worker/uniter/resolver.go

// Copyright 2015 Canonical Ltd.
// Licensed under the AGPLv3, see LICENCE file for details.

package uniter

import (
	"github.com/juju/errors"
	"gopkg.in/juju/charm.v6-unstable/hooks"

	"github.com/juju/juju/apiserver/params"
	"github.com/juju/juju/worker/uniter/hook"
	"github.com/juju/juju/worker/uniter/operation"
	"github.com/juju/juju/worker/uniter/remotestate"
	"github.com/juju/juju/worker/uniter/resolver"
)

// ResolverConfig defines configuration for the uniter resolver.
type ResolverConfig struct {
	ClearResolved       func() error
	ReportHookError     func(hook.Info) error
	FixDeployer         func() error
	ShouldRetryHooks    bool
	StartRetryHookTimer func()
	StopRetryHookTimer  func()
	Leadership          resolver.Resolver
	Actions             resolver.Resolver
	Relations           resolver.Resolver
	Storage             resolver.Resolver
	Commands            resolver.Resolver
}

type uniterResolver struct {
	config                ResolverConfig
	retryHookTimerStarted bool
}

// NewUniterResolver returns a new resolver.Resolver for the uniter.
func NewUniterResolver(cfg ResolverConfig) resolver.Resolver {
	return &uniterResolver{
		config:                cfg,
		retryHookTimerStarted: false,
	}
}

func (s *uniterResolver) NextOp(
	localState resolver.LocalState,
	remoteState remotestate.Snapshot,
	opFactory operation.Factory,
) (operation.Operation, error) {
	if remoteState.Life == params.Dead || localState.Stopped {
		return nil, resolver.ErrTerminate
	}

	if localState.Kind == operation.Upgrade {
		if localState.Conflicted {
			return s.nextOpConflicted(localState, remoteState, opFactory)
		}
		logger.Infof("resuming charm upgrade")
		return opFactory.NewUpgrade(localState.CharmURL)
	}

	if localState.Restart {
		// We've just run the upgrade op, which will change the
		// unit's charm URL. We need to restart the resolver
		// loop so that we start watching the correct events.
		return nil, resolver.ErrRestart
	}

	if localState.Kind == operation.Continue {
		if err := s.config.FixDeployer(); err != nil {
			return nil, errors.Trace(err)
		}
	}

	if s.retryHookTimerStarted && (localState.Kind != operation.RunHook || localState.Step != operation.Pending) {
		// The hook-retry timer is running, but there is no pending
		// hook operation. We're not in an error state, so stop the
		// timer now to reset the backoff state.
		s.config.StopRetryHookTimer()
		s.retryHookTimerStarted = false
	}

	op, err := s.config.Leadership.NextOp(localState, remoteState, opFactory)
	if errors.Cause(err) != resolver.ErrNoOperation {
		return op, err
	}

	op, err = s.config.Actions.NextOp(localState, remoteState, opFactory)
	if errors.Cause(err) != resolver.ErrNoOperation {
		return op, err
	}

	op, err = s.config.Commands.NextOp(localState, remoteState, opFactory)
	if errors.Cause(err) != resolver.ErrNoOperation {
		return op, err
	}

	op, err = s.config.Storage.NextOp(localState, remoteState, opFactory)
	if errors.Cause(err) != resolver.ErrNoOperation {
		return op, err
	}

	switch localState.Kind {
	case operation.RunHook:
		switch localState.Step {
		case operation.Pending:
			logger.Infof("awaiting error resolution for %q hook", localState.Hook.Kind)
			return s.nextOpHookError(localState, remoteState, opFactory)

		case operation.Queued:
			logger.Infof("found queued %q hook", localState.Hook.Kind)
			if localState.Hook.Kind == hooks.Install {
				// Special case: handle install in nextOp,
				// so we do nothing when the unit is dying.
				return s.nextOp(localState, remoteState, opFactory)
			}
			return opFactory.NewRunHook(*localState.Hook)

		case operation.Done:
			logger.Infof("committing %q hook", localState.Hook.Kind)
			return opFactory.NewSkipHook(*localState.Hook)

		default:
			return nil, errors.Errorf("unknown operation step %v", localState.Step)
		}

	case operation.Continue:
		logger.Infof("no operations in progress; waiting for changes")
		return s.nextOp(localState, remoteState, opFactory)

	default:
		return nil, errors.Errorf("unknown operation kind %v", localState.Kind)
	}
}

// nextOpConflicted is called after an upgrade operation has failed, and hasn't
// yet been resolved or reverted. When in this mode, the resolver will only
// consider those two possibilities for progressing.
func (s *uniterResolver) nextOpConflicted(
	localState resolver.LocalState,
	remoteState remotestate.Snapshot,
	opFactory operation.Factory,
) (operation.Operation, error) {
	if remoteState.ResolvedMode != params.ResolvedNone {
		if err := s.config.ClearResolved(); err != nil {
			return nil, errors.Trace(err)
		}
		return opFactory.NewResolvedUpgrade(localState.CharmURL)
	}
	if remoteState.ForceCharmUpgrade && charmModified(localState, remoteState) {
		return opFactory.NewRevertUpgrade(remoteState.CharmURL)
	}
	return nil, resolver.ErrWaiting
}

func (s *uniterResolver) nextOpHookError(
	localState resolver.LocalState,
	remoteState remotestate.Snapshot,
	opFactory operation.Factory,
) (operation.Operation, error) {

