downloader.go 76.3 KB
Newer Older
F
Felix Lange 已提交
1
// Copyright 2015 The go-ethereum Authors
2
// This file is part of the go-ethereum library.
F
Felix Lange 已提交
3
//
4
// The go-ethereum library is free software: you can redistribute it and/or modify
F
Felix Lange 已提交
5 6 7 8
// 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.
//
9
// The go-ethereum library is distributed in the hope that it will be useful,
F
Felix Lange 已提交
10
// but WITHOUT ANY WARRANTY; without even the implied warranty of
11
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
F
Felix Lange 已提交
12 13 14
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
15
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
F
Felix Lange 已提交
16

17
// Package downloader contains the manual full chain synchronisation.
18 19 20
package downloader

import (
21
	"errors"
22
	"fmt"
23
	"math/big"
24 25 26 27
	"sync"
	"sync/atomic"
	"time"

28
	"github.com/ethereum/go-ethereum"
29
	"github.com/ethereum/go-ethereum/common"
30
	"github.com/ethereum/go-ethereum/core/rawdb"
31
	"github.com/ethereum/go-ethereum/core/types"
32
	"github.com/ethereum/go-ethereum/eth/protocols/snap"
33
	"github.com/ethereum/go-ethereum/ethdb"
34
	"github.com/ethereum/go-ethereum/event"
35
	"github.com/ethereum/go-ethereum/log"
36
	"github.com/ethereum/go-ethereum/metrics"
37
	"github.com/ethereum/go-ethereum/params"
38
	"github.com/ethereum/go-ethereum/trie"
39 40
)

41
var (
42 43
	MaxBlockFetch   = 128 // Amount of blocks to be fetched per retrieval request
	MaxHeaderFetch  = 192 // Amount of block headers to be fetched per retrieval request
44
	MaxSkeletonSize = 128 // Number of header fetches to need for a skeleton assembly
45
	MaxReceiptFetch = 256 // Amount of transaction receipts to allow fetching per request
46
	MaxStateFetch   = 384 // Amount of node state values to allow fetching per request
47

48 49 50 51 52
	rttMinEstimate   = 2 * time.Second  // Minimum round-trip time to target for download requests
	rttMaxEstimate   = 20 * time.Second // Maximum round-trip time to target for download requests
	rttMinConfidence = 0.1              // Worse confidence factor in our estimated RTT value
	ttlScaling       = 3                // Constant scaling factor for RTT -> TTL conversion
	ttlLimit         = time.Minute      // Maximum TTL allowance to prevent reaching crazy timeouts
53 54 55 56

	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
57

58 59 60 61 62
	maxQueuedHeaders            = 32 * 1024                         // [eth/62] Maximum number of headers to queue for import (DOS protection)
	maxHeadersProcess           = 2048                              // Number of header download results to import at once into the chain
	maxResultsProcess           = 2048                              // Number of content download results to import at once into the chain
	fullMaxForkAncestry  uint64 = params.FullImmutabilityThreshold  // Maximum chain reorganisation (locally redeclared so tests can reduce it)
	lightMaxForkAncestry uint64 = params.LightImmutabilityThreshold // Maximum chain reorganisation (locally redeclared so tests can reduce it)
63

64 65 66
	reorgProtThreshold   = 48 // Threshold number of recent blocks to disable mini reorg protection
	reorgProtHeaderDelay = 2  // Number of headers to delay delivering to cover mini reorgs

67 68 69 70 71
	fsHeaderCheckFrequency = 100             // Verification frequency of the downloaded headers during fast sync
	fsHeaderSafetyNet      = 2048            // Number of headers to discard in case a chain violation is detected
	fsHeaderForceVerify    = 24              // Number of headers to verify before and after the pivot to accept it
	fsHeaderContCheck      = 3 * time.Second // Time interval to check for header continuations during state download
	fsMinFullBlocks        = 64              // Number of blocks to retrieve fully even in fast sync
72
)
73

74
var (
75 76 77 78
	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")
79
	errUnsyncedPeer            = errors.New("unsynced peer")
80 81 82 83 84 85 86 87 88 89
	errNoPeers                 = errors.New("no peers to keep download active")
	errTimeout                 = errors.New("timeout")
	errEmptyHeaderSet          = errors.New("empty header set by peer")
	errPeersUnavailable        = errors.New("no peers available or all tried for download")
	errInvalidAncestor         = errors.New("retrieved ancestor is invalid")
	errInvalidChain            = errors.New("retrieved hash chain is invalid")
	errInvalidBody             = errors.New("retrieved block body is invalid")
	errInvalidReceipt          = errors.New("retrieved receipt is invalid")
	errCancelStateFetch        = errors.New("state data download canceled (requested)")
	errCancelContentProcessing = errors.New("content processing canceled (requested)")
90
	errCanceled                = errors.New("syncing canceled (requested)")
91
	errNoSyncActive            = errors.New("no sync active")
92
	errTooOld                  = errors.New("peer's protocol version too old")
93 94
)

95
type Downloader struct {
96 97 98 99 100 101 102
	// WARNING: The `rttEstimate` and `rttConfidence` fields are accessed atomically.
	// On 32 bit platforms, only 64-bit aligned fields can be atomic. The struct is
	// guaranteed to be so aligned, so take advantage of that. For more information,
	// see https://golang.org/pkg/sync/atomic/#pkg-note-BUG.
	rttEstimate   uint64 // Round trip time to target for download requests
	rttConfidence uint64 // Confidence in the estimated RTT (unit: millionths to allow atomic ops)

103
	mode uint32         // Synchronisation mode defining the strategy used (per sync cycle), use d.getMode() to get the SyncMode
104
	mux  *event.TypeMux // Event multiplexer to announce sync operation events
105

106 107 108 109
	checkpoint uint64   // Checkpoint block number to enforce head against (e.g. fast sync)
	genesis    uint64   // Genesis block number to limit sync to (e.g. light client CHT)
	queue      *queue   // Scheduler for selecting the hashes to download
	peers      *peerSet // Set of active peers from which download can proceed
110 111

	stateDB    ethdb.Database  // Database to state sync into (and deduplicate via)
112
	stateBloom *trie.SyncBloom // Bloom filter for fast trie node and contract code existence checks
113

114
	// Statistics
115 116 117
	syncStatsChainOrigin uint64 // Origin block number where syncing started at
	syncStatsChainHeight uint64 // Highest block number known when syncing started
	syncStatsState       stateSyncStats
118
	syncStatsLock        sync.RWMutex // Lock protecting the sync stats fields
119

120
	lightchain LightChain
121
	blockchain BlockChain
122

123
	// Callbacks
124
	dropPeer peerDropFn // Drops a peer for misbehaving
125

126
	// Status
127 128 129
	synchroniseMock func(id string, hash common.Hash) error // Replacement for synchronise during testing
	synchronising   int32
	notified        int32
130
	committed       int32
131
	ancientLimit    uint64 // The maximum block number which can be regarded as ancient data.
132 133

	// Channels
134 135 136 137 138 139
	headerCh      chan dataPack        // Channel receiving inbound block headers
	bodyCh        chan dataPack        // Channel receiving inbound block bodies
	receiptCh     chan dataPack        // Channel receiving inbound receipts
	bodyWakeCh    chan bool            // Channel to signal the block body fetcher of new tasks
	receiptWakeCh chan bool            // Channel to signal the receipt fetcher of new tasks
	headerProcCh  chan []*types.Header // Channel to feed the header processor new tasks
140

141 142 143 144
	// State sync
	pivotHeader *types.Header // Pivot block header to dynamically push the syncing state root
	pivotLock   sync.RWMutex  // Lock protecting pivot header reads from updates

145 146
	snapSync       bool         // Whether to run state sync over the snap protocol
	SnapSyncer     *snap.Syncer // TODO(karalabe): make private! hack for now
147 148
	stateSyncStart chan *stateSync
	trackStateReq  chan *stateReq
149
	stateCh        chan dataPack // Channel receiving inbound node state data
150

151
	// Cancellation and termination
152 153 154 155
	cancelPeer string         // Identifier of the peer currently being used as the master (cancel on drop)
	cancelCh   chan struct{}  // Channel to cancel mid-flight syncs
	cancelLock sync.RWMutex   // Lock to protect the cancel channel and peer in delivers
	cancelWg   sync.WaitGroup // Make sure all fetcher goroutines have exited.
156

157
	quitCh   chan struct{} // Quit channel to signal termination
158
	quitLock sync.Mutex    // Lock to prevent double closes
159

160
	// Testing hooks
161 162 163 164
	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)
165 166
}

N
Nick Johnson 已提交
167
// LightChain encapsulates functions required to synchronise a light chain.
168 169
type LightChain interface {
	// HasHeader verifies a header's presence in the local chain.
170
	HasHeader(common.Hash, uint64) bool
171 172 173 174 175 176 177

	// GetHeaderByHash retrieves a header from the local chain.
	GetHeaderByHash(common.Hash) *types.Header

	// CurrentHeader retrieves the head header from the local chain.
	CurrentHeader() *types.Header

178 179
	// GetTd returns the total difficulty of a local block.
	GetTd(common.Hash, uint64) *big.Int
180 181 182 183

	// InsertHeaderChain inserts a batch of headers into the local chain.
	InsertHeaderChain([]*types.Header, int) (int, error)

184 185
	// SetHead rewinds the local chain to a new head.
	SetHead(uint64) error
186 187
}

N
Nick Johnson 已提交
188
// BlockChain encapsulates functions required to sync a (full or fast) blockchain.
189 190 191
type BlockChain interface {
	LightChain

192 193
	// HasBlock verifies a block's presence in the local chain.
	HasBlock(common.Hash, uint64) bool
194

195 196 197
	// HasFastBlock verifies a fast block's presence in the local chain.
	HasFastBlock(common.Hash, uint64) bool

198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
	// GetBlockByHash retrieves a block from the local chain.
	GetBlockByHash(common.Hash) *types.Block

	// CurrentBlock retrieves the head block from the local chain.
	CurrentBlock() *types.Block

	// CurrentFastBlock retrieves the head fast block from the local chain.
	CurrentFastBlock() *types.Block

	// FastSyncCommitHead directly commits the head block to a certain entity.
	FastSyncCommitHead(common.Hash) error

