moby/daemon/cluster/swarm.go

621 lines
17 KiB
Go
Raw Normal View History

package cluster // import "github.com/docker/docker/daemon/cluster"
import (
"context"
"fmt"
"net"
"strings"
"time"
"github.com/containerd/log"
"github.com/docker/docker/api/types/backend"
"github.com/docker/docker/api/types/container"
"github.com/docker/docker/api/types/filters"
types "github.com/docker/docker/api/types/swarm"
"github.com/docker/docker/daemon/cluster/convert"
"github.com/docker/docker/errdefs"
"github.com/docker/docker/opts"
"github.com/docker/docker/pkg/stack"
swarmapi "github.com/moby/swarmkit/v2/api"
"github.com/moby/swarmkit/v2/manager/encryption"
swarmnode "github.com/moby/swarmkit/v2/node"
"github.com/pkg/errors"
"google.golang.org/grpc"
)
// Init initializes new cluster from user provided request.
func (c *Cluster) Init(req types.InitRequest) (string, error) {
c.controlMutex.Lock()
defer c.controlMutex.Unlock()
if c.nr != nil {
if req.ForceNewCluster {
// Take c.mu temporarily to wait for presently running
// API handlers to finish before shutting down the node.
c.mu.Lock()
if !c.nr.nodeState.IsManager() {
c.mu.Unlock()
return "", errSwarmNotManager
}
c.mu.Unlock()
if err := c.nr.Stop(); err != nil {
return "", err
}
} else {
return "", errSwarmExists
}
}
if err := validateAndSanitizeInitRequest(&req); err != nil {
return "", errdefs.InvalidParameter(err)
}
listenHost, listenPort, err := resolveListenAddr(req.ListenAddr)
if err != nil {
return "", errdefs.InvalidParameter(err)
}
advertiseHost, advertisePort, err := c.resolveAdvertiseAddr(req.AdvertiseAddr, listenPort)
if err != nil {
return "", err
}
dataPathAddr, err := resolveDataPathAddr(req.DataPathAddr)
if err != nil {
return "", err
}
localAddr := listenHost
// If the local address is undetermined, the advertise address
// will be used as local address, if it belongs to this system.
// If the advertise address is not local, then we try to find
// a system address to use as local address. If this fails,
// we give up and ask the user to pass the listen address.
if net.ParseIP(localAddr).IsUnspecified() {
advertiseIP := net.ParseIP(advertiseHost)
found := false
for _, systemIP := range listSystemIPs() {
if systemIP.Equal(advertiseIP) {
localAddr = advertiseIP.String()
found = true
break
}
}
if !found {
ip, err := c.resolveSystemAddr()
if err != nil {
log.G(context.TODO()).Warnf("Could not find a local address: %v", err)
return "", errMustSpecifyListenAddr
}
localAddr = ip.String()
}
}
if err := validateDefaultAddrPool(req.DefaultAddrPool, req.SubnetSize); err != nil {
return "", err
}
port, err := getDataPathPort(req.DataPathPort)
if err != nil {
return "", err
}
nr, err := c.newNodeRunner(nodeStartConfig{
forceNewCluster: req.ForceNewCluster,
autolock: req.AutoLockManagers,
LocalAddr: localAddr,
ListenAddr: net.JoinHostPort(listenHost, listenPort),
AdvertiseAddr: net.JoinHostPort(advertiseHost, advertisePort),
DataPathAddr: dataPathAddr,
DefaultAddressPool: req.DefaultAddrPool,
SubnetSize: req.SubnetSize,
availability: req.Availability,
DataPathPort: port,
})
if err != nil {
return "", err
}
c.mu.Lock()
c.nr = nr
c.mu.Unlock()
if err := <-nr.Ready(); err != nil {
c.mu.Lock()
c.nr = nil
c.mu.Unlock()
if !req.ForceNewCluster { // if failure on first attempt don't keep state
if err := clearPersistentState(c.root); err != nil {
return "", err
}
}
return "", err
}
state := nr.State()
if state.swarmNode == nil { // should never happen but protect from panic
return "", errors.New("invalid cluster state for spec initialization")
}
if err := initClusterSpec(state.swarmNode, req.Spec); err != nil {
return "", err
}
return state.NodeID(), nil
}
// Join makes current Cluster part of an existing swarm cluster.
