core/state: convert prefetcher to concurrent per-trie loader

accounts/abi/bind/backends/simulated.go

@@ -125,10 +125,9 @@ func (b *SimulatedBackend) Rollback() {
 
 func (b *SimulatedBackend) rollback() {
 	blocks, _ := core.GenerateChain(b.config, b.blockchain.CurrentBlock(), ethash.NewFaker(), b.database, 1, func(int, *core.BlockGen) {})
-	stateDB, _ := b.blockchain.State()
 
 	b.pendingBlock = blocks[0]
-	b.pendingState, _ = state.New(b.pendingBlock.Root(), stateDB.Database(), nil)
+	b.pendingState, _ = state.New(b.pendingBlock.Root(), b.blockchain.StateCache(), nil)
 }
 
 // stateByBlockNumber retrieves a state by a given blocknumber.

core/blockchain.go

@@ -201,12 +201,11 @@ type BlockChain struct {
 	running       int32          // 0 if chain is running, 1 when stopped
 	procInterrupt int32          // interrupt signaler for block processing
 
-	engine         consensus.Engine
+	engine     consensus.Engine
-	validator      Validator             // Block and state validator interface
+	validator  Validator // Block and state validator interface
-	triePrefetcher *state.TriePrefetcher // Trie prefetcher interface
+	prefetcher Prefetcher
-	prefetcher     Prefetcher
+	processor  Processor // Block transaction processor interface
-	processor      Processor // Block transaction processor interface
+	vmConfig   vm.Config
-	vmConfig       vm.Config
 
 	shouldPreserve     func(*types.Block) bool        // Function used to determine whether should preserve the given block.
 	terminateInsert    func(common.Hash, uint64) bool // Testing hook used to terminate ancient receipt chain insertion.
@@ -250,15 +249,6 @@ func NewBlockChain(db ethdb.Database, cacheConfig *CacheConfig, chainConfig *par
 	}
 	bc.validator = NewBlockValidator(chainConfig, bc, engine)
 	bc.prefetcher = newStatePrefetcher(chainConfig, bc, engine)
-	tp := state.NewTriePrefetcher(bc.stateCache)
-
-	bc.wg.Add(1)
-	go func() {
-		tp.Loop()
-		bc.wg.Done()
-	}()
-	bc.triePrefetcher = tp
-
 	bc.processor = NewStateProcessor(chainConfig, bc, engine)
 
 	var err error
@@ -1001,9 +991,6 @@ func (bc *BlockChain) Stop() {
 	bc.scope.Close()
 	close(bc.quit)
 	bc.StopInsert()
-	if bc.triePrefetcher != nil {
-		bc.triePrefetcher.Close()
-	}
 	bc.wg.Wait()
 
 	// Ensure that the entirety of the state snapshot is journalled to disk.
@@ -1870,16 +1857,20 @@ func (bc *BlockChain) insertChain(chain types.Blocks, verifySeals bool) (int, er
 			parent = bc.GetHeader(block.ParentHash(), block.NumberU64()-1)
 		}
 		statedb, err := state.New(parent.Root, bc.stateCache, bc.snaps)
-		statedb.UsePrefetcher(bc.triePrefetcher)
 		if err != nil {
 			return it.index, err
 		}
+		// Enable prefetching to pull in trie node paths while processing transactions
+		statedb.StartPrefetcher("chain")
+		defer statedb.StopPrefetcher() // stopped on write anyway, defer meant to catch early error returns
+
 		// If we have a followup block, run that against the current state to pre-cache
 		// transactions and probabilistically some of the account/storage trie nodes.
 		var followupInterrupt uint32
 		if !bc.cacheConfig.TrieCleanNoPrefetch {
 			if followup, err := it.peek(); followup != nil && err == nil {
 				throwaway, _ := state.New(parent.Root, bc.stateCache, bc.snaps)
+
 				go func(start time.Time, followup *types.Block, throwaway *state.StateDB, interrupt *uint32) {
 					bc.prefetcher.Prefetch(followup, throwaway, bc.vmConfig, &followupInterrupt)
 
@@ -1933,7 +1924,6 @@ func (bc *BlockChain) insertChain(chain types.Blocks, verifySeals bool) (int, er
 		if err != nil {
 			return it.index, err
 		}
-
 		// Update the metrics touched during block commit
 		accountCommitTimer.Update(statedb.AccountCommits)   // Account commits are complete, we can mark them
 		storageCommitTimer.Update(statedb.StorageCommits)   // Storage commits are complete, we can mark them

core/state/state_object.go

@@ -162,7 +162,7 @@ func (s *stateObject) getTrie(db Database) Trie {
 		if s.data.Root != emptyRoot && s.db.prefetcher != nil {
 			// When the miner is creating the pending state, there is no
 			// prefetcher
-			s.trie = s.db.prefetcher.GetTrie(s.data.Root)
+			s.trie = s.db.prefetcher.trie(s.data.Root)
 		}
 		if s.trie == nil {
 			var err error
@@ -309,14 +309,16 @@ func (s *stateObject) setState(key, value common.Hash) {
 
