its description is amended accordingly, and its usage is handled as such (i.e. IsExist error is not ignored). For more details, a quote from my runc commit 6f82d4b (July 2015): TL;DR: check for IsExist(err) after a failed MkdirAll() is both redundant and wrong -- so two reasons to remove it. Quoting MkdirAll documentation: > MkdirAll creates a directory named path, along with any necessary > parents, and returns nil, or else returns an error. If path > is already a directory, MkdirAll does nothing and returns nil. This means two things: 1. If a directory to be created already exists, no error is returned. 2. If the error returned is IsExist (EEXIST), it means there exists a non-directory with the same name as MkdirAll need to use for directory. Example: we want to MkdirAll("a/b"), but file "a" (or "a/b") already exists, so MkdirAll fails. The above is a theory, based on quoted documentation and my UNIX knowledge. 3. In practice, though, current MkdirAll implementation [1] returns ENOTDIR in most of cases described in #2, with the exception when there is a race between MkdirAll and someone else creating the last component of MkdirAll argument as a file. In this very case MkdirAll() will indeed return EEXIST. Because of #1, IsExist check after MkdirAll is not needed. Because of #2 and #3, ignoring IsExist error is just plain wrong, as directory we require is not created. It's cleaner to report the error now. Note this error is all over the tree, I guess due to copy-paste, or trying to follow the same usage pattern as for Mkdir(), or some not quite correct examples on the Internet. [1] https://github.com/golang/go/blob/f9ed2f75/src/os/path.go Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
2017-09-25 19:39:36 +00:00
if err := idtools.MkdirAllAndChown(daemonRepo, 0700, rootIDs); err != nil {
return nil, err
}
// Create the directory where we'll store the runtime scripts (i.e. in
// order to support runtimeArgs)
daemonRuntimes := filepath.Join(config.Root, "runtimes")
Simplify/fix MkdirAll usage This subtle bug keeps lurking in because error checking for `Mkdir()` and `MkdirAll()` is slightly different wrt to `EEXIST`/`IsExist`: - for `Mkdir()`, `IsExist` error should (usually) be ignored (unless you want to make sure directory was not there before) as it means "the destination directory was already there" - for `MkdirAll()`, `IsExist` error should NEVER be ignored. Mostly, this commit just removes ignoring the IsExist error, as it should not be ignored. Also, there are a couple of cases then IsExist is handled as "directory already exist" which is wrong. As a result, some code that never worked as intended is now removed. NOTE that `idtools.MkdirAndChown()` behaves like `os.MkdirAll()` rather than `os.Mkdir()` -- so its description is amended accordingly, and its usage is handled as such (i.e. IsExist error is not ignored). For more details, a quote from my runc commit 6f82d4b (July 2015): TL;DR: check for IsExist(err) after a failed MkdirAll() is both redundant and wrong -- so two reasons to remove it. Quoting MkdirAll documentation: > MkdirAll creates a directory named path, along with any necessary > parents, and returns nil, or else returns an error. If path > is already a directory, MkdirAll does nothing and returns nil. This means two things: 1. If a directory to be created already exists, no error is returned. 2. If the error returned is IsExist (EEXIST), it means there exists a non-directory with the same name as MkdirAll need to use for directory. Example: we want to MkdirAll("a/b"), but file "a" (or "a/b") already exists, so MkdirAll fails. The above is a theory, based on quoted documentation and my UNIX knowledge. 3. In practice, though, current MkdirAll implementation [1] returns ENOTDIR in most of cases described in #2, with the exception when there is a race between MkdirAll and someone else creating the last component of MkdirAll argument as a file. In this very case MkdirAll() will indeed return EEXIST. Because of #1, IsExist check after MkdirAll is not needed. Because of #2 and #3, ignoring IsExist error is just plain wrong, as directory we require is not created. It's cleaner to report the error now. Note this error is all over the tree, I guess due to copy-paste, or trying to follow the same usage pattern as for Mkdir(), or some not quite correct examples on the Internet. [1] https://github.com/golang/go/blob/f9ed2f75/src/os/path.go Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
