kademlia.go

// Copyright 2020 The Swarm Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package kademlia

import (
	"context"
	random "crypto/rand"
	"encoding/json"
	"errors"
	"math"
	"math/big"
	"math/bits"
	"sync"
	"time"

	"github.com/ethsana/sana/pkg/addressbook"
	"github.com/ethsana/sana/pkg/discovery"
	"github.com/ethsana/sana/pkg/logging"
	"github.com/ethsana/sana/pkg/p2p"
	"github.com/ethsana/sana/pkg/shed"
	"github.com/ethsana/sana/pkg/swarm"
	"github.com/ethsana/sana/pkg/topology"
	im "github.com/ethsana/sana/pkg/topology/kademlia/internal/metrics"
	"github.com/ethsana/sana/pkg/topology/kademlia/internal/waitnext"
	"github.com/ethsana/sana/pkg/topology/pslice"
	ma "github.com/multiformats/go-multiaddr"
)

const (
	nnLowWatermark         = 2 // the number of peers in consecutive deepest bins that constitute as nearest neighbours
	maxConnAttempts        = 1 // when there is maxConnAttempts failed connect calls for a given peer it is considered non-connectable
	maxBootNodeAttempts    = 3 // how many attempts to dial to boot-nodes before giving up
	defaultBitSuffixLength = 3 // the number of bits used to create pseudo addresses for balancing

	addPeerBatchSize = 500

	peerConnectionAttemptTimeout = 5 * time.Second // Timeout for establishing a new connection with peer.
)

var (
	quickSaturationPeers        = 4
	saturationPeers             = 8
	overSaturationPeers         = 20
	bootNodeOverSaturationPeers = 20
	shortRetry                  = 30 * time.Second
	timeToRetry                 = 2 * shortRetry
	broadcastBinSize            = 4
)

var (
	errOverlayMismatch = errors.New("overlay mismatch")
	errPruneEntry      = errors.New("prune entry")
	errEmptyBin        = errors.New("empty bin")
)

type (
	binSaturationFunc  func(bin uint8, peers, connected *pslice.PSlice) (saturated bool, oversaturated bool)
	sanctionedPeerFunc func(peer swarm.Address) bool
)

var noopSanctionedPeerFn = func(_ swarm.Address) bool { return false }

// Options for injecting services to Kademlia.
type Options struct {
	SaturationFunc  binSaturationFunc
	Bootnodes       []ma.Multiaddr
	StandaloneMode  bool
	BootnodeMode    bool
	BitSuffixLength int
}

// Kad is the Swarm forwarding kademlia implementation.
type Kad struct {
	base              swarm.Address         // this node's overlay address
	discovery         discovery.Driver      // the discovery driver
	addressBook       addressbook.Interface // address book to get underlays
	p2p               p2p.Service           // p2p service to connect to nodes with
	saturationFunc    binSaturationFunc     // pluggable saturation function
	bitSuffixLength   int                   // additional depth of common prefix for bin
	commonBinPrefixes [][]swarm.Address     // list of address prefixes for each bin
	connectedPeers    *pslice.PSlice        // a slice of peers sorted and indexed by po, indexes kept in `bins`
	knownPeers        *pslice.PSlice        // both are po aware slice of addresses
	bootnodes         []ma.Multiaddr
	depth             uint8         // current neighborhood depth
	radius            uint8         // storage area of responsibility
	depthMu           sync.RWMutex  // protect depth changes
	manageC           chan struct{} // trigger the manage forever loop to connect to new peers
	peerSig           []chan struct{}
	peerSigMtx        sync.Mutex
	logger            logging.Logger // logger
	standalone        bool           // indicates whether the node is working in standalone mode
	bootnode          bool           // indicates whether the node is working in bootnode mode
	collector         *im.Collector
	quit              chan struct{} // quit channel
	halt              chan struct{} // halt channel
	done              chan struct{} // signal that `manage` has quit
	wg                sync.WaitGroup
	waitNext          *waitnext.WaitNext
	metrics           metrics
}

// New returns a new Kademlia.
func New(
	base swarm.Address,
	addressbook addressbook.Interface,
	discovery discovery.Driver,
	p2p p2p.Service,
	metricsDB *shed.DB,
	logger logging.Logger,
	o Options,
) *Kad {
	if o.SaturationFunc == nil {
		os := overSaturationPeers
		if o.BootnodeMode {
			os = bootNodeOverSaturationPeers
		}
		o.SaturationFunc = binSaturated(os)
	}
	if o.BitSuffixLength == 0 {
		o.BitSuffixLength = defaultBitSuffixLength
	}

	k := &Kad{
		base:              base,
		discovery:         discovery,
		addressBook:       addressbook,
		p2p:               p2p,
		saturationFunc:    o.SaturationFunc,
		bitSuffixLength:   o.BitSuffixLength,
		commonBinPrefixes: make([][]swarm.Address, int(swarm.MaxBins)),
		connectedPeers:    pslice.New(int(swarm.MaxBins), base),
		knownPeers:        pslice.New(int(swarm.MaxBins), base),
		bootnodes:         o.Bootnodes,
		manageC:           make(chan struct{}, 1),
		waitNext:          waitnext.New(),
		logger:            logger,
		standalone:        o.StandaloneMode,
		bootnode:          o.BootnodeMode,
		collector:         im.NewCollector(metricsDB),
		quit:              make(chan struct{}),
		halt:              make(chan struct{}),
		done:              make(chan struct{}),
		wg:                sync.WaitGroup{},
		metrics:           newMetrics(),
	}