 // finalise moves all dirty storage slots into the pending area to be hashed or
 // committed later. It is invoked at the end of every transaction.
-func (s *stateObject) finalise() {
+func (s *stateObject) finalise(prefetch bool) {
-	trieChanges := make([]common.Hash, 0, len(s.dirtyStorage))
+	slotsToPrefetch := make([][]byte, 0, len(s.dirtyStorage))
 	for key, value := range s.dirtyStorage {
 		s.pendingStorage[key] = value
-		trieChanges = append(trieChanges, key)
+		if value != s.originStorage[key] {
+			slotsToPrefetch = append(slotsToPrefetch, common.CopyBytes(key[:])) // Copy needed for closure
+		}
 	}
-	if len(trieChanges) > 0 && s.db.prefetcher != nil && s.data.Root != emptyRoot {
+	if s.db.prefetcher != nil && prefetch && len(slotsToPrefetch) > 0 && s.data.Root != emptyRoot {
-		s.db.prefetcher.PrefetchStorage(s.data.Root, trieChanges)
+		s.db.prefetcher.prefetch(s.data.Root, slotsToPrefetch)
 	}
 	if len(s.dirtyStorage) > 0 {
 		s.dirtyStorage = make(Storage)
@@ -327,7 +329,7 @@ func (s *stateObject) finalise() {
 // It will return nil if the trie has not been loaded and no changes have been made
 func (s *stateObject) updateTrie(db Database) Trie {
 	// Make sure all dirty slots are finalized into the pending storage area
-	s.finalise()
+	s.finalise(false) // Don't prefetch any more, pull directly if need be
 	if len(s.pendingStorage) == 0 {
 		return s.trie
 	}
@@ -340,6 +342,8 @@ func (s *stateObject) updateTrie(db Database) Trie {
 	// Insert all the pending updates into the trie
 	tr := s.getTrie(db)
 	hasher := s.db.hasher
+
+	usedStorage := make([][]byte, 0, len(s.pendingStorage))
 	for key, value := range s.pendingStorage {
 		// Skip noop changes, persist actual changes
 		if value == s.originStorage[key] {
@@ -366,6 +370,10 @@ func (s *stateObject) updateTrie(db Database) Trie {
 			}
 			storage[crypto.HashData(hasher, key[:])] = v // v will be nil if value is 0x00
 		}
+		usedStorage = append(usedStorage, common.CopyBytes(key[:])) // Copy needed for closure
+	}
+	if s.db.prefetcher != nil {
+		s.db.prefetcher.used(s.data.Root, usedStorage)
 	}
 	if len(s.pendingStorage) > 0 {
 		s.pendingStorage = make(Storage)

core/state/state_test.go

@@ -170,7 +170,7 @@ func TestSnapshot2(t *testing.T) {
 	state.setStateObject(so0)
 
 	root, _ := state.Commit(false)
-	state.Reset(root)
+	state, _ = New(root, state.db, state.snaps)
 
 	// and one with deleted == true
 	so1 := state.getStateObject(stateobjaddr1)

core/state/statedb.go

@@ -63,7 +63,7 @@ func (n *proofList) Delete(key []byte) error {
 // * Accounts
 type StateDB struct {
 	db           Database
-	prefetcher   *TriePrefetcher
+	prefetcher   *triePrefetcher
 	originalRoot common.Hash // The pre-state root, before any changes were made
 	trie         Trie
 	hasher       crypto.KeccakState
@@ -149,10 +149,25 @@ func New(root common.Hash, db Database, snaps *snapshot.Tree) (*StateDB, error)
 	return sdb, nil
 }
 
-func (s *StateDB) UsePrefetcher(prefetcher *TriePrefetcher) {
+// StartPrefetcher initializes a new trie prefetcher to pull in nodes from the
-	if prefetcher != nil {
+// state trie concurrently while the state is mutated so that when we reach the
-		s.prefetcher = prefetcher
+// commit phase, most of the needed data is already hot.
-		s.prefetcher.Resume(s.originalRoot)
+func (s *StateDB) StartPrefetcher(namespace string) {
+	if s.prefetcher != nil {
+		s.prefetcher.close()
+		s.prefetcher = nil
+	}
+	if s.snap != nil {
+		s.prefetcher = newTriePrefetcher(s.db, s.originalRoot, namespace)
+	}
+}
+
+// StopPrefetcher terminates a running prefetcher and reports any leftover stats
+// from the gathered metrics.
+func (s *StateDB) StopPrefetcher() {
+	if s.prefetcher != nil {
+		s.prefetcher.close()
+		s.prefetcher = nil
 	}
 }
 