2017-09-25 19:39:36 +00:00
if err := system.MkdirAll(daemonRuntimes, 0700, ""); err != nil {
return nil, err
}
if err := d.loadRuntimes(); err != nil {
return nil, err
}
if runtime.GOOS == "windows" {
Simplify/fix MkdirAll usage This subtle bug keeps lurking in because error checking for `Mkdir()` and `MkdirAll()` is slightly different wrt to `EEXIST`/`IsExist`: - for `Mkdir()`, `IsExist` error should (usually) be ignored (unless you want to make sure directory was not there before) as it means "the destination directory was already there" - for `MkdirAll()`, `IsExist` error should NEVER be ignored. Mostly, this commit just removes ignoring the IsExist error, as it should not be ignored. Also, there are a couple of cases then IsExist is handled as "directory already exist" which is wrong. As a result, some code that never worked as intended is now removed. NOTE that `idtools.MkdirAndChown()` behaves like `os.MkdirAll()` rather than `os.Mkdir()` -- so its description is amended accordingly, and its usage is handled as such (i.e. IsExist error is not ignored). For more details, a quote from my runc commit 6f82d4b (July 2015): TL;DR: check for IsExist(err) after a failed MkdirAll() is both redundant and wrong -- so two reasons to remove it. Quoting MkdirAll documentation: > MkdirAll creates a directory named path, along with any necessary > parents, and returns nil, or else returns an error. If path > is already a directory, MkdirAll does nothing and returns nil. This means two things: 1. If a directory to be created already exists, no error is returned. 2. If the error returned is IsExist (EEXIST), it means there exists a non-directory with the same name as MkdirAll need to use for directory. Example: we want to MkdirAll("a/b"), but file "a" (or "a/b") already exists, so MkdirAll fails. The above is a theory, based on quoted documentation and my UNIX knowledge. 3. In practice, though, current MkdirAll implementation [1] returns ENOTDIR in most of cases described in #2, with the exception when there is a race between MkdirAll and someone else creating the last component of MkdirAll argument as a file. In this very case MkdirAll() will indeed return EEXIST. Because of #1, IsExist check after MkdirAll is not needed. Because of #2 and #3, ignoring IsExist error is just plain wrong, as directory we require is not created. It's cleaner to report the error now. Note this error is all over the tree, I guess due to copy-paste, or trying to follow the same usage pattern as for Mkdir(), or some not quite correct examples on the Internet. [1] https://github.com/golang/go/blob/f9ed2f75/src/os/path.go Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
2017-09-25 19:39:36 +00:00
if err := system.MkdirAll(filepath.Join(config.Root, "credentialspecs"), 0, ""); err != nil {
return nil, err
}
}
// On Windows we don't support the environment variable, or a user supplied graphdriver
// as Windows has no choice in terms of which graphdrivers to use. It's a case of
// running Windows containers on Windows - windowsfilter, running Linux containers on Windows,
// lcow. Unix platforms however run a single graphdriver for all containers, and it can
// be set through an environment variable, a daemon start parameter, or chosen through
// initialization of the layerstore through driver priority order for example.
d.graphDrivers = make(map[string]string)
layerStores := make(map[string]layer.Store)
if runtime.GOOS == "windows" {
d.graphDrivers[runtime.GOOS] = "windowsfilter"
if system.LCOWSupported() {
d.graphDrivers["linux"] = "lcow"
}
} else {
driverName := os.Getenv("DOCKER_DRIVER")
if driverName == "" {
driverName = config.GraphDriver
} else {
logrus.Infof("Setting the storage driver from the $DOCKER_DRIVER environment variable (%s)", driverName)
}
d.graphDrivers[runtime.GOOS] = driverName // May still be empty. Layerstore init determines instead.
}
d.RegistryService = registryService
logger.RegisterPluginGetter(d.PluginStore)
metricsSockPath, err := d.listenMetricsSock()
if err != nil {
return nil, err
}
registerMetricsPluginCallback(d.PluginStore, metricsSockPath)
gopts := []grpc.DialOption{
grpc.WithInsecure(),
grpc.WithBackoffMaxDelay(3 * time.Second),
grpc.WithDialer(dialer.Dialer),
// TODO(stevvooe): We may need to allow configuration of this on the client.
