isle/go/daemon/daemon.go

// Package daemon implements the isle daemon, which is a long-running service
// managing all isle background tasks and sub-processes for a single network.
package daemon

import (
	"bytes"
	"context"
	"errors"
	"fmt"
	"io"
	"io/fs"
	"isle/bootstrap"
	"os"
	"sync"
	"time"

	"dev.mediocregopher.com/mediocre-go-lib.git/mlog"
)

// Daemon presents all functionality required for client frontends to interact
// with isle, typically via the unix socket.
type Daemon interface {

	// JoinNetwork joins the Daemon to an existing network using the given
	// Bootstrap.
	//
	// Errors:
	// - ErrAlreadyJoined
	JoinNetwork(context.Context, bootstrap.Bootstrap) error

	// GetGarageBootstrapHosts loads (and verifies) the <hostname>.json.signed
	// file for all hosts stored in garage.
	GetGarageBootstrapHosts(
		ctx context.Context,
	) (
		map[string]bootstrap.Host, error,
	)

	// Shutdown blocks until all resources held or created by the daemon,
	// including child processes it has started, have been cleaned up.
	//
	// If this returns an error then it's possible that child processes are
	// still running and are no longer managed.
	Shutdown() error
}

// Opts are optional parameters which can be passed in when initializing a new
// Daemon instance. A nil Opts is equivalent to a zero value.
type Opts struct {
	// Stdout and Stderr are what the associated outputs from child processes
	// will be directed to.
	Stdout, Stderr io.Writer

	EnvVars EnvVars // Defaults to that returned by GetEnvVars.
}

func (o *Opts) withDefaults() *Opts {
	if o == nil {
		o = new(Opts)
	}

	if o.Stdout == nil {
		o.Stdout = os.Stdout
	}

	if o.Stderr == nil {
		o.Stderr = os.Stderr
	}

	if o.EnvVars == (EnvVars{}) {
		o.EnvVars = GetEnvVars()
	}

	return o
}

const (
	daemonStateNoNetwork = iota
	daemonStateInitializing
	daemonStateOk
	daemonStateRestarting
	daemonStateShutdown
)

type daemon struct {
	logger        *mlog.Logger
	daemonConfig  Config
	envBinDirPath string
	opts          *Opts

	l             sync.RWMutex
	state         int
	children      *Children
	currBootstrap bootstrap.Bootstrap

	shutdownCh chan struct{}
	wg         sync.WaitGroup
}

// NewDaemon initializes and returns a Daemon instance which will manage all
// child processes and state required by the isle service, as well as an HTTP
// socket over which RPC calls will be served.
//
// Inner Children instance(s) will be wrapped such that they will be
// automatically shutdown and re-created whenever there's changes in the network
// which require the configuration to be changed (e.g. a new nebula lighthouse).
// During such an inner restart all methods will return ErrRestarting, except
// Shutdown which will block until the currently executing restart is finished
// and then shutdown cleanly from there.
//
// While still starting up the Daemon for the first time all methods will return
// ErrInitializing, except Shutdown which will block until initialization is
// canceled.
//
// TODO make daemon smarter, it currently restarts on _any_ change, but
// it should restart itself only when there's something actually requiring a
// restart.
func NewDaemon(
	logger *mlog.Logger,
	daemonConfig Config,
	envBinDirPath string,
	opts *Opts,
) (
	Daemon, error,
) {
	var (
		d = &daemon{
			logger:        logger,
			daemonConfig:  daemonConfig,
			envBinDirPath: envBinDirPath,
			opts:          opts.withDefaults(),
			shutdownCh:    make(chan struct{}),
		}
		bootstrapFilePath = bootstrap.StateDirPath(d.opts.EnvVars.StateDirPath)
	)

	currBootstrap, err := bootstrap.FromFile(bootstrapFilePath)
	if errors.Is(err, fs.ErrNotExist) {
		// daemon has never had a network created or joined
	} else if err != nil {
		return nil, fmt.Errorf(
			"loading bootstrap from %q: %w", bootstrapFilePath, err,
		)
	} else if err := d.initialize(currBootstrap); err != nil {
		return nil, fmt.Errorf("initializing with bootstrap: %w", err)
	}

	return d, nil
}

func (d *daemon) initialize(currBootstrap bootstrap.Bootstrap) error {
	// we update this Host's data using whatever configuration has been provided
	// by the daemon config. This way the daemon has the most up-to-date
	// possible bootstrap. This updated bootstrap will later get updated in
	// garage as a background daemon task, so other hosts will see it as well.
	currBootstrap, err := coalesceDaemonConfigAndBootstrap(
		d.daemonConfig, currBootstrap,
	)
	if err != nil {
		return fmt.Errorf("combining daemon configuration into bootstrap: %w", err)
	}

	err = writeBootstrapToStateDir(d.opts.EnvVars.StateDirPath, currBootstrap)
	if err != nil {
		return fmt.Errorf("writing bootstrap to state dir: %w", err)
	}

	d.currBootstrap = currBootstrap
	d.state = daemonStateInitializing

	ctx, cancel := context.WithCancel(context.Background())
	d.wg.Add(1)
	go func() {
		defer d.wg.Done()
		<-d.shutdownCh
		cancel()
	}()

	d.wg.Add(1)
	go func() {
		defer d.wg.Done()
		d.restartLoop(ctx)
		d.logger.Debug(ctx, "Daemon restart loop stopped")
	}()

	return nil
}

func withCurrBootstrap[Res any](
	d *daemon, fn func(bootstrap.Bootstrap) (Res, error),
) (Res, error) {
	var zero Res
	d.l.RLock()
	defer d.l.RUnlock()

	currBootstrap, state := d.currBootstrap, d.state

	switch state {
	case daemonStateNoNetwork:
		return zero, ErrNoNetwork
	case daemonStateInitializing:
		return zero, ErrInitializing
	case daemonStateOk:
		return fn(currBootstrap)
	case daemonStateRestarting:
		return zero, ErrRestarting
	case daemonStateShutdown:
		return zero, errors.New("already shutdown")
	default:
		panic(fmt.Sprintf("unknown state %d", d.state))
	}
}

