// recoverablePullError wraps the error gotten when trying to pull and image if
// the error is recoverable.
func (d *DockerDriver) recoverablePullError(err error, image string) error {
recoverable := true
if imageNotFoundMatcher.MatchString(err.Error()) {
recoverable = false
}
return dstructs.NewRecoverableError(fmt.Errorf("Failed to pull `%s`: %s", image, err), recoverable)
}
func TestClient_RestartTracker_StartError_Recoverable_Fail(t *testing.T) {
t.Parallel()
p := testPolicy(true, structs.RestartPolicyModeFail)
rt := newRestartTracker(p, structs.JobTypeSystem)
recErr := cstructs.NewRecoverableError(fmt.Errorf("foo"), true)
for i := 0; i < p.Attempts; i++ {
state, when := rt.SetStartError(recErr).GetState()
if state != structs.TaskRestarting {
t.Fatalf("NextRestart() returned %v, want %v", state, structs.TaskRestarting)
}
if !withinJitter(p.Delay, when) {
t.Fatalf("NextRestart() returned %v; want %v+jitter", when, p.Delay)
}
}
// Next restart should cause fail
if state, _ := rt.SetStartError(recErr).GetState(); state != structs.TaskNotRestarting {
t.Fatalf("NextRestart() returned %v; want %v", state, structs.TaskNotRestarting)
}
}
func (r *TaskRunner) run() {
// Predeclare things so we an jump to the RESTART
var handleEmpty bool
for {
// Download the task's artifacts
if !r.artifactsDownloaded && len(r.task.Artifacts) > 0 {
r.setState(structs.TaskStatePending, structs.NewTaskEvent(structs.TaskDownloadingArtifacts))
taskDir, ok := r.ctx.AllocDir.TaskDirs[r.task.Name]
if !ok {
err := fmt.Errorf("task directory couldn't be found")
r.setState(structs.TaskStateDead, structs.NewTaskEvent(structs.TaskDriverFailure).SetDriverError(err))
r.logger.Printf("[ERR] client: task directory for alloc %q task %q couldn't be found", r.alloc.ID, r.task.Name)
r.restartTracker.SetStartError(err)
goto RESTART
}
for _, artifact := range r.task.Artifacts {
if err := getter.GetArtifact(r.taskEnv, artifact, taskDir, r.logger); err != nil {
r.setState(structs.TaskStateDead,
structs.NewTaskEvent(structs.TaskArtifactDownloadFailed).SetDownloadError(err))
r.restartTracker.SetStartError(cstructs.NewRecoverableError(err, true))
goto RESTART
}
}
r.artifactsDownloaded = true
}
// Start the task if not yet started or it is being forced. This logic
// is necessary because in the case of a restore the handle already
// exists.
r.handleLock.Lock()
handleEmpty = r.handle == nil
r.handleLock.Unlock()
if handleEmpty {
startErr := r.startTask()
r.restartTracker.SetStartError(startErr)
if startErr != nil {
r.setState(structs.TaskStateDead, structs.NewTaskEvent(structs.TaskDriverFailure).SetDriverError(startErr))
goto RESTART
}
}
// Mark the task as started
r.setState(structs.TaskStateRunning, structs.NewTaskEvent(structs.TaskStarted))
// Wait for updates
WAIT:
for {
select {
case waitRes := <-r.handle.WaitCh():
if waitRes == nil {
panic("nil wait")
}
// Log whether the task was successful or not.
r.restartTracker.SetWaitResult(waitRes)
r.setState(structs.TaskStateDead, r.waitErrorToEvent(waitRes))
if !waitRes.Successful() {
r.logger.Printf("[INFO] client: task %q for alloc %q failed: %v", r.task.Name, r.alloc.ID, waitRes)
} else {
r.logger.Printf("[INFO] client: task %q for alloc %q completed successfully", r.task.Name, r.alloc.ID)
}
break WAIT
case update := <-r.updateCh:
if err := r.handleUpdate(update); err != nil {
r.logger.Printf("[ERR] client: update to task %q failed: %v", r.task.Name, err)
}
case <-r.destroyCh:
// Kill the task using an exponential backoff in-case of failures.
destroySuccess, err := r.handleDestroy()
if !destroySuccess {
// We couldn't successfully destroy the resource created.
r.logger.Printf("[ERR] client: failed to kill task %q. Resources may have been leaked: %v", r.task.Name, err)
}
// Store that the task has been destroyed and any associated error.
r.setState(structs.TaskStateDead, structs.NewTaskEvent(structs.TaskKilled).SetKillError(err))
return
}
}
RESTART:
state, when := r.restartTracker.GetState()
r.restartTracker.SetStartError(nil).SetWaitResult(nil)
reason := r.restartTracker.GetReason()
switch state {
case structs.TaskNotRestarting, structs.TaskTerminated:
r.logger.Printf("[INFO] client: Not restarting task: %v for alloc: %v ", r.task.Name, r.alloc.ID)
if state == structs.TaskNotRestarting {
r.setState(structs.TaskStateDead,
structs.NewTaskEvent(structs.TaskNotRestarting).
SetRestartReason(reason))
}
return
case structs.TaskRestarting:
r.logger.Printf("[INFO] client: Restarting task %q for alloc %q in %v", r.task.Name, r.alloc.ID, when)
r.setState(structs.TaskStatePending,
structs.NewTaskEvent(structs.TaskRestarting).
SetRestartDelay(when).
SetRestartReason(reason))
default:
r.logger.Printf("[ERR] client: restart tracker returned unknown state: %q", state)
return
}
// Sleep but watch for destroy events.
select {
case <-time.After(when):
case <-r.destroyCh:
}
// Destroyed while we were waiting to restart, so abort.
r.destroyLock.Lock()
destroyed := r.destroy
r.destroyLock.Unlock()
if destroyed {
r.logger.Printf("[DEBUG] client: Not restarting task: %v because it's destroyed by user", r.task.Name)
r.setState(structs.TaskStateDead, structs.NewTaskEvent(structs.TaskKilled))
return
}
// Clear the handle so a new driver will be created.
r.handleLock.Lock()
r.handle = nil
r.handleLock.Unlock()
}
}