grpc.WithDefaultCallOptions(grpc.MaxCallRecvMsgSize(defaults.DefaultMaxRecvMsgSize)),
grpc.WithDefaultCallOptions(grpc.MaxCallSendMsgSize(defaults.DefaultMaxSendMsgSize)),
}
if config.ContainerdAddr != "" {
d.containerdCli, err = containerd.New(config.ContainerdAddr, containerd.WithDefaultNamespace(ContainersNamespace), containerd.WithDialOpts(gopts), containerd.WithTimeout(60*time.Second))
if err != nil {
return nil, errors.Wrapf(err, "failed to dial %q", config.ContainerdAddr)
}
}
createPluginExec := func(m *plugin.Manager) (plugin.Executor, error) {
var pluginCli *containerd.Client
// Windows is not currently using containerd, keep the
// client as nil
if config.ContainerdAddr != "" {
pluginCli, err = containerd.New(config.ContainerdAddr, containerd.WithDefaultNamespace(pluginexec.PluginNamespace), containerd.WithDialOpts(gopts), containerd.WithTimeout(60*time.Second))
if err != nil {
return nil, errors.Wrapf(err, "failed to dial %q", config.ContainerdAddr)
}
}
return pluginexec.New(ctx, getPluginExecRoot(config.Root), pluginCli, m)
}
// Plugin system initialization should happen before restore. Do not change order.
d.pluginManager, err = plugin.NewManager(plugin.ManagerConfig{
Root: filepath.Join(config.Root, "plugins"),
ExecRoot: getPluginExecRoot(config.Root),
Store: d.PluginStore,
CreateExecutor: createPluginExec,
RegistryService: registryService,
LiveRestoreEnabled: config.LiveRestoreEnabled,
LogPluginEvent: d.LogPluginEvent, // todo: make private
AuthzMiddleware: config.AuthzMiddleware,
})
if err != nil {
return nil, errors.Wrap(err, "couldn't create plugin manager")
}
if err := d.setupDefaultLogConfig(); err != nil {
return nil, err
}
for operatingSystem, gd := range d.graphDrivers {
layerStores[operatingSystem], err = layer.NewStoreFromOptions(layer.StoreOptions{
Root: config.Root,
MetadataStorePathTemplate: filepath.Join(config.Root, "image", "%s", "layerdb"),
GraphDriver: gd,
GraphDriverOptions: config.GraphOptions,
IDMapping: idMapping,
PluginGetter: d.PluginStore,
ExperimentalEnabled: config.Experimental,
OS: operatingSystem,
})
if err != nil {
return nil, err
}
}
// As layerstore initialization may set the driver
for os := range d.graphDrivers {
d.graphDrivers[os] = layerStores[os].DriverName()
}
// Configure and validate the kernels security support. Note this is a Linux/FreeBSD
// operation only, so it is safe to pass *just* the runtime OS graphdriver.
if err := configureKernelSecuritySupport(config, d.graphDrivers[runtime.GOOS]); err != nil {
return nil, err
}
imageRoot := filepath.Join(config.Root, "image", d.graphDrivers[runtime.GOOS])
ifs, err := image.NewFSStoreBackend(filepath.Join(imageRoot, "imagedb"))
if err != nil {
return nil, err
}
lgrMap := make(map[string]image.LayerGetReleaser)
for os, ls := range layerStores {
lgrMap[os] = ls
}
imageStore, err := image.NewImageStore(ifs, lgrMap)
if err != nil {
return nil, err
}
d.volumes, err = volumesservice.NewVolumeService(config.Root, d.PluginStore, rootIDs, d)
if err != nil {
return nil, err
}
trustKey, err := loadOrCreateTrustKey(config.TrustKeyPath)
if err != nil {
return nil, err
}
trustDir := filepath.Join(config.Root, "trust")
if err := system.MkdirAll(trustDir, 0700, ""); err != nil {
return nil, err
}
// We have a single tag/reference store for the daemon globally. However, it's
// stored under the graphdriver. On host platforms which only support a single
// container OS, but multiple selectable graphdrivers, this means depending on which
// graphdriver is chosen, the global reference store is under there. For
// platforms which support multiple container operating systems, this is slightly
// more problematic as where does the global ref store get located? Fortunately,
// for Windows, which is currently the only daemon supporting multiple container
// operating systems, the list of graphdrivers available isn't user configurable.
// For backwards compatibility, we just put it under the windowsfilter
// directory regardless.
