2015-07-06 19:54:22 -05:00
// Copyright 2015 The go-ethereum Authors
2015-07-22 11:48:40 -05:00
// This file is part of the go-ethereum library.
2015-07-06 19:54:22 -05:00
//
2015-07-23 11:35:11 -05:00
// The go-ethereum library is free software: you can redistribute it and/or modify
2015-07-06 19:54:22 -05:00
// 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 11:48:40 -05:00
// The go-ethereum library is distributed in the hope that it will be useful,
2015-07-06 19:54:22 -05:00
// but WITHOUT ANY WARRANTY; without even the implied warranty of
2015-07-22 11:48:40 -05:00
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
2015-07-06 19:54:22 -05:00
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
2015-07-22 11:48:40 -05:00
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
2015-07-06 19:54:22 -05:00
2015-06-16 03:58:32 -05:00
// Package downloader contains the manual full chain synchronisation.
2015-04-12 05:38:25 -05:00
package downloader
import (
2015-10-13 04:04:25 -05:00
"crypto/rand"
2015-04-17 18:10:32 -05:00
"errors"
2015-10-13 04:04:25 -05:00
"fmt"
2015-06-12 05:35:29 -05:00
"math"
2015-07-29 05:20:54 -05:00
"math/big"
2015-04-12 05:38:25 -05:00
"sync"
"sync/atomic"
"time"
2016-09-06 04:39:14 -05:00
ethereum "github.com/ethereum/go-ethereum"
2015-04-12 05:38:25 -05:00
"github.com/ethereum/go-ethereum/common"
"github.com/ethereum/go-ethereum/core/types"
2015-10-05 11:37:56 -05:00
"github.com/ethereum/go-ethereum/ethdb"
2015-05-14 17:43:00 -05:00
"github.com/ethereum/go-ethereum/event"
2017-02-22 06:10:07 -06:00
"github.com/ethereum/go-ethereum/log"
2016-05-13 05:12:13 -05:00
"github.com/ethereum/go-ethereum/params"
2015-10-05 11:37:56 -05:00
"github.com/rcrowley/go-metrics"
2015-04-12 05:38:25 -05:00
)
2015-06-08 06:06:36 -05:00
var (
2015-09-28 11:27:31 -05:00
MaxHashFetch = 512 // Amount of hashes to be fetched per retrieval request
MaxBlockFetch = 128 // Amount of blocks to be fetched per retrieval request
MaxHeaderFetch = 192 // Amount of block headers to be fetched per retrieval request
2016-02-25 10:36:42 -06:00
MaxSkeletonSize = 128 // Number of header fetches to need for a skeleton assembly
2015-09-28 11:27:31 -05:00
MaxBodyFetch = 128 // Amount of block bodies to be fetched per retrieval request
MaxReceiptFetch = 256 // Amount of transaction receipts to allow fetching per request
2015-10-05 11:37:56 -05:00
MaxStateFetch = 384 // Amount of node state values to allow fetching per request
2015-09-28 11:27:31 -05:00
2017-01-04 13:17:24 -06:00
MaxForkAncestry = 3 * params . EpochDuration // Maximum chain reorganisation
rttMinEstimate = 2 * time . Second // Minimum round-trip time to target for download requests
rttMaxEstimate = 20 * time . Second // Maximum rount-trip time to target for download requests
rttMinConfidence = 0.1 // Worse confidence factor in our estimated RTT value
ttlScaling = 3 // Constant scaling factor for RTT -> TTL conversion
ttlLimit = time . Minute // Maximum TTL allowance to prevent reaching crazy timeouts
2016-06-01 10:07:25 -05:00
qosTuningPeers = 5 // Number of peers to tune based on (best peers)
qosConfidenceCap = 10 // Number of peers above which not to modify RTT confidence
qosTuningImpact = 0.25 // Impact that a new tuning target has on the previous value
2015-09-28 11:27:31 -05:00
2016-02-25 10:36:42 -06:00
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
2016-05-17 03:12:57 -05:00
maxResultsProcess = 2048 // Number of content download results to import at once into the chain
2015-09-28 11:27:31 -05:00
2016-10-31 06:55:12 -05:00
fsHeaderCheckFrequency = 100 // Verification frequency of the downloaded headers during fast sync
fsHeaderSafetyNet = 2048 // Number of headers to discard in case a chain violation is detected
fsHeaderForceVerify = 24 // Number of headers to verify before and after the pivot to accept it
2016-10-31 10:26:44 -05:00
fsPivotInterval = 256 // Number of headers out of which to randomize the pivot point
fsMinFullBlocks = 64 // Number of blocks to retrieve fully even in fast sync
2016-10-31 06:55:12 -05:00
fsCriticalTrials = uint32 ( 32 ) // Number of times to retry in the cricical section before bailing
2015-05-15 05:14:46 -05:00
)
2015-04-19 06:30:34 -05:00
2015-05-15 05:14:46 -05:00
var (
2016-02-25 10:36:42 -06:00
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" )
errNoPeers = errors . New ( "no peers to keep download active" )
errTimeout = errors . New ( "timeout" )
errEmptyHeaderSet = errors . New ( "empty header set by peer" )
errPeersUnavailable = errors . New ( "no peers available or all tried for download" )
errInvalidAncestor = errors . New ( "retrieved ancestor is invalid" )
errInvalidChain = errors . New ( "retrieved hash chain is invalid" )
errInvalidBlock = errors . New ( "retrieved block is invalid" )
errInvalidBody = errors . New ( "retrieved block body is invalid" )
errInvalidReceipt = errors . New ( "retrieved receipt is invalid" )
errCancelBlockFetch = errors . New ( "block download canceled (requested)" )
errCancelHeaderFetch = errors . New ( "block header download canceled (requested)" )
errCancelBodyFetch = errors . New ( "block body download canceled (requested)" )
errCancelReceiptFetch = errors . New ( "receipt download canceled (requested)" )
errCancelStateFetch = errors . New ( "state data download canceled (requested)" )
errCancelHeaderProcessing = errors . New ( "header processing canceled (requested)" )
errCancelContentProcessing = errors . New ( "content processing canceled (requested)" )
errNoSyncActive = errors . New ( "no sync active" )
2016-07-21 04:36:38 -05:00
errTooOld = errors . New ( "peer doesn't speak recent enough protocol version (need version >= 62)" )
2015-04-17 18:10:32 -05:00
)
2015-04-12 05:38:25 -05:00
type Downloader struct {
2016-06-02 04:37:14 -05:00
mode SyncMode // Synchronisation mode defining the strategy used (per sync cycle)
mux * event . TypeMux // Event multiplexer to announce sync operation events
2015-05-14 17:43:00 -05:00
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 07:26:03 -05:00
queue * queue // Scheduler for selecting the hashes to download
peers * peerSet // Set of active peers from which download can proceed
stateDB ethdb . Database
2015-04-12 05:38:25 -05:00
2016-06-02 04:37:14 -05:00
fsPivotLock * types . Header // Pivot header on critical section entry (cannot change between retries)
2016-10-31 06:55:12 -05:00
fsPivotFails uint32 // Number of subsequent fast sync failures in the critical section
2016-06-02 04:37:14 -05:00
2016-06-01 10:07:25 -05:00
rttEstimate uint64 // Round trip time to target for download requests
rttConfidence uint64 // Confidence in the estimated RTT (unit: millionths to allow atomic ops)
2015-06-17 16:04:57 -05:00
2015-06-09 17:20:35 -05:00
// Statistics
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
syncStatsChainOrigin uint64 // Origin block number where syncing started at
syncStatsChainHeight uint64 // Highest block number known when syncing started
syncStatsState stateSyncStats
2015-10-05 11:37:56 -05:00
syncStatsLock sync . RWMutex // Lock protecting the sync stats fields
2015-06-09 17:20:35 -05:00
2015-04-13 09:38:32 -05:00
// Callbacks
2015-12-29 06:01:08 -06:00
hasHeader headerCheckFn // Checks if a header is present in the chain
hasBlockAndState blockAndStateCheckFn // Checks if a block and associated state is present in the chain
getHeader headerRetrievalFn // Retrieves a header from the chain
getBlock blockRetrievalFn // Retrieves a block from the chain
headHeader headHeaderRetrievalFn // Retrieves the head header from the chain
headBlock headBlockRetrievalFn // Retrieves the head block from the chain
headFastBlock headFastBlockRetrievalFn // Retrieves the head fast-sync block from the chain
commitHeadBlock headBlockCommitterFn // Commits a manually assembled block as the chain head
getTd tdRetrievalFn // Retrieves the TD of a block from the chain
insertHeaders headerChainInsertFn // Injects a batch of headers into the chain
insertBlocks blockChainInsertFn // Injects a batch of blocks into the chain
insertReceipts receiptChainInsertFn // Injects a batch of blocks and their receipts into the chain
rollback chainRollbackFn // Removes a batch of recently added chain links
dropPeer peerDropFn // Drops a peer for misbehaving
2015-04-12 05:38:25 -05:00
2015-04-13 09:38:32 -05:00
// Status
2015-06-11 10:13:13 -05:00
synchroniseMock func ( id string , hash common . Hash ) error // Replacement for synchronise during testing
synchronising int32
notified int32
2015-04-13 09:38:32 -05:00
// Channels
2016-02-25 10:36:42 -06:00
headerCh chan dataPack // [eth/62] Channel receiving inbound block headers
bodyCh chan dataPack // [eth/62] Channel receiving inbound block bodies
receiptCh chan dataPack // [eth/63] Channel receiving inbound receipts
bodyWakeCh chan bool // [eth/62] Channel to signal the block body fetcher of new tasks
receiptWakeCh chan bool // [eth/63] Channel to signal the receipt fetcher of new tasks
headerProcCh chan [ ] * types . Header // [eth/62] Channel to feed the header processor new tasks
2015-05-13 05:47:21 -05:00
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 07:26:03 -05:00
// for stateFetcher
stateSyncStart chan * stateSync
trackStateReq chan * stateReq
stateCh chan dataPack // [eth/63] Channel receiving inbound node state data
2016-07-26 05:07:12 -05:00
// Cancellation and termination
cancelPeer string // Identifier of the peer currently being used as the master (cancel on drop)
2015-05-13 05:47:21 -05:00
cancelCh chan struct { } // Channel to cancel mid-flight syncs
2016-07-26 05:07:12 -05:00
cancelLock sync . RWMutex // Lock to protect the cancel channel and peer in delivers
2015-08-14 13:25:41 -05:00
2016-06-01 10:07:25 -05:00
quitCh chan struct { } // Quit channel to signal termination
quitLock sync . RWMutex // Lock to prevent double closes
2015-08-14 13:25:41 -05:00
// Testing hooks
2015-09-28 11:27:31 -05:00
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)
2015-05-26 06:00:21 -05:00
}
2015-06-11 07:56:08 -05:00
// New creates a new downloader to fetch hashes and blocks from remote peers.