	// InsertChain inserts a batch of blocks into the local chain.
	InsertChain(types.Blocks) (int, error)

	// InsertReceiptChain inserts a batch of receipts into the local chain.
214
	InsertReceiptChain(types.Blocks, []types.Receipts, uint64) (int, error)
215 216
}

217
// New creates a new downloader to fetch hashes and blocks from remote peers.
218
func New(checkpoint uint64, stateDb ethdb.Database, stateBloom *trie.SyncBloom, mux *event.TypeMux, chain BlockChain, lightchain LightChain, dropPeer peerDropFn) *Downloader {
219 220 221
	if lightchain == nil {
		lightchain = chain
	}
222
	dl := &Downloader{
223
		stateDB:        stateDb,
224
		stateBloom:     stateBloom,
225
		mux:            mux,
226
		checkpoint:     checkpoint,
227
		queue:          newQueue(blockCacheMaxItems, blockCacheInitialItems),
228 229 230
		peers:          newPeerSet(),
		rttEstimate:    uint64(rttMaxEstimate),
		rttConfidence:  uint64(1000000),
231
		blockchain:     chain,
232 233 234 235 236 237 238 239 240 241
		lightchain:     lightchain,
		dropPeer:       dropPeer,
		headerCh:       make(chan dataPack, 1),
		bodyCh:         make(chan dataPack, 1),
		receiptCh:      make(chan dataPack, 1),
		bodyWakeCh:     make(chan bool, 1),
		receiptWakeCh:  make(chan bool, 1),
		headerProcCh:   make(chan []*types.Header, 1),
		quitCh:         make(chan struct{}),
		stateCh:        make(chan dataPack),
242
		SnapSyncer:     snap.NewSyncer(stateDb, stateBloom),
243
		stateSyncStart: make(chan *stateSync),
244
		syncStatsState: stateSyncStats{
245
			processed: rawdb.ReadFastTrieProgress(stateDb),
246 247
		},
		trackStateReq: make(chan *stateReq),
248
	}
249
	go dl.qosTuner()
250
	go dl.stateFetcher()
251
	return dl
252 253
}

254 255 256
// 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.
257
//
L
Leif Jurvetson 已提交
258
// In addition, during the state download phase of fast synchronisation the number
259 260
// of processed and the total number of known states are also returned. Otherwise
// these are zero.
261
func (d *Downloader) Progress() ethereum.SyncProgress {
262
	// Lock the current stats and return the progress
263 264
	d.syncStatsLock.RLock()
	defer d.syncStatsLock.RUnlock()
265

266
	current := uint64(0)
267
	mode := d.getMode()
268
	switch {
269
	case d.blockchain != nil && mode == FullSync:
270
		current = d.blockchain.CurrentBlock().NumberU64()
271
	case d.blockchain != nil && mode == FastSync:
272
		current = d.blockchain.CurrentFastBlock().NumberU64()
273
	case d.lightchain != nil:
274
		current = d.lightchain.CurrentHeader().Number.Uint64()
275
	default:
276
		log.Error("Unknown downloader chain/mode combo", "light", d.lightchain != nil, "full", d.blockchain != nil, "mode", mode)
277
	}
278 279 280 281
	return ethereum.SyncProgress{
		StartingBlock: d.syncStatsChainOrigin,
		CurrentBlock:  current,
		HighestBlock:  d.syncStatsChainHeight,
282 283
		PulledStates:  d.syncStatsState.processed,
		KnownStates:   d.syncStatsState.processed + d.syncStatsState.pending,
284
	}
O
obscuren 已提交
285 286
}

287
// Synchronising returns whether the downloader is currently retrieving blocks.
288
func (d *Downloader) Synchronising() bool {
289
	return atomic.LoadInt32(&d.synchronising) > 0
290 291
}

292 293
// RegisterPeer injects a new download peer into the set of block source to be
// used for fetching hashes and blocks from.
294 295 296 297 298 299 300 301
func (d *Downloader) RegisterPeer(id string, version uint, peer Peer) error {
	var logger log.Logger
	if len(id) < 16 {
		// Tests use short IDs, don't choke on them
		logger = log.New("peer", id)
	} else {
		logger = log.New("peer", id[:16])
	}
302
	logger.Trace("Registering sync peer")
303
	if err := d.peers.Register(newPeerConnection(id, version, peer, logger)); err != nil {
304
		logger.Error("Failed to register sync peer", "err", err)
305 306
		return err
	}
307 308
	d.qosReduceConfidence()

309 310 311
	return nil
}

N
Nick Johnson 已提交
312
// RegisterLightPeer injects a light client peer, wrapping it so it appears as a regular peer.
313
func (d *Downloader) RegisterLightPeer(id string, version uint, peer LightPeer) error {
314 315 316
	return d.RegisterPeer(id, version, &lightPeerWrapper{peer})
}

317
// UnregisterPeer remove a peer from the known list, preventing any action from
318 319
// the specified peer. An effort is also made to return any pending fetches into
// the queue.
320
func (d *Downloader) UnregisterPeer(id string) error {
321
	// Unregister the peer from the active peer set and revoke any fetch tasks
322 323 324 325 326 327 328
	var logger log.Logger
	if len(id) < 16 {
		// Tests use short IDs, don't choke on them
		logger = log.New("peer", id)
	} else {
		logger = log.New("peer", id[:16])
	}
329
	logger.Trace("Unregistering sync peer")
330
	if err := d.peers.Unregister(id); err != nil {
331
		logger.Error("Failed to unregister sync peer", "err", err)
332 333
		return err
	}
334
	d.queue.Revoke(id)
335

336
	return nil
337 338
}

339 340
// 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.
341 342
func (d *Downloader) Synchronise(id string, head common.Hash, td *big.Int, mode SyncMode) error {
	err := d.synchronise(id, head, td, mode)
343

344
	switch err {
345 346 347
	case nil, errBusy, errCanceled:
		return err
	}
348

349 350 351
	if errors.Is(err, errInvalidChain) || errors.Is(err, errBadPeer) || errors.Is(err, errTimeout) ||
		errors.Is(err, errStallingPeer) || errors.Is(err, errUnsyncedPeer) || errors.Is(err, errEmptyHeaderSet) ||
		errors.Is(err, errPeersUnavailable) || errors.Is(err, errTooOld) || errors.Is(err, errInvalidAncestor) {
352 353 354 355 356 357 358 359 360 361
		log.Warn("Synchronisation failed, dropping peer", "peer", id, "err", err)
		if d.dropPeer == nil {
			// The dropPeer method is nil when `--copydb` is used for a local copy.
			// Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored
			log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", id)
		} else {
			d.dropPeer(id)
		}
		return err
	}
362
	log.Warn("Synchronisation failed, retrying", "err", err)
363
	return err
364 365 366
}

// synchronise will select the peer and use it for synchronising. If an empty string is given
367
// it will use the best peer possible and synchronize if its TD is higher than our own. If any of the
368
// checks fail an error will be returned. This method is synchronous
369
func (d *Downloader) synchronise(id string, hash common.Hash, td *big.Int, mode SyncMode) error {
L
Leif Jurvetson 已提交
370
	// Mock out the synchronisation if testing
371 372 373
	if d.synchroniseMock != nil {
		return d.synchroniseMock(id, hash)
	}
374
	// Make sure only one goroutine is ever allowed past this point at once
375
	if !atomic.CompareAndSwapInt32(&d.synchronising, 0, 1) {
376
		return errBusy
377
	}
378
	defer atomic.StoreInt32(&d.synchronising, 0)
379

380 381
	// Post a user notification of the sync (only once per session)
	if atomic.CompareAndSwapInt32(&d.notified, 0, 1) {
382
		log.Info("Block synchronisation started")
383
	}
384
	// If we are already full syncing, but have a fast-sync bloom filter laying
385
	// around, make sure it doesn't use memory any more. This is a special case
386 387 388 389
	// when the user attempts to fast sync a new empty network.
	if mode == FullSync && d.stateBloom != nil {
		d.stateBloom.Close()
	}
390 391 392 393 394 395 396 397 398 399
	// If snap sync was requested, create the snap scheduler and switch to fast
	// sync mode. Long term we could drop fast sync or merge the two together,
	// but until snap becomes prevalent, we should support both. TODO(karalabe).
	if mode == SnapSync {
		if !d.snapSync {
			log.Warn("Enabling snapshot sync prototype")
			d.snapSync = true
		}
		mode = FastSync
	}
400
	// Reset the queue, peer set and wake channels to clean any internal leftover state
401
	d.queue.Reset(blockCacheMaxItems, blockCacheInitialItems)
402
	d.peers.Reset()
403

404
	for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
405 406 407 408
		select {
		case <-ch:
		default:
		}
409
	}
410
	for _, ch := range []chan dataPack{d.headerCh, d.bodyCh, d.receiptCh} {
411 412 413 414 415 416 417 418
		for empty := false; !empty; {
			select {
			case <-ch:
			default:
				empty = true
			}
		}
	}
419 420 421 422 423 424 425
	for empty := false; !empty; {
		select {
		case <-d.headerProcCh:
		default:
			empty = true
		}
	}
426
	// Create cancel channel for aborting mid-flight and mark the master peer
427 428
	d.cancelLock.Lock()
	d.cancelCh = make(chan struct{})
429
	d.cancelPeer = id
430 431
	d.cancelLock.Unlock()

432
	defer d.Cancel() // No matter what, we can't leave the cancel channel open
433

434 435
	// Atomically set the requested sync mode
	atomic.StoreUint32(&d.mode, uint32(mode))
436

437
	// Retrieve the origin peer and initiate the downloading process
438
	p := d.peers.Peer(id)
439
	if p == nil {
440
		return errUnknownPeer
441
	}
442
	return d.syncWithPeer(p, hash, td)
443 444
}

445 446 447 448
func (d *Downloader) getMode() SyncMode {
	return SyncMode(atomic.LoadUint32(&d.mode))
}

