2015-07-07 03:54:22 +03:00
// Copyright 2015 The go-ethereum Authors
2015-07-22 19:48:40 +03:00
// This file is part of the go-ethereum library.
2015-07-07 03:54:22 +03:00
//
2015-07-23 19:35:11 +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.
//
2015-07-22 19:48:40 +03:00
// 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
2015-07-22 19:48:40 +03:00
// 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
2015-07-22 19:48:40 +03:00
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
2015-07-07 03:54:22 +03:00
2015-06-16 11:58:32 +03:00
// Package downloader contains the manual full chain synchronisation.
2015-04-12 13:38:25 +03:00
package downloader
import (
2015-04-18 02:10:32 +03:00
"errors"
2015-10-13 12:04:25 +03:00
"fmt"
2015-07-29 13:20:54 +03:00
"math/big"
2015-04-12 13:38:25 +03:00
"sync"
"sync/atomic"
"time"
2019-05-15 14:33:33 +03:00
"github.com/ethereum/go-ethereum"
2015-04-12 13:38:25 +03:00
"github.com/ethereum/go-ethereum/common"
2018-05-07 14:35:06 +03:00
"github.com/ethereum/go-ethereum/core/rawdb"
2021-04-29 17:33:45 +03:00
"github.com/ethereum/go-ethereum/core/state/snapshot"
2015-04-12 13:38:25 +03:00
"github.com/ethereum/go-ethereum/core/types"
2020-12-14 12:27:15 +03:00
"github.com/ethereum/go-ethereum/eth/protocols/snap"
2015-10-05 19:37:56 +03:00
"github.com/ethereum/go-ethereum/ethdb"
2015-05-15 01:43:00 +03:00
"github.com/ethereum/go-ethereum/event"
2017-02-22 15:10:07 +03:00
"github.com/ethereum/go-ethereum/log"
2016-05-13 13:12:13 +03:00
"github.com/ethereum/go-ethereum/params"
2024-02-13 16:49:53 +03:00
"github.com/ethereum/go-ethereum/triedb"
2015-04-12 13:38:25 +03:00
)
2015-06-08 14:06:36 +03:00
var (
2015-09-28 19:27:31 +03:00
MaxBlockFetch = 128 // Amount of blocks to be fetched per retrieval request
MaxHeaderFetch = 192 // Amount of block headers to be fetched per retrieval request
2016-02-25 18:36:42 +02:00
MaxSkeletonSize = 128 // Number of header fetches to need for a skeleton assembly
2015-09-28 19:27:31 +03:00
MaxReceiptFetch = 256 // Amount of transaction receipts to allow fetching per request
2020-07-13 12:02:54 +03:00
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
fullMaxForkAncestry uint64 = params . FullImmutabilityThreshold // Maximum chain reorganisation (locally redeclared so tests can reduce it)
lightMaxForkAncestry uint64 = params . LightImmutabilityThreshold // Maximum chain reorganisation (locally redeclared so tests can reduce it)
2015-09-28 19:27:31 +03:00
2018-10-04 16:36:59 +03:00
reorgProtThreshold = 48 // Threshold number of recent blocks to disable mini reorg protection
reorgProtHeaderDelay = 2 // Number of headers to delay delivering to cover mini reorgs
2023-05-03 12:58:39 +03:00
fsHeaderSafetyNet = 2048 // Number of headers to discard in case a chain violation is detected
fsHeaderContCheck = 3 * time . Second // Time interval to check for header continuations during state download
fsMinFullBlocks = 64 // Number of blocks to retrieve fully even in snap sync
2015-05-15 13:14:46 +03:00
)
2015-04-19 14:30:34 +03:00
2015-05-15 13:14:46 +03:00
var (
2016-02-25 18:36:42 +02:00
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" )
2019-04-16 13:20:38 +03:00
errUnsyncedPeer = errors . New ( "unsynced peer" )
2016-02-25 18:36:42 +02:00
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" )
errInvalidBody = errors . New ( "retrieved block body is invalid" )
errInvalidReceipt = errors . New ( "retrieved receipt is invalid" )
errCancelStateFetch = errors . New ( "state data download canceled (requested)" )
errCancelContentProcessing = errors . New ( "content processing canceled (requested)" )
2019-06-05 15:00:46 +03:00
errCanceled = errors . New ( "syncing canceled (requested)" )
2021-01-06 09:37:45 +03:00
errTooOld = errors . New ( "peer's protocol version too old" )
2021-01-21 00:45:01 +03:00
errNoAncestorFound = errors . New ( "no common ancestor found" )
2022-04-04 10:10:16 +03:00
errNoPivotHeader = errors . New ( "pivot header is not found" )
2022-03-11 15:14:45 +03:00
ErrMergeTransition = errors . New ( "legacy sync reached the merge" )
2015-04-18 02:10:32 +03:00
)
2021-11-26 14:26:03 +03:00
// peerDropFn is a callback type for dropping a peer detected as malicious.
type peerDropFn func ( id string )
2022-07-25 16:51:04 +03:00
// badBlockFn is a callback for the async beacon sync to notify the caller that
// the origin header requested to sync to, produced a chain with a bad block.
type badBlockFn func ( invalid * types . Header , origin * types . Header )
2021-12-01 21:18:12 +03:00
// headerTask is a set of downloaded headers to queue along with their precomputed
// hashes to avoid constant rehashing.
type headerTask struct {
headers [ ] * types . Header
hashes [ ] common . Hash
}
2015-04-12 13:38:25 +03:00
type Downloader struct {
2023-04-03 22:48:10 +03:00
mode atomic . Uint32 // Synchronisation mode defining the strategy used (per sync cycle), use d.getMode() to get the SyncMode
2016-06-02 12:37:14 +03:00
mux * event . TypeMux // Event multiplexer to announce sync operation events
2015-05-15 01:43:00 +03:00
2023-04-24 09:37:10 +03:00
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
2019-05-13 15:28:01 +03:00
2021-12-03 13:32:41 +03:00
stateDB ethdb . Database // Database to state sync into (and deduplicate via)
2015-04-12 13:38:25 +03:00
2015-06-10 01:20:35 +03:00
// Statistics
2021-11-26 14:26:03 +03:00
syncStatsChainOrigin uint64 // Origin block number where syncing started at
syncStatsChainHeight uint64 // Highest block number known when syncing started
2015-10-05 19:37:56 +03:00
syncStatsLock sync . RWMutex // Lock protecting the sync stats fields
2015-06-10 01:20:35 +03:00
2017-06-27 18:15:29 +03:00
lightchain LightChain
2017-07-03 17:17:12 +03:00
blockchain BlockChain
2017-06-27 18:15:29 +03:00
2015-04-13 17:38:32 +03:00
// Callbacks
2017-06-27 18:15:29 +03:00
dropPeer peerDropFn // Drops a peer for misbehaving
2022-07-25 16:51:04 +03:00
badBlock badBlockFn // Reports a block as rejected by the chain
2015-04-12 13:38:25 +03:00
2015-04-13 17:38:32 +03:00
// Status
2015-06-11 18:13:13 +03:00
synchroniseMock func ( id string , hash common . Hash ) error // Replacement for synchronise during testing
2023-04-03 22:48:10 +03:00
synchronising atomic . Bool
notified atomic . Bool
committed atomic . Bool
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
ancientLimit uint64 // The maximum block number which can be regarded as ancient data.
2015-04-13 17:38:32 +03:00
// Channels
2021-12-01 21:18:12 +03:00
headerProcCh chan * headerTask // Channel to feed the header processor new tasks
2015-05-13 13:47:21 +03:00
2022-03-11 15:14:45 +03:00
// Skeleton sync
skeleton * skeleton // Header skeleton to backfill the chain with (eth2 mode)
2020-09-08 11:13:16 +03:00
// State sync
pivotHeader * types . Header // Pivot block header to dynamically push the syncing state root
pivotLock sync . RWMutex // Lock protecting pivot header reads from updates
2020-12-14 12:27:15 +03:00
SnapSyncer * snap . Syncer // TODO(karalabe): make private! hack for now
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 chan * stateSync
2016-07-26 13:07:12 +03:00
// Cancellation and termination
2018-04-16 11:37:48 +03:00
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.
2015-08-14 21:25:41 +03:00
2016-06-01 18:07:25 +03:00
quitCh chan struct { } // Quit channel to signal termination
2020-10-13 11:58:41 +03:00
quitLock sync . Mutex // Lock to prevent double closes
2016-06-01 18:07:25 +03:00
2015-08-14 21:25:41 +03:00
// Testing hooks
2015-09-28 19:27:31 +03:00
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)
2023-02-21 13:17:34 +03:00
// Progress reporting metrics
syncStartBlock uint64 // Head snap block when Geth was started
syncStartTime time . Time // Time instance when chain sync started
syncLogTime time . Time // Time instance when status was last reported
2015-05-26 14:00:21 +03:00
}
2017-07-05 12:42:37 +03:00
// LightChain encapsulates functions required to synchronise a light chain.
2017-06-27 18:15:29 +03:00
type LightChain interface {
// HasHeader verifies a header's presence in the local chain.
2018-02-05 19:40:32 +03:00
HasHeader ( common . Hash , uint64 ) bool
2017-06-27 18:15:29 +03:00
// 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
2018-01-30 19:39:32 +03:00
// GetTd returns the total difficulty of a local block.
GetTd ( common . Hash , uint64 ) * big . Int
2017-06-27 18:15:29 +03:00
// InsertHeaderChain inserts a batch of headers into the local chain.
2023-05-03 12:58:39 +03:00
InsertHeaderChain ( [ ] * types . Header ) ( int , error )
2017-06-27 18:15:29 +03:00
2020-08-20 13:01:24 +03:00
// SetHead rewinds the local chain to a new head.
SetHead ( uint64 ) error
2017-06-27 18:15:29 +03:00
}
2021-11-26 14:26:03 +03:00
// BlockChain encapsulates functions required to sync a (full or snap) blockchain.
2017-06-27 18:15:29 +03:00
type BlockChain interface {
LightChain
2018-02-11 15:43:56 +03:00
// HasBlock verifies a block's presence in the local chain.
HasBlock ( common . Hash , uint64 ) bool
2017-06-27 18:15:29 +03:00
2021-11-26 14:26:03 +03:00
// HasFastBlock verifies a snap block's presence in the local chain.
2018-11-16 14:15:05 +03:00
HasFastBlock ( common . Hash , uint64 ) bool
2017-06-27 18:15:29 +03:00
// GetBlockByHash retrieves a block from the local chain.
GetBlockByHash ( common . Hash ) * types . Block
// CurrentBlock retrieves the head block from the local chain.
2023-03-02 09:29:15 +03:00
CurrentBlock ( ) * types . Header
2017-06-27 18:15:29 +03:00
2023-03-02 09:29:15 +03:00
// CurrentSnapBlock retrieves the head snap block from the local chain.
CurrentSnapBlock ( ) * types . Header
2017-06-27 18:15:29 +03:00
2021-11-26 14:26:03 +03:00
// SnapSyncCommitHead directly commits the head block to a certain entity.
SnapSyncCommitHead ( common . Hash ) error
2017-06-27 18:15:29 +03:00
// 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.
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
InsertReceiptChain ( types . Blocks , [ ] types . Receipts , uint64 ) ( int , error )
2021-04-29 17:33:45 +03:00
// Snapshots returns the blockchain snapshot tree to paused it during sync.
Snapshots ( ) * snapshot . Tree
2022-11-28 16:31:28 +03:00
// TrieDB retrieves the low level trie database used for interacting
// with trie nodes.
2024-02-13 16:49:53 +03:00
TrieDB ( ) * triedb . Database
2017-06-27 18:15:29 +03:00
}
2015-06-11 15:56:08 +03:00
// New creates a new downloader to fetch hashes and blocks from remote peers.
