github.com/ebakus/go-ebakus@v1.0.5-0.20200520105415-dbccef9ec421/eth/downloader/downloader.go (about) 1 // Copyright 2019 The ebakus/go-ebakus Authors 2 // This file is part of the ebakus/go-ebakus library. 3 // 4 // The ebakus/go-ebakus library is free software: you can redistribute it and/or modify 5 // it under the terms of the GNU Lesser General Public License as published by 6 // the Free Software Foundation, either version 3 of the License, or 7 // (at your option) any later version. 8 // 9 // The ebakus/go-ebakus library is distributed in the hope that it will be useful, 10 // but WITHOUT ANY WARRANTY; without even the implied warranty of 11 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 // GNU Lesser General Public License for more details. 13 // 14 // You should have received a copy of the GNU Lesser General Public License 15 // along with the ebakus/go-ebakus library. If not, see <http://www.gnu.org/licenses/>. 16 17 // Package downloader contains the manual full chain synchronisation. 18 package downloader 19 20 import ( 21 "errors" 22 "fmt" 23 "math/big" 24 "sync" 25 "sync/atomic" 26 "time" 27 28 "github.com/ebakus/go-ebakus" 29 "github.com/ebakus/go-ebakus/common" 30 "github.com/ebakus/go-ebakus/core/rawdb" 31 "github.com/ebakus/go-ebakus/core/types" 32 "github.com/ebakus/go-ebakus/ethdb" 33 "github.com/ebakus/go-ebakus/event" 34 "github.com/ebakus/go-ebakus/log" 35 "github.com/ebakus/go-ebakus/metrics" 36 "github.com/ebakus/go-ebakus/params" 37 "github.com/ebakus/go-ebakus/trie" 38 ) 39 40 var ( 41 MaxHashFetch = 512 // Amount of hashes to be fetched per retrieval request 42 MaxBlockFetch = 128 // Amount of blocks to be fetched per retrieval request 43 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 MaxBodyFetch = 128 // Amount of block bodies to be fetched per retrieval request 46 MaxReceiptFetch = 256 // Amount of transaction receipts to allow fetching per request 47 MaxStateFetch = 384 // Amount of node state values to allow fetching per request 48 49 rttMinEstimate = 2 * time.Second // Minimum round-trip time to target for download requests 50 rttMaxEstimate = 20 * time.Second // Maximum round-trip time to target for download requests 51 rttMinConfidence = 0.1 // Worse confidence factor in our estimated RTT value 52 ttlScaling = 3 // Constant scaling factor for RTT -> TTL conversion 53 ttlLimit = time.Minute // Maximum TTL allowance to prevent reaching crazy timeouts 54 55 qosTuningPeers = 5 // Number of peers to tune based on (best peers) 56 qosConfidenceCap = 10 // Number of peers above which not to modify RTT confidence 57 qosTuningImpact = 0.25 // Impact that a new tuning target has on the previous value 58 59 maxQueuedHeaders = 32 * 1024 // [eth/62] Maximum number of headers to queue for import (DOS protection) 60 maxHeadersProcess = 2048 // Number of header download results to import at once into the chain 61 maxResultsProcess = 2048 // Number of content download results to import at once into the chain 62 maxForkAncestry uint64 = params.ImmutabilityThreshold // Maximum chain reorganisation (locally redeclared so tests can reduce it) 63 64 reorgProtThreshold = 48 // Threshold number of recent blocks to disable mini reorg protection 65 reorgProtHeaderDelay = 2 // Number of headers to delay delivering to cover mini reorgs 66 67 fsHeaderCheckFrequency = 100 // Verification frequency of the downloaded headers during fast sync 68 fsHeaderSafetyNet = 2048 // Number of headers to discard in case a chain violation is detected 69 fsHeaderForceVerify = 24 // Number of headers to verify before and after the pivot to accept it 70 fsHeaderContCheck = 3 * time.Second // Time interval to check for header continuations during state download 71 fsMinFullBlocks = 64 // Number of blocks to retrieve fully even in fast sync 72 ) 73 74 var ( 75 errBusy = errors.New("busy") 76 errUnknownPeer = errors.New("peer is unknown or unhealthy") 77 errBadPeer = errors.New("action from bad peer ignored") 78 errStallingPeer = errors.New("peer is stalling") 79 errUnsyncedPeer = errors.New("unsynced peer") 80 errNoPeers = errors.New("no peers to keep download active") 81 errTimeout = errors.New("timeout") 82 errEmptyHeaderSet = errors.New("empty header set by peer") 83 errPeersUnavailable = errors.New("no peers available or all tried for download") 84 errInvalidAncestor = errors.New("retrieved ancestor is invalid") 85 errInvalidChain = errors.New("retrieved hash chain is invalid") 86 errInvalidBody = errors.New("retrieved block body is invalid") 87 errInvalidReceipt = errors.New("retrieved receipt is invalid") 88 errCancelStateFetch = errors.New("state data download canceled (requested)") 89 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 doesn't speak recent enough protocol version (need version >= 62)") 93 ) 94 95 type Downloader struct { 96 // WARNING: The `rttEstimate` and `rttConfidence` fields are accessed atomically. 97 // On 32 bit platforms, only 64-bit aligned fields can be atomic. The struct is 98 // guaranteed to be so aligned, so take advantage of that. For more information, 99 // see https://golang.org/pkg/sync/atomic/#pkg-note-BUG. 100 rttEstimate uint64 // Round trip time to target for download requests 101 rttConfidence uint64 // Confidence in the estimated RTT (unit: millionths to allow atomic ops) 102 103 mode SyncMode // Synchronisation mode defining the strategy used (per sync cycle) 104 mux *event.TypeMux // Event multiplexer to announce sync operation events 105 106 checkpoint uint64 // Checkpoint block number to enforce head against (e.g. fast sync) 107 genesis uint64 // Genesis block number to limit sync to (e.g. light client CHT) 108 queue *queue // Scheduler for selecting the hashes to download 109 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 existence checks 113 114 // Statistics 115 syncStatsChainOrigin uint64 // Origin block number where syncing started at 116 syncStatsChainHeight uint64 // Highest block number known when syncing started 117 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 synchroniseMock func(id string, hash common.Hash) error // Replacement for synchronise during testing 128 synchronising int32 129 notified int32 130 committed int32 131 ancientLimit uint64 // The maximum block number which can be regarded as ancient data. 132 133 // Channels 134 headerCh chan dataPack // [eth/62] Channel receiving inbound block headers 135 bodyCh chan dataPack // [eth/62] Channel receiving inbound block bodies 136 receiptCh chan dataPack // [eth/63] Channel receiving inbound receipts 137 bodyWakeCh chan bool // [eth/62] Channel to signal the block body fetcher of new tasks 138 receiptWakeCh chan bool // [eth/63] Channel to signal the receipt fetcher of new tasks 139 headerProcCh chan []*types.Header // [eth/62] Channel to feed the header processor new tasks 140 141 // for stateFetcher 142 stateSyncStart chan *stateSync 143 trackStateReq chan *stateReq 144 stateCh chan dataPack // [eth/63] Channel receiving inbound node state data 145 146 // Cancellation and termination 147 cancelPeer string // Identifier of the peer currently being used as the master (cancel on drop) 148 cancelCh chan struct{} // Channel to cancel mid-flight syncs 149 cancelLock sync.RWMutex // Lock to protect the cancel channel and peer in delivers 150 cancelWg sync.WaitGroup // Make sure all fetcher goroutines have exited. 151 152 quitCh chan struct{} // Quit channel to signal termination 153 quitLock sync.RWMutex // Lock to prevent double closes 154 155 // Testing hooks 156 syncInitHook func(uint64, uint64) // Method to call upon initiating a new sync run 157 bodyFetchHook func([]*types.Header) // Method to call upon starting a block body fetch 158 receiptFetchHook func([]*types.Header) // Method to call upon starting a receipt fetch 159 chainInsertHook func([]*fetchResult) // Method to call upon inserting a chain of blocks (possibly in multiple invocations) 160 } 161 162 // LightChain encapsulates functions required to synchronise a light chain. 163 type LightChain interface { 164 // HasHeader verifies a header's presence in the local chain. 165 HasHeader(common.Hash, uint64) bool 166 167 // GetHeaderByHash retrieves a header from the local chain. 168 GetHeaderByHash(common.Hash) *types.Header 169 170 // CurrentHeader retrieves the head header from the local chain. 171 CurrentHeader() *types.Header 172 173 // InsertHeaderChain inserts a batch of headers into the local chain. 174 InsertHeaderChain([]*types.Header, int) (int, error) 175 176 // Rollback removes a few recently added elements from the local chain. 177 Rollback([]common.Hash) 178 } 179 180 // BlockChain encapsulates functions required to sync a (full or fast) blockchain. 181 type BlockChain interface { 182 LightChain 183 184 // HasBlock verifies a block's presence in the local chain. 185 HasBlock(common.Hash, uint64) bool 186 187 // HasFastBlock verifies a fast block's presence in the local chain. 188 HasFastBlock(common.Hash, uint64) bool 189 190 // GetBlockByHash retrieves a block from the local chain. 