449 450
// syncWithPeer starts a block synchronization based on the hash chain from the
// specified peer and head hash.
451
func (d *Downloader) syncWithPeer(p *peerConnection, hash common.Hash, td *big.Int) (err error) {
O
obscuren 已提交
452
	d.mux.Post(StartEvent{})
453 454 455
	defer func() {
		// reset on error
		if err != nil {
456 457
			d.mux.Post(FailedEvent{err})
		} else {
458 459
			latest := d.lightchain.CurrentHeader()
			d.mux.Post(DoneEvent{latest})
460 461
		}
	}()
462
	if p.version < 64 {
463
		return fmt.Errorf("%w: advertized %d < required %d", errTooOld, p.version, 64)
464
	}
465
	mode := d.getMode()
466

467
	log.Debug("Synchronising with the network", "peer", p.id, "eth", p.version, "head", hash, "td", td, "mode", mode)
468
	defer func(start time.Time) {
469
		log.Debug("Synchronisation terminated", "elapsed", common.PrettyDuration(time.Since(start)))
470
	}(time.Now())
471

472
	// Look up the sync boundaries: the common ancestor and the target block
473
	latest, pivot, err := d.fetchHead(p)
474 475 476
	if err != nil {
		return err
	}
477 478 479 480 481 482 483
	if mode == FastSync && pivot == nil {
		// If no pivot block was returned, the head is below the min full block
		// threshold (i.e. new chian). In that case we won't really fast sync
		// anyway, but still need a valid pivot block to avoid some code hitting
		// nil panics on an access.
		pivot = d.blockchain.CurrentBlock().Header()
	}
484
	height := latest.Number.Uint64()
485

486
	origin, err := d.findAncestor(p, latest)
487 488 489 490 491 492 493 494 495
	if err != nil {
		return err
	}
	d.syncStatsLock.Lock()
	if d.syncStatsChainHeight <= origin || d.syncStatsChainOrigin > origin {
		d.syncStatsChainOrigin = origin
	}
	d.syncStatsChainHeight = height
	d.syncStatsLock.Unlock()
496

497
	// Ensure our origin point is below any fast sync pivot point
498
	if mode == FastSync {
499 500
		if height <= uint64(fsMinFullBlocks) {
			origin = 0
501
		} else {
502 503 504
			pivotNumber := pivot.Number.Uint64()
			if pivotNumber <= origin {
				origin = pivotNumber - 1
505
			}
506 507
			// Write out the pivot into the database so a rollback beyond it will
			// reenable fast sync
508
			rawdb.WriteLastPivotNumber(d.stateDB, pivotNumber)
509
		}
510
	}
511
	d.committed = 1
512
	if mode == FastSync && pivot.Number.Uint64() != 0 {
513 514
		d.committed = 0
	}
515
	if mode == FastSync {
516
		// Set the ancient data limitation.
517 518 519
		// If we are running fast sync, all block data older than ancientLimit will be
		// written to the ancient store. More recent data will be written to the active
		// database and will wait for the freezer to migrate.
520
		//
521 522 523
		// If there is a checkpoint available, then calculate the ancientLimit through
		// that. Otherwise calculate the ancient limit through the advertised height
		// of the remote peer.
524
		//
525 526 527 528 529
		// The reason for picking checkpoint first is that a malicious peer can give us
		// a fake (very high) height, forcing the ancient limit to also be very high.
		// The peer would start to feed us valid blocks until head, resulting in all of
		// the blocks might be written into the ancient store. A following mini-reorg
		// could cause issues.
530
		if d.checkpoint != 0 && d.checkpoint > fullMaxForkAncestry+1 {
531
			d.ancientLimit = d.checkpoint
532 533
		} else if height > fullMaxForkAncestry+1 {
			d.ancientLimit = height - fullMaxForkAncestry - 1
534 535
		} else {
			d.ancientLimit = 0
536 537
		}
		frozen, _ := d.stateDB.Ancients() // Ignore the error here since light client can also hit here.
538

539 540 541 542 543 544 545 546 547 548
		// If a part of blockchain data has already been written into active store,
		// disable the ancient style insertion explicitly.
		if origin >= frozen && frozen != 0 {
			d.ancientLimit = 0
			log.Info("Disabling direct-ancient mode", "origin", origin, "ancient", frozen-1)
		} else if d.ancientLimit > 0 {
			log.Debug("Enabling direct-ancient mode", "ancient", d.ancientLimit)
		}
		// Rewind the ancient store and blockchain if reorg happens.
		if origin+1 < frozen {
549 550
			if err := d.lightchain.SetHead(origin + 1); err != nil {
				return err
551 552 553
			}
		}
	}
554
	// Initiate the sync using a concurrent header and content retrieval algorithm
555
	d.queue.Prepare(origin+1, mode)
556 557
	if d.syncInitHook != nil {
		d.syncInitHook(origin, height)
558
	}
559
	fetchers := []func() error{
560 561 562 563
		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) },
564
	}
565
	if mode == FastSync {
566 567 568 569 570
		d.pivotLock.Lock()
		d.pivotHeader = pivot
		d.pivotLock.Unlock()

		fetchers = append(fetchers, func() error { return d.processFastSyncContent() })
571
	} else if mode == FullSync {
572 573
		fetchers = append(fetchers, d.processFullSyncContent)
	}
574
	return d.spawnSync(fetchers)
575 576 577 578
}

// spawnSync runs d.process and all given fetcher functions to completion in
// separate goroutines, returning the first error that appears.
579 580
func (d *Downloader) spawnSync(fetchers []func() error) error {
	errc := make(chan error, len(fetchers))
581
	d.cancelWg.Add(len(fetchers))
582 583
	for _, fn := range fetchers {
		fn := fn
584
		go func() { defer d.cancelWg.Done(); errc <- fn() }()
585 586 587
	}
	// Wait for the first error, then terminate the others.
	var err error
588 589
	for i := 0; i < len(fetchers); i++ {
		if i == len(fetchers)-1 {
590 591 592 593 594
			// 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()
		}
595
		if err = <-errc; err != nil && err != errCanceled {
596 597 598 599
			break
		}
	}
	d.queue.Close()
600
	d.Cancel()
601
	return err
602 603
}

604 605 606 607
// cancel aborts all of the operations and resets the queue. However, cancel does
// not wait for the running download goroutines to finish. This method should be
// used when cancelling the downloads from inside the downloader.
func (d *Downloader) cancel() {
608
	// Close the current cancel channel
609
	d.cancelLock.Lock()
610 611
	defer d.cancelLock.Unlock()

612 613 614 615 616 617 618
	if d.cancelCh != nil {
		select {
		case <-d.cancelCh:
			// Channel was already closed
		default:
			close(d.cancelCh)
		}
619
	}
620 621 622 623 624 625
}

// Cancel aborts all of the operations and waits for all download goroutines to
// finish before returning.
func (d *Downloader) Cancel() {
	d.cancel()
626
	d.cancelWg.Wait()
627 628
}

629
// Terminate interrupts the downloader, canceling all pending operations.
630
// The downloader cannot be reused after calling Terminate.
631
func (d *Downloader) Terminate() {
632 633 634 635 636 637 638
	// Close the termination channel (make sure double close is allowed)
	d.quitLock.Lock()
	select {
	case <-d.quitCh:
	default:
		close(d.quitCh)
	}
639 640 641
	if d.stateBloom != nil {
		d.stateBloom.Close()
	}
642 643 644
	d.quitLock.Unlock()

	// Cancel any pending download requests
645
	d.Cancel()
646 647
}

648 649 650 651 652
// fetchHead retrieves the head header and prior pivot block (if available) from
// a remote peer.
func (d *Downloader) fetchHead(p *peerConnection) (head *types.Header, pivot *types.Header, err error) {
	p.log.Debug("Retrieving remote chain head")
	mode := d.getMode()
653 654

	// Request the advertised remote head block and wait for the response
655 656 657 658 659 660
	latest, _ := p.peer.Head()
	fetch := 1
	if mode == FastSync {
		fetch = 2 // head + pivot headers
	}
	go p.peer.RequestHeadersByHash(latest, fetch, fsMinFullBlocks-1, true)
661

662 663
	ttl := d.requestTTL()
	timeout := time.After(ttl)
664 665 666
	for {
		select {
		case <-d.cancelCh:
667
			return nil, nil, errCanceled
668

669
		case packet := <-d.headerCh:
670
			// Discard anything not from the origin peer
671
			if packet.PeerId() != p.id {
672
				log.Debug("Received headers from incorrect peer", "peer", packet.PeerId())
673 674
				break
			}
675
			// Make sure the peer gave us at least one and at most the requested headers
676
			headers := packet.(*headerPack).headers
677 678
			if len(headers) == 0 || len(headers) > fetch {
				return nil, nil, fmt.Errorf("%w: returned headers %d != requested %d", errBadPeer, len(headers), fetch)
679
			}
680 681 682
			// The first header needs to be the head, validate against the checkpoint
			// and request. If only 1 header was returned, make sure there's no pivot
			// or there was not one requested.
683
			head := headers[0]
684
			if (mode == FastSync || mode == LightSync) && head.Number.Uint64() < d.checkpoint {
685 686 687 688 689 690 691 692 693 694 695 696 697 698
				return nil, nil, fmt.Errorf("%w: remote head %d below checkpoint %d", errUnsyncedPeer, head.Number, d.checkpoint)
			}
			if len(headers) == 1 {
				if mode == FastSync && head.Number.Uint64() > uint64(fsMinFullBlocks) {
					return nil, nil, fmt.Errorf("%w: no pivot included along head header", errBadPeer)
				}
				p.log.Debug("Remote head identified, no pivot", "number", head.Number, "hash", head.Hash())
				return head, nil, nil
			}
			// At this point we have 2 headers in total and the first is the
			// validated head of the chian. Check the pivot number and return,
			pivot := headers[1]
			if pivot.Number.Uint64() != head.Number.Uint64()-uint64(fsMinFullBlocks) {
				return nil, nil, fmt.Errorf("%w: remote pivot %d != requested %d", errInvalidChain, pivot.Number, head.Number.Uint64()-uint64(fsMinFullBlocks))
699
			}
700
			return head, pivot, nil
701

702
		case <-timeout:
P
Péter Szilágyi 已提交
703
			p.log.Debug("Waiting for head header timed out", "elapsed", ttl)
704
			return nil, nil, errTimeout
705

706
		case <-d.bodyCh:
707 708
		case <-d.receiptCh:
			// Out of bounds delivery, ignore
709 710 711 712
		}
	}
}