@@ -167,37 +182,6 @@ func (s *StateDB) Error() error {
 	return s.dbErr
 }
 
-// Reset clears out all ephemeral state objects from the state db, but keeps
-// the underlying state trie to avoid reloading data for the next operations.
-func (s *StateDB) Reset(root common.Hash) error {
-	tr, err := s.db.OpenTrie(root)
-	if err != nil {
-		return err
-	}
-	s.trie = tr
-	s.stateObjects = make(map[common.Address]*stateObject)
-	s.stateObjectsPending = make(map[common.Address]struct{})
-	s.stateObjectsDirty = make(map[common.Address]struct{})
-	s.thash = common.Hash{}
-	s.bhash = common.Hash{}
-	s.txIndex = 0
-	s.logs = make(map[common.Hash][]*types.Log)
-	s.logSize = 0
-	s.preimages = make(map[common.Hash][]byte)
-	s.clearJournalAndRefund()
-
-	if s.snaps != nil {
-		s.snapAccounts, s.snapDestructs, s.snapStorage = nil, nil, nil
-		if s.snap = s.snaps.Snapshot(root); s.snap != nil {
-			s.snapDestructs = make(map[common.Hash]struct{})
-			s.snapAccounts = make(map[common.Hash][]byte)
-			s.snapStorage = make(map[common.Hash]map[common.Hash][]byte)
-		}
-	}
-	s.accessList = newAccessList()
-	return nil
-}
-
 func (s *StateDB) AddLog(log *types.Log) {
 	s.journal.append(addLogChange{txhash: s.thash})
 
@@ -737,6 +721,13 @@ func (s *StateDB) Copy() *StateDB {
 	// However, it doesn't cost us much to copy an empty list, so we do it anyway
 	// to not blow up if we ever decide copy it in the middle of a transaction
 	state.accessList = s.accessList.Copy()
+
+	// If there's a prefetcher running, make an inactive copy of it that can
+	// only access data but does not actively preload (since the user will not
+	// know that they need to explicitly terminate an active copy).
+	if s.prefetcher != nil {
+		state.prefetcher = s.prefetcher.copy()
+	}
 	return state
 }
 
@@ -773,7 +764,7 @@ func (s *StateDB) GetRefund() uint64 {
 // the journal as well as the refunds. Finalise, however, will not push any updates
 // into the tries just yet. Only IntermediateRoot or Commit will do that.
 func (s *StateDB) Finalise(deleteEmptyObjects bool) {
-	var addressesToPrefetch []common.Address
+	addressesToPrefetch := make([][]byte, 0, len(s.journal.dirties))
 	for addr := range s.journal.dirties {
 		obj, exist := s.stateObjects[addr]
 		if !exist {
@@ -798,21 +789,19 @@ func (s *StateDB) Finalise(deleteEmptyObjects bool) {
 				delete(s.snapStorage, obj.addrHash)        // Clear out any previously updated storage data (may be recreated via a ressurrect)
 			}
 		} else {
-			obj.finalise()
+			obj.finalise(true) // Prefetch slots in the background
 		}
 		s.stateObjectsPending[addr] = struct{}{}
 		s.stateObjectsDirty[addr] = struct{}{}
+
 		// At this point, also ship the address off to the precacher. The precacher
 		// will start loading tries, and when the change is eventually committed,
 		// the commit-phase will be a lot faster
-		if s.prefetcher != nil {
+		addressesToPrefetch = append(addressesToPrefetch, common.CopyBytes(addr[:])) // Copy needed for closure
-			addressesToPrefetch = append(addressesToPrefetch, addr)
-		}
 	}
-	if s.prefetcher != nil {
+	if s.prefetcher != nil && len(addressesToPrefetch) > 0 {
-		s.prefetcher.PrefetchAddresses(addressesToPrefetch)
+		s.prefetcher.prefetch(s.originalRoot, addressesToPrefetch)
 	}
-
 	// Invalidate journal because reverting across transactions is not allowed.
 	s.clearJournalAndRefund()
 }
@@ -824,29 +813,49 @@ func (s *StateDB) IntermediateRoot(deleteEmptyObjects bool) common.Hash {
 	// Finalise all the dirty storage states and write them into the tries
 	s.Finalise(deleteEmptyObjects)
 
