plugeth/eth/sync.go
Felix Lange d6c5f2417c
eth: improve shutdown synchronization (#20695)
* eth: improve shutdown synchronization

Most goroutines started by eth.Ethereum didn't have any shutdown sync at
all, which lead to weird error messages when quitting the client.

This change improves the clean shutdown path by stopping all internal
components in dependency order and waiting for them to actually be
stopped before shutdown is considered done. In particular, we now stop
everything related to peers before stopping 'resident' parts such as
core.BlockChain.

* eth: rewrite sync controller

* eth: remove sync start debug message

* eth: notify chainSyncer about new peers after handshake

* eth: move downloader.Cancel call into chainSyncer

* eth: make post-sync block broadcast synchronous

* eth: add comments

* core: change blockchain stop message

* eth: change closeBloomHandler channel type
2020-03-27 15:03:20 +02:00

321 lines
9.3 KiB
Go

// Copyright 2015 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 eth
import (
"math/big"
"math/rand"
"sync/atomic"
"time"
"github.com/ethereum/go-ethereum/common"
"github.com/ethereum/go-ethereum/core/types"
"github.com/ethereum/go-ethereum/eth/downloader"
"github.com/ethereum/go-ethereum/log"
"github.com/ethereum/go-ethereum/p2p/enode"
)
const (
forceSyncCycle = 10 * time.Second // Time interval to force syncs, even if few peers are available
defaultMinSyncPeers = 5 // Amount of peers desired to start syncing
// This is the target size for the packs of transactions sent by txsyncLoop64.
// A pack can get larger than this if a single transactions exceeds this size.
txsyncPackSize = 100 * 1024
)
type txsync struct {
p *peer
txs []*types.Transaction
}
// syncTransactions starts sending all currently pending transactions to the given peer.
func (pm *ProtocolManager) syncTransactions(p *peer) {
// Assemble the set of transaction to broadcast or announce to the remote
// peer. Fun fact, this is quite an expensive operation as it needs to sort
// the transactions if the sorting is not cached yet. However, with a random
// order, insertions could overflow the non-executable queues and get dropped.
//
// TODO(karalabe): Figure out if we could get away with random order somehow
var txs types.Transactions
pending, _ := pm.txpool.Pending()
for _, batch := range pending {
txs = append(txs, batch...)
}
if len(txs) == 0 {
return
}
// The eth/65 protocol introduces proper transaction announcements, so instead
// of dripping transactions across multiple peers, just send the entire list as
// an announcement and let the remote side decide what they need (likely nothing).
if p.version >= eth65 {
hashes := make([]common.Hash, len(txs))
for i, tx := range txs {
hashes[i] = tx.Hash()
}
p.AsyncSendPooledTransactionHashes(hashes)
return
}
// Out of luck, peer is running legacy protocols, drop the txs over
select {
case pm.txsyncCh <- &txsync{p: p, txs: txs}:
case <-pm.quitSync:
}
}
// txsyncLoop64 takes care of the initial transaction sync for each new
// connection. When a new peer appears, we relay all currently pending
// transactions. In order to minimise egress bandwidth usage, we send
// the transactions in small packs to one peer at a time.
func (pm *ProtocolManager) txsyncLoop64() {
defer pm.wg.Done()
var (
pending = make(map[enode.ID]*txsync)
sending = false // whether a send is active
pack = new(txsync) // the pack that is being sent
done = make(chan error, 1) // result of the send
)
// send starts a sending a pack of transactions from the sync.
send := func(s *txsync) {
if s.p.version >= eth65 {
panic("initial transaction syncer running on eth/65+")
}
// Fill pack with transactions up to the target size.
size := common.StorageSize(0)
pack.p = s.p
pack.txs = pack.txs[:0]
for i := 0; i < len(s.txs) && size < txsyncPackSize; i++ {
pack.txs = append(pack.txs, s.txs[i])
size += s.txs[i].Size()
}
// Remove the transactions that will be sent.
s.txs = s.txs[:copy(s.txs, s.txs[len(pack.txs):])]
if len(s.txs) == 0 {
delete(pending, s.p.ID())
}
// Send the pack in the background.
s.p.Log().Trace("Sending batch of transactions", "count", len(pack.txs), "bytes", size)
sending = true
go func() { done <- pack.p.SendTransactions64(pack.txs) }()
}
// pick chooses the next pending sync.
pick := func() *txsync {
if len(pending) == 0 {
return nil
}
n := rand.Intn(len(pending)) + 1
for _, s := range pending {
if n--; n == 0 {
return s
}
}
return nil
}
for {
select {
case s := <-pm.txsyncCh:
pending[s.p.ID()] = s
if !sending {
send(s)
}
case err := <-done:
sending = false
// Stop tracking peers that cause send failures.
if err != nil {