	if k.bitSuffixLength > 0 {
		k.generateCommonBinPrefixes()
	}

	return k
}

func (k *Kad) generateCommonBinPrefixes() {
	bitCombinationsCount := int(math.Pow(2, float64(k.bitSuffixLength)))
	bitSufixes := make([]uint8, bitCombinationsCount)

	for i := 0; i < bitCombinationsCount; i++ {
		bitSufixes[i] = uint8(i)
	}

	addr := swarm.MustParseHexAddress(k.base.String())
	addrBytes := addr.Bytes()
	_ = addrBytes

	binPrefixes := k.commonBinPrefixes

	// copy base address
	for i := range binPrefixes {
		binPrefixes[i] = make([]swarm.Address, bitCombinationsCount)
	}

	for i := range binPrefixes {
		for j := range binPrefixes[i] {
			pseudoAddrBytes := make([]byte, len(k.base.Bytes()))
			copy(pseudoAddrBytes, k.base.Bytes())
			binPrefixes[i][j] = swarm.NewAddress(pseudoAddrBytes)
		}
	}

	for i := range binPrefixes {
		for j := range binPrefixes[i] {
			pseudoAddrBytes := binPrefixes[i][j].Bytes()

			// flip first bit for bin
			indexByte, posBit := i/8, i%8
			if hasBit(bits.Reverse8(pseudoAddrBytes[indexByte]), uint8(posBit)) {
				pseudoAddrBytes[indexByte] = bits.Reverse8(clearBit(bits.Reverse8(pseudoAddrBytes[indexByte]), uint8(posBit)))
			} else {
				pseudoAddrBytes[indexByte] = bits.Reverse8(setBit(bits.Reverse8(pseudoAddrBytes[indexByte]), uint8(posBit)))
			}

			// set pseudo suffix
			bitSuffixPos := k.bitSuffixLength - 1
			for l := i + 1; l < i+k.bitSuffixLength+1; l++ {
				index, pos := l/8, l%8

				if hasBit(bitSufixes[j], uint8(bitSuffixPos)) {
					pseudoAddrBytes[index] = bits.Reverse8(setBit(bits.Reverse8(pseudoAddrBytes[index]), uint8(pos)))
				} else {
					pseudoAddrBytes[index] = bits.Reverse8(clearBit(bits.Reverse8(pseudoAddrBytes[index]), uint8(pos)))
				}

				bitSuffixPos--
			}

			// clear rest of the bits
			for l := i + k.bitSuffixLength + 1; l < len(pseudoAddrBytes)*8; l++ {
				index, pos := l/8, l%8
				pseudoAddrBytes[index] = bits.Reverse8(clearBit(bits.Reverse8(pseudoAddrBytes[index]), uint8(pos)))
			}
		}
	}
}

// Clears the bit at pos in n.
func clearBit(n, pos uint8) uint8 {
	mask := ^(uint8(1) << pos)
	return n & mask
}

// Sets the bit at pos in the integer n.
func setBit(n, pos uint8) uint8 {
	return n | 1<<pos
}

func hasBit(n, pos uint8) bool {
	return n&(1<<pos) > 0
}

// peerConnInfo groups necessary fields needed to create a connection.
type peerConnInfo struct {
	po   uint8
	addr swarm.Address
}

// connectBalanced attempts to connect to the balanced peers first.
func (k *Kad) connectBalanced(wg *sync.WaitGroup, peerConnChan chan<- *peerConnInfo) {
	skipPeers := func(peer swarm.Address) bool {
		if k.waitNext.Waiting(peer) {
			k.metrics.TotalBeforeExpireWaits.Inc()
			return true
		}
		return false
	}

	for i := range k.commonBinPrefixes {
		if i >= int(k.NeighborhoodDepth()) {
			continue
		}
		for j := range k.commonBinPrefixes[i] {
			pseudoAddr := k.commonBinPrefixes[i][j]

			closestConnectedPeer, err := closestPeer(k.connectedPeers, pseudoAddr, noopSanctionedPeerFn)
			if err != nil {
				if errors.Is(err, topology.ErrNotFound) {
					break
				}
				k.logger.Errorf("closest connected peer: %v", err)
				continue
			}

			closestConnectedPO := swarm.ExtendedProximity(closestConnectedPeer.Bytes(), pseudoAddr.Bytes())
			if int(closestConnectedPO) >= i+k.bitSuffixLength+1 {
				continue
			}

			// Connect to closest known peer which we haven't tried connecting to recently.
			closestKnownPeer, err := closestPeer(k.knownPeers, pseudoAddr, skipPeers)
			if err != nil {
				if errors.Is(err, topology.ErrNotFound) {
					break
				}
				k.logger.Errorf("closest known peer: %v", err)
				continue
			}

			if k.connectedPeers.Exists(closestKnownPeer) {
				continue
			}

			closestKnownPeerPO := swarm.ExtendedProximity(closestKnownPeer.Bytes(), pseudoAddr.Bytes())
			if int(closestKnownPeerPO) < i+k.bitSuffixLength+1 {
				continue
			}

			select {
			case <-k.quit:
				return
			default:
				wg.Add(1)
				peerConnChan <- &peerConnInfo{
					po:   swarm.Proximity(k.base.Bytes(), closestKnownPeer.Bytes()),
					addr: closestKnownPeer,
				}
			}
			break
		}
	}
}