-	// Now we're about to start to write changes to the trie. The trie is so
+	// If there was a trie prefetcher operating, it gets aborted and irrevocably
-	// far _untouched_. We can check with the prefetcher, if it can give us
+	// modified after we start retrieving tries. Remove it from the statedb after
-	// a trie which has the same root, but also has some content loaded into it.
+	// this round of use.
-	// If so, use that one instead.
+	//
+	// This is weird pre-byzantium since the first tx runs with a prefetcher and
+	// the remainder without, but pre-byzantium even the initial prefetcher is
+	// useless, so no sleep lost.
+	prefetcher := s.prefetcher
 	if s.prefetcher != nil {
-		s.prefetcher.Pause()
+		defer func() {
-		// We only want to do this _once_, if someone calls IntermediateRoot again,
+			s.prefetcher.close()
-		// we shouldn't fetch the trie again
+			s.prefetcher = nil
-		if s.originalRoot != (common.Hash{}) {
+		}()
-			if trie := s.prefetcher.GetTrie(s.originalRoot); trie != nil {
+	}
-				s.trie = trie
+	// Although naively it makes sense to retrieve the account trie and then do
-			}
+	// the contract storage and account updates sequentially, that short circuits
-			s.originalRoot = common.Hash{}
+	// the account prefetcher. Instead, let's process all the storage updates
+	// first, giving the account prefeches just a few more milliseconds of time
+	// to pull useful data from disk.
+	for addr := range s.stateObjectsPending {
+		if obj := s.stateObjects[addr]; !obj.deleted {
+			obj.updateRoot(s.db)
+		}
+	}
+	// Now we're about to start to write changes to the trie. The trie is so far
+	// _untouched_. We can check with the prefetcher, if it can give us a trie
+	// which has the same root, but also has some content loaded into it.
+	if prefetcher != nil {
+		if trie := prefetcher.trie(s.originalRoot); trie != nil {
+			s.trie = trie
 		}
 	}
+	usedAddrs := make([][]byte, 0, len(s.stateObjectsPending))
 	for addr := range s.stateObjectsPending {
-		obj := s.stateObjects[addr]
+		if obj := s.stateObjects[addr]; obj.deleted {
-		if obj.deleted {
 			s.deleteStateObject(obj)
 		} else {
-			obj.updateRoot(s.db)
 			s.updateStateObject(obj)
 		}
+		usedAddrs = append(usedAddrs, common.CopyBytes(addr[:])) // Copy needed for closure
+	}
+	if prefetcher != nil {
+		prefetcher.used(s.originalRoot, usedAddrs)
 	}
 	if len(s.stateObjectsPending) > 0 {
 		s.stateObjectsPending = make(map[common.Address]struct{})

core/state/statedb_test.go

@@ -474,7 +474,7 @@ func TestTouchDelete(t *testing.T) {
 	s := newStateTest()
 	s.state.GetOrNewStateObject(common.Address{})
 	root, _ := s.state.Commit(false)
-	s.state.Reset(root)
+	s.state, _ = New(root, s.state.db, s.state.snaps)
 
 	snapshot := s.state.Snapshot()
 	s.state.AddBalance(common.Address{}, new(big.Int))
@@ -676,7 +676,7 @@ func TestDeleteCreateRevert(t *testing.T) {
 	state.SetBalance(addr, big.NewInt(1))
 
 	root, _ := state.Commit(false)
-	state.Reset(root)
+	state, _ = New(root, state.db, state.snaps)
 
 	// Simulate self-destructing in one transaction, then create-reverting in another
 	state.Suicide(addr)
@@ -688,7 +688,7 @@ func TestDeleteCreateRevert(t *testing.T) {
 
 	// Commit the entire state and make sure we don't crash and have the correct state
 	root, _ = state.Commit(true)
-	state.Reset(root)
+	state, _ = New(root, state.db, state.snaps)
 
 	if state.getStateObject(addr) != nil {
 		t.Fatalf("self-destructed contract came alive")

core/state/trie_prefetcher.go

@@ -17,233 +17,318 @@
 package state
 
 import (
+	"sync"
+
 	"github.com/ethereum/go-ethereum/common"
 	"github.com/ethereum/go-ethereum/log"
 	"github.com/ethereum/go-ethereum/metrics"
 )
 
 var (
-	// trieDeliveryMeter counts how many times the prefetcher was unable to supply
+	// triePrefetchMetricsPrefix is the prefix under which to publis the metrics.
-	// the statedb with a prefilled trie. This meter should be zero -- if it's not, that
+	triePrefetchMetricsPrefix = "trie/prefetch/"
-	// needs to be investigated
-	trieDeliveryMissMeter = metrics.NewRegisteredMeter("trie/prefetch/deliverymiss", nil)
-
-	triePrefetchFetchMeter = metrics.NewRegisteredMeter("trie/prefetch/fetch", nil)
-	triePrefetchSkipMeter  = metrics.NewRegisteredMeter("trie/prefetch/skip", nil)
-	triePrefetchDropMeter  = metrics.NewRegisteredMeter("trie/prefetch/drop", nil)
 )
 