191 GetBlockByHash(common.Hash) *types.Block 192 193 // CurrentBlock retrieves the head block from the local chain. 194 CurrentBlock() *types.Block 195 196 // CurrentFastBlock retrieves the head fast block from the local chain. 197 CurrentFastBlock() *types.Block 198 199 // FastSyncCommitHead directly commits the head block to a certain entity. 200 FastSyncCommitHead(common.Hash) error 201 202 // InsertChain inserts a batch of blocks into the local chain. 203 InsertChain(types.Blocks) (int, error) 204 205 // InsertReceiptChain inserts a batch of receipts into the local chain. 206 InsertReceiptChain(types.Blocks, []types.Receipts, uint64) (int, error) 207 } 208 209 // New creates a new downloader to fetch hashes and blocks from remote peers. 210 func New(checkpoint uint64, stateDb ethdb.Database, stateBloom *trie.SyncBloom, mux *event.TypeMux, chain BlockChain, lightchain LightChain, dropPeer peerDropFn) *Downloader { 211 if lightchain == nil { 212 lightchain = chain 213 } 214 dl := &Downloader{ 215 stateDB: stateDb, 216 stateBloom: stateBloom, 217 mux: mux, 218 checkpoint: checkpoint, 219 queue: newQueue(), 220 peers: newPeerSet(), 221 rttEstimate: uint64(rttMaxEstimate), 222 rttConfidence: uint64(1000000), 223 blockchain: chain, 224 lightchain: lightchain, 225 dropPeer: dropPeer, 226 headerCh: make(chan dataPack, 1), 227 bodyCh: make(chan dataPack, 1), 228 receiptCh: make(chan dataPack, 1), 229 bodyWakeCh: make(chan bool, 1), 230 receiptWakeCh: make(chan bool, 1), 231 headerProcCh: make(chan []*types.Header, 1), 232 quitCh: make(chan struct{}), 233 stateCh: make(chan dataPack), 234 stateSyncStart: make(chan *stateSync), 235 syncStatsState: stateSyncStats{ 236 processed: rawdb.ReadFastTrieProgress(stateDb), 237 }, 238 trackStateReq: make(chan *stateReq), 239 } 240 go dl.qosTuner() 241 go dl.stateFetcher() 242 return dl 243 } 244 245 // Progress retrieves the synchronisation boundaries, specifically the origin 246 // block where synchronisation started at (may have failed/suspended); the block 247 // or header sync is currently at; and the latest known block which the sync targets. 248 // 249 // In addition, during the state download phase of fast synchronisation the number 250 // of processed and the total number of known states are also returned. Otherwise 251 // these are zero. 252 func (d *Downloader) Progress() ebakus.SyncProgress { 253 // Lock the current stats and return the progress 254 d.syncStatsLock.RLock() 255 defer d.syncStatsLock.RUnlock() 256 257 current := uint64(0) 258 switch { 259 case d.blockchain != nil && d.mode == FullSync: 260 current = d.blockchain.CurrentBlock().NumberU64() 261 case d.blockchain != nil && d.mode == FastSync: 262 current = d.blockchain.CurrentFastBlock().NumberU64() 263 case d.lightchain != nil: 264 current = d.lightchain.CurrentHeader().Number.Uint64() 265 default: 266 log.Error("Unknown downloader chain/mode combo", "light", d.lightchain != nil, "full", d.blockchain != nil, "mode", d.mode) 267 } 268 return ebakus.SyncProgress{ 269 StartingBlock: d.syncStatsChainOrigin, 270 CurrentBlock: current, 271 HighestBlock: d.syncStatsChainHeight, 272 PulledStates: d.syncStatsState.processed, 273 KnownStates: d.syncStatsState.processed + d.syncStatsState.pending, 274 } 275 } 276 277 // Synchronising returns whether the downloader is currently retrieving blocks. 278 func (d *Downloader) Synchronising() bool { 279 return atomic.LoadInt32(&d.synchronising) > 0 280 } 281 282 // RegisterPeer injects a new download peer into the set of block source to be 283 // used for fetching hashes and blocks from. 284 func (d *Downloader) RegisterPeer(id string, version int, peer Peer) error { 285 logger := log.New("peer", id) 286 logger.Trace("Registering sync peer") 287 if err := d.peers.Register(newPeerConnection(id, version, peer, logger)); err != nil { 288 logger.Error("Failed to register sync peer", "err", err) 289 return err 290 } 291 d.qosReduceConfidence() 292 293 return nil 294 } 295 296 // RegisterLightPeer injects a light client peer, wrapping it so it appears as a regular peer. 297 func (d *Downloader) RegisterLightPeer(id string, version int, peer LightPeer) error { 298 return d.RegisterPeer(id, version, &lightPeerWrapper{peer}) 299 } 300 301 // UnregisterPeer remove a peer from the known list, preventing any action from 302 // the specified peer. An effort is also made to return any pending fetches into 303 // the queue. 304 func (d *Downloader) UnregisterPeer(id string) error { 305 // Unregister the peer from the active peer set and revoke any fetch tasks 306 logger := log.New("peer", id) 307 logger.Trace("Unregistering sync peer") 308 if err := d.peers.Unregister(id); err != nil { 309 logger.Error("Failed to unregister sync peer", "err", err) 310 return err 311 } 312 d.queue.Revoke(id) 313 314 return nil 315 } 316 317 // Synchronise tries to sync up our local block chain with a remote peer, both 318 // adding various sanity checks as well as wrapping it with various log entries. 319 func (d *Downloader) Synchronise(id string, head common.Hash, td *big.Int, mode SyncMode) error { 320 err := d.synchronise(id, head, td, mode) 321 switch err { 322 case nil: 323 case errBusy, errCanceled: 324 325 case errTimeout, errBadPeer, errStallingPeer, errUnsyncedPeer, 326 errEmptyHeaderSet, errPeersUnavailable, errTooOld, 327 errInvalidAncestor, errInvalidChain: 328 log.Warn("Synchronisation failed, dropping peer", "peer", id, "err", err) 329 if d.dropPeer == nil { 330 // The dropPeer method is nil when `--copydb` is used for a local copy. 331 // Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored 332 log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", id) 333 } else { 334 d.dropPeer(id) 335 } 336 default: 337 log.Warn("Synchronisation failed, retrying", "err", err) 338 } 339 return err 340 } 341 342 // synchronise will select the peer and use it for synchronising. If an empty string is given 343 // it will use the best peer possible and synchronize if its TD is higher than our own. If any of the 344 // checks fail an error will be returned. This method is synchronous 345 func (d *Downloader) synchronise(id string, hash common.Hash, td *big.Int, mode SyncMode) error { 346 // Mock out the synchronisation if testing 347 if d.synchroniseMock != nil { 348 return d.synchroniseMock(id, hash) 349 } 350 // Make sure only one goroutine is ever allowed past this point at once 351 if !atomic.CompareAndSwapInt32(&d.synchronising, 0, 1) { 352 return errBusy 353 } 354 defer atomic.StoreInt32(&d.synchronising, 0) 355 356 // Post a user notification of the sync (only once per session) 357 if atomic.CompareAndSwapInt32(&d.notified, 0, 1) { 358 log.Info("Block synchronisation started") 359 } 360 // If we are already full syncing, but have a fast-sync bloom filter laying 361 // around, make sure it does't use memory any more. This is a special case 362 // when the user attempts to fast sync a new empty network. 363 if mode == FullSync && d.stateBloom != nil { 364 d.stateBloom.Close() 365 } 366 // Reset the queue, peer set and wake channels to clean any internal leftover state 367 d.queue.Reset() 368 d.peers.Reset() 369 370 for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} { 371 select { 372 case <-ch: 373 default: 374 } 375 } 376 for _, ch := range []chan dataPack{d.headerCh, d.bodyCh, d.receiptCh} { 377 for empty := false; !empty; { 378 select { 379 case <-ch: 380 default: 381 empty = true 382 } 383 } 384 } 385 for empty := false; !empty; { 386 select { 387 case <-d.headerProcCh: 388 default: 389 empty = true 390 } 391 } 392 // Create cancel channel for aborting mid-flight and mark the master peer 393 d.cancelLock.Lock() 394 d.cancelCh = make(chan struct{}) 395 d.cancelPeer = id 396 d.cancelLock.Unlock() 397 398 defer d.Cancel() // No matter what, we can't leave the cancel channel open 399 400 // Set the requested sync mode, unless it's forbidden 401 d.mode = mode 402 403 // Retrieve the origin peer and initiate the downloading process 404 p := d.peers.Peer(id) 405 if p == nil { 406 return errUnknownPeer 407 } 408 return d.syncWithPeer(p, hash, td) 409 } 410 411 // syncWithPeer starts a block synchronization based on the hash chain from the 412 // specified peer and head hash. 413 func (d *Downloader) syncWithPeer(p *peerConnection, hash common.Hash, td *big.Int) (err error) { 414 d.mux.Post(StartEvent{}) 415 defer func() { 416 // reset on error 417 if err != nil { 418 d.mux.Post(FailedEvent{err}) 419 } else { 420 latest := d.lightchain.CurrentHeader() 421 d.mux.Post(DoneEvent{latest}) 422 } 423 }() 424 if p.version < 62 { 425 return errTooOld 426 } 427 428 log.Info("Synchronising with the network", "peer", p.id, "eth", p.version, "head", hash, "td", td, "mode", d.mode) 429 defer func(start time.Time) { 430 log.Debug("Synchronisation terminated", "elapsed", common.PrettyDuration(time.Since(start))) 431 }(time.Now()) 432 433 // Look up the sync boundaries: the common ancestor and the target block 434 latest, err := d.fetchHeight(p) 435 if err != nil { 436 return err 437 } 438 height := latest.Number.Uint64() 439 440 origin, err := d.findAncestor(p, latest) 441 if err != nil { 442 return err 443 } 444 d.syncStatsLock.Lock() 445 if d.syncStatsChainHeight <= origin || d.syncStatsChainOrigin > origin { 446 d.syncStatsChainOrigin = origin 447 } 448 d.syncStatsChainHeight = height 449 d.syncStatsLock.Unlock() 450 451 // Ensure our origin point is below any fast sync pivot point 452 pivot := uint64(0) 453 if d.mode == FastSync { 454 if height <= uint64(fsMinFullBlocks) { 455 origin = 0 456 } else { 457 pivot = height - uint64(fsMinFullBlocks) 458 if pivot <= origin { 459 origin = pivot - 1 460 } 461 } 462 } 463 d.committed = 1 464 if d.mode == FastSync && pivot != 0 { 465 d.committed = 0 466 } 467 if d.mode == FastSync { 468 // Set the ancient data limitation. 