nebula/interface.go

package nebula

import (
	"context"
	"errors"
	"fmt"
	"net/netip"
	"os"
	"runtime"
	"sync/atomic"
	"time"

	"github.com/gaissmai/bart"
	"github.com/rcrowley/go-metrics"
	"github.com/sirupsen/logrus"
	"github.com/slackhq/nebula/config"
	"github.com/slackhq/nebula/firewall"
	"github.com/slackhq/nebula/header"
	"github.com/slackhq/nebula/overlay"
	"github.com/slackhq/nebula/udp"
)

const mtu = 9001
const virtioNetHdrLen = overlay.VirtioNetHdrLen

type InterfaceConfig struct {
	HostMap            *HostMap
	Outside            udp.Conn
	Inside             overlay.Device
	pki                *PKI
	Cipher             string
	Firewall           *Firewall
	ServeDns           bool
	HandshakeManager   *HandshakeManager
	lightHouse         *LightHouse
	connectionManager  *connectionManager
	DropLocalBroadcast bool
	DropMulticast      bool
	routines           int
	MessageMetrics     *MessageMetrics
	version            string
	relayManager       *relayManager
	punchy             *Punchy

	tryPromoteEvery uint32
	reQueryEvery    uint32
	reQueryWait     time.Duration

	ConntrackCacheTimeout time.Duration
	l                     *logrus.Logger
}

type batchMetrics struct {
	udpReadSize      metrics.Histogram
	encryptionTime   metrics.Histogram // Time spent in encryption (including lock waits)
	batchSize        metrics.Histogram // Dynamic batch sizes being used
	lockAcquisitions metrics.Counter   // Number of lock acquisitions (should be minimal)
}

type Interface struct {
	hostMap               *HostMap
	outside               udp.Conn
	inside                overlay.Device
	pki                   *PKI
	firewall              *Firewall
	connectionManager     *connectionManager
	handshakeManager      *HandshakeManager
	serveDns              bool
	createTime            time.Time
	lightHouse            *LightHouse
	myBroadcastAddrsTable *bart.Lite
	myVpnAddrs            []netip.Addr // A list of addresses assigned to us via our certificate
	myVpnAddrsTable       *bart.Lite
	myVpnNetworks         []netip.Prefix // A list of networks assigned to us via our certificate
	myVpnNetworksTable    *bart.Lite
	dropLocalBroadcast    bool
	dropMulticast         bool
	routines              int
	disconnectInvalid     atomic.Bool
	closed                atomic.Bool
	relayManager          *relayManager

	tryPromoteEvery atomic.Uint32
	reQueryEvery    atomic.Uint32
	reQueryWait     atomic.Int64

	sendRecvErrorConfig   recvErrorConfig
	acceptRecvErrorConfig recvErrorConfig

	// rebindCount is used to decide if an active tunnel should trigger a punch notification through a lighthouse
	rebindCount int8
	version     string

	conntrackCacheTimeout time.Duration

	writers []udp.Conn
	readers []overlay.BatchReadWriter

	metricHandshakes    metrics.Histogram
	messageMetrics      *MessageMetrics
	cachedPacketMetrics *cachedPacketMetrics
	batchMetrics        *batchMetrics

	l *logrus.Logger
}

type EncWriter interface {
	SendVia(via *HostInfo,
		relay *Relay,
		ad,
		nb,
		out []byte,
		nocopy bool,
	)
	SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnAddr netip.Addr, p, nb, out []byte)
	SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p, nb, out []byte)
	Handshake(vpnAddr netip.Addr)
	GetHostInfo(vpnAddr netip.Addr) *HostInfo
	GetCertState() *CertState
}

type recvErrorConfig uint8

const (
	recvErrorAlways recvErrorConfig = iota
	recvErrorNever
	recvErrorPrivate
)

func (s recvErrorConfig) ShouldRecvError(endpoint netip.AddrPort) bool {
	switch s {
	case recvErrorPrivate:
		return endpoint.Addr().IsPrivate()
	case recvErrorAlways:
		return true
	case recvErrorNever:
		return false
	default:
		panic(fmt.Errorf("invalid recvErrorConfig value: %d", s))
	}
}

func (s recvErrorConfig) String() string {
	switch s {
	case recvErrorAlways:
		return "always"
	case recvErrorNever:
		return "never"
	case recvErrorPrivate:
		return "private"
	default:
		return fmt.Sprintf("invalid(%d)", s)
	}
}

