go-ethereum/eth/downloader/downloader.go

1856 lines
69 KiB
Go
Raw Normal View History

2015-07-07 03:54:22 +03:00
// Copyright 2015 The go-ethereum Authors
// This file is part of the go-ethereum library.
2015-07-07 03:54:22 +03:00
//
// The go-ethereum library is free software: you can redistribute it and/or modify
2015-07-07 03:54:22 +03:00
// 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,
2015-07-07 03:54:22 +03:00
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
2015-07-07 03:54:22 +03:00
// 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/>.
2015-07-07 03:54:22 +03:00
// Package downloader contains the manual full chain synchronisation.
2015-04-12 13:38:25 +03:00
package downloader
import (
"errors"
"fmt"
"math/big"
2015-04-12 13:38:25 +03:00
"sync"
"sync/atomic"
"time"
"github.com/ethereum/go-ethereum"
2015-04-12 13:38:25 +03:00
"github.com/ethereum/go-ethereum/common"
"github.com/ethereum/go-ethereum/core/rawdb"
2015-04-12 13:38:25 +03:00
"github.com/ethereum/go-ethereum/core/types"
"github.com/ethereum/go-ethereum/ethdb"
"github.com/ethereum/go-ethereum/event"
"github.com/ethereum/go-ethereum/log"
"github.com/ethereum/go-ethereum/metrics"
"github.com/ethereum/go-ethereum/params"
"github.com/ethereum/go-ethereum/trie"
2015-04-12 13:38:25 +03:00
)
var (
MaxHashFetch = 512 // Amount of hashes to be fetched per retrieval request
MaxBlockFetch = 128 // Amount of blocks to be fetched per retrieval request
MaxHeaderFetch = 192 // Amount of block headers to be fetched per retrieval request
MaxSkeletonSize = 128 // Number of header fetches to need for a skeleton assembly
MaxBodyFetch = 128 // Amount of block bodies to be fetched per retrieval request
MaxReceiptFetch = 256 // Amount of transaction receipts to allow fetching per request
MaxStateFetch = 384 // Amount of node state values to allow fetching per request
rttMinEstimate = 2 * time.Second // Minimum round-trip time to target for download requests
rttMaxEstimate = 20 * time.Second // Maximum round-trip time to target for download requests
rttMinConfidence = 0.1 // Worse confidence factor in our estimated RTT value
ttlScaling = 3 // Constant scaling factor for RTT -> TTL conversion
ttlLimit = time.Minute // Maximum TTL allowance to prevent reaching crazy timeouts
qosTuningPeers = 5 // Number of peers to tune based on (best peers)
qosConfidenceCap = 10 // Number of peers above which not to modify RTT confidence
qosTuningImpact = 0.25 // Impact that a new tuning target has on the previous value
maxQueuedHeaders = 32 * 1024 // [eth/62] Maximum number of headers to queue for import (DOS protection)
maxHeadersProcess = 2048 // Number of header download results to import at once into the chain
maxResultsProcess = 2048 // Number of content download results to import at once into the chain
maxForkAncestry uint64 = params.ImmutabilityThreshold // Maximum chain reorganisation (locally redeclared so tests can reduce it)
reorgProtThreshold = 48 // Threshold number of recent blocks to disable mini reorg protection
reorgProtHeaderDelay = 2 // Number of headers to delay delivering to cover mini reorgs
fsHeaderCheckFrequency = 100 // Verification frequency of the downloaded headers during fast sync
fsHeaderSafetyNet = 2048 // Number of headers to discard in case a chain violation is detected
fsHeaderForceVerify = 24 // Number of headers to verify before and after the pivot to accept it
fsHeaderContCheck = 3 * time.Second // Time interval to check for header continuations during state download
fsMinFullBlocks = 64 // Number of blocks to retrieve fully even in fast sync
)
var (
errBusy = errors.New("busy")
errUnknownPeer = errors.New("peer is unknown or unhealthy")
errBadPeer = errors.New("action from bad peer ignored")
errStallingPeer = errors.New("peer is stalling")
errUnsyncedPeer = errors.New("unsynced peer")
errNoPeers = errors.New("no peers to keep download active")
errTimeout = errors.New("timeout")
errEmptyHeaderSet = errors.New("empty header set by peer")
errPeersUnavailable = errors.New("no peers available or all tried for download")
errInvalidAncestor = errors.New("retrieved ancestor is invalid")
errInvalidChain = errors.New("retrieved hash chain is invalid")
errInvalidBlock = errors.New("retrieved block is invalid")
errInvalidBody = errors.New("retrieved block body is invalid")
errInvalidReceipt = errors.New("retrieved receipt is invalid")
errCancelBlockFetch = errors.New("block download canceled (requested)")
errCancelHeaderFetch = errors.New("block header download canceled (requested)")
errCancelBodyFetch = errors.New("block body download canceled (requested)")
errCancelReceiptFetch = errors.New("receipt download canceled (requested)")
errCancelStateFetch = errors.New("state data download canceled (requested)")
errCancelHeaderProcessing = errors.New("header processing canceled (requested)")
errCancelContentProcessing = errors.New("content processing canceled (requested)")
errNoSyncActive = errors.New("no sync active")
errTooOld = errors.New("peer doesn't speak recent enough protocol version (need version >= 62)")
)
2015-04-12 13:38:25 +03:00
type Downloader struct {
mode SyncMode // Synchronisation mode defining the strategy used (per sync cycle)
mux *event.TypeMux // Event multiplexer to announce sync operation events
checkpoint uint64 // Checkpoint block number to enforce head against (e.g. fast sync)
genesis uint64 // Genesis block number to limit sync to (e.g. light client CHT)
queue *queue // Scheduler for selecting the hashes to download
peers *peerSet // Set of active peers from which download can proceed
stateDB ethdb.Database // Database to state sync into (and deduplicate via)
stateBloom *trie.SyncBloom // Bloom filter for fast trie node existence checks
2015-04-12 13:38:25 +03:00
rttEstimate uint64 // Round trip time to target for download requests
rttConfidence uint64 // Confidence in the estimated RTT (unit: millionths to allow atomic ops)
// Statistics
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
syncStatsChainOrigin uint64 // Origin block number where syncing started at
syncStatsChainHeight uint64 // Highest block number known when syncing started
syncStatsState stateSyncStats
syncStatsLock sync.RWMutex // Lock protecting the sync stats fields
lightchain LightChain
blockchain BlockChain
// Callbacks
dropPeer peerDropFn // Drops a peer for misbehaving
2015-04-12 13:38:25 +03:00
// Status
synchroniseMock func(id string, hash common.Hash) error // Replacement for synchronise during testing
synchronising int32
notified int32
committed int32
ancientLimit uint64 // The maximum block number which can be regarded as ancient data.
// Channels
headerCh chan dataPack // [eth/62] Channel receiving inbound block headers
bodyCh chan dataPack // [eth/62] Channel receiving inbound block bodies
receiptCh chan dataPack // [eth/63] Channel receiving inbound receipts
bodyWakeCh chan bool // [eth/62] Channel to signal the block body fetcher of new tasks
receiptWakeCh chan bool // [eth/63] Channel to signal the receipt fetcher of new tasks
headerProcCh chan []*types.Header // [eth/62] Channel to feed the header processor new tasks
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
// for stateFetcher
stateSyncStart chan *stateSync
trackStateReq chan *stateReq
stateCh chan dataPack // [eth/63] Channel receiving inbound node state data
// Cancellation and termination
cancelPeer string // Identifier of the peer currently being used as the master (cancel on drop)
cancelCh chan struct{} // Channel to cancel mid-flight syncs
cancelLock sync.RWMutex // Lock to protect the cancel channel and peer in delivers
cancelWg sync.WaitGroup // Make sure all fetcher goroutines have exited.
quitCh chan struct{} // Quit channel to signal termination
quitLock sync.RWMutex // Lock to prevent double closes
// Testing hooks
syncInitHook func(uint64, uint64) // Method to call upon initiating a new sync run
bodyFetchHook func([]*types.Header) // Method to call upon starting a block body fetch
receiptFetchHook func([]*types.Header) // Method to call upon starting a receipt fetch
chainInsertHook func([]*fetchResult) // Method to call upon inserting a chain of blocks (possibly in multiple invocations)
}
2017-07-05 12:42:37 +03:00
// LightChain encapsulates functions required to synchronise a light chain.
type LightChain interface {
// HasHeader verifies a header's presence in the local chain.
HasHeader(common.Hash, uint64) bool
// GetHeaderByHash retrieves a header from the local chain.
GetHeaderByHash(common.Hash) *types.Header
// CurrentHeader retrieves the head header from the local chain.
CurrentHeader() *types.Header
// GetTd returns the total difficulty of a local block.
GetTd(common.Hash, uint64) *big.Int
// InsertHeaderChain inserts a batch of headers into the local chain.
InsertHeaderChain([]*types.Header, int) (int, error)
// Rollback removes a few recently added elements from the local chain.
Rollback([]common.Hash)
}
2017-07-05 12:42:37 +03:00
// BlockChain encapsulates functions required to sync a (full or fast) blockchain.
type BlockChain interface {
LightChain
// HasBlock verifies a block's presence in the local chain.
