// handleSessionExit processes the result from the session command, records it in persistent
// maner and determines if the Executor should exit
func (t *tether) handleSessionExit(session *SessionConfig) {
defer trace.End(trace.Begin("handling exit of session " + session.ID))
session.Lock()
defer session.Unlock()
if session.wait != nil {
session.wait.Wait()
}
// reader must be closed before calling wait, and this must interrupt the underlying Read
// or Wait will not return.
session.Reader.Close()
session.Cmd.Wait()
// close down the outputs
session.Outwriter.Close()
session.Errwriter.Close()
// close the signaling channel (it is nil for detached sessions) and set it to nil (for restart)
if session.ClearToLaunch != nil {
close(session.ClearToLaunch)
session.ClearToLaunch = nil
}
// Remove associated PID file
cmdname := path.Base(session.Cmd.Path)
_ = os.Remove(fmt.Sprintf("%s.pid", path.Join(PIDFileDir(), cmdname)))
// set the stop time
session.StopTime = time.Now().UTC().Unix()
// this returns an arbitrary closure for invocation after the session status update
f := t.ops.HandleSessionExit(t.config, session)
// FIXME: we cannot have this embedded knowledge of the extraconfig encoding pattern, but not
// currently sure how to expose it neatly via a utility function
extraconfig.EncodeWithPrefix(t.sink, session, fmt.Sprintf("guestinfo.vice..sessions|%s", session.ID))
if f != nil {
f()
}
}
// handleSessionExit processes the result from the session command, records it in persistent
// maner and determines if the Executor should exit
func (t *tether) handleSessionExit(session *SessionConfig) {
defer trace.End(trace.Begin("handling exit of session " + session.ID))
// close down the IO
session.Reader.Close()
// live.outwriter.Close()
// live.errwriter.Close()
// flush session log output
// record exit status
// FIXME: we cannot have this embedded knowledge of the extraconfig encoding pattern, but not
// currently sure how to expose it neatly via a utility function
extraconfig.EncodeWithPrefix(t.sink, session.ExitStatus, fmt.Sprintf("guestinfo..sessions|%s.status", session.ID))
log.Infof("%s exit code: %d", session.ID, session.ExitStatus)
if t.ops.HandleSessionExit(t.config, session) {
t.Stop()
}
}
// handleSessionExit processes the result from the session command, records it in persistent
// maner and determines if the Executor should exit
func (t *tether) handleSessionExit(session *SessionConfig) {
defer trace.End(trace.Begin("handling exit of session " + session.ID))
// stdio must be closed before calling wait or Wait hangs indefinitely
session.Reader.Close()
session.Cmd.Wait()
// close down the outputs
session.Outwriter.Close()
session.Errwriter.Close()
// Log a death record trimming records if need be
logs := session.Diagnostics.ExitLogs
logCount := len(logs)
if logCount >= MaxDeathRecords {
logs = logs[logCount-MaxDeathRecords+1:]
}
session.Diagnostics.ExitLogs = append(logs, executor.ExitLog{
Time: time.Now(),
ExitStatus: session.ExitStatus,
// We don't have any message for now
})
// this returns an arbitrary closure for invocation after the session status update
f := t.ops.HandleSessionExit(t.config, session)
log.Infof("%s exit code: %d", session.ID, session.ExitStatus)
// record exit status
// FIXME: we cannot have this embedded knowledge of the extraconfig encoding pattern, but not
// currently sure how to expose it neatly via a utility function
extraconfig.EncodeWithPrefix(t.sink, session, fmt.Sprintf("guestinfo..sessions|%s", session.ID))
if f != nil {
f()
}
}
// launch will launch the command defined in the session.
// This will return an error if the session fails to launch
func (t *tether) launch(session *SessionConfig) error {
defer trace.End(trace.Begin("launching session " + session.ID))
// encode the result whether success or error
defer func() {
extraconfig.EncodeWithPrefix(t.sink, session, fmt.Sprintf("guestinfo.vice..sessions|%s", session.ID))
}()
session.Lock()
defer session.Unlock()
// Special case here because UID/GID lookup need to be done
// on the appliance...