2023-04-24 09:37:10 +03:00
func New ( stateDb ethdb . Database , mux * event . TypeMux , chain BlockChain , lightchain LightChain , dropPeer peerDropFn , success func ( ) ) * Downloader {
2017-06-27 18:15:29 +03:00
if lightchain == nil {
lightchain = chain
}
2016-06-01 18:07:25 +03:00
dl := & Downloader {
2017-06-28 15:25:08 +03:00
stateDB : stateDb ,
mux : mux ,
2020-09-02 12:01:46 +03:00
queue : newQueue ( blockCacheMaxItems , blockCacheInitialItems ) ,
2017-06-28 15:25:08 +03:00
peers : newPeerSet ( ) ,
2017-07-03 17:17:12 +03:00
blockchain : chain ,
2017-06-28 15:25:08 +03:00
lightchain : lightchain ,
dropPeer : dropPeer ,
2021-12-01 21:18:12 +03:00
headerProcCh : make ( chan * headerTask , 1 ) ,
2017-06-28 15:25:08 +03:00
quitCh : make ( chan struct { } ) ,
2022-11-28 16:31:28 +03:00
SnapSyncer : snap . NewSyncer ( stateDb , chain . TrieDB ( ) . Scheme ( ) ) ,
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 ) ,
2023-03-02 09:29:15 +03:00
syncStartBlock : chain . CurrentSnapBlock ( ) . Number . Uint64 ( ) ,
2015-04-12 13:38:25 +03:00
}
2023-02-21 13:17:34 +03:00
// Create the post-merge skeleton syncer and start the process
2022-03-11 15:14:45 +03:00
dl . skeleton = newSkeleton ( stateDb , dl . peers , dropPeer , newBeaconBackfiller ( dl , success ) )
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 ( )
2016-06-01 18:07:25 +03:00
return dl
2015-04-12 13:38:25 +03:00
}
2015-10-13 12:04:25 +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.
2016-02-10 11:56:15 +02:00
//
2021-11-26 14:26:03 +03:00
// In addition, during the state download phase of snap synchronisation the number
2016-02-10 11:56:15 +02:00
// of processed and the total number of known states are also returned. Otherwise
// these are zero.
2016-09-06 12:39:14 +03:00
func ( d * Downloader ) Progress ( ) ethereum . SyncProgress {
2016-02-10 11:56:15 +02:00
// Lock the current stats and return the progress
2015-09-09 19:02:54 +03:00
d . syncStatsLock . RLock ( )
defer d . syncStatsLock . RUnlock ( )
2015-06-10 01:20:35 +03:00
2015-10-13 12:04:25 +03:00
current := uint64 ( 0 )
2020-06-30 20:43:29 +03:00
mode := d . getMode ( )
2019-05-13 15:28:01 +03:00
switch {
2020-06-30 20:43:29 +03:00
case d . blockchain != nil && mode == FullSync :
2023-03-02 09:29:15 +03:00
current = d . blockchain . CurrentBlock ( ) . Number . Uint64 ( )
2021-11-26 14:26:03 +03:00
case d . blockchain != nil && mode == SnapSync :
2023-03-02 09:29:15 +03:00
current = d . blockchain . CurrentSnapBlock ( ) . Number . Uint64 ( )
2019-05-13 15:28:01 +03:00
case d . lightchain != nil :
2017-06-27 18:15:29 +03:00
current = d . lightchain . CurrentHeader ( ) . Number . Uint64 ( )
2019-05-13 15:28:01 +03:00
default :
2020-06-30 20:43:29 +03:00
log . Error ( "Unknown downloader chain/mode combo" , "light" , d . lightchain != nil , "full" , d . blockchain != nil , "mode" , mode )
2015-10-13 12:04:25 +03:00
}
2021-11-26 14:26:03 +03:00
progress , pending := d . SnapSyncer . Progress ( )
2016-09-06 12:39:14 +03:00
return ethereum . SyncProgress {
2021-11-26 14:26:03 +03:00
StartingBlock : d . syncStatsChainOrigin ,
CurrentBlock : current ,
HighestBlock : d . syncStatsChainHeight ,
SyncedAccounts : progress . AccountSynced ,
SyncedAccountBytes : uint64 ( progress . AccountBytes ) ,
SyncedBytecodes : progress . BytecodeSynced ,
SyncedBytecodeBytes : uint64 ( progress . BytecodeBytes ) ,
SyncedStorage : progress . StorageSynced ,
SyncedStorageBytes : uint64 ( progress . StorageBytes ) ,
HealedTrienodes : progress . TrienodeHealSynced ,
HealedTrienodeBytes : uint64 ( progress . TrienodeHealBytes ) ,
HealedBytecodes : progress . BytecodeHealSynced ,
HealedBytecodeBytes : uint64 ( progress . BytecodeHealBytes ) ,
HealingTrienodes : pending . TrienodeHeal ,
HealingBytecode : pending . BytecodeHeal ,
2016-09-06 12:39:14 +03:00
}
2015-04-19 22:45:58 +03:00
}
2015-05-11 14:26:20 +03:00
// RegisterPeer injects a new download peer into the set of block source to be
// used for fetching hashes and blocks from.
2020-12-14 12:27:15 +03:00
func ( d * Downloader ) RegisterPeer ( id string , version uint , peer Peer ) error {
var logger log . Logger
if len ( id ) < 16 {
// Tests use short IDs, don't choke on them
logger = log . New ( "peer" , id )
} else {
2021-01-25 09:17:05 +03:00
logger = log . New ( "peer" , id [ : 8 ] )
2020-12-14 12:27:15 +03:00
}
2017-02-24 19:23:03 +03:00
logger . Trace ( "Registering sync peer" )
2017-06-28 15:25:08 +03:00
if err := d . peers . Register ( newPeerConnection ( id , version , peer , logger ) ) ; err != nil {
2017-02-27 14:17:58 +03:00
logger . Error ( "Failed to register sync peer" , "err" , err )
2015-05-11 14:26:20 +03:00
return err
}
2015-04-12 13:38:25 +03:00
return nil
}
2015-05-11 14:26:20 +03:00
// UnregisterPeer remove a peer from the known list, preventing any action from
2015-09-28 19:27:31 +03:00
// the specified peer. An effort is also made to return any pending fetches into
// the queue.
2015-05-11 14:26:20 +03:00
func ( d * Downloader ) UnregisterPeer ( id string ) error {
2016-07-26 13:07:12 +03:00
// Unregister the peer from the active peer set and revoke any fetch tasks
2020-12-14 12:27:15 +03:00
var logger log . Logger
if len ( id ) < 16 {
// Tests use short IDs, don't choke on them
logger = log . New ( "peer" , id )
} else {
2021-01-25 09:17:05 +03:00
logger = log . New ( "peer" , id [ : 8 ] )
2020-12-14 12:27:15 +03:00
}
2017-02-24 19:23:03 +03:00
logger . Trace ( "Unregistering sync peer" )
2015-05-11 14:26:20 +03:00
if err := d . peers . Unregister ( id ) ; err != nil {
2017-02-27 14:17:58 +03:00
logger . Error ( "Failed to unregister sync peer" , "err" , err )
2015-05-11 14:26:20 +03:00
return err
}
2015-09-28 19:27:31 +03:00
d . queue . Revoke ( id )
2016-07-26 13:07:12 +03:00
2015-05-11 14:26:20 +03:00
return nil
2015-04-12 13:38:25 +03:00
}
2022-03-11 15:14:45 +03:00
// LegacySync tries to sync up our local block chain with a remote peer, both
2015-06-11 15:56:08 +03:00
// adding various sanity checks as well as wrapping it with various log entries.
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) LegacySync ( id string , head common . Hash , td , ttd * big . Int , mode SyncMode ) error {
err := d . synchronise ( id , head , td , ttd , mode , false , nil )
2020-05-29 12:12:43 +03:00
2015-10-28 17:41:01 +03:00
switch err {
2020-05-29 12:12:43 +03:00
case nil , errBusy , errCanceled :
return err
}
2020-07-09 00:08:08 +03:00
if errors . Is ( err , errInvalidChain ) || errors . Is ( err , errBadPeer ) || errors . Is ( err , errTimeout ) ||
errors . Is ( err , errStallingPeer ) || errors . Is ( err , errUnsyncedPeer ) || errors . Is ( err , errEmptyHeaderSet ) ||
errors . Is ( err , errPeersUnavailable ) || errors . Is ( err , errTooOld ) || errors . Is ( err , errInvalidAncestor ) {
2020-05-29 12:12:43 +03:00
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 )
}
return err
}
2022-03-11 15:14:45 +03:00
if errors . Is ( err , ErrMergeTransition ) {
return err // This is an expected fault, don't keep printing it in a spin-loop
}
2020-07-09 00:08:08 +03:00
log . Warn ( "Synchronisation failed, retrying" , "err" , err )
2015-10-28 17:41:01 +03:00
return err
2015-06-11 15:56:08 +03:00
}
// synchronise will select the peer and use it for synchronising. If an empty string is given
2017-11-16 14:14:51 +03:00
// it will use the best peer possible and synchronize if its TD is higher than our own. If any of the
2015-04-24 15:40:32 +03:00
// checks fail an error will be returned. This method is synchronous
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) synchronise ( id string , hash common . Hash , td , ttd * big . Int , mode SyncMode , beaconMode bool , beaconPing chan struct { } ) error {
// The beacon header syncer is async. It will start this synchronization and
2022-04-20 17:17:29 +03:00
// will continue doing other tasks. However, if synchronization needs to be
2022-03-11 15:14:45 +03:00
// cancelled, the syncer needs to know if we reached the startup point (and
2022-08-19 09:00:21 +03:00
// inited the cancel channel) or not yet. Make sure that we'll signal even in
2022-03-11 15:14:45 +03:00
// case of a failure.
if beaconPing != nil {
defer func ( ) {
select {
case <- beaconPing : // already notified
default :
close ( beaconPing ) // weird exit condition, notify that it's safe to cancel (the nothing)
}
} ( )
}
2016-03-15 20:27:49 +02:00
// Mock out the synchronisation if testing
2015-06-11 18:13:13 +03:00
if d . synchroniseMock != nil {
return d . synchroniseMock ( id , hash )
}
2015-05-07 21:07:20 +03:00
// Make sure only one goroutine is ever allowed past this point at once
2023-04-03 22:48:10 +03:00
if ! d . synchronising . CompareAndSwap ( false , true ) {
2015-06-11 15:56:08 +03:00
return errBusy
2015-04-19 14:30:34 +03:00
}
2023-04-03 22:48:10 +03:00
defer d . synchronising . Store ( false )
2015-04-24 15:40:32 +03:00
2015-05-13 16:03:05 +03:00
// Post a user notification of the sync (only once per session)
2023-04-03 22:48:10 +03:00
if d . notified . CompareAndSwap ( false , true ) {
2017-02-24 19:23:03 +03:00
log . Info ( "Block synchronisation started" )
2015-05-13 16:03:05 +03:00
}
2020-12-14 12:27:15 +03:00
if mode == SnapSync {
2023-09-17 17:35:09 +03:00
// Snap sync will directly modify the persistent state, making the entire
// trie database unusable until the state is fully synced. To prevent any
// subsequent state reads, explicitly disable the trie database and state
// syncer is responsible to address and correct any state missing.
if d . blockchain . TrieDB ( ) . Scheme ( ) == rawdb . PathScheme {
core, accounts, eth, trie: handle genesis state missing (#28171)
* core, accounts, eth, trie: handle genesis state missing
* core, eth, trie: polish
* core: manage txpool subscription in mainpool
* eth/backend: fix test
* cmd, eth: fix test
* core/rawdb, trie/triedb/pathdb: address comments
* eth, trie: address comments
* eth: inline the function
* eth: use synced flag
* core/txpool: revert changes in txpool
* core, eth, trie: rename functions
2023-09-28 10:00:53 +03:00
if err := d . blockchain . TrieDB ( ) . Disable ( ) ; err != nil {
return err
}
2023-09-17 17:35:09 +03:00
}
// Snap sync uses the snapshot namespace to store potentially flaky data until
2021-11-26 14:26:03 +03:00
// sync completely heals and finishes. Pause snapshot maintenance in the mean-
// time to prevent access.