HasBlock(common.Hash, uint64) bool
// HasFastBlock verifies a fast block's presence in the local chain.
HasFastBlock(common.Hash, uint64) bool
// GetBlockByHash retrieves a block from the local chain.
GetBlockByHash(common.Hash) *types.Block
// CurrentBlock retrieves the head block from the local chain.
CurrentBlock() *types.Block
// CurrentFastBlock retrieves the head fast block from the local chain.
CurrentFastBlock() *types.Block
// FastSyncCommitHead directly commits the head block to a certain entity.
FastSyncCommitHead(common.Hash) error
// InsertChain inserts a batch of blocks into the local chain.
InsertChain(types.Blocks) (int, error)
// InsertReceiptChain inserts a batch of receipts into the local chain.
InsertReceiptChain(types.Blocks, []types.Receipts, uint64) (int, error)
}
// New creates a new downloader to fetch hashes and blocks from remote peers.
func New(checkpoint uint64, stateDb ethdb.Database, stateBloom *trie.SyncBloom, mux *event.TypeMux, chain BlockChain, lightchain LightChain, dropPeer peerDropFn) *Downloader {
if lightchain == nil {
lightchain = chain
}
dl := &Downloader{
stateDB: stateDb,
stateBloom: stateBloom,
mux: mux,
checkpoint: checkpoint,
queue: newQueue(),
peers: newPeerSet(),
rttEstimate: uint64(rttMaxEstimate),
rttConfidence: uint64(1000000),
blockchain: chain,
lightchain: lightchain,
dropPeer: dropPeer,
headerCh: make(chan dataPack, 1),
bodyCh: make(chan dataPack, 1),
receiptCh: make(chan dataPack, 1),
bodyWakeCh: make(chan bool, 1),
receiptWakeCh: make(chan bool, 1),
headerProcCh: make(chan []*types.Header, 1),
quitCh: make(chan struct{}),
stateCh: make(chan dataPack),
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
stateSyncStart: make(chan *stateSync),
syncStatsState: stateSyncStats{
processed: rawdb.ReadFastTrieProgress(stateDb),
},
trackStateReq: make(chan *stateReq),
2015-04-12 13:38:25 +03:00
}
go dl.qosTuner()
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
go dl.stateFetcher()
return dl
2015-04-12 13:38:25 +03:00
}
// Progress retrieves the synchronisation boundaries, specifically the origin
// block where synchronisation started at (may have failed/suspended); the block
// or header sync is currently at; and the latest known block which the sync targets.
//
2016-03-15 20:27:49 +02:00
// In addition, during the state download phase of fast synchronisation the number
// of processed and the total number of known states are also returned. Otherwise
// these are zero.
func (d *Downloader) Progress() ethereum.SyncProgress {
// Lock the current stats and return the progress
d.syncStatsLock.RLock()
defer d.syncStatsLock.RUnlock()
current := uint64(0)
switch {
case d.blockchain != nil && d.mode == FullSync:
current = d.blockchain.CurrentBlock().NumberU64()
case d.blockchain != nil && d.mode == FastSync:
current = d.blockchain.CurrentFastBlock().NumberU64()
case d.lightchain != nil:
current = d.lightchain.CurrentHeader().Number.Uint64()
default:
log.Error("Unknown downloader chain/mode combo", "light", d.lightchain != nil, "full", d.blockchain != nil, "mode", d.mode)
}
return ethereum.SyncProgress{
StartingBlock: d.syncStatsChainOrigin,
CurrentBlock: current,
HighestBlock: d.syncStatsChainHeight,
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
PulledStates: d.syncStatsState.processed,
KnownStates: d.syncStatsState.processed + d.syncStatsState.pending,
}
2015-04-19 22:45:58 +03:00
}
// Synchronising returns whether the downloader is currently retrieving blocks.
func (d *Downloader) Synchronising() bool {
return atomic.LoadInt32(&d.synchronising) > 0
}
// RegisterPeer injects a new download peer into the set of block source to be
// used for fetching hashes and blocks from.
func (d *Downloader) RegisterPeer(id string, version int, peer Peer) error {
logger := log.New("peer", id)
logger.Trace("Registering sync peer")
if err := d.peers.Register(newPeerConnection(id, version, peer, logger)); err != nil {
logger.Error("Failed to register sync peer", "err", err)
return err
}
d.qosReduceConfidence()
2015-04-12 13:38:25 +03:00
return nil
}
2017-07-05 12:42:37 +03:00
// RegisterLightPeer injects a light client peer, wrapping it so it appears as a regular peer.
func (d *Downloader) RegisterLightPeer(id string, version int, peer LightPeer) error {
return d.RegisterPeer(id, version, &lightPeerWrapper{peer})
}
// UnregisterPeer remove a peer from the known list, preventing any action from
// the specified peer. An effort is also made to return any pending fetches into
// the queue.
func (d *Downloader) UnregisterPeer(id string) error {
// Unregister the peer from the active peer set and revoke any fetch tasks
logger := log.New("peer", id)
logger.Trace("Unregistering sync peer")
if err := d.peers.Unregister(id); err != nil {
logger.Error("Failed to unregister sync peer", "err", err)
return err
}
d.queue.Revoke(id)
// If this peer was the master peer, abort sync immediately
d.cancelLock.RLock()
master := id == d.cancelPeer
d.cancelLock.RUnlock()
if master {
d.cancel()
}
return nil
2015-04-12 13:38:25 +03:00
}
// Synchronise tries to sync up our local block chain with a remote peer, both
// adding various sanity checks as well as wrapping it with various log entries.
func (d *Downloader) Synchronise(id string, head common.Hash, td *big.Int, mode SyncMode) error {
err := d.synchronise(id, head, td, mode)
switch err {
case nil:
case errBusy:
case errTimeout, errBadPeer, errStallingPeer, errUnsyncedPeer,
errEmptyHeaderSet, errPeersUnavailable, errTooOld,
errInvalidAncestor, errInvalidChain:
log.Warn("Synchronisation failed, dropping peer", "peer", id, "err", err)
if d.dropPeer == nil {
// The dropPeer method is nil when `--copydb` is used for a local copy.
// Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored
log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", id)
} else {
d.dropPeer(id)
}
default:
log.Warn("Synchronisation failed, retrying", "err", err)
}
return err
}
// synchronise will select the peer and use it for synchronising. If an empty string is given
// it will use the best peer possible and synchronize if its TD is higher than our own. If any of the
// checks fail an error will be returned. This method is synchronous
func (d *Downloader) synchronise(id string, hash common.Hash, td *big.Int, mode SyncMode) error {
2016-03-15 20:27:49 +02:00
// Mock out the synchronisation if testing
if d.synchroniseMock != nil {
return d.synchroniseMock(id, hash)
}
// Make sure only one goroutine is ever allowed past this point at once
if !atomic.CompareAndSwapInt32(&d.synchronising, 0, 1) {
return errBusy
}
defer atomic.StoreInt32(&d.synchronising, 0)
// Post a user notification of the sync (only once per session)
if atomic.CompareAndSwapInt32(&d.notified, 0, 1) {
log.Info("Block synchronisation started")
}
// If we are already full syncing, but have a fast-sync bloom filter laying
// around, make sure it does't use memory any more. This is a special case
// when the user attempts to fast sync a new empty network.
if mode == FullSync && d.stateBloom != nil {
d.stateBloom.Close()
}
// Reset the queue, peer set and wake channels to clean any internal leftover state
d.queue.Reset()
d.peers.Reset()
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
select {
case <-ch:
default:
}
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for _, ch := range []chan dataPack{d.headerCh, d.bodyCh, d.receiptCh} {
for empty := false; !empty; {
select {
case <-ch:
default:
empty = true
}
}
}
for empty := false; !empty; {
select {
case <-d.headerProcCh:
default:
empty = true
}
}
// Create cancel channel for aborting mid-flight and mark the master peer
d.cancelLock.Lock()
d.cancelCh = make(chan struct{})
d.cancelPeer = id
d.cancelLock.Unlock()
defer d.Cancel() // No matter what, we can't leave the cancel channel open
// Set the requested sync mode, unless it's forbidden
d.mode = mode
// Retrieve the origin peer and initiate the downloading process
p := d.peers.Peer(id)
if p == nil {
return errUnknownPeer
}
return d.syncWithPeer(p, hash, td)
}
// syncWithPeer starts a block synchronization based on the hash chain from the
// specified peer and head hash.
