path: root/p2p/discover/table.go

// Package discover implements the Node Discovery Protocol.
//
// The Node Discovery protocol provides a way to find RLPx nodes that
// can be connected to. It uses a Kademlia-like protocol to maintain a
// distributed database of the IDs and endpoints of all listening
// nodes.
package discover

import (
    "crypto/rand"
    "encoding/binary"
    "net"
    "sort"
    "sync"
    "time"

    "github.com/ethereum/go-ethereum/common"
    "github.com/ethereum/go-ethereum/crypto"
    "github.com/ethereum/go-ethereum/logger"
    "github.com/ethereum/go-ethereum/logger/glog"
)

const (
    alpha      = 3  // Kademlia concurrency factor
    bucketSize = 16 // Kademlia bucket size
    hashBits   = len(common.Hash{}) * 8
    nBuckets   = hashBits + 1 // Number of buckets

    maxBondingPingPongs = 16
    maxFindnodeFailures = 5
)

type Table struct {
    mutex   sync.Mutex        // protects buckets, their content, and nursery
    buckets [nBuckets]*bucket // index of known nodes by distance
    nursery []*Node           // bootstrap nodes
    db      *nodeDB           // database of known nodes

    bondmu    sync.Mutex
    bonding   map[NodeID]*bondproc
    bondslots chan struct{} // limits total number of active bonding processes

    net  transport
    self *Node // metadata of the local node
}

type bondproc struct {
    err  error
    n    *Node
    done chan struct{}
}

// transport is implemented by the UDP transport.
// it is an interface so we can test without opening lots of UDP
// sockets and without generating a private key.
type transport interface {
    ping(NodeID, *net.UDPAddr) error
    waitping(NodeID) error
    findnode(toid NodeID, addr *net.UDPAddr, target NodeID) ([]*Node, error)
    close()
}

// bucket contains nodes, ordered by their last activity.
// the entry that was most recently active is the last element
// in entries.
type bucket struct {
    lastLookup time.Time
    entries    []*Node
}

func newTable(t transport, ourID NodeID, ourAddr *net.UDPAddr, nodeDBPath string) *Table {
    // If no node database was given, use an in-memory one
    db, err := newNodeDB(nodeDBPath, Version, ourID)
    if err != nil {
        glog.V(logger.Warn).Infoln("Failed to open node database:", err)
        db, _ = newNodeDB("", Version, ourID)
    }
    tab := &Table{
        net:       t,
        db:        db,
        self:      newNode(ourID, ourAddr.IP, uint16(ourAddr.Port), uint16(ourAddr.Port)),
        bonding:   make(map[NodeID]*bondproc),
        bondslots: make(chan struct{}, maxBondingPingPongs),
    }
    for i := 0; i < cap(tab.bondslots); i++ {
        tab.bondslots <- struct{}{}
    }
    for i := range tab.buckets {
        tab.buckets[i] = new(bucket)
    }
    return tab
}

// Self returns the local node.
// The returned node should not be modified by the caller.
func (tab *Table) Self() *Node {
    return tab.self
}

// ReadRandomNodes fills the given slice with random nodes from the
// table. It will not write the same node more than once. The nodes in
// the slice are copies and can be modified by the caller.
func (tab *Table) ReadRandomNodes(buf []*Node) (n int) {
    tab.mutex.Lock()
    defer tab.mutex.Unlock()
    // TODO: tree-based buckets would help here
    // Find all non-empty buckets and get a fresh slice of their entries.
    var buckets [][]*Node
    for _, b := range tab.buckets {
        if len(b.entries) > 0 {
            buckets = append(buckets, b.entries[:])
        }
    }
    if len(buckets) == 0 {
        return 0
    }
    // Shuffle the buckets.
    for i := uint32(len(buckets)) - 1; i > 0; i-- {
        j := randUint(i)
        buckets[i], buckets[j] = buckets[j], buckets[i]
    }
    // Move head of each bucket into buf, removing buckets that become empty.
    var i, j int
    for ; i < len(buf); i, j = i+1, (j+1)%len(buckets) {
        b := buckets[j]
        buf[i] = &(*b[0])
        buckets[j] = b[1:]
        if len(b) == 1 {
            buckets = append(buckets[:j], buckets[j+1:]...)
        }
        if len(buckets) == 0 {
            break
        }
    }
    return i + 1
}

func randUint(max uint32) uint32 {
    if max == 0 {
        return 0
    }
    var b [4]byte
    rand.Read(b[:])
    return binary.BigEndian.Uint32(b[:]) % max
}