func (c *Cluster) Join(req types.JoinRequest) error {
c.controlMutex.Lock()
defer c.controlMutex.Unlock()
c.mu.Lock()
if c.nr != nil {
c.mu.Unlock()
return errors.WithStack(errSwarmExists)
}
c.mu.Unlock()
if err := validateAndSanitizeJoinRequest(&req); err != nil {
return errdefs.InvalidParameter(err)
}
listenHost, listenPort, err := resolveListenAddr(req.ListenAddr)
if err != nil {
return err
}
var advertiseAddr string
if req.AdvertiseAddr != "" {
advertiseHost, advertisePort, err := c.resolveAdvertiseAddr(req.AdvertiseAddr, listenPort)
// For joining, we don't need to provide an advertise address,
// since the remote side can detect it.
if err == nil {
advertiseAddr = net.JoinHostPort(advertiseHost, advertisePort)
}
}
dataPathAddr, err := resolveDataPathAddr(req.DataPathAddr)
if err != nil {
return err
}
nr, err := c.newNodeRunner(nodeStartConfig{
RemoteAddr: req.RemoteAddrs[0],
ListenAddr: net.JoinHostPort(listenHost, listenPort),
AdvertiseAddr: advertiseAddr,
DataPathAddr: dataPathAddr,
joinAddr: req.RemoteAddrs[0],
joinToken: req.JoinToken,
availability: req.Availability,
})
if err != nil {
return err
}
c.mu.Lock()
c.nr = nr
c.mu.Unlock()
timeout := time.NewTimer(swarmConnectTimeout)
defer timeout.Stop()
select {
case <-timeout.C:
return errSwarmJoinTimeoutReached
case err := <-nr.Ready():
if err != nil {
c.mu.Lock()
c.nr = nil
c.mu.Unlock()
if err := clearPersistentState(c.root); err != nil {
return err
}
}
return err
}
}
// Inspect retrieves the configuration properties of a managed swarm cluster.
func (c *Cluster) Inspect() (types.Swarm, error) {
var swarm types.Swarm
if err := c.lockedManagerAction(func(ctx context.Context, state nodeState) error {
s, err := c.inspect(ctx, state)
if err != nil {
return err
}
swarm = s
return nil
}); err != nil {
return types.Swarm{}, err
}
return swarm, nil
}
func (c *Cluster) inspect(ctx context.Context, state nodeState) (types.Swarm, error) {
s, err := getSwarm(ctx, state.controlClient)
if err != nil {
return types.Swarm{}, err
}
return convert.SwarmFromGRPC(*s), nil
}
// Update updates configuration of a managed swarm cluster.
func (c *Cluster) Update(version uint64, spec types.Spec, flags types.UpdateFlags) error {
return c.lockedManagerAction(func(ctx context.Context, state nodeState) error {
swarm, err := getSwarm(ctx, state.controlClient)
if err != nil {
return err
}
// Validate spec name.
if spec.Annotations.Name == "" {
spec.Annotations.Name = "default"
} else if spec.Annotations.Name != "default" {
return errdefs.InvalidParameter(errors.New(`swarm spec must be named "default"`))
}
// In update, client should provide the complete spec of the swarm, including
// Name and Labels. If a field is specified with 0 or nil, then the default value
// will be used to swarmkit.
clusterSpec, err := convert.SwarmSpecToGRPC(spec)
if err != nil {
return errdefs.InvalidParameter(err)
}
_, err = state.controlClient.UpdateCluster(
ctx,
&swarmapi.UpdateClusterRequest{
ClusterID: swarm.ID,
Spec: &clusterSpec,
ClusterVersion: &swarmapi.Version{
Index: version,
},
Rotation: swarmapi.KeyRotation{
WorkerJoinToken: flags.RotateWorkerToken,
ManagerJoinToken: flags.RotateManagerToken,
ManagerUnlockKey: flags.RotateManagerUnlockKey,
},
},
)
return err
})
}
// GetUnlockKey returns the unlock key for the swarm.
func (c *Cluster) GetUnlockKey() (string, error) {
var resp *swarmapi.GetUnlockKeyResponse
if err := c.lockedManagerAction(func(ctx context.Context, state nodeState) error {
client := swarmapi.NewCAClient(state.grpcConn)
r, err := client.GetUnlockKey(ctx, &swarmapi.GetUnlockKeyRequest{})
if err != nil {
return err
}
resp = r
return nil
}); err != nil {
return "", err
}
if len(resp.UnlockKey) == 0 {
// no key
return "", nil
}
return encryption.HumanReadableKey(resp.UnlockKey), nil
}
// UnlockSwarm provides a key to decrypt data that is encrypted at rest.
func (c *Cluster) UnlockSwarm(req types.UnlockRequest) error {
c.controlMutex.Lock()
defer c.controlMutex.Unlock()
c.mu.RLock()
state := c.currentNodeState()
if !state.IsActiveManager() {
// when manager is not active,
// unless it is locked, otherwise return error.
