quorum/les/clientpool.go

707 lines
23 KiB
Go

// Copyright 2019 The go-ethereum Authors
// This file is part of the go-ethereum library.
//
// The go-ethereum library is free software: you can redistribute it and/or modify
// it under the terms of the GNU Lesser General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// The go-ethereum library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
package les
import (
"encoding/binary"
"io"
"math"
"sync"
"time"
"github.com/ethereum/go-ethereum/common"
"github.com/ethereum/go-ethereum/common/mclock"
"github.com/ethereum/go-ethereum/common/prque"
"github.com/ethereum/go-ethereum/ethdb"
"github.com/ethereum/go-ethereum/log"
"github.com/ethereum/go-ethereum/p2p/enode"
"github.com/ethereum/go-ethereum/rlp"
"github.com/hashicorp/golang-lru"
)
const (
negBalanceExpTC = time.Hour // time constant for exponentially reducing negative balance
fixedPointMultiplier = 0x1000000 // constant to convert logarithms to fixed point format
lazyQueueRefresh = time.Second * 10 // refresh period of the connected queue
persistCumulativeTimeRefresh = time.Minute * 5 // refresh period of the cumulative running time persistence
posBalanceCacheLimit = 8192 // the maximum number of cached items in positive balance queue
negBalanceCacheLimit = 8192 // the maximum number of cached items in negative balance queue
// connectedBias is applied to already connected clients So that
// already connected client won't be kicked out very soon and we
// can ensure all connected clients can have enough time to request
// or sync some data.
//
// todo(rjl493456442) make it configurable. It can be the option of
// free trial time!
connectedBias = time.Minute * 3
)
// clientPool implements a client database that assigns a priority to each client
// based on a positive and negative balance. Positive balance is externally assigned
// to prioritized clients and is decreased with connection time and processed
// requests (unless the price factors are zero). If the positive balance is zero
// then negative balance is accumulated.
//
// Balance tracking and priority calculation for connected clients is done by
// balanceTracker. connectedQueue ensures that clients with the lowest positive or
// highest negative balance get evicted when the total capacity allowance is full
// and new clients with a better balance want to connect.
//
// Already connected nodes receive a small bias in their favor in order to avoid
// accepting and instantly kicking out clients. In theory, we try to ensure that
// each client can have several minutes of connection time.
//
// Balances of disconnected clients are stored in nodeDB including positive balance
// and negative banalce. Negative balance is transformed into a logarithmic form
// with a constantly shifting linear offset in order to implement an exponential
// decrease. Besides nodeDB will have a background thread to check the negative
// balance of disconnected client. If the balance is low enough, then the record
// will be dropped.
type clientPool struct {
ndb *nodeDB
lock sync.Mutex
clock mclock.Clock
stopCh chan struct{}
closed bool
removePeer func(enode.ID)
connectedMap map[enode.ID]*clientInfo
connectedQueue *prque.LazyQueue
posFactors, negFactors priceFactors
connLimit int // The maximum number of connections that clientpool can support
capLimit uint64 // The maximum cumulative capacity that clientpool can support
connectedCap uint64 // The sum of the capacity of the current clientpool connected
freeClientCap uint64 // The capacity value of each free client
startTime mclock.AbsTime // The timestamp at which the clientpool started running
cumulativeTime int64 // The cumulative running time of clientpool at the start point.
disableBias bool // Disable connection bias(used in testing)
}
// clientPeer represents a client in the pool.
// Positive balances are assigned to node key while negative balances are assigned
// to freeClientId. Currently network IP address without port is used because
// clients have a limited access to IP addresses while new node keys can be easily
// generated so it would be useless to assign a negative value to them.