469 // If we are running fast sync, all block data older than ancientLimit will be 470 // written to the ancient store. More recent data will be written to the active 471 // database and will wait for the freezer to migrate. 472 // 473 // If there is a checkpoint available, then calculate the ancientLimit through 474 // that. Otherwise calculate the ancient limit through the advertised height 475 // of the remote peer. 476 // 477 // The reason for picking checkpoint first is that a malicious peer can give us 478 // a fake (very high) height, forcing the ancient limit to also be very high. 479 // The peer would start to feed us valid blocks until head, resulting in all of 480 // the blocks might be written into the ancient store. A following mini-reorg 481 // could cause issues. 482 if d.checkpoint != 0 && d.checkpoint > maxForkAncestry+1 { 483 d.ancientLimit = d.checkpoint 484 } else if height > maxForkAncestry+1 { 485 d.ancientLimit = height - maxForkAncestry - 1 486 } 487 frozen, _ := d.stateDB.Ancients() // Ignore the error here since light client can also hit here. 488 // If a part of blockchain data has already been written into active store, 489 // disable the ancient style insertion explicitly. 490 if origin >= frozen && frozen != 0 { 491 d.ancientLimit = 0 492 log.Info("Disabling direct-ancient mode", "origin", origin, "ancient", frozen-1) 493 } else if d.ancientLimit > 0 { 494 log.Debug("Enabling direct-ancient mode", "ancient", d.ancientLimit) 495 } 496 // Rewind the ancient store and blockchain if reorg happens. 497 if origin+1 < frozen { 498 var hashes []common.Hash 499 for i := origin + 1; i < d.lightchain.CurrentHeader().Number.Uint64(); i++ { 500 hashes = append(hashes, rawdb.ReadCanonicalHash(d.stateDB, i)) 501 } 502 d.lightchain.Rollback(hashes) 503 } 504 } 505 // Initiate the sync using a concurrent header and content retrieval algorithm 506 d.queue.Prepare(origin+1, d.mode) 507 if d.syncInitHook != nil { 508 d.syncInitHook(origin, height) 509 } 510 fetchers := []func() error{ 511 func() error { return d.fetchHeaders(p, origin+1, pivot) }, // Headers are always retrieved 512 func() error { return d.fetchBodies(origin + 1) }, // Bodies are retrieved during normal and fast sync 513 func() error { return d.fetchReceipts(origin + 1) }, // Receipts are retrieved during fast sync 514 func() error { return d.processHeaders(origin+1, pivot, td) }, 515 } 516 if d.mode == FastSync { 517 fetchers = append(fetchers, func() error { return d.processFastSyncContent(latest) }) 518 } else if d.mode == FullSync { 519 fetchers = append(fetchers, d.processFullSyncContent) 520 } 521 return d.spawnSync(fetchers) 522 } 523 524 // spawnSync runs d.process and all given fetcher functions to completion in 525 // separate goroutines, returning the first error that appears. 526 func (d *Downloader) spawnSync(fetchers []func() error) error { 527 errc := make(chan error, len(fetchers)) 528 d.cancelWg.Add(len(fetchers)) 529 for _, fn := range fetchers { 530 fn := fn 531 go func() { defer d.cancelWg.Done(); errc <- fn() }() 532 } 533 // Wait for the first error, then terminate the others. 534 var err error 535 for i := 0; i < len(fetchers); i++ { 536 if i == len(fetchers)-1 { 537 // Close the queue when all fetchers have exited. 538 // This will cause the block processor to end when 539 // it has processed the queue. 540 d.queue.Close() 541 } 542 if err = <-errc; err != nil && err != errCanceled { 543 break 544 } 545 } 546 d.queue.Close() 547 d.Cancel() 548 return err 549 } 550 551 // cancel aborts all of the operations and resets the queue. However, cancel does 552 // not wait for the running download goroutines to finish. This method should be 553 // used when cancelling the downloads from inside the downloader. 554 func (d *Downloader) cancel() { 555 // Close the current cancel channel 556 d.cancelLock.Lock() 557 if d.cancelCh != nil { 558 select { 559 case <-d.cancelCh: 560 // Channel was already closed 561 default: 562 close(d.cancelCh) 563 } 564 } 565 d.cancelLock.Unlock() 566 } 567 568 // Cancel aborts all of the operations and waits for all download goroutines to 569 // finish before returning. 570 func (d *Downloader) Cancel() { 571 d.cancel() 572 d.cancelWg.Wait() 573 574 d.ancientLimit = 0 575 log.Debug("Reset ancient limit to zero") 576 } 577 578 // Terminate interrupts the downloader, canceling all pending operations. 579 // The downloader cannot be reused after calling Terminate. 580 func (d *Downloader) Terminate() { 581 // Close the termination channel (make sure double close is allowed) 582 d.quitLock.Lock() 583 select { 584 case <-d.quitCh: 585 default: 586 close(d.quitCh) 587 } 588 d.quitLock.Unlock() 589 590 // Cancel any pending download requests 591 d.Cancel() 592 } 593 594 // fetchHeight retrieves the head header of the remote peer to aid in estimating 595 // the total time a pending synchronisation would take. 596 func (d *Downloader) fetchHeight(p *peerConnection) (*types.Header, error) { 597 p.log.Debug("Retrieving remote chain height") 598 599 // Request the advertised remote head block and wait for the response 600 head, _ := p.peer.Head() 601 go p.peer.RequestHeadersByHash(head, 1, 0, false) 602 603 ttl := d.requestTTL() 604 timeout := time.After(ttl) 605 for { 606 select { 607 case <-d.cancelCh: 608 return nil, errCanceled 609 610 case packet := <-d.headerCh: 611 // Discard anything not from the origin peer 612 if packet.PeerId() != p.id { 613 log.Debug("Received headers from incorrect peer", "peer", packet.PeerId()) 614 break 615 } 616 // Make sure the peer actually gave something valid 617 headers := packet.(*headerPack).headers 618 if len(headers) != 1 { 619 p.log.Debug("Multiple headers for single request", "headers", len(headers)) 620 return nil, errBadPeer 621 } 622 head := headers[0] 623 if (d.mode == FastSync || d.mode == LightSync) && head.Number.Uint64() < d.checkpoint { 624 p.log.Warn("Remote head below checkpoint", "number", head.Number, "hash", head.Hash()) 625 return nil, errUnsyncedPeer 626 } 627 p.log.Debug("Remote head header identified", "number", head.Number, "hash", head.Hash()) 628 return head, nil 629 630 case <-timeout: 631 p.log.Debug("Waiting for head header timed out", "elapsed", ttl) 632 return nil, errTimeout 633 634 case <-d.bodyCh: 635 case <-d.receiptCh: 636 // Out of bounds delivery, ignore 637 } 638 } 639 } 640 641 // calculateRequestSpan calculates what headers to request from a peer when trying to determine the 642 // common ancestor. 643 // It returns parameters to be used for peer.RequestHeadersByNumber: 644 // from - starting block number 645 // count - number of headers to request 646 // skip - number of headers to skip 647 // and also returns 'max', the last block which is expected to be returned by the remote peers, 648 // given the (from,count,skip) 649 func calculateRequestSpan(remoteHeight, localHeight uint64) (int64, int, int, uint64) { 650 var ( 651 from int 652 count int 653 MaxCount = MaxHeaderFetch / 16 654 ) 655 // requestHead is the highest block that we will ask for. If requestHead is not offset, 656 // the highest block that we will get is 16 blocks back from head, which means we 657 // will fetch 14 or 15 blocks unnecessarily in the case the height difference 658 // between us and the peer is 1-2 blocks, which is most common 659 requestHead := int(remoteHeight) - 1 660 if requestHead < 0 { 661 requestHead = 0 662 } 663 // requestBottom is the lowest block we want included in the query 664 // Ideally, we want to include just below own head 665 requestBottom := int(localHeight - 1) 666 if requestBottom < 0 { 667 requestBottom = 0 668 } 669 totalSpan := requestHead - requestBottom 670 span := 1 + totalSpan/MaxCount 671 if span < 2 { 672 span = 2 673 } 674 if span > 16 { 675 span = 16 676 } 677 678 count = 1 + totalSpan/span 679 if count > MaxCount { 680 count = MaxCount 681 } 682 if count < 2 { 683 count = 2 684 } 685 from = requestHead - (count-1)*span 686 if from < 0 { 687 from = 0 688 } 689 max := from + (count-1)*span 690 return int64(from), count, span - 1, uint64(max) 691 } 692 693 // findAncestor tries to locate the common ancestor link of the local chain and 694 // a remote peers blockchain. In the general case when our node was in sync and 695 // on the correct chain, checking the top N links should already get us a match. 