// connectNeighbours attempts to connect to the neighbours
// which were not considered by the connectBalanced method.
func (k *Kad) connectNeighbours(wg *sync.WaitGroup, peerConnChan, peerConnChan2 chan<- *peerConnInfo) {
	const multiplePeerThreshold = 8

	sent := 0
	_ = k.knownPeers.EachBinRev(func(addr swarm.Address, po uint8) (bool, bool, error) {
		depth := k.NeighborhoodDepth()

		if depth > po || po >= depth+multiplePeerThreshold {
			return false, true, nil
		}

		if len(k.connectedPeers.BinPeers(po)) >= overSaturationPeers-1 {
			return false, true, nil
		}

		if k.connectedPeers.Exists(addr) {
			return false, false, nil
		}

		if k.waitNext.Waiting(addr) {
			k.metrics.TotalBeforeExpireWaits.Inc()
			return false, false, nil
		}

		select {
		case <-k.quit:
			return true, false, nil
		default:
			wg.Add(1)
			peerConnChan <- &peerConnInfo{
				po:   po,
				addr: addr,
			}
			sent++
		}

		// We want to sent number of attempts equal to saturationPeers
		// in order to speed up the topology build.
		next := sent == saturationPeers
		if next {
			sent = 0
		}
		return false, next, nil
	})

	_ = k.knownPeers.EachBinRev(func(addr swarm.Address, po uint8) (bool, bool, error) {
		depth := k.NeighborhoodDepth()

		if po < depth+multiplePeerThreshold {
			return false, true, nil
		}

		if k.connectedPeers.Exists(addr) {
			return false, false, nil
		}

		if k.waitNext.Waiting(addr) {
			k.metrics.TotalBeforeExpireWaits.Inc()
			return false, false, nil
		}

		select {
		case <-k.quit:
			return true, false, nil
		default:
			wg.Add(1)
			peerConnChan2 <- &peerConnInfo{
				po:   po,
				addr: addr,
			}
		}

		// The bin could be saturated or not, so a decision cannot
		// be made before checking the next peer, so we iterate to next.
		return false, true, nil
	})
}

// connectionAttemptsHandler handles the connection attempts
// to peers sent by the producers to the peerConnChan.
func (k *Kad) connectionAttemptsHandler(ctx context.Context, wg *sync.WaitGroup, peerConnChan, peerConnChan2 <-chan *peerConnInfo) {
	connect := func(peer *peerConnInfo) {
		bzzAddr, err := k.addressBook.Get(peer.addr)
		switch {
		case errors.Is(err, addressbook.ErrNotFound):
			k.logger.Debugf("kademlia: empty address book entry for peer %q", peer.addr)
			k.knownPeers.Remove(peer.addr)
			return
		case err != nil:
			k.logger.Debugf("kademlia: failed to get address book entry for peer %q: %v", peer.addr, err)
			return
		}

		remove := func(peer *peerConnInfo) {
			k.waitNext.Remove(peer.addr)
			k.knownPeers.Remove(peer.addr)
			if err := k.addressBook.Remove(peer.addr); err != nil {
				k.logger.Debugf("kademlia: could not remove peer %q from addressbook", peer.addr)
			}
		}

		switch err = k.connect(ctx, peer.addr, bzzAddr.Underlay); {
		case errors.Is(err, errPruneEntry):
			k.logger.Debugf("kademlia: dial to light node with overlay %q and underlay %q", peer.addr, bzzAddr.Underlay)
			remove(peer)
			return
		case errors.Is(err, errOverlayMismatch):
			k.logger.Debugf("kademlia: overlay mismatch has occurred to an overlay %q with underlay %q", peer.addr, bzzAddr.Underlay)
			remove(peer)
			return
		case err != nil:
			k.logger.Debugf("kademlia: peer not reachable from kademlia %q: %v", bzzAddr, err)
			k.logger.Warningf("peer not reachable when attempting to connect")
			return
		}

		k.waitNext.Set(peer.addr, time.Now().Add(shortRetry), 0)

		k.connectedPeers.Add(peer.addr)

		k.metrics.TotalOutboundConnections.Inc()
		k.collector.Record(peer.addr, im.PeerLogIn(time.Now(), im.PeerConnectionDirectionOutbound))

		k.depthMu.Lock()
		k.depth = recalcDepth(k.connectedPeers, k.radius)
		k.depthMu.Unlock()

		k.logger.Debugf("kademlia: connected to peer: %q in bin: %d", peer.addr, peer.po)
		k.notifyManageLoop()
		k.notifyPeerSig()
	}

	var (
		// The inProgress helps to avoid making a connection
		// to a peer who has the connection already in progress.
		inProgress   = make(map[string]bool)
		inProgressMu sync.Mutex
	)
	connAttempt := func(peerConnChan <-chan *peerConnInfo) {
		for {
			select {
			case <-k.quit:
				return
			case peer := <-peerConnChan:
				addr := peer.addr.String()

				if k.waitNext.Waiting(peer.addr) {
					k.metrics.TotalBeforeExpireWaits.Inc()
					wg.Done()
					continue
				}

				inProgressMu.Lock()
				if !inProgress[addr] {
					inProgress[addr] = true
					inProgressMu.Unlock()
					connect(peer)
					inProgressMu.Lock()
					delete(inProgress, addr)
				}
				inProgressMu.Unlock()
				wg.Done()
			}
		}
	}
	for i := 0; i < 64; i++ {
		go connAttempt(peerConnChan)
	}
	for i := 0; i < 8; i++ {
		go connAttempt(peerConnChan2)
	}
}