func (d *Downloader) syncWithPeer(p *peerConnection, hash common.Hash, td *big.Int) (err error) {
d.mux.Post(StartEvent{})
defer func() {
// reset on error
if err != nil {
d.mux.Post(FailedEvent{err})
} else {
cmd,eth: 16400 Add an option to stop geth once in sync. WIP for light mode (#17321) * cmd, eth: Added in the flag to step geth once sync based on input * cmd, eth: 16400 Add an option to stop geth once in sync. * cmd: 16400 Add an option to stop geth once in sync. WIP * cmd/geth/main, les/fletcher: added in light mode support * cmd/geth/main, les/fletcher: Cleaned Comments and code for light mode * cmd: 16400 Fixed formatting issue and cleaned code * cmd, eth, les: 16400 Fixed formatting issues * cmd, eth, les: Performed gofmt to update formatting * cmd, eth, les: Fixed bugs resulting formatting * cmd/geth, eth/, les: switched to downloader event * eth: Fixed styling and gen_config * eth/: Fix nil error in config file * cmd/geth: Updated countdown log * les/fetcher.go: Removed depcreated channel * eth/downloader.go: Removed deprecated select * cmd/geth, cmd/utils: Fixed minor issues * eth: Reverted config files to proper format * eth: Fixed typo in config file * cmd/geth, eth/down: Updated code to use header time stamp * eth/downloader: Changed the time threshold to 10 minutes * cmd/geth, eth/downloader: Updated downloading event to pass latest header * cmd/geth: Updated main to use right timer object * cmd/geth: Removed unused failed event * cmd/geth: added in correct time field with type assertion * cmd/geth, cmd/utils: Updated flag to use boolean * cmd/geth, cmd/utils, eth/downloader: Cleaned up code based on recommendations * cmd/geth: Removed unneeded import * cmd/geth, eth/downloader: fixed event field and suggested changes * cmd/geth, cmd/utils: Updated flag and linting issue
2019-01-30 10:40:36 +03:00
latest := d.lightchain.CurrentHeader()
d.mux.Post(DoneEvent{latest})
}
}()
if p.version < 62 {
return errTooOld
}
log.Debug("Synchronising with the network", "peer", p.id, "eth", p.version, "head", hash, "td", td, "mode", d.mode)
defer func(start time.Time) {
log.Debug("Synchronisation terminated", "elapsed", common.PrettyDuration(time.Since(start)))
}(time.Now())
// Look up the sync boundaries: the common ancestor and the target block
latest, err := d.fetchHeight(p)
if err != nil {
return err
}
height := latest.Number.Uint64()
2018-11-12 16:18:56 +03:00
origin, err := d.findAncestor(p, latest)
if err != nil {
return err
}
d.syncStatsLock.Lock()
if d.syncStatsChainHeight <= origin || d.syncStatsChainOrigin > origin {
d.syncStatsChainOrigin = origin
}
d.syncStatsChainHeight = height
d.syncStatsLock.Unlock()
// Ensure our origin point is below any fast sync pivot point
pivot := uint64(0)
if d.mode == FastSync {
if height <= uint64(fsMinFullBlocks) {
origin = 0
} else {
pivot = height - uint64(fsMinFullBlocks)
if pivot <= origin {
origin = pivot - 1
}
}
}
d.committed = 1
if d.mode == FastSync && pivot != 0 {
d.committed = 0
}
if d.mode == FastSync {
// Set the ancient data limitation.
// If we are running fast sync, all block data older than ancientLimit will be
// written to the ancient store. More recent data will be written to the active
// database and will wait for the freezer to migrate.
//
// If there is a checkpoint available, then calculate the ancientLimit through
// that. Otherwise calculate the ancient limit through the advertised height
// of the remote peer.
//
// The reason for picking checkpoint first is that a malicious peer can give us
// a fake (very high) height, forcing the ancient limit to also be very high.
// The peer would start to feed us valid blocks until head, resulting in all of
// the blocks might be written into the ancient store. A following mini-reorg
// could cause issues.
if d.checkpoint != 0 && d.checkpoint > maxForkAncestry+1 {
d.ancientLimit = d.checkpoint
} else if height > maxForkAncestry+1 {
d.ancientLimit = height - maxForkAncestry - 1
}
frozen, _ := d.stateDB.Ancients() // Ignore the error here since light client can also hit here.
// If a part of blockchain data has already been written into active store,
// disable the ancient style insertion explicitly.
if origin >= frozen && frozen != 0 {
d.ancientLimit = 0
log.Info("Disabling direct-ancient mode", "origin", origin, "ancient", frozen-1)
} else if d.ancientLimit > 0 {
log.Debug("Enabling direct-ancient mode", "ancient", d.ancientLimit)
}
// Rewind the ancient store and blockchain if reorg happens.
if origin+1 < frozen {
var hashes []common.Hash
for i := origin + 1; i < d.lightchain.CurrentHeader().Number.Uint64(); i++ {
hashes = append(hashes, rawdb.ReadCanonicalHash(d.stateDB, i))
}
d.lightchain.Rollback(hashes)
}
}
// Initiate the sync using a concurrent header and content retrieval algorithm
d.queue.Prepare(origin+1, d.mode)
if d.syncInitHook != nil {
d.syncInitHook(origin, height)
2015-04-12 13:38:25 +03:00
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
fetchers := []func() error{
func() error { return d.fetchHeaders(p, origin+1, pivot) }, // Headers are always retrieved
func() error { return d.fetchBodies(origin + 1) }, // Bodies are retrieved during normal and fast sync
func() error { return d.fetchReceipts(origin + 1) }, // Receipts are retrieved during fast sync
func() error { return d.processHeaders(origin+1, pivot, td) },
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
if d.mode == FastSync {
fetchers = append(fetchers, func() error { return d.processFastSyncContent(latest) })
} else if d.mode == FullSync {
fetchers = append(fetchers, d.processFullSyncContent)
}
return d.spawnSync(fetchers)
}
// spawnSync runs d.process and all given fetcher functions to completion in
// separate goroutines, returning the first error that appears.
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
func (d *Downloader) spawnSync(fetchers []func() error) error {
errc := make(chan error, len(fetchers))
d.cancelWg.Add(len(fetchers))
for _, fn := range fetchers {
fn := fn
go func() { defer d.cancelWg.Done(); errc <- fn() }()
}
// Wait for the first error, then terminate the others.
var err error
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for i := 0; i < len(fetchers); i++ {
if i == len(fetchers)-1 {
// Close the queue when all fetchers have exited.
// This will cause the block processor to end when
// it has processed the queue.
d.queue.Close()
}
if err = <-errc; err != nil {
break
}
}
d.queue.Close()
d.Cancel()
return err
2015-04-12 13:38:25 +03:00
}
// cancel aborts all of the operations and resets the queue. However, cancel does
// not wait for the running download goroutines to finish. This method should be
// used when cancelling the downloads from inside the downloader.
func (d *Downloader) cancel() {
// Close the current cancel channel
d.cancelLock.Lock()
if d.cancelCh != nil {
select {
case <-d.cancelCh:
// Channel was already closed
default:
close(d.cancelCh)
}
}
d.cancelLock.Unlock()
}
// Cancel aborts all of the operations and waits for all download goroutines to
// finish before returning.
func (d *Downloader) Cancel() {
d.cancel()
d.cancelWg.Wait()
d.ancientLimit = 0
log.Debug("Reset ancient limit to zero")
}
// Terminate interrupts the downloader, canceling all pending operations.
// The downloader cannot be reused after calling Terminate.
func (d *Downloader) Terminate() {
// Close the termination channel (make sure double close is allowed)
d.quitLock.Lock()
select {
case <-d.quitCh:
default:
close(d.quitCh)
}
d.quitLock.Unlock()
// Cancel any pending download requests
d.Cancel()
}
// fetchHeight retrieves the head header of the remote peer to aid in estimating
// the total time a pending synchronisation would take.
func (d *Downloader) fetchHeight(p *peerConnection) (*types.Header, error) {
p.log.Debug("Retrieving remote chain height")
// Request the advertised remote head block and wait for the response
head, _ := p.peer.Head()
go p.peer.RequestHeadersByHash(head, 1, 0, false)
ttl := d.requestTTL()
timeout := time.After(ttl)
for {
select {
case <-d.cancelCh:
return nil, errCancelBlockFetch
case packet := <-d.headerCh:
// Discard anything not from the origin peer
if packet.PeerId() != p.id {
log.Debug("Received headers from incorrect peer", "peer", packet.PeerId())
break
}
// Make sure the peer actually gave something valid
headers := packet.(*headerPack).headers
if len(headers) != 1 {
p.log.Debug("Multiple headers for single request", "headers", len(headers))
return nil, errBadPeer
}
head := headers[0]
if (d.mode == FastSync || d.mode == LightSync) && head.Number.Uint64() < d.checkpoint {
p.log.Warn("Remote head below checkpoint", "number", head.Number, "hash", head.Hash())
return nil, errUnsyncedPeer
}
p.log.Debug("Remote head header identified", "number", head.Number, "hash", head.Hash())
return head, nil
case <-timeout:
p.log.Debug("Waiting for head header timed out", "elapsed", ttl)
return nil, errTimeout
case <-d.bodyCh:
case <-d.receiptCh:
// Out of bounds delivery, ignore
}
}
}
2018-11-12 16:18:56 +03:00
// calculateRequestSpan calculates what headers to request from a peer when trying to determine the
// common ancestor.