-// TriePrefetcher is an active prefetcher, which receives accounts or storage
+// triePrefetcher is an active prefetcher, which receives accounts or storage
-// items on two channels, and does trie-loading of the items.
+// items and does trie-loading of them. The goal is to get as much useful content
-// The goal is to get as much useful content into the caches as possible
+// into the caches as possible.
-type TriePrefetcher struct {
+//
-	requestCh  chan (fetchRequest) // Chan to receive requests for data to fetch
+// Note, the prefetcher's API is not thread safe.
-	cmdCh      chan (*cmd)         // Chan to control activity, pause/new root
+type triePrefetcher struct {
-	quitCh     chan (struct{})
+	db       Database                    // Database to fetch trie nodes through
-	deliveryCh chan (struct{})
+	root     common.Hash                 // Root hash of theaccount trie for metrics
-	db         Database
+	fetches  map[common.Hash]Trie        // Partially or fully fetcher tries
-
+	fetchers map[common.Hash]*subfetcher // Subfetchers for each trie
-	paused bool
+
-
+	deliveryMissMeter metrics.Meter
-	storageTries    map[common.Hash]Trie
+	accountLoadMeter  metrics.Meter
-	accountTrie     Trie
+	accountDupMeter   metrics.Meter
-	accountTrieRoot common.Hash
+	accountSkipMeter  metrics.Meter
+	accountWasteMeter metrics.Meter
+	storageLoadMeter  metrics.Meter
+	storageDupMeter   metrics.Meter
+	storageSkipMeter  metrics.Meter
+	storageWasteMeter metrics.Meter
 }
 
-func NewTriePrefetcher(db Database) *TriePrefetcher {
+// newTriePrefetcher
-	return &TriePrefetcher{
+func newTriePrefetcher(db Database, root common.Hash, namespace string) *triePrefetcher {
-		requestCh:  make(chan fetchRequest, 200),
+	prefix := triePrefetchMetricsPrefix + namespace
-		cmdCh:      make(chan *cmd),
+	p := &triePrefetcher{
-		quitCh:     make(chan struct{}),
+		db:       db,
-		deliveryCh: make(chan struct{}),
+		root:     root,
-		db:         db,
+		fetchers: make(map[common.Hash]*subfetcher), // Active prefetchers use the fetchers map
+
+		deliveryMissMeter: metrics.GetOrRegisterMeter(prefix+"/deliverymiss", nil),
+		accountLoadMeter:  metrics.GetOrRegisterMeter(prefix+"/account/load", nil),
+		accountDupMeter:   metrics.GetOrRegisterMeter(prefix+"/account/dup", nil),
+		accountSkipMeter:  metrics.GetOrRegisterMeter(prefix+"/account/skip", nil),
+		accountWasteMeter: metrics.GetOrRegisterMeter(prefix+"/account/waste", nil),
+		storageLoadMeter:  metrics.GetOrRegisterMeter(prefix+"/storage/load", nil),
+		storageDupMeter:   metrics.GetOrRegisterMeter(prefix+"/storage/dup", nil),
+		storageSkipMeter:  metrics.GetOrRegisterMeter(prefix+"/storage/skip", nil),
+		storageWasteMeter: metrics.GetOrRegisterMeter(prefix+"/storage/waste", nil),
 	}
+	return p
 }
 
-type cmd struct {
+// close iterates over all the subfetchers, aborts any that were left spinning
-	root common.Hash
+// and reports the stats to the metrics subsystem.
-}
+func (p *triePrefetcher) close() {
+	for _, fetcher := range p.fetchers {
+		fetcher.abort() // safe to do multiple times
 
-type fetchRequest struct {
+		if metrics.Enabled {
-	slots       []common.Hash
+			if fetcher.root == p.root {
-	storageRoot *common.Hash
+				p.accountLoadMeter.Mark(int64(len(fetcher.seen)))
-	addresses   []common.Address
+				p.accountDupMeter.Mark(int64(fetcher.dups))
-}
+				p.accountSkipMeter.Mark(int64(len(fetcher.tasks)))
 