// notifyManageLoop notifies kademlia manage loop.
func (k *Kad) notifyManageLoop() {
	select {
	case k.manageC <- struct{}{}:
	default:
	}
}

// manage is a forever loop that manages the connection to new peers
// once they get added or once others leave.
func (k *Kad) manage() {
	defer k.wg.Done()
	defer close(k.done)
	defer k.logger.Debugf("kademlia manage loop exited")

	ctx, cancel := context.WithCancel(context.Background())
	go func() {
		<-k.quit
		cancel()
	}()

	// The wg makes sure that we wait for all the connection attempts,
	// spun up by goroutines, to finish before we try the boot-nodes.
	var wg sync.WaitGroup
	peerConnChan := make(chan *peerConnInfo)
	peerConnChan2 := make(chan *peerConnInfo)
	go k.connectionAttemptsHandler(ctx, &wg, peerConnChan, peerConnChan2)

	for {
		select {
		case <-k.quit:
			return
		case <-time.After(15 * time.Second):
			start := time.Now()
			if err := k.collector.Flush(); err != nil {
				k.metrics.InternalMetricsFlushTotalErrors.Inc()
				k.logger.Debugf("kademlia: unable to flush metrics counters to the persistent store: %v", err)
			} else {
				k.metrics.InternalMetricsFlushTime.Observe(float64(time.Since(start).Nanoseconds()))
			}
			k.notifyManageLoop()
		case <-k.manageC:
			start := time.Now()

			select {
			case <-k.halt:
				// halt stops dial-outs while shutting down
				return
			case <-k.quit:
				return
			default:
			}

			if k.standalone {
				continue
			}

			oldDepth := k.NeighborhoodDepth()
			k.connectNeighbours(&wg, peerConnChan, peerConnChan2)
			k.connectBalanced(&wg, peerConnChan2)
			wg.Wait()

			k.depthMu.Lock()
			depth := k.depth
			radius := k.radius
			k.depthMu.Unlock()

			k.logger.Tracef(
				"kademlia: connector took %s to finish: old depth %d; new depth %d",
				time.Since(start),
				oldDepth,
				depth,
			)

			k.metrics.CurrentDepth.Set(float64(depth))
			k.metrics.CurrentRadius.Set(float64(radius))
			k.metrics.CurrentlyKnownPeers.Set(float64(k.knownPeers.Length()))
			k.metrics.CurrentlyConnectedPeers.Set(float64(k.connectedPeers.Length()))

			if k.connectedPeers.Length() == 0 {
				select {
				case <-k.halt:
					continue
				default:
				}
				k.logger.Debug("kademlia: no connected peers, trying bootnodes")
				k.connectBootNodes(ctx)
			}
		}
	}
}

func (k *Kad) Start(_ context.Context) error {
	k.wg.Add(1)
	go k.manage()

	go func() {
		select {
		case <-k.halt:
			return
		case <-k.quit:
			return
		default:
		}
		var (
			start     = time.Now()
			addresses []swarm.Address
		)

		err := k.addressBook.IterateOverlays(func(addr swarm.Address) (stop bool, err error) {
			addresses = append(addresses, addr)
			if len(addresses) == addPeerBatchSize {
				k.AddPeers(addresses...)
				addresses = nil
			}
			return false, nil
		})
		if err != nil {
			k.logger.Errorf("addressbook overlays: %w", err)
			return
		}
		k.AddPeers(addresses...)
		k.metrics.StartAddAddressBookOverlaysTime.Observe(float64(time.Since(start).Nanoseconds()))
	}()

	// trigger the first manage loop immediately so that
	// we can start connecting to the bootnode quickly
	k.notifyManageLoop()

	return nil
}

func (k *Kad) connectBootNodes(ctx context.Context) {
	var attempts, connected int
	totalAttempts := maxBootNodeAttempts * len(k.bootnodes)

	ctx, cancel := context.WithTimeout(ctx, 15*time.Second)
	defer cancel()

	for _, addr := range k.bootnodes {
		if attempts >= totalAttempts || connected >= 3 {
			return
		}

		if _, err := p2p.Discover(ctx, addr, func(addr ma.Multiaddr) (stop bool, err error) {
			k.logger.Tracef("connecting to bootnode %s", addr)
			if attempts >= maxBootNodeAttempts {
				return true, nil
			}
			bzzAddress, err := k.p2p.Connect(ctx, addr)

			attempts++
			k.metrics.TotalBootNodesConnectionAttempts.Inc()

			if err != nil {
				if !errors.Is(err, p2p.ErrAlreadyConnected) {
					k.logger.Debugf("connect fail %s: %v", addr, err)
					k.logger.Warningf("connect to bootnode %s", addr)
					return false, err
				}
				k.logger.Debugf("connect to bootnode fail: %v", err)
				return false, nil
			}

			if err := k.connected(ctx, bzzAddress.Overlay); err != nil {
				return false, err
			}
			k.logger.Tracef("connected to bootnode %s", addr)
			connected++
			// connect to max 3 bootnodes
			return connected >= 3, nil
		}); err != nil && !errors.Is(err, context.Canceled) {
			k.logger.Debugf("discover fail %s: %v", addr, err)
			k.logger.Warningf("discover to bootnode %s", addr)
			return
		}
	}
}

// binSaturated indicates whether a certain bin is saturated or not.
// when a bin is not saturated it means we would like to proactively
// initiate connections to other peers in the bin.
func binSaturated(oversaturationAmount int) binSaturationFunc {
	return func(bin uint8, peers, connected *pslice.PSlice) (bool, bool) {
		potentialDepth := recalcDepth(peers, swarm.MaxPO)

