func (s *resolverSuite) TestRunHookStopRetryTimer(c *gc.C) {
s.reportHookError = func(hook.Info) error { return nil }
localState := resolver.LocalState{
CharmModifiedVersion: s.charmModifiedVersion,
CharmURL: s.charmURL,
State: operation.State{
Kind: operation.RunHook,
Step: operation.Pending,
Installed: true,
Started: true,
Hook: &hook.Info{
Kind: hooks.ConfigChanged,
},
},
}
_, err := s.resolver.NextOp(localState, s.remoteState, s.opFactory)
c.Assert(err, gc.Equals, resolver.ErrNoOperation)
s.stub.CheckCallNames(c, "StartRetryHookTimer")
localState.Kind = operation.Continue
_, err = s.resolver.NextOp(localState, s.remoteState, s.opFactory)
c.Assert(err, gc.Equals, resolver.ErrNoOperation)
s.stub.CheckCallNames(c, "StartRetryHookTimer", "StopRetryHookTimer")
}
func (s *resolverSuite) TestHookErrorStartRetryTimerAgain(c *gc.C) {
s.reportHookError = func(hook.Info) error { return nil }
localState := resolver.LocalState{
CharmModifiedVersion: s.charmModifiedVersion,
CharmURL: s.charmURL,
State: operation.State{
Kind: operation.RunHook,
Step: operation.Pending,
Installed: true,
Started: true,
Hook: &hook.Info{
Kind: hooks.ConfigChanged,
},
},
}
_, err := s.resolver.NextOp(localState, s.remoteState, s.opFactory)
c.Assert(err, gc.Equals, resolver.ErrNoOperation)
s.stub.CheckCallNames(c, "StartRetryHookTimer")
s.remoteState.RetryHookVersion = 1
op, err := s.resolver.NextOp(localState, s.remoteState, s.opFactory)
c.Assert(err, jc.ErrorIsNil)
c.Assert(op.String(), gc.Equals, "run config-changed hook")
s.stub.CheckCallNames(c, "StartRetryHookTimer") // no change
localState.RetryHookVersion = 1
_, err = s.resolver.NextOp(localState, s.remoteState, s.opFactory)
c.Assert(err, gc.Equals, resolver.ErrNoOperation)
s.stub.CheckCallNames(c, "StartRetryHookTimer", "StartRetryHookTimer")
}
func (u *Uniter) loop(unitTag names.UnitTag) (err error) {
if err := u.init(unitTag); err != nil {
if err == worker.ErrTerminateAgent {
return err
}
return fmt.Errorf("failed to initialize uniter for %q: %v", unitTag, err)
}
logger.Infof("unit %q started", u.unit)
// Install is a special case, as it must run before there
// is any remote state, and before the remote state watcher
// is started.
var charmURL *corecharm.URL
var charmModifiedVersion int
opState := u.operationExecutor.State()
if opState.Kind == operation.Install {
logger.Infof("resuming charm install")
op, err := u.operationFactory.NewInstall(opState.CharmURL)
if err != nil {
return errors.Trace(err)
}
if err := u.operationExecutor.Run(op); err != nil {
return errors.Trace(err)
}
charmURL = opState.CharmURL
} else {
curl, err := u.unit.CharmURL()
if err != nil {
return errors.Trace(err)
}
charmURL = curl
svc, err := u.unit.Service()
if err != nil {
return errors.Trace(err)
}
charmModifiedVersion, err = svc.CharmModifiedVersion()
if err != nil {
return errors.Trace(err)
}
}
var (
watcher *remotestate.RemoteStateWatcher
watcherMu sync.Mutex
)
logger.Infof("hooks are retried %v", u.hookRetryStrategy.ShouldRetry)
retryHookChan := make(chan struct{}, 1)
retryHookTimer := utils.NewBackoffTimer(utils.BackoffTimerConfig{
Min: u.hookRetryStrategy.MinRetryTime,
Max: u.hookRetryStrategy.MaxRetryTime,
Jitter: u.hookRetryStrategy.JitterRetryTime,
Factor: u.hookRetryStrategy.RetryTimeFactor,
Func: func() {
// Don't try to send on the channel if it's already full
// This can happen if the timer fires off before the event is consumed
// by the resolver loop
select {
case retryHookChan <- struct{}{}:
default:
}
},
Clock: u.clock,
})
defer func() {
// Whenever we exit the uniter we want to stop a potentially
// running timer so it doesn't trigger for nothing.
