mesh-host/internal/link/run.go

package link

import (
	"context"
	"crypto/ed25519"
	"encoding/json"
	"errors"
	"fmt"
	"net/url"
	"time"

	amqp "github.com/rabbitmq/amqp091-go"
)

// ErrForged is what a node returns for a declaration whose signature is not the mesh's.
//
// Its own error, and it must never be confused with a malformed message. novox/hq ADR 0004
// requires a host to tell *this is not from the mesh I joined* apart from *this is malformed*:
// the first means somebody is trying, the second means something is broken.
var ErrForged = errors.New("this declaration was not signed by the mesh this node joined")

// AliveEvery is how often a node says it is there.
//
// Often enough that "no word for five minutes" means something, rarely enough that a hundred
// nodes are not a hundred messages a second. The mesh reads absence rather than presence, so what
// matters is the interval being known and steady.
const AliveEvery = 60 * time.Second

// Membership is what a node needs to reach its mesh again, held by the caller.
type Membership struct {
	Node        string
	Broker      string
	Fingerprint string
	Password    string
	Signer      ed25519.PublicKey
}

// Applier is what the host does with a declaration that has been proved to come from the mesh.
//
// It receives the signature as well as the declaration, so the host can keep both: what it was
// told is kept signed and verified again when it is read back, which means the file on disk is
// trusted for the same reason the message was rather than for being local.
type Applier func(ctx context.Context, declaration, signature []byte) Report

// Announce is how the link says what is happening, so a node running unattended leaves an
// account of it. Nil is allowed and means say nothing.
type Announce func(string)

// Hold keeps this node in the mesh, reconnecting for as long as it is asked to.
//
// Disconnection is an ordinary situation and not a failure (novox/hq ADR 0004), so this does not
// give up. A laptop shut for a week comes back and reconnects; it does not come back needing
// somebody to start it again.
//
// The backoff exists because the two common reasons differ in how long they last: a broker
// restarting is back in seconds, and a machine that has moved to a network with no route may be
// hours. Retrying every second for hours is a node shouting into nothing; waiting a minute after
// a broker blip is a node that is needlessly late. So it starts fast and slows down, and resets
// once a connection has actually held.
// Roused is a channel that says the machine has reason to believe its link is stale — it woke
// from suspend, or its network changed.
//
// **The machine knows before any timeout does.** A suspended laptop's connection is dead the
// moment it wakes, and heartbeats find that out in twenty or thirty seconds; for that time the
// node believes it is in the mesh and is not, which is the one state this design says must never
// be indistinguishable from being connected. Nothing new listens on the node to arrange it — the
// signal a service manager already sends is enough (novox/hq ADR 0004).
//
// Nil is allowed and means nothing ever rouses it, which is every machine that does not suspend.
type Roused <-chan struct{}

func Hold(ctx context.Context, m Membership, apply Applier, say Announce, timeout time.Duration) error {
	return HoldRoused(ctx, m, apply, say, timeout, nil)
}

// HoldRoused is Hold, told when the machine has reason to think its link is stale.
func HoldRoused(ctx context.Context, m Membership, apply Applier, say Announce,
	timeout time.Duration, roused Roused) error {

	return holdWith(ctx, func(ctx context.Context) error {
		return Run(ctx, m, apply, say, timeout)
	}, say, roused)
}

// attempt is one try at holding the link open, returning when it ends for any reason.
//
// Named so the loop below can be driven without a broker. What the loop decides — when to wait,
// how long, what being roused does — is the part with the reasoning in it, and it was reachable
// only through a real connection before.
type attempt func(context.Context) error

func holdWith(ctx context.Context, run attempt, say Announce, roused Roused) error {
	const (
		first = 2 * time.Second
		most  = 2 * time.Minute
		// A connection that lasted this long counts as having worked, so the next failure starts
		// from the bottom again. Without it a node that reconnects and immediately drops climbs
		// to the maximum and stays there, long after whatever caused it went away.
		settled = 30 * time.Second
	)
	wait := first

	for {
		began := time.Now()