if snapshots := d . blockchain . Snapshots ( ) ; snapshots != nil { // Only nil in tests
snapshots . Disable ( )
2020-12-14 12:27:15 +03:00
}
}
2015-09-28 19:27:31 +03:00
// Reset the queue, peer set and wake channels to clean any internal leftover state
2020-09-02 12:01:46 +03:00
d . queue . Reset ( blockCacheMaxItems , blockCacheInitialItems )
2015-05-11 14:26:20 +03:00
d . peers . Reset ( )
2015-05-08 17:21:11 +03:00
2021-11-26 14:26:03 +03:00
for _ , ch := range [ ] chan bool { d . queue . blockWakeCh , d . queue . receiptWakeCh } {
2015-09-28 19:27:31 +03:00
select {
case <- ch :
default :
}
2015-09-23 12:39:17 +03:00
}
2016-02-25 18:36:42 +02:00
for empty := false ; ! empty ; {
select {
case <- d . headerProcCh :
default :
empty = true
}
}
2016-07-26 13:07:12 +03:00
// Create cancel channel for aborting mid-flight and mark the master peer
2015-06-18 00:04:57 +03:00
d . cancelLock . Lock ( )
d . cancelCh = make ( chan struct { } )
2016-07-26 13:07:12 +03:00
d . cancelPeer = id
2015-06-18 00:04:57 +03:00
d . cancelLock . Unlock ( )
2017-03-22 03:37:24 +03:00
defer d . Cancel ( ) // No matter what, we can't leave the cancel channel open
2016-05-30 12:01:50 +03:00
2020-06-30 20:43:29 +03:00
// Atomically set the requested sync mode
2023-04-03 22:48:10 +03:00
d . mode . Store ( uint32 ( mode ) )
2018-02-05 19:40:32 +03:00
2015-05-07 21:07:20 +03:00
// Retrieve the origin peer and initiate the downloading process
2022-03-11 15:14:45 +03:00
var p * peerConnection
if ! beaconMode { // Beacon mode doesn't need a peer to sync from
p = d . peers . Peer ( id )
if p == nil {
return errUnknownPeer
}
2015-04-13 17:38:32 +03:00
}
2022-03-11 15:14:45 +03:00
if beaconPing != nil {
close ( beaconPing )
}
return d . syncWithPeer ( p , hash , td , ttd , beaconMode )
2015-05-01 01:23:51 +03:00
}
2020-06-30 20:43:29 +03:00
func ( d * Downloader ) getMode ( ) SyncMode {
2023-04-03 22:48:10 +03:00
return SyncMode ( d . mode . Load ( ) )
2020-06-30 20:43:29 +03:00
}
2015-05-11 14:26:20 +03:00
// syncWithPeer starts a block synchronization based on the hash chain from the
// specified peer and head hash.
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) syncWithPeer ( p * peerConnection , hash common . Hash , td , ttd * big . Int , beaconMode bool ) ( err error ) {
2015-05-16 13:29:19 +03:00
d . mux . Post ( StartEvent { } )
2015-05-01 01:23:51 +03:00
defer func ( ) {
// reset on error
if err != nil {
2015-05-15 01:43:00 +03:00
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 } )
2015-05-01 01:23:51 +03:00
}
} ( )
2020-06-30 20:43:29 +03:00
mode := d . getMode ( )
2015-04-24 15:40:32 +03:00
2022-03-11 15:14:45 +03:00
if ! beaconMode {
log . Debug ( "Synchronising with the network" , "peer" , p . id , "eth" , p . version , "head" , hash , "td" , td , "mode" , mode )
} else {
log . Debug ( "Backfilling with the network" , "mode" , mode )
}
2015-09-30 19:23:31 +03:00
defer func ( start time . Time ) {
2019-05-15 14:33:33 +03:00
log . Debug ( "Synchronisation terminated" , "elapsed" , common . PrettyDuration ( time . Since ( start ) ) )
2015-09-30 19:23:31 +03:00
} ( time . Now ( ) )
2015-08-14 21:25:41 +03:00
2016-07-21 12:36:38 +03:00
// Look up the sync boundaries: the common ancestor and the target block
2023-02-23 14:22:41 +03:00
var latest , pivot , final * types . Header
2022-03-11 15:14:45 +03:00
if ! beaconMode {
// In legacy mode, use the master peer to retrieve the headers from
latest , pivot , err = d . fetchHead ( p )
if err != nil {
return err
}
} else {
2023-01-03 12:03:44 +03:00
// In beacon mode, use the skeleton chain to retrieve the headers from
2023-02-23 14:22:41 +03:00
latest , _ , final , err = d . skeleton . Bounds ( )
2022-03-11 15:14:45 +03:00
if err != nil {
return err
}
if latest . Number . Uint64 ( ) > uint64 ( fsMinFullBlocks ) {
2022-04-04 10:10:16 +03:00
number := latest . Number . Uint64 ( ) - uint64 ( fsMinFullBlocks )
// Retrieve the pivot header from the skeleton chain segment but
// fallback to local chain if it's not found in skeleton space.
if pivot = d . skeleton . Header ( number ) ; pivot == nil {
2023-02-23 14:22:41 +03:00
_ , oldest , _ , _ := d . skeleton . Bounds ( ) // error is already checked
2022-04-14 09:49:23 +03:00
if number < oldest . Number . Uint64 ( ) {
count := int ( oldest . Number . Uint64 ( ) - number ) // it's capped by fsMinFullBlocks
headers := d . readHeaderRange ( oldest , count )
if len ( headers ) == count {
pivot = headers [ len ( headers ) - 1 ]
log . Warn ( "Retrieved pivot header from local" , "number" , pivot . Number , "hash" , pivot . Hash ( ) , "latest" , latest . Number , "oldest" , oldest . Number )
}
}
2022-04-04 10:10:16 +03:00
}
// Print an error log and return directly in case the pivot header
// is still not found. It means the skeleton chain is not linked
// correctly with local chain.
if pivot == nil {
log . Error ( "Pivot header is not found" , "number" , number )
return errNoPivotHeader
}
2022-03-11 15:14:45 +03:00
}
2016-07-21 12:36:38 +03:00
}
2022-03-11 15:14:45 +03:00
// If no pivot block was returned, the head is below the min full block
// threshold (i.e. new chain). In that case we won't really snap sync
// anyway, but still need a valid pivot block to avoid some code hitting
// nil panics on access.
2021-11-26 14:26:03 +03:00
if mode == SnapSync && pivot == nil {
2023-03-02 09:29:15 +03:00
pivot = d . blockchain . CurrentBlock ( )
2020-09-08 11:13:16 +03:00
}
2016-07-21 12:36:38 +03:00
height := latest . Number . Uint64 ( )
2015-09-09 19:02:54 +03:00
2022-03-11 15:14:45 +03:00
var origin uint64
if ! beaconMode {
// In legacy mode, reach out to the network and find the ancestor
origin , err = d . findAncestor ( p , latest )
if err != nil {
return err
}
} else {
// In beacon mode, use the skeleton chain for the ancestor lookup
2022-03-21 11:43:54 +03:00
origin , err = d . findBeaconAncestor ( )
if err != nil {
return err
}
2016-07-21 12:36:38 +03:00
}
d . syncStatsLock . Lock ( )
if d . syncStatsChainHeight <= origin || d . syncStatsChainOrigin > origin {
d . syncStatsChainOrigin = origin
}
d . syncStatsChainHeight = height
d . syncStatsLock . Unlock ( )
2016-05-27 14:26:00 +03:00
2021-11-26 14:26:03 +03:00
// Ensure our origin point is below any snap sync pivot point
if mode == SnapSync {
2018-02-05 19:40:32 +03:00
if height <= uint64 ( fsMinFullBlocks ) {
origin = 0
2016-07-21 12:36:38 +03:00
} else {
2020-09-08 11:13:16 +03:00
pivotNumber := pivot . Number . Uint64 ( )
if pivotNumber <= origin {
origin = pivotNumber - 1
2016-07-21 12:36:38 +03:00
}
2020-08-20 13:01:24 +03:00
// Write out the pivot into the database so a rollback beyond it will
2021-11-26 14:26:03 +03:00
// reenable snap sync
2020-09-08 11:13:16 +03:00
rawdb . WriteLastPivotNumber ( d . stateDB , pivotNumber )
2015-09-09 19:02:54 +03:00
}
2016-07-21 12:36:38 +03:00
}
2023-04-03 22:48:10 +03:00
d . committed . Store ( true )
2021-11-26 14:26:03 +03:00
if mode == SnapSync && pivot . Number . Uint64 ( ) != 0 {
2023-04-03 22:48:10 +03:00
d . committed . Store ( false )
2018-02-05 19:40:32 +03:00
}
2021-11-26 14:26:03 +03:00
if mode == SnapSync {
2023-02-23 14:22:41 +03:00
// Set the ancient data limitation. If we are running snap 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.
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
//
2023-02-23 14:22:41 +03:00
// If the network is post-merge, use either the last announced finalized
// block as the ancient limit, or if we haven't yet received one, the head-
// a max fork ancestry limit. One quirky case if we've already passed the
// finalized block, in which case the skeleton.Bounds will return nil and
// we'll revert to head - 90K. That's fine, we're finishing sync anyway.
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
//
2023-02-23 14:22:41 +03:00
// For non-merged networks, 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. This most is mostly a fallback for
2023-12-18 11:35:12 +03:00
// legacy networks, but should eventually be dropped. TODO(karalabe).
2023-02-23 14:22:41 +03:00
if beaconMode {
// Beacon sync, use the latest finalized block as the ancient limit
// or a reasonable height if no finalized block is yet announced.
if final != nil {
d . ancientLimit = final . Number . Uint64 ( )
} else if height > fullMaxForkAncestry + 1 {
d . ancientLimit = height - fullMaxForkAncestry - 1
} else {
d . ancientLimit = 0
}
2020-10-09 09:58:30 +03:00
} else {
2023-04-24 09:37:10 +03:00
// Legacy sync, use the best announcement we have from the remote peer.
// TODO(karalabe): Drop this pathway.
if height > fullMaxForkAncestry + 1 {
2023-02-23 14:22:41 +03:00
d . ancientLimit = height - fullMaxForkAncestry - 1
} else {
d . ancientLimit = 0
}
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
}
frozen , _ := d . stateDB . Ancients ( ) // Ignore the error here since light client can also hit here.
