// RenewTLSConfig will continuously monitor for the necessity of renewing the local certificates, either by
// issuing them locally if key-material is available, or requesting them from a remote CA.
func RenewTLSConfig(ctx context.Context, s *SecurityConfig, remotes remotes.Remotes, renew <-chan struct{}) <-chan CertificateUpdate {
updates := make(chan CertificateUpdate)
go func() {
var retry time.Duration
defer close(updates)
for {
ctx = log.WithModule(ctx, "tls")
log := log.G(ctx).WithFields(logrus.Fields{
"node.id": s.ClientTLSCreds.NodeID(),
"node.role": s.ClientTLSCreds.Role(),
})
// Our starting default will be 5 minutes
retry = 5 * time.Minute
// Since the expiration of the certificate is managed remotely we should update our
// retry timer on every iteration of this loop.
// Retrieve the current certificate expiration information.
validFrom, validUntil, err := readCertValidity(s.KeyReader())
if err != nil {
// We failed to read the expiration, let's stick with the starting default
log.Errorf("failed to read the expiration of the TLS certificate in: %s", s.KeyReader().Target())
updates <- CertificateUpdate{Err: errors.New("failed to read certificate expiration")}
} else {
// If we have an expired certificate, we let's stick with the starting default in
// the hope that this is a temporary clock skew.
if validUntil.Before(time.Now()) {
log.WithError(err).Errorf("failed to create a new client TLS config")
updates <- CertificateUpdate{Err: errors.New("TLS certificate is expired")}
} else {
// Random retry time between 50% and 80% of the total time to expiration
retry = calculateRandomExpiry(validFrom, validUntil)
}
}
log.WithFields(logrus.Fields{
"time": time.Now().Add(retry),
}).Debugf("next certificate renewal scheduled")
select {
case <-time.After(retry):
log.Infof("renewing certificate")
case <-renew:
log.Infof("forced certificate renewal")
case <-ctx.Done():
log.Infof("shuting down certificate renewal routine")
return
}
// ignore errors - it will just try again laster
if err := RenewTLSConfigNow(ctx, s, remotes); err != nil {
updates <- CertificateUpdate{Err: err}
} else {
updates <- CertificateUpdate{Role: s.ClientTLSCreds.Role()}
}
}
}()
return updates
}
// RenewTLSConfig will continuously monitor for the necessity of renewing the local certificates, either by
// issuing them locally if key-material is available, or requesting them from a remote CA.
func RenewTLSConfig(ctx context.Context, s *SecurityConfig, baseCertDir string, remotes remotes.Remotes, renew <-chan struct{}) <-chan CertificateUpdate {
paths := NewConfigPaths(baseCertDir)
updates := make(chan CertificateUpdate)
go func() {
var retry time.Duration
defer close(updates)
for {
ctx = log.WithModule(ctx, "tls")
log := log.G(ctx).WithFields(logrus.Fields{
"node.id": s.ClientTLSCreds.NodeID(),
"node.role": s.ClientTLSCreds.Role(),
})
// Our starting default will be 5 minutes
retry = 5 * time.Minute
// Since the expiration of the certificate is managed remotely we should update our
// retry timer on every iteration of this loop.
// Retrieve the time until the certificate expires.
expiresIn, err := readCertExpiration(paths.Node)
if err != nil {
// We failed to read the expiration, let's stick with the starting default
log.Errorf("failed to read the expiration of the TLS certificate in: %s", paths.Node.Cert)
updates <- CertificateUpdate{Err: fmt.Errorf("failed to read certificate expiration")}
} else {
// If we have an expired certificate, we let's stick with the starting default in
// the hope that this is a temporary clock skew.
if expiresIn.Minutes() < 0 {
log.WithError(err).Errorf("failed to create a new client TLS config")
updates <- CertificateUpdate{Err: fmt.Errorf("TLS certificate is expired")}
} else {
// Random retry time between 50% and 80% of the total time to expiration
retry = calculateRandomExpiry(expiresIn)
}
}
log.WithFields(logrus.Fields{
"time": time.Now().Add(retry),
}).Debugf("next certificate renewal scheduled")
select {
case <-time.After(retry):
log.Infof("renewing certificate")
case <-renew:
log.Infof("forced certificate renewal")
case <-ctx.Done():
log.Infof("shuting down certificate renewal routine")
return
}
// Let's request new certs. Renewals don't require a token.
