// runs the main kubelet loop, closing the kubeletFinished chan when the loop exits.
// never returns.
func (ms *MinionServer) Run(hks hyperkube.Interface, _ []string) error {
if ms.privateMountNS {
// only the Linux version will do anything
enterPrivateMountNamespace()
}
// create apiserver client
clientConfig, err := ms.KubeletExecutorServer.CreateAPIServerClientConfig()
if err != nil {
// required for k8sm since we need to send api.Binding information
// back to the apiserver
log.Fatalf("No API client: %v", err)
}
ms.clientConfig = clientConfig
// run subprocesses until ms.done is closed on return of this function
defer close(ms.done)
if ms.runProxy {
go runtime.Until(ms.launchProxyServer, 5*time.Second, ms.done)
}
go runtime.Until(ms.launchExecutorServer, 5*time.Second, ms.done)
// wait until minion exit is requested
// don't close ms.exit here to avoid panics of go routines writing an error to it
return <-ms.exit
}
func (m *MockScheduler) Run(done <-chan struct{}) {
_ = m.Called()
runtime.Until(func() {
time.Sleep(time.Second)
}, time.Second, done)
return
}
// perform one-time initialization actions upon the first registration event received from Mesos.
func (k *framework) onInitialRegistration(driver bindings.SchedulerDriver) {
defer close(k.registration)
if k.failoverTimeout > 0 {
refreshInterval := k.schedulerConfig.FrameworkIdRefreshInterval.Duration
if k.failoverTimeout < k.schedulerConfig.FrameworkIdRefreshInterval.Duration.Seconds() {
refreshInterval = time.Duration(math.Max(1, k.failoverTimeout/2)) * time.Second
}
// wait until we've written the framework ID at least once before proceeding
firstStore := make(chan struct{})
go runtime.Until(func() {
// only close firstStore once
select {
case <-firstStore:
default:
defer close(firstStore)
}
err := k.storeFrameworkId(context.TODO(), k.frameworkId.GetValue())
if err != nil {
log.Errorf("failed to store framework ID: %v", err)
if err == frameworkid.ErrMismatch {
// we detected a framework ID in storage that doesn't match what we're trying
// to save. this is a dangerous state:
// (1) perhaps we failed to initially recover the framework ID and so mesos
// issued us a new one. now that we're trying to save it there's a mismatch.
// (2) we've somehow bungled the framework ID and we're out of alignment with
// what mesos is expecting.
// (3) multiple schedulers were launched at the same time, and both have
// registered with mesos (because when they each checked, there was no ID in
// storage, so they asked for a new one). one of them has already written the
// ID to storage -- we lose.
log.Error("aborting due to framework ID mismatch")
driver.Abort()
}
}
}, refreshInterval, k.terminate)
// wait for the first store attempt of the framework ID
select {
case <-firstStore:
case <-k.terminate:
}
}
r1 := k.makeTaskRegistryReconciler()
r2 := k.makePodRegistryReconciler()
k.tasksReconciler = taskreconciler.New(k.asRegisteredMaster, taskreconciler.MakeComposite(k.terminate, r1, r2),
k.reconcileCooldown, k.schedulerConfig.ExplicitReconciliationAbortTimeout.Duration, k.terminate)
go k.tasksReconciler.Run(driver, k.terminate)
if k.reconcileInterval > 0 {
ri := time.Duration(k.reconcileInterval) * time.Second
time.AfterFunc(k.schedulerConfig.InitialImplicitReconciliationDelay.Duration, func() { runtime.Until(k.tasksReconciler.RequestImplicit, ri, k.terminate) })
log.Infof("will perform implicit task reconciliation at interval: %v after %v", ri, k.schedulerConfig.InitialImplicitReconciliationDelay.Duration)
}
k.installDebugHandlers(k.mux)
}
// perform one-time initialization actions upon the first registration event received from Mesos.