type clientPeer interface {
ID() enode.ID
freeClientId() string
updateCapacity(uint64)
}
// clientInfo represents a connected client
type clientInfo struct {
address string
id enode.ID
capacity uint64
priority bool
pool *clientPool
peer clientPeer
queueIndex int // position in connectedQueue
balanceTracker balanceTracker
}
// connSetIndex callback updates clientInfo item index in connectedQueue
func connSetIndex(a interface{}, index int) {
a.(*clientInfo).queueIndex = index
}
// connPriority callback returns actual priority of clientInfo item in connectedQueue
func connPriority(a interface{}, now mclock.AbsTime) int64 {
c := a.(*clientInfo)
return c.balanceTracker.getPriority(now)
}
// connMaxPriority callback returns estimated maximum priority of clientInfo item in connectedQueue
func connMaxPriority(a interface{}, until mclock.AbsTime) int64 {
c := a.(*clientInfo)
pri := c.balanceTracker.estimatedPriority(until, true)
c.balanceTracker.addCallback(balanceCallbackQueue, pri+1, func() {
c.pool.lock.Lock()
if c.queueIndex != -1 {
c.pool.connectedQueue.Update(c.queueIndex)
}
c.pool.lock.Unlock()
})
return pri
}
// priceFactors determine the pricing policy (may apply either to positive or
// negative balances which may have different factors).
// - timeFactor is cost unit per nanosecond of connection time
// - capacityFactor is cost unit per nanosecond of connection time per 1000000 capacity
// - requestFactor is cost unit per request "realCost" unit
type priceFactors struct {
timeFactor, capacityFactor, requestFactor float64
}
// newClientPool creates a new client pool
func newClientPool(db ethdb.Database, freeClientCap uint64, clock mclock.Clock, removePeer func(enode.ID)) *clientPool {
ndb := newNodeDB(db, clock)
pool := &clientPool{
ndb: ndb,
clock: clock,
connectedMap: make(map[enode.ID]*clientInfo),
connectedQueue: prque.NewLazyQueue(connSetIndex, connPriority, connMaxPriority, clock, lazyQueueRefresh),
freeClientCap: freeClientCap,
removePeer: removePeer,
startTime: clock.Now(),
cumulativeTime: ndb.getCumulativeTime(),
stopCh: make(chan struct{}),
}
// If the negative balance of free client is even lower than 1,
// delete this entry.
ndb.nbEvictCallBack = func(now mclock.AbsTime, b negBalance) bool {
balance := math.Exp(float64(b.logValue-pool.logOffset(now)) / fixedPointMultiplier)
return balance <= 1
}
go func() {
for {
select {
case <-clock.After(lazyQueueRefresh):
pool.lock.Lock()
pool.connectedQueue.Refresh()
pool.lock.Unlock()
case <-clock.After(persistCumulativeTimeRefresh):
pool.ndb.setCumulativeTime(pool.logOffset(clock.Now()))
case <-pool.stopCh:
return
}
}
}()
return pool
}
// stop shuts the client pool down
func (f *clientPool) stop() {
close(f.stopCh)
f.lock.Lock()
f.closed = true
f.lock.Unlock()
f.ndb.setCumulativeTime(f.logOffset(f.clock.Now()))
f.ndb.close()
}
// connect should be called after a successful handshake. If the connection was
// rejected, there is no need to call disconnect.
func (f *clientPool) connect(peer clientPeer, capacity uint64) bool {
f.lock.Lock()
defer f.lock.Unlock()
// Short circuit if clientPool is already closed.
if f.closed {
return false
}
// Dedup connected peers.
id, freeID := peer.ID(), peer.freeClientId()
if _, ok := f.connectedMap[id]; ok {
clientRejectedMeter.Mark(1)
log.Debug("Client already connected", "address", freeID, "id", peerIdToString(id))
return false
}
// Create a clientInfo but do not add it yet
var (
posBalance uint64
negBalance uint64
now = f.clock.Now()
)
pb := f.ndb.getOrNewPB(id)
posBalance = pb.value
e := &clientInfo{pool: f, peer: peer, address: freeID, queueIndex: -1, id: id, priority: posBalance != 0}
nb := f.ndb.getOrNewNB(freeID)
if nb.logValue != 0 {
negBalance = uint64(math.Exp(float64(nb.logValue-f.logOffset(now)) / fixedPointMultiplier))
negBalance *= uint64(time.Second)
}
// If the client is a free client, assign with a low free capacity,
// Otherwise assign with the given value(priority client)
if !e.priority {
capacity = f.freeClientCap
}
// Ensure the capacity will never lower than the free capacity.
if capacity < f.freeClientCap {
capacity = f.freeClientCap
}
e.capacity = capacity
// Starts a balance tracker
e.balanceTracker.init(f.clock, capacity)
e.balanceTracker.setBalance(posBalance, negBalance)
f.setClientPriceFactors(e)
// If the number of clients already connected in the clientpool exceeds its
// capacity, evict some clients with lowest priority.