retryHookTimer.Reset()
}()
restartWatcher := func() error {
watcherMu.Lock()
defer watcherMu.Unlock()
if watcher != nil {
// watcher added to catacomb, will kill uniter if there's an error.
worker.Stop(watcher)
}
var err error
watcher, err = remotestate.NewWatcher(
remotestate.WatcherConfig{
State: remotestate.NewAPIState(u.st),
LeadershipTracker: u.leadershipTracker,
UnitTag: unitTag,
UpdateStatusChannel: u.updateStatusAt,
CommandChannel: u.commandChannel,
RetryHookChannel: retryHookChan,
})
if err != nil {
return errors.Trace(err)
}
if err := u.catacomb.Add(watcher); err != nil {
return errors.Trace(err)
}
return nil
}
onIdle := func() error {
opState := u.operationExecutor.State()
if opState.Kind != operation.Continue {
// We should only set idle status if we're in
// the "Continue" state, which indicates that
// there is nothing to do and we're not in an
// error state.
return nil
}
return setAgentStatus(u, status.StatusIdle, "", nil)
}
clearResolved := func() error {
if err := u.unit.ClearResolved(); err != nil {
return errors.Trace(err)
}
watcher.ClearResolvedMode()
return nil
}
for {
if err = restartWatcher(); err != nil {
err = errors.Annotate(err, "(re)starting watcher")
break
}
uniterResolver := NewUniterResolver(ResolverConfig{
ClearResolved: clearResolved,
ReportHookError: u.reportHookError,
FixDeployer: u.deployer.Fix,
ShouldRetryHooks: u.hookRetryStrategy.ShouldRetry,
StartRetryHookTimer: retryHookTimer.Start,
StopRetryHookTimer: retryHookTimer.Reset,
Actions: actions.NewResolver(),
Leadership: uniterleadership.NewResolver(),
Relations: relation.NewRelationsResolver(u.relations),
Storage: storage.NewResolver(u.storage),
Commands: runcommands.NewCommandsResolver(
u.commands, watcher.CommandCompleted,
),
})
// We should not do anything until there has been a change
// to the remote state. The watcher will trigger at least
// once initially.
select {
case <-u.catacomb.Dying():
return u.catacomb.ErrDying()
case <-watcher.RemoteStateChanged():
}
localState := resolver.LocalState{
CharmURL: charmURL,
CharmModifiedVersion: charmModifiedVersion,
}
for err == nil {
err = resolver.Loop(resolver.LoopConfig{
Resolver: uniterResolver,
Watcher: watcher,
Executor: u.operationExecutor,
Factory: u.operationFactory,
Abort: u.catacomb.Dying(),
OnIdle: onIdle,
CharmDirGuard: u.charmDirGuard,
}, &localState)
switch cause := errors.Cause(err); cause {
case nil:
// Loop back around.
case resolver.ErrLoopAborted:
err = u.catacomb.ErrDying()
case operation.ErrNeedsReboot:
err = worker.ErrRebootMachine
case operation.ErrHookFailed:
// Loop back around. The resolver can tell that it is in
// an error state by inspecting the operation state.
err = nil
case resolver.ErrTerminate:
err = u.terminate()
case resolver.ErrRestart:
// make sure we update the two values used above in
// creating LocalState.
charmURL = localState.CharmURL
charmModifiedVersion = localState.CharmModifiedVersion
// leave err assigned, causing loop to break
default:
// We need to set conflicted from here, because error
// handling is outside of the resolver's control.
if operation.IsDeployConflictError(cause) {
localState.Conflicted = true
err = setAgentStatus(u, status.StatusError, "upgrade failed", nil)
} else {
reportAgentError(u, "resolver loop error", err)
}
}
}
if errors.Cause(err) != resolver.ErrRestart {
break
}
}
logger.Infof("unit %q shutting down: %s", u.unit, err)
return err
}
func (u *Uniter) loop(unitTag names.UnitTag) (err error) {
if err := u.init(unitTag); err != nil {
if err == worker.ErrTerminateAgent {
return err
}
return fmt.Errorf("failed to initialize uniter for %q: %v", unitTag, err)
}
logger.Infof("unit %q started", u.unit)
// Install is a special case, as it must run before there
// is any remote state, and before the remote state watcher
// is started.