func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
	if c.Outside == nil {
		return nil, errors.New("no outside connection")
	}
	if c.Inside == nil {
		return nil, errors.New("no inside interface (tun)")
	}
	if c.pki == nil {
		return nil, errors.New("no certificate state")
	}
	if c.Firewall == nil {
		return nil, errors.New("no firewall rules")
	}
	if c.connectionManager == nil {
		return nil, errors.New("no connection manager")
	}

	cs := c.pki.getCertState()
	ifce := &Interface{
		pki:                   c.pki,
		hostMap:               c.HostMap,
		outside:               c.Outside,
		inside:                c.Inside,
		firewall:              c.Firewall,
		serveDns:              c.ServeDns,
		handshakeManager:      c.HandshakeManager,
		createTime:            time.Now(),
		lightHouse:            c.lightHouse,
		dropLocalBroadcast:    c.DropLocalBroadcast,
		dropMulticast:         c.DropMulticast,
		routines:              c.routines,
		version:               c.version,
		writers:               make([]udp.Conn, c.routines),
		readers:               make([]overlay.BatchReadWriter, c.routines),
		myVpnNetworks:         cs.myVpnNetworks,
		myVpnNetworksTable:    cs.myVpnNetworksTable,
		myVpnAddrs:            cs.myVpnAddrs,
		myVpnAddrsTable:       cs.myVpnAddrsTable,
		myBroadcastAddrsTable: cs.myVpnBroadcastAddrsTable,
		relayManager:          c.relayManager,
		connectionManager:     c.connectionManager,
		conntrackCacheTimeout: c.ConntrackCacheTimeout,

		metricHandshakes: metrics.GetOrRegisterHistogram("handshakes", nil, metrics.NewExpDecaySample(1028, 0.015)),
		messageMetrics:   c.MessageMetrics,
		cachedPacketMetrics: &cachedPacketMetrics{
			sent:    metrics.GetOrRegisterCounter("hostinfo.cached_packets.sent", nil),
			dropped: metrics.GetOrRegisterCounter("hostinfo.cached_packets.dropped", nil),
		},
		batchMetrics: &batchMetrics{
			udpReadSize:      metrics.GetOrRegisterHistogram("batch.udp_read_size", nil, metrics.NewUniformSample(1024)),
			encryptionTime:   metrics.GetOrRegisterHistogram("batch.encryption_time_ns", nil, metrics.NewUniformSample(1024)),
			batchSize:        metrics.GetOrRegisterHistogram("batch.size", nil, metrics.NewUniformSample(1024)),
			lockAcquisitions: metrics.GetOrRegisterCounter("batch.lock_acquisitions", nil),
		},

		l: c.l,
	}

	ifce.tryPromoteEvery.Store(c.tryPromoteEvery)
	ifce.reQueryEvery.Store(c.reQueryEvery)
	ifce.reQueryWait.Store(int64(c.reQueryWait))

	ifce.connectionManager.intf = ifce

	return ifce, nil
}

// activate creates the interface on the host. After the interface is created, any
// other services that want to bind listeners to its IP may do so successfully. However,
// the interface isn't going to process anything until run() is called.
func (f *Interface) activate() {
	// actually turn on tun dev

	addr, err := f.outside.LocalAddr()
	if err != nil {
		f.l.WithError(err).Error("Failed to get udp listen address")
	}

	f.l.WithField("interface", f.inside.Name()).WithField("networks", f.myVpnNetworks).
		WithField("build", f.version).WithField("udpAddr", addr).
		WithField("boringcrypto", boringEnabled()).
		Info("Nebula interface is active")

	if f.routines > 1 {
		if !f.inside.SupportsMultiqueue() || !f.outside.SupportsMultipleReaders() {
			f.routines = 1
			f.l.Warn("routines is not supported on this platform, falling back to a single routine")
		}
	}

	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))

	// Prepare n tun queues
	var reader overlay.BatchReadWriter = f.inside
	for i := 0; i < f.routines; i++ {
		if i > 0 {
			reader, err = f.inside.NewMultiQueueReader()
			if err != nil {
				f.l.Fatal(err)
			}
		}
		f.readers[i] = reader
	}

	if err := f.inside.Activate(); err != nil {
		f.inside.Close()
		f.l.Fatal(err)
	}
}

func (f *Interface) run() {
	// Launch n queues to read packets from udp
	for i := 0; i < f.routines; i++ {
		go f.listenOut(i)
	}

	// Launch n queues to read packets from tun dev
	for i := 0; i < f.routines; i++ {
		go f.listenIn(f.readers[i], i)
	}
}

func (f *Interface) listenOut(i int) {
	runtime.LockOSThread()

	var li udp.Conn
	if i > 0 {
		li = f.writers[i]
	} else {
		li = f.outside
	}

	ctCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
	lhh := f.lightHouse.NewRequestHandler()