//
// If the priority of the newly added client is lower than the priority of
// all connected clients, the client is rejected.
newCapacity := f.connectedCap + capacity
newCount := f.connectedQueue.Size() + 1
if newCapacity > f.capLimit || newCount > f.connLimit {
var (
kickList []*clientInfo
kickPriority int64
)
f.connectedQueue.MultiPop(func(data interface{}, priority int64) bool {
c := data.(*clientInfo)
kickList = append(kickList, c)
kickPriority = priority
newCapacity -= c.capacity
newCount--
return newCapacity > f.capLimit || newCount > f.connLimit
})
bias := connectedBias
if f.disableBias {
bias = 0
}
if newCapacity > f.capLimit || newCount > f.connLimit || (e.balanceTracker.estimatedPriority(now+mclock.AbsTime(bias), false)-kickPriority) > 0 {
for _, c := range kickList {
f.connectedQueue.Push(c)
}
clientRejectedMeter.Mark(1)
log.Debug("Client rejected", "address", freeID, "id", peerIdToString(id))
return false
}
// accept new client, drop old ones
for _, c := range kickList {
f.dropClient(c, now, true)
}
}
// Register new client to connection queue.
f.connectedMap[id] = e
f.connectedQueue.Push(e)
f.connectedCap += e.capacity
// If the current client is a paid client, monitor the status of client,
// downgrade it to normal client if positive balance is used up.
if e.priority {
e.balanceTracker.addCallback(balanceCallbackZero, 0, func() { f.balanceExhausted(id) })
}
// If the capacity of client is not the default value(free capacity), notify
// it to update capacity.
if e.capacity != f.freeClientCap {
e.peer.updateCapacity(e.capacity)
}
totalConnectedGauge.Update(int64(f.connectedCap))
clientConnectedMeter.Mark(1)
log.Debug("Client accepted", "address", freeID)
return true
}
// disconnect should be called when a connection is terminated. If the disconnection
// was initiated by the pool itself using disconnectFn then calling disconnect is
// not necessary but permitted.
func (f *clientPool) disconnect(p clientPeer) {
f.lock.Lock()
defer f.lock.Unlock()
// Short circuit if client pool is already closed.
if f.closed {
return
}
// Short circuit if the peer hasn't been registered.
e := f.connectedMap[p.ID()]
if e == nil {
log.Debug("Client not connected", "address", p.freeClientId(), "id", peerIdToString(p.ID()))
return
}
f.dropClient(e, f.clock.Now(), false)
}
// dropClient removes a client from the connected queue and finalizes its balance.
// If kick is true then it also initiates the disconnection.
func (f *clientPool) dropClient(e *clientInfo, now mclock.AbsTime, kick bool) {
if _, ok := f.connectedMap[e.id]; !ok {
return
}
f.finalizeBalance(e, now)
f.connectedQueue.Remove(e.queueIndex)
delete(f.connectedMap, e.id)
f.connectedCap -= e.capacity
totalConnectedGauge.Update(int64(f.connectedCap))
if kick {
clientKickedMeter.Mark(1)
log.Debug("Client kicked out", "address", e.address)
f.removePeer(e.id)
} else {
clientDisconnectedMeter.Mark(1)
log.Debug("Client disconnected", "address", e.address)
}
}
// finalizeBalance stops the balance tracker, retrieves the final balances and
// stores them in posBalanceQueue and negBalanceQueue
func (f *clientPool) finalizeBalance(c *clientInfo, now mclock.AbsTime) {
c.balanceTracker.stop(now)
pos, neg := c.balanceTracker.getBalance(now)
pb, nb := f.ndb.getOrNewPB(c.id), f.ndb.getOrNewNB(c.address)
pb.value = pos
f.ndb.setPB(c.id, pb)
neg /= uint64(time.Second) // Convert the expanse to second level.
if neg > 1 {
nb.logValue = int64(math.Log(float64(neg))*fixedPointMultiplier) + f.logOffset(now)
f.ndb.setNB(c.address, nb)
} else {
f.ndb.delNB(c.address) // Negative balance is small enough, drop it directly.