696 // In the rare scenario when we ended up on a long reorganisation (i.e. none of 697 // the head links match), we do a binary search to find the common ancestor. 698 func (d *Downloader) findAncestor(p *peerConnection, remoteHeader *types.Header) (uint64, error) { 699 // Figure out the valid ancestor range to prevent rewrite attacks 700 var ( 701 floor = int64(-1) 702 localHeight uint64 703 remoteHeight = remoteHeader.Number.Uint64() 704 ) 705 switch d.mode { 706 case FullSync: 707 localHeight = d.blockchain.CurrentBlock().NumberU64() 708 case FastSync: 709 localHeight = d.blockchain.CurrentFastBlock().NumberU64() 710 default: 711 localHeight = d.lightchain.CurrentHeader().Number.Uint64() 712 } 713 p.log.Debug("Looking for common ancestor", "local", localHeight, "remote", remoteHeight) 714 715 // Recap floor value for binary search 716 if localHeight >= maxForkAncestry { 717 // We're above the max reorg threshold, find the earliest fork point 718 floor = int64(localHeight - maxForkAncestry) 719 } 720 // If we're doing a light sync, ensure the floor doesn't go below the CHT, as 721 // all headers before that point will be missing. 722 if d.mode == LightSync { 723 // If we dont know the current CHT position, find it 724 if d.genesis == 0 { 725 header := d.lightchain.CurrentHeader() 726 for header != nil { 727 d.genesis = header.Number.Uint64() 728 if floor >= int64(d.genesis)-1 { 729 break 730 } 731 header = d.lightchain.GetHeaderByHash(header.ParentHash) 732 } 733 } 734 // We already know the "genesis" block number, cap floor to that 735 if floor < int64(d.genesis)-1 { 736 floor = int64(d.genesis) - 1 737 } 738 } 739 740 from, count, skip, max := calculateRequestSpan(remoteHeight, localHeight) 741 742 p.log.Trace("Span searching for common ancestor", "count", count, "from", from, "skip", skip) 743 go p.peer.RequestHeadersByNumber(uint64(from), count, skip, false) 744 745 // Wait for the remote response to the head fetch 746 number, hash := uint64(0), common.Hash{} 747 748 ttl := d.requestTTL() 749 timeout := time.After(ttl) 750 751 for finished := false; !finished; { 752 select { 753 case <-d.cancelCh: 754 return 0, errCanceled 755 756 case packet := <-d.headerCh: 757 // Discard anything not from the origin peer 758 if packet.PeerId() != p.id { 759 log.Debug("Received headers from incorrect peer", "peer", packet.PeerId()) 760 break 761 } 762 // Make sure the peer actually gave something valid 763 headers := packet.(*headerPack).headers 764 if len(headers) == 0 { 765 p.log.Warn("Empty head header set") 766 return 0, errEmptyHeaderSet 767 } 768 // Make sure the peer's reply conforms to the request 769 for i, header := range headers { 770 expectNumber := from + int64(i)*int64(skip+1) 771 if number := header.Number.Int64(); number != expectNumber { 772 p.log.Warn("Head headers broke chain ordering", "index", i, "requested", expectNumber, "received", number) 773 return 0, errInvalidChain 774 } 775 } 776 // Check if a common ancestor was found 777 finished = true 778 for i := len(headers) - 1; i >= 0; i-- { 779 // Skip any headers that underflow/overflow our requested set 780 if headers[i].Number.Int64() < from || headers[i].Number.Uint64() > max { 781 continue 782 } 783 // Otherwise check if we already know the header or not 784 h := headers[i].Hash() 785 n := headers[i].Number.Uint64() 786 787 var known bool 788 switch d.mode { 789 case FullSync: 790 known = d.blockchain.HasBlock(h, n) 791 case FastSync: 792 known = d.blockchain.HasFastBlock(h, n) 793 default: 794 known = d.lightchain.HasHeader(h, n) 795 } 796 if known { 797 number, hash = n, h 798 break 799 } 800 } 801 802 case <-timeout: 803 p.log.Debug("Waiting for head header timed out", "elapsed", ttl) 804 return 0, errTimeout 805 806 case <-d.bodyCh: 807 case <-d.receiptCh: 808 // Out of bounds delivery, ignore 809 } 810 } 811 // If the head fetch already found an ancestor, return 812 if hash != (common.Hash{}) { 813 if int64(number) <= floor { 814 p.log.Warn("Ancestor below allowance", "number", number, "hash", hash, "allowance", floor) 815 return 0, errInvalidAncestor 816 } 817 p.log.Debug("Found common ancestor", "number", number, "hash", hash) 818 return number, nil 819 } 820 // Ancestor not found, we need to binary search over our chain 821 start, end := uint64(0), remoteHeight 822 if floor > 0 { 823 start = uint64(floor) 824 } 825 p.log.Trace("Binary searching for common ancestor", "start", start, "end", end) 826 827 for start+1 < end { 828 // Split our chain interval in two, and request the hash to cross check 829 check := (start + end) / 2 830 831 ttl := d.requestTTL() 832 timeout := time.After(ttl) 833 834 go p.peer.RequestHeadersByNumber(check, 1, 0, false) 835 836 // Wait until a reply arrives to this request 837 for arrived := false; !arrived; { 838 select { 839 case <-d.cancelCh: 840 return 0, errCanceled 841 842 case packer := <-d.headerCh: 843 // Discard anything not from the origin peer 844 if packer.PeerId() != p.id { 845 log.Debug("Received headers from incorrect peer", "peer", packer.PeerId()) 846 break 847 } 848 // Make sure the peer actually gave something valid 849 headers := packer.(*headerPack).headers 850 if len(headers) != 1 { 851 p.log.Debug("Multiple headers for single request", "headers", len(headers)) 852 return 0, errBadPeer 853 } 854 arrived = true 855 856 // Modify the search interval based on the response 857 h := headers[0].Hash() 858 n := headers[0].Number.Uint64() 859 860 var known bool 861 switch d.mode { 862 case FullSync: 863 known = d.blockchain.HasBlock(h, n) 864 case FastSync: 865 known = d.blockchain.HasFastBlock(h, n) 866 default: 867 known = d.lightchain.HasHeader(h, n) 868 } 869 if !known { 870 end = check 871 break 872 } 873 header := d.lightchain.GetHeaderByHash(h) // Independent of sync mode, header surely exists 874 if header.Number.Uint64() != check { 875 p.log.Debug("Received non requested header", "number", header.Number, "hash", header.Hash(), "request", check) 876 return 0, errBadPeer 877 } 878 start = check 879 hash = h 880 881 case <-timeout: 882 p.log.Debug("Waiting for search header timed out", "elapsed", ttl) 883 return 0, errTimeout 884 885 case <-d.bodyCh: 886 case <-d.receiptCh: 887 // Out of bounds delivery, ignore 888 } 889 } 890 } 891 // Ensure valid ancestry and return 892 if int64(start) <= floor { 893 p.log.Warn("Ancestor below allowance", "number", start, "hash", hash, "allowance", floor) 894 return 0, errInvalidAncestor 895 } 896 p.log.Debug("Found common ancestor", "number", start, "hash", hash) 897 return start, nil 898 } 899 900 // fetchHeaders keeps retrieving headers concurrently from the number 901 // requested, until no more are returned, potentially throttling on the way. To 902 // facilitate concurrency but still protect against malicious nodes sending bad 903 // headers, we construct a header chain skeleton using the "origin" peer we are 904 // syncing with, and fill in the missing headers using anyone else. Headers from 905 // other peers are only accepted if they map cleanly to the skeleton. If no one 906 // can fill in the skeleton - not even the origin peer - it's assumed invalid and 907 // the origin is dropped. 908 func (d *Downloader) fetchHeaders(p *peerConnection, from uint64, pivot uint64) error { 909 p.log.Debug("Directing header downloads", "origin", from) 910 defer p.log.Debug("Header download terminated") 911 912 // Create a timeout timer, and the associated header fetcher 913 skeleton := true // Skeleton assembly phase or finishing up 914 request := time.Now() // time of the last skeleton fetch request 915 timeout := time.NewTimer(0) // timer to dump a non-responsive active peer 916 <-timeout.C // timeout channel should be initially empty 917 defer timeout.Stop() 918 919 var ttl time.Duration 920 getHeaders := func(from uint64) { 921 request = time.Now() 922 923 ttl = d.requestTTL() 924 timeout.Reset(ttl) 925 926 if skeleton { 927 p.log.Trace("Fetching skeleton headers", "count", MaxHeaderFetch, "from", from) 928 go p.peer.RequestHeadersByNumber(from+uint64(MaxHeaderFetch)-1, MaxSkeletonSize, MaxHeaderFetch-1, false) 929 } else { 930 p.log.Trace("Fetching full headers", "count", MaxHeaderFetch, "from", from) 931 go p.peer.RequestHeadersByNumber(from, MaxHeaderFetch, 0, false) 932 } 933 } 934 // Start pulling the header chain skeleton until all is done 935 ancestor := from 936 getHeaders(from) 937 938 for { 939 select { 940 case <-d.cancelCh: 941 return errCanceled 942 943 case packet := <-d.headerCh: 944 // Make sure the active peer is giving us the skeleton headers 945 if packet.PeerId() != p.id { 946 log.Debug("Received skeleton from incorrect peer", "peer", packet.PeerId()) 947 break 948 } 949 headerReqTimer.UpdateSince(request) 950 timeout.Stop() 951 952 // If the skeleton's finished, pull any remaining head headers directly from the origin 953 if packet.Items() == 0 && skeleton { 954 skeleton = false 955 getHeaders(from) 956 continue 957 } 958 // If no more headers are inbound, notify the content fetchers and return 959 if packet.Items() == 0 { 960 // Don't abort header fetches while the pivot is downloading 961 if atomic.LoadInt32(&d.committed) == 0 && pivot <= from { 962 p.log.Debug("No headers, waiting for pivot commit") 963 select { 964 case <-time.After(fsHeaderContCheck): 965 getHeaders(from) 966 continue 967 case <-d.cancelCh: 968 return errCanceled 969 } 970 } 971 // Pivot done (or not in fast sync) and no more headers, terminate the process 972 p.log.Debug("No more headers available") 973 select { 974 case d.headerProcCh <- nil: 975 return nil 976 case <-d.cancelCh: 977 return errCanceled 978 } 979 } 980 headers := packet.