// Close terminates the network listener and flushes the node database.
func (tab *Table) Close() {
    tab.net.close()
    tab.db.close()
}

// Bootstrap sets the bootstrap nodes. These nodes are used to connect
// to the network if the table is empty. Bootstrap will also attempt to
// fill the table by performing random lookup operations on the
// network.
func (tab *Table) Bootstrap(nodes []*Node) {
    tab.mutex.Lock()
    // TODO: maybe filter nodes with bad fields (nil, etc.) to avoid strange crashes
    tab.nursery = make([]*Node, 0, len(nodes))
    for _, n := range nodes {
        cpy := *n
        cpy.sha = crypto.Sha3Hash(n.ID[:])
        tab.nursery = append(tab.nursery, &cpy)
    }
    tab.mutex.Unlock()
    tab.refresh()
}

// Lookup performs a network search for nodes close
// to the given target. It approaches the target by querying
// nodes that are closer to it on each iteration.
// The given target does not need to be an actual node
// identifier.
func (tab *Table) Lookup(targetID NodeID) []*Node {
    var (
        target         = crypto.Sha3Hash(targetID[:])
        asked          = make(map[NodeID]bool)
        seen           = make(map[NodeID]bool)
        reply          = make(chan []*Node, alpha)
        pendingQueries = 0
    )
    // don't query further if we hit ourself.
    // unlikely to happen often in practice.
    asked[tab.self.ID] = true

    tab.mutex.Lock()
    // update last lookup stamp (for refresh logic)
    tab.buckets[logdist(tab.self.sha, target)].lastLookup = time.Now()
    // generate initial result set
    result := tab.closest(target, bucketSize)
    tab.mutex.Unlock()

    for {
        // ask the alpha closest nodes that we haven't asked yet
        for i := 0; i < len(result.entries) && pendingQueries < alpha; i++ {
            n := result.entries[i]
            if !asked[n.ID] {
                asked[n.ID] = true
                pendingQueries++
                go func() {
                    // Find potential neighbors to bond with
                    r, err := tab.net.findnode(n.ID, n.addr(), targetID)
                    if err != nil {
                        // Bump the failure counter to detect and evacuate non-bonded entries
                        fails := tab.db.findFails(n.ID) + 1
                        tab.db.updateFindFails(n.ID, fails)
                        glog.V(logger.Detail).Infof("Bumping failures for %x: %d", n.ID[:8], fails)

                        if fails > maxFindnodeFailures {
                            glog.V(logger.Detail).Infof("Evacuating node %x: %d findnode failures", n.ID[:8], fails)
                            tab.del(n)
                        }
                    }
                    reply <- tab.bondall(r)
                }()
            }
        }
        if pendingQueries == 0 {
            // we have asked all closest nodes, stop the search
            break
        }
        // wait for the next reply
        for _, n := range <-reply {
            if n != nil && !seen[n.ID] {
                seen[n.ID] = true
                result.push(n, bucketSize)
            }
        }
        pendingQueries--
    }
    return result.entries
}