2016-01-13 12:35:48 -06:00
func New ( mode SyncMode , stateDb ethdb . Database , mux * event . TypeMux , hasHeader headerCheckFn , hasBlockAndState blockAndStateCheckFn ,
2015-12-29 06:01:08 -06:00
getHeader headerRetrievalFn , getBlock blockRetrievalFn , headHeader headHeaderRetrievalFn , headBlock headBlockRetrievalFn ,
headFastBlock headFastBlockRetrievalFn , commitHeadBlock headBlockCommitterFn , getTd tdRetrievalFn , insertHeaders headerChainInsertFn ,
insertBlocks blockChainInsertFn , insertReceipts receiptChainInsertFn , rollback chainRollbackFn , dropPeer peerDropFn ) * Downloader {
2015-09-28 11:27:31 -05:00
2016-06-01 10:07:25 -05:00
dl := & Downloader {
2016-01-13 12:35:48 -06:00
mode : mode ,
2015-12-29 06:01:08 -06:00
mux : mux ,
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 07:26:03 -05:00
queue : newQueue ( ) ,
2015-12-29 06:01:08 -06:00
peers : newPeerSet ( ) ,
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 07:26:03 -05:00
stateDB : stateDb ,
2016-06-01 10:07:25 -05:00
rttEstimate : uint64 ( rttMaxEstimate ) ,
rttConfidence : uint64 ( 1000000 ) ,
2015-12-29 06:01:08 -06:00
hasHeader : hasHeader ,
hasBlockAndState : hasBlockAndState ,
getHeader : getHeader ,
getBlock : getBlock ,
headHeader : headHeader ,
headBlock : headBlock ,
headFastBlock : headFastBlock ,
commitHeadBlock : commitHeadBlock ,
getTd : getTd ,
insertHeaders : insertHeaders ,
insertBlocks : insertBlocks ,
insertReceipts : insertReceipts ,
rollback : rollback ,
dropPeer : dropPeer ,
headerCh : make ( chan dataPack , 1 ) ,
bodyCh : make ( chan dataPack , 1 ) ,
receiptCh : make ( chan dataPack , 1 ) ,
bodyWakeCh : make ( chan bool , 1 ) ,
receiptWakeCh : make ( chan bool , 1 ) ,
2016-02-25 10:36:42 -06:00
headerProcCh : make ( chan [ ] * types . Header , 1 ) ,
2016-06-01 10:07:25 -05:00
quitCh : make ( chan struct { } ) ,
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 07:26:03 -05:00
// for stateFetcher
stateSyncStart : make ( chan * stateSync ) ,
trackStateReq : make ( chan * stateReq ) ,
stateCh : make ( chan dataPack ) ,
2015-04-12 05:38:25 -05:00
}
2016-06-01 10:07:25 -05:00
go dl . qosTuner ( )
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
go dl . stateFetcher ( )
2016-06-01 10:07:25 -05:00
return dl
2015-04-12 05:38:25 -05:00
}
2015-10-13 04:04:25 -05:00
// 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 03:56:15 -06:00
//
2016-03-15 13:27:49 -05:00
// In addition, during the state download phase of fast synchronisation the number
2016-02-10 03:56:15 -06:00
// of processed and the total number of known states are also returned. Otherwise
// these are zero.
2016-09-06 04:39:14 -05:00
func ( d * Downloader ) Progress ( ) ethereum . SyncProgress {
2016-02-10 03:56:15 -06:00
// Lock the current stats and return the progress
2015-09-09 11:02:54 -05:00
d . syncStatsLock . RLock ( )
defer d . syncStatsLock . RUnlock ( )
2015-06-09 17:20:35 -05:00
2015-10-13 04:04:25 -05:00
current := uint64 ( 0 )
switch d . mode {
case FullSync :
current = d . headBlock ( ) . NumberU64 ( )
case FastSync :
current = d . headFastBlock ( ) . NumberU64 ( )
case LightSync :
current = d . headHeader ( ) . Number . Uint64 ( )
}
2016-09-06 04:39:14 -05:00
return ethereum . SyncProgress {
StartingBlock : d . syncStatsChainOrigin ,
CurrentBlock : current ,
HighestBlock : d . syncStatsChainHeight ,
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
PulledStates : d . syncStatsState . processed ,
KnownStates : d . syncStatsState . processed + d . syncStatsState . pending ,
2016-09-06 04:39:14 -05:00
}
2015-04-19 14:45:58 -05:00
}
2015-06-12 05:35:29 -05:00
// Synchronising returns whether the downloader is currently retrieving blocks.
2015-05-14 17:43:00 -05:00
func ( d * Downloader ) Synchronising ( ) bool {
2015-11-13 10:08:15 -06:00
return atomic . LoadInt32 ( & d . synchronising ) > 0
2015-05-14 17:43:00 -05:00
}
2015-05-11 06:26:20 -05:00
// RegisterPeer injects a new download peer into the set of block source to be
// used for fetching hashes and blocks from.
2016-07-25 07:14:14 -05:00
func ( d * Downloader ) RegisterPeer ( id string , version int , currentHead currentHeadRetrievalFn ,
2015-10-05 11:37:56 -05:00
getRelHeaders relativeHeaderFetcherFn , getAbsHeaders absoluteHeaderFetcherFn , getBlockBodies blockBodyFetcherFn ,
getReceipts receiptFetcherFn , getNodeData stateFetcherFn ) error {
2015-08-14 13:25:41 -05:00
2017-02-24 10:23:03 -06:00
logger := log . New ( "peer" , id )
logger . Trace ( "Registering sync peer" )
if err := d . peers . Register ( newPeer ( id , version , currentHead , getRelHeaders , getAbsHeaders , getBlockBodies , getReceipts , getNodeData , logger ) ) ; err != nil {
2017-02-27 05:17:58 -06:00
logger . Error ( "Failed to register sync peer" , "err" , err )
2015-05-11 06:26:20 -05:00
return err
}
2016-06-01 10:07:25 -05:00
d . qosReduceConfidence ( )
2015-04-12 05:38:25 -05:00
return nil
}
2015-05-11 06:26:20 -05:00
// UnregisterPeer remove a peer from the known list, preventing any action from
2015-09-28 11:27:31 -05:00
// the specified peer. An effort is also made to return any pending fetches into
// the queue.
2015-05-11 06:26:20 -05:00
func ( d * Downloader ) UnregisterPeer ( id string ) error {
2016-07-26 05:07:12 -05:00
// Unregister the peer from the active peer set and revoke any fetch tasks
2017-02-24 10:23:03 -06:00
logger := log . New ( "peer" , id )
logger . Trace ( "Unregistering sync peer" )
2015-05-11 06:26:20 -05:00
if err := d . peers . Unregister ( id ) ; err != nil {
2017-02-27 05:17:58 -06:00
logger . Error ( "Failed to unregister sync peer" , "err" , err )
2015-05-11 06:26:20 -05:00
return err
}
2015-09-28 11:27:31 -05:00
d . queue . Revoke ( id )
2016-07-26 05:07:12 -05:00
// If this peer was the master peer, abort sync immediately
d . cancelLock . RLock ( )
master := id == d . cancelPeer
d . cancelLock . RUnlock ( )
if master {
2017-03-21 19:37:24 -05:00
d . Cancel ( )
2016-07-26 05:07:12 -05:00
}
2015-05-11 06:26:20 -05:00
return nil
2015-04-12 05:38:25 -05:00
}
2015-06-11 07:56:08 -05:00
// Synchronise tries to sync up our local block chain with a remote peer, both
// adding various sanity checks as well as wrapping it with various log entries.
2015-10-28 09:41:01 -05:00
func ( d * Downloader ) Synchronise ( id string , head common . Hash , td * big . Int , mode SyncMode ) error {
err := d . synchronise ( id , head , td , mode )
switch err {
2015-06-11 07:56:08 -05:00
case nil :
case errBusy :
2016-07-21 04:36:38 -05:00
case errTimeout , errBadPeer , errStallingPeer ,
errEmptyHeaderSet , errPeersUnavailable , errTooOld ,
errInvalidAncestor , errInvalidChain :
2017-02-27 05:17:58 -06:00
log . Warn ( "Synchronisation failed, dropping peer" , "peer" , id , "err" , err )
2015-06-11 07:56:08 -05:00
d . dropPeer ( id )
default :
2017-02-27 05:17:58 -06:00
log . Warn ( "Synchronisation failed, retrying" , "err" , err )
2015-06-11 07:56:08 -05:00
}
2015-10-28 09:41:01 -05:00
return err
2015-06-11 07:56:08 -05:00
}
// synchronise will select the peer and use it for synchronising. If an empty string is given
2015-05-06 07:32:53 -05:00
// it will use the best peer possible and synchronize if it's TD is higher than our own. If any of the
2015-04-24 07:40:32 -05:00
// checks fail an error will be returned. This method is synchronous
2015-10-13 04:04:25 -05:00
func ( d * Downloader ) synchronise ( id string , hash common . Hash , td * big . Int , mode SyncMode ) error {
2016-03-15 13:27:49 -05:00
// Mock out the synchronisation if testing
2015-06-11 10:13:13 -05:00
if d . synchroniseMock != nil {
return d . synchroniseMock ( id , hash )
}
2015-05-07 13:07:20 -05:00
// Make sure only one goroutine is ever allowed past this point at once
2015-05-08 07:22:48 -05:00
if ! atomic . CompareAndSwapInt32 ( & d . synchronising , 0 , 1 ) {
2015-06-11 07:56:08 -05:00
return errBusy
2015-04-19 06:30:34 -05:00
}
2015-05-08 07:22:48 -05:00
defer atomic . StoreInt32 ( & d . synchronising , 0 )
2015-04-24 07:40:32 -05:00
2015-05-13 08:03:05 -05:00
// Post a user notification of the sync (only once per session)
if atomic . CompareAndSwapInt32 ( & d . notified , 0 , 1 ) {
2017-02-24 10:23:03 -06:00
log . Info ( "Block synchronisation started" )
2015-05-13 08:03:05 -05:00
}
2015-09-28 11:27:31 -05:00
// Reset the queue, peer set and wake channels to clean any internal leftover state
2015-05-08 09:21:11 -05:00
d . queue . Reset ( )
2015-05-11 06:26:20 -05:00
d . peers . Reset ( )
2015-05-08 09:21:11 -05:00
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 07:26:03 -05:00
for _ , ch := range [ ] chan bool { d . bodyWakeCh , d . receiptWakeCh } {
2015-09-28 11:27:31 -05:00
select {
case <- ch :
default :
}
2015-09-23 04:39:17 -05:00
}
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 07:26:03 -05:00
for _ , ch := range [ ] chan dataPack { d . headerCh , d . bodyCh , d . receiptCh } {
2016-06-02 04:37:14 -05:00
for empty := false ; ! empty ; {
select {
case <- ch :
default :
empty = true
}
}
}
2016-02-25 10:36:42 -06:00
for empty := false ; ! empty ; {
select {
case <- d . headerProcCh :
default :
empty = true
}
}
2016-07-26 05:07:12 -05:00
// Create cancel channel for aborting mid-flight and mark the master peer
2015-06-17 16:04:57 -05:00
d . cancelLock . Lock ( )
d . cancelCh = make ( chan struct { } )
2016-07-26 05:07:12 -05:00
d . cancelPeer = id
2015-06-17 16:04:57 -05:00
d . cancelLock . Unlock ( )
2017-03-21 19:37:24 -05:00
defer d . Cancel ( ) // No matter what, we can't leave the cancel channel open
2016-05-30 04:01:50 -05:00
2015-10-13 04:04:25 -05:00
// Set the requested sync mode, unless it's forbidden
d . mode = mode
2016-10-31 06:55:12 -05:00
if d . mode == FastSync && atomic . LoadUint32 ( & d . fsPivotFails ) >= fsCriticalTrials {
2015-10-13 04:04:25 -05:00
d . mode = FullSync
}
2015-05-07 13:07:20 -05:00
// Retrieve the origin peer and initiate the downloading process
2015-05-11 06:26:20 -05:00
p := d . peers . Peer ( id )
2015-04-24 07:40:32 -05:00
if p == nil {
2015-04-24 08:37:32 -05:00
return errUnknownPeer
2015-04-13 09:38:32 -05:00
}
2015-07-29 05:20:54 -05:00
return d . syncWithPeer ( p , hash , td )
2015-04-30 17:23:51 -05:00
}
2015-05-11 06:26:20 -05:00
// syncWithPeer starts a block synchronization based on the hash chain from the
// specified peer and head hash.