(*headerPack).headers 981 982 // If we received a skeleton batch, resolve internals concurrently 983 if skeleton { 984 filled, proced, err := d.fillHeaderSkeleton(from, headers) 985 if err != nil { 986 p.log.Debug("Skeleton chain invalid", "err", err) 987 return errInvalidChain 988 } 989 headers = filled[proced:] 990 from += uint64(proced) 991 } else { 992 // If we're closing in on the chain head, but haven't yet reached it, delay 993 // the last few headers so mini reorgs on the head don't cause invalid hash 994 // chain errors. 995 if n := len(headers); n > 0 { 996 // Retrieve the current head we're at 997 var head uint64 998 if d.mode == LightSync { 999 head = d.lightchain.CurrentHeader().Number.Uint64() 1000 } else { 1001 head = d.blockchain.CurrentFastBlock().NumberU64() 1002 if full := d.blockchain.CurrentBlock().NumberU64(); head < full { 1003 head = full 1004 } 1005 } 1006 // If the head is below the common ancestor, we're actually deduplicating 1007 // already existing chain segments, so use the ancestor as the fake head. 1008 // Otherwise we might end up delaying header deliveries pointlessly. 1009 if head < ancestor { 1010 head = ancestor 1011 } 1012 // If the head is way older than this batch, delay the last few headers 1013 if head+uint64(reorgProtThreshold) < headers[n-1].Number.Uint64() { 1014 delay := reorgProtHeaderDelay 1015 if delay > n { 1016 delay = n 1017 } 1018 headers = headers[:n-delay] 1019 } 1020 } 1021 } 1022 // Insert all the new headers and fetch the next batch 1023 if len(headers) > 0 { 1024 p.log.Trace("Scheduling new headers", "count", len(headers), "from", from) 1025 select { 1026 case d.headerProcCh <- headers: 1027 case <-d.cancelCh: 1028 return errCanceled 1029 } 1030 from += uint64(len(headers)) 1031 getHeaders(from) 1032 } else { 1033 // No headers delivered, or all of them being delayed, sleep a bit and retry 1034 p.log.Trace("All headers delayed, waiting") 1035 select { 1036 case <-time.After(fsHeaderContCheck): 1037 getHeaders(from) 1038 continue 1039 case <-d.cancelCh: 1040 return errCanceled 1041 } 1042 } 1043 1044 case <-timeout.C: 1045 if d.dropPeer == nil { 1046 // The dropPeer method is nil when `--copydb` is used for a local copy. 1047 // Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored 1048 p.log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", p.id) 1049 break 1050 } 1051 // Header retrieval timed out, consider the peer bad and drop 1052 p.log.Debug("Header request timed out", "elapsed", ttl) 1053 headerTimeoutMeter.Mark(1) 1054 d.dropPeer(p.id) 1055 1056 // Finish the sync gracefully instead of dumping the gathered data though 1057 for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} { 1058 select { 1059 case ch <- false: 1060 case <-d.cancelCh: 1061 } 1062 } 1063 select { 1064 case d.headerProcCh <- nil: 1065 case <-d.cancelCh: 1066 } 1067 return errBadPeer 1068 } 1069 } 1070 } 1071 1072 // fillHeaderSkeleton concurrently retrieves headers from all our available peers 1073 // and maps them to the provided skeleton header chain. 1074 // 1075 // Any partial results from the beginning of the skeleton is (if possible) forwarded 1076 // immediately to the header processor to keep the rest of the pipeline full even 1077 // in the case of header stalls. 1078 // 1079 // The method returns the entire filled skeleton and also the number of headers 1080 // already forwarded for processing. 1081 func (d *Downloader) fillHeaderSkeleton(from uint64, skeleton []*types.Header) ([]*types.Header, int, error) { 1082 log.Debug("Filling up skeleton", "from", from) 1083 d.queue.ScheduleSkeleton(from, skeleton) 1084 1085 var ( 1086 deliver = func(packet dataPack) (int, error) { 1087 pack := packet.(*headerPack) 1088 return d.queue.DeliverHeaders(pack.peerID, pack.headers, d.headerProcCh) 1089 } 1090 expire = func() map[string]int { return d.queue.ExpireHeaders(d.requestTTL()) } 1091 throttle = func() bool { return false } 1092 reserve = func(p *peerConnection, count int) (*fetchRequest, bool, error) { 1093 return d.queue.ReserveHeaders(p, count), false, nil 1094 } 1095 fetch = func(p *peerConnection, req *fetchRequest) error { return p.FetchHeaders(req.From, MaxHeaderFetch) } 1096 capacity = func(p *peerConnection) int { return p.HeaderCapacity(d.requestRTT()) } 1097 setIdle = func(p *peerConnection, accepted int) { p.SetHeadersIdle(accepted) } 1098 ) 1099 err := d.fetchParts(d.headerCh, deliver, d.queue.headerContCh, expire, 1100 d.queue.PendingHeaders, d.queue.InFlightHeaders, throttle, reserve, 1101 nil, fetch, d.queue.CancelHeaders, capacity, d.peers.HeaderIdlePeers, setIdle, "headers") 1102 1103 log.Debug("Skeleton fill terminated", "err", err) 1104 1105 filled, proced := d.queue.RetrieveHeaders() 1106 return filled, proced, err 1107 } 1108 1109 // fetchBodies iteratively downloads the scheduled block bodies, taking any 1110 // available peers, reserving a chunk of blocks for each, waiting for delivery 1111 // and also periodically checking for timeouts. 1112 func (d *Downloader) fetchBodies(from uint64) error { 1113 log.Debug("Downloading block bodies", "origin", from) 1114 1115 var ( 1116 deliver = func(packet dataPack) (int, error) { 1117 pack := packet.(*bodyPack) 1118 return d.queue.DeliverBodies(pack.peerID, pack.transactions) 1119 } 1120 expire = func() map[string]int { return d.queue.ExpireBodies(d.requestTTL()) } 1121 fetch = func(p *peerConnection, req *fetchRequest) error { return p.FetchBodies(req) } 1122 capacity = func(p *peerConnection) int { return p.BlockCapacity(d.requestRTT()) } 1123 setIdle = func(p *peerConnection, accepted int) { p.SetBodiesIdle(accepted) } 1124 ) 1125 err := d.fetchParts(d.bodyCh, deliver, d.bodyWakeCh, expire, 1126 d.queue.PendingBlocks, d.queue.InFlightBlocks, d.queue.ShouldThrottleBlocks, d.queue.ReserveBodies, 1127 d.bodyFetchHook, fetch, d.queue.CancelBodies, capacity, d.peers.BodyIdlePeers, setIdle, "bodies") 1128 1129 log.Debug("Block body download terminated", "err", err) 1130 return err 1131 } 1132 1133 // fetchReceipts iteratively downloads the scheduled block receipts, taking any 1134 // available peers, reserving a chunk of receipts for each, waiting for delivery 1135 // and also periodically checking for timeouts. 1136 func (d *Downloader) fetchReceipts(from uint64) error { 1137 log.Debug("Downloading transaction receipts", "origin", from) 1138 1139 var ( 1140 deliver = func(packet dataPack) (int, error) { 1141 pack := packet.(*receiptPack) 1142 return d.queue.DeliverReceipts(pack.peerID, pack.receipts) 1143 } 1144 expire = func() map[string]int { return d.queue.ExpireReceipts(d.requestTTL()) } 1145 fetch = func(p *peerConnection, req *fetchRequest) error { return p.FetchReceipts(req) } 1146 capacity = func(p *peerConnection) int { return p.ReceiptCapacity(d.requestRTT()) } 1147 setIdle = func(p *peerConnection, accepted int) { p.SetReceiptsIdle(accepted) } 1148 ) 1149 err := d.fetchParts(d.receiptCh, deliver, d.receiptWakeCh, expire, 1150 d.queue.PendingReceipts, d.queue.InFlightReceipts, d.queue.ShouldThrottleReceipts, d.queue.ReserveReceipts, 1151 d.receiptFetchHook, fetch, d.queue.CancelReceipts, capacity, d.peers.ReceiptIdlePeers, setIdle, "receipts") 1152 1153 log.Debug("Transaction receipt download terminated", "err", err) 1154 return err 1155 } 1156 1157 // fetchParts iteratively downloads scheduled block parts, taking any available 1158 // peers, reserving a chunk of fetch requests for each, waiting for delivery and 1159 // also periodically checking for timeouts. 