-func (p *TriePrefetcher) Loop() {
+				for _, key := range fetcher.used {
-	var (
+					delete(fetcher.seen, string(key))
-		accountTrieRoot common.Hash
-		accountTrie     Trie
-		storageTries    map[common.Hash]Trie
-
-		err error
-		// Some tracking of performance
-		skipped int64
-		fetched int64
-
-		paused = true
-	)
-	// The prefetcher loop has two distinct phases:
-	// 1: Paused: when in this state, the accumulated tries are accessible to outside
-	// callers.
-	// 2: Active prefetching, awaiting slots and accounts to prefetch
-	for {
-		select {
-		case <-p.quitCh:
-			return
-		case cmd := <-p.cmdCh:
-			// Clear out any old requests
-		drain:
-			for {
-				select {
-				case req := <-p.requestCh:
-					if req.slots != nil {
-						skipped += int64(len(req.slots))
-					} else {
-						skipped += int64(len(req.addresses))
-					}
-				default:
-					break drain
 				}
-			}
+				p.accountWasteMeter.Mark(int64(len(fetcher.seen)))
-			if paused {
-				// Clear old data
-				p.storageTries = nil
-				p.accountTrie = nil
-				p.accountTrieRoot = common.Hash{}
-				// Resume again
-				storageTries = make(map[common.Hash]Trie)
-				accountTrieRoot = cmd.root
-				accountTrie, err = p.db.OpenTrie(accountTrieRoot)
-				if err != nil {
-					log.Error("Trie prefetcher failed opening trie", "root", accountTrieRoot, "err", err)
-				}
-				if accountTrieRoot == (common.Hash{}) {
-					log.Error("Trie prefetcher unpaused with bad root")
-				}
-				paused = false
 			} else {
-				// Update metrics at new block events
+				p.storageLoadMeter.Mark(int64(len(fetcher.seen)))
-				triePrefetchFetchMeter.Mark(fetched)
+				p.storageDupMeter.Mark(int64(fetcher.dups))
-				triePrefetchSkipMeter.Mark(skipped)
+				p.storageSkipMeter.Mark(int64(len(fetcher.tasks)))
-				fetched, skipped = 0, 0
+
-				// Make the tries accessible
+				for _, key := range fetcher.used {
-				p.accountTrie = accountTrie
+					delete(fetcher.seen, string(key))
-				p.storageTries = storageTries
-				p.accountTrieRoot = accountTrieRoot
-				if cmd.root != (common.Hash{}) {
-					log.Error("Trie prefetcher paused with non-empty root")
-				}
-				paused = true
-			}
-			p.deliveryCh <- struct{}{}
-		case req := <-p.requestCh:
-			if paused {
-				continue
-			}
-			if sRoot := req.storageRoot; sRoot != nil {
-				// Storage slots to fetch
-				var (
-					storageTrie Trie
-					err         error
-				)
-				if storageTrie = storageTries[*sRoot]; storageTrie == nil {
-					if storageTrie, err = p.db.OpenTrie(*sRoot); err != nil {
-						log.Warn("trie prefetcher failed opening storage trie", "root", *sRoot, "err", err)
-						skipped += int64(len(req.slots))
-						continue
-					}
-					storageTries[*sRoot] = storageTrie
-				}
-				for _, key := range req.slots {
-					storageTrie.TryGet(key[:])
-				}
-				fetched += int64(len(req.slots))
-			} else { // an account
-				for _, addr := range req.addresses {
-					accountTrie.TryGet(addr[:])
 				}
-				fetched += int64(len(req.addresses))
+				p.storageWasteMeter.Mark(int64(len(fetcher.seen)))
 			}
 		}
 	}
+	// Clear out all fetchers (will crash on a second call, deliberate)
+	p.fetchers = nil
 }
 
-// Close stops the prefetcher
+// copy creates a deep-but-inactive copy of the trie prefetcher. Any trie data
-func (p *TriePrefetcher) Close() {
+// already loaded will be copied over, but no goroutines will be started. This
-	if p.quitCh != nil {
+// is mostly used in the miner which creates a copy of it's actively mutated
-		close(p.quitCh)
+// state to be sealed while it may further mutate the state.
-		p.quitCh = nil
+func (p *triePrefetcher) copy() *triePrefetcher {
+	copy := &triePrefetcher{
+		db:      p.db,
+		root:    p.root,
+		fetches: make(map[common.Hash]Trie), // Active prefetchers use the fetches map
+
+		deliveryMissMeter: p.deliveryMissMeter,
+		accountLoadMeter:  p.accountLoadMeter,
+		accountDupMeter:   p.accountDupMeter,
+		accountSkipMeter:  p.accountSkipMeter,
+		accountWasteMeter: p.accountWasteMeter,
+		storageLoadMeter:  p.storageLoadMeter,
+		storageDupMeter:   p.storageDupMeter,
+		storageSkipMeter:  p.storageSkipMeter,
+		storageWasteMeter: p.storageWasteMeter,
+	}
+	// If the prefetcher is already a copy, duplicate the data
+	if p.fetches != nil {
+		for root, fetch := range p.fetches {
+			copy.fetches[root] = p.db.CopyTrie(fetch)
+		}
+		return copy
+	}
+	// Otherwise we're copying an active fetcher, retrieve the current states
+	for root, fetcher := range p.fetchers {
+		copy.fetches[root] = fetcher.peek()
 	}
+	return copy
 }
 