// It returns parameters to be used for peer.RequestHeadersByNumber:
// from - starting block number
// count - number of headers to request
// skip - number of headers to skip
// and also returns 'max', the last block which is expected to be returned by the remote peers,
// given the (from,count,skip)
func calculateRequestSpan(remoteHeight, localHeight uint64) (int64, int, int, uint64) {
var (
from int
count int
MaxCount = MaxHeaderFetch / 16
)
// requestHead is the highest block that we will ask for. If requestHead is not offset,
// the highest block that we will get is 16 blocks back from head, which means we
// will fetch 14 or 15 blocks unnecessarily in the case the height difference
// between us and the peer is 1-2 blocks, which is most common
requestHead := int(remoteHeight) - 1
if requestHead < 0 {
requestHead = 0
}
// requestBottom is the lowest block we want included in the query
// Ideally, we want to include just below own head
requestBottom := int(localHeight - 1)
if requestBottom < 0 {
requestBottom = 0
}
totalSpan := requestHead - requestBottom
span := 1 + totalSpan/MaxCount
if span < 2 {
span = 2
}
if span > 16 {
span = 16
}
count = 1 + totalSpan/span
if count > MaxCount {
count = MaxCount
}
if count < 2 {
count = 2
}
from = requestHead - (count-1)*span
if from < 0 {
from = 0
}
max := from + (count-1)*span
return int64(from), count, span - 1, uint64(max)
}
// findAncestor tries to locate the common ancestor link of the local chain and
// a remote peers blockchain. In the general case when our node was in sync and
// on the correct chain, checking the top N links should already get us a match.
// In the rare scenario when we ended up on a long reorganisation (i.e. none of
// the head links match), we do a binary search to find the common ancestor.
2018-11-12 16:18:56 +03:00
func (d *Downloader) findAncestor(p *peerConnection, remoteHeader *types.Header) (uint64, error) {
// Figure out the valid ancestor range to prevent rewrite attacks
2018-11-12 16:18:56 +03:00
var (
floor = int64(-1)
localHeight uint64
remoteHeight = remoteHeader.Number.Uint64()
)
switch d.mode {
case FullSync:
localHeight = d.blockchain.CurrentBlock().NumberU64()
case FastSync:
localHeight = d.blockchain.CurrentFastBlock().NumberU64()
default:
localHeight = d.lightchain.CurrentHeader().Number.Uint64()
}
2018-11-12 16:18:56 +03:00
p.log.Debug("Looking for common ancestor", "local", localHeight, "remote", remoteHeight)
// Recap floor value for binary search
if localHeight >= maxForkAncestry {
// We're above the max reorg threshold, find the earliest fork point
floor = int64(localHeight - maxForkAncestry)
}
// If we're doing a light sync, ensure the floor doesn't go below the CHT, as
// all headers before that point will be missing.
if d.mode == LightSync {
// If we dont know the current CHT position, find it
if d.genesis == 0 {
header := d.lightchain.CurrentHeader()
for header != nil {
d.genesis = header.Number.Uint64()
if floor >= int64(d.genesis)-1 {
break
}
header = d.lightchain.GetHeaderByHash(header.ParentHash)
}
}
// We already know the "genesis" block number, cap floor to that
if floor < int64(d.genesis)-1 {
floor = int64(d.genesis) - 1
}
}
2018-11-12 16:18:56 +03:00
from, count, skip, max := calculateRequestSpan(remoteHeight, localHeight)
p.log.Trace("Span searching for common ancestor", "count", count, "from", from, "skip", skip)
2018-11-12 16:18:56 +03:00
go p.peer.RequestHeadersByNumber(uint64(from), count, skip, false)
// Wait for the remote response to the head fetch
number, hash := uint64(0), common.Hash{}
ttl := d.requestTTL()
timeout := time.After(ttl)
for finished := false; !finished; {
select {
case <-d.cancelCh:
return 0, errCancelHeaderFetch
case packet := <-d.headerCh:
// Discard anything not from the origin peer
if packet.PeerId() != p.id {
log.Debug("Received headers from incorrect peer", "peer", packet.PeerId())
break
}
// Make sure the peer actually gave something valid
headers := packet.(*headerPack).headers
if len(headers) == 0 {
p.log.Warn("Empty head header set")
return 0, errEmptyHeaderSet
}
// Make sure the peer's reply conforms to the request
2018-11-12 16:18:56 +03:00
for i, header := range headers {
expectNumber := from + int64(i)*int64(skip+1)
2018-11-12 16:18:56 +03:00
if number := header.Number.Int64(); number != expectNumber {
p.log.Warn("Head headers broke chain ordering", "index", i, "requested", expectNumber, "received", number)
return 0, errInvalidChain
}
}
// Check if a common ancestor was found
finished = true
for i := len(headers) - 1; i >= 0; i-- {
// Skip any headers that underflow/overflow our requested set
2018-11-12 16:18:56 +03:00
if headers[i].Number.Int64() < from || headers[i].Number.Uint64() > max {
continue
}
// Otherwise check if we already know the header or not
h := headers[i].Hash()
n := headers[i].Number.Uint64()
var known bool
switch d.mode {
case FullSync:
known = d.blockchain.HasBlock(h, n)
case FastSync:
known = d.blockchain.HasFastBlock(h, n)
default:
known = d.lightchain.HasHeader(h, n)
}
if known {
number, hash = n, h
break
}
}
case <-timeout:
p.log.Debug("Waiting for head header timed out", "elapsed", ttl)
return 0, errTimeout
case <-d.bodyCh:
case <-d.receiptCh:
// Out of bounds delivery, ignore
}
}
// If the head fetch already found an ancestor, return
if hash != (common.Hash{}) {
if int64(number) <= floor {
p.log.Warn("Ancestor below allowance", "number", number, "hash", hash, "allowance", floor)
return 0, errInvalidAncestor
}
p.log.Debug("Found common ancestor", "number", number, "hash", hash)
return number, nil
}
// Ancestor not found, we need to binary search over our chain
2018-11-12 16:18:56 +03:00
start, end := uint64(0), remoteHeight
if floor > 0 {
start = uint64(floor)
}
p.log.Trace("Binary searching for common ancestor", "start", start, "end", end)
for start+1 < end {
// Split our chain interval in two, and request the hash to cross check
check := (start + end) / 2
ttl := d.requestTTL()
timeout := time.After(ttl)
go p.peer.RequestHeadersByNumber(check, 1, 0, false)
// Wait until a reply arrives to this request
for arrived := false; !arrived; {
select {
case <-d.cancelCh:
return 0, errCancelHeaderFetch
case packer := <-d.headerCh:
// Discard anything not from the origin peer
if packer.PeerId() != p.id {
log.Debug("Received headers from incorrect peer", "peer", packer.PeerId())
break
}
// Make sure the peer actually gave something valid
headers := packer.(*headerPack).headers
if len(headers) != 1 {
p.log.Debug("Multiple headers for single request", "headers", len(headers))
return 0, errBadPeer
}
arrived = true
// Modify the search interval based on the response
h := headers[0].Hash()
n := headers[0].Number.Uint64()
var known bool
switch d.mode {
case FullSync:
known = d.blockchain.HasBlock(h, n)
case FastSync:
known = d.blockchain.HasFastBlock(h, n)
default:
known = d.lightchain.HasHeader(h, n)
}
if !known {
end = check
break
}
header := d.lightchain.GetHeaderByHash(h) // Independent of sync mode, header surely exists
if header.Number.Uint64() != check {
p.log.Debug("Received non requested header", "number", header.Number, "hash", header.Hash(), "request", check)
return 0, errBadPeer
}
start = check
hash = h
case <-timeout:
p.log.Debug("Waiting for search header timed out", "elapsed", ttl)
return 0, errTimeout
case <-d.bodyCh:
case <-d.receiptCh:
// Out of bounds delivery, ignore
}
}
}
// Ensure valid ancestry and return
if int64(start) <= floor {
p.log.Warn("Ancestor below allowance", "number", start, "hash", hash, "allowance", floor)
return 0, errInvalidAncestor
}
p.log.Debug("Found common ancestor", "number", start, "hash", hash)
return start, nil
}
// fetchHeaders keeps retrieving headers concurrently from the number
// requested, until no more are returned, potentially throttling on the way. To
// facilitate concurrency but still protect against malicious nodes sending bad
// headers, we construct a header chain skeleton using the "origin" peer we are
// syncing with, and fill in the missing headers using anyone else. Headers from
2016-05-17 11:12:57 +03:00
// other peers are only accepted if they map cleanly to the skeleton. If no one
// can fill in the skeleton - not even the origin peer - it's assumed invalid and
// the origin is dropped.