713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735
// calculateRequestSpan calculates what headers to request from a peer when trying to determine the
// common ancestor.
// It returns parameters to be used for peer.RequestHeadersByNumber:
//  from - starting block number
//  count - number of headers to request
//  skip - number of headers to skip
// and also returns 'max', the last block which is expected to be returned by the remote peers,
// given the (from,count,skip)
func calculateRequestSpan(remoteHeight, localHeight uint64) (int64, int, int, uint64) {
	var (
		from     int
		count    int
		MaxCount = MaxHeaderFetch / 16
	)
	// requestHead is the highest block that we will ask for. If requestHead is not offset,
	// the highest block that we will get is 16 blocks back from head, which means we
	// will fetch 14 or 15 blocks unnecessarily in the case the height difference
	// between us and the peer is 1-2 blocks, which is most common
	requestHead := int(remoteHeight) - 1
	if requestHead < 0 {
		requestHead = 0
	}
	// requestBottom is the lowest block we want included in the query
736
	// Ideally, we want to include the one just below our own head
737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
	requestBottom := int(localHeight - 1)
	if requestBottom < 0 {
		requestBottom = 0
	}
	totalSpan := requestHead - requestBottom
	span := 1 + totalSpan/MaxCount
	if span < 2 {
		span = 2
	}
	if span > 16 {
		span = 16
	}

	count = 1 + totalSpan/span
	if count > MaxCount {
		count = MaxCount
	}
	if count < 2 {
		count = 2
	}
	from = requestHead - (count-1)*span
	if from < 0 {
		from = 0
	}
	max := from + (count-1)*span
	return int64(from), count, span - 1, uint64(max)
}

765
// findAncestor tries to locate the common ancestor link of the local chain and
766
// a remote peers blockchain. In the general case when our node was in sync and
767
// on the correct chain, checking the top N links should already get us a match.
768
// In the rare scenario when we ended up on a long reorganisation (i.e. none of
769
// the head links match), we do a binary search to find the common ancestor.
770
func (d *Downloader) findAncestor(p *peerConnection, remoteHeader *types.Header) (uint64, error) {
771
	// Figure out the valid ancestor range to prevent rewrite attacks
772 773 774 775 776
	var (
		floor        = int64(-1)
		localHeight  uint64
		remoteHeight = remoteHeader.Number.Uint64()
	)
777 778
	mode := d.getMode()
	switch mode {
779 780 781 782 783 784
	case FullSync:
		localHeight = d.blockchain.CurrentBlock().NumberU64()
	case FastSync:
		localHeight = d.blockchain.CurrentFastBlock().NumberU64()
	default:
		localHeight = d.lightchain.CurrentHeader().Number.Uint64()
785
	}
786
	p.log.Debug("Looking for common ancestor", "local", localHeight, "remote", remoteHeight)
787 788

	// Recap floor value for binary search
789 790 791 792
	maxForkAncestry := fullMaxForkAncestry
	if d.getMode() == LightSync {
		maxForkAncestry = lightMaxForkAncestry
	}
793
	if localHeight >= maxForkAncestry {
794
		// We're above the max reorg threshold, find the earliest fork point
795
		floor = int64(localHeight - maxForkAncestry)
796 797 798
	}
	// If we're doing a light sync, ensure the floor doesn't go below the CHT, as
	// all headers before that point will be missing.
799
	if mode == LightSync {
800
		// If we don't know the current CHT position, find it
801 802 803 804 805 806
		if d.genesis == 0 {
			header := d.lightchain.CurrentHeader()
			for header != nil {
				d.genesis = header.Number.Uint64()
				if floor >= int64(d.genesis)-1 {
					break
807
				}
808
				header = d.lightchain.GetHeaderByHash(header.ParentHash)
809
			}
810 811 812 813
		}
		// We already know the "genesis" block number, cap floor to that
		if floor < int64(d.genesis)-1 {
			floor = int64(d.genesis) - 1
814
		}
815
	}
816

817
	from, count, skip, max := calculateRequestSpan(remoteHeight, localHeight)
818 819

	p.log.Trace("Span searching for common ancestor", "count", count, "from", from, "skip", skip)
820
	go p.peer.RequestHeadersByNumber(uint64(from), count, skip, false)
821 822 823

	// Wait for the remote response to the head fetch
	number, hash := uint64(0), common.Hash{}
824 825 826

	ttl := d.requestTTL()
	timeout := time.After(ttl)
827 828 829 830

	for finished := false; !finished; {
		select {
		case <-d.cancelCh:
831
			return 0, errCanceled
832

833
		case packet := <-d.headerCh:
834
			// Discard anything not from the origin peer
835
			if packet.PeerId() != p.id {
836
				log.Debug("Received headers from incorrect peer", "peer", packet.PeerId())
837 838 839
				break
			}
			// Make sure the peer actually gave something valid
840
			headers := packet.(*headerPack).headers
841
			if len(headers) == 0 {
P
Péter Szilágyi 已提交
842
				p.log.Warn("Empty head header set")
843 844
				return 0, errEmptyHeaderSet
			}
845
			// Make sure the peer's reply conforms to the request
846
			for i, header := range headers {
847
				expectNumber := from + int64(i)*int64(skip+1)
848 849
				if number := header.Number.Int64(); number != expectNumber {
					p.log.Warn("Head headers broke chain ordering", "index", i, "requested", expectNumber, "received", number)
850
					return 0, fmt.Errorf("%w: %v", errInvalidChain, errors.New("head headers broke chain ordering"))
851 852
				}
			}
853 854 855
			// Check if a common ancestor was found
			finished = true
			for i := len(headers) - 1; i >= 0; i-- {
856
				// Skip any headers that underflow/overflow our requested set
857
				if headers[i].Number.Int64() < from || headers[i].Number.Uint64() > max {
858 859 860
					continue
				}
				// Otherwise check if we already know the header or not
861 862
				h := headers[i].Hash()
				n := headers[i].Number.Uint64()
863 864

				var known bool
865
				switch mode {
866 867 868 869 870 871 872 873
				case FullSync:
					known = d.blockchain.HasBlock(h, n)
				case FastSync:
					known = d.blockchain.HasFastBlock(h, n)
				default:
					known = d.lightchain.HasHeader(h, n)
				}
				if known {
874
					number, hash = n, h
875 876 877 878
					break
				}
			}

879
		case <-timeout:
P
Péter Szilágyi 已提交
880
			p.log.Debug("Waiting for head header timed out", "elapsed", ttl)
881 882
			return 0, errTimeout

883
		case <-d.bodyCh:
884 885
		case <-d.receiptCh:
			// Out of bounds delivery, ignore
886 887 888
		}
	}
	// If the head fetch already found an ancestor, return
889
	if hash != (common.Hash{}) {
890
		if int64(number) <= floor {
P
Péter Szilágyi 已提交
891
			p.log.Warn("Ancestor below allowance", "number", number, "hash", hash, "allowance", floor)
892 893
			return 0, errInvalidAncestor
		}
P
Péter Szilágyi 已提交
894
		p.log.Debug("Found common ancestor", "number", number, "hash", hash)
895 896 897
		return number, nil
	}
	// Ancestor not found, we need to binary search over our chain
898
	start, end := uint64(0), remoteHeight
899 900 901
	if floor > 0 {
		start = uint64(floor)
	}
902 903
	p.log.Trace("Binary searching for common ancestor", "start", start, "end", end)

904 905 906 907
	for start+1 < end {
		// Split our chain interval in two, and request the hash to cross check
		check := (start + end) / 2

908 909 910
		ttl := d.requestTTL()
		timeout := time.After(ttl)

911
		go p.peer.RequestHeadersByNumber(check, 1, 0, false)
912 913 914 915 916

		// Wait until a reply arrives to this request
		for arrived := false; !arrived; {
			select {
			case <-d.cancelCh:
917
				return 0, errCanceled
918

919
			case packet := <-d.headerCh:
920
				// Discard anything not from the origin peer
921 922
				if packet.PeerId() != p.id {
					log.Debug("Received headers from incorrect peer", "peer", packet.PeerId())
923 924 925
					break
				}
				// Make sure the peer actually gave something valid
926
				headers := packet.(*headerPack).headers
927
				if len(headers) != 1 {
928
					p.log.Warn("Multiple headers for single request", "headers", len(headers))
929
					return 0, fmt.Errorf("%w: multiple headers (%d) for single request", errBadPeer, len(headers))
930 931 932 933
				}
				arrived = true

				// Modify the search interval based on the response
934 935
				h := headers[0].Hash()
				n := headers[0].Number.Uint64()
936 937

				var known bool
938
				switch mode {
939 940 941 942 943 944 945 946
				case FullSync:
					known = d.blockchain.HasBlock(h, n)
				case FastSync:
					known = d.blockchain.HasFastBlock(h, n)
				default:
					known = d.lightchain.HasHeader(h, n)
				}
				if !known {
947 948 949
					end = check
					break
				}
950
				header := d.lightchain.GetHeaderByHash(h) // Independent of sync mode, header surely exists
951
				if header.Number.Uint64() != check {
952
					p.log.Warn("Received non requested header", "number", header.Number, "hash", header.Hash(), "request", check)
953
					return 0, fmt.Errorf("%w: non-requested header (%d)", errBadPeer, header.Number)
954 955
				}
				start = check
956
				hash = h
957

958
			case <-timeout:
P
Péter Szilágyi 已提交
959
				p.log.Debug("Waiting for search header timed out", "elapsed", ttl)
960 961
				return 0, errTimeout

962
			case <-d.bodyCh:
963 964
			case <-d.receiptCh:
				// Out of bounds delivery, ignore
965 966 967
			}
		}
	}
968 969
	// Ensure valid ancestry and return
	if int64(start) <= floor {
P
Péter Szilágyi 已提交
970
		p.log.Warn("Ancestor below allowance", "number", start, "hash", hash, "allowance", floor)
971 972
		return 0, errInvalidAncestor
	}
P
Péter Szilágyi 已提交
973
	p.log.Debug("Found common ancestor", "number", start, "hash", hash)
974 975 976
	return start, nil
}