if len(session.User) > 0 {
session.Cmd.SysProcAttr = getUserSysProcAttr(session.User)
}
session.Cmd.Env = t.ops.ProcessEnv(session.Cmd.Env)
session.Cmd.Stdout = session.Outwriter
session.Cmd.Stderr = session.Errwriter
session.Cmd.Stdin = session.Reader
resolved, err := lookPath(session.Cmd.Path, session.Cmd.Env, session.Cmd.Dir)
if err != nil {
log.Errorf("Path lookup failed for %s: %s", session.Cmd.Path, err)
session.Started = err.Error()
return err
}
log.Debugf("Resolved %s to %s", session.Cmd.Path, resolved)
session.Cmd.Path = resolved
// block until we have a connection
if session.RunBlock && session.ClearToLaunch != nil {
log.Debugf("Waiting clear signal to launch %s", session.ID)
select {
case <-t.ctx.Done():
log.Warnf("Waiting to launch %s canceled, bailing out", session.ID)
return nil
case <-session.ClearToLaunch:
log.Debugf("Received the clear signal to launch %s", session.ID)
}
}
pid := 0
// Use the mutex to make creating a child and adding the child pid into the
// childPidTable appear atomic to the reaper function. Use a anonymous function
// so we can defer unlocking locally
// logging is done after the function to keep the locked time as low as possible
err = func() error {
t.config.pidMutex.Lock()
defer t.config.pidMutex.Unlock()
if !session.Tty {
err = session.Cmd.Start()
} else {
err = establishPty(session)
}
if err != nil {
return err
}
pid = session.Cmd.Process.Pid
t.config.pids[pid] = session
return nil
}()
if err != nil {
detail := fmt.Sprintf("failed to start container process: %s", err)
log.Error(detail)
// Set the Started key to the undecorated error message
session.Started = err.Error()
return errors.New(detail)
}
// Set the Started key to "true" - this indicates a successful launch
session.Started = "true"
// Write the PID to the associated PID file
cmdname := path.Base(session.Cmd.Path)
err = ioutil.WriteFile(fmt.Sprintf("%s.pid", path.Join(PIDFileDir(), cmdname)),
[]byte(fmt.Sprintf("%d", pid)),
0644)
if err != nil {
log.Errorf("Unable to write PID file for %s: %s", cmdname, err)
}
log.Debugf("Launched command with pid %d", pid)
return nil
}
func (t *tether) processSessions() error {
type results struct {
id string
path string
err error
}
// so that we can launch multiple sessions in parallel
var wg sync.WaitGroup
// to collect the errors back from them
resultsCh := make(chan results, len(t.config.Sessions))
// process the sessions and launch if needed
for id, session := range t.config.Sessions {
session.Lock()
log.Debugf("Processing config for session %s", id)
var proc = session.Cmd.Process
// check if session is alive and well
if proc != nil && proc.Signal(syscall.Signal(0)) == nil {
log.Debugf("Process for session %s is already running (pid: %d)", id, proc.Pid)
session.Unlock()
continue
}
// check if session has never been started or is configured for restart
if proc == nil || session.Restart {
if proc == nil {
log.Infof("Launching process for session %s", id)
} else {
session.Diagnostics.ResurrectionCount++
// FIXME: we cannot have this embedded knowledge of the extraconfig encoding pattern, but not
// currently sure how to expose it neatly via a utility function
extraconfig.EncodeWithPrefix(t.sink, session, fmt.Sprintf("guestinfo.vice..sessions|%s", id))
log.Warnf("Re-launching process for session %s (count: %d)", id, session.Diagnostics.ResurrectionCount)
session.Cmd = *restartableCmd(&session.Cmd)
}
session.Unlock()
wg.Add(1)
go func(session *SessionConfig) {
defer wg.Done()
resultsCh <- results{
id: session.ID,
path: session.Cmd.Path,
err: t.launch(session),
}
}(session)
}
}
wg.Wait()
// close the channel
close(resultsCh)
// iterate over the results
for result := range resultsCh {
if result.err != nil {
detail := fmt.Errorf("failed to launch %s for %s: %s", result.path, result.id, result.err)
return detail
}
}
return nil
}
// launch will launch the command defined in the session.