-// Resume causes the prefetcher to clear out old data, and get ready to
+// prefetch schedules a batch of trie items to prefetch.
-// fetch data concerning the new root
+func (p *triePrefetcher) prefetch(root common.Hash, keys [][]byte) {
-func (p *TriePrefetcher) Resume(root common.Hash) {
+	// If the prefetcher is an inactive one, bail out
-	p.paused = false
+	if p.fetches != nil {
-	p.cmdCh <- &cmd{
+		return
-		root: root,
+	}
+	// Active fetcher, schedule the retrievals
+	fetcher := p.fetchers[root]
+	if fetcher == nil {
+		fetcher = newSubfetcher(p.db, root)
+		p.fetchers[root] = fetcher
 	}
-	// Wait for it
+	fetcher.schedule(keys)
-	<-p.deliveryCh
 }
 
-// Pause causes the prefetcher to pause prefetching, and make tries
+// trie returns the trie matching the root hash, or nil if the prefetcher doesn't
-// accessible to callers via GetTrie
+// have it.
-func (p *TriePrefetcher) Pause() {
+func (p *triePrefetcher) trie(root common.Hash) Trie {
-	if p.paused {
+	// If the prefetcher is inactive, return from existing deep copies
-		return
+	if p.fetches != nil {
+		trie := p.fetches[root]
+		if trie == nil {
+			p.deliveryMissMeter.Mark(1)
+			return nil
+		}
+		return p.db.CopyTrie(trie)
 	}
-	p.paused = true
+	// Otherwise the prefetcher is active, bail if no trie was prefetched for this root
-	p.cmdCh <- &cmd{
+	fetcher := p.fetchers[root]
-		root: common.Hash{},
+	if fetcher == nil {
+		p.deliveryMissMeter.Mark(1)
+		return nil
 	}
-	// Wait for it
+	// Interrupt the prefetcher if it's by any chance still running and return
-	<-p.deliveryCh
+	// a copy of any pre-loaded trie.
+	fetcher.abort() // safe to do multiple times
+
+	trie := fetcher.peek()
+	if trie == nil {
+		p.deliveryMissMeter.Mark(1)
+		return nil
+	}
+	return trie
 }
 
-// PrefetchAddresses adds an address for prefetching
+// used marks a batch of state items used to allow creating statistics as to
-func (p *TriePrefetcher) PrefetchAddresses(addresses []common.Address) {
+// how useful or wasteful the prefetcher is.
-	cmd := fetchRequest{
+func (p *triePrefetcher) used(root common.Hash, used [][]byte) {
-		addresses: addresses,
+	if fetcher := p.fetchers[root]; fetcher != nil {
+		fetcher.used = used
 	}
-	// We do an async send here, to not cause the caller to block
+}
-	//p.requestCh <- cmd
+
+// subfetcher is a trie fetcher goroutine responsible for pulling entries for a
+// single trie. It is spawned when a new root is encountered and lives until the
+// main prefetcher is paused and either all requested items are processed or if
+// the trie being worked on is retrieved from the prefetcher.
+type subfetcher struct {
+	db   Database    // Database to load trie nodes through
+	root common.Hash // Root hash of the trie to prefetch
+	trie Trie        // Trie being populated with nodes
+
+	tasks [][]byte   // Items queued up for retrieval
+	lock  sync.Mutex // Lock protecting the task queue
+
+	wake chan struct{}  // Wake channel if a new task is scheduled
+	stop chan struct{}  // Channel to interrupt processing
+	term chan struct{}  // Channel to signal iterruption
+	copy chan chan Trie // Channel to request a copy of the current trie
+
+	seen map[string]struct{} // Tracks the entries already loaded
+	dups int                 // Number of duplicate preload tasks
+	used [][]byte            // Tracks the entries used in the end
+}
+
+// newSubfetcher creates a goroutine to prefetch state items belonging to a
+// particular root hash.
+func newSubfetcher(db Database, root common.Hash) *subfetcher {
+	sf := &subfetcher{
+		db:   db,
+		root: root,
+		wake: make(chan struct{}, 1),
+		stop: make(chan struct{}),
+		term: make(chan struct{}),
+		copy: make(chan chan Trie),
+		seen: make(map[string]struct{}),
+	}
+	go sf.loop()
+	return sf
+}
+
+// schedule adds a batch of trie keys to the queue to prefetch.
+func (sf *subfetcher) schedule(keys [][]byte) {
+	// Append the tasks to the current queue
+	sf.lock.Lock()
+	sf.tasks = append(sf.tasks, keys...)
+	sf.lock.Unlock()
+
+	// Notify the prefetcher, it's fine if it's already terminated
 	select {
-	case p.requestCh <- cmd:
+	case sf.wake <- struct{}{}:
 	default:
-		triePrefetchDropMeter.Mark(int64(len(addresses)))
 	}
 }
 