2015-07-29 05:20:54 -05:00
func ( d * Downloader ) syncWithPeer ( p * peer , hash common . Hash , td * big . Int ) ( err error ) {
2015-05-16 05:29:19 -05:00
d . mux . Post ( StartEvent { } )
2015-04-30 17:23:51 -05:00
defer func ( ) {
// reset on error
if err != nil {
2015-05-14 17:43:00 -05:00
d . mux . Post ( FailedEvent { err } )
} else {
d . mux . Post ( DoneEvent { } )
2015-04-30 17:23:51 -05:00
}
} ( )
2016-07-21 04:36:38 -05:00
if p . version < 62 {
return errTooOld
}
2015-04-24 07:40:32 -05:00
2017-02-27 09:06:40 -06:00
log . Debug ( "Synchronising with the network" , "peer" , p . id , "eth" , p . version , "head" , hash , "td" , td , "mode" , d . mode )
2015-09-30 11:23:31 -05:00
defer func ( start time . Time ) {
2017-02-24 10:23:03 -06:00
log . Debug ( "Synchronisation terminated" , "elapsed" , time . Since ( start ) )
2015-09-30 11:23:31 -05:00
} ( time . Now ( ) )
2015-08-14 13:25:41 -05:00
2016-07-21 04:36:38 -05:00
// Look up the sync boundaries: the common ancestor and the target block
latest , err := d . fetchHeight ( p )
if err != nil {
return err
}
height := latest . Number . Uint64 ( )
2015-09-09 11:02:54 -05:00
2016-07-21 04:36:38 -05:00
origin , err := d . findAncestor ( p , height )
if err != nil {
return err
}
d . syncStatsLock . Lock ( )
if d . syncStatsChainHeight <= origin || d . syncStatsChainOrigin > origin {
d . syncStatsChainOrigin = origin
}
d . syncStatsChainHeight = height
d . syncStatsLock . Unlock ( )
2016-05-27 06:26:00 -05:00
2016-07-21 04:36:38 -05:00
// Initiate the sync using a concurrent header and content retrieval algorithm
pivot := uint64 ( 0 )
switch d . mode {
case LightSync :
pivot = height
case FastSync :
// Calculate the new fast/slow sync pivot point
if d . fsPivotLock == nil {
pivotOffset , err := rand . Int ( rand . Reader , big . NewInt ( int64 ( fsPivotInterval ) ) )
if err != nil {
panic ( fmt . Sprintf ( "Failed to access crypto random source: %v" , err ) )
2015-10-13 04:04:25 -05:00
}
2016-07-21 04:36:38 -05:00
if height > uint64 ( fsMinFullBlocks ) + pivotOffset . Uint64 ( ) {
pivot = height - uint64 ( fsMinFullBlocks ) - pivotOffset . Uint64 ( )
2015-10-13 04:04:25 -05:00
}
2016-07-21 04:36:38 -05:00
} else {
// Pivot point locked in, use this and do not pick a new one!
pivot = d . fsPivotLock . Number . Uint64 ( )
2015-09-28 11:27:31 -05:00
}
2016-07-21 04:36:38 -05:00
// If the point is below the origin, move origin back to ensure state download
if pivot < origin {
if pivot > 0 {
origin = pivot - 1
} else {
origin = 0
}
2015-09-09 11:02:54 -05:00
}
2017-02-24 10:23:03 -06:00
log . Debug ( "Fast syncing until pivot block" , "pivot" , pivot )
2016-07-21 04:36:38 -05:00
}
d . queue . Prepare ( origin + 1 , d . mode , pivot , latest )
if d . syncInitHook != nil {
d . syncInitHook ( origin , height )
2015-04-12 05:38:25 -05:00
}
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 07:26:03 -05:00
fetchers := [ ] func ( ) error {
func ( ) error { return d . fetchHeaders ( p , origin + 1 ) } , // Headers are always retrieved
func ( ) error { return d . fetchBodies ( origin + 1 ) } , // Bodies are retrieved during normal and fast sync
func ( ) error { return d . fetchReceipts ( origin + 1 ) } , // Receipts are retrieved during fast sync
func ( ) error { return d . processHeaders ( origin + 1 , td ) } ,
}
if d . mode == FastSync {
fetchers = append ( fetchers , func ( ) error { return d . processFastSyncContent ( latest ) } )
} else if d . mode == FullSync {
fetchers = append ( fetchers , d . processFullSyncContent )
}
err = d . spawnSync ( fetchers )
if err != nil && d . mode == FastSync && d . fsPivotLock != nil {
// If sync failed in the critical section, bump the fail counter.
atomic . AddUint32 ( & d . fsPivotFails , 1 )
}
return err
2015-11-13 10:08:15 -06:00
}
// 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 07:26:03 -05:00
func ( d * Downloader ) spawnSync ( fetchers [ ] func ( ) error ) error {
2015-11-13 10:08:15 -06:00
var wg sync . WaitGroup
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 07:26:03 -05:00
errc := make ( chan error , len ( fetchers ) )
wg . Add ( len ( fetchers ) )
2015-11-13 10:08:15 -06:00
for _ , fn := range fetchers {
fn := fn
go func ( ) { defer wg . Done ( ) ; errc <- fn ( ) } ( )
}
// Wait for the first error, then terminate the others.
var err error
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
for i := 0 ; i < len ( fetchers ) ; i ++ {
if i == len ( fetchers ) - 1 {
2015-11-13 10:08:15 -06:00
// 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 ( )
}
if err = <- errc ; err != nil {
break
}
}
d . queue . Close ( )
2017-03-21 19:37:24 -05:00
d . Cancel ( )
2015-11-13 10:08:15 -06:00
wg . Wait ( )
return err
2015-04-12 05:38:25 -05:00
}
2017-03-21 19:37:24 -05:00
// Cancel cancels all of the operations and resets the queue. It returns true
2015-05-09 17:34:07 -05:00
// if the cancel operation was completed.
2017-03-21 19:37:24 -05:00
func ( d * Downloader ) Cancel ( ) {
2015-05-13 06:01:08 -05:00
// Close the current cancel channel
2015-05-15 11:43:42 -05:00
d . cancelLock . Lock ( )
2015-06-12 05:35:29 -05:00
if d . cancelCh != nil {
select {
case <- d . cancelCh :
// Channel was already closed
default :
close ( d . cancelCh )
}
2015-05-15 11:43:42 -05:00
}
d . cancelLock . Unlock ( )
2015-05-09 17:34:07 -05:00
}
2015-06-17 16:04:57 -05:00
// Terminate interrupts the downloader, canceling all pending operations.
2015-11-13 10:08:15 -06:00
// The downloader cannot be reused after calling Terminate.
2015-06-17 16:04:57 -05:00
func ( d * Downloader ) Terminate ( ) {
2016-06-01 10:07:25 -05:00
// Close the termination channel (make sure double close is allowed)
d . quitLock . Lock ( )
select {
case <- d . quitCh :
default :
close ( d . quitCh )
}
d . quitLock . Unlock ( )
// Cancel any pending download requests
2017-03-21 19:37:24 -05:00
d . Cancel ( )
2015-06-17 16:04:57 -05:00
}
2015-09-09 11:02:54 -05:00
// fetchHeight retrieves the head header of the remote peer to aid in estimating
// the total time a pending synchronisation would take.
2016-05-27 06:26:00 -05:00
func ( d * Downloader ) fetchHeight ( p * peer ) ( * types . Header , error ) {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Retrieving remote chain height" )
2015-09-09 11:02:54 -05:00
// Request the advertised remote head block and wait for the response
2016-07-25 07:14:14 -05:00
head , _ := p . currentHead ( )
go p . getRelHeaders ( head , 1 , 0 , false )
2015-09-09 11:02:54 -05:00
2017-02-24 10:23:03 -06:00
ttl := d . requestTTL ( )
timeout := time . After ( ttl )
2015-09-09 11:02:54 -05:00
for {
select {
case <- d . cancelCh :
2016-05-27 06:26:00 -05:00
return nil , errCancelBlockFetch
2015-09-09 11:02:54 -05:00
2015-10-05 11:37:56 -05:00
case packet := <- d . headerCh :
2015-09-09 11:02:54 -05:00
// Discard anything not from the origin peer
2015-10-05 11:37:56 -05:00
if packet . PeerId ( ) != p . id {
2017-02-24 10:23:03 -06:00
log . Debug ( "Received headers from incorrect peer" , "peer" , packet . PeerId ( ) )
2015-09-09 11:02:54 -05:00
break
}
// Make sure the peer actually gave something valid
2015-10-05 11:37:56 -05:00
headers := packet . ( * headerPack ) . headers
2015-09-09 11:02:54 -05:00
if len ( headers ) != 1 {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Multiple headers for single request" , "headers" , len ( headers ) )
2016-05-27 06:26:00 -05:00
return nil , errBadPeer
2015-09-09 11:02:54 -05:00
}
2017-02-24 10:23:03 -06:00
head := headers [ 0 ]
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Remote head header identified" , "number" , head . Number , "hash" , head . Hash ( ) )
2017-02-24 10:23:03 -06:00
return head , nil
2015-09-09 11:02:54 -05:00
2015-11-17 15:55:32 -06:00
case <- timeout :
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Waiting for head header timed out" , "elapsed" , ttl )
2016-05-27 06:26:00 -05:00
return nil , errTimeout
2015-11-17 15:55:32 -06:00
2015-09-09 11:02:54 -05:00
case <- d . bodyCh :
2015-11-17 15:55:32 -06:00
case <- d . receiptCh :
// Out of bounds delivery, ignore
2015-09-09 11:02:54 -05:00
}
}
}
2015-09-28 11:27:31 -05:00
// findAncestor tries to locate the common ancestor link of the local chain and
2015-08-14 13:25:41 -05:00
// a remote peers blockchain. In the general case when our node was in sync and
2015-09-28 11:27:31 -05:00
// on the correct chain, checking the top N links should already get us a match.
2016-03-15 13:55:39 -05:00
// In the rare scenario when we ended up on a long reorganisation (i.e. none of
2015-09-28 11:27:31 -05:00
// the head links match), we do a binary search to find the common ancestor.