977 978 979 980 981
// 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
982
// other peers are only accepted if they map cleanly to the skeleton. If no one
983 984
// can fill in the skeleton - not even the origin peer - it's assumed invalid and
// the origin is dropped.
985
func (d *Downloader) fetchHeaders(p *peerConnection, from uint64) error {
P
Péter Szilágyi 已提交
986 987
	p.log.Debug("Directing header downloads", "origin", from)
	defer p.log.Debug("Header download terminated")
988

989 990
	// Create a timeout timer, and the associated header fetcher
	skeleton := true            // Skeleton assembly phase or finishing up
991
	pivoting := false           // Whether the next request is pivot verification
992
	request := time.Now()       // time of the last skeleton fetch request
993 994 995 996
	timeout := time.NewTimer(0) // timer to dump a non-responsive active peer
	<-timeout.C                 // timeout channel should be initially empty
	defer timeout.Stop()

997
	var ttl time.Duration
998
	getHeaders := func(from uint64) {
999
		request = time.Now()
1000 1001 1002

		ttl = d.requestTTL()
		timeout.Reset(ttl)
1003

1004
		if skeleton {
P
Péter Szilágyi 已提交
1005
			p.log.Trace("Fetching skeleton headers", "count", MaxHeaderFetch, "from", from)
1006
			go p.peer.RequestHeadersByNumber(from+uint64(MaxHeaderFetch)-1, MaxSkeletonSize, MaxHeaderFetch-1, false)
1007
		} else {
P
Péter Szilágyi 已提交
1008
			p.log.Trace("Fetching full headers", "count", MaxHeaderFetch, "from", from)
1009
			go p.peer.RequestHeadersByNumber(from, MaxHeaderFetch, 0, false)
1010
		}
1011
	}
1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
	getNextPivot := func() {
		pivoting = true
		request = time.Now()

		ttl = d.requestTTL()
		timeout.Reset(ttl)

		d.pivotLock.RLock()
		pivot := d.pivotHeader.Number.Uint64()
		d.pivotLock.RUnlock()

		p.log.Trace("Fetching next pivot header", "number", pivot+uint64(fsMinFullBlocks))
		go p.peer.RequestHeadersByNumber(pivot+uint64(fsMinFullBlocks), 2, fsMinFullBlocks-9, false) // move +64 when it's 2x64-8 deep
	}
1026
	// Start pulling the header chain skeleton until all is done
1027
	ancestor := from
1028 1029
	getHeaders(from)

1030
	mode := d.getMode()
1031 1032 1033
	for {
		select {
		case <-d.cancelCh:
1034
			return errCanceled
1035

1036
		case packet := <-d.headerCh:
1037
			// Make sure the active peer is giving us the skeleton headers
1038
			if packet.PeerId() != p.id {
1039
				log.Debug("Received skeleton from incorrect peer", "peer", packet.PeerId())
1040 1041
				break
			}
1042
			headerReqTimer.UpdateSince(request)
1043 1044
			timeout.Stop()

1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082
			// If the pivot is being checked, move if it became stale and run the real retrieval
			var pivot uint64

			d.pivotLock.RLock()
			if d.pivotHeader != nil {
				pivot = d.pivotHeader.Number.Uint64()
			}
			d.pivotLock.RUnlock()

			if pivoting {
				if packet.Items() == 2 {
					// Retrieve the headers and do some sanity checks, just in case
					headers := packet.(*headerPack).headers

					if have, want := headers[0].Number.Uint64(), pivot+uint64(fsMinFullBlocks); have != want {
						log.Warn("Peer sent invalid next pivot", "have", have, "want", want)
						return fmt.Errorf("%w: next pivot number %d != requested %d", errInvalidChain, have, want)
					}
					if have, want := headers[1].Number.Uint64(), pivot+2*uint64(fsMinFullBlocks)-8; have != want {
						log.Warn("Peer sent invalid pivot confirmer", "have", have, "want", want)
						return fmt.Errorf("%w: next pivot confirmer number %d != requested %d", errInvalidChain, have, want)
					}
					log.Warn("Pivot seemingly stale, moving", "old", pivot, "new", headers[0].Number)
					pivot = headers[0].Number.Uint64()

					d.pivotLock.Lock()
					d.pivotHeader = headers[0]
					d.pivotLock.Unlock()

					// Write out the pivot into the database so a rollback beyond
					// it will reenable fast sync and update the state root that
					// the state syncer will be downloading.
					rawdb.WriteLastPivotNumber(d.stateDB, pivot)
				}
				pivoting = false
				getHeaders(from)
				continue
			}
1083
			// If the skeleton's finished, pull any remaining head headers directly from the origin
1084
			if skeleton && packet.Items() == 0 {
1085 1086 1087 1088
				skeleton = false
				getHeaders(from)
				continue
			}
1089
			// If no more headers are inbound, notify the content fetchers and return
1090
			if packet.Items() == 0 {
1091 1092 1093 1094 1095 1096 1097 1098
				// Don't abort header fetches while the pivot is downloading
				if atomic.LoadInt32(&d.committed) == 0 && pivot <= from {
					p.log.Debug("No headers, waiting for pivot commit")
					select {
					case <-time.After(fsHeaderContCheck):
						getHeaders(from)
						continue
					case <-d.cancelCh:
1099
						return errCanceled
1100 1101 1102
					}
				}
				// Pivot done (or not in fast sync) and no more headers, terminate the process
P
Péter Szilágyi 已提交
1103
				p.log.Debug("No more headers available")
1104 1105 1106 1107
				select {
				case d.headerProcCh <- nil:
					return nil
				case <-d.cancelCh:
1108
					return errCanceled
1109
				}
1110
			}
1111
			headers := packet.(*headerPack).headers
1112

1113 1114
			// If we received a skeleton batch, resolve internals concurrently
			if skeleton {
1115
				filled, proced, err := d.fillHeaderSkeleton(from, headers)
1116
				if err != nil {
P
Péter Szilágyi 已提交
1117
					p.log.Debug("Skeleton chain invalid", "err", err)
1118
					return fmt.Errorf("%w: %v", errInvalidChain, err)
1119
				}
1120 1121
				headers = filled[proced:]
				from += uint64(proced)
1122 1123 1124 1125 1126 1127
			} else {
				// If we're closing in on the chain head, but haven't yet reached it, delay
				// the last few headers so mini reorgs on the head don't cause invalid hash
				// chain errors.
				if n := len(headers); n > 0 {
					// Retrieve the current head we're at
1128
					var head uint64
1129
					if mode == LightSync {
1130 1131 1132 1133 1134 1135 1136
						head = d.lightchain.CurrentHeader().Number.Uint64()
					} else {
						head = d.blockchain.CurrentFastBlock().NumberU64()
						if full := d.blockchain.CurrentBlock().NumberU64(); head < full {
							head = full
						}
					}
1137 1138 1139 1140 1141 1142
					// If the head is below the common ancestor, we're actually deduplicating
					// already existing chain segments, so use the ancestor as the fake head.
					// Otherwise we might end up delaying header deliveries pointlessly.
					if head < ancestor {
						head = ancestor
					}
1143 1144 1145 1146 1147 1148 1149 1150 1151
					// If the head is way older than this batch, delay the last few headers
					if head+uint64(reorgProtThreshold) < headers[n-1].Number.Uint64() {
						delay := reorgProtHeaderDelay
						if delay > n {
							delay = n
						}
						headers = headers[:n-delay]
					}
				}
1152
			}
1153
			// Insert all the new headers and fetch the next batch
1154
			if len(headers) > 0 {
P
Péter Szilágyi 已提交
1155
				p.log.Trace("Scheduling new headers", "count", len(headers), "from", from)
1156 1157 1158
				select {
				case d.headerProcCh <- headers:
				case <-d.cancelCh:
1159
					return errCanceled
1160 1161
				}
				from += uint64(len(headers))
1162 1163 1164 1165 1166 1167 1168 1169

				// If we're still skeleton filling fast sync, check pivot staleness
				// before continuing to the next skeleton filling
				if skeleton && pivot > 0 {
					getNextPivot()
				} else {
					getHeaders(from)
				}
1170 1171 1172 1173 1174 1175 1176 1177
			} else {
				// No headers delivered, or all of them being delayed, sleep a bit and retry
				p.log.Trace("All headers delayed, waiting")
				select {
				case <-time.After(fsHeaderContCheck):
					getHeaders(from)
					continue
				case <-d.cancelCh:
1178
					return errCanceled
1179
				}
1180
			}
1181 1182

		case <-timeout.C:
1183 1184 1185 1186 1187 1188
			if d.dropPeer == nil {
				// The dropPeer method is nil when `--copydb` is used for a local copy.
				// Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored
				p.log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", p.id)
				break
			}
1189
			// Header retrieval timed out, consider the peer bad and drop
P
Péter Szilágyi 已提交
1190
			p.log.Debug("Header request timed out", "elapsed", ttl)
1191
			headerTimeoutMeter.Mark(1)
1192 1193 1194
			d.dropPeer(p.id)

			// Finish the sync gracefully instead of dumping the gathered data though
1195
			for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
1196 1197 1198 1199
				select {
				case ch <- false:
				case <-d.cancelCh:
				}
1200
			}
1201 1202 1203 1204
			select {
			case d.headerProcCh <- nil:
			case <-d.cancelCh:
			}
1205
			return fmt.Errorf("%w: header request timed out", errBadPeer)
1206 1207 1208 1209
		}
	}
}