1160 // 1161 // As the scheduling/timeout logic mostly is the same for all downloaded data 1162 // types, this method is used by each for data gathering and is instrumented with 1163 // various callbacks to handle the slight differences between processing them. 1164 // 1165 // The instrumentation parameters: 1166 // - errCancel: error type to return if the fetch operation is cancelled (mostly makes logging nicer) 1167 // - deliveryCh: channel from which to retrieve downloaded data packets (merged from all concurrent peers) 1168 // - deliver: processing callback to deliver data packets into type specific download queues (usually within `queue`) 1169 // - wakeCh: notification channel for waking the fetcher when new tasks are available (or sync completed) 1170 // - expire: task callback method to abort requests that took too long and return the faulty peers (traffic shaping) 1171 // - pending: task callback for the number of requests still needing download (detect completion/non-completability) 1172 // - inFlight: task callback for the number of in-progress requests (wait for all active downloads to finish) 1173 // - throttle: task callback to check if the processing queue is full and activate throttling (bound memory use) 1174 // - reserve: task callback to reserve new download tasks to a particular peer (also signals partial completions) 1175 // - fetchHook: tester callback to notify of new tasks being initiated (allows testing the scheduling logic) 1176 // - fetch: network callback to actually send a particular download request to a physical remote peer 1177 // - cancel: task callback to abort an in-flight download request and allow rescheduling it (in case of lost peer) 1178 // - capacity: network callback to retrieve the estimated type-specific bandwidth capacity of a peer (traffic shaping) 1179 // - idle: network callback to retrieve the currently (type specific) idle peers that can be assigned tasks 1180 // - setIdle: network callback to set a peer back to idle and update its estimated capacity (traffic shaping) 1181 // - kind: textual label of the type being downloaded to display in log mesages 1182 func (d *Downloader) fetchParts(deliveryCh chan dataPack, deliver func(dataPack) (int, error), wakeCh chan bool, 1183 expire func() map[string]int, pending func() int, inFlight func() bool, throttle func() bool, reserve func(*peerConnection, int) (*fetchRequest, bool, error), 1184 fetchHook func([]*types.Header), fetch func(*peerConnection, *fetchRequest) error, cancel func(*fetchRequest), capacity func(*peerConnection) int, 1185 idle func() ([]*peerConnection, int), setIdle func(*peerConnection, int), kind string) error { 1186 1187 // Create a ticker to detect expired retrieval tasks 1188 ticker := time.NewTicker(100 * time.Millisecond) 1189 defer ticker.Stop() 1190 1191 update := make(chan struct{}, 1) 1192 1193 // Prepare the queue and fetch block parts until the block header fetcher's done 1194 finished := false 1195 for { 1196 select { 1197 case <-d.cancelCh: 1198 return errCanceled 1199 1200 case packet := <-deliveryCh: 1201 // If the peer was previously banned and failed to deliver its pack 1202 // in a reasonable time frame, ignore its message. 1203 if peer := d.peers.Peer(packet.PeerId()); peer != nil { 1204 // Deliver the received chunk of data and check chain validity 1205 accepted, err := deliver(packet) 1206 if err == errInvalidChain { 1207 return err 1208 } 1209 // Unless a peer delivered something completely else than requested (usually 1210 // caused by a timed out request which came through in the end), set it to 1211 // idle. If the delivery's stale, the peer should have already been idled. 1212 if err != errStaleDelivery { 1213 setIdle(peer, accepted) 1214 } 1215 // Issue a log to the user to see what's going on 1216 switch { 1217 case err == nil && packet.Items() == 0: 1218 peer.log.Trace("Requested data not delivered", "type", kind) 1219 case err == nil: 1220 peer.log.Trace("Delivered new batch of data", "type", kind, "count", packet.Stats()) 1221 default: 1222 peer.log.Trace("Failed to deliver retrieved data", "type", kind, "err", err) 1223 } 1224 } 1225 // Blocks assembled, try to update the progress 1226 select { 1227 case update <- struct{}{}: 1228 default: 1229 } 1230 1231 case cont := <-wakeCh: 1232 // The header fetcher sent a continuation flag, check if it's done 1233 if !cont { 1234 finished = true 1235 } 1236 // Headers arrive, try to update the progress 1237 select { 1238 case update <- struct{}{}: 1239 default: 1240 } 1241 1242 case <-ticker.C: 1243 // Sanity check update the progress 1244 select { 1245 case update <- struct{}{}: 1246 default: 1247 } 1248 1249 case <-update: 1250 // Short circuit if we lost all our peers 1251 if d.peers.Len() == 0 { 1252 return errNoPeers 1253 } 1254 // Check for fetch request timeouts and demote the responsible peers 1255 for pid, fails := range expire() { 1256 if peer := d.peers.Peer(pid); peer != nil { 1257 // If a lot of retrieval elements expired, we might have overestimated the remote peer or perhaps 1258 // ourselves. Only reset to minimal throughput but don't drop just yet. If even the minimal times 1259 // out that sync wise we need to get rid of the peer. 1260 // 1261 // The reason the minimum threshold is 2 is because the downloader tries to estimate the bandwidth 1262 // and latency of a peer separately, which requires pushing the measures capacity a bit and seeing 1263 // how response times reacts, to it always requests one more than the minimum (i.e. min 2). 1264 if fails > 2 { 1265 peer.log.Trace("Data delivery timed out", "type", kind) 1266 setIdle(peer, 0) 1267 } else { 1268 peer.log.Debug("Stalling delivery, dropping", "type", kind) 1269 1270 if d.dropPeer == nil { 1271 // The dropPeer method is nil when `--copydb` is used for a local copy. 1272 // Timeouts can occur if e.g. compaction hits at the wrong time, and can be ignored 1273 peer.log.Warn("Downloader wants to drop peer, but peerdrop-function is not set", "peer", pid) 1274 } else { 1275 d.dropPeer(pid) 1276 1277 // If this peer was the master peer, abort sync immediately 1278 d.cancelLock.RLock() 1279 master := pid == d.cancelPeer 1280 d.cancelLock.RUnlock() 1281 1282 if master { 1283 d.cancel() 1284 return errTimeout 1285 } 1286 } 1287 } 1288 } 1289 } 1290 // If there's nothing more to fetch, wait or terminate 1291 if pending() == 0 { 1292 if !inFlight() && finished { 1293 log.Debug("Data fetching completed", "type", kind) 1294 return nil 1295 } 1296 break 1297 } 1298 // Send a download request to all idle peers, until throttled 1299 progressed, throttled, running := false, false, inFlight() 1300 idles, total := idle() 1301 1302 for _, peer := range idles { 1303 // Short circuit if throttling activated 1304 if throttle() { 1305 throttled = true 1306 break 1307 } 1308 // Short circuit if there is no more available task. 1309 if pending() == 0 { 1310 break 1311 } 1312 // Reserve a chunk of fetches for a peer. A nil can mean either that 1313 // no more headers are available, or that the peer is known not to 1314 // have them. 1315 request, progress, err := reserve(peer, capacity(peer)) 1316 if err != nil { 1317 return err 1318 } 1319 if progress { 1320 progressed = true 1321 } 1322 if request == nil { 1323 continue 1324 } 1325 if request.From > 0 { 1326 peer.log.Trace("Requesting new batch of data", "type", kind, "from", request.From) 1327 } else { 1328 peer.log.Trace("Requesting new batch of data", "type", kind, "count", len(request.Headers), "from", request.Headers[0].Number) 1329 } 1330 // Fetch the chunk and make sure any errors return the hashes to the queue 1331 if fetchHook != nil { 1332 fetchHook(request.Headers) 1333 } 1334 if err := fetch(peer, request); err != nil { 1335 // Although we could try and make an attempt to fix this, this error really 1336 // means that we've double allocated a fetch task to a peer. If that is the 1337 // case, the internal state of the downloader and the queue is very wrong so 1338 // better hard crash and note the error instead of silently accumulating into 1339 // a much bigger issue. 1340 panic(fmt.Sprintf("%v: %s fetch assignment failed", peer, kind)) 1341 } 1342 running = true 1343 } 1344 // Make sure that we have peers available for fetching. If all peers have been tried 1345 // and all failed throw an error 1346 if !progressed && !throttled && !running && len(idles) == total && pending() > 0 { 1347 return errPeersUnavailable 1348 } 1349 } 1350 } 1351 } 1352 1353 // processHeaders takes batches of retrieved headers from an input channel and 1354 // keeps processing and scheduling them into the header chain and downloader's 1355 // queue until the stream ends or a failure occurs. 