		// short circuit for bins which are >= depth
		if bin >= potentialDepth {
			return false, false
		}

		// lets assume for now that the minimum number of peers in a bin
		// would be 2, under which we would always want to connect to new peers
		// obviously this should be replaced with a better optimization
		// the iterator is used here since when we check if a bin is saturated,
		// the plain number of size of bin might not suffice (for example for squared
		// gaps measurement)

		size := 0
		_ = connected.EachBin(func(_ swarm.Address, po uint8) (bool, bool, error) {
			if po == bin {
				size++
			}
			return false, false, nil
		})

		return size >= saturationPeers, size >= oversaturationAmount
	}
}

// recalcDepth calculates and returns the kademlia depth.
func recalcDepth(peers *pslice.PSlice, radius uint8) uint8 {
	// handle edge case separately
	if peers.Length() <= nnLowWatermark {
		return 0
	}
	var (
		peersCtr                     = uint(0)
		candidate                    = uint8(0)
		shallowestEmpty, noEmptyBins = peers.ShallowestEmpty()
	)

	shallowestUnsaturated := uint8(0)
	binCount := 0
	_ = peers.EachBinRev(func(_ swarm.Address, bin uint8) (bool, bool, error) {
		if bin == shallowestUnsaturated {
			binCount++
			return false, false, nil
		}
		if bin > shallowestUnsaturated && binCount < quickSaturationPeers {
			// this means we have less than quickSaturation in the previous bin
			// therefore we can return assuming that bin is the unsaturated one.
			return true, false, nil
		}
		shallowestUnsaturated = bin
		binCount = 1

		return false, false, nil
	})

	// if there are some empty bins and the shallowestEmpty is
	// smaller than the shallowestUnsaturated then set shallowest
	// unsaturated to the empty bin.
	if !noEmptyBins && shallowestEmpty < shallowestUnsaturated {
		shallowestUnsaturated = shallowestEmpty
	}

	_ = peers.EachBin(func(_ swarm.Address, po uint8) (bool, bool, error) {
		peersCtr++
		if peersCtr >= nnLowWatermark {
			candidate = po
			return true, false, nil
		}
		return false, false, nil
	})
	if shallowestUnsaturated > candidate {
		if radius < candidate {
			return radius
		}
		return candidate
	}

	if radius < shallowestUnsaturated {
		return radius
	}
	return shallowestUnsaturated
}

// connect connects to a peer and gossips its address to our connected peers,
// as well as sends the peers we are connected to to the newly connected peer
func (k *Kad) connect(ctx context.Context, peer swarm.Address, ma ma.Multiaddr) error {
	k.logger.Infof("attempting to connect to peer %q", peer)

	ctx, cancel := context.WithTimeout(ctx, peerConnectionAttemptTimeout)
	defer cancel()

	k.metrics.TotalOutboundConnectionAttempts.Inc()

	switch i, err := k.p2p.Connect(ctx, ma); {
	case errors.Is(err, p2p.ErrDialLightNode):
		return errPruneEntry
	case errors.Is(err, p2p.ErrAlreadyConnected):
		if !i.Overlay.Equal(peer) {
			return errOverlayMismatch
		}
		return nil
	case errors.Is(err, context.Canceled):
		return err
	case err != nil:
		k.logger.Debugf("could not connect to peer %q: %v", peer, err)

		retryTime := time.Now().Add(timeToRetry)
		var e *p2p.ConnectionBackoffError
		failedAttempts := 0
		if errors.As(err, &e) {
			retryTime = e.TryAfter()
		} else {
			failedAttempts = k.waitNext.Attempts(peer)
			failedAttempts++
		}

		k.metrics.TotalOutboundConnectionFailedAttempts.Inc()
		k.collector.Record(peer, im.IncSessionConnectionRetry())

		k.collector.Inspect(peer, func(ss *im.Snapshot) {
			quickPrune := ss == nil || ss.HasAtMaxOneConnectionAttempt()
			if (k.connectedPeers.Length() > 0 && quickPrune) || failedAttempts >= maxConnAttempts {
				k.waitNext.Remove(peer)
				k.knownPeers.Remove(peer)
				if err := k.addressBook.Remove(peer); err != nil {
					k.logger.Debugf("could not remove peer from addressbook: %q", peer)
				}
				k.logger.Debugf("kademlia pruned peer from address book %q", peer)
			} else {
				k.waitNext.Set(peer, retryTime, failedAttempts)
			}
		})

		return err
	case !i.Overlay.Equal(peer):
		_ = k.p2p.Disconnect(peer)
		_ = k.p2p.Disconnect(i.Overlay)
		return errOverlayMismatch
	}

	return k.Announce(ctx, peer, true)
}

// Announce a newly connected peer to our connected peers, but also
// notify the peer about our already connected peers
func (k *Kad) Announce(ctx context.Context, peer swarm.Address, fullnode bool) error {
	var addrs []swarm.Address

	for bin := uint8(0); bin < swarm.MaxBins; bin++ {

		connectedPeers, err := randomSubset(k.connectedPeers.BinPeers(bin), broadcastBinSize)
		if err != nil {
			return err
		}

		for _, connectedPeer := range connectedPeers {
			if connectedPeer.Equal(peer) {
				continue
			}

			addrs = append(addrs, connectedPeer)