rootCA := s.RootCA()
tlsKeyPair, err := rootCA.RequestAndSaveNewCertificates(ctx,
paths.Node,
"",
remotes,
s.ClientTLSCreds,
nil)
if err != nil {
log.WithError(err).Errorf("failed to renew the certificate")
updates <- CertificateUpdate{Err: err}
continue
}
clientTLSConfig, err := NewClientTLSConfig(tlsKeyPair, rootCA.Pool, CARole)
if err != nil {
log.WithError(err).Errorf("failed to create a new client config")
updates <- CertificateUpdate{Err: err}
}
serverTLSConfig, err := NewServerTLSConfig(tlsKeyPair, rootCA.Pool)
if err != nil {
log.WithError(err).Errorf("failed to create a new server config")
updates <- CertificateUpdate{Err: err}
}
err = s.ClientTLSCreds.LoadNewTLSConfig(clientTLSConfig)
if err != nil {
log.WithError(err).Errorf("failed to update the client credentials")
updates <- CertificateUpdate{Err: err}
}
// Update the external CA to use the new client TLS
// config using a copy without a serverName specified.
s.externalCA.UpdateTLSConfig(&tls.Config{
Certificates: clientTLSConfig.Certificates,
RootCAs: clientTLSConfig.RootCAs,
MinVersion: tls.VersionTLS12,
})
err = s.ServerTLSCreds.LoadNewTLSConfig(serverTLSConfig)
if err != nil {
log.WithError(err).Errorf("failed to update the server TLS credentials")
updates <- CertificateUpdate{Err: err}
}
updates <- CertificateUpdate{Role: s.ClientTLSCreds.Role()}
}
}()
return updates
}
// Assignments is a stream of assignments for a node. Each message contains
// either full list of tasks and secrets for the node, or an incremental update.
func (d *Dispatcher) Assignments(r *api.AssignmentsRequest, stream api.Dispatcher_AssignmentsServer) error {
nodeInfo, err := ca.RemoteNode(stream.Context())
if err != nil {
return err
}
nodeID := nodeInfo.NodeID
dctx, err := d.isRunningLocked()
if err != nil {
return err
}
fields := logrus.Fields{
"node.id": nodeID,
"node.session": r.SessionID,
"method": "(*Dispatcher).Assignments",
}
if nodeInfo.ForwardedBy != nil {
fields["forwarder.id"] = nodeInfo.ForwardedBy.NodeID
}
log := log.G(stream.Context()).WithFields(fields)
log.Debugf("")
if _, err = d.nodes.GetWithSession(nodeID, r.SessionID); err != nil {
return err
}
var (
sequence int64
appliesTo string
initial api.AssignmentsMessage
)
tasksMap := make(map[string]*api.Task)
tasksUsingSecret := make(map[string]map[string]struct{})
sendMessage := func(msg api.AssignmentsMessage, assignmentType api.AssignmentsMessage_Type) error {
sequence++
msg.AppliesTo = appliesTo
msg.ResultsIn = strconv.FormatInt(sequence, 10)
appliesTo = msg.ResultsIn
msg.Type = assignmentType
if err := stream.Send(&msg); err != nil {
return err
}
return nil
}
// returns a slice of new secrets to send down
addSecretsForTask := func(readTx store.ReadTx, t *api.Task) []*api.Secret {
container := t.Spec.GetContainer()
if container == nil {
return nil
}
var newSecrets []*api.Secret
for _, secretRef := range container.Secrets {
// Empty ID prefix will return all secrets. Bail if there is no SecretID
if secretRef.SecretID == "" {
log.Debugf("invalid secret reference")
continue
}
secretID := secretRef.SecretID
log := log.WithFields(logrus.Fields{
"secret.id": secretID,
"secret.name": secretRef.SecretName,
})
if len(tasksUsingSecret[secretID]) == 0 {
tasksUsingSecret[secretID] = make(map[string]struct{})
secrets, err := store.FindSecrets(readTx, store.ByIDPrefix(secretID))
if err != nil {
log.WithError(err).Errorf("error retrieving secret")
continue
}
if len(secrets) != 1 {
log.Debugf("secret not found")
continue
}
// If the secret was found and there was one result
// (there should never be more than one because of the
// uniqueness constraint), add this secret to our
// initial set that we send down.