1356 func (d *Downloader) processHeaders(origin uint64, pivot uint64, td *big.Int) error { 1357 // Keep a count of uncertain headers to roll back 1358 var rollback []*types.Header 1359 defer func() { 1360 if len(rollback) > 0 { 1361 // Flatten the headers and roll them back 1362 hashes := make([]common.Hash, len(rollback)) 1363 for i, header := range rollback { 1364 hashes[i] = header.Hash() 1365 } 1366 lastHeader, lastFastBlock, lastBlock := d.lightchain.CurrentHeader().Number, common.Big0, common.Big0 1367 if d.mode != LightSync { 1368 lastFastBlock = d.blockchain.CurrentFastBlock().Number() 1369 lastBlock = d.blockchain.CurrentBlock().Number() 1370 } 1371 d.lightchain.Rollback(hashes) 1372 curFastBlock, curBlock := common.Big0, common.Big0 1373 if d.mode != LightSync { 1374 curFastBlock = d.blockchain.CurrentFastBlock().Number() 1375 curBlock = d.blockchain.CurrentBlock().Number() 1376 } 1377 log.Warn("Rolled back headers", "count", len(hashes), 1378 "header", fmt.Sprintf("%d->%d", lastHeader, d.lightchain.CurrentHeader().Number), 1379 "fast", fmt.Sprintf("%d->%d", lastFastBlock, curFastBlock), 1380 "block", fmt.Sprintf("%d->%d", lastBlock, curBlock)) 1381 } 1382 }() 1383 1384 // Wait for batches of headers to process 1385 gotHeaders := false 1386 1387 for { 1388 select { 1389 case <-d.cancelCh: 1390 return errCanceled 1391 1392 case headers := <-d.headerProcCh: 1393 // Terminate header processing if we synced up 1394 if len(headers) == 0 { 1395 // Notify everyone that headers are fully processed 1396 for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} { 1397 select { 1398 case ch <- false: 1399 case <-d.cancelCh: 1400 } 1401 } 1402 // If no headers were retrieved at all, the peer violated its TD promise that it had a 1403 // better chain compared to ours. The only exception is if its promised blocks were 1404 // already imported by other means (e.g. fetcher): 1405 // 1406 // R <remote peer>, L <local node>: Both at block 10 1407 // R: Mine block 11, and propagate it to L 1408 // L: Queue block 11 for import 1409 // L: Notice that R's head and TD increased compared to ours, start sync 1410 // L: Import of block 11 finishes 1411 // L: Sync begins, and finds common ancestor at 11 1412 // L: Request new headers up from 11 (R's TD was higher, it must have something) 1413 // R: Nothing to give 1414 if d.mode != LightSync { 1415 if !gotHeaders && td.Cmp(d.blockchain.CurrentBlock().Number()) > 0 { 1416 return errStallingPeer 1417 } 1418 } 1419 // If fast or light syncing, ensure promised headers are indeed delivered. This is 1420 // needed to detect scenarios where an attacker feeds a bad pivot and then bails out 1421 // of delivering the post-pivot blocks that would flag the invalid content. 1422 // 1423 // This check cannot be executed "as is" for full imports, since blocks may still be 1424 // queued for processing when the header download completes. However, as long as the 1425 // peer gave us something useful, we're already happy/progressed (above check). 1426 if d.mode == FastSync || d.mode == LightSync { 1427 if td.Cmp(d.lightchain.CurrentHeader().Number) > 0 { 1428 return errStallingPeer 1429 } 1430 } 1431 // Disable any rollback and return 1432 rollback = nil 1433 return nil 1434 } 1435 // Otherwise split the chunk of headers into batches and process them 1436 gotHeaders = true 1437 for len(headers) > 0 { 1438 // Terminate if something failed in between processing chunks 1439 select { 1440 case <-d.cancelCh: 1441 return errCanceled 1442 default: 1443 } 1444 // Select the next chunk of headers to import 1445 limit := maxHeadersProcess 1446 if limit > len(headers) { 1447 limit = len(headers) 1448 } 1449 chunk := headers[:limit] 1450 // In case of header only syncing, validate the chunk immediately 1451 if d.mode == FastSync || d.mode == LightSync { 1452 // Collect the yet unknown headers to mark them as uncertain 1453 unknown := make([]*types.Header, 0, len(chunk)) 1454 for _, header := range chunk { 1455 if !d.lightchain.HasHeader(header.Hash(), header.Number.Uint64()) { 1456 unknown = append(unknown, header) 1457 } 1458 } 1459 // If we're importing pure headers, verify based on their recentness 1460 frequency := fsHeaderCheckFrequency 1461 if chunk[len(chunk)-1].Number.Uint64()+uint64(fsHeaderForceVerify) > pivot { 1462 frequency = 1 1463 } 1464 if n, err := d.lightchain.InsertHeaderChain(chunk, frequency); err != nil { 1465 // If some headers were inserted, add them too to the rollback list 1466 if n > 0 { 1467 rollback = append(rollback, chunk[:n]...) 1468 } 1469 log.Debug("Invalid header encountered", "number", chunk[n].Number, "hash", chunk[n].Hash(), "err", err) 1470 return errInvalidChain 1471 } 1472 // All verifications passed, store newly found uncertain headers 1473 rollback = append(rollback, unknown...) 1474 if len(rollback) > fsHeaderSafetyNet { 1475 rollback = append(rollback[:0], rollback[len(rollback)-fsHeaderSafetyNet:]...) 1476 } 1477 } 1478 // Unless we're doing light chains, schedule the headers for associated content retrieval 1479 if d.mode == FullSync || d.mode == FastSync { 1480 // If we've reached the allowed number of pending headers, stall a bit 1481 for d.queue.PendingBlocks() >= maxQueuedHeaders || d.queue.PendingReceipts() >= maxQueuedHeaders { 1482 select { 1483 case <-d.cancelCh: 1484 return errCanceled 1485 case <-time.After(time.Second): 1486 } 1487 } 1488 // Otherwise insert the headers for content retrieval 1489 inserts := d.queue.Schedule(chunk, origin) 1490 if len(inserts) != len(chunk) { 1491 log.Debug("Stale headers") 1492 return errBadPeer 1493 } 1494 } 1495 headers = headers[limit:] 1496 origin += uint64(limit) 1497 } 1498 // Update the highest block number we know if a higher one is found. 1499 d.syncStatsLock.Lock() 1500 if d.syncStatsChainHeight < origin { 1501 d.syncStatsChainHeight = origin - 1 1502 } 1503 d.syncStatsLock.Unlock() 1504 1505 // Signal the content downloaders of the availablility of new tasks 1506 for _, ch := range []chan bool{d.bodyWakeCh, d.receiptWakeCh} { 1507 select { 1508 case ch <- true: 1509 default: 1510 } 1511 } 1512 } 1513 } 1514 } 1515 1516 // processFullSyncContent takes fetch results from the queue and imports them into the chain. 1517 func (d *Downloader) processFullSyncContent() error { 1518 for { 1519 results := d.queue.Results(true) 1520 if len(results) == 0 { 1521 return nil 1522 } 1523 if d.chainInsertHook != nil { 1524 d.chainInsertHook(results) 1525 } 1526 if err := d.importBlockResults(results); err != nil { 1527 return err 1528 } 1529 } 1530 } 1531 1532 func (d *Downloader) importBlockResults(results []*fetchResult) error { 1533 // Check for any early termination requests 1534 if len(results) == 0 { 1535 return nil 1536 } 1537 select { 1538 case <-d.quitCh: 1539 return errCancelContentProcessing 1540 default: 1541 } 1542 // Retrieve the a batch of results to import 1543 first, last := results[0].Header, results[len(results)-1].Header 1544 log.Debug("Inserting downloaded chain", "items", len(results), 1545 "firstnum", first.Number, "firsthash", first.Hash(), 1546 "lastnum", last.Number, "lasthash", last.Hash(), 1547 ) 1548 blocks := make([]*types.Block, len(results)) 1549 for i, result := range results { 1550 blocks[i] = types.NewBlockWithHeader(result.Header).WithBody(result.Transactions) 1551 } 1552 if index, err := d.blockchain.InsertChain(blocks); err != nil { 1553 if index < len(results) { 1554 log.Debug("Downloaded item processing failed", "number", results[index].Header.Number, "hash", results[index].Header.Hash(), "err", err) 1555 } else { 1556 // The InsertChain method in blockchain.go will sometimes return an out-of-bounds index, 1557 // when it needs to preprocess blocks to import a sidechain. 1558 // The importer will put together a new list of blocks to import, which is a superset 1559 // of the blocks delivered from the downloader, and the indexing will be off. 1560 log.Debug("Downloaded item processing failed on sidechain import", "index", index, "err", err) 1561 } 1562 return errInvalidChain 1563 } 1564 return nil 1565 } 1566 1567 // processFastSyncContent takes fetch results from the queue and writes them to the 1568 // database. It also controls the synchronisation of state nodes of the pivot block. 1569 func (d *Downloader) processFastSyncContent(latest *types.Header) error { 1570 // Start syncing state of the reported head block. This should get us most of 1571 // the state of the pivot block. 1572 sync := d.syncState(latest.Root) 1573 defer sync.Cancel() 1574 closeOnErr := func(s *stateSync) { 1575 if err := s.Wait(); err != nil && err != errCancelStateFetch && err != errCanceled { 1576 d.queue.Close() // wake up Results 1577 } 1578 } 1579 go closeOnErr(sync) 1580 // Figure out the ideal pivot block. Note, that this goalpost may move if the 1581 // sync takes long enough for the chain head to move significantly. 1582 pivot := uint64(0) 1583 if height := latest.Number.Uint64(); height > uint64(fsMinFullBlocks) { 1584 pivot = height - uint64(fsMinFullBlocks) 1585 } 1586 // To cater for moving pivot points, track the pivot block and subsequently 1587 // accumulated download results separately. 1588 var ( 1589 oldPivot *fetchResult // Locked in pivot block, might change eventually 1590 oldTail []*fetchResult // Downloaded content after the pivot 1591 ) 1592 for { 1593 // Wait for the next batch of downloaded data to be available, and if the pivot 1594 // block became stale, move the goalpost 1595 results := d.queue.Results(oldPivot == nil) // Block if we're not monitoring pivot staleness 1596 if len(results) == 0 { 1597 // If pivot sync is done, stop 1598 if oldPivot == nil { 1599 return sync.Cancel() 1600 } 1601 // If sync failed, stop 1602 select { 1603 case <-d.cancelCh: 1604 sync.Cancel() 1605 return errCanceled 1606 default: 1607 } 1608 } 1609 if d.chainInsertHook != nil { 1610 d.chainInsertHook(results) 1611 } 1612 if oldPivot != nil { 1613 results = append(append([]*fetchResult{oldPivot}, oldTail...), results...) 