refStoreLocation := filepath.Join(imageRoot, `repositories.json`)
rs, err := refstore.NewReferenceStore(refStoreLocation)
if err != nil {
return nil, fmt.Errorf("Couldn't create reference store repository: %s", err)
}
distributionMetadataStore, err := dmetadata.NewFSMetadataStore(filepath.Join(imageRoot, "distribution"))
if err != nil {
return nil, err
}
// No content-addressability migration on Windows as it never supported pre-CA
if runtime.GOOS != "windows" {
migrationStart := time.Now()
if err := v1.Migrate(config.Root, d.graphDrivers[runtime.GOOS], layerStores[runtime.GOOS], imageStore, rs, distributionMetadataStore); err != nil {
logrus.Errorf("Graph migration failed: %q. Your old graph data was found to be too inconsistent for upgrading to content-addressable storage. Some of the old data was probably not upgraded. We recommend starting over with a clean storage directory if possible.", err)
}
logrus.Infof("Graph migration to content-addressability took %.2f seconds", time.Since(migrationStart).Seconds())
}
// Discovery is only enabled when the daemon is launched with an address to advertise. When
// initialized, the daemon is registered and we can store the discovery backend as it's read-only
if err := d.initDiscovery(config); err != nil {
return nil, err
}
sysInfo := sysinfo.New(false)
// Check if Devices cgroup is mounted, it is hard requirement for container security,
// on Linux.
if runtime.GOOS == "linux" && !sysInfo.CgroupDevicesEnabled {
return nil, errors.New("Devices cgroup isn't mounted")
}
d.ID = trustKey.PublicKey().KeyID()
d.repository = daemonRepo
d.containers = container.NewMemoryStore()
if d.containersReplica, err = container.NewViewDB(); err != nil {
return nil, err
}
d.execCommands = exec.NewStore()
d.idIndex = truncindex.NewTruncIndex([]string{})
d.statsCollector = d.newStatsCollector(1 * time.Second)
d.EventsService = events.New()
d.root = config.Root
d.idMapping = idMapping
d.seccompEnabled = sysInfo.Seccomp
d.apparmorEnabled = sysInfo.AppArmor
d.linkIndex = newLinkIndex()
// TODO: imageStore, distributionMetadataStore, and ReferenceStore are only
// used above to run migration. They could be initialized in ImageService
// if migration is called from daemon/images. layerStore might move as well.
d.imageService = images.NewImageService(images.ImageServiceConfig{
ContainerStore: d.containers,
DistributionMetadataStore: distributionMetadataStore,
EventsService: d.EventsService,
ImageStore: imageStore,
LayerStores: layerStores,
MaxConcurrentDownloads: *config.MaxConcurrentDownloads,
MaxConcurrentUploads: *config.MaxConcurrentUploads,
ReferenceStore: rs,
RegistryService: registryService,
TrustKey: trustKey,
})
go d.execCommandGC()
d.containerd, err = libcontainerd.NewClient(ctx, d.containerdCli, filepath.Join(config.ExecRoot, "containerd"), ContainersNamespace, d)
if err != nil {
return nil, err
}
if err := d.restore(); err != nil {
return nil, err
}
close(d.startupDone)
// FIXME: this method never returns an error
info, _ := d.SystemInfo()
engineInfo.WithValues(
dockerversion.Version,
dockerversion.GitCommit,
info.Architecture,
info.Driver,
info.KernelVersion,
info.OperatingSystem,
info.OSType,
info.ID,
).Set(1)
engineCpus.Set(float64(info.NCPU))
engineMemory.Set(float64(info.MemTotal))
gd := ""
for os, driver := range d.graphDrivers {
if len(gd) > 0 {
gd += ", "
}
gd += driver
if len(d.graphDrivers) > 1 {
gd = fmt.Sprintf("%s (%s)", gd, os)
}
}
logrus.WithFields(logrus.Fields{
"version": dockerversion.Version,
"commit": dockerversion.GitCommit,
"graphdriver(s)": gd,
}).Info("Docker daemon")
return d, nil
}
// DistributionServices returns services controlling daemon storage
func (daemon *Daemon) DistributionServices() images.DistributionServices {
return daemon.imageService.DistributionServices()
}
func (daemon *Daemon) waitForStartupDone() {
<-daemon.startupDone
}
func (daemon *Daemon) shutdownContainer(c *container.Container) error {
stopTimeout := c.StopTimeout()
// If container failed to exit in stopTimeout seconds of SIGTERM, then using the force
if err := daemon.containerStop(c, stopTimeout); err != nil {
return fmt.Errorf("Failed to stop container %s with error: %v", c.ID, err)
}
// Wait without timeout for the container to exit.
// Ignore the result.
<-c.Wait(context.Background(), container.WaitConditionNotRunning)
return nil
}
// ShutdownTimeout returns the timeout (in seconds) before containers are forcibly
// killed during shutdown. The default timeout can be configured both on the daemon
// and per container, and the longest timeout will be used. A grace-period of
// 5 seconds is added to the configured timeout.