2016-05-13 05:12:13 -05:00
func ( d * Downloader ) findAncestor ( p * peer , height uint64 ) ( uint64 , error ) {
// Figure out the valid ancestor range to prevent rewrite attacks
floor , ceil := int64 ( - 1 ) , d . headHeader ( ) . Number . Uint64 ( )
2017-02-24 10:23:03 -06:00
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Looking for common ancestor" , "local" , ceil , "remote" , height )
2015-09-28 11:27:31 -05:00
if d . mode == FullSync {
2016-05-13 05:12:13 -05:00
ceil = d . headBlock ( ) . NumberU64 ( )
2015-09-30 11:23:31 -05:00
} else if d . mode == FastSync {
2016-05-13 05:12:13 -05:00
ceil = d . headFastBlock ( ) . NumberU64 ( )
}
if ceil >= MaxForkAncestry {
floor = int64 ( ceil - MaxForkAncestry )
}
// Request the topmost blocks to short circuit binary ancestor lookup
head := ceil
if head > height {
head = height
2015-09-28 11:27:31 -05:00
}
2016-07-26 03:15:38 -05:00
from := int64 ( head ) - int64 ( MaxHeaderFetch )
2015-08-14 13:25:41 -05:00
if from < 0 {
from = 0
}
2016-07-26 03:15:38 -05:00
// Span out with 15 block gaps into the future to catch bad head reports
limit := 2 * MaxHeaderFetch / 16
count := 1 + int ( ( int64 ( ceil ) - from ) / 16 )
if count > limit {
count = limit
}
go p . getAbsHeaders ( uint64 ( from ) , count , 15 , false )
2015-08-14 13:25:41 -05:00
// Wait for the remote response to the head fetch
number , hash := uint64 ( 0 ) , common . Hash { }
2017-02-24 10:23:03 -06:00
ttl := d . requestTTL ( )
timeout := time . After ( ttl )
2015-08-14 13:25:41 -05:00
for finished := false ; ! finished ; {
select {
case <- d . cancelCh :
2016-07-21 04:36:38 -05:00
return 0 , errCancelHeaderFetch
2015-08-14 13:25:41 -05:00
2015-10-05 11:37:56 -05:00
case packet := <- d . headerCh :
2015-08-14 13:25:41 -05:00
// Discard anything not from the origin peer
2015-10-05 11:37:56 -05:00
if packet . PeerId ( ) != p . id {
2017-02-24 10:23:03 -06:00
log . Debug ( "Received headers from incorrect peer" , "peer" , packet . PeerId ( ) )
2015-08-14 13:25:41 -05:00
break
}
// Make sure the peer actually gave something valid
2015-10-05 11:37:56 -05:00
headers := packet . ( * headerPack ) . headers
2015-08-14 13:25:41 -05:00
if len ( headers ) == 0 {
2017-03-02 07:06:16 -06:00
p . log . Warn ( "Empty head header set" )
2015-08-14 13:25:41 -05:00
return 0 , errEmptyHeaderSet
}
2016-02-16 02:36:26 -06:00
// Make sure the peer's reply conforms to the request
for i := 0 ; i < len ( headers ) ; i ++ {
2016-07-26 03:15:38 -05:00
if number := headers [ i ] . Number . Int64 ( ) ; number != from + int64 ( i ) * 16 {
2017-03-02 07:06:16 -06:00
p . log . Warn ( "Head headers broke chain ordering" , "index" , i , "requested" , from + int64 ( i ) * 16 , "received" , number )
2016-02-16 02:36:26 -06:00
return 0 , errInvalidChain
}
}
2015-08-14 13:25:41 -05:00
// Check if a common ancestor was found
finished = true
for i := len ( headers ) - 1 ; i >= 0 ; i -- {
2016-02-16 02:36:26 -06:00
// Skip any headers that underflow/overflow our requested set
2016-07-26 03:15:38 -05:00
if headers [ i ] . Number . Int64 ( ) < from || headers [ i ] . Number . Uint64 ( ) > ceil {
2016-02-16 02:36:26 -06:00
continue
}
// Otherwise check if we already know the header or not
2016-04-19 04:27:37 -05:00
if ( d . mode == FullSync && d . hasBlockAndState ( headers [ i ] . Hash ( ) ) ) || ( d . mode != FullSync && d . hasHeader ( headers [ i ] . Hash ( ) ) ) {
2015-08-14 13:25:41 -05:00
number , hash = headers [ i ] . Number . Uint64 ( ) , headers [ i ] . Hash ( )
2016-07-26 03:15:38 -05:00
// If every header is known, even future ones, the peer straight out lied about its head
if number > height && i == limit - 1 {
2017-03-02 07:06:16 -06:00
p . log . Warn ( "Lied about chain head" , "reported" , height , "found" , number )
2016-07-26 03:15:38 -05:00
return 0 , errStallingPeer
}
2015-08-14 13:25:41 -05:00
break
}
}
2015-11-17 15:55:32 -06:00
case <- timeout :
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Waiting for head header timed out" , "elapsed" , ttl )
2015-11-17 15:55:32 -06:00
return 0 , errTimeout
2015-08-14 13:25:41 -05:00
case <- d . bodyCh :
2015-11-17 15:55:32 -06:00
case <- d . receiptCh :
// Out of bounds delivery, ignore
2015-08-14 13:25:41 -05:00
}
}
// If the head fetch already found an ancestor, return
if ! common . EmptyHash ( hash ) {
2016-05-13 05:12:13 -05:00
if int64 ( number ) <= floor {
2017-03-02 07:06:16 -06:00
p . log . Warn ( "Ancestor below allowance" , "number" , number , "hash" , hash , "allowance" , floor )
2016-05-13 05:12:13 -05:00
return 0 , errInvalidAncestor
}
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Found common ancestor" , "number" , number , "hash" , hash )
2015-08-14 13:25:41 -05:00
return number , nil
}
// Ancestor not found, we need to binary search over our chain
start , end := uint64 ( 0 ) , head
2016-05-13 05:12:13 -05:00
if floor > 0 {
start = uint64 ( floor )
}
2015-08-14 13:25:41 -05:00
for start + 1 < end {
// Split our chain interval in two, and request the hash to cross check
check := ( start + end ) / 2
2017-02-24 10:23:03 -06:00
ttl := d . requestTTL ( )
timeout := time . After ( ttl )
2015-08-14 13:25:41 -05:00
go p . getAbsHeaders ( uint64 ( check ) , 1 , 0 , false )
// Wait until a reply arrives to this request
for arrived := false ; ! arrived ; {
select {
case <- d . cancelCh :
2016-07-21 04:36:38 -05:00
return 0 , errCancelHeaderFetch
2015-08-14 13:25:41 -05:00
2015-10-05 11:37:56 -05:00
case packer := <- d . headerCh :
2015-08-14 13:25:41 -05:00
// Discard anything not from the origin peer
2015-10-05 11:37:56 -05:00
if packer . PeerId ( ) != p . id {
2017-02-24 10:23:03 -06:00
log . Debug ( "Received headers from incorrect peer" , "peer" , packer . PeerId ( ) )
2015-08-14 13:25:41 -05:00
break
}
// Make sure the peer actually gave something valid
2015-10-05 11:37:56 -05:00
headers := packer . ( * headerPack ) . headers
2015-08-14 13:25:41 -05:00
if len ( headers ) != 1 {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Multiple headers for single request" , "headers" , len ( headers ) )
2015-08-14 13:25:41 -05:00
return 0 , errBadPeer
}
arrived = true
// Modify the search interval based on the response
2015-12-29 06:01:08 -06:00
if ( d . mode == FullSync && ! d . hasBlockAndState ( headers [ 0 ] . Hash ( ) ) ) || ( d . mode != FullSync && ! d . hasHeader ( headers [ 0 ] . Hash ( ) ) ) {
2015-08-14 13:25:41 -05:00
end = check
break
}
2015-09-28 11:27:31 -05:00
header := d . getHeader ( headers [ 0 ] . Hash ( ) ) // Independent of sync mode, header surely exists
if header . Number . Uint64 ( ) != check {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Received non requested header" , "number" , header . Number , "hash" , header . Hash ( ) , "request" , check )
2015-08-14 13:25:41 -05:00
return 0 , errBadPeer
}
start = check
2015-11-17 15:55:32 -06:00
case <- timeout :
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Waiting for search header timed out" , "elapsed" , ttl )
2015-11-17 15:55:32 -06:00
return 0 , errTimeout
2015-08-14 13:25:41 -05:00
case <- d . bodyCh :
2015-11-17 15:55:32 -06:00
case <- d . receiptCh :
// Out of bounds delivery, ignore
2015-08-14 13:25:41 -05:00
}
}
}
2016-05-13 05:12:13 -05:00
// Ensure valid ancestry and return
if int64 ( start ) <= floor {
2017-03-02 07:06:16 -06:00
p . log . Warn ( "Ancestor below allowance" , "number" , start , "hash" , hash , "allowance" , floor )
2016-05-13 05:12:13 -05:00
return 0 , errInvalidAncestor
}
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Found common ancestor" , "number" , start , "hash" , hash )
2015-08-14 13:25:41 -05:00
return start , nil
}
2016-02-25 10:36:42 -06:00
// 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 03:12:57 -05:00
// other peers are only accepted if they map cleanly to the skeleton. If no one
2016-02-25 10:36:42 -06:00
// can fill in the skeleton - not even the origin peer - it's assumed invalid and
// the origin is dropped.