}
}
// balanceExhausted callback is called by balanceTracker when positive balance is exhausted.
// It revokes priority status and also reduces the client capacity if necessary.
func (f *clientPool) balanceExhausted(id enode.ID) {
f.lock.Lock()
defer f.lock.Unlock()
c := f.connectedMap[id]
if c == nil || !c.priority {
return
}
c.priority = false
if c.capacity != f.freeClientCap {
f.connectedCap += f.freeClientCap - c.capacity
totalConnectedGauge.Update(int64(f.connectedCap))
c.capacity = f.freeClientCap
c.peer.updateCapacity(c.capacity)
}
f.ndb.delPB(id)
}
// setConnLimit sets the maximum number and total capacity of connected clients,
// dropping some of them if necessary.
func (f *clientPool) setLimits(totalConn int, totalCap uint64) {
f.lock.Lock()
defer f.lock.Unlock()
f.connLimit = totalConn
f.capLimit = totalCap
if f.connectedCap > f.capLimit || f.connectedQueue.Size() > f.connLimit {
f.connectedQueue.MultiPop(func(data interface{}, priority int64) bool {
f.dropClient(data.(*clientInfo), mclock.Now(), true)
return f.connectedCap > f.capLimit || f.connectedQueue.Size() > f.connLimit
})
}
}
// requestCost feeds request cost after serving a request from the given peer.
func (f *clientPool) requestCost(p *peer, cost uint64) {
f.lock.Lock()
defer f.lock.Unlock()
info, exist := f.connectedMap[p.ID()]
if !exist || f.closed {
return
}
info.balanceTracker.requestCost(cost)
}
// logOffset calculates the time-dependent offset for the logarithmic
// representation of negative balance
//
// From another point of view, the result returned by the function represents
// the total time that the clientpool is cumulatively running(total_hours/multiplier).
func (f *clientPool) logOffset(now mclock.AbsTime) int64 {
// Note: fixedPointMultiplier acts as a multiplier here; the reason for dividing the divisor
// is to avoid int64 overflow. We assume that int64(negBalanceExpTC) >> fixedPointMultiplier.
cumulativeTime := int64((time.Duration(now - f.startTime)) / (negBalanceExpTC / fixedPointMultiplier))
return f.cumulativeTime + cumulativeTime
}
// setPriceFactors changes pricing factors for both positive and negative balances.
// Applies to connected clients and also future connections.
func (f *clientPool) setPriceFactors(posFactors, negFactors priceFactors) {
f.lock.Lock()
defer f.lock.Unlock()
f.posFactors, f.negFactors = posFactors, negFactors
for _, c := range f.connectedMap {
f.setClientPriceFactors(c)
}
}
// setClientPriceFactors sets the pricing factors for an individual connected client
func (f *clientPool) setClientPriceFactors(c *clientInfo) {
c.balanceTracker.setFactors(true, f.negFactors.timeFactor+float64(c.capacity)*f.negFactors.capacityFactor/1000000, f.negFactors.requestFactor)
c.balanceTracker.setFactors(false, f.posFactors.timeFactor+float64(c.capacity)*f.posFactors.capacityFactor/1000000, f.posFactors.requestFactor)
}
// addBalance updates the positive balance of a client.
// If setTotal is false then the given amount is added to the balance.