newSecrets = append(newSecrets, secrets[0])
}
tasksUsingSecret[secretID][t.ID] = struct{}{}
}
return newSecrets
}
// TODO(aaronl): Also send node secrets that should be exposed to
// this node.
nodeTasks, cancel, err := store.ViewAndWatch(
d.store,
func(readTx store.ReadTx) error {
tasks, err := store.FindTasks(readTx, store.ByNodeID(nodeID))
if err != nil {
return err
}
for _, t := range tasks {
// We only care about tasks that are ASSIGNED or
// higher. If the state is below ASSIGNED, the
// task may not meet the constraints for this
// node, so we have to be careful about sending
// secrets associated with it.
if t.Status.State < api.TaskStateAssigned {
continue
}
tasksMap[t.ID] = t
taskChange := &api.AssignmentChange{
Assignment: &api.Assignment{
Item: &api.Assignment_Task{
Task: t,
},
},
Action: api.AssignmentChange_AssignmentActionUpdate,
}
initial.Changes = append(initial.Changes, taskChange)
// Only send secrets down if these tasks are in < RUNNING
if t.Status.State <= api.TaskStateRunning {
newSecrets := addSecretsForTask(readTx, t)
for _, secret := range newSecrets {
secretChange := &api.AssignmentChange{
Assignment: &api.Assignment{
Item: &api.Assignment_Secret{
Secret: secret,
},
},
Action: api.AssignmentChange_AssignmentActionUpdate,
}
initial.Changes = append(initial.Changes, secretChange)
}
}
}
return nil
},
state.EventUpdateTask{Task: &api.Task{NodeID: nodeID},
Checks: []state.TaskCheckFunc{state.TaskCheckNodeID}},
state.EventDeleteTask{Task: &api.Task{NodeID: nodeID},
Checks: []state.TaskCheckFunc{state.TaskCheckNodeID}},
state.EventUpdateSecret{},
state.EventDeleteSecret{},
)
if err != nil {
return err
}
defer cancel()
if err := sendMessage(initial, api.AssignmentsMessage_COMPLETE); err != nil {
return err
}
for {
// Check for session expiration
if _, err := d.nodes.GetWithSession(nodeID, r.SessionID); err != nil {
return err
}
// bursty events should be processed in batches and sent out together
var (
update api.AssignmentsMessage
modificationCnt int
batchingTimer *time.Timer
batchingTimeout <-chan time.Time
updateTasks = make(map[string]*api.Task)
updateSecrets = make(map[string]*api.Secret)
removeTasks = make(map[string]struct{})
removeSecrets = make(map[string]struct{})
)
oneModification := func() {
modificationCnt++
if batchingTimer != nil {
batchingTimer.Reset(batchingWaitTime)
} else {
batchingTimer = time.NewTimer(batchingWaitTime)
batchingTimeout = batchingTimer.C
}
}
// Release the secrets references from this task
releaseSecretsForTask := func(t *api.Task) bool {
var modified bool
container := t.Spec.GetContainer()
if container == nil {
return modified
}
for _, secretRef := range container.Secrets {
secretID := secretRef.SecretID
delete(tasksUsingSecret[secretID], t.ID)
if len(tasksUsingSecret[secretID]) == 0 {
// No tasks are using the secret anymore
delete(tasksUsingSecret, secretID)
removeSecrets[secretID] = struct{}{}
modified = true
}
}
return modified
}
// The batching loop waits for 50 ms after the most recent
// change, or until modificationBatchLimit is reached. The
// worst case latency is modificationBatchLimit * batchingWaitTime,
// which is 10 seconds.