2020-08-20 13:01:24 +03:00
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
// 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 {
2021-11-22 12:11:59 +03:00
if err := d . lightchain . SetHead ( origin ) ; err != nil {
2020-08-20 13:01:24 +03:00
return err
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
}
2024-01-31 11:57:33 +03:00
log . Info ( "Truncated excess ancient chain segment" , "oldhead" , frozen - 1 , "newhead" , origin )
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
}
}
2018-02-05 19:40:32 +03:00
// Initiate the sync using a concurrent header and content retrieval algorithm
2020-06-30 20:43:29 +03:00
d . queue . Prepare ( origin + 1 , mode )
2016-07-21 12:36:38 +03:00
if d . syncInitHook != nil {
d . syncInitHook ( origin , height )
2015-04-12 13:38:25 +03:00
}
2022-03-11 15:14:45 +03:00
var headerFetcher func ( ) error
if ! beaconMode {
// In legacy mode, headers are retrieved from the network
headerFetcher = func ( ) error { return d . fetchHeaders ( p , origin + 1 , latest . Number . Uint64 ( ) ) }
} else {
// In beacon mode, headers are served by the skeleton syncer
headerFetcher = func ( ) error { return d . fetchBeaconHeaders ( origin + 1 ) }
}
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 {
2022-03-11 15:14:45 +03:00
headerFetcher , // Headers are always retrieved
func ( ) error { return d . fetchBodies ( origin + 1 , beaconMode ) } , // Bodies are retrieved during normal and snap sync
func ( ) error { return d . fetchReceipts ( origin + 1 , beaconMode ) } , // Receipts are retrieved during snap sync
func ( ) error { return d . processHeaders ( origin + 1 , td , ttd , beaconMode ) } ,
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
}
2021-11-26 14:26:03 +03:00
if mode == SnapSync {
2020-09-08 11:13:16 +03:00
d . pivotLock . Lock ( )
d . pivotHeader = pivot
d . pivotLock . Unlock ( )
2021-11-26 14:26:03 +03:00
fetchers = append ( fetchers , func ( ) error { return d . processSnapSyncContent ( ) } )
2020-06-30 20:43:29 +03:00
} else if mode == FullSync {
2022-03-11 15:14:45 +03:00
fetchers = append ( fetchers , func ( ) error { return d . processFullSyncContent ( ttd , beaconMode ) } )
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
}
2018-02-05 19:40:32 +03:00
return d . spawnSync ( fetchers )
2015-11-13 18:08:15 +02:00
}
// 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 ) )
2018-04-16 11:37:48 +03:00
d . cancelWg . Add ( len ( fetchers ) )
2015-11-13 18:08:15 +02:00
for _ , fn := range fetchers {
fn := fn
2018-04-16 11:37:48 +03:00
go func ( ) { defer d . cancelWg . Done ( ) ; errc <- fn ( ) } ( )
2015-11-13 18:08:15 +02:00
}
// 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 {
2015-11-13 18:08:15 +02:00
// 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 ( )
}
2023-05-05 08:55:32 +03:00
if got := <- errc ; got != nil {
err = got
if got != errCanceled {
break // receive a meaningful error, bubble it up
}
2015-11-13 18:08:15 +02:00
}
}
d . queue . Close ( )
2017-03-22 03:37:24 +03:00
d . Cancel ( )
2015-11-13 18:08:15 +02:00
return err
2015-04-12 13:38:25 +03:00
}
2018-04-23 10:01:21 +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 ( ) {
2015-05-13 14:01:08 +03:00
// Close the current cancel channel
2015-05-15 19:43:42 +03:00
d . cancelLock . Lock ( )
2020-04-27 11:22:15 +03:00
defer d . cancelLock . Unlock ( )
2015-06-12 13:35:29 +03:00
if d . cancelCh != nil {
select {
case <- d . cancelCh :
// Channel was already closed
default :
close ( d . cancelCh )
}
2015-05-15 19:43:42 +03:00
}
2018-04-23 10:01:21 +03:00
}
// Cancel aborts all of the operations and waits for all download goroutines to
// finish before returning.
func ( d * Downloader ) Cancel ( ) {
d . cancel ( )
2018-04-16 11:37:48 +03:00
d . cancelWg . Wait ( )
2015-05-10 01:34:07 +03:00
}
2015-06-18 00:04:57 +03:00
// Terminate interrupts the downloader, canceling all pending operations.
2015-11-13 18:08:15 +02:00
// The downloader cannot be reused after calling Terminate.
2015-06-18 00:04:57 +03:00
func ( d * Downloader ) Terminate ( ) {
2016-06-01 18:07:25 +03:00
// Close the termination channel (make sure double close is allowed)
d . quitLock . Lock ( )
select {
case <- d . quitCh :
default :
close ( d . quitCh )
2022-03-11 15:14:45 +03:00
// Terminate the internal beacon syncer
d . skeleton . Terminate ( )
2016-06-01 18:07:25 +03:00
}
d . quitLock . Unlock ( )
// Cancel any pending download requests
2017-03-22 03:37:24 +03:00
d . Cancel ( )
2015-06-18 00:04:57 +03:00
}
2020-09-08 11:13:16 +03:00
// fetchHead retrieves the head header and prior pivot block (if available) from
// a remote peer.
func ( d * Downloader ) fetchHead ( p * peerConnection ) ( head * types . Header , pivot * types . Header , err error ) {
p . log . Debug ( "Retrieving remote chain head" )
mode := d . getMode ( )
2015-09-09 19:02:54 +03:00
// Request the advertised remote head block and wait for the response
2020-09-08 11:13:16 +03:00
latest , _ := p . peer . Head ( )
fetch := 1
2021-11-26 14:26:03 +03:00
if mode == SnapSync {
2020-09-08 11:13:16 +03:00
fetch = 2 // head + pivot headers
}
2021-12-01 21:18:12 +03:00
headers , hashes , err := d . fetchHeadersByHash ( p , latest , fetch , fsMinFullBlocks - 1 , true )
2021-11-26 14:26:03 +03:00
if err != nil {
return nil , nil , err
}
// Make sure the peer gave us at least one and at most the requested headers
if len ( headers ) == 0 || len ( headers ) > fetch {
return nil , nil , fmt . Errorf ( "%w: returned headers %d != requested %d" , errBadPeer , len ( headers ) , fetch )
}
2023-04-24 09:37:10 +03:00
// The first header needs to be the head, validate against the request. If
// only 1 header was returned, make sure there's no pivot or there was not
// one requested.
2021-11-26 14:26:03 +03:00
head = headers [ 0 ]
if len ( headers ) == 1 {
if mode == SnapSync && head . Number . Uint64 ( ) > uint64 ( fsMinFullBlocks ) {
return nil , nil , fmt . Errorf ( "%w: no pivot included along head header" , errBadPeer )
2015-09-09 19:02:54 +03:00
}
2021-12-01 21:18:12 +03:00
p . log . Debug ( "Remote head identified, no pivot" , "number" , head . Number , "hash" , hashes [ 0 ] )
2021-11-26 14:26:03 +03:00
return head , nil , nil
}
// At this point we have 2 headers in total and the first is the
// validated head of the chain. Check the pivot number and return,
pivot = headers [ 1 ]
if pivot . Number . Uint64 ( ) != head . Number . Uint64 ( ) - uint64 ( fsMinFullBlocks ) {
return nil , nil , fmt . Errorf ( "%w: remote pivot %d != requested %d" , errInvalidChain , pivot . Number , head . Number . Uint64 ( ) - uint64 ( fsMinFullBlocks ) )
2015-09-09 19:02:54 +03:00
}
2021-11-26 14:26:03 +03:00
return head , pivot , nil
2015-09-09 19:02:54 +03:00
}
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:
2022-09-10 14:25:40 +03:00
//
// from - starting block number
// count - number of headers to request
// skip - number of headers to skip
//
2018-11-12 16:18:56 +03:00
// 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
2020-05-06 16:35:04 +03:00
// Ideally, we want to include the one just below our own head
2018-11-12 16:18:56 +03:00
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 )
}
2015-09-28 19:27:31 +03:00
// findAncestor tries to locate the common ancestor link of the local chain and
2015-08-14 21:25:41 +03:00
// a remote peers blockchain. In the general case when our node was in sync and
2015-09-28 19:27:31 +03:00
// on the correct chain, checking the top N links should already get us a match.
2016-03-15 20:55:39 +02:00
// In the rare scenario when we ended up on a long reorganisation (i.e. none of
2015-09-28 19:27:31 +03:00
// 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 ) {
2016-05-13 13:12:13 +03:00
// 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 ( )
)
2020-06-30 20:43:29 +03:00
mode := d . getMode ( )
switch mode {
2018-11-12 16:18:56 +03:00
case FullSync :
2023-03-02 09:29:15 +03:00
localHeight = d . blockchain . CurrentBlock ( ) . Number . Uint64 ( )
2021-11-26 14:26:03 +03:00
case SnapSync :
2023-03-02 09:29:15 +03:00
localHeight = d . blockchain . CurrentSnapBlock ( ) . Number . Uint64 ( )
2018-11-12 16:18:56 +03:00
default :
localHeight = d . lightchain . CurrentHeader ( ) . Number . Uint64 ( )
2016-05-13 13:12:13 +03:00
}
2018-11-12 16:18:56 +03:00
p . log . Debug ( "Looking for common ancestor" , "local" , localHeight , "remote" , remoteHeight )
2019-03-14 13:19:03 +03:00
// Recap floor value for binary search
2020-07-13 12:02:54 +03:00
maxForkAncestry := fullMaxForkAncestry
if d . getMode ( ) == LightSync {
maxForkAncestry = lightMaxForkAncestry
}
2019-05-15 14:33:33 +03:00
if localHeight >= maxForkAncestry {
2018-11-28 14:31:42 +03:00
// We're above the max reorg threshold, find the earliest fork point
2019-05-15 14:33:33 +03:00
floor = int64 ( localHeight - maxForkAncestry )
2019-03-14 13:19:03 +03:00
}
// 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.
2020-06-30 20:43:29 +03:00
if mode == LightSync {
2020-05-25 11:21:28 +03:00
// If we don't know the current CHT position, find it
2019-03-14 13:19:03 +03:00
if d . genesis == 0 {
header := d . lightchain . CurrentHeader ( )
for header != nil {
d . genesis = header . Number . Uint64 ( )
if floor >= int64 ( d . genesis ) - 1 {
break
2018-11-28 14:31:42 +03:00
}
2019-03-14 13:19:03 +03:00
header = d . lightchain . GetHeaderByHash ( header . ParentHash )
2018-11-28 14:31:42 +03:00
}
2019-03-14 13:19:03 +03:00
}
// We already know the "genesis" block number, cap floor to that
if floor < int64 ( d . genesis ) - 1 {
floor = int64 ( d . genesis ) - 1
2018-11-28 14:31:42 +03:00
}
2016-05-13 13:12:13 +03:00
}
2019-03-14 13:19:03 +03:00
2021-01-21 00:45:01 +03:00
ancestor , err := d . findAncestorSpanSearch ( p , mode , remoteHeight , localHeight , floor )
if err == nil {
return ancestor , nil
}
// The returned error was not nil.
// If the error returned does not reflect that a common ancestor was not found, return it.