// If setTotal is true then amount represents the total amount ever added to the
// given ID and positive balance is increased by (amount-lastTotal) while lastTotal
// is updated to amount. This method also allows removing positive balance.
func (f *clientPool) addBalance(id enode.ID, amount uint64, setTotal bool) {
f.lock.Lock()
defer f.lock.Unlock()
pb := f.ndb.getOrNewPB(id)
c := f.connectedMap[id]
if c != nil {
posBalance, negBalance := c.balanceTracker.getBalance(f.clock.Now())
pb.value = posBalance
defer func() {
c.balanceTracker.setBalance(pb.value, negBalance)
if !c.priority && pb.value > 0 {
// The capacity should be adjusted based on the requirement,
// but we have no idea about the new capacity, need a second
// call to udpate it.
c.priority = true
c.balanceTracker.addCallback(balanceCallbackZero, 0, func() { f.balanceExhausted(id) })
}
}()
}
if setTotal {
if pb.value+amount > pb.lastTotal {
pb.value += amount - pb.lastTotal
} else {
pb.value = 0
}
pb.lastTotal = amount
} else {
pb.value += amount
pb.lastTotal += amount
}
f.ndb.setPB(id, pb)
}
// posBalance represents a recently accessed positive balance entry
type posBalance struct {
value, lastTotal uint64
}
// EncodeRLP implements rlp.Encoder
func (e *posBalance) EncodeRLP(w io.Writer) error {
return rlp.Encode(w, []interface{}{e.value, e.lastTotal})
}
// DecodeRLP implements rlp.Decoder
func (e *posBalance) DecodeRLP(s *rlp.Stream) error {
var entry struct {
Value, LastTotal uint64
}
if err := s.Decode(&entry); err != nil {
return err
}
e.value = entry.Value
e.lastTotal = entry.LastTotal
return nil
}
// negBalance represents a negative balance entry of a disconnected client
type negBalance struct{ logValue int64 }
// EncodeRLP implements rlp.Encoder
func (e *negBalance) EncodeRLP(w io.Writer) error {
return rlp.Encode(w, []interface{}{uint64(e.logValue)})
}
// DecodeRLP implements rlp.Decoder
func (e *negBalance) DecodeRLP(s *rlp.Stream) error {
var entry struct {
LogValue uint64
}
if err := s.Decode(&entry); err != nil {
return err
}
e.logValue = int64(entry.LogValue)
return nil
}
const (
// nodeDBVersion is the version identifier of the node data in db
nodeDBVersion = 0
// dbCleanupCycle is the cycle of db for useless data cleanup
dbCleanupCycle = time.Hour
)
var (
positiveBalancePrefix = []byte("pb:") // dbVersion(uint16 big endian) + positiveBalancePrefix + id -> balance
negativeBalancePrefix = []byte("nb:") // dbVersion(uint16 big endian) + negativeBalancePrefix + ip -> balance
cumulativeRunningTimeKey = []byte("cumulativeTime:") // dbVersion(uint16 big endian) + cumulativeRunningTimeKey -> cumulativeTime
)
type nodeDB struct {
db ethdb.Database
pcache *lru.Cache
ncache *lru.Cache
auxbuf []byte // 37-byte auxiliary buffer for key encoding
verbuf [2]byte // 2-byte auxiliary buffer for db version
nbEvictCallBack func(mclock.AbsTime, negBalance) bool // Callback to determine whether the negative balance can be evicted.
clock mclock.Clock
closeCh chan struct{}
cleanupHook func() // Test hook used for testing
}
func newNodeDB(db ethdb.Database, clock mclock.Clock) *nodeDB {
pcache, _ := lru.New(posBalanceCacheLimit)
ncache, _ := lru.New(negBalanceCacheLimit)
ndb := &nodeDB{
db: db,
pcache: pcache,
ncache: ncache,
auxbuf: make([]byte, 37),
clock: clock,
closeCh: make(chan struct{}),
}
binary.BigEndian.PutUint16(ndb.verbuf[:], uint16(nodeDBVersion))
go ndb.expirer()
return ndb
}
func (db *nodeDB) close() {
close(db.closeCh)
}
func (db *nodeDB) key(id []byte, neg bool) []byte {
prefix := positiveBalancePrefix
if neg {
prefix = negativeBalancePrefix
}
if len(prefix)+len(db.verbuf)+len(id) > len(db.auxbuf) {
db.auxbuf = append(db.auxbuf, make([]byte, len(prefix)+len(db.verbuf)+len(id)-len(db.auxbuf))...)