		// The link runs under a context this loop can cancel, so being roused ends the current
		// attempt rather than only shortening the wait after it.
		//
		// **That is the whole of it.** After a resume the socket looks perfectly healthy from
		// inside this process — there is no error and no close, because nothing has tried to
		// send anything. It is heartbeats that eventually discover it, twenty or thirty seconds
		// later. A machine that knows it just woke does not have to wait to be told.
		//
		// A rouse that turns out to be spurious costs one reconnect, which is cheap and
		// idempotent: the node redeclares its queue and anything unacknowledged is redelivered.
		// The alternative costs half a minute of believing it is in a mesh it has left.
		trying, done := context.WithCancel(ctx)
		if roused != nil {
			go func() {
				select {
				case <-trying.Done():
				case <-roused:
					say("woken or moved — dropping the link and opening it again")
					done()
				}
			}()
		}
		err := run(trying)
		done()
		if ctx.Err() != nil {
			return nil
		}
		if time.Since(began) > settled {
			wait = first
		}

		switch {
		case errors.Is(err, ErrWrongCertificate):
			// Said in full every time rather than folded into a retry count. This does not mean
			// the network is down; it means what answered is not the mesh this node joined, and
			// no amount of waiting fixes it. The node keeps running what it was last told, which
			// is the right thing to do while somebody works out what happened.
			say("the broker is not the one this node joined: " + err.Error())
			say("this will not fix itself. This node keeps running what it was last told.")
		case err != nil:
			say(fmt.Sprintf("disconnected: %v — trying again in %s", err, wait))
		default:
			say(fmt.Sprintf("the link closed — trying again in %s", wait))
		}

		select {
		case <-ctx.Done():
			return nil
		case <-roused:
			// And it does not serve out a wait computed for a broker that was restarting, either.
			//
			// The backoff is not *reset* by this. Being roused says the machine changed, not that
			// whatever was refusing the connection has stopped — a laptop woken repeatedly on a
			// network with no route would otherwise retry at full speed for as long as somebody
			// keeps opening the lid.
			say("woken or moved — trying again now")
		case <-time.After(wait):
		}
		if wait *= 2; wait > most {
			wait = most
		}
	}
}

// Run holds the link open once, applying what arrives and reporting what happened.
//
// Outbound only, and nothing listens on this machine. Returns when the link ends, for any reason;
// Hold is what decides whether to open it again.
func Run(ctx context.Context, m Membership, apply Applier, say Announce, timeout time.Duration) error {
	if say == nil {
		say = func(string) {}
	}
	config, err := PinnedConfig(m.Fingerprint)
	if err != nil {
		return err
	}

	dsn := fmt.Sprintf("amqps://%s:%s@%s/",
		url.QueryEscape(m.Node), url.QueryEscape(m.Password), m.Broker)
	conn, err := amqp.DialConfig(dsn, amqp.Config{
		TLSClientConfig: config,
		Dial:            amqp.DefaultDial(timeout),
		// Kept short so a node that has silently lost its route notices, rather than holding a
		// connection the broker forgot about and believing it is still in the mesh.
		Heartbeat: 10 * time.Second,
	})
	if err != nil {
		if errors.Is(err, ErrWrongCertificate) {
			return err
		}
		return fmt.Errorf("cannot reach the broker at %s: %w", m.Broker, err)
	}
	defer conn.Close()

	channel, err := conn.Channel()
	if err != nil {
		return err
	}
	defer channel.Close()

	queue := QueueFor(m.Node)
	if _, err := channel.QueueDeclare(queue, true, false, false, false, nil); err != nil {
		return fmt.Errorf("cannot declare this node's queue %s: %w", queue, err)
	}

	// One at a time. A declaration is applied to a machine, and applying two at once would race
	// on the same filesystem — so the broker holds the next one until this one is finished,
	// where it survives a restart.
	if err := channel.Qos(1, 0, false); err != nil {
		return err
	}

	deliveries, err := channel.ConsumeWithContext(ctx, queue, "", false, false, false, false, nil)
	if err != nil {
		return err
	}
	// Said, because it is the event anybody watching actually wants. Without it a node logs
	// every failure and nothing on success, so a log full of "trying again" and then silence
	// reads as still broken when it means the opposite.
	say("in the mesh, consuming " + queue)

	// A word every so often, so the mesh can tell a node that is quiet from one that is gone.
	// Cheap on purpose: it carries a name and nothing else, because anything more would be a
	// report, and reports are rare where this is constant.
	beat := time.NewTicker(AliveEvery)
	defer beat.Stop()
	publishAlive(ctx, channel, m, say, timeout)

	closed := conn.NotifyClose(make(chan *amqp.Error, 1))