1614 } 1615 // Split around the pivot block and process the two sides via fast/full sync 1616 if atomic.LoadInt32(&d.committed) == 0 { 1617 latest = results[len(results)-1].Header 1618 if height := latest.Number.Uint64(); height > pivot+2*uint64(fsMinFullBlocks) { 1619 log.Warn("Pivot became stale, moving", "old", pivot, "new", height-uint64(fsMinFullBlocks)) 1620 pivot = height - uint64(fsMinFullBlocks) 1621 } 1622 } 1623 P, beforeP, afterP := splitAroundPivot(pivot, results) 1624 if err := d.commitFastSyncData(beforeP, sync); err != nil { 1625 return err 1626 } 1627 if P != nil { 1628 // If new pivot block found, cancel old state retrieval and restart 1629 if oldPivot != P { 1630 sync.Cancel() 1631 1632 sync = d.syncState(P.Header.Root) 1633 defer sync.Cancel() 1634 go closeOnErr(sync) 1635 oldPivot = P 1636 } 1637 // Wait for completion, occasionally checking for pivot staleness 1638 select { 1639 case <-sync.done: 1640 if sync.err != nil { 1641 return sync.err 1642 } 1643 if err := d.commitPivotBlock(P); err != nil { 1644 return err 1645 } 1646 oldPivot = nil 1647 1648 case <-time.After(time.Second): 1649 oldTail = afterP 1650 continue 1651 } 1652 } 1653 // Fast sync done, pivot commit done, full import 1654 if err := d.importBlockResults(afterP); err != nil { 1655 return err 1656 } 1657 } 1658 } 1659 1660 func splitAroundPivot(pivot uint64, results []*fetchResult) (p *fetchResult, before, after []*fetchResult) { 1661 for _, result := range results { 1662 num := result.Header.Number.Uint64() 1663 switch { 1664 case num < pivot: 1665 before = append(before, result) 1666 case num == pivot: 1667 p = result 1668 default: 1669 after = append(after, result) 1670 } 1671 } 1672 return p, before, after 1673 } 1674 1675 func (d *Downloader) commitFastSyncData(results []*fetchResult, stateSync *stateSync) error { 1676 // Check for any early termination requests 1677 if len(results) == 0 { 1678 return nil 1679 } 1680 select { 1681 case <-d.quitCh: 1682 return errCancelContentProcessing 1683 case <-stateSync.done: 1684 if err := stateSync.Wait(); err != nil { 1685 return err 1686 } 1687 default: 1688 } 1689 // Retrieve the a batch of results to import 1690 first, last := results[0].Header, results[len(results)-1].Header 1691 log.Debug("Inserting fast-sync blocks", "items", len(results), 1692 "firstnum", first.Number, "firsthash", first.Hash(), 1693 "lastnumn", last.Number, "lasthash", last.Hash(), 1694 ) 1695 blocks := make([]*types.Block, len(results)) 1696 receipts := make([]types.Receipts, len(results)) 1697 for i, result := range results { 1698 blocks[i] = types.NewBlockWithHeader(result.Header).WithBody(result.Transactions) 1699 receipts[i] = result.Receipts 1700 } 1701 if index, err := d.blockchain.InsertReceiptChain(blocks, receipts, d.ancientLimit); err != nil { 1702 log.Debug("Downloaded item processing failed", "number", results[index].Header.Number, "hash", results[index].Header.Hash(), "err", err) 1703 return errInvalidChain 1704 } 1705 return nil 1706 } 1707 1708 func (d *Downloader) commitPivotBlock(result *fetchResult) error { 1709 block := types.NewBlockWithHeader(result.Header).WithBody(result.Transactions) 1710 log.Debug("Committing fast sync pivot as new head", "number", block.Number(), "hash", block.Hash()) 1711 1712 // Commit the pivot block as the new head, will require full sync from here on 1713 if _, err := d.blockchain.InsertReceiptChain([]*types.Block{block}, []types.Receipts{result.Receipts}, d.ancientLimit); err != nil { 1714 return err 1715 } 1716 if err := d.blockchain.FastSyncCommitHead(block.Hash()); err != nil { 1717 return err 1718 } 1719 atomic.StoreInt32(&d.committed, 1) 1720 1721 // If we had a bloom filter for the state sync, deallocate it now. Note, we only 1722 // deallocate internally, but keep the empty wrapper. This ensures that if we do 1723 // a rollback after committing the pivot and restarting fast sync, we don't end 1724 // up using a nil bloom. Empty bloom is fine, it just returns that it does not 1725 // have the info we need, so reach down to the database instead. 1726 if d.stateBloom != nil { 1727 d.stateBloom.Close() 1728 } 1729 return nil 1730 } 1731 1732 // DeliverHeaders injects a new batch of block headers received from a remote 1733 // node into the download schedule. 1734 func (d *Downloader) DeliverHeaders(id string, headers []*types.Header) (err error) { 1735 return d.deliver(id, d.headerCh, &headerPack{id, headers}, headerInMeter, headerDropMeter) 1736 } 1737 1738 // DeliverBodies injects a new batch of block bodies received from a remote node. 1739 func (d *Downloader) DeliverBodies(id string, transactions [][]*types.Transaction) (err error) { 1740 return d.deliver(id, d.bodyCh, &bodyPack{id, transactions}, bodyInMeter, bodyDropMeter) 1741 } 1742 1743 // DeliverReceipts injects a new batch of receipts received from a remote node. 1744 func (d *Downloader) DeliverReceipts(id string, receipts [][]*types.Receipt) (err error) { 1745 return d.deliver(id, d.receiptCh, &receiptPack{id, receipts}, receiptInMeter, receiptDropMeter) 1746 } 1747 1748 // DeliverNodeData injects a new batch of node state data received from a remote node. 1749 func (d *Downloader) DeliverNodeData(id string, data [][]byte) (err error) { 1750 return d.deliver(id, d.stateCh, &statePack{id, data}, stateInMeter, stateDropMeter) 1751 } 1752 1753 // deliver injects a new batch of data received from a remote node. 1754 func (d *Downloader) deliver(id string, destCh chan dataPack, packet dataPack, inMeter, dropMeter metrics.Meter) (err error) { 1755 // Update the delivery metrics for both good and failed deliveries 1756 inMeter.Mark(int64(packet.Items())) 1757 defer func() { 1758 if err != nil { 1759 dropMeter.Mark(int64(packet.Items())) 1760 } 1761 }() 1762 // Deliver or abort if the sync is canceled while queuing 1763 d.cancelLock.RLock() 1764 cancel := d.cancelCh 1765 d.cancelLock.RUnlock() 1766 if cancel == nil { 1767 return errNoSyncActive 1768 } 1769 select { 1770 case destCh <- packet: 1771 return nil 1772 case <-cancel: 1773 return errNoSyncActive 1774 } 1775 } 1776 1777 // qosTuner is the quality of service tuning loop that occasionally gathers the 1778 // peer latency statistics and updates the estimated request round trip time. 1779 func (d *Downloader) qosTuner() { 1780 for { 1781 // Retrieve the current median RTT and integrate into the previoust target RTT 1782 rtt := time.Duration((1-qosTuningImpact)*float64(atomic.LoadUint64(&d.rttEstimate)) + qosTuningImpact*float64(d.peers.medianRTT())) 1783 atomic.StoreUint64(&d.rttEstimate, uint64(rtt)) 1784 1785 // A new RTT cycle passed, increase our confidence in the estimated RTT 1786 conf := atomic.LoadUint64(&d.rttConfidence) 1787 conf = conf + (1000000-conf)/2 1788 atomic.StoreUint64(&d.rttConfidence, conf) 1789 1790 // Log the new QoS values and sleep until the next RTT 1791 log.Debug("Recalculated downloader QoS values", "rtt", rtt, "confidence", float64(conf)/1000000.0, "ttl", d.requestTTL()) 1792 select { 1793 case <-d.quitCh: 1794 return 1795 case <-time.After(rtt): 1796 } 1797 } 1798 } 1799 1800 // qosReduceConfidence is meant to be called when a new peer joins the downloader's 1801 // peer set, needing to reduce the confidence we have in out QoS estimates. 1802 func (d *Downloader) qosReduceConfidence() { 1803 // If we have a single peer, confidence is always 1 1804 peers := uint64(d.peers.Len()) 1805 if peers == 0 { 1806 // Ensure peer connectivity races don't catch us off guard 1807 return 1808 } 1809 if peers == 1 { 1810 atomic.StoreUint64(&d.rttConfidence, 1000000) 1811 return 1812 } 1813 // If we have a ton of peers, don't drop confidence) 1814 if peers >= uint64(qosConfidenceCap) { 1815 return 1816 } 1817 // Otherwise drop the confidence factor 1818 conf := atomic.LoadUint64(&d.rttConfidence) * (peers - 1) / peers 1819 if float64(conf)/1000000 < rttMinConfidence { 1820 conf = uint64(rttMinConfidence * 1000000) 1821 } 1822 atomic.StoreUint64(&d.rttConfidence, conf) 1823 1824 rtt := time.Duration(atomic.LoadUint64(&d.rttEstimate)) 1825 log.Debug("Relaxed downloader QoS values", "rtt", rtt, "confidence", float64(conf)/1000000.0, "ttl", d.requestTTL()) 1826 } 1827 1828 // requestRTT returns the current target round trip time for a download request 1829 // to complete in. 1830 // 1831 // Note, the returned RTT is .9 of the actually estimated RTT. The reason is that 1832 // the downloader tries to adapt queries to the RTT, so multiple RTT values can 1833 // be adapted to, but smaller ones are preferred (stabler download stream). 1834 func (d *Downloader) requestRTT() time.Duration { 1835 return time.Duration(atomic.LoadUint64(&d.rttEstimate)) * 9 / 10 1836 } 1837 1838 // requestTTL returns the current timeout allowance for a single download request 1839 // to finish under. 1840 func (d *Downloader) requestTTL() time.Duration { 1841 var ( 1842 rtt = time.Duration(atomic.LoadUint64(&d.rttEstimate)) 1843 conf = float64(atomic.LoadUint64(&d.rttConfidence)) / 1000000.0 1844 ) 1845 ttl := time.Duration(ttlScaling) * time.Duration(float64(rtt)/conf) 1846 if ttl > ttlLimit { 1847 ttl = ttlLimit 1848 } 1849 return ttl 1850 }