1210 1211
// fillHeaderSkeleton concurrently retrieves headers from all our available peers
// and maps them to the provided skeleton header chain.
1212 1213 1214 1215 1216
//
// 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.
//
Y
Yusup 已提交
1217
// The method returns the entire filled skeleton and also the number of headers
1218 1219
// already forwarded for processing.
func (d *Downloader) fillHeaderSkeleton(from uint64, skeleton []*types.Header) ([]*types.Header, int, error) {
1220
	log.Debug("Filling up skeleton", "from", from)
1221 1222 1223 1224 1225
	d.queue.ScheduleSkeleton(from, skeleton)

	var (
		deliver = func(packet dataPack) (int, error) {
			pack := packet.(*headerPack)
1226
			return d.queue.DeliverHeaders(pack.peerID, pack.headers, d.headerProcCh)
1227
		}
1228 1229 1230
		expire  = func() map[string]int { return d.queue.ExpireHeaders(d.requestTTL()) }
		reserve = func(p *peerConnection, count int) (*fetchRequest, bool, bool) {
			return d.queue.ReserveHeaders(p, count), false, false
1231
		}
1232 1233
		fetch    = func(p *peerConnection, req *fetchRequest) error { return p.FetchHeaders(req.From, MaxHeaderFetch) }
		capacity = func(p *peerConnection) int { return p.HeaderCapacity(d.requestRTT()) }
1234 1235 1236
		setIdle  = func(p *peerConnection, accepted int, deliveryTime time.Time) {
			p.SetHeadersIdle(accepted, deliveryTime)
		}
1237
	)
1238
	err := d.fetchParts(d.headerCh, deliver, d.queue.headerContCh, expire,
1239
		d.queue.PendingHeaders, d.queue.InFlightHeaders, reserve,
1240
		nil, fetch, d.queue.CancelHeaders, capacity, d.peers.HeaderIdlePeers, setIdle, "headers")
1241

1242
	log.Debug("Skeleton fill terminated", "err", err)
1243 1244 1245

	filled, proced := d.queue.RetrieveHeaders()
	return filled, proced, err
1246 1247
}

1248 1249 1250 1251
// 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 {
1252
	log.Debug("Downloading block bodies", "origin", from)
1253

1254
	var (
1255
		deliver = func(packet dataPack) (int, error) {
1256
			pack := packet.(*bodyPack)
1257
			return d.queue.DeliverBodies(pack.peerID, pack.transactions, pack.uncles)
1258
		}
1259
		expire   = func() map[string]int { return d.queue.ExpireBodies(d.requestTTL()) }
1260 1261
		fetch    = func(p *peerConnection, req *fetchRequest) error { return p.FetchBodies(req) }
		capacity = func(p *peerConnection) int { return p.BlockCapacity(d.requestRTT()) }
1262
		setIdle  = func(p *peerConnection, accepted int, deliveryTime time.Time) { p.SetBodiesIdle(accepted, deliveryTime) }
1263
	)
1264
	err := d.fetchParts(d.bodyCh, deliver, d.bodyWakeCh, expire,
1265
		d.queue.PendingBlocks, d.queue.InFlightBlocks, d.queue.ReserveBodies,
1266
		d.bodyFetchHook, fetch, d.queue.CancelBodies, capacity, d.peers.BodyIdlePeers, setIdle, "bodies")
1267

1268
	log.Debug("Block body download terminated", "err", err)
1269 1270 1271 1272 1273 1274 1275
	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 {
1276
	log.Debug("Downloading transaction receipts", "origin", from)
1277 1278

	var (
1279
		deliver = func(packet dataPack) (int, error) {
1280
			pack := packet.(*receiptPack)
1281
			return d.queue.DeliverReceipts(pack.peerID, pack.receipts)
1282
		}
1283
		expire   = func() map[string]int { return d.queue.ExpireReceipts(d.requestTTL()) }
1284 1285
		fetch    = func(p *peerConnection, req *fetchRequest) error { return p.FetchReceipts(req) }
		capacity = func(p *peerConnection) int { return p.ReceiptCapacity(d.requestRTT()) }
1286 1287 1288
		setIdle  = func(p *peerConnection, accepted int, deliveryTime time.Time) {
			p.SetReceiptsIdle(accepted, deliveryTime)
		}
1289
	)
1290
	err := d.fetchParts(d.receiptCh, deliver, d.receiptWakeCh, expire,
1291
		d.queue.PendingReceipts, d.queue.InFlightReceipts, d.queue.ReserveReceipts,
1292
		d.receiptFetchHook, fetch, d.queue.CancelReceipts, capacity, d.peers.ReceiptIdlePeers, setIdle, "receipts")
1293

1294
	log.Debug("Transaction receipt download terminated", "err", err)
1295 1296 1297 1298 1299 1300
	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.
1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318
//
// 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)
P
Péter Szilágyi 已提交
1319
//  - capacity:    network callback to retrieve the estimated type-specific bandwidth capacity of a peer (traffic shaping)
1320 1321
//  - 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)
1322
//  - kind:        textual label of the type being downloaded to display in log messages
1323
func (d *Downloader) fetchParts(deliveryCh chan dataPack, deliver func(dataPack) (int, error), wakeCh chan bool,
1324
	expire func() map[string]int, pending func() int, inFlight func() bool, reserve func(*peerConnection, int) (*fetchRequest, bool, bool),
1325
	fetchHook func([]*types.Header), fetch func(*peerConnection, *fetchRequest) error, cancel func(*fetchRequest), capacity func(*peerConnection) int,
1326
	idle func() ([]*peerConnection, int), setIdle func(*peerConnection, int, time.Time), kind string) error {
1327

1328
	// Create a ticker to detect expired retrieval tasks
1329 1330 1331 1332 1333
	ticker := time.NewTicker(100 * time.Millisecond)
	defer ticker.Stop()

	update := make(chan struct{}, 1)

1334
	// Prepare the queue and fetch block parts until the block header fetcher's done
1335 1336 1337 1338
	finished := false
	for {
		select {
		case <-d.cancelCh:
1339
			return errCanceled
1340

1341
		case packet := <-deliveryCh:
1342
			deliveryTime := time.Now()
1343 1344
			// If the peer was previously banned and failed to deliver its pack
			// in a reasonable time frame, ignore its message.
1345
			if peer := d.peers.Peer(packet.PeerId()); peer != nil {
1346 1347
				// Deliver the received chunk of data and check chain validity
				accepted, err := deliver(packet)
1348
				if errors.Is(err, errInvalidChain) {
1349
					return err
1350 1351 1352 1353
				}
				// 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.
1354
				if !errors.Is(err, errStaleDelivery) {
1355
					setIdle(peer, accepted, deliveryTime)
1356 1357 1358 1359
				}
				// Issue a log to the user to see what's going on
				switch {
				case err == nil && packet.Items() == 0:
P
Péter Szilágyi 已提交
1360
					peer.log.Trace("Requested data not delivered", "type", kind)
1361
				case err == nil:
P
Péter Szilágyi 已提交
1362
					peer.log.Trace("Delivered new batch of data", "type", kind, "count", packet.Stats())
1363
				default:
P
Péter Szilágyi 已提交
1364
					peer.log.Trace("Failed to deliver retrieved data", "type", kind, "err", err)
1365 1366 1367 1368 1369 1370 1371 1372
				}
			}
			// Blocks assembled, try to update the progress
			select {
			case update <- struct{}{}:
			default:
			}

1373
		case cont := <-wakeCh:
1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395
			// 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
			}
1396
			// Check for fetch request timeouts and demote the responsible peers
1397
			for pid, fails := range expire() {
1398
				if peer := d.peers.Peer(pid); peer != nil {
1399 1400 1401 1402 1403 1404 1405 1406
					// 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 {
P
Péter Szilágyi 已提交
1407
						peer.log.Trace("Data delivery timed out", "type", kind)
1408
						setIdle(peer, 0, time.Now())
1409
					} else {
P
Péter Szilágyi 已提交
1410
						peer.log.Debug("Stalling delivery, dropping", "type", kind)
1411

1412 1413 1414 1415 1416 1417
						if d.dropPeer == nil {
							// The dropPeer method is nil when `--copydb` is used for a local copy.
							// Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored
							peer.log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", pid)
						} else {
							d.dropPeer(pid)
1418 1419 1420 1421 1422 1423 1424 1425 1426 1427

							// If this peer was the master peer, abort sync immediately
							d.cancelLock.RLock()
							master := pid == d.cancelPeer
							d.cancelLock.RUnlock()

							if master {
								d.cancel()
								return errTimeout
							}
1428
						}
1429
					}
1430 1431
				}
			}
1432 1433
			// If there's nothing more to fetch, wait or terminate
			if pending() == 0 {
1434
				if !inFlight() && finished {
1435
					log.Debug("Data fetching completed", "type", kind)
1436 1437 1438 1439 1440
					return nil
				}
				break
			}
			// Send a download request to all idle peers, until throttled
1441
			progressed, throttled, running := false, false, inFlight()
1442
			idles, total := idle()
1443
			pendCount := pending()
1444
			for _, peer := range idles {
1445
				// Short circuit if throttling activated
1446
				if throttled {
1447 1448
					break
				}
1449
				// Short circuit if there is no more available task.
1450
				if pendCount = pending(); pendCount == 0 {
1451 1452
					break
				}
1453 1454
				// 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
1455
				// have them.
1456
				request, progress, throttle := reserve(peer, capacity(peer))
1457 1458
				if progress {
					progressed = true
1459
				}
1460 1461 1462 1463
				if throttle {
					throttled = true
					throttleCounter.Inc(1)
				}
1464 1465 1466
				if request == nil {
					continue
				}
1467
				if request.From > 0 {
P
Péter Szilágyi 已提交
1468
					peer.log.Trace("Requesting new batch of data", "type", kind, "from", request.From)
1469
				} else {
1470
					peer.log.Trace("Requesting new batch of data", "type", kind, "count", len(request.Headers), "from", request.Headers[0].Number)
1471
				}
1472
				// Fetch the chunk and make sure any errors return the hashes to the queue
1473 1474
				if fetchHook != nil {
					fetchHook(request.Headers)
1475
				}
1476
				if err := fetch(peer, request); err != nil {
1477 1478 1479 1480 1481
					// 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.
1482
					panic(fmt.Sprintf("%v: %s fetch assignment failed", peer, kind))
1483
				}
1484
				running = true
1485 1486 1487
			}
			// Make sure that we have peers available for fetching. If all peers have been tried
			// and all failed throw an error
1488
			if !progressed && !throttled && !running && len(idles) == total && pendCount > 0 {
1489 1490 1491 1492 1493 1494
				return errPeersUnavailable
			}
		}
	}
}