func (d *Downloader) fetchHeaders(p *peerConnection, from uint64, pivot uint64) error {
p.log.Debug("Directing header downloads", "origin", from)
defer p.log.Debug("Header download terminated")
// Create a timeout timer, and the associated header fetcher
skeleton := true // Skeleton assembly phase or finishing up
request := time.Now() // time of the last skeleton fetch request
timeout := time.NewTimer(0) // timer to dump a non-responsive active peer
<-timeout.C // timeout channel should be initially empty
defer timeout.Stop()
var ttl time.Duration
getHeaders := func(from uint64) {
2016-05-17 11:12:57 +03:00
request = time.Now()
ttl = d.requestTTL()
timeout.Reset(ttl)
2016-05-17 11:12:57 +03:00
if skeleton {
p.log.Trace("Fetching skeleton headers", "count", MaxHeaderFetch, "from", from)
go p.peer.RequestHeadersByNumber(from+uint64(MaxHeaderFetch)-1, MaxSkeletonSize, MaxHeaderFetch-1, false)
} else {
p.log.Trace("Fetching full headers", "count", MaxHeaderFetch, "from", from)
go p.peer.RequestHeadersByNumber(from, MaxHeaderFetch, 0, false)
}
}
// Start pulling the header chain skeleton until all is done
ancestor := from
getHeaders(from)
for {
select {
case <-d.cancelCh:
return errCancelHeaderFetch
case packet := <-d.headerCh:
// Make sure the active peer is giving us the skeleton headers
if packet.PeerId() != p.id {
log.Debug("Received skeleton from incorrect peer", "peer", packet.PeerId())
break
}
headerReqTimer.UpdateSince(request)
timeout.Stop()
// If the skeleton's finished, pull any remaining head headers directly from the origin
if packet.Items() == 0 && skeleton {
skeleton = false
getHeaders(from)
continue
}
// If no more headers are inbound, notify the content fetchers and return
if packet.Items() == 0 {
// Don't abort header fetches while the pivot is downloading
if atomic.LoadInt32(&d.committed) == 0 && pivot <= from {
p.log.Debug("No headers, waiting for pivot commit")
select {
case <-time.After(fsHeaderContCheck):
getHeaders(from)
continue
case <-d.cancelCh:
return errCancelHeaderFetch
}
}
// Pivot done (or not in fast sync) and no more headers, terminate the process
p.log.Debug("No more headers available")
select {
case d.headerProcCh <- nil:
return nil
case <-d.cancelCh:
return errCancelHeaderFetch
}
}
headers := packet.(*headerPack).headers
// If we received a skeleton batch, resolve internals concurrently
if skeleton {
filled, proced, err := d.fillHeaderSkeleton(from, headers)
if err != nil {
p.log.Debug("Skeleton chain invalid", "err", err)
return errInvalidChain
}
headers = filled[proced:]
from += uint64(proced)
} else {
// If we're closing in on the chain head, but haven't yet reached it, delay
// the last few headers so mini reorgs on the head don't cause invalid hash
// chain errors.
if n := len(headers); n > 0 {
// Retrieve the current head we're at
head := uint64(0)
if d.mode == LightSync {
head = d.lightchain.CurrentHeader().Number.Uint64()
} else {
head = d.blockchain.CurrentFastBlock().NumberU64()
if full := d.blockchain.CurrentBlock().NumberU64(); head < full {
head = full
}
}
// If the head is below the common ancestor, we're actually deduplicating
// already existing chain segments, so use the ancestor as the fake head.
// Otherwise we might end up delaying header deliveries pointlessly.
if head < ancestor {
head = ancestor
}
// If the head is way older than this batch, delay the last few headers
if head+uint64(reorgProtThreshold) < headers[n-1].Number.Uint64() {
delay := reorgProtHeaderDelay
if delay > n {
delay = n
}
headers = headers[:n-delay]
}
}
}
// Insert all the new headers and fetch the next batch
if len(headers) > 0 {
p.log.Trace("Scheduling new headers", "count", len(headers), "from", from)
select {
case d.headerProcCh <- headers:
case <-d.cancelCh:
return errCancelHeaderFetch
}
from += uint64(len(headers))
getHeaders(from)
} else {
// No headers delivered, or all of them being delayed, sleep a bit and retry
p.log.Trace("All headers delayed, waiting")
select {
case <-time.After(fsHeaderContCheck):
getHeaders(from)
continue
case <-d.cancelCh:
return errCancelHeaderFetch
}
}
case <-timeout.C:
if d.dropPeer == nil {
// The dropPeer method is nil when `--copydb` is used for a local copy.
// Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored
p.log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", p.id)
break
}
// Header retrieval timed out, consider the peer bad and drop
p.log.Debug("Header request timed out", "elapsed", ttl)
headerTimeoutMeter.Mark(1)
d.dropPeer(p.id)
// Finish the sync gracefully instead of dumping the gathered data though
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
select {
case ch <- false:
case <-d.cancelCh:
}
}
select {
case d.headerProcCh <- nil:
case <-d.cancelCh:
}
return errBadPeer
}
}
}
// fillHeaderSkeleton concurrently retrieves headers from all our available peers
// and maps them to the provided skeleton header chain.
//
// Any partial results from the beginning of the skeleton is (if possible) forwarded
// immediately to the header processor to keep the rest of the pipeline full even
// in the case of header stalls.
//
2018-04-04 13:25:02 +03:00
// The method returns the entire filled skeleton and also the number of headers
// already forwarded for processing.
func (d *Downloader) fillHeaderSkeleton(from uint64, skeleton []*types.Header) ([]*types.Header, int, error) {
log.Debug("Filling up skeleton", "from", from)
d.queue.ScheduleSkeleton(from, skeleton)
var (
deliver = func(packet dataPack) (int, error) {
pack := packet.(*headerPack)
return d.queue.DeliverHeaders(pack.peerID, pack.headers, d.headerProcCh)
}
expire = func() map[string]int { return d.queue.ExpireHeaders(d.requestTTL()) }
throttle = func() bool { return false }
reserve = func(p *peerConnection, count int) (*fetchRequest, bool, error) {
return d.queue.ReserveHeaders(p, count), false, nil
}
fetch = func(p *peerConnection, req *fetchRequest) error { return p.FetchHeaders(req.From, MaxHeaderFetch) }
capacity = func(p *peerConnection) int { return p.HeaderCapacity(d.requestRTT()) }
setIdle = func(p *peerConnection, accepted int) { p.SetHeadersIdle(accepted) }
)
err := d.fetchParts(errCancelHeaderFetch, d.headerCh, deliver, d.queue.headerContCh, expire,
d.queue.PendingHeaders, d.queue.InFlightHeaders, throttle, reserve,
nil, fetch, d.queue.CancelHeaders, capacity, d.peers.HeaderIdlePeers, setIdle, "headers")
log.Debug("Skeleton fill terminated", "err", err)
filled, proced := d.queue.RetrieveHeaders()
return filled, proced, err
}
// fetchBodies iteratively downloads the scheduled block bodies, taking any
// available peers, reserving a chunk of blocks for each, waiting for delivery
// and also periodically checking for timeouts.
func (d *Downloader) fetchBodies(from uint64) error {
log.Debug("Downloading block bodies", "origin", from)
var (
deliver = func(packet dataPack) (int, error) {
pack := packet.(*bodyPack)
return d.queue.DeliverBodies(pack.peerID, pack.transactions, pack.uncles)
}
expire = func() map[string]int { return d.queue.ExpireBodies(d.requestTTL()) }
fetch = func(p *peerConnection, req *fetchRequest) error { return p.FetchBodies(req) }
capacity = func(p *peerConnection) int { return p.BlockCapacity(d.requestRTT()) }
setIdle = func(p *peerConnection, accepted int) { p.SetBodiesIdle(accepted) }
)
err := d.fetchParts(errCancelBodyFetch, d.bodyCh, deliver, d.bodyWakeCh, expire,
d.queue.PendingBlocks, d.queue.InFlightBlocks, d.queue.ShouldThrottleBlocks, d.queue.ReserveBodies,
d.bodyFetchHook, fetch, d.queue.CancelBodies, capacity, d.peers.BodyIdlePeers, setIdle, "bodies")
log.Debug("Block body download terminated", "err", err)
return err
}
// fetchReceipts iteratively downloads the scheduled block receipts, taking any
// available peers, reserving a chunk of receipts for each, waiting for delivery
// and also periodically checking for timeouts.
func (d *Downloader) fetchReceipts(from uint64) error {
log.Debug("Downloading transaction receipts", "origin", from)
var (
deliver = func(packet dataPack) (int, error) {
pack := packet.(*receiptPack)
return d.queue.DeliverReceipts(pack.peerID, pack.receipts)
}
expire = func() map[string]int { return d.queue.ExpireReceipts(d.requestTTL()) }
fetch = func(p *peerConnection, req *fetchRequest) error { return p.FetchReceipts(req) }
capacity = func(p *peerConnection) int { return p.ReceiptCapacity(d.requestRTT()) }
setIdle = func(p *peerConnection, accepted int) { p.SetReceiptsIdle(accepted) }
)
err := d.fetchParts(errCancelReceiptFetch, d.receiptCh, deliver, d.receiptWakeCh, expire,
d.queue.PendingReceipts, d.queue.InFlightReceipts, d.queue.ShouldThrottleReceipts, d.queue.ReserveReceipts,
d.receiptFetchHook, fetch, d.queue.CancelReceipts, capacity, d.peers.ReceiptIdlePeers, setIdle, "receipts")
log.Debug("Transaction receipt download terminated", "err", err)
return err
}
// fetchParts iteratively downloads scheduled block parts, taking any available
// peers, reserving a chunk of fetch requests for each, waiting for delivery and
// also periodically checking for timeouts.
2016-05-17 11:12:57 +03:00
//
// As the scheduling/timeout logic mostly is the same for all downloaded data
// types, this method is used by each for data gathering and is instrumented with
// various callbacks to handle the slight differences between processing them.