func (k *framework) onInitialRegistration(driver bindings.SchedulerDriver) {
defer close(k.registration)
if k.failoverTimeout > 0 {
refreshInterval := k.schedulerConfig.FrameworkIdRefreshInterval.Duration
if k.failoverTimeout < k.schedulerConfig.FrameworkIdRefreshInterval.Duration.Seconds() {
refreshInterval = time.Duration(math.Max(1, k.failoverTimeout/2)) * time.Second
}
go runtime.Until(func() {
k.storeFrameworkId(k.frameworkId.GetValue())
}, refreshInterval, k.terminate)
}
r1 := k.makeTaskRegistryReconciler()
r2 := k.makePodRegistryReconciler()
k.tasksReconciler = taskreconciler.New(k.asRegisteredMaster, taskreconciler.MakeComposite(k.terminate, r1, r2),
k.reconcileCooldown, k.schedulerConfig.ExplicitReconciliationAbortTimeout.Duration, k.terminate)
go k.tasksReconciler.Run(driver, k.terminate)
if k.reconcileInterval > 0 {
ri := time.Duration(k.reconcileInterval) * time.Second
time.AfterFunc(k.schedulerConfig.InitialImplicitReconciliationDelay.Duration, func() { runtime.Until(k.tasksReconciler.RequestImplicit, ri, k.terminate) })
log.Infof("will perform implicit task reconciliation at interval: %v after %v", ri, k.schedulerConfig.InitialImplicitReconciliationDelay.Duration)
}
k.installDebugHandlers(k.mux)
}
func (self *procImpl) begin() runtime.Signal {
if !self.state.transition(stateNew, stateRunning) {
panic(fmt.Errorf("failed to transition from New to Idle state"))
}
defer log.V(2).Infof("started process %d", self.pid)
var entered runtime.Latch
// execute actions on the backlog chan
return runtime.After(func() {
runtime.Until(func() {
if entered.Acquire() {
close(self.running)
self.wg.Add(1)
}
for action := range self.backlog {
select {
case <-self.terminate:
return
default:
// signal to indicate there's room in the backlog now
self.changed.Broadcast()
// rely on Until to handle action panics
action()
}
}
}, self.actionHandlerCrashDelay, self.terminate)
}).Then(func() {
log.V(2).Infof("finished processing action backlog for process %d", self.pid)
if !entered.Acquire() {
self.wg.Done()
}
})
}
// Notify runs Elect() on m, and calls Start()/Stop() on s when the
// elected master starts/stops matching 'id'. Never returns.
func Notify(m MasterElector, path, id string, s Service, abort <-chan struct{}) {
n := ¬ifier{id: Master(id), service: s, masters: make(chan Master, 1)}
finished := runtime.After(func() {
runtime.Until(func() {
for {
w := m.Elect(path, id)
for {
select {
case <-abort:
return
case event, open := <-w.ResultChan():
if !open {
break
}
if event.Type != watch.Modified {
continue
}
electedMaster, ok := event.Object.(Master)
if !ok {
glog.Errorf("Unexpected object from election channel: %v", event.Object)
break
}
sendElected:
for {
select {
case <-abort:
return
case n.masters <- electedMaster:
break sendElected
default: // ring full, discard old value and add the new
select {
case <-abort:
return
case <-n.masters:
default: // ring was cleared for us?!
}
}
}
}
}
}
}, 0, abort)
})
runtime.Until(func() { n.serviceLoop(finished) }, 0, abort)
}
func (s *SchedulerServer) Run(hks hyperkube.Interface, _ []string) error {
if n := len(s.frameworkRoles); n == 0 || n > 2 || (n == 2 && s.frameworkRoles[0] != "*" && s.frameworkRoles[1] != "*") {
log.Fatalf(`only one custom role allowed in addition to "*"`)
}
fwSet := sets.NewString(s.frameworkRoles...)
podSet := sets.NewString(s.defaultPodRoles...)