if err := c.errNoManager(state); err != errSwarmLocked {
c.mu.RUnlock()
return err
}
} else {
// when manager is active, return an error of "not locked"
c.mu.RUnlock()
return notLockedError{}
}
// only when swarm is locked, code running reaches here
nr := c.nr
c.mu.RUnlock()
key, err := encryption.ParseHumanReadableKey(req.UnlockKey)
if err != nil {
return errdefs.InvalidParameter(err)
}
config := nr.config
config.lockKey = key
if err := nr.Stop(); err != nil {
return err
}
nr, err = c.newNodeRunner(config)
if err != nil {
return err
}
c.mu.Lock()
c.nr = nr
c.mu.Unlock()
if err := <-nr.Ready(); err != nil {
if errors.Is(err, errSwarmLocked) {
return invalidUnlockKey{}
}
return errors.Errorf("swarm component could not be started: %v", err)
}
return nil
}
// Leave shuts down Cluster and removes current state.
func (c *Cluster) Leave(ctx context.Context, force bool) error {
c.controlMutex.Lock()
defer c.controlMutex.Unlock()
c.mu.Lock()
nr := c.nr
if nr == nil {
c.mu.Unlock()
return errors.WithStack(errNoSwarm)
}
state := c.currentNodeState()
c.mu.Unlock()
if errors.Is(state.err, errSwarmLocked) && !force {
// leave a locked swarm without --force is not allowed
return errors.WithStack(notAvailableError("Swarm is encrypted and locked. Please unlock it first or use `--force` to ignore this message."))
}
if state.IsManager() && !force {
msg := "You are attempting to leave the swarm on a node that is participating as a manager. "
if state.IsActiveManager() {
active, reachable, unreachable, err := managerStats(state.controlClient, state.NodeID())
if err == nil {
if active && removingManagerCausesLossOfQuorum(reachable, unreachable) {
if isLastManager(reachable, unreachable) {
msg += "Removing the last manager erases all current state of the swarm. Use `--force` to ignore this message. "
return errors.WithStack(notAvailableError(msg))
}
msg += fmt.Sprintf("Removing this node leaves %v managers out of %v. Without a Raft quorum your swarm will be inaccessible. ", reachable-1, reachable+unreachable)
}
}
} else {
msg += "Doing so may lose the consensus of your cluster. "
}
msg += "The only way to restore a swarm that has lost consensus is to reinitialize it with `--force-new-cluster`. Use `--force` to suppress this message."
return errors.WithStack(notAvailableError(msg))
}
// release readers in here
if err := nr.Stop(); err != nil {
log.G(ctx).Errorf("failed to shut down cluster node: %v", err)
stack.Dump()
return err
}
c.mu.Lock()
c.nr = nil
c.mu.Unlock()
if nodeID := state.NodeID(); nodeID != "" {
nodeContainers, err := c.listContainerForNode(ctx, nodeID)
if err != nil {
return err
}
for _, id := range nodeContainers {
if err := c.config.Backend.ContainerRm(id, &backend.ContainerRmConfig{ForceRemove: true}); err != nil {
log.G(ctx).Errorf("error removing %v: %v", id, err)
}
}
}
// todo: cleanup optional?
if err := clearPersistentState(c.root); err != nil {
return err
}
c.config.Backend.DaemonLeavesCluster()
return nil
}
// Info returns information about the current cluster state.
func (c *Cluster) Info(ctx context.Context) types.Info {
info := types.Info{
NodeAddr: c.GetAdvertiseAddress(),
}
c.mu.RLock()
defer c.mu.RUnlock()
state := c.currentNodeState()
info.LocalNodeState = state.status
if state.err != nil {
info.Error = state.err.Error()
}
ctx, cancel := c.getRequestContext(ctx)
defer cancel()
if state.IsActiveManager() {
info.ControlAvailable = true
swarm, err := c.inspect(ctx, state)
if err != nil {
info.Error = err.Error()
}
info.Cluster = &swarm.ClusterInfo
if r, err := state.controlClient.ListNodes(
ctx, &swarmapi.ListNodesRequest{},
grpc.MaxCallRecvMsgSize(defaultRecvSizeForListResponse),
); err != nil {
info.Error = err.Error()
} else {
info.Nodes = len(r.Nodes)
for _, n := range r.Nodes {
if n.ManagerStatus != nil {
info.Managers = info.Managers + 1
}
}
}
switch info.LocalNodeState {
case types.LocalNodeStateInactive, types.LocalNodeStateLocked, types.LocalNodeStateError:
// nothing to do
default:
if info.Managers == 2 {
const warn string = `WARNING: Running Swarm in a two-manager configuration. This configuration provides
no fault tolerance, and poses a high risk to lose control over the cluster.