func ( d * Downloader ) fetchHeaders ( p * peer , from uint64 ) error {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Directing header downloads" , "origin" , from )
defer p . log . Debug ( "Header download terminated" )
2015-08-14 13:25:41 -05:00
2016-02-25 10:36:42 -06:00
// Create a timeout timer, and the associated header fetcher
skeleton := true // Skeleton assembly phase or finishing up
request := time . Now ( ) // time of the last skeleton fetch request
2015-08-14 13:25:41 -05:00
timeout := time . NewTimer ( 0 ) // timer to dump a non-responsive active peer
<- timeout . C // timeout channel should be initially empty
defer timeout . Stop ( )
2017-02-24 10:23:03 -06:00
var ttl time . Duration
2015-08-14 13:25:41 -05:00
getHeaders := func ( from uint64 ) {
2016-05-17 03:12:57 -05:00
request = time . Now ( )
2017-02-24 10:23:03 -06:00
ttl = d . requestTTL ( )
timeout . Reset ( ttl )
2016-05-17 03:12:57 -05:00
2016-02-25 10:36:42 -06:00
if skeleton {
2017-03-02 07:06:16 -06:00
p . log . Trace ( "Fetching skeleton headers" , "count" , MaxHeaderFetch , "from" , from )
2016-02-25 10:36:42 -06:00
go p . getAbsHeaders ( from + uint64 ( MaxHeaderFetch ) - 1 , MaxSkeletonSize , MaxHeaderFetch - 1 , false )
} else {
2017-03-02 07:06:16 -06:00
p . log . Trace ( "Fetching full headers" , "count" , MaxHeaderFetch , "from" , from )
2016-02-25 10:36:42 -06:00
go p . getAbsHeaders ( from , MaxHeaderFetch , 0 , false )
}
2015-08-14 13:25:41 -05:00
}
2016-02-25 10:36:42 -06:00
// Start pulling the header chain skeleton until all is done
2015-08-14 13:25:41 -05:00
getHeaders ( from )
for {
select {
case <- d . cancelCh :
return errCancelHeaderFetch
2015-10-05 11:37:56 -05:00
case packet := <- d . headerCh :
2016-02-25 10:36:42 -06:00
// Make sure the active peer is giving us the skeleton headers
2015-10-05 11:37:56 -05:00
if packet . PeerId ( ) != p . id {
2017-02-24 10:23:03 -06:00
log . Debug ( "Received skeleton from incorrect peer" , "peer" , packet . PeerId ( ) )
2015-08-14 13:25:41 -05:00
break
}
2015-08-25 05:57:49 -05:00
headerReqTimer . UpdateSince ( request )
2015-08-14 13:25:41 -05:00
timeout . Stop ( )
2016-02-25 10:36:42 -06:00
// If the skeleton's finished, pull any remaining head headers directly from the origin
if packet . Items ( ) == 0 && skeleton {
skeleton = false
getHeaders ( from )
continue
}
2015-09-28 11:27:31 -05:00
// If no more headers are inbound, notify the content fetchers and return
2015-10-05 11:37:56 -05:00
if packet . Items ( ) == 0 {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "No more headers available" )
2016-06-02 04:37:14 -05:00
select {
case d . headerProcCh <- nil :
return nil
case <- d . cancelCh :
return errCancelHeaderFetch
}
2015-08-14 13:25:41 -05:00
}
2015-10-05 11:37:56 -05:00
headers := packet . ( * headerPack ) . headers
2015-08-14 13:25:41 -05:00
2016-02-25 10:36:42 -06:00
// If we received a skeleton batch, resolve internals concurrently
if skeleton {
2016-04-19 04:27:37 -05:00
filled , proced , err := d . fillHeaderSkeleton ( from , headers )
2016-02-25 10:36:42 -06:00
if err != nil {
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Skeleton chain invalid" , "err" , err )
2015-09-30 11:23:31 -05:00
return errInvalidChain
}
2016-04-19 04:27:37 -05:00
headers = filled [ proced : ]
from += uint64 ( proced )
2015-08-14 13:25:41 -05:00
}
2016-02-25 10:36:42 -06:00
// Insert all the new headers and fetch the next batch
2016-04-19 04:27:37 -05:00
if len ( headers ) > 0 {
2017-03-02 07:06:16 -06:00
p . log . Trace ( "Scheduling new headers" , "count" , len ( headers ) , "from" , from )
2016-04-19 04:27:37 -05:00
select {
case d . headerProcCh <- headers :
case <- d . cancelCh :
return errCancelHeaderFetch
}
from += uint64 ( len ( headers ) )
2016-02-29 06:22:28 -06:00
}
2015-08-14 13:25:41 -05:00
getHeaders ( from )
case <- timeout . C :
// Header retrieval timed out, consider the peer bad and drop
2017-03-02 07:06:16 -06:00
p . log . Debug ( "Header request timed out" , "elapsed" , ttl )
2015-08-25 05:57:49 -05:00
headerTimeoutMeter . Mark ( 1 )
2015-08-14 13:25:41 -05:00
d . dropPeer ( p . id )
// Finish the sync gracefully instead of dumping the gathered data though
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
for _ , ch := range [ ] chan bool { d . bodyWakeCh , d . receiptWakeCh } {
2015-09-28 11:27:31 -05:00
select {
case ch <- false :
case <- d . cancelCh :
}
2015-08-14 13:25:41 -05:00
}
2016-02-25 10:36:42 -06:00
select {
case d . headerProcCh <- nil :
case <- d . cancelCh :
}
return errBadPeer
2015-08-14 13:25:41 -05:00
}
}
}
2016-02-25 10:36:42 -06:00
// fillHeaderSkeleton concurrently retrieves headers from all our available peers
// and maps them to the provided skeleton header chain.
2016-04-19 04:27:37 -05:00
//
// 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.
//
// The method returs the entire filled skeleton and also the number of headers
// already forwarded for processing.
func ( d * Downloader ) fillHeaderSkeleton ( from uint64 , skeleton [ ] * types . Header ) ( [ ] * types . Header , int , error ) {
2017-02-24 10:23:03 -06:00
log . Debug ( "Filling up skeleton" , "from" , from )
2016-02-25 10:36:42 -06:00
d . queue . ScheduleSkeleton ( from , skeleton )
var (
deliver = func ( packet dataPack ) ( int , error ) {
pack := packet . ( * headerPack )
2016-04-19 04:27:37 -05:00
return d . queue . DeliverHeaders ( pack . peerId , pack . headers , d . headerProcCh )
2016-02-25 10:36:42 -06:00
}
2016-06-01 10:07:25 -05:00
expire = func ( ) map [ string ] int { return d . queue . ExpireHeaders ( d . requestTTL ( ) ) }
2016-02-25 10:36:42 -06:00
throttle = func ( ) bool { return false }
reserve = func ( p * peer , count int ) ( * fetchRequest , bool , error ) {
return d . queue . ReserveHeaders ( p , count ) , false , nil
}
fetch = func ( p * peer , req * fetchRequest ) error { return p . FetchHeaders ( req . From , MaxHeaderFetch ) }
2016-06-01 10:07:25 -05:00
capacity = func ( p * peer ) int { return p . HeaderCapacity ( d . requestRTT ( ) ) }
2016-02-25 10:36:42 -06:00
setIdle = func ( p * peer , accepted int ) { p . SetHeadersIdle ( accepted ) }
)
err := d . fetchParts ( errCancelHeaderFetch , d . headerCh , deliver , d . queue . headerContCh , expire ,
d . queue . PendingHeaders , d . queue . InFlightHeaders , throttle , reserve ,
2017-02-24 10:23:03 -06:00
nil , fetch , d . queue . CancelHeaders , capacity , d . peers . HeaderIdlePeers , setIdle , "headers" )
2016-02-25 10:36:42 -06:00
2017-02-27 05:17:58 -06:00
log . Debug ( "Skeleton fill terminated" , "err" , err )
2016-04-19 04:27:37 -05:00
filled , proced := d . queue . RetrieveHeaders ( )
return filled , proced , err
2016-02-25 10:36:42 -06:00
}
2015-08-14 13:25:41 -05:00
// 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.
func ( d * Downloader ) fetchBodies ( from uint64 ) error {
2017-02-24 10:23:03 -06:00
log . Debug ( "Downloading block bodies" , "origin" , from )
2015-08-14 13:25:41 -05:00
2015-09-28 11:27:31 -05:00
var (
2015-10-29 11:37:26 -05:00
deliver = func ( packet dataPack ) ( int , error ) {
2015-09-28 11:27:31 -05:00
pack := packet . ( * bodyPack )
2015-10-05 11:37:56 -05:00
return d . queue . DeliverBodies ( pack . peerId , pack . transactions , pack . uncles )
2015-09-28 11:27:31 -05:00
}
2016-06-01 10:07:25 -05:00
expire = func ( ) map [ string ] int { return d . queue . ExpireBodies ( d . requestTTL ( ) ) }
2015-09-28 11:27:31 -05:00
fetch = func ( p * peer , req * fetchRequest ) error { return p . FetchBodies ( req ) }
2016-06-01 10:07:25 -05:00
capacity = func ( p * peer ) int { return p . BlockCapacity ( d . requestRTT ( ) ) }
2015-10-29 11:37:26 -05:00
setIdle = func ( p * peer , accepted int ) { p . SetBodiesIdle ( accepted ) }
2015-09-28 11:27:31 -05:00
)
2015-10-05 11:37:56 -05:00
err := d . fetchParts ( errCancelBodyFetch , d . bodyCh , deliver , d . bodyWakeCh , expire ,
2015-10-13 04:04:25 -05:00
d . queue . PendingBlocks , d . queue . InFlightBlocks , d . queue . ShouldThrottleBlocks , d . queue . ReserveBodies ,
2017-02-24 10:23:03 -06:00
d . bodyFetchHook , fetch , d . queue . CancelBodies , capacity , d . peers . BodyIdlePeers , setIdle , "bodies" )
2015-09-28 11:27:31 -05:00
2017-02-27 05:17:58 -06:00
log . Debug ( "Block body download terminated" , "err" , err )
2015-09-28 11:27:31 -05:00
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.
func ( d * Downloader ) fetchReceipts ( from uint64 ) error {
2017-02-24 10:23:03 -06:00
log . Debug ( "Downloading transaction receipts" , "origin" , from )
2015-09-28 11:27:31 -05:00
var (
2015-10-29 11:37:26 -05:00
deliver = func ( packet dataPack ) ( int , error ) {
2015-09-28 11:27:31 -05:00
pack := packet . ( * receiptPack )
return d . queue . DeliverReceipts ( pack . peerId , pack . receipts )
}
2016-06-01 10:07:25 -05:00
expire = func ( ) map [ string ] int { return d . queue . ExpireReceipts ( d . requestTTL ( ) ) }
2015-09-28 11:27:31 -05:00
fetch = func ( p * peer , req * fetchRequest ) error { return p . FetchReceipts ( req ) }
2016-06-01 10:07:25 -05:00
capacity = func ( p * peer ) int { return p . ReceiptCapacity ( d . requestRTT ( ) ) }
2015-10-29 11:37:26 -05:00
setIdle = func ( p * peer , accepted int ) { p . SetReceiptsIdle ( accepted ) }
2015-09-28 11:27:31 -05:00
)
2015-10-05 11:37:56 -05:00
err := d . fetchParts ( errCancelReceiptFetch , d . receiptCh , deliver , d . receiptWakeCh , expire ,
2015-10-13 04:04:25 -05:00
d . queue . PendingReceipts , d . queue . InFlightReceipts , d . queue . ShouldThrottleReceipts , d . queue . ReserveReceipts ,
2017-02-24 10:23:03 -06:00
d . receiptFetchHook , fetch , d . queue . CancelReceipts , capacity , d . peers . ReceiptIdlePeers , setIdle , "receipts" )
2015-09-28 11:27:31 -05:00
2017-02-27 05:17:58 -06:00
log . Debug ( "Transaction receipt download terminated" , "err" , err )
2015-09-28 11:27:31 -05:00
return err
}
// fetchParts iteratively downloads scheduled block parts, taking any available
// peers, reserving a chunk of fetch requests for each, waiting for delivery and
// also periodically checking for timeouts.
2016-05-17 03:12:57 -05:00
//
// As the scheduling/timeout logic mostly is the same for all downloaded data
// types, this method is used by each for data gathering and is instrumented with
// various callbacks to handle the slight differences between processing them.