// If the error reflects that a common ancestor was not found, continue to binary search,
// where the error value will be reassigned.
if ! errors . Is ( err , errNoAncestorFound ) {
return 0 , err
}
ancestor , err = d . findAncestorBinarySearch ( p , mode , remoteHeight , floor )
if err != nil {
return 0 , err
}
return ancestor , nil
}
2021-11-26 14:26:03 +03:00
func ( d * Downloader ) findAncestorSpanSearch ( p * peerConnection , mode SyncMode , remoteHeight , localHeight uint64 , floor int64 ) ( uint64 , error ) {
2018-11-12 16:18:56 +03:00
from , count , skip , max := calculateRequestSpan ( remoteHeight , localHeight )
2018-11-16 14:15:05 +03:00
p . log . Trace ( "Span searching for common ancestor" , "count" , count , "from" , from , "skip" , skip )
2021-12-01 21:18:12 +03:00
headers , hashes , err := d . fetchHeadersByNumber ( p , uint64 ( from ) , count , skip , false )
2021-11-26 14:26:03 +03:00
if err != nil {
return 0 , err
}
2015-08-14 21:25:41 +03:00
// Wait for the remote response to the head fetch
number , hash := uint64 ( 0 ) , common . Hash { }
2017-02-24 19:23:03 +03:00
2021-11-26 14:26:03 +03:00
// Make sure the peer actually gave something valid
if len ( headers ) == 0 {
p . log . Warn ( "Empty head header set" )
return 0 , errEmptyHeaderSet
}
// Make sure the peer's reply conforms to the request
for i , header := range headers {
expectNumber := from + int64 ( i ) * int64 ( skip + 1 )
if number := header . Number . Int64 ( ) ; number != expectNumber {
p . log . Warn ( "Head headers broke chain ordering" , "index" , i , "requested" , expectNumber , "received" , number )
return 0 , fmt . Errorf ( "%w: %v" , errInvalidChain , errors . New ( "head headers broke chain ordering" ) )
}
}
// Check if a common ancestor was found
for i := len ( headers ) - 1 ; i >= 0 ; i -- {
// Skip any headers that underflow/overflow our requested set
if headers [ i ] . Number . Int64 ( ) < from || headers [ i ] . Number . Uint64 ( ) > max {
continue
}
// Otherwise check if we already know the header or not
2021-12-01 21:18:12 +03:00
h := hashes [ i ]
2021-11-26 14:26:03 +03:00
n := headers [ i ] . Number . Uint64 ( )
var known bool
switch mode {
case FullSync :
known = d . blockchain . HasBlock ( h , n )
case SnapSync :
known = d . blockchain . HasFastBlock ( h , n )
default :
known = d . lightchain . HasHeader ( h , n )
}
if known {
number , hash = n , h
break
2015-08-14 21:25:41 +03:00
}
}
// If the head fetch already found an ancestor, return
2018-05-29 13:42:21 +03:00
if hash != ( common . Hash { } ) {
2016-05-13 13:12:13 +03:00
if int64 ( number ) <= floor {
2017-03-02 16:06:16 +03:00
p . log . Warn ( "Ancestor below allowance" , "number" , number , "hash" , hash , "allowance" , floor )
2016-05-13 13:12:13 +03:00
return 0 , errInvalidAncestor
}
2017-03-02 16:06:16 +03:00
p . log . Debug ( "Found common ancestor" , "number" , number , "hash" , hash )
2015-08-14 21:25:41 +03:00
return number , nil
}
2021-01-21 00:45:01 +03:00
return 0 , errNoAncestorFound
}
2021-11-26 14:26:03 +03:00
func ( d * Downloader ) findAncestorBinarySearch ( p * peerConnection , mode SyncMode , remoteHeight uint64 , floor int64 ) ( uint64 , error ) {
2021-01-21 00:45:01 +03:00
hash := common . Hash { }
2015-08-14 21:25:41 +03:00
// Ancestor not found, we need to binary search over our chain
2018-11-12 16:18:56 +03:00
start , end := uint64 ( 0 ) , remoteHeight
2016-05-13 13:12:13 +03:00
if floor > 0 {
start = uint64 ( floor )
}
2018-11-16 14:15:05 +03:00
p . log . Trace ( "Binary searching for common ancestor" , "start" , start , "end" , end )
2015-08-14 21:25:41 +03:00
for start + 1 < end {
// Split our chain interval in two, and request the hash to cross check
check := ( start + end ) / 2
2021-12-01 21:18:12 +03:00
headers , hashes , err := d . fetchHeadersByNumber ( p , check , 1 , 0 , false )
2021-11-26 14:26:03 +03:00
if err != nil {
return 0 , err
}
// Make sure the peer actually gave something valid
if len ( headers ) != 1 {
p . log . Warn ( "Multiple headers for single request" , "headers" , len ( headers ) )
return 0 , fmt . Errorf ( "%w: multiple headers (%d) for single request" , errBadPeer , len ( headers ) )
}
// Modify the search interval based on the response
2021-12-01 21:18:12 +03:00
h := hashes [ 0 ]
2021-11-26 14:26:03 +03:00
n := headers [ 0 ] . Number . Uint64 ( )
var known bool
switch mode {
case FullSync :
known = d . blockchain . HasBlock ( h , n )
case SnapSync :
known = d . blockchain . HasFastBlock ( h , n )
default :
known = d . lightchain . HasHeader ( h , n )
2015-08-14 21:25:41 +03:00
}
2021-11-26 14:26:03 +03:00
if ! known {
end = check
continue
}
header := d . lightchain . GetHeaderByHash ( h ) // Independent of sync mode, header surely exists
if header . Number . Uint64 ( ) != check {
p . log . Warn ( "Received non requested header" , "number" , header . Number , "hash" , header . Hash ( ) , "request" , check )
return 0 , fmt . Errorf ( "%w: non-requested header (%d)" , errBadPeer , header . Number )
}
start = check
hash = h
2015-08-14 21:25:41 +03:00
}
2016-05-13 13:12:13 +03:00
// Ensure valid ancestry and return
if int64 ( start ) <= floor {
2017-03-02 16:06:16 +03:00
p . log . Warn ( "Ancestor below allowance" , "number" , start , "hash" , hash , "allowance" , floor )
2016-05-13 13:12:13 +03:00
return 0 , errInvalidAncestor
}
2017-03-02 16:06:16 +03:00
p . log . Debug ( "Found common ancestor" , "number" , start , "hash" , hash )
2015-08-14 21:25:41 +03:00
return start , nil
}
2016-02-25 18:36:42 +02:00
// 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
2016-02-25 18:36:42 +02:00
// can fill in the skeleton - not even the origin peer - it's assumed invalid and
// the origin is dropped.
2021-11-26 14:26:03 +03:00
func ( d * Downloader ) fetchHeaders ( p * peerConnection , from uint64 , head uint64 ) error {
2017-03-02 16:06:16 +03:00
p . log . Debug ( "Directing header downloads" , "origin" , from )
defer p . log . Debug ( "Header download terminated" )
2015-08-14 21:25:41 +03:00
2021-11-26 14:26:03 +03:00
// Start pulling the header chain skeleton until all is done
var (
skeleton = true // Skeleton assembly phase or finishing up
pivoting = false // Whether the next request is pivot verification
ancestor = from
mode = d . getMode ( )
)
for {
// Pull the next batch of headers, it either:
// - Pivot check to see if the chain moved too far
// - Skeleton retrieval to permit concurrent header fetches
// - Full header retrieval if we're near the chain head
var (
headers [ ] * types . Header
2021-12-01 21:18:12 +03:00
hashes [ ] common . Hash
2021-11-26 14:26:03 +03:00
err error
)
switch {
case pivoting :
d . pivotLock . RLock ( )
pivot := d . pivotHeader . Number . Uint64 ( )
d . pivotLock . RUnlock ( )
2017-02-24 19:23:03 +03:00
2021-11-26 14:26:03 +03:00
p . log . Trace ( "Fetching next pivot header" , "number" , pivot + uint64 ( fsMinFullBlocks ) )
2021-12-01 21:18:12 +03:00
headers , hashes , err = d . fetchHeadersByNumber ( p , pivot + uint64 ( fsMinFullBlocks ) , 2 , fsMinFullBlocks - 9 , false ) // move +64 when it's 2x64-8 deep
2016-05-17 11:12:57 +03:00
2021-11-26 14:26:03 +03:00
case skeleton :
2017-03-02 16:06:16 +03:00
p . log . Trace ( "Fetching skeleton headers" , "count" , MaxHeaderFetch , "from" , from )
2021-12-01 21:18:12 +03:00
headers , hashes , err = d . fetchHeadersByNumber ( p , from + uint64 ( MaxHeaderFetch ) - 1 , MaxSkeletonSize , MaxHeaderFetch - 1 , false )
2021-11-26 14:26:03 +03:00
default :
2017-03-02 16:06:16 +03:00
p . log . Trace ( "Fetching full headers" , "count" , MaxHeaderFetch , "from" , from )
2021-12-01 21:18:12 +03:00
headers , hashes , err = d . fetchHeadersByNumber ( p , from , MaxHeaderFetch , 0 , false )
2016-02-25 18:36:42 +02:00
}
2021-11-26 14:26:03 +03:00
switch err {
case nil :
// Headers retrieved, continue with processing
2020-09-08 11:13:16 +03:00
2021-11-26 14:26:03 +03:00
case errCanceled :
// Sync cancelled, no issue, propagate up
return err
2015-08-14 21:25:41 +03:00
2021-11-26 14:26:03 +03:00
default :
// Header retrieval either timed out, or the peer failed in some strange way
// (e.g. disconnect). Consider the master peer bad and drop
d . dropPeer ( p . id )
2015-08-14 21:25:41 +03:00
2021-11-26 14:26:03 +03:00
// Finish the sync gracefully instead of dumping the gathered data though
for _ , ch := range [ ] chan bool { d . queue . blockWakeCh , d . queue . receiptWakeCh } {
select {
case ch <- false :
case <- d . cancelCh :
}
2015-08-14 21:25:41 +03:00
}
2021-11-26 14:26:03 +03:00
select {
case d . headerProcCh <- nil :
case <- d . cancelCh :
2020-09-08 11:13:16 +03:00
}
2021-11-26 14:26:03 +03:00
return fmt . Errorf ( "%w: header request failed: %v" , errBadPeer , err )
}
// If the pivot is being checked, move if it became stale and run the real retrieval
var pivot uint64
2020-09-08 11:13:16 +03:00
2021-11-26 14:26:03 +03:00
d . pivotLock . RLock ( )
if d . pivotHeader != nil {
pivot = d . pivotHeader . Number . Uint64 ( )
}
d . pivotLock . RUnlock ( )
2020-09-08 11:13:16 +03:00
2021-11-26 14:26:03 +03:00
if pivoting {
if len ( headers ) == 2 {
if have , want := headers [ 0 ] . Number . Uint64 ( ) , pivot + uint64 ( fsMinFullBlocks ) ; have != want {
log . Warn ( "Peer sent invalid next pivot" , "have" , have , "want" , want )
return fmt . Errorf ( "%w: next pivot number %d != requested %d" , errInvalidChain , have , want )
}
if have , want := headers [ 1 ] . Number . Uint64 ( ) , pivot + 2 * uint64 ( fsMinFullBlocks ) - 8 ; have != want {
log . Warn ( "Peer sent invalid pivot confirmer" , "have" , have , "want" , want )
return fmt . Errorf ( "%w: next pivot confirmer number %d != requested %d" , errInvalidChain , have , want )
}
log . Warn ( "Pivot seemingly stale, moving" , "old" , pivot , "new" , headers [ 0 ] . Number )
pivot = headers [ 0 ] . Number . Uint64 ( )
2020-09-08 11:13:16 +03:00
2021-11-26 14:26:03 +03:00
d . pivotLock . Lock ( )
d . pivotHeader = headers [ 0 ]
d . pivotLock . Unlock ( )
2020-09-08 11:13:16 +03:00
2021-11-26 14:26:03 +03:00
// Write out the pivot into the database so a rollback beyond
// it will reenable snap sync and update the state root that
// the state syncer will be downloading.