//
// The instrumentation parameters:
// - errCancel: error type to return if the fetch operation is cancelled (mostly makes logging nicer)
// - deliveryCh: channel from which to retrieve downloaded data packets (merged from all concurrent peers)
// - deliver: processing callback to deliver data packets into type specific download queues (usually within `queue`)
// - wakeCh: notification channel for waking the fetcher when new tasks are available (or sync completed)
// - expire: task callback method to abort requests that took too long and return the faulty peers (traffic shaping)
// - pending: task callback for the number of requests still needing download (detect completion/non-completability)
// - inFlight: task callback for the number of in-progress requests (wait for all active downloads to finish)
// - throttle: task callback to check if the processing queue is full and activate throttling (bound memory use)
// - reserve: task callback to reserve new download tasks to a particular peer (also signals partial completions)
// - fetchHook: tester callback to notify of new tasks being initiated (allows testing the scheduling logic)
// - fetch: network callback to actually send a particular download request to a physical remote peer
// - cancel: task callback to abort an in-flight download request and allow rescheduling it (in case of lost peer)
2017-01-06 20:44:35 +03:00
// - capacity: network callback to retrieve the estimated type-specific bandwidth capacity of a peer (traffic shaping)
2016-05-17 11:12:57 +03:00
// - idle: network callback to retrieve the currently (type specific) idle peers that can be assigned tasks
// - setIdle: network callback to set a peer back to idle and update its estimated capacity (traffic shaping)
// - kind: textual label of the type being downloaded to display in log mesages
func (d *Downloader) fetchParts(errCancel error, deliveryCh chan dataPack, deliver func(dataPack) (int, error), wakeCh chan bool,
expire func() map[string]int, pending func() int, inFlight func() bool, throttle func() bool, reserve func(*peerConnection, int) (*fetchRequest, bool, error),
fetchHook func([]*types.Header), fetch func(*peerConnection, *fetchRequest) error, cancel func(*fetchRequest), capacity func(*peerConnection) int,
idle func() ([]*peerConnection, int), setIdle func(*peerConnection, int), kind string) error {
// Create a ticker to detect expired retrieval tasks
ticker := time.NewTicker(100 * time.Millisecond)
defer ticker.Stop()
update := make(chan struct{}, 1)
// Prepare the queue and fetch block parts until the block header fetcher's done
finished := false
for {
select {
case <-d.cancelCh:
return errCancel
case packet := <-deliveryCh:
// If the peer was previously banned and failed to deliver its pack
// in a reasonable time frame, ignore its message.
if peer := d.peers.Peer(packet.PeerId()); peer != nil {
// Deliver the received chunk of data and check chain validity
accepted, err := deliver(packet)
if err == errInvalidChain {
return err
}
// Unless a peer delivered something completely else than requested (usually
// caused by a timed out request which came through in the end), set it to
// idle. If the delivery's stale, the peer should have already been idled.
if err != errStaleDelivery {
setIdle(peer, accepted)
}
// Issue a log to the user to see what's going on
switch {
case err == nil && packet.Items() == 0:
peer.log.Trace("Requested data not delivered", "type", kind)
case err == nil:
peer.log.Trace("Delivered new batch of data", "type", kind, "count", packet.Stats())
default:
peer.log.Trace("Failed to deliver retrieved data", "type", kind, "err", err)
}
}
// Blocks assembled, try to update the progress
select {
case update <- struct{}{}:
default:
}
case cont := <-wakeCh:
// The header fetcher sent a continuation flag, check if it's done
if !cont {
finished = true
}
// Headers arrive, try to update the progress
select {
case update <- struct{}{}:
default:
}
case <-ticker.C:
// Sanity check update the progress
select {
case update <- struct{}{}:
default:
}
case <-update:
// Short circuit if we lost all our peers
if d.peers.Len() == 0 {
return errNoPeers
}
// Check for fetch request timeouts and demote the responsible peers
for pid, fails := range expire() {
if peer := d.peers.Peer(pid); peer != nil {
// If a lot of retrieval elements expired, we might have overestimated the remote peer or perhaps
// ourselves. Only reset to minimal throughput but don't drop just yet. If even the minimal times
// out that sync wise we need to get rid of the peer.
//
// The reason the minimum threshold is 2 is because the downloader tries to estimate the bandwidth
// and latency of a peer separately, which requires pushing the measures capacity a bit and seeing
// how response times reacts, to it always requests one more than the minimum (i.e. min 2).
if fails > 2 {
peer.log.Trace("Data delivery timed out", "type", kind)
setIdle(peer, 0)
} else {
peer.log.Debug("Stalling delivery, dropping", "type", kind)
if d.dropPeer == nil {
// The dropPeer method is nil when `--copydb` is used for a local copy.
// Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored
peer.log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", pid)
} else {
d.dropPeer(pid)
}
}
}
}
// If there's nothing more to fetch, wait or terminate
if pending() == 0 {
if !inFlight() && finished {
log.Debug("Data fetching completed", "type", kind)
return nil
}
break
}
// Send a download request to all idle peers, until throttled
progressed, throttled, running := false, false, inFlight()
idles, total := idle()
for _, peer := range idles {
// Short circuit if throttling activated
if throttle() {
throttled = true
break
}
// Short circuit if there is no more available task.
if pending() == 0 {
break
}
// Reserve a chunk of fetches for a peer. A nil can mean either that
// no more headers are available, or that the peer is known not to
// have them.
request, progress, err := reserve(peer, capacity(peer))
if err != nil {
return err
}
if progress {
progressed = true
}
if request == nil {
continue
}
if request.From > 0 {
peer.log.Trace("Requesting new batch of data", "type", kind, "from", request.From)
} else {
peer.log.Trace("Requesting new batch of data", "type", kind, "count", len(request.Headers), "from", request.Headers[0].Number)
}
// Fetch the chunk and make sure any errors return the hashes to the queue
if fetchHook != nil {
fetchHook(request.Headers)
}
if err := fetch(peer, request); err != nil {
// Although we could try and make an attempt to fix this, this error really
// means that we've double allocated a fetch task to a peer. If that is the
// case, the internal state of the downloader and the queue is very wrong so
// better hard crash and note the error instead of silently accumulating into
// a much bigger issue.
panic(fmt.Sprintf("%v: %s fetch assignment failed", peer, kind))
}
running = true
}
// Make sure that we have peers available for fetching. If all peers have been tried
// and all failed throw an error
if !progressed && !throttled && !running && len(idles) == total && pending() > 0 {
return errPeersUnavailable
}
}
}
}
// processHeaders takes batches of retrieved headers from an input channel and
// keeps processing and scheduling them into the header chain and downloader's
// queue until the stream ends or a failure occurs.
func (d *Downloader) processHeaders(origin uint64, pivot uint64, td *big.Int) error {
// Keep a count of uncertain headers to roll back
var rollback []*types.Header
defer func() {
if len(rollback) > 0 {
// Flatten the headers and roll them back
hashes := make([]common.Hash, len(rollback))
for i, header := range rollback {
hashes[i] = header.Hash()
}
lastHeader, lastFastBlock, lastBlock := d.lightchain.CurrentHeader().Number, common.Big0, common.Big0
if d.mode != LightSync {
lastFastBlock = d.blockchain.CurrentFastBlock().Number()
lastBlock = d.blockchain.CurrentBlock().Number()
}
d.lightchain.Rollback(hashes)
curFastBlock, curBlock := common.Big0, common.Big0
if d.mode != LightSync {
curFastBlock = d.blockchain.CurrentFastBlock().Number()
curBlock = d.blockchain.CurrentBlock().Number()
}
log.Warn("Rolled back headers", "count", len(hashes),
"header", fmt.Sprintf("%d->%d", lastHeader, d.lightchain.CurrentHeader().Number),
"fast", fmt.Sprintf("%d->%d", lastFastBlock, curFastBlock),
"block", fmt.Sprintf("%d->%d", lastBlock, curBlock))
}
}()
// Wait for batches of headers to process
gotHeaders := false
for {
select {
case <-d.cancelCh:
return errCancelHeaderProcessing
case headers := <-d.headerProcCh:
// Terminate header processing if we synced up
if len(headers) == 0 {
// Notify everyone that headers are fully processed
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
select {
case ch <- false:
case <-d.cancelCh:
}
}
// If no headers were retrieved at all, the peer violated its TD promise that it had a
// better chain compared to ours. The only exception is if its promised blocks were
// already imported by other means (e.g. fetcher):
//
// R <remote peer>, L <local node>: Both at block 10
// R: Mine block 11, and propagate it to L
// L: Queue block 11 for import
// L: Notice that R's head and TD increased compared to ours, start sync
// L: Import of block 11 finishes
// L: Sync begins, and finds common ancestor at 11
// L: Request new headers up from 11 (R's TD was higher, it must have something)
// R: Nothing to give
if d.mode != LightSync {
head := d.blockchain.CurrentBlock()
if !gotHeaders && td.Cmp(d.blockchain.GetTd(head.Hash(), head.NumberU64())) > 0 {
return errStallingPeer
}
}
// If fast or light syncing, ensure promised headers are indeed delivered. This is
// needed to detect scenarios where an attacker feeds a bad pivot and then bails out
// of delivering the post-pivot blocks that would flag the invalid content.