			if !fullnode {
				// we continue here so we dont gossip
				// about lightnodes to others.
				continue
			}
			go func(connectedPeer swarm.Address) {
				if err := k.discovery.BroadcastPeers(ctx, connectedPeer, peer); err != nil {
					k.logger.Debugf("could not gossip peer %s to peer %s: %v", peer, connectedPeer, err)
				}
			}(connectedPeer)
		}
	}

	if len(addrs) == 0 {
		return nil
	}

	err := k.discovery.BroadcastPeers(ctx, peer, addrs...)
	if err != nil {
		k.logger.Errorf("kademlia: could not broadcast to peer %s", peer)
		_ = k.p2p.Disconnect(peer)
	}

	return err
}

// AddPeers adds peers to the knownPeers list.
// This does not guarantee that a connection will immediately
// be made to the peer.
func (k *Kad) AddPeers(addrs ...swarm.Address) {
	k.knownPeers.Add(addrs...)
	k.notifyManageLoop()
}

func (k *Kad) Pick(peer p2p.Peer) bool {
	k.metrics.PickCalls.Inc()
	if k.bootnode {
		// shortcircuit for bootnode mode - always accept connections,
		// at least until we find a better solution.
		return true
	}
	po := swarm.Proximity(k.base.Bytes(), peer.Address.Bytes())
	_, oversaturated := k.saturationFunc(po, k.knownPeers, k.connectedPeers)
	// pick the peer if we are not oversaturated
	if !oversaturated {
		return true
	}
	k.metrics.PickCallsFalse.Inc()
	return false
}

// Connected is called when a peer has dialed in.
// If forceConnection is true `overSaturated` is ignored for non-bootnodes.
func (k *Kad) Connected(ctx context.Context, peer p2p.Peer, forceConnection bool) error {
	address := peer.Address
	po := swarm.Proximity(k.base.Bytes(), address.Bytes())

	if _, overSaturated := k.saturationFunc(po, k.knownPeers, k.connectedPeers); overSaturated {
		if k.bootnode {
			randPeer, err := k.randomPeer(po)
			if err != nil {
				return err
			}
			_ = k.p2p.Disconnect(randPeer)
			return k.connected(ctx, address)
		}
		if !forceConnection {
			return topology.ErrOversaturated
		}
	}

	return k.connected(ctx, address)
}

func (k *Kad) connected(ctx context.Context, addr swarm.Address) error {
	if err := k.Announce(ctx, addr, true); err != nil {
		return err
	}

	k.knownPeers.Add(addr)
	k.connectedPeers.Add(addr)

	k.metrics.TotalInboundConnections.Inc()
	k.collector.Record(addr, im.PeerLogIn(time.Now(), im.PeerConnectionDirectionInbound))

	k.waitNext.Remove(addr)

	k.depthMu.Lock()
	k.depth = recalcDepth(k.connectedPeers, k.radius)
	k.depthMu.Unlock()

	k.notifyManageLoop()
	k.notifyPeerSig()
	return nil
}

// Disconnected is called when peer disconnects.
func (k *Kad) Disconnected(peer p2p.Peer) {
	k.logger.Debugf("kademlia: disconnected peer %s", peer.Address)

	k.connectedPeers.Remove(peer.Address)

	k.waitNext.SetTryAfter(peer.Address, time.Now().Add(timeToRetry))

	k.metrics.TotalInboundDisconnections.Inc()
	k.collector.Record(peer.Address, im.PeerLogOut(time.Now()))

	k.depthMu.Lock()
	k.depth = recalcDepth(k.connectedPeers, k.radius)
	k.depthMu.Unlock()

	k.notifyManageLoop()
	k.notifyPeerSig()
}

func (k *Kad) notifyPeerSig() {
	k.peerSigMtx.Lock()
	defer k.peerSigMtx.Unlock()

	for _, c := range k.peerSig {
		// Every peerSig channel has a buffer capacity of 1,
		// so every receiver will get the signal even if the
		// select statement has the default case to avoid blocking.
		select {
		case c <- struct{}{}:
		default:
		}
	}
}

func closestPeer(peers *pslice.PSlice, addr swarm.Address, spf sanctionedPeerFunc) (swarm.Address, error) {
	closest := swarm.ZeroAddress
	err := peers.EachBinRev(func(peer swarm.Address, po uint8) (bool, bool, error) {
		// check whether peer is sanctioned
		if spf(peer) {
			return false, false, nil
		}
		if closest.IsZero() {
			closest = peer
			return false, false, nil
		}
		dcmp, err := swarm.DistanceCmp(addr.Bytes(), closest.Bytes(), peer.Bytes())
		if err != nil {
			return false, false, err
		}
		switch dcmp {
		case 0:
			// do nothing
		case -1:
			// current peer is closer
			closest = peer
		case 1:
			// closest is already closer to chunk
			// do nothing
		}
		return false, false, nil
	})
	if err != nil {
		return swarm.ZeroAddress, err
	}

	// check if found
	if closest.IsZero() {
		return swarm.ZeroAddress, topology.ErrNotFound
	}

	return closest, nil
}

func isIn(a swarm.Address, addresses []p2p.Peer) bool {
	for _, v := range addresses {
		if v.Address.Equal(a) {
			return true
		}
	}
	return false
}