rawdb . WriteLastPivotNumber ( d . stateDB , pivot )
2020-09-08 11:13:16 +03:00
}
2021-11-26 14:26:03 +03:00
// Disable the pivot check and fetch the next batch of headers
pivoting = false
continue
}
// If the skeleton's finished, pull any remaining head headers directly from the origin
if skeleton && len ( headers ) == 0 {
// A malicious node might withhold advertised headers indefinitely
if from + uint64 ( MaxHeaderFetch ) - 1 <= head {
p . log . Warn ( "Peer withheld skeleton headers" , "advertised" , head , "withheld" , from + uint64 ( MaxHeaderFetch ) - 1 )
return fmt . Errorf ( "%w: withheld skeleton headers: advertised %d, withheld #%d" , errStallingPeer , head , from + uint64 ( MaxHeaderFetch ) - 1 )
2016-02-25 18:36:42 +02:00
}
2021-11-26 14:26:03 +03:00
p . log . Debug ( "No skeleton, fetching headers directly" )
skeleton = false
continue
}
// If no more headers are inbound, notify the content fetchers and return
if len ( headers ) == 0 {
// Don't abort header fetches while the pivot is downloading
2023-04-03 22:48:10 +03:00
if ! d . committed . Load ( ) && pivot <= from {
2021-11-26 14:26:03 +03:00
p . log . Debug ( "No headers, waiting for pivot commit" )
2016-06-02 12:37:14 +03:00
select {
2021-11-26 14:26:03 +03:00
case <- time . After ( fsHeaderContCheck ) :
continue
2016-06-02 12:37:14 +03:00
case <- d . cancelCh :
2019-06-05 15:00:46 +03:00
return errCanceled
2016-06-02 12:37:14 +03:00
}
2015-08-14 21:25:41 +03:00
}
2021-11-26 14:26:03 +03:00
// Pivot done (or not in snap 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 errCanceled
2015-08-14 21:25:41 +03:00
}
2021-11-26 14:26:03 +03:00
}
// If we received a skeleton batch, resolve internals concurrently
var progressed bool
if skeleton {
2021-12-01 21:18:12 +03:00
filled , hashset , proced , err := d . fillHeaderSkeleton ( from , headers )
2021-11-26 14:26:03 +03:00
if err != nil {
p . log . Debug ( "Skeleton chain invalid" , "err" , err )
return fmt . Errorf ( "%w: %v" , errInvalidChain , err )
}
headers = filled [ proced : ]
2021-12-01 21:18:12 +03:00
hashes = hashset [ proced : ]
2021-11-26 14:26:03 +03:00
progressed = proced > 0
from += uint64 ( proced )
} else {
// A malicious node might withhold advertised headers indefinitely
if n := len ( headers ) ; n < MaxHeaderFetch && headers [ n - 1 ] . Number . Uint64 ( ) < head {
p . log . Warn ( "Peer withheld headers" , "advertised" , head , "delivered" , headers [ n - 1 ] . Number . Uint64 ( ) )
return fmt . Errorf ( "%w: withheld headers: advertised %d, delivered %d" , errStallingPeer , head , headers [ n - 1 ] . Number . Uint64 ( ) )
}
// 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
var head uint64
if mode == LightSync {
head = d . lightchain . CurrentHeader ( ) . Number . Uint64 ( )
2020-09-08 11:13:16 +03:00
} else {
2023-03-02 09:29:15 +03:00
head = d . blockchain . CurrentSnapBlock ( ) . Number . Uint64 ( )
if full := d . blockchain . CurrentBlock ( ) . Number . Uint64 ( ) ; head < full {
2021-11-26 14:26:03 +03:00
head = full
}
2020-09-08 11:13:16 +03:00
}
2021-11-26 14:26:03 +03:00
// 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
2018-10-04 16:36:59 +03:00
}
2021-11-26 14:26:03 +03:00
// 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 ]
2021-12-01 21:18:12 +03:00
hashes = hashes [ : n - delay ]
2015-09-28 19:27:31 +03:00
}
2015-08-14 21:25:41 +03:00
}
2021-11-26 14:26:03 +03:00
}
2023-01-03 12:03:44 +03:00
// If no headers have been delivered, or all of them have been delayed,
2021-11-26 14:26:03 +03:00
// sleep a bit and retry. Take care with headers already consumed during
// skeleton filling
if len ( headers ) == 0 && ! progressed {
p . log . Trace ( "All headers delayed, waiting" )
2016-02-25 18:36:42 +02:00
select {
2021-11-26 14:26:03 +03:00
case <- time . After ( fsHeaderContCheck ) :
continue
2016-02-25 18:36:42 +02:00
case <- d . cancelCh :
2021-11-26 14:26:03 +03:00
return errCanceled
2016-02-25 18:36:42 +02:00
}
2021-11-26 14:26:03 +03:00
}
// Insert any remaining new headers and fetch the next batch
if len ( headers ) > 0 {
p . log . Trace ( "Scheduling new headers" , "count" , len ( headers ) , "from" , from )
select {
2021-12-01 21:18:12 +03:00
case d . headerProcCh <- & headerTask {
headers : headers ,
hashes : hashes ,
} :
2021-11-26 14:26:03 +03:00
case <- d . cancelCh :
return errCanceled
}
from += uint64 ( len ( headers ) )
}
// If we're still skeleton filling snap sync, check pivot staleness
// before continuing to the next skeleton filling
if skeleton && pivot > 0 {
pivoting = true
2015-08-14 21:25:41 +03:00
}
}
}
2016-02-25 18:36:42 +02:00
// fillHeaderSkeleton concurrently retrieves headers from all our available peers
// and maps them to the provided skeleton header chain.
2016-04-19 12:27:37 +03:00
//
// 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
2016-04-19 12:27:37 +03:00
// already forwarded for processing.
2021-12-01 21:18:12 +03:00
func ( d * Downloader ) fillHeaderSkeleton ( from uint64 , skeleton [ ] * types . Header ) ( [ ] * types . Header , [ ] common . Hash , int , error ) {
2017-02-24 19:23:03 +03:00
log . Debug ( "Filling up skeleton" , "from" , from )
2016-02-25 18:36:42 +02:00
d . queue . ScheduleSkeleton ( from , skeleton )
2022-03-11 15:14:45 +03:00
err := d . concurrentFetch ( ( * headerQueue ) ( d ) , false )
2021-11-26 14:26:03 +03:00
if err != nil {
log . Debug ( "Skeleton fill failed" , "err" , err )
}
2021-12-01 21:18:12 +03:00
filled , hashes , proced := d . queue . RetrieveHeaders ( )
2021-11-26 14:26:03 +03:00
if err == nil {
log . Debug ( "Skeleton fill succeeded" , "filled" , len ( filled ) , "processed" , proced )
}
2021-12-01 21:18:12 +03:00
return filled , hashes , proced , err
2016-02-25 18:36:42 +02:00
}
2015-08-14 21:25:41 +03:00
// 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.
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) fetchBodies ( from uint64 , beaconMode bool ) error {
2017-02-24 19:23:03 +03:00
log . Debug ( "Downloading block bodies" , "origin" , from )
2022-03-11 15:14:45 +03:00
err := d . concurrentFetch ( ( * bodyQueue ) ( d ) , beaconMode )
2015-09-28 19:27:31 +03:00
2017-02-27 14:17:58 +03:00
log . Debug ( "Block body download terminated" , "err" , err )
2015-09-28 19:27:31 +03:00
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.
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) fetchReceipts ( from uint64 , beaconMode bool ) error {
2021-11-26 14:26:03 +03:00
log . Debug ( "Downloading receipts" , "origin" , from )
2022-03-11 15:14:45 +03:00
err := d . concurrentFetch ( ( * receiptQueue ) ( d ) , beaconMode )
2015-09-28 19:27:31 +03:00
2021-11-26 14:26:03 +03:00
log . Debug ( "Receipt download terminated" , "err" , err )
2015-09-28 19:27:31 +03:00
return err
}
2016-02-25 18:36:42 +02:00
// 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.
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) processHeaders ( origin uint64 , td , ttd * big . Int , beaconMode bool ) error {
2020-07-24 10:46:26 +03:00
var (
2023-09-26 16:29:13 +03:00
mode = d . getMode ( )
gotHeaders = false // Wait for batches of headers to process
2020-07-24 10:46:26 +03:00
)
2016-02-25 18:36:42 +02:00
for {
select {
case <- d . cancelCh :
2019-06-05 15:00:46 +03:00
return errCanceled
2016-02-25 18:36:42 +02:00
2021-12-01 21:18:12 +03:00
case task := <- d . headerProcCh :
2016-02-25 18:36:42 +02:00
// Terminate header processing if we synced up
2021-12-01 21:18:12 +03:00
if task == nil || len ( task . headers ) == 0 {
2016-02-25 18:36:42 +02:00
// Notify everyone that headers are fully processed
2021-11-26 14:26:03 +03:00
for _ , ch := range [ ] chan bool { d . queue . blockWakeCh , d . queue . receiptWakeCh } {
2016-02-25 18:36:42 +02:00
select {
case ch <- false :
case <- d . cancelCh :
}
}
2022-03-11 15:14:45 +03:00
// If we're in legacy sync mode, we need to check total difficulty
// violations from malicious peers. That is not needed in beacon
// mode and we can skip to terminating sync.
if ! beaconMode {
// 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 mode != LightSync {
head := d . blockchain . CurrentBlock ( )
2023-03-02 09:29:15 +03:00
if ! gotHeaders && td . Cmp ( d . blockchain . GetTd ( head . Hash ( ) , head . Number . Uint64 ( ) ) ) > 0 {
2022-03-11 15:14:45 +03:00
return errStallingPeer
}
2016-01-13 20:35:48 +02:00
}
2022-03-11 15:14:45 +03:00
// If snap 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 mode == SnapSync || mode == LightSync {
head := d . lightchain . CurrentHeader ( )
if td . Cmp ( d . lightchain . GetTd ( head . Hash ( ) , head . Number . Uint64 ( ) ) ) > 0 {
return errStallingPeer
}
2016-02-25 18:36:42 +02:00
}
}
return nil
}
// Otherwise split the chunk of headers into batches and process them
2021-12-01 21:18:12 +03:00
headers , hashes := task . headers , task . hashes
2016-02-25 18:36:42 +02:00
gotHeaders = true
for len ( headers ) > 0 {
// Terminate if something failed in between processing chunks
select {
case <- d . cancelCh :
2019-06-05 15:00:46 +03:00
return errCanceled
2016-02-25 18:36:42 +02:00
default :
}
// Select the next chunk of headers to import
limit := maxHeadersProcess
if limit > len ( headers ) {
limit = len ( headers )
}
2021-12-01 21:18:12 +03:00
chunkHeaders := headers [ : limit ]
chunkHashes := hashes [ : limit ]
2020-08-20 13:01:24 +03:00
2016-02-25 18:36:42 +02:00
// In case of header only syncing, validate the chunk immediately
2021-11-26 14:26:03 +03:00
if mode == SnapSync || mode == LightSync {
2022-03-11 15:14:45 +03:00
// Although the received headers might be all valid, a legacy
// PoW/PoA sync must not accept post-merge headers. Make sure
// that any transition is rejected at this point.
var (
rejected [ ] * types . Header
td * big . Int
)
if ! beaconMode && ttd != nil {
td = d . blockchain . GetTd ( chunkHeaders [ 0 ] . ParentHash , chunkHeaders [ 0 ] . Number . Uint64 ( ) - 1 )
if td == nil {
// This should never really happen, but handle gracefully for now
log . Error ( "Failed to retrieve parent header TD" , "number" , chunkHeaders [ 0 ] . Number . Uint64 ( ) - 1 , "hash" , chunkHeaders [ 0 ] . ParentHash )
return fmt . Errorf ( "%w: parent TD missing" , errInvalidChain )
}
for i , header := range chunkHeaders {
td = new ( big . Int ) . Add ( td , header . Difficulty )
if td . Cmp ( ttd ) >= 0 {
// Terminal total difficulty reached, allow the last header in
if new ( big . Int ) . Sub ( td , header . Difficulty ) . Cmp ( ttd ) < 0 {
chunkHeaders , rejected = chunkHeaders [ : i + 1 ] , chunkHeaders [ i + 1 : ]
if len ( rejected ) > 0 {
// Make a nicer user log as to the first TD truly rejected
td = new ( big . Int ) . Add ( td , rejected [ 0 ] . Difficulty )
}
} else {
chunkHeaders , rejected = chunkHeaders [ : i ] , chunkHeaders [ i : ]
}
break
}
}
}
2022-03-22 20:58:05 +03:00
if len ( chunkHeaders ) > 0 {
2023-05-03 12:58:39 +03:00
if n , err := d . lightchain . InsertHeaderChain ( chunkHeaders ) ; err != nil {
2022-03-22 20:58:05 +03:00
log . Warn ( "Invalid header encountered" , "number" , chunkHeaders [ n ] . Number , "hash" , chunkHashes [ n ] , "parent" , chunkHeaders [ n ] . ParentHash , "err" , err )
return fmt . Errorf ( "%w: %v" , errInvalidChain , err )
2016-02-25 18:36:42 +02:00
}
}
2022-03-11 15:14:45 +03:00
if len ( rejected ) != 0 {
log . Info ( "Legacy sync reached merge threshold" , "number" , rejected [ 0 ] . Number , "hash" , rejected [ 0 ] . Hash ( ) , "td" , td , "ttd" , ttd )
return ErrMergeTransition
}
2016-02-25 18:36:42 +02:00
}
// Unless we're doing light chains, schedule the headers for associated content retrieval
2021-11-26 14:26:03 +03:00
if mode == FullSync || mode == SnapSync {
2016-02-25 18:36:42 +02:00
// If we've reached the allowed number of pending headers, stall a bit
2021-11-26 14:26:03 +03:00
for d . queue . PendingBodies ( ) >= maxQueuedHeaders || d . queue . PendingReceipts ( ) >= maxQueuedHeaders {
2016-02-25 18:36:42 +02:00
select {
case <- d . cancelCh :
2019-06-05 15:00:46 +03:00
return errCanceled
2016-02-25 18:36:42 +02:00
case <- time . After ( time . Second ) :
}
}
// Otherwise insert the headers for content retrieval
2021-12-01 21:18:12 +03:00
inserts := d . queue . Schedule ( chunkHeaders , chunkHashes , origin )
if len ( inserts ) != len ( chunkHeaders ) {
2020-07-09 00:08:08 +03:00
return fmt . Errorf ( "%w: stale headers" , errBadPeer )
2016-02-25 18:36:42 +02:00
}
}
headers = headers [ limit : ]
2021-12-01 21:18:12 +03:00
hashes = hashes [ limit : ]
2016-02-25 18:36:42 +02:00
origin += uint64 ( limit )
}
2018-03-09 12:51:30 +03:00
// 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 ( )
2022-08-19 09:00:21 +03:00
// Signal the content downloaders of the availability of new tasks
2021-11-26 14:26:03 +03:00
for _ , ch := range [ ] chan bool { d . queue . blockWakeCh , d . queue . receiptWakeCh } {
2016-02-25 18:36:42 +02:00
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.