//
// The instrumentation parameters:
// - errCancel: error type to return if the fetch operation is cancelled (mostly makes logging nicer)
// - deliveryCh: channel from which to retrieve downloaded data packets (merged from all concurrent peers)
// - deliver: processing callback to deliver data packets into type specific download queues (usually within `queue`)
// - wakeCh: notification channel for waking the fetcher when new tasks are available (or sync completed)
// - expire: task callback method to abort requests that took too long and return the faulty peers (traffic shaping)
// - pending: task callback for the number of requests still needing download (detect completion/non-completability)
// - inFlight: task callback for the number of in-progress requests (wait for all active downloads to finish)
// - throttle: task callback to check if the processing queue is full and activate throttling (bound memory use)
// - reserve: task callback to reserve new download tasks to a particular peer (also signals partial completions)
// - fetchHook: tester callback to notify of new tasks being initiated (allows testing the scheduling logic)
// - fetch: network callback to actually send a particular download request to a physical remote peer
// - cancel: task callback to abort an in-flight download request and allow rescheduling it (in case of lost peer)
2017-01-06 11:44:35 -06:00
// - capacity: network callback to retrieve the estimated type-specific bandwidth capacity of a peer (traffic shaping)
2016-05-17 03:12:57 -05:00
// - idle: network callback to retrieve the currently (type specific) idle peers that can be assigned tasks
// - setIdle: network callback to set a peer back to idle and update its estimated capacity (traffic shaping)
// - kind: textual label of the type being downloaded to display in log mesages
2015-10-29 11:37:26 -05:00
func ( d * Downloader ) fetchParts ( errCancel error , deliveryCh chan dataPack , deliver func ( dataPack ) ( int , error ) , wakeCh chan bool ,
expire func ( ) map [ string ] int , pending func ( ) int , inFlight func ( ) bool , throttle func ( ) bool , reserve func ( * peer , int ) ( * fetchRequest , bool , error ) ,
2015-10-07 04:14:30 -05:00
fetchHook func ( [ ] * types . Header ) , fetch func ( * peer , * fetchRequest ) error , cancel func ( * fetchRequest ) , capacity func ( * peer ) int ,
2015-10-29 11:37:26 -05:00
idle func ( ) ( [ ] * peer , int ) , setIdle func ( * peer , int ) , kind string ) error {
2015-09-28 11:27:31 -05:00
2015-10-13 04:04:25 -05:00
// Create a ticker to detect expired retrieval tasks
2015-08-14 13:25:41 -05:00
ticker := time . NewTicker ( 100 * time . Millisecond )
defer ticker . Stop ( )
update := make ( chan struct { } , 1 )
2015-09-28 11:27:31 -05:00
// Prepare the queue and fetch block parts until the block header fetcher's done
2015-08-14 13:25:41 -05:00
finished := false
for {
select {
case <- d . cancelCh :
2015-09-28 11:27:31 -05:00
return errCancel
2015-08-14 13:25:41 -05:00
2015-09-28 11:27:31 -05:00
case packet := <- deliveryCh :
2015-08-14 13:25:41 -05:00
// If the peer was previously banned and failed to deliver it's pack
// in a reasonable time frame, ignore it's message.
2015-09-28 11:27:31 -05:00
if peer := d . peers . Peer ( packet . PeerId ( ) ) ; peer != nil {
2015-10-29 11:37:26 -05:00
// Deliver the received chunk of data and check chain validity
accepted , err := deliver ( packet )
if err == errInvalidChain {
2015-08-14 13:25:41 -05:00
return err
2015-10-29 11:37:26 -05:00
}
// Unless a peer delivered something completely else than requested (usually
// caused by a timed out request which came through in the end), set it to
// idle. If the delivery's stale, the peer should have already been idled.
if err != errStaleDelivery {
setIdle ( peer , accepted )
}
// Issue a log to the user to see what's going on
switch {
case err == nil && packet . Items ( ) == 0 :
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Requested data not delivered" , "type" , kind )
2015-10-29 11:37:26 -05:00
case err == nil :
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Delivered new batch of data" , "type" , kind , "count" , packet . Stats ( ) )
2015-08-14 13:25:41 -05:00
default :
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Failed to deliver retrieved data" , "type" , kind , "err" , err )
2015-08-14 13:25:41 -05:00
}
}
// Blocks assembled, try to update the progress
select {
case update <- struct { } { } :
default :
}
2015-09-28 11:27:31 -05:00
case cont := <- wakeCh :
2015-08-14 13:25:41 -05:00
// The header fetcher sent a continuation flag, check if it's done
if ! cont {
finished = true
}
// Headers arrive, try to update the progress
select {
case update <- struct { } { } :
default :
}
case <- ticker . C :
// Sanity check update the progress
select {
case update <- struct { } { } :
default :
}
case <- update :
// Short circuit if we lost all our peers
if d . peers . Len ( ) == 0 {
return errNoPeers
}
2015-09-28 11:27:31 -05:00
// Check for fetch request timeouts and demote the responsible peers
2015-10-29 11:37:26 -05:00
for pid , fails := range expire ( ) {
2015-08-14 13:25:41 -05:00
if peer := d . peers . Peer ( pid ) ; peer != nil {
2016-07-22 09:55:46 -05:00
// If a lot of retrieval elements expired, we might have overestimated the remote peer or perhaps
// ourselves. Only reset to minimal throughput but don't drop just yet. If even the minimal times
// out that sync wise we need to get rid of the peer.
//
// The reason the minimum threshold is 2 is because the downloader tries to estimate the bandwidth
// and latency of a peer separately, which requires pushing the measures capacity a bit and seeing
// how response times reacts, to it always requests one more than the minimum (i.e. min 2).
if fails > 2 {
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Data delivery timed out" , "type" , kind )
2015-10-29 11:37:26 -05:00
setIdle ( peer , 0 )
} else {
2017-03-02 07:06:16 -06:00
peer . log . Debug ( "Stalling delivery, dropping" , "type" , kind )
2015-10-29 11:37:26 -05:00
d . dropPeer ( pid )
}
2015-08-14 13:25:41 -05:00
}
}
2015-09-28 11:27:31 -05:00
// If there's nothing more to fetch, wait or terminate
if pending ( ) == 0 {
2015-10-07 04:14:30 -05:00
if ! inFlight ( ) && finished {
2017-02-24 10:23:03 -06:00
log . Debug ( "Data fetching completed" , "type" , kind )
2015-08-14 13:25:41 -05:00
return nil
}
break
}
// Send a download request to all idle peers, until throttled
2015-10-07 04:14:30 -05:00
progressed , throttled , running := false , false , inFlight ( )
2015-09-28 11:27:31 -05:00
idles , total := idle ( )
for _ , peer := range idles {
2015-08-14 13:25:41 -05:00
// Short circuit if throttling activated
2015-09-28 11:27:31 -05:00
if throttle ( ) {
2015-08-14 13:25:41 -05:00
throttled = true
break
}
2015-09-28 11:27:31 -05:00
// Reserve a chunk of fetches for a peer. A nil can mean either that
// no more headers are available, or that the peer is known not to
2015-08-14 13:25:41 -05:00
// have them.
2015-09-28 11:27:31 -05:00
request , progress , err := reserve ( peer , capacity ( peer ) )
2015-08-14 13:25:41 -05:00
if err != nil {
return err
}
2015-09-28 11:27:31 -05:00
if progress {
progressed = true
2015-08-14 13:25:41 -05:00
}
if request == nil {
continue
}
2017-02-24 10:23:03 -06:00
if request . From > 0 {
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Requesting new batch of data" , "type" , kind , "from" , request . From )
2017-02-24 10:23:03 -06:00
} else if len ( request . Headers ) > 0 {
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Requesting new batch of data" , "type" , kind , "count" , len ( request . Headers ) , "from" , request . Headers [ 0 ] . Number )
2017-02-24 10:23:03 -06:00
} else {
2017-03-02 07:06:16 -06:00
peer . log . Trace ( "Requesting new batch of data" , "type" , kind , "count" , len ( request . Hashes ) )
2017-02-24 10:23:03 -06:00
}
2015-08-14 13:25:41 -05:00
// Fetch the chunk and make sure any errors return the hashes to the queue
2015-09-28 11:27:31 -05:00
if fetchHook != nil {
fetchHook ( request . Headers )
2015-08-14 13:25:41 -05:00
}
2015-09-28 11:27:31 -05:00
if err := fetch ( peer , request ) ; err != nil {
2015-10-13 04:04:25 -05:00
// Although we could try and make an attempt to fix this, this error really
// means that we've double allocated a fetch task to a peer. If that is the
// case, the internal state of the downloader and the queue is very wrong so
// better hard crash and note the error instead of silently accumulating into
// a much bigger issue.
2017-02-24 10:23:03 -06:00
panic ( fmt . Sprintf ( "%v: %s fetch assignment failed" , peer , kind ) )
2015-08-14 13:25:41 -05:00
}
2015-10-07 04:14:30 -05:00
running = true
2015-08-14 13:25:41 -05:00
}
// Make sure that we have peers available for fetching. If all peers have been tried
// and all failed throw an error
2015-10-07 04:14:30 -05:00
if ! progressed && ! throttled && ! running && len ( idles ) == total && pending ( ) > 0 {
2015-08-14 13:25:41 -05:00
return errPeersUnavailable
}
}
}
}
2016-02-25 10:36:42 -06:00
// 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.
func ( d * Downloader ) processHeaders ( origin uint64 , td * big . Int ) error {
// Calculate the pivoting point for switching from fast to slow sync
pivot := d . queue . FastSyncPivot ( )
// Keep a count of uncertain headers to roll back
rollback := [ ] * types . Header { }
defer func ( ) {
if len ( rollback ) > 0 {
// Flatten the headers and roll them back
hashes := make ( [ ] common . Hash , len ( rollback ) )
for i , header := range rollback {
hashes [ i ] = header . Hash ( )
}
2016-01-13 12:35:48 -06:00
lastHeader , lastFastBlock , lastBlock := d . headHeader ( ) . Number , common . Big0 , common . Big0
if d . headFastBlock != nil {
lastFastBlock = d . headFastBlock ( ) . Number ( )
}
if d . headBlock != nil {
lastBlock = d . headBlock ( ) . Number ( )
}
2016-02-25 10:36:42 -06:00
d . rollback ( hashes )
2016-01-13 12:35:48 -06:00
curFastBlock , curBlock := common . Big0 , common . Big0
if d . headFastBlock != nil {
curFastBlock = d . headFastBlock ( ) . Number ( )
}
if d . headBlock != nil {
curBlock = d . headBlock ( ) . Number ( )
}
2017-02-24 10:23:03 -06:00
log . Warn ( "Rolled back headers" , "count" , len ( hashes ) ,
"header" , fmt . Sprintf ( "%d->%d" , lastHeader , d . headHeader ( ) . Number ) ,
"fast" , fmt . Sprintf ( "%d->%d" , lastFastBlock , curFastBlock ) ,
"block" , fmt . Sprintf ( "%d->%d" , lastBlock , curBlock ) )
2016-02-25 10:36:42 -06:00
2016-06-02 04:37:14 -05:00
// If we're already past the pivot point, this could be an attack, thread carefully
2016-02-25 10:36:42 -06:00
if rollback [ len ( rollback ) - 1 ] . Number . Uint64 ( ) > pivot {
2016-06-02 04:37:14 -05:00
// If we didn't ever fail, lock in te pivot header (must! not! change!)