// creates a new bootstrap file using available information from the network. If
// the new bootstrap file is different than the existing one, the existing one
// is overwritten and true is returned.
func (d *daemon) checkBootstrap(
	ctx context.Context, hostBootstrap bootstrap.Bootstrap,
) (
	bootstrap.Bootstrap, bool, error,
) {

	thisHost := hostBootstrap.ThisHost()

	newHosts, err := getGarageBootstrapHosts(ctx, d.logger, hostBootstrap)
	if err != nil {
		return bootstrap.Bootstrap{}, false, fmt.Errorf("getting hosts from garage: %w", err)
	}

	// the daemon's view of this host's bootstrap info takes precedence over
	// whatever is in garage
	newHosts[thisHost.Name] = thisHost

	newHostsHash, err := bootstrap.HostsHash(newHosts)
	if err != nil {
		return bootstrap.Bootstrap{}, false, fmt.Errorf("calculating hash of new hosts: %w", err)
	}

	currHostsHash, err := bootstrap.HostsHash(hostBootstrap.Hosts)
	if err != nil {
		return bootstrap.Bootstrap{}, false, fmt.Errorf("calculating hash of current hosts: %w", err)
	}

	if bytes.Equal(newHostsHash, currHostsHash) {
		return hostBootstrap, false, nil
	}

	hostBootstrap.Hosts = newHosts

	return hostBootstrap, true, nil
}

// blocks until a new bootstrap is available or context is canceled
func (d *daemon) watchForChanges(ctx context.Context) bootstrap.Bootstrap {

	ticker := time.NewTicker(3 * time.Minute)
	defer ticker.Stop()

	for {
		select {

		case <-ctx.Done():
			return bootstrap.Bootstrap{}

		case <-ticker.C:

			d.logger.Info(ctx, "Checking for changes to bootstrap")

			newBootstrap, changed, err := d.checkBootstrap(
				ctx, d.currBootstrap,
			)
			if err != nil {
				d.logger.Error(ctx, "Checking bootstrap for changes failed", err)
				continue
			} else if !changed {
				continue
			}

			err = writeBootstrapToStateDir(d.opts.EnvVars.StateDirPath, newBootstrap)
			if err != nil {
				d.logger.Error(ctx, "Writing new bootstrap to disk failed", err)
				continue
			}

			return newBootstrap
		}
	}
}

func (d *daemon) restartLoop(ctx context.Context) {
	wait := func(d time.Duration) bool {
		select {
		case <-ctx.Done():
			return false
		case <-time.After(d):
			return true
		}
	}

	for {
		if ctx.Err() != nil {
			return
		}

		d.logger.Info(ctx, "Creating new daemon")
		children, err := NewChildren(
			ctx,
			d.logger.WithNamespace("daemon"),
			d.daemonConfig,
			d.currBootstrap,
			d.envBinDirPath,
			d.opts,
		)
		if errors.Is(err, context.Canceled) {
			return
		} else if err != nil {
			d.logger.Error(ctx, "failed to initialize daemon", err)
			if !wait(1 * time.Second) {
				return
			}
			continue
		}

		d.l.Lock()
		d.children = children
		d.state = daemonStateOk
		d.l.Unlock()

		if len(d.daemonConfig.Storage.Allocations) > 0 {
			if !until(
				ctx,
				d.logger,
				"Applying garage layout",
				func(ctx context.Context) error {
					return GarageApplyLayout(
						ctx, d.logger, d.currBootstrap, d.daemonConfig,
					)
				},
			) {
				return
			}
		}

		if !until(
			ctx,
			d.logger,
			"Updating host info in garage",
			func(ctx context.Context) error {
				return d.putGarageBoostrapHost(ctx)
			},
		) {
			return
		}

		newBootstrap := d.watchForChanges(ctx)
		if ctx.Err() != nil {
			return
		}

		d.logger.Info(ctx, "Bootstrap has changed, will restart daemon")
		d.l.Lock()
		d.currBootstrap = newBootstrap
		d.state = daemonStateRestarting
		d.l.Unlock()

		d.logger.Info(ctx, "Shutting down previous child processes")
		if err := d.children.Shutdown(); err != nil {
			d.logger.Fatal(ctx, "Failed to cleanly shutdown children, there may be orphaned child processes", err)
		}
	}
}

func (d *daemon) JoinNetwork(
	ctx context.Context, newBootstrap bootstrap.Bootstrap,
) error {
	d.l.Lock()
	defer d.l.Unlock()

	if d.state != daemonStateNoNetwork {
		return ErrAlreadyJoined
	}

	return d.initialize(newBootstrap)
}

func (d *daemon) GetGarageBootstrapHosts(
	ctx context.Context,
) (
	map[string]bootstrap.Host, error,
) {
	return withCurrBootstrap(d, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		map[string]bootstrap.Host, error,
	) {
		return getGarageBootstrapHosts(ctx, d.logger, currBootstrap)
	})
}

func (d *daemon) Shutdown() error {
	d.l.Lock()
	defer d.l.Unlock()

	close(d.shutdownCh)
	d.wg.Wait()
	d.state = daemonStateShutdown

	if d.children != nil {
		if err := d.children.Shutdown(); err != nil {
			return fmt.Errorf("shutting down children: %w", err)
		}
	}

	return nil
}