	// Report the hook error.
	if err := s.config.ReportHookError(*localState.Hook); err != nil {
		return nil, errors.Trace(err)
	}

	if remoteState.ForceCharmUpgrade && charmModified(localState, remoteState) {
		return opFactory.NewUpgrade(remoteState.CharmURL)
	}

	switch remoteState.ResolvedMode {
	case params.ResolvedNone:
		if remoteState.RetryHookVersion > localState.RetryHookVersion {
			// We've been asked to retry: clear the hook timer
			// started state so we'll restart it if this fails.
			//
			// If the hook fails again, we'll re-enter this method
			// with the retry hook versions equal and restart the
			// timer. If the hook succeeds, we'll enter nextOp
			// and stop the timer.
			s.retryHookTimerStarted = false
			return opFactory.NewRunHook(*localState.Hook)
		}
		if !s.retryHookTimerStarted && s.config.ShouldRetryHooks {
			// We haven't yet started a retry timer, so start one
			// now. If we retry and fail, retryHookTimerStarted is
			// cleared so that we'll still start it again.
			s.config.StartRetryHookTimer()
			s.retryHookTimerStarted = true
		}
		return nil, resolver.ErrNoOperation
	case params.ResolvedRetryHooks:
		s.config.StopRetryHookTimer()
		s.retryHookTimerStarted = false
		if err := s.config.ClearResolved(); err != nil {
			return nil, errors.Trace(err)
		}
		return opFactory.NewRunHook(*localState.Hook)
	case params.ResolvedNoHooks:
		s.config.StopRetryHookTimer()
		s.retryHookTimerStarted = false
		if err := s.config.ClearResolved(); err != nil {
			return nil, errors.Trace(err)
		}
		return opFactory.NewSkipHook(*localState.Hook)
	default:
		return nil, errors.Errorf(
			"unknown resolved mode %q", remoteState.ResolvedMode,
		)
	}
}

func charmModified(local resolver.LocalState, remote remotestate.Snapshot) bool {
	if *local.CharmURL != *remote.CharmURL {
		logger.Debugf("upgrade from %v to %v", local.CharmURL, remote.CharmURL)
		return true
	}

	if local.CharmModifiedVersion != remote.CharmModifiedVersion {
		logger.Debugf("upgrade from CharmModifiedVersion %v to %v", local.CharmModifiedVersion, remote.CharmModifiedVersion)
		return true
	}
	return false
}

func (s *uniterResolver) nextOp(
	localState resolver.LocalState,
	remoteState remotestate.Snapshot,
	opFactory operation.Factory,
) (operation.Operation, error) {

	switch remoteState.Life {
	case params.Alive:
	case params.Dying:
		// Normally we handle relations last, but if we're dying we
		// must ensure that all relations are broken first.
		op, err := s.config.Relations.NextOp(localState, remoteState, opFactory)
		if errors.Cause(err) != resolver.ErrNoOperation {
			return op, err
		}

		// We're not in a hook error and the unit is Dying,
		// so we should proceed to tear down.
		//
		// TODO(axw) move logic for cascading destruction of
		//           subordinates, relation units and storage
		//           attachments into state, via cleanups.
		if localState.Started {
			return opFactory.NewRunHook(hook.Info{Kind: hooks.Stop})
		}
		fallthrough
	case params.Dead:
		// The unit is dying/dead and stopped, so tell the uniter
		// to terminate.
		return nil, resolver.ErrTerminate
	}

	// Now that storage hooks have run at least once, before anything else,
	// we need to run the install hook.
	// TODO(cmars): remove !localState.Started. It's here as a temporary
	// measure because unit agent upgrades aren't being performed yet.
	if !localState.Installed && !localState.Started {
		return opFactory.NewRunHook(hook.Info{Kind: hooks.Install})
	}

	if charmModified(localState, remoteState) {
		return opFactory.NewUpgrade(remoteState.CharmURL)
	}

	if localState.ConfigVersion != remoteState.ConfigVersion {
		return opFactory.NewRunHook(hook.Info{Kind: hooks.ConfigChanged})
	}

	op, err := s.config.Relations.NextOp(localState, remoteState, opFactory)
	if errors.Cause(err) != resolver.ErrNoOperation {
		return op, err
	}

	// UpdateStatus hook runs if nothing else needs to.
	if localState.UpdateStatusVersion != remoteState.UpdateStatusVersion {
		return opFactory.NewRunHook(hook.Info{Kind: hooks.UpdateStatus})
	}

	return nil, resolver.ErrNoOperation
}