2016-10-31 06:55:12 -05:00
if atomic . LoadUint32 ( & d . fsPivotFails ) == 0 {
2016-06-02 04:37:14 -05:00
for _ , header := range rollback {
if header . Number . Uint64 ( ) == pivot {
2017-03-21 19:37:24 -05:00
log . Warn ( "Fast-sync pivot locked in" , "number" , pivot , "hash" , header . Hash ( ) )
2016-06-02 04:37:14 -05:00
d . fsPivotLock = header
}
}
}
2016-02-25 10:36:42 -06:00
}
}
} ( )
// Wait for batches of headers to process
gotHeaders := false
for {
select {
case <- d . cancelCh :
return errCancelHeaderProcessing
case headers := <- d . headerProcCh :
// Terminate header processing if we synced up
if len ( headers ) == 0 {
// Notify everyone that headers are fully processed
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
for _ , ch := range [ ] chan bool { d . bodyWakeCh , d . receiptWakeCh } {
2016-02-25 10:36:42 -06:00
select {
case ch <- false :
case <- d . cancelCh :
}
}
// If no headers were retrieved at all, the peer violated it's TD promise that it had a
// better chain compared to ours. The only exception is if it's promised blocks were
// already imported by other means (e.g. fecher):
//
// 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
2016-01-13 12:35:48 -06:00
if d . mode != LightSync {
if ! gotHeaders && td . Cmp ( d . getTd ( d . headBlock ( ) . Hash ( ) ) ) > 0 {
return errStallingPeer
}
2016-02-25 10:36:42 -06:00
}
// If fast or light syncing, ensure promised headers are indeed delivered. This is
// needed to detect scenarios where an attacker feeds a bad pivot and then bails out
// of delivering the post-pivot blocks that would flag the invalid content.
//
// This check cannot be executed "as is" for full imports, since blocks may still be
// queued for processing when the header download completes. However, as long as the
// peer gave us something useful, we're already happy/progressed (above check).
if d . mode == FastSync || d . mode == LightSync {
if td . Cmp ( d . getTd ( d . headHeader ( ) . Hash ( ) ) ) > 0 {
return errStallingPeer
}
}
// Disable any rollback and return
rollback = nil
return nil
}
// Otherwise split the chunk of headers into batches and process them
gotHeaders = true
for len ( headers ) > 0 {
// Terminate if something failed in between processing chunks
select {
case <- d . cancelCh :
return errCancelHeaderProcessing
default :
}
// Select the next chunk of headers to import
limit := maxHeadersProcess
if limit > len ( headers ) {
limit = len ( headers )
}
chunk := headers [ : limit ]
// In case of header only syncing, validate the chunk immediately
if d . mode == FastSync || d . mode == LightSync {
// Collect the yet unknown headers to mark them as uncertain
unknown := make ( [ ] * types . Header , 0 , len ( headers ) )
for _ , header := range chunk {
if ! d . hasHeader ( header . Hash ( ) ) {
unknown = append ( unknown , header )
}
}
// If we're importing pure headers, verify based on their recentness
frequency := fsHeaderCheckFrequency
if chunk [ len ( chunk ) - 1 ] . Number . Uint64 ( ) + uint64 ( fsHeaderForceVerify ) > pivot {
frequency = 1
}
if n , err := d . insertHeaders ( chunk , frequency ) ; err != nil {
// If some headers were inserted, add them too to the rollback list
if n > 0 {
rollback = append ( rollback , chunk [ : n ] ... )
}
2017-02-27 09:06:40 -06:00
log . Debug ( "Invalid header encountered" , "number" , chunk [ n ] . Number , "hash" , chunk [ n ] . Hash ( ) , "err" , err )
2016-02-25 10:36:42 -06:00
return errInvalidChain
}
// All verifications passed, store newly found uncertain headers
rollback = append ( rollback , unknown ... )
if len ( rollback ) > fsHeaderSafetyNet {
rollback = append ( rollback [ : 0 ] , rollback [ len ( rollback ) - fsHeaderSafetyNet : ] ... )
}
}
2016-06-02 04:37:14 -05:00
// If we're fast syncing and just pulled in the pivot, make sure it's the one locked in
if d . mode == FastSync && d . fsPivotLock != nil && chunk [ 0 ] . Number . Uint64 ( ) <= pivot && chunk [ len ( chunk ) - 1 ] . Number . Uint64 ( ) >= pivot {
if pivot := chunk [ int ( pivot - chunk [ 0 ] . Number . Uint64 ( ) ) ] ; pivot . Hash ( ) != d . fsPivotLock . Hash ( ) {
2017-02-27 09:06:40 -06:00
log . Warn ( "Pivot doesn't match locked in one" , "remoteNumber" , pivot . Number , "remoteHash" , pivot . Hash ( ) , "localNumber" , d . fsPivotLock . Number , "localHash" , d . fsPivotLock . Hash ( ) )
2016-06-02 04:37:14 -05:00
return errInvalidChain
}
}
2016-02-25 10:36:42 -06:00
// Unless we're doing light chains, schedule the headers for associated content retrieval
if d . mode == FullSync || d . mode == FastSync {
// If we've reached the allowed number of pending headers, stall a bit
for d . queue . PendingBlocks ( ) >= maxQueuedHeaders || d . queue . PendingReceipts ( ) >= maxQueuedHeaders {
select {
case <- d . cancelCh :
return errCancelHeaderProcessing
case <- time . After ( time . Second ) :
}
}
// Otherwise insert the headers for content retrieval
inserts := d . queue . Schedule ( chunk , origin )
if len ( inserts ) != len ( chunk ) {
2017-02-24 10:23:03 -06:00
log . Debug ( "Stale headers" )
2016-02-25 10:36:42 -06:00
return errBadPeer
}
}
headers = headers [ limit : ]
origin += uint64 ( limit )
}
// Signal the content downloaders of the availablility of new tasks
eth/downloader: separate state sync from queue (#14460)
* eth/downloader: separate state sync from queue
Scheduling of state node downloads hogged the downloader queue lock when
new requests were scheduled. This caused timeouts for other requests.
With this change, state sync is fully independent of all other downloads
and doesn't involve the queue at all.
State sync is started and checked on in processContent. This is slightly
awkward because processContent doesn't have a select loop. Instead, the
queue is closed by an auxiliary goroutine when state sync fails. We
tried several alternatives to this but settled on the current approach
because it's the least amount of change overall.
Handling of the pivot block has changed slightly: the queue previously
prevented import of pivot block receipts before the state of the pivot
block was available. In this commit, the receipt will be imported before
the state. This causes an annoyance where the pivot block is committed
as fast block head even when state downloads fail. Stay tuned for more
updates in this area ;)
* eth/downloader: remove cancelTimeout channel
* eth/downloader: retry state requests on timeout
* eth/downloader: improve comment
* eth/downloader: mark peers idle when state sync is done
* eth/downloader: move pivot block splitting to processContent
This change also ensures that pivot block receipts aren't imported
before the pivot block itself.
* eth/downloader: limit state node retries
* eth/downloader: improve state node error handling and retry check
* eth/downloader: remove maxStateNodeRetries
It fails the sync too much.
* eth/downloader: remove last use of cancelCh in statesync.go
Fixes TestDeliverHeadersHang*Fast and (hopefully)
the weird cancellation behaviour at the end of fast sync.
* eth/downloader: fix leak in runStateSync
* eth/downloader: don't run processFullSyncContent in LightSync mode
* eth/downloader: improve comments
* eth/downloader: fix vet, megacheck
* eth/downloader: remove unrequested tasks anyway
* eth/downloader, trie: various polishes around duplicate items
This commit explicitly tracks duplicate and unexpected state
delieveries done against a trie Sync structure, also adding
there to import info logs.
The commit moves the db batch used to commit trie changes one
level deeper so its flushed after every node insertion. This
is needed to avoid a lot of duplicate retrievals caused by
inconsistencies between Sync internals and database. A better
approach is to track not-yet-written states in trie.Sync and
flush on commit, but I'm focuing on correctness first now.
The commit fixes a regression around pivot block fail count.
The counter previously was reset to 1 if and only if a sync
cycle progressed (inserted at least 1 entry to the database).
The current code reset it already if a node was delivered,
which is not stong enough, because unless it ends up written
to disk, an attacker can just loop and attack ad infinitum.
The commit also fixes a regression around state deliveries
and timeouts. The old downloader tracked if a delivery is
stale (none of the deliveries were requestedt), in which
case it didn't mark the node idle and did not send further
requests, since it signals a past timeout. The current code
did mark it idle even on stale deliveries, which eventually
caused two requests to be in flight at the same time, making
the deliveries always stale and mass duplicating retrievals
between multiple peers.
* eth/downloader: fix state request leak
This commit fixes the hang seen sometimes while doing the state
sync. The cause of the hang was a rare combination of events:
request state data from peer, peer drops and reconnects almost
immediately. This caused a new download task to be assigned to
the peer, overwriting the old one still waiting for a timeout,
which in turned leaked the requests out, never to be retried.
The fix is to ensure that a task assignment moves any pending
one back into the retry queue.
The commit also fixes a regression with peer dropping due to
stalls. The current code considered a peer stalling if they
timed out delivering 1 item. However, the downloader never
requests only one, the minimum is 2 (attempt to fine tune
estimated latency/bandwidth). The fix is simply to drop if
a timeout is detected at 2 items.
Apart from the above bugfixes, the commit contains some code
polishes I made while debugging the hang.
* core, eth, trie: support batched trie sync db writes
* trie: rename SyncMemCache to syncMemBatch
2017-06-22 07:26:03 -05:00
for _ , ch := range [ ] chan bool { d . bodyWakeCh , d . receiptWakeCh } {
2016-02-25 10:36:42 -06:00
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 07:26:03 -05:00
// processFullSyncContent takes fetch results from the queue and imports them into the chain.
func ( d * Downloader ) processFullSyncContent ( ) error {
2015-06-12 05:35:29 -05:00
for {
2015-11-13 10:08:15 -06:00
results := d . queue . WaitResults ( )
2015-09-28 11:27:31 -05:00
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 07:26:03 -05:00
return nil
2015-06-12 05:35:29 -05:00
}
2015-08-14 13:25:41 -05:00
if d . chainInsertHook != nil {
2015-09-28 11:27:31 -05:00
d . chainInsertHook ( results )
2015-08-14 13:25:41 -05:00
}
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 07:26:03 -05:00
if err := d . importBlockResults ( results ) ; err != nil {
return err
}
}
}
func ( d * Downloader ) importBlockResults ( results [ ] * fetchResult ) error {
for len ( results ) != 0 {
// Check for any termination requests. This makes clean shutdown faster.
select {
case <- d . quitCh :
return errCancelContentProcessing
default :
}
// Retrieve the a batch of results to import
items := int ( math . Min ( float64 ( len ( results ) ) , float64 ( maxResultsProcess ) ) )
first , last := results [ 0 ] . Header , results [ items - 1 ] . Header
2017-02-24 10:23:03 -06:00
log . Debug ( "Inserting downloaded chain" , "items" , len ( results ) ,
2017-02-27 09:06:40 -06:00
"firstnum" , first . Number , "firsthash" , first . Hash ( ) ,
"lastnum" , last . Number , "lasthash" , last . Hash ( ) ,
)
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 07:26:03 -05:00
blocks := make ( [ ] * types . Block , items )
for i , result := range results [ : items ] {
blocks [ i ] = types . NewBlockWithHeader ( result . Header ) . WithBody ( result . Transactions , result . Uncles )
}
if index , err := d . insertBlocks ( blocks ) ; err != nil {
log . Debug ( "Downloaded item processing failed" , "number" , results [ index ] . Header . Number , "hash" , results [ index ] . Header . Hash ( ) , "err" , err )
return errInvalidChain
}
// Shift the results to the next batch
results = results [ items : ]
}
return nil
}
// processFastSyncContent takes fetch results from the queue and writes them to the
// database. It also controls the synchronisation of state nodes of the pivot block.
func ( d * Downloader ) processFastSyncContent ( latest * types . Header ) error {
// Start syncing state of the reported head block.