if !fwSet.IsSuperset(podSet) {
log.Fatalf("all default pod roles %q must be included in framework roles %q", s.defaultPodRoles, s.frameworkRoles)
}
// get scheduler low-level config
sc := schedcfg.CreateDefaultConfig()
if s.schedulerConfigFileName != "" {
f, err := os.Open(s.schedulerConfigFileName)
if err != nil {
log.Fatalf("Cannot open scheduler config file: %v", err)
}
defer f.Close()
err = sc.Read(bufio.NewReader(f))
if err != nil {
log.Fatalf("Invalid scheduler config file: %v", err)
}
}
schedulerProcess, driverFactory, etcdClient, eid := s.bootstrap(hks, sc)
if s.enableProfiling {
profile.InstallHandler(s.mux)
}
go runtime.Until(func() {
log.V(1).Info("Starting HTTP interface")
log.Error(http.ListenAndServe(net.JoinHostPort(s.address.String(), strconv.Itoa(s.port)), s.mux))
}, sc.HttpBindInterval.Duration, schedulerProcess.Terminal())
if s.ha {
validation := ha.ValidationFunc(validateLeadershipTransition)
srv := ha.NewCandidate(schedulerProcess, driverFactory, validation)
path := meta.ElectionPath(s.frameworkName)
uuid := eid.GetValue() + ":" + uuid.New() // unique for each scheduler instance
log.Infof("registering for election at %v with id %v", path, uuid)
go election.Notify(
election.NewEtcdMasterElector(etcdClient),
path,
uuid,
srv,
nil)
} else {
log.Infoln("self-electing in non-HA mode")
schedulerProcess.Elect(driverFactory)
}
return s.awaitFailover(schedulerProcess, func() error { return s.failover(s.getDriver(), hks) })
}
func (s *offerStorage) Init(done <-chan struct{}) {
// zero delay, reap offers as soon as they expire
go runtime.Until(s.ageOffers, 0, done)
// cached offer ids for the purposes of listener notification
idCache := &stringsCache{
refill: func() sets.String {
result := sets.NewString()
for _, v := range s.offers.List() {
if offer, ok := v.(Perishable); ok {
result.Insert(offer.Id())
}
}
return result
},
ttl: offerIdCacheTTL,
}
go runtime.Until(func() { s.notifyListeners(idCache.Strings) }, notifyListenersDelay, done)
}
// Notify runs Elect() on m, and calls Start()/Stop() on s when the
// elected master starts/stops matching 'id'. Never returns.
func Notify(m MasterElector, path, id string, s Service, abort <-chan struct{}) {
n := ¬ifier{id: Master(id), service: s}
n.changed = make(chan struct{})
finished := runtime.After(func() {
runtime.Until(func() {
for {
w := m.Elect(path, id)
for {
select {
case <-abort:
return
case event, open := <-w.ResultChan():
if !open {
break
}
if event.Type != watch.Modified {
continue
}
electedMaster, ok := event.Object.(Master)
if !ok {
glog.Errorf("Unexpected object from election channel: %v", event.Object)
break
}
n.lock.Lock()
n.desired = electedMaster
n.lock.Unlock()
// notify serviceLoop, but don't block. If a change
// is queued already it will see the new n.desired.
select {
case n.changed <- struct{}{}:
}
}
}
}
}, 0, abort)
})
runtime.Until(func() { n.serviceLoop(finished) }, 0, abort)
}
// perform one-time initialization actions upon the first registration event received from Mesos.
func (k *KubernetesScheduler) onInitialRegistration(driver bindings.SchedulerDriver) {
defer close(k.registration)
if k.failoverTimeout > 0 {
refreshInterval := k.schedcfg.FrameworkIdRefreshInterval.Duration
if k.failoverTimeout < k.schedcfg.FrameworkIdRefreshInterval.Duration.Seconds() {
refreshInterval = time.Duration(math.Max(1, k.failoverTimeout/2)) * time.Second
}
go runtime.Until(k.storeFrameworkId, refreshInterval, k.terminate)
}
r1 := k.makeTaskRegistryReconciler()
r2 := k.makePodRegistryReconciler()
k.reconciler = newReconciler(k.asRegisteredMaster, k.makeCompositeReconciler(r1, r2),
k.reconcileCooldown, k.schedcfg.ExplicitReconciliationAbortTimeout.Duration, k.terminate)
go k.reconciler.Run(driver)
if k.reconcileInterval > 0 {
ri := time.Duration(k.reconcileInterval) * time.Second
time.AfterFunc(k.schedcfg.InitialImplicitReconciliationDelay.Duration, func() { runtime.Until(k.reconciler.RequestImplicit, ri, k.terminate) })
log.Infof("will perform implicit task reconciliation at interval: %v after %v", ri, k.schedcfg.InitialImplicitReconciliationDelay.Duration)
}
}
// currently monitors for "pod deleted" events, upon which handle()
// is invoked.