	// Published mandatory, so the broker hands back anything it cannot route rather than
	// dropping it. Without this a report goes to an exchange with no matching binding, the
	// publisher is told nothing, and the mesh believes this node never answered while the node
	// believes it did — which is what happened when `report` was left unbound on the other side.
	returned := channel.NotifyReturn(make(chan amqp.Return, 4))
	go func() {
		for r := range returned {
			say(fmt.Sprintf("the broker could not route this node's %s: %s (%d %s)",
				r.RoutingKey, r.Exchange, r.ReplyCode, r.ReplyText))
		}
	}()

	for {
		select {
		case <-ctx.Done():
			return nil
		case <-beat.C:
			publishAlive(ctx, channel, m, say, timeout)
		case reason := <-closed:
			return fmt.Errorf("the link closed: %v", reason)
		case delivery, ok := <-deliveries:
			if !ok {
				return errors.New("the broker stopped delivering")
			}
			report := handle(ctx, m, apply, delivery)
			switch {
			case report.Refused != "":
				say("refused a declaration: " + report.Refused)
			case len(report.Failed) > 0:
				say(fmt.Sprintf("applied %d and failed: %v", len(report.Applied), report.Failed))
			default:
				say(fmt.Sprintf("applied %d resource(s)", len(report.Applied)))
			}
			publishReport(ctx, channel, m, report, say, timeout)
			// Acknowledged after the report is published. A node that dies between applying and
			// reporting leaves the declaration on the broker and applies it again on return,
			// which is safe because applying is reconciliation — it converges rather than
			// repeating.
			_ = delivery.Ack(false)
		}
	}
}

func handle(ctx context.Context, m Membership, apply Applier, delivery amqp.Delivery) Report {
	return handleBody(ctx, m, delivery.Body, apply)
}

// handleBody is the whole of deciding whether to trust a message, separated from the broker so it
// can be tested as the security check it is rather than as message plumbing.
func handleBody(ctx context.Context, m Membership, body []byte, apply Applier) Report {
	var signed Signed
	if err := json.Unmarshal(body, &signed); err != nil {
		return Report{Node: m.Node, Refused: "this message is not a declaration: " + err.Error()}
	}

	// Before anything is read out of it, let alone applied. The host applies whatever the link
	// delivers, so this check is the difference between the mesh changing this machine and
	// anybody changing it.
	if !ed25519.Verify(m.Signer, signed.Declaration, signed.Signature) {
		return Report{Node: m.Node, Refused: ErrForged.Error()}
	}
	return apply(ctx, signed.Declaration, signed.Signature)
}

func publishReport(ctx context.Context, channel *amqp.Channel, m Membership, report Report,
	say Announce, timeout time.Duration) {
	report.Node = m.Node
	body, err := json.Marshal(report)
	if err != nil {
		say("cannot encode this node's own report: " + err.Error())
		return
	}
	publish, cancel := context.WithTimeout(ctx, timeout)
	defer cancel()

	// Said rather than swallowed. A report that fails to publish leaves the mesh believing this
	// node never answered, while the node believes it did — and the two would go on disagreeing
	// with nothing anywhere saying so. That shape of fault is the one this project keeps finding.
	if err := channel.PublishWithContext(publish, Exchange, KeyReport, true, false,
		amqp.Publishing{ContentType: "application/json", Body: body}); err != nil {
		say(fmt.Sprintf("applied, and could not tell the mesh: %v", err))
	}
}

// publishAlive says this node is here, and nothing else.
func publishAlive(ctx context.Context, channel *amqp.Channel, m Membership, say Announce,
	timeout time.Duration) {
	body, err := json.Marshal(Alive{Node: m.Node})
	if err != nil {
		return
	}
	publish, cancel := context.WithTimeout(ctx, timeout)
	defer cancel()
	// Not mandatory, unlike a report. Losing one is nothing: the next is a minute away, and the
	// mesh is reading a gap rather than counting arrivals. Insisting on delivery would turn a
	// harmless miss into a logged failure every minute.
	if err := channel.PublishWithContext(publish, Exchange, KeyAlive, false, false,
		amqp.Publishing{ContentType: "application/json", Body: body}); err != nil {
		say("could not tell the mesh this node is here: " + err.Error())
	}
}