}
copy(db.auxbuf[:len(db.verbuf)], db.verbuf[:])
copy(db.auxbuf[len(db.verbuf):len(db.verbuf)+len(prefix)], prefix)
copy(db.auxbuf[len(prefix)+len(db.verbuf):len(prefix)+len(db.verbuf)+len(id)], id)
return db.auxbuf[:len(prefix)+len(db.verbuf)+len(id)]
}
func (db *nodeDB) getCumulativeTime() int64 {
blob, err := db.db.Get(append(cumulativeRunningTimeKey, db.verbuf[:]...))
if err != nil || len(blob) == 0 {
return 0
}
return int64(binary.BigEndian.Uint64(blob))
}
func (db *nodeDB) setCumulativeTime(v int64) {
binary.BigEndian.PutUint64(db.auxbuf[:8], uint64(v))
db.db.Put(append(cumulativeRunningTimeKey, db.verbuf[:]...), db.auxbuf[:8])
}
func (db *nodeDB) getOrNewPB(id enode.ID) posBalance {
key := db.key(id.Bytes(), false)
item, exist := db.pcache.Get(string(key))
if exist {
return item.(posBalance)
}
var balance posBalance
if enc, err := db.db.Get(key); err == nil {
if err := rlp.DecodeBytes(enc, &balance); err != nil {
log.Error("Failed to decode positive balance", "err", err)
}
}
db.pcache.Add(string(key), balance)
return balance
}
func (db *nodeDB) setPB(id enode.ID, b posBalance) {
key := db.key(id.Bytes(), false)
enc, err := rlp.EncodeToBytes(&(b))
if err != nil {
log.Error("Failed to encode positive balance", "err", err)
return
}
db.db.Put(key, enc)
db.pcache.Add(string(key), b)
}
func (db *nodeDB) delPB(id enode.ID) {
key := db.key(id.Bytes(), false)
db.db.Delete(key)
db.pcache.Remove(string(key))
}
func (db *nodeDB) getOrNewNB(id string) negBalance {
key := db.key([]byte(id), true)
item, exist := db.ncache.Get(string(key))
if exist {
return item.(negBalance)
}
var balance negBalance
if enc, err := db.db.Get(key); err == nil {
if err := rlp.DecodeBytes(enc, &balance); err != nil {
log.Error("Failed to decode negative balance", "err", err)
}
}
db.ncache.Add(string(key), balance)
return balance
}
func (db *nodeDB) setNB(id string, b negBalance) {
key := db.key([]byte(id), true)
enc, err := rlp.EncodeToBytes(&(b))
if err != nil {
log.Error("Failed to encode negative balance", "err", err)
return
}
db.db.Put(key, enc)
db.ncache.Add(string(key), b)
}
func (db *nodeDB) delNB(id string) {
key := db.key([]byte(id), true)
db.db.Delete(key)
db.ncache.Remove(string(key))
}
func (db *nodeDB) expirer() {
for {
select {
case <-db.clock.After(dbCleanupCycle):
db.expireNodes()
case <-db.closeCh:
return
}
}
}
// expireNodes iterates the whole node db and checks whether the negative balance
// entry can deleted.
//
// The rationale behind this is: server doesn't need to keep the negative balance
// records if they are low enough.
func (db *nodeDB) expireNodes() {
var (
visited int
deleted int
start = time.Now()
)
iter := db.db.NewIteratorWithPrefix(append(db.verbuf[:], negativeBalancePrefix...))
for iter.Next() {
visited += 1
var balance negBalance
if err := rlp.DecodeBytes(iter.Value(), &balance); err != nil {
log.Error("Failed to decode negative balance", "err", err)
continue
}
if db.nbEvictCallBack != nil && db.nbEvictCallBack(db.clock.Now(), balance) {
deleted += 1
db.db.Delete(iter.Key())
}
}
// Invoke testing hook if it's not nil.
if db.cleanupHook != nil {
db.cleanupHook()
}
log.Debug("Expire nodes", "visited", visited, "deleted", deleted, "elapsed", common.PrettyDuration(time.Since(start)))
}