1495 1496 1497
// 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.
1498
func (d *Downloader) processHeaders(origin uint64, td *big.Int) error {
1499
	// Keep a count of uncertain headers to roll back
1500
	var (
1501
		rollback    uint64 // Zero means no rollback (fine as you can't unroll the genesis)
1502 1503 1504
		rollbackErr error
		mode        = d.getMode()
	)
1505
	defer func() {
1506
		if rollback > 0 {
1507
			lastHeader, lastFastBlock, lastBlock := d.lightchain.CurrentHeader().Number, common.Big0, common.Big0
1508
			if mode != LightSync {
1509 1510
				lastFastBlock = d.blockchain.CurrentFastBlock().Number()
				lastBlock = d.blockchain.CurrentBlock().Number()
1511
			}
1512 1513 1514 1515
			if err := d.lightchain.SetHead(rollback - 1); err != nil { // -1 to target the parent of the first uncertain block
				// We're already unwinding the stack, only print the error to make it more visible
				log.Error("Failed to roll back chain segment", "head", rollback-1, "err", err)
			}
1516
			curFastBlock, curBlock := common.Big0, common.Big0
1517
			if mode != LightSync {
1518 1519
				curFastBlock = d.blockchain.CurrentFastBlock().Number()
				curBlock = d.blockchain.CurrentBlock().Number()
1520
			}
1521
			log.Warn("Rolled back chain segment",
1522
				"header", fmt.Sprintf("%d->%d", lastHeader, d.lightchain.CurrentHeader().Number),
1523
				"fast", fmt.Sprintf("%d->%d", lastFastBlock, curFastBlock),
1524
				"block", fmt.Sprintf("%d->%d", lastBlock, curBlock), "reason", rollbackErr)
1525 1526 1527 1528 1529 1530 1531 1532
		}
	}()
	// Wait for batches of headers to process
	gotHeaders := false

	for {
		select {
		case <-d.cancelCh:
1533
			rollbackErr = errCanceled
1534
			return errCanceled
1535 1536 1537 1538 1539

		case headers := <-d.headerProcCh:
			// Terminate header processing if we synced up
			if len(headers) == 0 {
				// Notify everyone that headers are fully processed
1540
				for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
1541 1542 1543 1544 1545
					select {
					case ch <- false:
					case <-d.cancelCh:
					}
				}
1546 1547
				// If no headers were retrieved at all, the peer violated its TD promise that it had a
				// better chain compared to ours. The only exception is if its promised blocks were
1548
				// already imported by other means (e.g. fetcher):
1549 1550 1551 1552 1553 1554 1555 1556 1557
				//
				// 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
1558
				if mode != LightSync {
1559 1560
					head := d.blockchain.CurrentBlock()
					if !gotHeaders && td.Cmp(d.blockchain.GetTd(head.Hash(), head.NumberU64())) > 0 {
1561 1562
						return errStallingPeer
					}
1563 1564 1565 1566 1567 1568 1569 1570
				}
				// 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).
1571
				if mode == FastSync || mode == LightSync {
1572 1573
					head := d.lightchain.CurrentHeader()
					if td.Cmp(d.lightchain.GetTd(head.Hash(), head.Number.Uint64())) > 0 {
1574 1575 1576 1577
						return errStallingPeer
					}
				}
				// Disable any rollback and return
1578
				rollback = 0
1579 1580 1581 1582 1583 1584 1585 1586
				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:
1587
					rollbackErr = errCanceled
1588
					return errCanceled
1589 1590 1591 1592 1593 1594 1595 1596
				default:
				}
				// Select the next chunk of headers to import
				limit := maxHeadersProcess
				if limit > len(headers) {
					limit = len(headers)
				}
				chunk := headers[:limit]
1597

1598
				// In case of header only syncing, validate the chunk immediately
1599
				if mode == FastSync || mode == LightSync {
1600
					// If we're importing pure headers, verify based on their recentness
1601 1602 1603 1604 1605 1606 1607 1608
					var pivot uint64

					d.pivotLock.RLock()
					if d.pivotHeader != nil {
						pivot = d.pivotHeader.Number.Uint64()
					}
					d.pivotLock.RUnlock()

1609 1610 1611 1612
					frequency := fsHeaderCheckFrequency
					if chunk[len(chunk)-1].Number.Uint64()+uint64(fsHeaderForceVerify) > pivot {
						frequency = 1
					}
1613
					if n, err := d.lightchain.InsertHeaderChain(chunk, frequency); err != nil {
1614
						rollbackErr = err
1615 1616

						// If some headers were inserted, track them as uncertain
1617
						if (mode == FastSync || frequency > 1) && n > 0 && rollback == 0 {
1618
							rollback = chunk[0].Number.Uint64()
1619
						}
1620
						log.Warn("Invalid header encountered", "number", chunk[n].Number, "hash", chunk[n].Hash(), "parent", chunk[n].ParentHash, "err", err)
1621
						return fmt.Errorf("%w: %v", errInvalidChain, err)
1622
					}
1623
					// All verifications passed, track all headers within the alloted limits
1624 1625 1626 1627 1628 1629 1630
					if mode == FastSync {
						head := chunk[len(chunk)-1].Number.Uint64()
						if head-rollback > uint64(fsHeaderSafetyNet) {
							rollback = head - uint64(fsHeaderSafetyNet)
						} else {
							rollback = 1
						}
1631 1632 1633
					}
				}
				// Unless we're doing light chains, schedule the headers for associated content retrieval
1634
				if mode == FullSync || mode == FastSync {
1635 1636 1637 1638
					// 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:
1639
							rollbackErr = errCanceled
1640
							return errCanceled
1641 1642 1643 1644 1645 1646
						case <-time.After(time.Second):
						}
					}
					// Otherwise insert the headers for content retrieval
					inserts := d.queue.Schedule(chunk, origin)
					if len(inserts) != len(chunk) {
1647
						rollbackErr = fmt.Errorf("stale headers: len inserts %v len(chunk) %v", len(inserts), len(chunk))
1648
						return fmt.Errorf("%w: stale headers", errBadPeer)
1649 1650 1651 1652 1653
					}
				}
				headers = headers[limit:]
				origin += uint64(limit)
			}
1654 1655 1656 1657 1658 1659 1660
			// Update the highest block number we know if a higher one is found.
			d.syncStatsLock.Lock()
			if d.syncStatsChainHeight < origin {
				d.syncStatsChainHeight = origin - 1
			}
			d.syncStatsLock.Unlock()

1661
			// Signal the content downloaders of the availablility of new tasks
1662
			for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} {
1663 1664 1665 1666 1667 1668 1669 1670 1671
				select {
				case ch <- true:
				default:
				}
			}
		}
	}
}

1672 1673
// processFullSyncContent takes fetch results from the queue and imports them into the chain.
func (d *Downloader) processFullSyncContent() error {
1674
	for {
1675
		results := d.queue.Results(true)
1676
		if len(results) == 0 {
1677
			return nil
1678
		}
1679
		if d.chainInsertHook != nil {
1680
			d.chainInsertHook(results)
1681
		}
1682 1683 1684 1685 1686 1687 1688
		if err := d.importBlockResults(results); err != nil {
			return err
		}
	}
}

func (d *Downloader) importBlockResults(results []*fetchResult) error {
1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708
	// Check for any early termination requests
	if len(results) == 0 {
		return nil
	}
	select {
	case <-d.quitCh:
		return errCancelContentProcessing
	default:
	}
	// Retrieve the a batch of results to import
	first, last := results[0].Header, results[len(results)-1].Header
	log.Debug("Inserting downloaded chain", "items", len(results),
		"firstnum", first.Number, "firsthash", first.Hash(),
		"lastnum", last.Number, "lasthash", last.Hash(),
	)
	blocks := make([]*types.Block, len(results))
	for i, result := range results {
		blocks[i] = types.NewBlockWithHeader(result.Header).WithBody(result.Transactions, result.Uncles)
	}
	if index, err := d.blockchain.InsertChain(blocks); err != nil {
1709 1710 1711 1712 1713 1714 1715 1716 1717
		if index < len(results) {
			log.Debug("Downloaded item processing failed", "number", results[index].Header.Number, "hash", results[index].Header.Hash(), "err", err)
		} else {
			// The InsertChain method in blockchain.go will sometimes return an out-of-bounds index,
			// when it needs to preprocess blocks to import a sidechain.
			// The importer will put together a new list of blocks to import, which is a superset
			// of the blocks delivered from the downloader, and the indexing will be off.
			log.Debug("Downloaded item processing failed on sidechain import", "index", index, "err", err)
		}
1718
		return fmt.Errorf("%w: %v", errInvalidChain, err)
1719 1720 1721 1722 1723 1724
	}
	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.
1725
func (d *Downloader) processFastSyncContent() error {
1726 1727
	// Start syncing state of the reported head block. This should get us most of
	// the state of the pivot block.
1728 1729 1730 1731
	d.pivotLock.RLock()
	sync := d.syncState(d.pivotHeader.Root)
	d.pivotLock.RUnlock()

1732 1733 1734 1735 1736 1737 1738
	defer func() {
		// The `sync` object is replaced every time the pivot moves. We need to
		// defer close the very last active one, hence the lazy evaluation vs.
		// calling defer sync.Cancel() !!!
		sync.Cancel()
	}()

1739 1740
	closeOnErr := func(s *stateSync) {
		if err := s.Wait(); err != nil && err != errCancelStateFetch && err != errCanceled {
1741
			d.queue.Close() // wake up Results
1742
		}
1743 1744
	}
	go closeOnErr(sync)
1745