//
// A negative (-1) timeout means "indefinitely", which means that containers
// are not forcibly killed, and the daemon shuts down after all containers exit.
func (daemon *Daemon) ShutdownTimeout() int {
shutdownTimeout := daemon.configStore.ShutdownTimeout
if shutdownTimeout < 0 {
return -1
}
if daemon.containers == nil {
return shutdownTimeout
}
graceTimeout := 5
for _, c := range daemon.containers.List() {
stopTimeout := c.StopTimeout()
if stopTimeout < 0 {
return -1
}
if stopTimeout+graceTimeout > shutdownTimeout {
shutdownTimeout = stopTimeout + graceTimeout
}
}
return shutdownTimeout
}
// Shutdown stops the daemon.
func (daemon *Daemon) Shutdown() error {
daemon.shutdown = true
// Keep mounts and networking running on daemon shutdown if
// we are to keep containers running and restore them.
if daemon.configStore.LiveRestoreEnabled && daemon.containers != nil {
// check if there are any running containers, if none we should do some cleanup
if ls, err := daemon.Containers(&types.ContainerListOptions{}); len(ls) != 0 || err != nil {
// metrics plugins still need some cleanup
daemon.cleanupMetricsPlugins()
return nil
}
}
if daemon.containers != nil {
logrus.Debugf("daemon configured with a %d seconds minimum shutdown timeout", daemon.configStore.ShutdownTimeout)
logrus.Debugf("start clean shutdown of all containers with a %d seconds timeout...", daemon.ShutdownTimeout())
daemon.containers.ApplyAll(func(c *container.Container) {
if !c.IsRunning() {
return
}
logrus.Debugf("stopping %s", c.ID)
if err := daemon.shutdownContainer(c); err != nil {
logrus.Errorf("Stop container error: %v", err)
return
}
if mountid, err := daemon.imageService.GetLayerMountID(c.ID, c.OS); err == nil {
daemon.cleanupMountsByID(mountid)
}
logrus.Debugf("container stopped %s", c.ID)
})
}
if daemon.volumes != nil {
if err := daemon.volumes.Shutdown(); err != nil {
logrus.Errorf("Error shutting down volume store: %v", err)
}
}
if daemon.imageService != nil {
daemon.imageService.Cleanup()
}
// If we are part of a cluster, clean up cluster's stuff
if daemon.clusterProvider != nil {
logrus.Debugf("start clean shutdown of cluster resources...")
daemon.DaemonLeavesCluster()
}
daemon.cleanupMetricsPlugins()
// Shutdown plugins after containers and layerstore. Don't change the order.
daemon.pluginShutdown()
// trigger libnetwork Stop only if it's initialized
if daemon.netController != nil {
daemon.netController.Stop()
}
if daemon.containerdCli != nil {
daemon.containerdCli.Close()
}
return daemon.cleanupMounts()
}
// Mount sets container.BaseFS
// (is it not set coming in? why is it unset?)
func (daemon *Daemon) Mount(container *container.Container) error {
if container.RWLayer == nil {
return errors.New("RWLayer of container " + container.ID + " is unexpectedly nil")
}
dir, err := container.RWLayer.Mount(container.GetMountLabel())
if err != nil {
return err
}
logrus.Debugf("container mounted via layerStore: %v", dir)
if container.BaseFS != nil && container.BaseFS.Path() != dir.Path() {
// The mount path reported by the graph driver should always be trusted on Windows, since the
// volume path for a given mounted layer may change over time. This should only be an error
// on non-Windows operating systems.
if runtime.GOOS != "windows" {
daemon.Unmount(container)
return fmt.Errorf("Error: driver %s is returning inconsistent paths for container %s ('%s' then '%s')",
daemon.imageService.GraphDriverForOS(container.OS), container.ID, container.BaseFS, dir)
}
}
container.BaseFS = dir // TODO: combine these fields
return nil
}
// Unmount unsets the container base filesystem
func (daemon *Daemon) Unmount(container *container.Container) error {
if container.RWLayer == nil {
return errors.New("RWLayer of container " + container.ID + " is unexpectedly nil")
}
if err := container.RWLayer.Unmount(); err != nil {
logrus.Errorf("Error unmounting container %s: %s", container.ID, err)
return err
}
return nil
}
// Subnets return the IPv4 and IPv6 subnets of networks that are manager by Docker.