// refresh performs a lookup for a random target to keep buckets full.
func (tab *Table) refresh() {
    // The Kademlia paper specifies that the bucket refresh should
    // perform a refresh in the least recently used bucket. We cannot
    // adhere to this because the findnode target is a 512bit value
    // (not hash-sized) and it is not easily possible to generate a
    // sha3 preimage that falls into a chosen bucket.
    //
    // We perform a lookup with a random target instead.
    var target NodeID
    rand.Read(target[:])
    result := tab.Lookup(target)
    if len(result) == 0 {
        // Pick a batch of previously know seeds to lookup with
        seeds := tab.db.querySeeds(10)
        for _, seed := range seeds {
            glog.V(logger.Debug).Infoln("Seeding network with", seed)
        }
        // Bootstrap the table with a self lookup
        all := tab.bondall(append(tab.nursery, seeds...))
        tab.mutex.Lock()
        tab.add(all)
        tab.mutex.Unlock()
        tab.Lookup(tab.self.ID)
        // TODO: the Kademlia paper says that we're supposed to perform
        // random lookups in all buckets further away than our closest neighbor.
    }
}

// closest returns the n nodes in the table that are closest to the
// given id. The caller must hold tab.mutex.
func (tab *Table) closest(target common.Hash, nresults int) *nodesByDistance {
    // This is a very wasteful way to find the closest nodes but
    // obviously correct. I believe that tree-based buckets would make
    // this easier to implement efficiently.
    close := &nodesByDistance{target: target}
    for _, b := range tab.buckets {
        for _, n := range b.entries {
            close.push(n, nresults)
        }
    }
    return close
}

func (tab *Table) len() (n int) {
    for _, b := range tab.buckets {
        n += len(b.entries)
    }
    return n
}

// bondall bonds with all given nodes concurrently and returns
// those nodes for which bonding has probably succeeded.
func (tab *Table) bondall(nodes []*Node) (result []*Node) {
    rc := make(chan *Node, len(nodes))
    for i := range nodes {
        go func(n *Node) {
            nn, _ := tab.bond(false, n.ID, n.addr(), uint16(n.TCP))
            rc <- nn
        }(nodes[i])
    }
    for _ = range nodes {
        if n := <-rc; n != nil {
            result = append(result, n)
        }
    }
    return result
}

// bond ensures the local node has a bond with the given remote node.
// It also attempts to insert the node into the table if bonding succeeds.
// The caller must not hold tab.mutex.
//
// A bond is must be established before sending findnode requests.
// Both sides must have completed a ping/pong exchange for a bond to
// exist. The total number of active bonding processes is limited in
// order to restrain network use.
//
// bond is meant to operate idempotently in that bonding with a remote
// node which still remembers a previously established bond will work.
// The remote node will simply not send a ping back, causing waitping
// to time out.
//
// If pinged is true, the remote node has just pinged us and one half
// of the process can be skipped.
func (tab *Table) bond(pinged bool, id NodeID, addr *net.UDPAddr, tcpPort uint16) (*Node, error) {
    // Retrieve a previously known node and any recent findnode failures
    node, fails := tab.db.node(id), 0
    if node != nil {
        fails = tab.db.findFails(id)
    }
    // If the node is unknown (non-bonded) or failed (remotely unknown), bond from scratch
    var result error
    if node == nil || fails > 0 {
        glog.V(logger.Detail).Infof("Bonding %x: known=%v, fails=%v", id[:8], node != nil, fails)

        tab.bondmu.Lock()
        w := tab.bonding[id]
        if w != nil {
            // Wait for an existing bonding process to complete.
            tab.bondmu.Unlock()
            <-w.done
        } else {
            // Register a new bonding process.
            w = &bondproc{done: make(chan struct{})}
            tab.bonding[id] = w
            tab.bondmu.Unlock()
            // Do the ping/pong. The result goes into w.
            tab.pingpong(w, pinged, id, addr, tcpPort)
            // Unregister the process after it's done.
            tab.bondmu.Lock()
            delete(tab.bonding, id)
            tab.bondmu.Unlock()
        }
        // Retrieve the bonding results
        result = w.err
        if result == nil {
            node = w.n
        }
    }
    // Even if bonding temporarily failed, give the node a chance
    if node != nil {
        tab.mutex.Lock()
        defer tab.mutex.Unlock()

        b := tab.buckets[logdist(tab.self.sha, node.sha)]
        if !b.bump(node) {
            tab.pingreplace(node, b)
        }
        tab.db.updateFindFails(id, 0)
    }
    return node, result
}

func (tab *Table) pingpong(w *bondproc, pinged bool, id NodeID, addr *net.UDPAddr, tcpPort uint16) {
    // Request a bonding slot to limit network usage
    <-tab.bondslots
    defer func() { tab.bondslots <- struct{}{} }()