// ClosestPeer returns the closest peer to a given address.
func (k *Kad) ClosestPeer(addr swarm.Address, includeSelf bool, skipPeers ...swarm.Address) (swarm.Address, error) {
	if k.connectedPeers.Length() == 0 {
		return swarm.Address{}, topology.ErrNotFound
	}

	peers := k.p2p.Peers()
	var peersToDisconnect []swarm.Address
	closest := swarm.ZeroAddress

	if includeSelf {
		closest = k.base
	}

	err := k.connectedPeers.EachBinRev(func(peer swarm.Address, po uint8) (bool, bool, error) {
		if closest.IsZero() {
			closest = peer
		}

		for _, a := range skipPeers {
			if a.Equal(peer) {
				return false, false, nil
			}
		}

		// kludge: hotfix for topology peer inconsistencies bug
		if !isIn(peer, peers) {
			a := swarm.NewAddress(peer.Bytes())
			peersToDisconnect = append(peersToDisconnect, a)
			return false, false, nil
		}

		dcmp, err := swarm.DistanceCmp(addr.Bytes(), closest.Bytes(), peer.Bytes())
		if err != nil {
			return false, false, err
		}
		switch dcmp {
		case 0:
			// do nothing
		case -1:
			// current peer is closer
			closest = peer
		case 1:
			// closest is already closer to chunk
			// do nothing
		}
		return false, false, nil
	})
	if err != nil {
		return swarm.Address{}, err
	}

	if closest.IsZero() { // no peers
		return swarm.Address{}, topology.ErrNotFound // only for light nodes
	}

	for _, v := range peersToDisconnect {
		k.Disconnected(p2p.Peer{Address: v})
	}

	// check if self
	if closest.Equal(k.base) {
		return swarm.Address{}, topology.ErrWantSelf
	}

	return closest, nil
}

// IsWithinDepth returns if an address is within the neighborhood depth of a node.
func (k *Kad) IsWithinDepth(addr swarm.Address) bool {
	return swarm.Proximity(k.base.Bytes(), addr.Bytes()) >= k.NeighborhoodDepth()
}

// EachNeighbor iterates from closest bin to farthest of the neighborhood peers.
func (k *Kad) EachNeighbor(f topology.EachPeerFunc) error {
	depth := k.NeighborhoodDepth()
	fn := func(a swarm.Address, po uint8) (bool, bool, error) {
		if po < depth {
			return true, false, nil
		}
		return f(a, po)
	}
	return k.connectedPeers.EachBin(fn)
}

// EachNeighborRev iterates from farthest bin to closest of the neighborhood peers.
func (k *Kad) EachNeighborRev(f topology.EachPeerFunc) error {
	depth := k.NeighborhoodDepth()
	fn := func(a swarm.Address, po uint8) (bool, bool, error) {
		if po < depth {
			return false, true, nil
		}
		return f(a, po)
	}
	return k.connectedPeers.EachBinRev(fn)
}

// EachPeer iterates from closest bin to farthest.
func (k *Kad) EachPeer(f topology.EachPeerFunc) error {
	return k.connectedPeers.EachBin(f)
}

// EachPeerRev iterates from farthest bin to closest.
func (k *Kad) EachPeerRev(f topology.EachPeerFunc) error {
	return k.connectedPeers.EachBinRev(f)
}

// SubscribePeersChange returns the channel that signals when the connected peers
// set changes. Returned function is safe to be called multiple times.
func (k *Kad) SubscribePeersChange() (c <-chan struct{}, unsubscribe func()) {
	channel := make(chan struct{}, 1)
	var closeOnce sync.Once

	k.peerSigMtx.Lock()
	defer k.peerSigMtx.Unlock()

	k.peerSig = append(k.peerSig, channel)

	unsubscribe = func() {
		k.peerSigMtx.Lock()
		defer k.peerSigMtx.Unlock()

		for i, c := range k.peerSig {
			if c == channel {
				k.peerSig = append(k.peerSig[:i], k.peerSig[i+1:]...)
				break
			}
		}

		closeOnce.Do(func() { close(channel) })
	}

	return channel, unsubscribe
}

// NeighborhoodDepth returns the current Kademlia depth.
func (k *Kad) NeighborhoodDepth() uint8 {
	k.depthMu.RLock()
	defer k.depthMu.RUnlock()

	return k.depth
}

// IsBalanced returns if Kademlia is balanced to bin.
func (k *Kad) IsBalanced(bin uint8) bool {
	k.depthMu.RLock()
	defer k.depthMu.RUnlock()

	if int(bin) > len(k.commonBinPrefixes) {
		return false
	}

	// for each pseudo address
	for i := range k.commonBinPrefixes[bin] {
		pseudoAddr := k.commonBinPrefixes[bin][i]
		closestConnectedPeer, err := closestPeer(k.connectedPeers, pseudoAddr, noopSanctionedPeerFn)
		if err != nil {
			return false
		}

		closestConnectedPO := swarm.ExtendedProximity(closestConnectedPeer.Bytes(), pseudoAddr.Bytes())
		if int(closestConnectedPO) < int(bin)+k.bitSuffixLength+1 {
			return false
		}
	}

	return true
}

func (k *Kad) SetRadius(r uint8) {
	k.depthMu.Lock()
	defer k.depthMu.Unlock()
	if k.radius == r {
		return
	}
	k.radius = r
	oldD := k.depth
	k.depth = recalcDepth(k.connectedPeers, k.radius)
	if k.depth != oldD {
		k.notifyManageLoop()
	}
}

func (k *Kad) Snapshot() *topology.KadParams {
	var infos []topology.BinInfo
	for i := int(swarm.MaxPO); i >= 0; i-- {
		infos = append(infos, topology.BinInfo{})
	}

	ss := k.collector.Snapshot(time.Now())