1746
	// To cater for moving pivot points, track the pivot block and subsequently
Y
Yusup 已提交
1747
	// accumulated download results separately.
1748 1749 1750 1751
	var (
		oldPivot *fetchResult   // Locked in pivot block, might change eventually
		oldTail  []*fetchResult // Downloaded content after the pivot
	)
1752
	for {
1753 1754 1755
		// Wait for the next batch of downloaded data to be available, and if the pivot
		// block became stale, move the goalpost
		results := d.queue.Results(oldPivot == nil) // Block if we're not monitoring pivot staleness
1756
		if len(results) == 0 {
1757 1758
			// If pivot sync is done, stop
			if oldPivot == nil {
1759
				return sync.Cancel()
1760 1761 1762 1763
			}
			// If sync failed, stop
			select {
			case <-d.cancelCh:
1764
				sync.Cancel()
1765
				return errCanceled
1766 1767
			default:
			}
1768 1769 1770 1771
		}
		if d.chainInsertHook != nil {
			d.chainInsertHook(results)
		}
1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785
		// If we haven't downloaded the pivot block yet, check pivot staleness
		// notifications from the header downloader
		d.pivotLock.RLock()
		pivot := d.pivotHeader
		d.pivotLock.RUnlock()

		if oldPivot == nil {
			if pivot.Root != sync.root {
				sync.Cancel()
				sync = d.syncState(pivot.Root)

				go closeOnErr(sync)
			}
		} else {
1786 1787 1788 1789
			results = append(append([]*fetchResult{oldPivot}, oldTail...), results...)
		}
		// Split around the pivot block and process the two sides via fast/full sync
		if atomic.LoadInt32(&d.committed) == 0 {
1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804
			latest := results[len(results)-1].Header
			// If the height is above the pivot block by 2 sets, it means the pivot
			// become stale in the network and it was garbage collected, move to a
			// new pivot.
			//
			// Note, we have `reorgProtHeaderDelay` number of blocks withheld, Those
			// need to be taken into account, otherwise we're detecting the pivot move
			// late and will drop peers due to unavailable state!!!
			if height := latest.Number.Uint64(); height >= pivot.Number.Uint64()+2*uint64(fsMinFullBlocks)-uint64(reorgProtHeaderDelay) {
				log.Warn("Pivot became stale, moving", "old", pivot.Number.Uint64(), "new", height-uint64(fsMinFullBlocks)+uint64(reorgProtHeaderDelay))
				pivot = results[len(results)-1-fsMinFullBlocks+reorgProtHeaderDelay].Header // must exist as lower old pivot is uncommitted

				d.pivotLock.Lock()
				d.pivotHeader = pivot
				d.pivotLock.Unlock()
1805 1806 1807

				// Write out the pivot into the database so a rollback beyond it will
				// reenable fast sync
1808
				rawdb.WriteLastPivotNumber(d.stateDB, pivot.Number.Uint64())
1809 1810
			}
		}
1811
		P, beforeP, afterP := splitAroundPivot(pivot.Number.Uint64(), results)
1812
		if err := d.commitFastSyncData(beforeP, sync); err != nil {
1813 1814 1815
			return err
		}
		if P != nil {
1816 1817
			// If new pivot block found, cancel old state retrieval and restart
			if oldPivot != P {
1818 1819
				sync.Cancel()
				sync = d.syncState(P.Header.Root)
1820

1821
				go closeOnErr(sync)
1822 1823 1824 1825
				oldPivot = P
			}
			// Wait for completion, occasionally checking for pivot staleness
			select {
1826 1827 1828
			case <-sync.done:
				if sync.err != nil {
					return sync.err
1829 1830 1831 1832 1833 1834 1835 1836 1837
				}
				if err := d.commitPivotBlock(P); err != nil {
					return err
				}
				oldPivot = nil

			case <-time.After(time.Second):
				oldTail = afterP
				continue
1838
			}
1839
		}
1840
		// Fast sync done, pivot commit done, full import
1841 1842 1843 1844 1845 1846 1847
		if err := d.importBlockResults(afterP); err != nil {
			return err
		}
	}
}

func splitAroundPivot(pivot uint64, results []*fetchResult) (p *fetchResult, before, after []*fetchResult) {
1848 1849 1850 1851 1852 1853 1854 1855
	if len(results) == 0 {
		return nil, nil, nil
	}
	if lastNum := results[len(results)-1].Header.Number.Uint64(); lastNum < pivot {
		// the pivot is somewhere in the future
		return nil, results, nil
	}
	// This can also be optimized, but only happens very seldom
1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870
	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 {
1871 1872 1873 1874 1875 1876 1877 1878 1879 1880
	// Check for any early termination requests
	if len(results) == 0 {
		return nil
	}
	select {
	case <-d.quitCh:
		return errCancelContentProcessing
	case <-stateSync.done:
		if err := stateSync.Wait(); err != nil {
			return err
1881
		}
1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895
	default:
	}
	// Retrieve the a batch of results to import
	first, last := results[0].Header, results[len(results)-1].Header
	log.Debug("Inserting fast-sync blocks", "items", len(results),
		"firstnum", first.Number, "firsthash", first.Hash(),
		"lastnumn", last.Number, "lasthash", last.Hash(),
	)
	blocks := make([]*types.Block, len(results))
	receipts := make([]types.Receipts, len(results))
	for i, result := range results {
		blocks[i] = types.NewBlockWithHeader(result.Header).WithBody(result.Transactions, result.Uncles)
		receipts[i] = result.Receipts
	}
1896
	if index, err := d.blockchain.InsertReceiptChain(blocks, receipts, d.ancientLimit); err != nil {
1897
		log.Debug("Downloaded item processing failed", "number", results[index].Header.Number, "hash", results[index].Header.Hash(), "err", err)
1898
		return fmt.Errorf("%w: %v", errInvalidChain, err)
1899 1900 1901 1902 1903
	}
	return nil
}

func (d *Downloader) commitPivotBlock(result *fetchResult) error {
1904 1905
	block := types.NewBlockWithHeader(result.Header).WithBody(result.Transactions, result.Uncles)
	log.Debug("Committing fast sync pivot as new head", "number", block.Number(), "hash", block.Hash())
1906 1907

	// Commit the pivot block as the new head, will require full sync from here on
1908
	if _, err := d.blockchain.InsertReceiptChain([]*types.Block{block}, []types.Receipts{result.Receipts}, d.ancientLimit); err != nil {
1909 1910
		return err
	}
1911
	if err := d.blockchain.FastSyncCommitHead(block.Hash()); err != nil {
1912
		return err
1913
	}
1914
	atomic.StoreInt32(&d.committed, 1)
1915 1916 1917 1918 1919 1920 1921 1922 1923

	// If we had a bloom filter for the state sync, deallocate it now. Note, we only
	// deallocate internally, but keep the empty wrapper. This ensures that if we do
	// a rollback after committing the pivot and restarting fast sync, we don't end
	// up using a nil bloom. Empty bloom is fine, it just returns that it does not
	// have the info we need, so reach down to the database instead.
	if d.stateBloom != nil {
		d.stateBloom.Close()
	}
1924
	return nil
1925 1926
}

L
Leif Jurvetson 已提交
1927
// DeliverHeaders injects a new batch of block headers received from a remote
1928
// node into the download schedule.
1929 1930
func (d *Downloader) DeliverHeaders(id string, headers []*types.Header) error {
	return d.deliver(d.headerCh, &headerPack{id, headers}, headerInMeter, headerDropMeter)
1931 1932 1933
}

// DeliverBodies injects a new batch of block bodies received from a remote node.
1934 1935
func (d *Downloader) DeliverBodies(id string, transactions [][]*types.Transaction, uncles [][]*types.Header) error {
	return d.deliver(d.bodyCh, &bodyPack{id, transactions, uncles}, bodyInMeter, bodyDropMeter)
1936 1937 1938
}

// DeliverReceipts injects a new batch of receipts received from a remote node.
1939 1940
func (d *Downloader) DeliverReceipts(id string, receipts [][]*types.Receipt) error {
	return d.deliver(d.receiptCh, &receiptPack{id, receipts}, receiptInMeter, receiptDropMeter)
1941 1942 1943
}

// DeliverNodeData injects a new batch of node state data received from a remote node.
1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
func (d *Downloader) DeliverNodeData(id string, data [][]byte) error {
	return d.deliver(d.stateCh, &statePack{id, data}, stateInMeter, stateDropMeter)
}

// DeliverSnapPacket is invoked from a peer's message handler when it transmits a
// data packet for the local node to consume.
func (d *Downloader) DeliverSnapPacket(peer *snap.Peer, packet snap.Packet) error {
	switch packet := packet.(type) {
	case *snap.AccountRangePacket:
		hashes, accounts, err := packet.Unpack()
		if err != nil {
			return err
		}
		return d.SnapSyncer.OnAccounts(peer, packet.ID, hashes, accounts, packet.Proof)

	case *snap.StorageRangesPacket:
		hashset, slotset := packet.Unpack()
		return d.SnapSyncer.OnStorage(peer, packet.ID, hashset, slotset, packet.Proof)

	case *snap.ByteCodesPacket:
		return d.SnapSyncer.OnByteCodes(peer, packet.ID, packet.Codes)

	case *snap.TrieNodesPacket:
		return d.SnapSyncer.OnTrieNodes(peer, packet.ID, packet.Nodes)

	default:
		return fmt.Errorf("unexpected snap packet type: %T", packet)
	}
1972 1973 1974
}

// deliver injects a new batch of data received from a remote node.
1975
func (d *Downloader) deliver(destCh chan dataPack, packet dataPack, inMeter, dropMeter metrics.Meter) (err error) {
1976
	// Update the delivery metrics for both good and failed deliveries
1977
	inMeter.Mark(int64(packet.Items()))
1978 1979
	defer func() {
		if err != nil {
1980
			dropMeter.Mark(int64(packet.Items()))
1981 1982 1983 1984 1985 1986
		}
	}()
	// Deliver or abort if the sync is canceled while queuing
	d.cancelLock.RLock()
	cancel := d.cancelCh
	d.cancelLock.RUnlock()
1987 1988 1989
	if cancel == nil {
		return errNoSyncActive
	}
1990
	select {
1991
	case destCh <- packet:
1992 1993 1994 1995
		return nil
	case <-cancel:
		return errNoSyncActive
	}
1996
}
1997 1998 1999 2000 2001 2002

// 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
2003
		rtt := time.Duration((1-qosTuningImpact)*float64(atomic.LoadUint64(&d.rttEstimate)) + qosTuningImpact*float64(d.peers.medianRTT()))
2004 2005 2006 2007 2008 2009 2010 2011
		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
2012
		log.Debug("Recalculated downloader QoS values", "rtt", rtt, "confidence", float64(conf)/1000000.0, "ttl", d.requestTTL())
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025
		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())
2026 2027 2028 2029
	if peers == 0 {
		// Ensure peer connectivity races don't catch us off guard
		return
	}
2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045
	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))
2046
	log.Debug("Relaxed downloader QoS values", "rtt", rtt, "confidence", float64(conf)/1000000.0, "ttl", d.requestTTL())
2047 2048 2049 2050 2051 2052 2053
}

// 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
Y
Yusup 已提交
2054
// be adapted to, but smaller ones are preferred (stabler download stream).
2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071
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
}