// This will return an error if the session fails to launch
func (t *tether) launch(session *SessionConfig) error {
defer trace.End(trace.Begin("launching session " + session.ID))
// encode the result whether success or error
defer func() {
extraconfig.EncodeWithPrefix(t.sink, session.Started, fmt.Sprintf("guestinfo..sessions|%s.started", session.ID))
}()
logwriter, err := t.ops.SessionLog(session)
if err != nil {
detail := fmt.Sprintf("failed to get log writer for session: %s", err)
log.Error(detail)
session.Started = detail
return errors.New(detail)
}
// we store these outside of the session.Cmd struct so that there's consistent
// handling between tty & non-tty paths
session.Outwriter = logwriter
session.Errwriter = logwriter
session.Reader = dio.MultiReader()
session.Cmd.Env = t.ops.ProcessEnv(session.Cmd.Env)
session.Cmd.Stdout = session.Outwriter
session.Cmd.Stderr = session.Errwriter
session.Cmd.Stdin = session.Reader
resolved, err := lookPath(session.Cmd.Path, session.Cmd.Env, session.Cmd.Dir)
if err != nil {
log.Errorf("Path lookup failed for %s: %s", session.Cmd.Path, err)
session.Started = err.Error()
return err
}
log.Debugf("Resolved %s to %s", session.Cmd.Path, resolved)
session.Cmd.Path = resolved
// Use the mutex to make creating a child and adding the child pid into the
// childPidTable appear atomic to the reaper function. Use a anonymous function
// so we can defer unlocking locally
// logging is done after the function to keep the locked time as low as possible
err = func() error {
t.config.pidMutex.Lock()
defer t.config.pidMutex.Unlock()
log.Infof("Launching command %#v", session.Cmd.Args)
if !session.Tty {
err = session.Cmd.Start()
} else {
err = establishPty(session)
}
if err != nil {
return err
}
// ChildReaper will use this channel to inform us the wait status of the child.
t.config.pids[session.Cmd.Process.Pid] = session
return nil
}()
if err != nil {
detail := fmt.Sprintf("failed to start container process: %s", err)
log.Error(detail)
// Set the Started key to the undecorated error message
session.Started = err.Error()
return errors.New(detail)
}
// Set the Started key to "true" - this indicates a successful launch
session.Started = "true"
log.Debugf("Launched command with pid %d", session.Cmd.Process.Pid)
return nil
}
func (t *tether) Start() error {
defer trace.End(trace.Begin("main tether loop"))
t.setup()
defer t.cleanup()
// initial entry, so seed this
t.reload <- true
for _ = range t.reload {
log.Info("Loading main configuration")
// load the config - this modifies the structure values in place
extraconfig.Decode(t.src, t.config)
logConfig(t.config)
short := t.config.ID
if len(short) > shortLen {
short = short[:shortLen]
}
if err := t.ops.SetHostname(short, t.config.Name); err != nil {
detail := fmt.Sprintf("failed to set hostname: %s", err)
log.Error(detail)
// we don't attempt to recover from this - it's a fundemental misconfiguration
// so just exit
return errors.New(detail)
}
// process the networks then publish any dynamic data
for _, v := range t.config.Networks {
if err := t.ops.Apply(v); err != nil {
detail := fmt.Sprintf("failed to apply network endpoint config: %s", err)
log.Error(detail)
return errors.New(detail)
}
}
extraconfig.Encode(t.sink, t.config)
//process the filesystem mounts - this is performed after networks to allow for network mounts
for k, v := range t.config.Mounts {
if v.Source.Scheme != "label" {
detail := fmt.Sprintf("unsupported volume mount type for %s: %s", k, v.Source.Scheme)
log.Error(detail)
return errors.New(detail)
}
// this could block indefinitely while waiting for a volume to present
t.ops.MountLabel(context.Background(), v.Source.Path, v.Path)
}
// process the sessions and launch if needed
for id, session := range t.config.Sessions {
log.Debugf("Processing config for session %s", session.ID)
var proc = session.Cmd.Process
// check if session is alive and well
if proc != nil && proc.Signal(syscall.Signal(0)) == nil {
log.Debugf("Process for session %s is already running (pid: %d)", session.ID, proc.Pid)
continue
}
// check if session has never been started or is configured for restart
if proc == nil || session.Restart {
if proc == nil {
log.Infof("Launching process for session %s", session.ID)
} else {
session.Diagnostics.ResurrectionCount++
// FIXME: we cannot have this embedded knowledge of the extraconfig encoding pattern, but not
// currently sure how to expose it neatly via a utility function
extraconfig.EncodeWithPrefix(t.sink, session, fmt.Sprintf("guestinfo..sessions|%s", session.ID))
log.Warnf("Re-launching process for session %s (count: %d)", session.ID, session.Diagnostics.ResurrectionCount)
session.Cmd = *restartableCmd(&session.Cmd)
}
err := t.launch(session)
if err != nil {
detail := fmt.Sprintf("failed to launch %s for %s: %s", session.Cmd.Path, id, err)
log.Error(detail)
// TODO: check if failure to launch this is fatal to everything in this containerVM
// for now failure to launch at all is terminal
return errors.New(detail)
}
continue
}
log.Warnf("Process for session %s has exited (%d) and is not configured for restart", session.ID, session.ExitStatus)
}
for name, ext := range t.extensions {
log.Info("Passing config to " + name)
err := ext.Reload(t.config)
if err != nil {
log.Errorf("Failed to cleanly reload config for extension %s: %s", name, err)
return err
}
}
}
return nil
}