batchingLoop:
for modificationCnt < modificationBatchLimit {
select {
case event := <-nodeTasks:
switch v := event.(type) {
// We don't monitor EventCreateTask because tasks are
// never created in the ASSIGNED state. First tasks are
// created by the orchestrator, then the scheduler moves
// them to ASSIGNED. If this ever changes, we will need
// to monitor task creations as well.
case state.EventUpdateTask:
// We only care about tasks that are ASSIGNED or
// higher.
if v.Task.Status.State < api.TaskStateAssigned {
continue
}
if oldTask, exists := tasksMap[v.Task.ID]; exists {
// States ASSIGNED and below are set by the orchestrator/scheduler,
// not the agent, so tasks in these states need to be sent to the
// agent even if nothing else has changed.
if equality.TasksEqualStable(oldTask, v.Task) && v.Task.Status.State > api.TaskStateAssigned {
// this update should not trigger a task change for the agent
tasksMap[v.Task.ID] = v.Task
// If this task got updated to a final state, let's release
// the secrets that are being used by the task
if v.Task.Status.State > api.TaskStateRunning {
// If releasing the secrets caused a secret to be
// removed from an agent, mark one modification
if releaseSecretsForTask(v.Task) {
oneModification()
}
}
continue
}
} else if v.Task.Status.State <= api.TaskStateRunning {
// If this task wasn't part of the assignment set before, and it's <= RUNNING
// add the secrets it references to the secrets assignment.
// Task states > RUNNING are worker reported only, are never created in
// a > RUNNING state.
var newSecrets []*api.Secret
d.store.View(func(readTx store.ReadTx) {
newSecrets = addSecretsForTask(readTx, v.Task)
})
for _, secret := range newSecrets {
updateSecrets[secret.ID] = secret
}
}
tasksMap[v.Task.ID] = v.Task
updateTasks[v.Task.ID] = v.Task
oneModification()
case state.EventDeleteTask:
if _, exists := tasksMap[v.Task.ID]; !exists {
continue
}
removeTasks[v.Task.ID] = struct{}{}
delete(tasksMap, v.Task.ID)
// Release the secrets being used by this task
// Ignoring the return here. We will always mark
// this as a modification, since a task is being
// removed.
releaseSecretsForTask(v.Task)
oneModification()
// TODO(aaronl): For node secrets, we'll need to handle
// EventCreateSecret.
case state.EventUpdateSecret:
if _, exists := tasksUsingSecret[v.Secret.ID]; !exists {
continue
}
log.Debugf("Secret %s (ID: %d) was updated though it was still referenced by one or more tasks",
v.Secret.Spec.Annotations.Name, v.Secret.ID)
case state.EventDeleteSecret:
if _, exists := tasksUsingSecret[v.Secret.ID]; !exists {
continue
}
log.Debugf("Secret %s (ID: %d) was deleted though it was still referenced by one or more tasks",
v.Secret.Spec.Annotations.Name, v.Secret.ID)
}
case <-batchingTimeout:
break batchingLoop
case <-stream.Context().Done():
return stream.Context().Err()
case <-dctx.Done():
return dctx.Err()
}
}
if batchingTimer != nil {
batchingTimer.Stop()
}
if modificationCnt > 0 {
for id, task := range updateTasks {
if _, ok := removeTasks[id]; !ok {
taskChange := &api.AssignmentChange{
Assignment: &api.Assignment{
Item: &api.Assignment_Task{
Task: task,
},
},
Action: api.AssignmentChange_AssignmentActionUpdate,
}
update.Changes = append(update.Changes, taskChange)
}
}
for id, secret := range updateSecrets {
// If, due to multiple updates, this secret is no longer in use,
// don't send it down.