    // Ping the remote side and wait for a pong
    if w.err = tab.ping(id, addr); w.err != nil {
        close(w.done)
        return
    }
    if !pinged {
        // Give the remote node a chance to ping us before we start
        // sending findnode requests. If they still remember us,
        // waitping will simply time out.
        tab.net.waitping(id)
    }
    // Bonding succeeded, update the node database
    w.n = newNode(id, addr.IP, uint16(addr.Port), tcpPort)
    tab.db.updateNode(w.n)
    close(w.done)
}

func (tab *Table) pingreplace(new *Node, b *bucket) {
    if len(b.entries) == bucketSize {
        oldest := b.entries[bucketSize-1]
        if err := tab.ping(oldest.ID, oldest.addr()); err == nil {
            // The node responded, we don't need to replace it.
            return
        }
    } else {
        // Add a slot at the end so the last entry doesn't
        // fall off when adding the new node.
        b.entries = append(b.entries, nil)
    }
    copy(b.entries[1:], b.entries)
    b.entries[0] = new
}

// ping a remote endpoint and wait for a reply, also updating the node database
// accordingly.
func (tab *Table) ping(id NodeID, addr *net.UDPAddr) error {
    // Update the last ping and send the message
    tab.db.updateLastPing(id, time.Now())
    if err := tab.net.ping(id, addr); err != nil {
        return err
    }
    // Pong received, update the database and return
    tab.db.updateLastPong(id, time.Now())
    tab.db.ensureExpirer()

    return nil
}

// add puts the entries into the table if their corresponding
// bucket is not full. The caller must hold tab.mutex.
func (tab *Table) add(entries []*Node) {
outer:
    for _, n := range entries {
        if n.ID == tab.self.ID {
            // don't add self.
            continue
        }
        bucket := tab.buckets[logdist(tab.self.sha, n.sha)]
        for i := range bucket.entries {
            if bucket.entries[i].ID == n.ID {
                // already in bucket
                continue outer
            }
        }
        if len(bucket.entries) < bucketSize {
            bucket.entries = append(bucket.entries, n)
        }
    }
}

// del removes an entry from the node table (used to evacuate failed/non-bonded
// discovery peers).
func (tab *Table) del(node *Node) {
    tab.mutex.Lock()
    defer tab.mutex.Unlock()

    bucket := tab.buckets[logdist(tab.self.sha, node.sha)]
    for i := range bucket.entries {
        if bucket.entries[i].ID == node.ID {
            bucket.entries = append(bucket.entries[:i], bucket.entries[i+1:]...)
            return
        }
    }
}

func (b *bucket) bump(n *Node) bool {
    for i := range b.entries {
        if b.entries[i].ID == n.ID {
            // move it to the front
            copy(b.entries[1:], b.entries[:i])
            b.entries[0] = n
            return true
        }
    }
    return false
}

// nodesByDistance is a list of nodes, ordered by
// distance to target.
type nodesByDistance struct {
    entries []*Node
    target  common.Hash
}

// push adds the given node to the list, keeping the total size below maxElems.
func (h *nodesByDistance) push(n *Node, maxElems int) {
    ix := sort.Search(len(h.entries), func(i int) bool {
        return distcmp(h.target, h.entries[i].sha, n.sha) > 0
    })
    if len(h.entries) < maxElems {
        h.entries = append(h.entries, n)
    }
    if ix == len(h.entries) {
        // farther away than all nodes we already have.
        // if there was room for it, the node is now the last element.
    } else {
        // slide existing entries down to make room
        // this will overwrite the entry we just appended.
        copy(h.entries[ix+1:], h.entries[ix:])
        h.entries[ix] = n
    }
}