func (k *deleter) Run(updates <-chan queue.Entry, done <-chan struct{}) {
go runtime.Until(func() {
for {
entry := <-updates
pod := entry.Value().(*Pod)
if entry.Is(queue.DELETE_EVENT) {
if err := k.deleteOne(pod); err != nil {
log.Error(err)
}
} else if !entry.Is(queue.POP_EVENT) {
k.qr.updatesAvailable()
}
}
}, 1*time.Second, done)
}
// spawns a go-routine to watch for unscheduled pods and queue them up
// for scheduling. returns immediately.
func (q *queuer) Run(done <-chan struct{}) {
go runtime.Until(func() {
log.Info("Watching for newly created pods")
q.lock.Lock()
defer q.lock.Unlock()
for {
// limit blocking here for short intervals so that scheduling
// may proceed even if there have been no recent pod changes
p := q.podUpdates.Await(enqueuePopTimeout)
if p == nil {
signalled := runtime.After(q.deltaCond.Wait)
// we've yielded the lock
select {
case <-time.After(enqueueWaitTimeout):
q.deltaCond.Broadcast() // abort Wait()
<-signalled // wait for lock re-acquisition
log.V(4).Infoln("timed out waiting for a pod update")
case <-signalled:
// we've acquired the lock and there may be
// changes for us to process now
}
continue
}
pod := p.(*Pod)
if recoverAssignedSlave(pod.Pod) != "" {
log.V(3).Infof("dequeuing assigned pod for scheduling: %v", pod.Pod.Name)
q.dequeue(pod.GetUID())
} else if pod.InGracefulTermination() {
// pods which are pre-scheduled (i.e. NodeName is set) may be gracefully deleted,
// even though they are not running yet.
log.V(3).Infof("dequeuing graceful deleted pre-scheduled pod for scheduling: %v", pod.Pod.Name)
q.dequeue(pod.GetUID())
} else {
// use ReplaceExisting because we are always pushing the latest state
now := time.Now()
pod.deadline = &now
if q.podQueue.Offer(pod, queue.ReplaceExisting) {
q.unscheduledCond.Broadcast()
log.V(3).Infof("queued pod for scheduling: %v", pod.Pod.Name)
} else {
log.Warningf("failed to queue pod for scheduling: %v", pod.Pod.Name)
}
}
}
}, 1*time.Second, done)
}
func (r *clientRegistrator) Run(terminate <-chan struct{}) error {
loop := func() {
RegistrationLoop:
for {
obj := r.queue.Pop(terminate)
log.V(3).Infof("registration event observed")
if obj == nil {
break RegistrationLoop
}
select {
case <-terminate:
break RegistrationLoop
default:
}
rg := obj.(*registration)
n, needsUpdate := r.updateNecessary(rg.hostName, rg.labels)
if !needsUpdate {
log.V(2).Infof("no update needed, skipping for %s: %v", rg.hostName, rg.labels)
continue
}
if n == nil {
log.V(2).Infof("creating node %s with labels %v", rg.hostName, rg.labels)
_, err := CreateOrUpdate(r.client, rg.hostName, rg.labels, nil)
if err != nil {
log.Errorf("error creating the node %s: %v", rg.hostName, rg.labels)
}
} else {
log.V(2).Infof("updating node %s with labels %v", rg.hostName, rg.labels)
_, err := Update(r.client, rg.hostName, rg.labels, nil)
if err != nil && errors.IsNotFound(err) {
// last chance when our store was out of date
_, err = Create(r.client, rg.hostName, rg.labels, nil)
}
if err != nil {
log.Errorf("error updating the node %s: %v", rg.hostName, rg.labels)
}
}
}
}
go runtime.Until(loop, time.Second, terminate)
return nil
}
func (u *StatusUpdater) Run(terminate <-chan struct{}) error {
nodeStore := cache.NewStore(cache.MetaNamespaceKeyFunc)
nodeLW := cache.NewListWatchFromClient(u.client, "nodes", api.NamespaceAll, fields.Everything())
cache.NewReflector(nodeLW, &api.Node{}, nodeStore, u.relistPeriod).Run()
monitor := func() {
// build up a slave set of nodes without kubelet
slavesWithoutKubeletList, err := mesos.CloudProvider.ListWithoutKubelet()
if err != nil {
log.Errorf("Error while updating slave nodes: %v", err)
return
}
slavesWithoutKubelet := make(map[string]struct{}, len(slavesWithoutKubeletList))