if len(tasksUsingSecret[id]) == 0 {
// delete this secret for the secrets to be updated
// so that deleteSecrets knows the current list
delete(updateSecrets, id)
continue
}
secretChange := &api.AssignmentChange{
Assignment: &api.Assignment{
Item: &api.Assignment_Secret{
Secret: secret,
},
},
Action: api.AssignmentChange_AssignmentActionUpdate,
}
update.Changes = append(update.Changes, secretChange)
}
for id := range removeTasks {
taskChange := &api.AssignmentChange{
Assignment: &api.Assignment{
Item: &api.Assignment_Task{
Task: &api.Task{ID: id},
},
},
Action: api.AssignmentChange_AssignmentActionRemove,
}
update.Changes = append(update.Changes, taskChange)
}
for id := range removeSecrets {
// If this secret is also being sent on the updated set
// don't also add it to the removed set
if _, ok := updateSecrets[id]; ok {
continue
}
secretChange := &api.AssignmentChange{
Assignment: &api.Assignment{
Item: &api.Assignment_Secret{
Secret: &api.Secret{ID: id},
},
},
Action: api.AssignmentChange_AssignmentActionRemove,
}
update.Changes = append(update.Changes, secretChange)
}
if err := sendMessage(update, api.AssignmentsMessage_INCREMENTAL); err != nil {
return err
}
}
}
}
// RenewTLSConfig will continuously monitor for the necessity of renewing the local certificates, either by
// issuing them locally if key-material is available, or requesting them from a remote CA.
func RenewTLSConfig(ctx context.Context, s *SecurityConfig, connBroker *connectionbroker.Broker, renew <-chan struct{}) <-chan CertificateUpdate {
updates := make(chan CertificateUpdate)
go func() {
var retry time.Duration
expBackoff := events.NewExponentialBackoff(RenewTLSExponentialBackoff)
defer close(updates)
for {
ctx = log.WithModule(ctx, "tls")
log := log.G(ctx).WithFields(logrus.Fields{
"node.id": s.ClientTLSCreds.NodeID(),
"node.role": s.ClientTLSCreds.Role(),
})
// Our starting default will be 5 minutes
retry = 5 * time.Minute
// Since the expiration of the certificate is managed remotely we should update our
// retry timer on every iteration of this loop.
// Retrieve the current certificate expiration information.
validFrom, validUntil, err := readCertValidity(s.KeyReader())
if err != nil {
// We failed to read the expiration, let's stick with the starting default
log.Errorf("failed to read the expiration of the TLS certificate in: %s", s.KeyReader().Target())
select {
case updates <- CertificateUpdate{Err: errors.New("failed to read certificate expiration")}:
case <-ctx.Done():
log.Info("shutting down certificate renewal routine")
return
}
} else {
// If we have an expired certificate, try to renew immediately: the hope that this is a temporary clock skew, or
// we can issue our own TLS certs.
if validUntil.Before(time.Now()) {
log.Warn("the current TLS certificate is expired, so an attempt to renew it will be made immediately")
// retry immediately(ish) with exponential backoff
retry = expBackoff.Proceed(nil)
} else {
// Random retry time between 50% and 80% of the total time to expiration
retry = calculateRandomExpiry(validFrom, validUntil)
}
}
log.WithFields(logrus.Fields{
"time": time.Now().Add(retry),
}).Debugf("next certificate renewal scheduled for %v from now", retry)
select {
case <-time.After(retry):
log.Info("renewing certificate")
case <-renew:
log.Info("forced certificate renewal")
case <-ctx.Done():
log.Info("shutting down certificate renewal routine")
return
}
// ignore errors - it will just try again later
var certUpdate CertificateUpdate
if err := RenewTLSConfigNow(ctx, s, connBroker); err != nil {
certUpdate.Err = err
expBackoff.Failure(nil, nil)
} else {
certUpdate.Role = s.ClientTLSCreds.Role()
expBackoff = events.NewExponentialBackoff(RenewTLSExponentialBackoff)
}
select {
case updates <- certUpdate:
case <-ctx.Done():
log.Info("shutting down certificate renewal routine")
return
}
}
}()
return updates
}