//
// This check cannot be executed "as is" for full imports, since blocks may still be
// queued for processing when the header download completes. However, as long as the
// peer gave us something useful, we're already happy/progressed (above check).
if d.mode == FastSync || d.mode == LightSync {
head := d.lightchain.CurrentHeader()
if td.Cmp(d.lightchain.GetTd(head.Hash(), head.Number.Uint64())) > 0 {
return errStallingPeer
}
}
// Disable any rollback and return
rollback = nil
return nil
}
// Otherwise split the chunk of headers into batches and process them
gotHeaders = true
for len(headers) > 0 {
// Terminate if something failed in between processing chunks
select {
case <-d.cancelCh:
return errCancelHeaderProcessing
default:
}
// Select the next chunk of headers to import
limit := maxHeadersProcess
if limit > len(headers) {
limit = len(headers)
}
chunk := headers[:limit]
// In case of header only syncing, validate the chunk immediately
if d.mode == FastSync || d.mode == LightSync {
// Collect the yet unknown headers to mark them as uncertain
unknown := make([]*types.Header, 0, len(chunk))
for _, header := range chunk {
if !d.lightchain.HasHeader(header.Hash(), header.Number.Uint64()) {
unknown = append(unknown, header)
}
}
// If we're importing pure headers, verify based on their recentness
frequency := fsHeaderCheckFrequency
if chunk[len(chunk)-1].Number.Uint64()+uint64(fsHeaderForceVerify) > pivot {
frequency = 1
}
if n, err := d.lightchain.InsertHeaderChain(chunk, frequency); err != nil {
// If some headers were inserted, add them too to the rollback list
if n > 0 {
rollback = append(rollback, chunk[:n]...)
}
log.Debug("Invalid header encountered", "number", chunk[n].Number, "hash", chunk[n].Hash(), "err", err)
return errInvalidChain
}
// All verifications passed, store newly found uncertain headers
rollback = append(rollback, unknown...)
if len(rollback) > fsHeaderSafetyNet {
rollback = append(rollback[:0], rollback[len(rollback)-fsHeaderSafetyNet:]...)
}
}
// Unless we're doing light chains, schedule the headers for associated content retrieval
if d.mode == FullSync || d.mode == FastSync {
// If we've reached the allowed number of pending headers, stall a bit
for d.queue.PendingBlocks() >= maxQueuedHeaders || d.queue.PendingReceipts() >= maxQueuedHeaders {
select {
case <-d.cancelCh:
return errCancelHeaderProcessing
case <-time.After(time.Second):
}
}
// Otherwise insert the headers for content retrieval
inserts := d.queue.Schedule(chunk, origin)
if len(inserts) != len(chunk) {
log.Debug("Stale headers")
return errBadPeer
}
}
headers = headers[limit:]
origin += uint64(limit)
}
// Update the highest block number we know if a higher one is found.
d.syncStatsLock.Lock()
if d.syncStatsChainHeight < origin {
d.syncStatsChainHeight = origin - 1
}
d.syncStatsLock.Unlock()
// Signal the content downloaders of the availablility of new tasks
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
select {
case ch <- true:
default:
}
}
}
}
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
// processFullSyncContent takes fetch results from the queue and imports them into the chain.
func (d *Downloader) processFullSyncContent() error {
for {
results := d.queue.Results(true)
if len(results) == 0 {
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
return nil
}
if d.chainInsertHook != nil {
d.chainInsertHook(results)
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
if err := d.importBlockResults(results); err != nil {
return err
}
}
}
func (d *Downloader) importBlockResults(results []*fetchResult) error {
// Check for any early termination requests
if len(results) == 0 {
return nil
}
select {
case <-d.quitCh:
return errCancelContentProcessing
default:
}
// Retrieve the a batch of results to import
first, last := results[0].Header, results[len(results)-1].Header
log.Debug("Inserting downloaded chain", "items", len(results),
"firstnum", first.Number, "firsthash", first.Hash(),
"lastnum", last.Number, "lasthash", last.Hash(),
)
blocks := make([]*types.Block, len(results))
for i, result := range results {
blocks[i] = types.NewBlockWithHeader(result.Header).WithBody(result.Transactions, result.Uncles)
}
if index, err := d.blockchain.InsertChain(blocks); err != nil {
if index < len(results) {
log.Debug("Downloaded item processing failed", "number", results[index].Header.Number, "hash", results[index].Header.Hash(), "err", err)
} else {
// The InsertChain method in blockchain.go will sometimes return an out-of-bounds index,
// when it needs to preprocess blocks to import a sidechain.
// The importer will put together a new list of blocks to import, which is a superset
// of the blocks delivered from the downloader, and the indexing will be off.
log.Debug("Downloaded item processing failed on sidechain import", "index", index, "err", err)
}
return errInvalidChain
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
return nil
}
// processFastSyncContent takes fetch results from the queue and writes them to the
// database. It also controls the synchronisation of state nodes of the pivot block.
func (d *Downloader) processFastSyncContent(latest *types.Header) error {
// Start syncing state of the reported head block. This should get us most of
// the state of the pivot block.
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
stateSync := d.syncState(latest.Root)
defer stateSync.Cancel()
go func() {
if err := stateSync.Wait(); err != nil && err != errCancelStateFetch {
d.queue.Close() // wake up Results
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
}()
// Figure out the ideal pivot block. Note, that this goalpost may move if the
// sync takes long enough for the chain head to move significantly.
pivot := uint64(0)
if height := latest.Number.Uint64(); height > uint64(fsMinFullBlocks) {
pivot = height - uint64(fsMinFullBlocks)
}
// To cater for moving pivot points, track the pivot block and subsequently
2018-04-04 13:25:02 +03:00
// accumulated download results separately.
var (
oldPivot *fetchResult // Locked in pivot block, might change eventually
oldTail []*fetchResult // Downloaded content after the pivot
)
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
for {
// Wait for the next batch of downloaded data to be available, and if the pivot
// block became stale, move the goalpost
results := d.queue.Results(oldPivot == nil) // Block if we're not monitoring pivot staleness
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
if len(results) == 0 {
// If pivot sync is done, stop
if oldPivot == nil {
return stateSync.Cancel()
}
// If sync failed, stop
select {
case <-d.cancelCh:
return stateSync.Cancel()
default:
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
if d.chainInsertHook != nil {
d.chainInsertHook(results)
}
if oldPivot != nil {
results = append(append([]*fetchResult{oldPivot}, oldTail...), results...)
}
// Split around the pivot block and process the two sides via fast/full sync
if atomic.LoadInt32(&d.committed) == 0 {
latest = results[len(results)-1].Header
if height := latest.Number.Uint64(); height > pivot+2*uint64(fsMinFullBlocks) {
log.Warn("Pivot became stale, moving", "old", pivot, "new", height-uint64(fsMinFullBlocks))
pivot = height - uint64(fsMinFullBlocks)
}
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
P, beforeP, afterP := splitAroundPivot(pivot, results)
if err := d.commitFastSyncData(beforeP, stateSync); err != nil {
return err
}
if P != nil {
// If new pivot block found, cancel old state retrieval and restart
if oldPivot != P {
stateSync.Cancel()
stateSync = d.syncState(P.Header.Root)
defer stateSync.Cancel()
go func() {
if err := stateSync.Wait(); err != nil && err != errCancelStateFetch {
d.queue.Close() // wake up Results
}
}()
oldPivot = P
}
// Wait for completion, occasionally checking for pivot staleness
select {
case <-stateSync.done:
if stateSync.err != nil {
return stateSync.err
}
if err := d.commitPivotBlock(P); err != nil {
return err
}
oldPivot = nil
case <-time.After(time.Second):
oldTail = afterP
continue
}
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
// Fast sync done, pivot commit done, full import
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
if err := d.importBlockResults(afterP); err != nil {
return err
}
}
}
func splitAroundPivot(pivot uint64, results []*fetchResult) (p *fetchResult, before, after []*fetchResult) {
for _, result := range results {
num := result.Header.Number.Uint64()
switch {
case num < pivot:
before = append(before, result)
case num == pivot:
p = result
default:
after = append(after, result)
}
}
return p, before, after
}
func (d *Downloader) commitFastSyncData(results []*fetchResult, stateSync *stateSync) error {
// Check for any early termination requests
if len(results) == 0 {
return nil
}
select {
case <-d.quitCh:
return errCancelContentProcessing
case <-stateSync.done:
if err := stateSync.Wait(); err != nil {
return err
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
default:
}
// Retrieve the a batch of results to import
first, last := results[0].Header, results[len(results)-1].Header
log.Debug("Inserting fast-sync blocks", "items", len(results),
"firstnum", first.Number, "firsthash", first.Hash(),
"lastnumn", last.Number, "lasthash", last.Hash(),
)
blocks := make([]*types.Block, len(results))
receipts := make([]types.Receipts, len(results))
for i, result := range results {
blocks[i] = types.NewBlockWithHeader(result.Header).WithBody(result.Transactions, result.Uncles)
receipts[i] = result.Receipts
}
if index, err := d.blockchain.InsertReceiptChain(blocks, receipts, d.ancientLimit); err != nil {
log.Debug("Downloaded item processing failed", "number", results[index].Header.Number, "hash", results[index].Header.Hash(), "err", err)
return errInvalidChain
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
}
return nil
}
func (d *Downloader) commitPivotBlock(result *fetchResult) error {
block := types.NewBlockWithHeader(result.Header).WithBody(result.Transactions, result.Uncles)