Refer to https://docs.docker.com/engine/swarm/admin_guide/ to configure the
Swarm for fault-tolerance.`
info.Warnings = append(info.Warnings, warn)
}
}
}
if state.swarmNode != nil {
for _, r := range state.swarmNode.Remotes() {
info.RemoteManagers = append(info.RemoteManagers, types.Peer{NodeID: r.NodeID, Addr: r.Addr})
}
info.NodeID = state.swarmNode.NodeID()
}
return info
}
// Status returns a textual representation of the node's swarm status and role (manager/worker)
func (c *Cluster) Status() string {
c.mu.RLock()
s := c.currentNodeState()
c.mu.RUnlock()
state := string(s.status)
if s.status == types.LocalNodeStateActive {
if s.IsActiveManager() || s.IsManager() {
state += "/manager"
} else {
state += "/worker"
}
}
return state
}
func validateAndSanitizeInitRequest(req *types.InitRequest) error {
var err error
req.ListenAddr, err = validateAddr(req.ListenAddr)
if err != nil {
return fmt.Errorf("invalid ListenAddr %q: %v", req.ListenAddr, err)
}
if req.Spec.Annotations.Name == "" {
req.Spec.Annotations.Name = "default"
} else if req.Spec.Annotations.Name != "default" {
return errors.New(`swarm spec must be named "default"`)
}
return nil
}
func validateAndSanitizeJoinRequest(req *types.JoinRequest) error {
var err error
req.ListenAddr, err = validateAddr(req.ListenAddr)
if err != nil {
return fmt.Errorf("invalid ListenAddr %q: %v", req.ListenAddr, err)
}
if len(req.RemoteAddrs) == 0 {
return errors.New("at least 1 RemoteAddr is required to join")
}
for i := range req.RemoteAddrs {
req.RemoteAddrs[i], err = validateAddr(req.RemoteAddrs[i])
if err != nil {
return fmt.Errorf("invalid remoteAddr %q: %v", req.RemoteAddrs[i], err)
}
}
return nil
}
func validateAddr(addr string) (string, error) {
if addr == "" {
return addr, errors.New("invalid empty address")
}
newaddr, err := opts.ParseTCPAddr(addr, defaultAddr)
if err != nil {
// TODO(thaJeztah) why are we ignoring the error here? Is this to allow "non-tcp" addresses?
return addr, nil
}
return strings.TrimPrefix(newaddr, "tcp://"), nil
}
func initClusterSpec(node *swarmnode.Node, spec types.Spec) error {
ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
defer cancel()
for conn := range node.ListenControlSocket(ctx) {
if ctx.Err() != nil {
return ctx.Err()
}
if conn != nil {
client := swarmapi.NewControlClient(conn)
var cluster *swarmapi.Cluster
for i := 0; ; i++ {
lcr, err := client.ListClusters(ctx, &swarmapi.ListClustersRequest{})
if err != nil {
return fmt.Errorf("error on listing clusters: %v", err)
}
if len(lcr.Clusters) == 0 {
if i < 10 {
time.Sleep(200 * time.Millisecond)
continue
}
return errors.New("empty list of clusters was returned")
}
cluster = lcr.Clusters[0]
break
}
// In init, we take the initial default values from swarmkit, and merge
// any non nil or 0 value from spec to GRPC spec. This will leave the
// default value alone.
// Note that this is different from Update(), as in Update() we expect
// user to specify the complete spec of the cluster (as they already know
// the existing one and knows which field to update)
clusterSpec, err := convert.MergeSwarmSpecToGRPC(spec, cluster.Spec)
if err != nil {
return fmt.Errorf("error updating cluster settings: %v", err)
}
_, err = client.UpdateCluster(ctx, &swarmapi.UpdateClusterRequest{
ClusterID: cluster.ID,
ClusterVersion: &cluster.Meta.Version,
Spec: &clusterSpec,
})
if err != nil {
return fmt.Errorf("error updating cluster settings: %v", err)
}
return nil
}
}
return ctx.Err()
}
func (c *Cluster) listContainerForNode(ctx context.Context, nodeID string) ([]string, error) {
var ids []string
containers, err := c.config.Backend.Containers(ctx, &container.ListOptions{
Filters: filters.NewArgs(filters.Arg("label", "com.docker.swarm.node.id="+nodeID)),
})
if err != nil {
return []string{}, err
}
for _, c := range containers {
ids = append(ids, c.ID)
}
return ids, nil
}