	// Pre-allocate output buffers for batch processing
	batchSize := li.BatchSize()
	outs := make([][]byte, batchSize)
	for idx := range outs {
		// Allocate full buffer with virtio header space
		outs[idx] = make([]byte, virtioNetHdrLen, virtioNetHdrLen+udp.MTU)
	}

	h := &header.H{}
	fwPacket := &firewall.Packet{}
	nb := make([]byte, 12)

	li.ListenOutBatch(func(addrs []netip.AddrPort, payloads [][]byte, count int) {
		f.readOutsidePacketsBatch(addrs, payloads, count, outs[:count], nb, i, h, fwPacket, lhh, ctCache.Get(f.l))

	})
}

func (f *Interface) listenIn(reader overlay.BatchReadWriter, i int) {
	runtime.LockOSThread()

	batchSize := reader.BatchSize()

	// Allocate buffers for batch reading
	bufs := make([][]byte, batchSize)
	for idx := range bufs {
		bufs[idx] = make([]byte, mtu)
	}
	sizes := make([]int, batchSize)

	// Allocate output buffers for batch processing (one per packet)
	// Each has virtio header headroom to avoid copies on write
	outs := make([][]byte, batchSize)
	for idx := range outs {
		outBuf := make([]byte, virtioNetHdrLen+mtu)
		outs[idx] = outBuf[virtioNetHdrLen:] // Slice starting after headroom
	}

	// Pre-allocate batch accumulation buffers for sending
	batchPackets := make([][]byte, 0, batchSize)
	batchAddrs := make([]netip.AddrPort, 0, batchSize)

	// Pre-allocate nonce buffer (reused for all encryptions)
	nb := make([]byte, 12)

	conntrackCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)

	for {
		n, err := reader.BatchRead(bufs, sizes)
		if err != nil {
			if errors.Is(err, os.ErrClosed) && f.closed.Load() {
				return
			}

			f.l.WithError(err).Error("Error while batch reading outbound packets")
			// This only seems to happen when something fatal happens to the fd, so exit.
			os.Exit(2)
		}

		// Process all packets in the batch at once
		f.consumeInsidePackets(bufs, sizes, n, outs, nb, i, conntrackCache.Get(f.l), &batchPackets, &batchAddrs)
	}
}

func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
	c.RegisterReloadCallback(f.reloadFirewall)
	c.RegisterReloadCallback(f.reloadSendRecvError)
	c.RegisterReloadCallback(f.reloadAcceptRecvError)
	c.RegisterReloadCallback(f.reloadDisconnectInvalid)
	c.RegisterReloadCallback(f.reloadMisc)

	for _, udpConn := range f.writers {
		c.RegisterReloadCallback(udpConn.ReloadConfig)
	}
}

func (f *Interface) reloadDisconnectInvalid(c *config.C) {
	initial := c.InitialLoad()
	if initial || c.HasChanged("pki.disconnect_invalid") {
		f.disconnectInvalid.Store(c.GetBool("pki.disconnect_invalid", true))
		if !initial {
			f.l.Infof("pki.disconnect_invalid changed to %v", f.disconnectInvalid.Load())
		}
	}
}

func (f *Interface) reloadFirewall(c *config.C) {
	//TODO: need to trigger/detect if the certificate changed too
	if c.HasChanged("firewall") == false {
		f.l.Debug("No firewall config change detected")
		return
	}

	fw, err := NewFirewallFromConfig(f.l, f.pki.getCertState(), c)
	if err != nil {
		f.l.WithError(err).Error("Error while creating firewall during reload")
		return
	}

	oldFw := f.firewall
	conntrack := oldFw.Conntrack
	conntrack.Lock()
	defer conntrack.Unlock()

	fw.rulesVersion = oldFw.rulesVersion + 1
	// If rulesVersion is back to zero, we have wrapped all the way around. Be
	// safe and just reset conntrack in this case.
	if fw.rulesVersion == 0 {
		f.l.WithField("firewallHashes", fw.GetRuleHashes()).
			WithField("oldFirewallHashes", oldFw.GetRuleHashes()).
			WithField("rulesVersion", fw.rulesVersion).
			Warn("firewall rulesVersion has overflowed, resetting conntrack")
	} else {
		fw.Conntrack = conntrack
	}