	_ = k.connectedPeers.EachBin(func(addr swarm.Address, po uint8) (bool, bool, error) {
		infos[po].BinConnected++
		infos[po].ConnectedPeers = append(
			infos[po].ConnectedPeers,
			&topology.PeerInfo{
				Address: addr,
				Metrics: createMetricsSnapshotView(ss[addr.ByteString()]),
			},
		)
		return false, false, nil
	})

	// output (k.knownPeers ¬ k.connectedPeers) here to not repeat the peers we already have in the connected peers list
	_ = k.knownPeers.EachBin(func(addr swarm.Address, po uint8) (bool, bool, error) {
		infos[po].BinPopulation++

		for _, v := range infos[po].ConnectedPeers {
			// peer already connected, don't show in the known peers list
			if v.Address.Equal(addr) {
				return false, false, nil
			}
		}

		infos[po].DisconnectedPeers = append(
			infos[po].DisconnectedPeers,
			&topology.PeerInfo{
				Address: addr,
				Metrics: createMetricsSnapshotView(ss[addr.ByteString()]),
			},
		)
		return false, false, nil
	})

	return &topology.KadParams{
		Base:           k.base.String(),
		Population:     k.knownPeers.Length(),
		Connected:      k.connectedPeers.Length(),
		Timestamp:      time.Now(),
		NNLowWatermark: nnLowWatermark,
		Depth:          k.NeighborhoodDepth(),
		Bins: topology.KadBins{
			Bin0:  infos[0],
			Bin1:  infos[1],
			Bin2:  infos[2],
			Bin3:  infos[3],
			Bin4:  infos[4],
			Bin5:  infos[5],
			Bin6:  infos[6],
			Bin7:  infos[7],
			Bin8:  infos[8],
			Bin9:  infos[9],
			Bin10: infos[10],
			Bin11: infos[11],
			Bin12: infos[12],
			Bin13: infos[13],
			Bin14: infos[14],
			Bin15: infos[15],
			Bin16: infos[16],
			Bin17: infos[17],
			Bin18: infos[18],
			Bin19: infos[19],
			Bin20: infos[20],
			Bin21: infos[21],
			Bin22: infos[22],
			Bin23: infos[23],
			Bin24: infos[24],
			Bin25: infos[25],
			Bin26: infos[26],
			Bin27: infos[27],
			Bin28: infos[28],
			Bin29: infos[29],
			Bin30: infos[30],
			Bin31: infos[31],
		},
	}
}

// String returns a string represenstation of Kademlia.
func (k *Kad) String() string {
	j := k.Snapshot()
	b, err := json.MarshalIndent(j, "", "  ")
	if err != nil {
		k.logger.Errorf("could not marshal kademlia into json: %v", err)
		return ""
	}
	return string(b)
}

// Halt stops outgoing connections from happening.
// This is needed while we shut down, so that further topology
// changes do not happen while we shut down.
func (k *Kad) Halt() {
	close(k.halt)
}

// Close shuts down kademlia.
func (k *Kad) Close() error {
	k.logger.Info("kademlia shutting down")
	close(k.quit)
	cc := make(chan struct{})

	go func() {
		k.wg.Wait()
		close(cc)
	}()

	select {
	case <-cc:
	case <-time.After(peerConnectionAttemptTimeout):
		k.logger.Warning("kademlia shutting down with announce goroutines")
	}

	select {
	case <-k.done:
	case <-time.After(5 * time.Second):
		k.logger.Warning("kademlia manage loop did not shut down properly")
	}

	k.logger.Info("kademlia persisting peer metrics")
	if err := k.collector.Finalize(time.Now()); err != nil {
		k.logger.Debugf("kademlia: unable to finalize open sessions: %v", err)
	}

	return nil
}

func randomSubset(addrs []swarm.Address, count int) ([]swarm.Address, error) {
	if count >= len(addrs) {
		return addrs, nil
	}

	for i := 0; i < len(addrs); i++ {
		b, err := random.Int(random.Reader, big.NewInt(int64(len(addrs))))
		if err != nil {
			return nil, err
		}
		j := int(b.Int64())
		addrs[i], addrs[j] = addrs[j], addrs[i]
	}

	return addrs[:count], nil
}

func (k *Kad) randomPeer(bin uint8) (swarm.Address, error) {
	peers := k.connectedPeers.BinPeers(bin)

	if len(peers) == 0 {
		return swarm.ZeroAddress, errEmptyBin
	}

	rndIndx, err := random.Int(random.Reader, big.NewInt(int64(len(peers))))
	if err != nil {
		return swarm.ZeroAddress, err
	}

	return peers[rndIndx.Int64()], nil
}

// createMetricsSnapshotView creates new topology.MetricSnapshotView from the
// given metrics.Snapshot and rounds all the timestamps and durations to its
// nearest second.
func createMetricsSnapshotView(ss *im.Snapshot) *topology.MetricSnapshotView {
	if ss == nil {
		return nil
	}
	return &topology.MetricSnapshotView{
		LastSeenTimestamp:          time.Unix(0, ss.LastSeenTimestamp).Unix(),
		SessionConnectionRetry:     ss.SessionConnectionRetry,
		ConnectionTotalDuration:    ss.ConnectionTotalDuration.Truncate(time.Second).Seconds(),
		SessionConnectionDuration:  ss.SessionConnectionDuration.Truncate(time.Second).Seconds(),
		SessionConnectionDirection: string(ss.SessionConnectionDirection),
	}
}