func (daemon *Daemon) Subnets() ([]net.IPNet, []net.IPNet) {
var v4Subnets []net.IPNet
var v6Subnets []net.IPNet
Split advertised address from listen address There are currently problems with "swarm init" and "swarm join" when an explicit --listen-addr flag is not provided. swarmkit defaults to finding the IP address associated with the default route, and in cloud setups this is often the wrong choice. Introduce a notion of "advertised address", with the client flag --advertise-addr, and the daemon flag --swarm-default-advertise-addr to provide a default. The default listening address is now 0.0.0.0, but a valid advertised address must be detected or specified. If no explicit advertised address is specified, error out if there is more than one usable candidate IP address on the system. This requires a user to explicitly choose instead of letting swarmkit make the wrong choice. For the purposes of this autodetection, we ignore certain interfaces that are unlikely to be relevant (currently docker*). The user is also required to choose a listen address on swarm init if they specify an explicit advertise address that is a hostname or an IP address that's not local to the system. This is a requirement for overlay networking. Also support specifying interface names to --listen-addr, --advertise-addr, and the daemon flag --swarm-default-advertise-addr. This will fail if the interface has multiple IP addresses (unless it has a single IPv4 address and a single IPv6 address - then we resolve the tie in favor of IPv4). This change also exposes the node's externally-reachable address in docker info, as requested by #24017. Make corresponding API and CLI docs changes. Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
2016-07-01 01:07:35 +00:00
managedNetworks := daemon.netController.Networks()
for _, managedNetwork := range managedNetworks {
v4infos, v6infos := managedNetwork.Info().IpamInfo()
for _, info := range v4infos {
if info.IPAMData.Pool != nil {
v4Subnets = append(v4Subnets, *info.IPAMData.Pool)
Split advertised address from listen address There are currently problems with "swarm init" and "swarm join" when an explicit --listen-addr flag is not provided. swarmkit defaults to finding the IP address associated with the default route, and in cloud setups this is often the wrong choice. Introduce a notion of "advertised address", with the client flag --advertise-addr, and the daemon flag --swarm-default-advertise-addr to provide a default. The default listening address is now 0.0.0.0, but a valid advertised address must be detected or specified. If no explicit advertised address is specified, error out if there is more than one usable candidate IP address on the system. This requires a user to explicitly choose instead of letting swarmkit make the wrong choice. For the purposes of this autodetection, we ignore certain interfaces that are unlikely to be relevant (currently docker*). The user is also required to choose a listen address on swarm init if they specify an explicit advertise address that is a hostname or an IP address that's not local to the system. This is a requirement for overlay networking. Also support specifying interface names to --listen-addr, --advertise-addr, and the daemon flag --swarm-default-advertise-addr. This will fail if the interface has multiple IP addresses (unless it has a single IPv4 address and a single IPv6 address - then we resolve the tie in favor of IPv4). This change also exposes the node's externally-reachable address in docker info, as requested by #24017. Make corresponding API and CLI docs changes. Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
2016-07-01 01:07:35 +00:00
}
}
for _, info := range v6infos {
if info.IPAMData.Pool != nil {
v6Subnets = append(v6Subnets, *info.IPAMData.Pool)
Split advertised address from listen address There are currently problems with "swarm init" and "swarm join" when an explicit --listen-addr flag is not provided. swarmkit defaults to finding the IP address associated with the default route, and in cloud setups this is often the wrong choice. Introduce a notion of "advertised address", with the client flag --advertise-addr, and the daemon flag --swarm-default-advertise-addr to provide a default. The default listening address is now 0.0.0.0, but a valid advertised address must be detected or specified. If no explicit advertised address is specified, error out if there is more than one usable candidate IP address on the system. This requires a user to explicitly choose instead of letting swarmkit make the wrong choice. For the purposes of this autodetection, we ignore certain interfaces that are unlikely to be relevant (currently docker*). The user is also required to choose a listen address on swarm init if they specify an explicit advertise address that is a hostname or an IP address that's not local to the system. This is a requirement for overlay networking. Also support specifying interface names to --listen-addr, --advertise-addr, and the daemon flag --swarm-default-advertise-addr. This will fail if the interface has multiple IP addresses (unless it has a single IPv4 address and a single IPv6 address - then we resolve the tie in favor of IPv4). This change also exposes the node's externally-reachable address in docker info, as requested by #24017. Make corresponding API and CLI docs changes. Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
2016-07-01 01:07:35 +00:00
}
}
}
return v4Subnets, v6Subnets
Split advertised address from listen address There are currently problems with "swarm init" and "swarm join" when an explicit --listen-addr flag is not provided. swarmkit defaults to finding the IP address associated with the default route, and in cloud setups this is often the wrong choice. Introduce a notion of "advertised address", with the client flag --advertise-addr, and the daemon flag --swarm-default-advertise-addr to provide a default. The default listening address is now 0.0.0.0, but a valid advertised address must be detected or specified. If no explicit advertised address is specified, error out if there is more than one usable candidate IP address on the system. This requires a user to explicitly choose instead of letting swarmkit make the wrong choice. For the purposes of this autodetection, we ignore certain interfaces that are unlikely to be relevant (currently docker*). The user is also required to choose a listen address on swarm init if they specify an explicit advertise address that is a hostname or an IP address that's not local to the system. This is a requirement for overlay networking. Also support specifying interface names to --listen-addr, --advertise-addr, and the daemon flag --swarm-default-advertise-addr. This will fail if the interface has multiple IP addresses (unless it has a single IPv4 address and a single IPv6 address - then we resolve the tie in favor of IPv4). This change also exposes the node's externally-reachable address in docker info, as requested by #24017. Make corresponding API and CLI docs changes. Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
2016-07-01 01:07:35 +00:00
}
// prepareTempDir prepares and returns the default directory to use
// for temporary files.