// This should get us most of the state of the pivot block.
stateSync := d . syncState ( latest . Root )
defer stateSync . Cancel ( )
go func ( ) {
if err := stateSync . Wait ( ) ; err != nil {
d . queue . Close ( ) // wake up WaitResults
}
} ( )
pivot := d . queue . FastSyncPivot ( )
for {
results := d . queue . WaitResults ( )
if len ( results ) == 0 {
return stateSync . Cancel ( )
}
if d . chainInsertHook != nil {
d . chainInsertHook ( results )
}
P , beforeP , afterP := splitAroundPivot ( pivot , results )
if err := d . commitFastSyncData ( beforeP , stateSync ) ; err != nil {
return err
}
if P != nil {
stateSync . Cancel ( )
if err := d . commitPivotBlock ( P ) ; err != nil {
return err
2015-06-12 05:35:29 -05:00
}
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 07:26:03 -05:00
}
if err := d . importBlockResults ( afterP ) ; err != nil {
return err
}
}
}
func splitAroundPivot ( pivot uint64 , results [ ] * fetchResult ) ( p * fetchResult , before , after [ ] * fetchResult ) {
for _ , result := range results {
num := result . Header . Number . Uint64 ( )
switch {
case num < pivot :
before = append ( before , result )
case num == pivot :
p = result
default :
after = append ( after , result )
}
}
return p , before , after
}
func ( d * Downloader ) commitFastSyncData ( results [ ] * fetchResult , stateSync * stateSync ) error {
for len ( results ) != 0 {
// Check for any termination requests.
select {
case <- d . quitCh :
return errCancelContentProcessing
case <- stateSync . done :
if err := stateSync . Wait ( ) ; err != nil {
return err
2015-06-12 05:35:29 -05:00
}
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 07:26:03 -05:00
default :
2015-06-12 05:35:29 -05:00
}
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 07:26:03 -05:00
// Retrieve the a batch of results to import
items := int ( math . Min ( float64 ( len ( results ) ) , float64 ( maxResultsProcess ) ) )
first , last := results [ 0 ] . Header , results [ items - 1 ] . Header
log . Debug ( "Inserting fast-sync blocks" , "items" , len ( results ) ,
"firstnum" , first . Number , "firsthash" , first . Hash ( ) ,
"lastnumn" , last . Number , "lasthash" , last . Hash ( ) ,
)
blocks := make ( [ ] * types . Block , items )
receipts := make ( [ ] types . Receipts , items )
for i , result := range results [ : items ] {
blocks [ i ] = types . NewBlockWithHeader ( result . Header ) . WithBody ( result . Transactions , result . Uncles )
receipts [ i ] = result . Receipts
}
if index , err := d . insertReceipts ( blocks , receipts ) ; err != nil {
log . Debug ( "Downloaded item processing failed" , "number" , results [ index ] . Header . Number , "hash" , results [ index ] . Header . Hash ( ) , "err" , err )
return errInvalidChain
}
// Shift the results to the next batch
results = results [ items : ]
}
return nil
}
func ( d * Downloader ) commitPivotBlock ( result * fetchResult ) error {
b := types . NewBlockWithHeader ( result . Header ) . WithBody ( result . Transactions , result . Uncles )
// Sync the pivot block state. This should complete reasonably quickly because
// we've already synced up to the reported head block state earlier.
if err := d . syncState ( b . Root ( ) ) . Wait ( ) ; err != nil {
return err
}
log . Debug ( "Committing fast sync pivot as new head" , "number" , b . Number ( ) , "hash" , b . Hash ( ) )
if _ , err := d . insertReceipts ( [ ] * types . Block { b } , [ ] types . Receipts { result . Receipts } ) ; err != nil {
return err
2015-06-12 05:35:29 -05:00
}
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 07:26:03 -05:00
return d . commitHeadBlock ( b . Hash ( ) )
2015-06-12 05:35:29 -05:00
}
2016-03-15 13:27:49 -05:00
// DeliverHeaders injects a new batch of block headers received from a remote
2015-08-14 13:25:41 -05:00
// node into the download schedule.
2015-08-25 05:57:49 -05:00
func ( d * Downloader ) DeliverHeaders ( id string , headers [ ] * types . Header ) ( err error ) {
2015-10-05 11:37:56 -05:00
return d . deliver ( id , d . headerCh , & headerPack { id , headers } , headerInMeter , headerDropMeter )
2015-08-14 13:25:41 -05:00
}
// DeliverBodies injects a new batch of block bodies received from a remote node.
2015-08-25 05:57:49 -05:00
func ( d * Downloader ) DeliverBodies ( id string , transactions [ ] [ ] * types . Transaction , uncles [ ] [ ] * types . Header ) ( err error ) {
2015-10-05 11:37:56 -05:00
return d . deliver ( id , d . bodyCh , & bodyPack { id , transactions , uncles } , bodyInMeter , bodyDropMeter )
2015-09-28 11:27:31 -05:00
}
// DeliverReceipts injects a new batch of receipts received from a remote node.
func ( d * Downloader ) DeliverReceipts ( id string , receipts [ ] [ ] * types . Receipt ) ( err error ) {
2015-10-05 11:37:56 -05:00
return d . deliver ( id , d . receiptCh , & receiptPack { id , receipts } , receiptInMeter , receiptDropMeter )
}
// DeliverNodeData injects a new batch of node state data received from a remote node.
func ( d * Downloader ) DeliverNodeData ( id string , data [ ] [ ] byte ) ( err error ) {
return d . deliver ( id , d . stateCh , & statePack { id , data } , stateInMeter , stateDropMeter )
}
// deliver injects a new batch of data received from a remote node.
func ( d * Downloader ) deliver ( id string , destCh chan dataPack , packet dataPack , inMeter , dropMeter metrics . Meter ) ( err error ) {
2015-09-28 11:27:31 -05:00
// Update the delivery metrics for both good and failed deliveries
2015-10-05 11:37:56 -05:00
inMeter . Mark ( int64 ( packet . Items ( ) ) )
2015-09-28 11:27:31 -05:00
defer func ( ) {
if err != nil {
2015-10-05 11:37:56 -05:00
dropMeter . Mark ( int64 ( packet . Items ( ) ) )
2015-09-28 11:27:31 -05:00
}
} ( )
// Deliver or abort if the sync is canceled while queuing
d . cancelLock . RLock ( )
cancel := d . cancelCh
d . cancelLock . RUnlock ( )
2015-11-13 10:08:15 -06:00
if cancel == nil {
return errNoSyncActive
}
2015-09-28 11:27:31 -05:00
select {
2015-10-05 11:37:56 -05:00
case destCh <- packet :
2015-05-13 05:47:21 -05:00
return nil
case <- cancel :
return errNoSyncActive
}
2015-04-12 05:38:25 -05:00
}
2016-06-01 10:07:25 -05:00
// qosTuner is the quality of service tuning loop that occasionally gathers the
// peer latency statistics and updates the estimated request round trip time.
func ( d * Downloader ) qosTuner ( ) {
for {
// Retrieve the current median RTT and integrate into the previoust target RTT
rtt := time . Duration ( float64 ( 1 - qosTuningImpact ) * float64 ( atomic . LoadUint64 ( & d . rttEstimate ) ) + qosTuningImpact * float64 ( d . peers . medianRTT ( ) ) )
atomic . StoreUint64 ( & d . rttEstimate , uint64 ( rtt ) )
// A new RTT cycle passed, increase our confidence in the estimated RTT
conf := atomic . LoadUint64 ( & d . rttConfidence )
conf = conf + ( 1000000 - conf ) / 2
atomic . StoreUint64 ( & d . rttConfidence , conf )
// Log the new QoS values and sleep until the next RTT
2017-02-24 10:23:03 -06:00
log . Debug ( "Recalculated downloader QoS values" , "rtt" , rtt , "confidence" , float64 ( conf ) / 1000000.0 , "ttl" , d . requestTTL ( ) )
2016-06-01 10:07:25 -05:00
select {
case <- d . quitCh :
return
case <- time . After ( rtt ) :
}
}
}
// qosReduceConfidence is meant to be called when a new peer joins the downloader's
// peer set, needing to reduce the confidence we have in out QoS estimates.
func ( d * Downloader ) qosReduceConfidence ( ) {
// If we have a single peer, confidence is always 1
peers := uint64 ( d . peers . Len ( ) )
2017-05-02 08:14:35 -05:00
if peers == 0 {
// Ensure peer connectivity races don't catch us off guard
return
}
2016-06-01 10:07:25 -05:00
if peers == 1 {
atomic . StoreUint64 ( & d . rttConfidence , 1000000 )
return
}
// If we have a ton of peers, don't drop confidence)
if peers >= uint64 ( qosConfidenceCap ) {
return
}
// Otherwise drop the confidence factor
conf := atomic . LoadUint64 ( & d . rttConfidence ) * ( peers - 1 ) / peers
if float64 ( conf ) / 1000000 < rttMinConfidence {
conf = uint64 ( rttMinConfidence * 1000000 )
}
atomic . StoreUint64 ( & d . rttConfidence , conf )
rtt := time . Duration ( atomic . LoadUint64 ( & d . rttEstimate ) )
2017-02-24 10:23:03 -06:00
log . Debug ( "Relaxed downloader QoS values" , "rtt" , rtt , "confidence" , float64 ( conf ) / 1000000.0 , "ttl" , d . requestTTL ( ) )
2016-06-01 10:07:25 -05:00
}
// requestRTT returns the current target round trip time for a download request
// to complete in.
//
// Note, the returned RTT is .9 of the actually estimated RTT. The reason is that
// the downloader tries to adapt queries to the RTT, so multiple RTT values can
// be adapted to, but smaller ones are preffered (stabler download stream).
func ( d * Downloader ) requestRTT ( ) time . Duration {
return time . Duration ( atomic . LoadUint64 ( & d . rttEstimate ) ) * 9 / 10
}
// requestTTL returns the current timeout allowance for a single download request
// to finish under.
func ( d * Downloader ) requestTTL ( ) time . Duration {
var (
rtt = time . Duration ( atomic . LoadUint64 ( & d . rttEstimate ) )
conf = float64 ( atomic . LoadUint64 ( & d . rttConfidence ) ) / 1000000.0
)
ttl := time . Duration ( ttlScaling ) * time . Duration ( float64 ( rtt ) / conf )
if ttl > ttlLimit {
ttl = ttlLimit
}
return ttl
}