var charmURL *corecharm.URL
opState := u.operationExecutor.State()
if opState.Kind == operation.Install {
logger.Infof("resuming charm install")
op, err := u.operationFactory.NewInstall(opState.CharmURL)
if err != nil {
return errors.Trace(err)
}
if err := u.operationExecutor.Run(op); err != nil {
return errors.Trace(err)
}
charmURL = opState.CharmURL
} else {
curl, err := u.unit.CharmURL()
if err != nil {
return errors.Trace(err)
}
charmURL = curl
}
var (
watcher *remotestate.RemoteStateWatcher
watcherMu sync.Mutex
)
restartWatcher := func() error {
watcherMu.Lock()
defer watcherMu.Unlock()
if watcher != nil {
if err := watcher.Stop(); err != nil {
return errors.Trace(err)
}
}
var err error
watcher, err = remotestate.NewWatcher(
remotestate.WatcherConfig{
State: remotestate.NewAPIState(u.st),
LeadershipTracker: u.leadershipTracker,
UnitTag: unitTag,
UpdateStatusChannel: u.updateStatusAt,
})
if err != nil {
return errors.Trace(err)
}
// Stop the uniter if the watcher fails. The watcher may be
// stopped cleanly, so only kill the tomb if the error is
// non-nil.
go func(w *remotestate.RemoteStateWatcher) {
if err := w.Wait(); err != nil {
u.tomb.Kill(err)
}
}(watcher)
return nil
}
// watcher may be replaced, so use a closure.
u.addCleanup(func() error {
watcherMu.Lock()
defer watcherMu.Unlock()
if watcher != nil {
return watcher.Stop()
}
return nil
})
onIdle := func() error {
opState := u.operationExecutor.State()
if opState.Kind != operation.Continue {
// We should only set idle status if we're in
// the "Continue" state, which indicates that
// there is nothing to do and we're not in an
// error state.
return nil
}
return setAgentStatus(u, params.StatusIdle, "", nil)
}
clearResolved := func() error {
if err := u.unit.ClearResolved(); err != nil {
return errors.Trace(err)
}
watcher.ClearResolvedMode()
return nil
}
for {
if err = restartWatcher(); err != nil {
err = errors.Annotate(err, "(re)starting watcher")
break
}
uniterResolver := &uniterResolver{
clearResolved: clearResolved,
reportHookError: u.reportHookError,
fixDeployer: u.deployer.Fix,
actionsResolver: actions.NewResolver(),
leadershipResolver: uniterleadership.NewResolver(),
relationsResolver: relation.NewRelationsResolver(u.relations),
storageResolver: storage.NewResolver(u.storage),
}
// We should not do anything until there has been a change
// to the remote state. The watcher will trigger at least
// once initially.
select {
case <-u.tomb.Dying():
return tomb.ErrDying
case <-watcher.RemoteStateChanged():
}
localState := resolver.LocalState{CharmURL: charmURL}
for err == nil {
err = resolver.Loop(resolver.LoopConfig{
Resolver: uniterResolver,
Watcher: watcher,
Executor: u.operationExecutor,
Factory: u.operationFactory,
Dying: u.tomb.Dying(),
OnIdle: onIdle,
CharmDirLocker: u.charmDirLocker,
}, &localState)
switch cause := errors.Cause(err); cause {
case nil:
// Loop back around.
case tomb.ErrDying:
err = tomb.ErrDying
case operation.ErrNeedsReboot:
err = worker.ErrRebootMachine
case operation.ErrHookFailed:
// Loop back around. The resolver can tell that it is in
// an error state by inspecting the operation state.
err = nil
case resolver.ErrTerminate:
err = u.terminate()
case resolver.ErrRestart:
charmURL = localState.CharmURL
// leave err assigned, causing loop to break
default:
// We need to set conflicted from here, because error
// handling is outside of the resolver's control.
if operation.IsDeployConflictError(cause) {
localState.Conflicted = true
err = setAgentStatus(u, params.StatusError, "upgrade failed", nil)
} else {
reportAgentError(u, "resolver loop error", err)
}
}
}
if errors.Cause(err) != resolver.ErrRestart {
break
}
}
logger.Infof("unit %q shutting down: %s", u.unit, err)
return err
}