// If it doesn't exist, it is created. If it exists, its content is removed.
func prepareTempDir(rootDir string, rootIdentity idtools.Identity) (string, error) {
var tmpDir string
if tmpDir = os.Getenv("DOCKER_TMPDIR"); tmpDir == "" {
tmpDir = filepath.Join(rootDir, "tmp")
newName := tmpDir + "-old"
if err := os.Rename(tmpDir, newName); err == nil {
go func() {
if err := os.RemoveAll(newName); err != nil {
logrus.Warnf("failed to delete old tmp directory: %s", newName)
}
}()
} else if !os.IsNotExist(err) {
logrus.Warnf("failed to rename %s for background deletion: %s. Deleting synchronously", tmpDir, err)
if err := os.RemoveAll(tmpDir); err != nil {
logrus.Warnf("failed to delete old tmp directory: %s", tmpDir)
}
}
}
// We don't remove the content of tmpdir if it's not the default,
// it may hold things that do not belong to us.
return tmpDir, idtools.MkdirAllAndChown(tmpDir, 0700, rootIdentity)
}
func (daemon *Daemon) setGenericResources(conf *config.Config) error {
genericResources, err := config.ParseGenericResources(conf.NodeGenericResources)
if err != nil {
return err
}
daemon.genericResources = genericResources
return nil
}
func setDefaultMtu(conf *config.Config) {
// do nothing if the config does not have the default 0 value.
if conf.Mtu != 0 {
return
}
conf.Mtu = config.DefaultNetworkMtu
}
// IsShuttingDown tells whether the daemon is shutting down or not
func (daemon *Daemon) IsShuttingDown() bool {
return daemon.shutdown
}
// initDiscovery initializes the discovery watcher for this daemon.
func (daemon *Daemon) initDiscovery(conf *config.Config) error {
advertise, err := config.ParseClusterAdvertiseSettings(conf.ClusterStore, conf.ClusterAdvertise)
if err != nil {
if err == discovery.ErrDiscoveryDisabled {
return nil
}
return err
}
conf.ClusterAdvertise = advertise
discoveryWatcher, err := discovery.Init(conf.ClusterStore, conf.ClusterAdvertise, conf.ClusterOpts)
if err != nil {
return fmt.Errorf("discovery initialization failed (%v)", err)
}
daemon.discoveryWatcher = discoveryWatcher
return nil
}
func isBridgeNetworkDisabled(conf *config.Config) bool {
return conf.BridgeConfig.Iface == config.DisableNetworkBridge
}
func (daemon *Daemon) networkOptions(dconfig *config.Config, pg plugingetter.PluginGetter, activeSandboxes map[string]interface{}) ([]nwconfig.Option, error) {
options := []nwconfig.Option{}
if dconfig == nil {
return options, nil
}
options = append(options, nwconfig.OptionExperimental(dconfig.Experimental))
options = append(options, nwconfig.OptionDataDir(dconfig.Root))
options = append(options, nwconfig.OptionExecRoot(dconfig.GetExecRoot()))
dd := runconfig.DefaultDaemonNetworkMode()
dn := runconfig.DefaultDaemonNetworkMode().NetworkName()
options = append(options, nwconfig.OptionDefaultDriver(string(dd)))
options = append(options, nwconfig.OptionDefaultNetwork(dn))
if strings.TrimSpace(dconfig.ClusterStore) != "" {
kv := strings.Split(dconfig.ClusterStore, "://")
if len(kv) != 2 {
return nil, errors.New("kv store daemon config must be of the form KV-PROVIDER://KV-URL")
}
options = append(options, nwconfig.OptionKVProvider(kv[0]))
options = append(options, nwconfig.OptionKVProviderURL(kv[1]))
}
if len(dconfig.ClusterOpts) > 0 {
options = append(options, nwconfig.OptionKVOpts(dconfig.ClusterOpts))
}
if daemon.discoveryWatcher != nil {
options = append(options, nwconfig.OptionDiscoveryWatcher(daemon.discoveryWatcher))
}
if dconfig.ClusterAdvertise != "" {
options = append(options, nwconfig.OptionDiscoveryAddress(dconfig.ClusterAdvertise))
}
options = append(options, nwconfig.OptionLabels(dconfig.Labels))
options = append(options, driverOptions(dconfig)...)