	f.firewall = fw

	oldFw.Destroy()
	f.l.WithField("firewallHashes", fw.GetRuleHashes()).
		WithField("oldFirewallHashes", oldFw.GetRuleHashes()).
		WithField("rulesVersion", fw.rulesVersion).
		Info("New firewall has been installed")
}

func (f *Interface) reloadSendRecvError(c *config.C) {
	if c.InitialLoad() || c.HasChanged("listen.send_recv_error") {
		stringValue := c.GetString("listen.send_recv_error", "always")

		switch stringValue {
		case "always":
			f.sendRecvErrorConfig = recvErrorAlways
		case "never":
			f.sendRecvErrorConfig = recvErrorNever
		case "private":
			f.sendRecvErrorConfig = recvErrorPrivate
		default:
			if c.GetBool("listen.send_recv_error", true) {
				f.sendRecvErrorConfig = recvErrorAlways
			} else {
				f.sendRecvErrorConfig = recvErrorNever
			}
		}

		f.l.WithField("sendRecvError", f.sendRecvErrorConfig.String()).
			Info("Loaded send_recv_error config")
	}
}

func (f *Interface) reloadAcceptRecvError(c *config.C) {
	if c.InitialLoad() || c.HasChanged("listen.accept_recv_error") {
		stringValue := c.GetString("listen.accept_recv_error", "always")

		switch stringValue {
		case "always":
			f.acceptRecvErrorConfig = recvErrorAlways
		case "never":
			f.acceptRecvErrorConfig = recvErrorNever
		case "private":
			f.acceptRecvErrorConfig = recvErrorPrivate
		default:
			if c.GetBool("listen.accept_recv_error", true) {
				f.acceptRecvErrorConfig = recvErrorAlways
			} else {
				f.acceptRecvErrorConfig = recvErrorNever
			}
		}

		f.l.WithField("acceptRecvError", f.acceptRecvErrorConfig.String()).
			Info("Loaded accept_recv_error config")
	}
}

func (f *Interface) reloadMisc(c *config.C) {
	if c.HasChanged("counters.try_promote") {
		n := c.GetUint32("counters.try_promote", defaultPromoteEvery)
		f.tryPromoteEvery.Store(n)
		f.l.Info("counters.try_promote has changed")
	}

	if c.HasChanged("counters.requery_every_packets") {
		n := c.GetUint32("counters.requery_every_packets", defaultReQueryEvery)
		f.reQueryEvery.Store(n)
		f.l.Info("counters.requery_every_packets has changed")
	}

	if c.HasChanged("timers.requery_wait_duration") {
		n := c.GetDuration("timers.requery_wait_duration", defaultReQueryWait)
		f.reQueryWait.Store(int64(n))
		f.l.Info("timers.requery_wait_duration has changed")
	}
}

func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
	ticker := time.NewTicker(i)
	defer ticker.Stop()

	udpStats := udp.NewUDPStatsEmitter(f.writers)

	certExpirationGauge := metrics.GetOrRegisterGauge("certificate.ttl_seconds", nil)
	certInitiatingVersion := metrics.GetOrRegisterGauge("certificate.initiating_version", nil)
	certMaxVersion := metrics.GetOrRegisterGauge("certificate.max_version", nil)

	for {
		select {
		case <-ctx.Done():
			return
		case <-ticker.C:
			f.firewall.EmitStats()
			f.handshakeManager.EmitStats()
			udpStats()

			certState := f.pki.getCertState()
			defaultCrt := certState.GetDefaultCertificate()
			certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
			certInitiatingVersion.Update(int64(defaultCrt.Version()))

			// Report the max certificate version we are capable of using
			if certState.v2Cert != nil {
				certMaxVersion.Update(int64(certState.v2Cert.Version()))
			} else {
				certMaxVersion.Update(int64(certState.v1Cert.Version()))
			}
		}
	}
}

func (f *Interface) GetHostInfo(vpnIp netip.Addr) *HostInfo {
	return f.hostMap.QueryVpnAddr(vpnIp)
}

func (f *Interface) GetCertState() *CertState {
	return f.pki.getCertState()
}

func (f *Interface) Close() error {
	f.closed.Store(true)

	for _, u := range f.writers {
		err := u.Close()
		if err != nil {
			f.l.WithError(err).Error("Error while closing udp socket")
		}
	}
	for i, r := range f.readers {
		if i == 0 {
			continue // f.readers[0] is f.inside, which we want to save for last
		}
		if err := r.Close(); err != nil {
			f.l.WithError(err).Error("Error while closing tun reader")
		}
	}

	// Release the tun device
	return f.inside.Close()
}