log.Debug("Committing fast sync pivot as new head", "number", block.Number(), "hash", block.Hash())
// Commit the pivot block as the new head, will require full sync from here on
if _, err := d.blockchain.InsertReceiptChain([]*types.Block{block}, []types.Receipts{result.Receipts}, d.ancientLimit); err != nil {
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
return err
}
if err := d.blockchain.FastSyncCommitHead(block.Hash()); err != nil {
eth/downloader: separate state sync from queue (#14460) * eth/downloader: separate state sync from queue Scheduling of state node downloads hogged the downloader queue lock when new requests were scheduled. This caused timeouts for other requests. With this change, state sync is fully independent of all other downloads and doesn't involve the queue at all. State sync is started and checked on in processContent. This is slightly awkward because processContent doesn't have a select loop. Instead, the queue is closed by an auxiliary goroutine when state sync fails. We tried several alternatives to this but settled on the current approach because it's the least amount of change overall. Handling of the pivot block has changed slightly: the queue previously prevented import of pivot block receipts before the state of the pivot block was available. In this commit, the receipt will be imported before the state. This causes an annoyance where the pivot block is committed as fast block head even when state downloads fail. Stay tuned for more updates in this area ;) * eth/downloader: remove cancelTimeout channel * eth/downloader: retry state requests on timeout * eth/downloader: improve comment * eth/downloader: mark peers idle when state sync is done * eth/downloader: move pivot block splitting to processContent This change also ensures that pivot block receipts aren't imported before the pivot block itself. * eth/downloader: limit state node retries * eth/downloader: improve state node error handling and retry check * eth/downloader: remove maxStateNodeRetries It fails the sync too much. * eth/downloader: remove last use of cancelCh in statesync.go Fixes TestDeliverHeadersHang*Fast and (hopefully) the weird cancellation behaviour at the end of fast sync. * eth/downloader: fix leak in runStateSync * eth/downloader: don't run processFullSyncContent in LightSync mode * eth/downloader: improve comments * eth/downloader: fix vet, megacheck * eth/downloader: remove unrequested tasks anyway * eth/downloader, trie: various polishes around duplicate items This commit explicitly tracks duplicate and unexpected state delieveries done against a trie Sync structure, also adding there to import info logs. The commit moves the db batch used to commit trie changes one level deeper so its flushed after every node insertion. This is needed to avoid a lot of duplicate retrievals caused by inconsistencies between Sync internals and database. A better approach is to track not-yet-written states in trie.Sync and flush on commit, but I'm focuing on correctness first now. The commit fixes a regression around pivot block fail count. The counter previously was reset to 1 if and only if a sync cycle progressed (inserted at least 1 entry to the database). The current code reset it already if a node was delivered, which is not stong enough, because unless it ends up written to disk, an attacker can just loop and attack ad infinitum. The commit also fixes a regression around state deliveries and timeouts. The old downloader tracked if a delivery is stale (none of the deliveries were requestedt), in which case it didn't mark the node idle and did not send further requests, since it signals a past timeout. The current code did mark it idle even on stale deliveries, which eventually caused two requests to be in flight at the same time, making the deliveries always stale and mass duplicating retrievals between multiple peers. * eth/downloader: fix state request leak This commit fixes the hang seen sometimes while doing the state sync. The cause of the hang was a rare combination of events: request state data from peer, peer drops and reconnects almost immediately. This caused a new download task to be assigned to the peer, overwriting the old one still waiting for a timeout, which in turned leaked the requests out, never to be retried. The fix is to ensure that a task assignment moves any pending one back into the retry queue. The commit also fixes a regression with peer dropping due to stalls. The current code considered a peer stalling if they timed out delivering 1 item. However, the downloader never requests only one, the minimum is 2 (attempt to fine tune estimated latency/bandwidth). The fix is simply to drop if a timeout is detected at 2 items. Apart from the above bugfixes, the commit contains some code polishes I made while debugging the hang. * core, eth, trie: support batched trie sync db writes * trie: rename SyncMemCache to syncMemBatch
2017-06-22 15:26:03 +03:00
return err
}
atomic.StoreInt32(&d.committed, 1)
// If we had a bloom filter for the state sync, deallocate it now. Note, we only
// deallocate internally, but keep the empty wrapper. This ensures that if we do
// a rollback after committing the pivot and restarting fast sync, we don't end
// up using a nil bloom. Empty bloom is fine, it just returns that it does not
// have the info we need, so reach down to the database instead.
if d.stateBloom != nil {
d.stateBloom.Close()
}
return nil
}
2016-03-15 20:27:49 +02:00
// DeliverHeaders injects a new batch of block headers received from a remote
// node into the download schedule.
func (d *Downloader) DeliverHeaders(id string, headers []*types.Header) (err error) {
return d.deliver(id, d.headerCh, &headerPack{id, headers}, headerInMeter, headerDropMeter)
}
// DeliverBodies injects a new batch of block bodies received from a remote node.
func (d *Downloader) DeliverBodies(id string, transactions [][]*types.Transaction, uncles [][]*types.Header) (err error) {
return d.deliver(id, d.bodyCh, &bodyPack{id, transactions, uncles}, bodyInMeter, bodyDropMeter)
}
// DeliverReceipts injects a new batch of receipts received from a remote node.
func (d *Downloader) DeliverReceipts(id string, receipts [][]*types.Receipt) (err error) {
return d.deliver(id, d.receiptCh, &receiptPack{id, receipts}, receiptInMeter, receiptDropMeter)
}
// DeliverNodeData injects a new batch of node state data received from a remote node.
func (d *Downloader) DeliverNodeData(id string, data [][]byte) (err error) {
return d.deliver(id, d.stateCh, &statePack{id, data}, stateInMeter, stateDropMeter)
}
// deliver injects a new batch of data received from a remote node.
func (d *Downloader) deliver(id string, destCh chan dataPack, packet dataPack, inMeter, dropMeter metrics.Meter) (err error) {
// Update the delivery metrics for both good and failed deliveries
inMeter.Mark(int64(packet.Items()))
defer func() {
if err != nil {
dropMeter.Mark(int64(packet.Items()))
}
}()
// Deliver or abort if the sync is canceled while queuing
d.cancelLock.RLock()
cancel := d.cancelCh
d.cancelLock.RUnlock()
if cancel == nil {
return errNoSyncActive
}
select {
case destCh <- packet:
return nil
case <-cancel:
return errNoSyncActive
}
2015-04-12 13:38:25 +03:00
}
// qosTuner is the quality of service tuning loop that occasionally gathers the
// peer latency statistics and updates the estimated request round trip time.
func (d *Downloader) qosTuner() {
for {
// Retrieve the current median RTT and integrate into the previoust target RTT
rtt := time.Duration((1-qosTuningImpact)*float64(atomic.LoadUint64(&d.rttEstimate)) + qosTuningImpact*float64(d.peers.medianRTT()))
atomic.StoreUint64(&d.rttEstimate, uint64(rtt))
// A new RTT cycle passed, increase our confidence in the estimated RTT
conf := atomic.LoadUint64(&d.rttConfidence)
conf = conf + (1000000-conf)/2
atomic.StoreUint64(&d.rttConfidence, conf)
// Log the new QoS values and sleep until the next RTT
log.Debug("Recalculated downloader QoS values", "rtt", rtt, "confidence", float64(conf)/1000000.0, "ttl", d.requestTTL())
select {
case <-d.quitCh:
return
case <-time.After(rtt):
}
}
}
// qosReduceConfidence is meant to be called when a new peer joins the downloader's
// peer set, needing to reduce the confidence we have in out QoS estimates.
func (d *Downloader) qosReduceConfidence() {
// If we have a single peer, confidence is always 1
peers := uint64(d.peers.Len())
if peers == 0 {
// Ensure peer connectivity races don't catch us off guard
return
}
if peers == 1 {
atomic.StoreUint64(&d.rttConfidence, 1000000)
return
}
// If we have a ton of peers, don't drop confidence)
if peers >= uint64(qosConfidenceCap) {
return
}
// Otherwise drop the confidence factor
conf := atomic.LoadUint64(&d.rttConfidence) * (peers - 1) / peers
if float64(conf)/1000000 < rttMinConfidence {
conf = uint64(rttMinConfidence * 1000000)
}
atomic.StoreUint64(&d.rttConfidence, conf)
rtt := time.Duration(atomic.LoadUint64(&d.rttEstimate))
log.Debug("Relaxed downloader QoS values", "rtt", rtt, "confidence", float64(conf)/1000000.0, "ttl", d.requestTTL())
}
// requestRTT returns the current target round trip time for a download request
// to complete in.
//
// Note, the returned RTT is .9 of the actually estimated RTT. The reason is that
// the downloader tries to adapt queries to the RTT, so multiple RTT values can
2018-04-04 13:25:02 +03:00
// be adapted to, but smaller ones are preferred (stabler download stream).
func (d *Downloader) requestRTT() time.Duration {
return time.Duration(atomic.LoadUint64(&d.rttEstimate)) * 9 / 10
}
// requestTTL returns the current timeout allowance for a single download request
// to finish under.
func (d *Downloader) requestTTL() time.Duration {
var (
rtt = time.Duration(atomic.LoadUint64(&d.rttEstimate))
conf = float64(atomic.LoadUint64(&d.rttConfidence)) / 1000000.0
)
ttl := time.Duration(ttlScaling) * time.Duration(float64(rtt)/conf)
if ttl > ttlLimit {
ttl = ttlLimit
}
return ttl
}