2022-03-11 15:14:45 +03:00
func ( d * Downloader ) processFullSyncContent ( ttd * big . Int , beaconMode bool ) error {
2015-06-12 13:35:29 +03:00
for {
2018-02-05 19:40:32 +03:00
results := d . queue . Results ( true )
2015-09-28 19:27:31 +03:00
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
2015-06-12 13:35:29 +03:00
}
2015-08-14 21:25:41 +03:00
if d . chainInsertHook != nil {
2015-09-28 19:27:31 +03:00
d . chainInsertHook ( results )
2015-08-14 21:25:41 +03:00
}
2022-03-11 15:14:45 +03:00
// Although the received blocks might be all valid, a legacy PoW/PoA sync
// must not accept post-merge blocks. Make sure that pre-merge blocks are
// imported, but post-merge ones are rejected.
var (
rejected [ ] * fetchResult
td * big . Int
)
if ! beaconMode && ttd != nil {
td = d . blockchain . GetTd ( results [ 0 ] . Header . ParentHash , results [ 0 ] . Header . Number . Uint64 ( ) - 1 )
if td == nil {
// This should never really happen, but handle gracefully for now
log . Error ( "Failed to retrieve parent block TD" , "number" , results [ 0 ] . Header . Number . Uint64 ( ) - 1 , "hash" , results [ 0 ] . Header . ParentHash )
return fmt . Errorf ( "%w: parent TD missing" , errInvalidChain )
}
for i , result := range results {
td = new ( big . Int ) . Add ( td , result . Header . Difficulty )
if td . Cmp ( ttd ) >= 0 {
// Terminal total difficulty reached, allow the last block in
if new ( big . Int ) . Sub ( td , result . Header . Difficulty ) . Cmp ( ttd ) < 0 {
results , rejected = results [ : i + 1 ] , results [ i + 1 : ]
if len ( rejected ) > 0 {
// Make a nicer user log as to the first TD truly rejected
td = new ( big . Int ) . Add ( td , rejected [ 0 ] . Header . Difficulty )
}
} else {
results , rejected = results [ : i ] , results [ i : ]
}
break
}
}
}
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
}
2022-03-11 15:14:45 +03:00
if len ( rejected ) != 0 {
log . Info ( "Legacy sync reached merge threshold" , "number" , rejected [ 0 ] . Header . Number , "hash" , rejected [ 0 ] . Header . Hash ( ) , "td" , td , "ttd" , ttd )
return ErrMergeTransition
}
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 ) importBlockResults ( results [ ] * fetchResult ) error {
2018-02-05 19:40:32 +03:00
// Check for any early termination requests
if len ( results ) == 0 {
return nil
}
select {
case <- d . quitCh :
return errCancelContentProcessing
default :
}
2022-07-25 16:51:04 +03:00
// Retrieve a batch of results to import
2018-02-05 19:40:32 +03:00
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 {
2023-01-25 17:32:25 +03:00
blocks [ i ] = types . NewBlockWithHeader ( result . Header ) . WithBody ( result . Transactions , result . Uncles ) . WithWithdrawals ( result . Withdrawals )
2018-02-05 19:40:32 +03:00
}
all: core rework for the merge transition (#23761)
* all: work for eth1/2 transtition
* consensus/beacon, eth: change beacon difficulty to 0
* eth: updates
* all: add terminalBlockDifficulty config, fix rebasing issues
* eth: implemented merge interop spec
* internal/ethapi: update to v1.0.0.alpha.2
This commit updates the code to the new spec, moving payloadId into
it's own object. It also fixes an issue with finalizing an empty blockhash.
It also properly sets the basefee
* all: sync polishes, other fixes + refactors
* core, eth: correct semantics for LeavePoW, EnterPoS
* core: fixed rebasing artifacts
* core: light: performance improvements
* core: use keyed field (f)
* core: eth: fix compilation issues + tests
* eth/catalyst: dbetter error codes
* all: move Merger to consensus/, remove reliance on it in bc
* all: renamed EnterPoS and LeavePoW to ReachTDD and FinalizePoS
* core: make mergelogs a function
* core: use InsertChain instead of InsertBlock
* les: drop merger from lightchain object
* consensus: add merger
* core: recoverAncestors in catalyst mode
* core: fix nitpick
* all: removed merger from beacon, use TTD, nitpicks
* consensus: eth: add docstring, removed unnecessary code duplication
* consensus/beacon: better comment
* all: easy to fix nitpicks by karalabe
* consensus/beacon: verify known headers to be sure
* core: comments
* core: eth: don't drop peers who advertise blocks, nitpicks
* core: never add beacon blocks to the future queue
* core: fixed nitpicks
* consensus/beacon: simplify IsTTDReached check
* consensus/beacon: correct IsTTDReached check
Co-authored-by: rjl493456442 <garyrong0905@gmail.com>
Co-authored-by: Péter Szilágyi <peterke@gmail.com>
2021-11-26 14:23:02 +03:00
// Downloaded blocks are always regarded as trusted after the
// transition. Because the downloaded chain is guided by the
// consensus-layer.
2018-02-05 19:40:32 +03:00
if index , err := d . blockchain . InsertChain ( blocks ) ; err != nil {
2018-12-20 12:46:08 +03:00
if index < len ( results ) {
log . Debug ( "Downloaded item processing failed" , "number" , results [ index ] . Header . Number , "hash" , results [ index ] . Header . Hash ( ) , "err" , err )
2022-07-25 16:51:04 +03:00
// In post-merge, notify the engine API of encountered bad chains
if d . badBlock != nil {
2023-02-23 14:22:41 +03:00
head , _ , _ , err := d . skeleton . Bounds ( )
2022-07-25 16:51:04 +03:00
if err != nil {
log . Error ( "Failed to retrieve beacon bounds for bad block reporting" , "err" , err )
} else {
d . badBlock ( blocks [ index ] . Header ( ) , head )
}
}
2018-12-20 12:46:08 +03:00
} 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 )
}
2020-05-29 12:12:43 +03:00
return fmt . Errorf ( "%w: %v" , errInvalidChain , 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
}
return nil
}
2021-11-26 14:26:03 +03:00
// processSnapSyncContent takes fetch results from the queue and writes them to the
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
// database. It also controls the synchronisation of state nodes of the pivot block.
2021-11-26 14:26:03 +03:00
func ( d * Downloader ) processSnapSyncContent ( ) error {
2018-02-05 19:40:32 +03:00
// Start syncing state of the reported head block. This should get us most of
// the state of the pivot block.
2020-09-08 11:13:16 +03:00
d . pivotLock . RLock ( )
sync := d . syncState ( d . pivotHeader . Root )
d . pivotLock . RUnlock ( )
2020-08-26 13:05:06 +03:00
defer func ( ) {
// The `sync` object is replaced every time the pivot moves. We need to
// defer close the very last active one, hence the lazy evaluation vs.
// calling defer sync.Cancel() !!!
sync . Cancel ( )
} ( )
2019-10-25 14:17:32 +03:00
closeOnErr := func ( s * stateSync ) {
2021-02-16 17:11:33 +03:00
if err := s . Wait ( ) ; err != nil && err != errCancelStateFetch && err != errCanceled && err != snap . ErrCancelled {
2018-10-23 14:21:16 +03:00
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
}
2019-10-25 14:17:32 +03:00
}
go closeOnErr ( sync )
2020-08-20 13:01:24 +03:00
2018-02-05 19:40:32 +03:00
// To cater for moving pivot points, track the pivot block and subsequently
2018-04-04 13:25:02 +03:00
// accumulated download results separately.
2023-09-15 15:06:25 +03:00
//
// These will be nil up to the point where we reach the pivot, and will only
// be set temporarily if the synced blocks are piling up, but the pivot is
// still busy downloading. In that case, we need to occasionally check for
// pivot moves, so need to unblock the loop. These fields will accumulate
// the results in the meantime.
//
// Note, there's no issue with memory piling up since after 64 blocks the
// pivot will forcefully move so these accumulators will be dropped.
2018-02-05 19:40:32 +03:00
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 {
2023-09-15 15:06:25 +03:00
// Wait for the next batch of downloaded data to be available. If we have
// not yet reached the pivot point, wait blockingly as there's no need to
// spin-loop check for pivot moves. If we reached the pivot but have not
// yet processed it, check for results async, so we might notice pivot
// moves while state syncing. If the pivot was passed fully, block again
// as there's no more reason to check for pivot moves at all.
results := d . queue . Results ( oldPivot == 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
if len ( results ) == 0 {
2018-02-05 19:40:32 +03:00
// If pivot sync is done, stop
2023-09-15 15:06:25 +03:00
if d . committed . Load ( ) {
2023-02-21 13:17:34 +03:00
d . reportSnapSyncProgress ( true )
2019-10-25 14:17:32 +03:00
return sync . Cancel ( )
2018-02-05 19:40:32 +03:00
}
// If sync failed, stop
select {
case <- d . cancelCh :
2019-10-25 14:17:32 +03:00
sync . Cancel ( )
2019-06-05 15:00:46 +03:00
return errCanceled
2018-02-05 19:40:32 +03:00
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 )
}
2023-02-21 13:17:34 +03:00
d . reportSnapSyncProgress ( false )
2020-09-08 11:13:16 +03:00
// If we haven't downloaded the pivot block yet, check pivot staleness
// notifications from the header downloader
d . pivotLock . RLock ( )
pivot := d . pivotHeader
d . pivotLock . RUnlock ( )
2023-09-15 15:06:25 +03:00
if oldPivot == nil { // no results piling up, we can move the pivot
if ! d . committed . Load ( ) { // not yet passed the pivot, we can move the pivot
if pivot . Root != sync . root { // pivot position changed, we can move the pivot
sync . Cancel ( )
sync = d . syncState ( pivot . Root )
2020-09-08 11:13:16 +03:00
2023-09-15 15:06:25 +03:00
go closeOnErr ( sync )
}
2020-09-08 11:13:16 +03:00
}
2023-09-15 15:06:25 +03:00
} else { // results already piled up, consume before handling pivot move
2018-02-05 19:40:32 +03:00
results = append ( append ( [ ] * fetchResult { oldPivot } , oldTail ... ) , results ... )
}
2021-11-26 14:26:03 +03:00
// Split around the pivot block and process the two sides via snap/full sync
2023-04-03 22:48:10 +03:00
if ! d . committed . Load ( ) {
2020-09-08 11:13:16 +03:00
latest := results [ len ( results ) - 1 ] . Header
// If the height is above the pivot block by 2 sets, it means the pivot
2023-09-15 15:06:25 +03:00
// become stale in the network, and it was garbage collected, move to a
2020-09-08 11:13:16 +03:00
// new pivot.