pack.p.Log().Debug("Transaction send failed", "err", err)
delete(pending, pack.p.ID())
}
// Schedule the next send.
if s := pick(); s != nil {
send(s)
}
case <-pm.quitSync:
return
}
}
}
// chainSyncer coordinates blockchain sync components.
type chainSyncer struct {
pm *ProtocolManager
force *time.Timer
forced bool // true when force timer fired
peerEventCh chan struct{}
doneCh chan error // non-nil when sync is running
}
// chainSyncOp is a scheduled sync operation.
type chainSyncOp struct {
mode downloader.SyncMode
peer *peer
td *big.Int
head common.Hash
}
// newChainSyncer creates a chainSyncer.
func newChainSyncer(pm *ProtocolManager) *chainSyncer {
return &chainSyncer{
pm: pm,
peerEventCh: make(chan struct{}),
}
}
// handlePeerEvent notifies the syncer about a change in the peer set.
// This is called for new peers and every time a peer announces a new
// chain head.
func (cs *chainSyncer) handlePeerEvent(p *peer) bool {
select {
case cs.peerEventCh <- struct{}{}:
return true
case <-cs.pm.quitSync:
return false
}
}
// loop runs in its own goroutine and launches the sync when necessary.
func (cs *chainSyncer) loop() {
defer cs.pm.wg.Done()
cs.pm.blockFetcher.Start()
cs.pm.txFetcher.Start()
defer cs.pm.blockFetcher.Stop()
defer cs.pm.txFetcher.Stop()
defer cs.pm.downloader.Terminate()
// The force timer lowers the peer count threshold down to one when it fires.
// This ensures we'll always start sync even if there aren't enough peers.
cs.force = time.NewTimer(forceSyncCycle)
defer cs.force.Stop()
for {
if op := cs.nextSyncOp(); op != nil {
cs.startSync(op)
}
select {
case <-cs.peerEventCh:
// Peer information changed, recheck.
case <-cs.doneCh:
cs.doneCh = nil
cs.force.Reset(forceSyncCycle)
cs.forced = false
case <-cs.force.C:
cs.forced = true
case <-cs.pm.quitSync:
if cs.doneCh != nil {
cs.pm.downloader.Cancel()
<-cs.doneCh
}
return
}
}
}
// nextSyncOp determines whether sync is required at this time.
func (cs *chainSyncer) nextSyncOp() *chainSyncOp {
if cs.doneCh != nil {
return nil // Sync already running.
}
// Ensure we're at mininum peer count.
minPeers := defaultMinSyncPeers
if cs.forced {
minPeers = 1
} else if minPeers > cs.pm.maxPeers {
minPeers = cs.pm.maxPeers
}
if cs.pm.peers.Len() < minPeers {
return nil
}
// We have enough peers, check TD.
peer := cs.pm.peers.BestPeer()
if peer == nil {
return nil
}
mode, ourTD := cs.modeAndLocalHead()
op := peerToSyncOp(mode, peer)
if op.td.Cmp(ourTD) <= 0 {
return nil // We're in sync.
}
return op
}
func peerToSyncOp(mode downloader.SyncMode, p *peer) *chainSyncOp {
peerHead, peerTD := p.Head()
return &chainSyncOp{mode: mode, peer: p, td: peerTD, head: peerHead}
}
func (cs *chainSyncer) modeAndLocalHead() (downloader.SyncMode, *big.Int) {
if atomic.LoadUint32(&cs.pm.fastSync) == 1 {
block := cs.pm.blockchain.CurrentFastBlock()
td := cs.pm.blockchain.GetTdByHash(block.Hash())
return downloader.FastSync, td
} else {
head := cs.pm.blockchain.CurrentHeader()
td := cs.pm.blockchain.GetTd(head.Hash(), head.Number.Uint64())
return downloader.FullSync, td
}
}
// startSync launches doSync in a new goroutine.
func (cs *chainSyncer) startSync(op *chainSyncOp) {
cs.doneCh = make(chan error, 1)
go func() { cs.doneCh <- cs.pm.doSync(op) }()
}
// doSync synchronizes the local blockchain with a remote peer.
func (pm *ProtocolManager) doSync(op *chainSyncOp) error {
// Run the sync cycle, and disable fast sync if we're past the pivot block
err := pm.downloader.Synchronise(op.peer.id, op.head, op.td, op.mode)
if err != nil {
return err
}
if atomic.LoadUint32(&pm.fastSync) == 1 {
log.Info("Fast sync complete, auto disabling")
atomic.StoreUint32(&pm.fastSync, 0)
}
// If we've successfully finished a sync cycle and passed any required checkpoint,
// enable accepting transactions from the network.
head := pm.blockchain.CurrentBlock()
if head.NumberU64() >= pm.checkpointNumber {
// Checkpoint passed, sanity check the timestamp to have a fallback mechanism
// for non-checkpointed (number = 0) private networks.
if head.Time() >= uint64(time.Now().AddDate(0, -1, 0).Unix()) {
atomic.StoreUint32(&pm.acceptTxs, 1)
}
}
if head.NumberU64() > 0 {
// We've completed a sync cycle, notify all peers of new state. This path is
// essential in star-topology networks where a gateway node needs to notify
// all its out-of-date peers of the availability of a new block. This failure
// scenario will most often crop up in private and hackathon networks with
// degenerate connectivity, but it should be healthy for the mainnet too to
// more reliably update peers or the local TD state.
pm.BroadcastBlock(head, false)
}
return nil
}