if len(dconfig.NetworkConfig.DefaultAddressPools.Value()) > 0 {
options = append(options, nwconfig.OptionDefaultAddressPoolConfig(dconfig.NetworkConfig.DefaultAddressPools.Value()))
}
if daemon.configStore != nil && daemon.configStore.LiveRestoreEnabled && len(activeSandboxes) != 0 {
options = append(options, nwconfig.OptionActiveSandboxes(activeSandboxes))
}
if pg != nil {
options = append(options, nwconfig.OptionPluginGetter(pg))
}
options = append(options, nwconfig.OptionNetworkControlPlaneMTU(dconfig.NetworkControlPlaneMTU))
return options, nil
}
// GetCluster returns the cluster
func (daemon *Daemon) GetCluster() Cluster {
return daemon.cluster
}
// SetCluster sets the cluster
func (daemon *Daemon) SetCluster(cluster Cluster) {
daemon.cluster = cluster
}
func (daemon *Daemon) pluginShutdown() {
manager := daemon.pluginManager
// Check for a valid manager object. In error conditions, daemon init can fail
// and shutdown called, before plugin manager is initialized.
if manager != nil {
manager.Shutdown()
}
}
// PluginManager returns current pluginManager associated with the daemon
func (daemon *Daemon) PluginManager() *plugin.Manager { // set up before daemon to avoid this method
return daemon.pluginManager
}
// PluginGetter returns current pluginStore associated with the daemon
func (daemon *Daemon) PluginGetter() *plugin.Store {
return daemon.PluginStore
}
// CreateDaemonRoot creates the root for the daemon
func CreateDaemonRoot(config *config.Config) error {
// get the canonical path to the Docker root directory
var realRoot string
if _, err := os.Stat(config.Root); err != nil && os.IsNotExist(err) {
realRoot = config.Root
} else {
realRoot, err = getRealPath(config.Root)
if err != nil {
return fmt.Errorf("Unable to get the full path to root (%s): %s", config.Root, err)
}
}
idMapping, err := setupRemappedRoot(config)
if err != nil {
return err
}
return setupDaemonRoot(config, realRoot, idMapping.RootPair())
}
// checkpointAndSave grabs a container lock to safely call container.CheckpointTo
func (daemon *Daemon) checkpointAndSave(container *container.Container) error {
container.Lock()
defer container.Unlock()
if err := container.CheckpointTo(daemon.containersReplica); err != nil {
return fmt.Errorf("Error saving container state: %v", err)
}
return nil
}
// because the CLI sends a -1 when it wants to unset the swappiness value
// we need to clear it on the server side
func fixMemorySwappiness(resources *containertypes.Resources) {
if resources.MemorySwappiness != nil && *resources.MemorySwappiness == -1 {
resources.MemorySwappiness = nil
}
}
// GetAttachmentStore returns current attachment store associated with the daemon
func (daemon *Daemon) GetAttachmentStore() *network.AttachmentStore {
return &daemon.attachmentStore
}
// IdentityMapping returns uid/gid mapping or a SID (in the case of Windows) for the builder
func (daemon *Daemon) IdentityMapping() *idtools.IdentityMapping {
return daemon.idMapping
}
// ImageService returns the Daemon's ImageService
func (daemon *Daemon) ImageService() *images.ImageService {
return daemon.imageService
}
// BuilderBackend returns the backend used by builder
func (daemon *Daemon) BuilderBackend() builder.Backend {
return struct {
*Daemon
*images.ImageService
}{daemon, daemon.imageService}
}