for _, s := range slavesWithoutKubeletList {
slavesWithoutKubelet[s] = struct{}{}
}
// update status for nodes which do not have a kubelet running and
// which are still existing as slave. This status update must be done
// before the node controller counts down the NodeMonitorGracePeriod
obj, err := nodeLW.List()
if err != nil {
log.Errorf("Error listing the nodes for status updates: %v", err)
}
nl, _ := obj.(*api.NodeList)
nodes := nl.Items
for i := range nodes {
if _, ok := slavesWithoutKubelet[nodes[i].Spec.ExternalID]; !ok {
// let the kubelet do its job updating the status, or the
// node controller will remove this node if the node does not even
// exist anymore
continue
}
err := u.updateStatus(&nodes[i])
if err != nil {
log.Errorf("Error updating node status: %v", err)
}
}
}
go runtime.Until(monitor, u.heartBeatPeriod, terminate)
return nil
}
func (u *StatusUpdater) Run(terminate <-chan struct{}) error {
nodeStore := cache.NewStore(cache.MetaNamespaceKeyFunc)
nodeLW := cache.NewListWatchFromClient(u.client.CoreClient, "nodes", api.NamespaceAll, fields.Everything())
cache.NewReflector(nodeLW, &api.Node{}, nodeStore, u.relistPeriod).Run()
monitor := func() {
// build up a set of listed slave nodes without a kubelet
slaves, err := mesos.CloudProvider.ListWithoutKubelet()
if err != nil {
log.Errorf("Error listing slaves without kubelet: %v", err)
return
}
slavesWithoutKubelet := sets.NewString(slaves...)
// update status for nodes which do not have a kubelet running and
// which are still existing as slave. This status update must be done
// before the node controller counts down the NodeMonitorGracePeriod
nodes := nodeStore.List()
for _, n := range nodes {
node := n.(*api.Node)
if !slavesWithoutKubelet.Has(node.Spec.ExternalID) {
// let the kubelet do its job updating the status, or the
// node controller will remove this node if the node does not even
// exist anymore
continue
}
err := u.updateStatus(node)
if err != nil {
log.Errorf("Error updating node status: %v", err)
}
}
}
go runtime.Until(monitor, u.heartBeatPeriod, terminate)
return nil
}
func (s *SchedulerServer) Run(hks hyperkube.Interface, _ []string) error {
// get scheduler low-level config
sc := schedcfg.CreateDefaultConfig()
if s.SchedulerConfigFileName != "" {
f, err := os.Open(s.SchedulerConfigFileName)
if err != nil {
log.Fatalf("Cannot open scheduler config file: %v", err)
}
err = sc.Read(bufio.NewReader(f))
if err != nil {
log.Fatalf("Invalid scheduler config file: %v", err)
}
}
schedulerProcess, driverFactory, etcdClient, eid := s.bootstrap(hks, sc)
if s.EnableProfiling {
profile.InstallHandler(s.mux)
}
go runtime.Until(func() {
log.V(1).Info("Starting HTTP interface")
log.Error(http.ListenAndServe(net.JoinHostPort(s.Address.String(), strconv.Itoa(s.Port)), s.mux))
}, sc.HttpBindInterval.Duration, schedulerProcess.Terminal())
if s.HA {
validation := ha.ValidationFunc(validateLeadershipTransition)
srv := ha.NewCandidate(schedulerProcess, driverFactory, validation)
path := fmt.Sprintf(meta.DefaultElectionFormat, s.FrameworkName)
sid := uid.New(eid.Group(), "").String()
log.Infof("registering for election at %v with id %v", path, sid)
go election.Notify(election.NewEtcdMasterElector(etcdClient), path, sid, srv, nil)
} else {
log.Infoln("self-electing in non-HA mode")
schedulerProcess.Elect(driverFactory)
}
return s.awaitFailover(schedulerProcess, func() error { return s.failover(s.getDriver(), hks) })
}
func (s *schedulingPlugin) Run(done <-chan struct{}) {
defer close(s.starting)
go runtime.Until(s.scheduleOne, pluginRecoveryDelay, done)
}
func (s *controller) Run(done <-chan struct{}) {
defer close(s.started)
go runtime.Until(s.scheduleOne, recoveryDelay, done)
}