//
// Note, we have `reorgProtHeaderDelay` number of blocks withheld, Those
// need to be taken into account, otherwise we're detecting the pivot move
// late and will drop peers due to unavailable state!!!
if height := latest . Number . Uint64 ( ) ; height >= pivot . Number . Uint64 ( ) + 2 * uint64 ( fsMinFullBlocks ) - uint64 ( reorgProtHeaderDelay ) {
log . Warn ( "Pivot became stale, moving" , "old" , pivot . Number . Uint64 ( ) , "new" , height - uint64 ( fsMinFullBlocks ) + uint64 ( reorgProtHeaderDelay ) )
pivot = results [ len ( results ) - 1 - fsMinFullBlocks + reorgProtHeaderDelay ] . Header // must exist as lower old pivot is uncommitted
d . pivotLock . Lock ( )
d . pivotHeader = pivot
d . pivotLock . Unlock ( )
2020-08-20 13:01:24 +03:00
// Write out the pivot into the database so a rollback beyond it will
2021-11-26 14:26:03 +03:00
// reenable snap sync
2020-09-08 11:13:16 +03:00
rawdb . WriteLastPivotNumber ( d . stateDB , pivot . Number . Uint64 ( ) )
2018-02-05 19:40:32 +03:00
}
}
2020-09-08 11:13:16 +03:00
P , beforeP , afterP := splitAroundPivot ( pivot . Number . Uint64 ( ) , results )
2021-11-26 14:26:03 +03:00
if err := d . commitSnapSyncData ( beforeP , sync ) ; 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 P != nil {
2018-02-05 19:40:32 +03:00
// If new pivot block found, cancel old state retrieval and restart
if oldPivot != P {
2019-10-25 14:17:32 +03:00
sync . Cancel ( )
sync = d . syncState ( P . Header . Root )
2020-08-26 13:05:06 +03:00
2019-10-25 14:17:32 +03:00
go closeOnErr ( sync )
2018-02-05 19:40:32 +03:00
oldPivot = P
}
// Wait for completion, occasionally checking for pivot staleness
select {
2019-10-25 14:17:32 +03:00
case <- sync . done :
if sync . err != nil {
return sync . err
2018-02-05 19:40:32 +03:00
}
if err := d . commitPivotBlock ( P ) ; err != nil {
return err
}
oldPivot = nil
case <- time . After ( time . Second ) :
oldTail = afterP
continue
2015-06-12 13:35:29 +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
}
2018-02-05 19:40:32 +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 ) {
2020-07-24 10:46:26 +03:00
if len ( results ) == 0 {
return nil , nil , nil
}
if lastNum := results [ len ( results ) - 1 ] . Header . Number . Uint64 ( ) ; lastNum < pivot {
// the pivot is somewhere in the future
return nil , results , nil
}
// This can also be optimized, but only happens very seldom
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 _ , 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
}
2021-11-26 14:26:03 +03:00
func ( d * Downloader ) commitSnapSyncData ( results [ ] * fetchResult , stateSync * stateSync ) error {
2018-02-05 19:40:32 +03:00
// 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
}
2018-02-05 19:40:32 +03:00
default :
}
2023-02-21 13:17:34 +03:00
// Retrieve the batch of results to import
2018-02-05 19:40:32 +03:00
first , last := results [ 0 ] . Header , results [ len ( results ) - 1 ] . Header
2021-11-26 14:26:03 +03:00
log . Debug ( "Inserting snap-sync blocks" , "items" , len ( results ) ,
2018-02-05 19:40:32 +03:00
"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 {
2023-01-25 17:32:25 +03:00
blocks [ i ] = types . NewBlockWithHeader ( result . Header ) . WithBody ( result . Transactions , result . Uncles ) . WithWithdrawals ( result . Withdrawals )
2018-02-05 19:40:32 +03:00
receipts [ i ] = result . Receipts
}
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
if index , err := d . blockchain . InsertReceiptChain ( blocks , receipts , d . ancientLimit ) ; err != nil {
2018-02-05 19:40:32 +03:00
log . Debug ( "Downloaded item processing failed" , "number" , results [ index ] . Header . Number , "hash" , results [ index ] . Header . Hash ( ) , "err" , err )
2020-05-29 12:12:43 +03:00
return fmt . Errorf ( "%w: %v" , errInvalidChain , 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
}
return nil
}
func ( d * Downloader ) commitPivotBlock ( result * fetchResult ) error {
2023-01-25 17:32:25 +03:00
block := types . NewBlockWithHeader ( result . Header ) . WithBody ( result . Transactions , result . Uncles ) . WithWithdrawals ( result . Withdrawals )
2021-11-26 14:26:03 +03:00
log . Debug ( "Committing snap sync pivot as new head" , "number" , block . Number ( ) , "hash" , block . Hash ( ) )
2019-05-13 15:28:01 +03:00
// Commit the pivot block as the new head, will require full sync from here on
all: integrate the freezer with fast sync
* all: freezer style syncing
core, eth, les, light: clean up freezer relative APIs
core, eth, les, trie, ethdb, light: clean a bit
core, eth, les, light: add unit tests
core, light: rewrite setHead function
core, eth: fix downloader unit tests
core: add receipt chain insertion test
core: use constant instead of hardcoding table name
core: fix rollback
core: fix setHead
core/rawdb: remove canonical block first and then iterate side chain
core/rawdb, ethdb: add hasAncient interface
eth/downloader: calculate ancient limit via cht first
core, eth, ethdb: lots of fixes
* eth/downloader: print ancient disable log only for fast sync
2019-04-25 17:59:48 +03:00
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
}
2021-11-26 14:26:03 +03:00
if err := d . blockchain . SnapSyncCommitHead ( 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
2015-06-12 13:35:29 +03:00
}
2023-04-03 22:48:10 +03:00
d . committed . Store ( true )
2018-02-05 19:40:32 +03:00
return nil
2015-06-12 13:35:29 +03:00
}
2020-12-14 12:27:15 +03:00
// DeliverSnapPacket is invoked from a peer's message handler when it transmits a
// data packet for the local node to consume.
func ( d * Downloader ) DeliverSnapPacket ( peer * snap . Peer , packet snap . Packet ) error {
switch packet := packet . ( type ) {
case * snap . AccountRangePacket :
hashes , accounts , err := packet . Unpack ( )
if err != nil {
return err
}
return d . SnapSyncer . OnAccounts ( peer , packet . ID , hashes , accounts , packet . Proof )
case * snap . StorageRangesPacket :
hashset , slotset := packet . Unpack ( )
return d . SnapSyncer . OnStorage ( peer , packet . ID , hashset , slotset , packet . Proof )
case * snap . ByteCodesPacket :
return d . SnapSyncer . OnByteCodes ( peer , packet . ID , packet . Codes )
case * snap . TrieNodesPacket :
return d . SnapSyncer . OnTrieNodes ( peer , packet . ID , packet . Nodes )
default :
return fmt . Errorf ( "unexpected snap packet type: %T" , packet )
}
2015-10-05 19:37:56 +03:00
}
2022-04-14 09:49:23 +03:00
// readHeaderRange returns a list of headers, using the given last header as the base,
// and going backwards towards genesis. This method assumes that the caller already has
// placed a reasonable cap on count.
func ( d * Downloader ) readHeaderRange ( last * types . Header , count int ) [ ] * types . Header {
var (
current = last
headers [ ] * types . Header
)
for {
parent := d . lightchain . GetHeaderByHash ( current . ParentHash )
if parent == nil {
break // The chain is not continuous, or the chain is exhausted
}
headers = append ( headers , parent )
if len ( headers ) >= count {
break
}
current = parent
}
return headers
}
2023-02-21 13:17:34 +03:00
// reportSnapSyncProgress calculates various status reports and provides it to the user.
func ( d * Downloader ) reportSnapSyncProgress ( force bool ) {
// Initialize the sync start time if it's the first time we're reporting
if d . syncStartTime . IsZero ( ) {
d . syncStartTime = time . Now ( ) . Add ( - time . Millisecond ) // -1ms offset to avoid division by zero
}
// Don't report all the events, just occasionally
if ! force && time . Since ( d . syncLogTime ) < 8 * time . Second {
return
}
// Don't report anything until we have a meaningful progress
var (
headerBytes , _ = d . stateDB . AncientSize ( rawdb . ChainFreezerHeaderTable )
bodyBytes , _ = d . stateDB . AncientSize ( rawdb . ChainFreezerBodiesTable )
receiptBytes , _ = d . stateDB . AncientSize ( rawdb . ChainFreezerReceiptTable )
)
syncedBytes := common . StorageSize ( headerBytes + bodyBytes + receiptBytes )
if syncedBytes == 0 {
return
}
var (
header = d . blockchain . CurrentHeader ( )
2023-03-02 09:29:15 +03:00
block = d . blockchain . CurrentSnapBlock ( )
2023-02-21 13:17:34 +03:00
)
2023-03-02 09:29:15 +03:00
syncedBlocks := block . Number . Uint64 ( ) - d . syncStartBlock
2023-02-21 13:17:34 +03:00
if syncedBlocks == 0 {
return
}
// Retrieve the current chain head and calculate the ETA
2023-02-23 14:22:41 +03:00
latest , _ , _ , err := d . skeleton . Bounds ( )
2023-02-21 13:17:34 +03:00
if err != nil {
// We're going to cheat for non-merged networks, but that's fine
latest = d . pivotHeader
}
if latest == nil {
// This should really never happen, but add some defensive code for now.
// TODO(karalabe): Remove it eventually if we don't see it blow.
log . Error ( "Nil latest block in sync progress report" )
return
}
var (
2023-03-02 09:29:15 +03:00
left = latest . Number . Uint64 ( ) - block . Number . Uint64 ( )
2023-02-21 13:17:34 +03:00
eta = time . Since ( d . syncStartTime ) / time . Duration ( syncedBlocks ) * time . Duration ( left )
2023-03-02 09:29:15 +03:00
progress = fmt . Sprintf ( "%.2f%%" , float64 ( block . Number . Uint64 ( ) ) * 100 / float64 ( latest . Number . Uint64 ( ) ) )
2023-02-21 13:17:34 +03:00
headers = fmt . Sprintf ( "%v@%v" , log . FormatLogfmtUint64 ( header . Number . Uint64 ( ) ) , common . StorageSize ( headerBytes ) . TerminalString ( ) )
2023-03-02 09:29:15 +03:00
bodies = fmt . Sprintf ( "%v@%v" , log . FormatLogfmtUint64 ( block . Number . Uint64 ( ) ) , common . StorageSize ( bodyBytes ) . TerminalString ( ) )
receipts = fmt . Sprintf ( "%v@%v" , log . FormatLogfmtUint64 ( block . Number . Uint64 ( ) ) , common . StorageSize ( receiptBytes ) . TerminalString ( ) )
2023-02-21 13:17:34 +03:00
)
log . Info ( "Syncing: chain download in progress" , "synced" , progress , "chain" , syncedBytes , "headers" , headers , "bodies" , bodies , "receipts" , receipts , "eta" , common . PrettyDuration ( eta ) )
d . syncLogTime = time . Now ( )
}