-// PrefetchStorage adds a storage root and a set of keys for prefetching
+// peek tries to retrieve a deep copy of the fetcher's trie in whatever form it
-func (p *TriePrefetcher) PrefetchStorage(root common.Hash, slots []common.Hash) {
+// is currently.
-	cmd := fetchRequest{
+func (sf *subfetcher) peek() Trie {
-		storageRoot: &root,
+	ch := make(chan Trie)
-		slots:       slots,
+	select {
+	case sf.copy <- ch:
+		// Subfetcher still alive, return copy from it
+		return <-ch
+
+	case <-sf.term:
+		// Subfetcher already terminated, return a copy directly
+		if sf.trie == nil {
+			return nil
+		}
+		return sf.db.CopyTrie(sf.trie)
 	}
-	// We do an async send here, to not cause the caller to block
+}
-	//p.requestCh <- cmd
+
+// abort interrupts the subfetcher immediately. It is safe to call abort multiple
+// times but it is not thread safe.
+func (sf *subfetcher) abort() {
 	select {
-	case p.requestCh <- cmd:
+	case <-sf.stop:
 	default:
-		triePrefetchDropMeter.Mark(int64(len(slots)))
+		close(sf.stop)
 	}
+	<-sf.term
 }
 
-// GetTrie returns the trie matching the root hash, or nil if the prefetcher
+// loop waits for new tasks to be scheduled and keeps loading them until it runs
-// doesn't have it.
+// out of tasks or its underlying trie is retrieved for committing.
-func (p *TriePrefetcher) GetTrie(root common.Hash) Trie {
+func (sf *subfetcher) loop() {
-	if root == p.accountTrieRoot {
+	// No matter how the loop stops, signal anyone waiting that it's terminated
-		return p.accountTrie
+	defer close(sf.term)
-	}
+
-	if storageTrie, ok := p.storageTries[root]; ok {
+	// Start by opening the trie and stop processing if it fails
-		// Two accounts may well have the same storage root, but we cannot allow
+	trie, err := sf.db.OpenTrie(sf.root)
-		// them both to make updates to the same trie instance. Therefore,
+	if err != nil {
-		// we need to either delete the trie now, or deliver a copy of the trie.
+		log.Warn("Trie prefetcher failed opening trie", "root", sf.root, "err", err)
-		delete(p.storageTries, root)
+		return
-		return storageTrie
+	}
-	}
+	sf.trie = trie
-	trieDeliveryMissMeter.Mark(1)
+
-	return nil
+	// Trie opened successfully, keep prefetching items
+	for {
+		select {
+		case <-sf.wake:
+			// Subfetcher was woken up, retrieve any tasks to avoid spinning the lock
+			sf.lock.Lock()
+			tasks := sf.tasks
+			sf.tasks = nil
+			sf.lock.Unlock()
+
+			// Prefetch any tasks until the loop is interrupted
+			for i, task := range tasks {
+				select {
+				case <-sf.stop:
+					// If termination is requested, add any leftover back and return
+					sf.lock.Lock()
+					sf.tasks = append(sf.tasks, tasks[i:]...)
+					sf.lock.Unlock()
+					return
+
+				case ch := <-sf.copy:
+					// Somebody wants a copy of the current trie, grant them
+					ch <- sf.db.CopyTrie(sf.trie)
+
+				default:
+					// No termination request yet, prefetch the next entry
+					taskid := string(task)
+					if _, ok := sf.seen[taskid]; ok {
+						sf.dups++
+					} else {
+						sf.trie.TryGet(task)
+						sf.seen[taskid] = struct{}{}
+					}
+				}
+			}
+
+		case ch := <-sf.copy:
+			// Somebody wants a copy of the current trie, grant them
+			ch <- sf.db.CopyTrie(sf.trie)
+
+		case <-sf.stop:
+			// Termination is requested, abort and leave remaining tasks
+			return
+		}
+	}
 }

eth/api_tracer.go

@@ -299,7 +299,8 @@ func (api *PrivateDebugAPI) traceChain(ctx context.Context, start, end *types.Bl
 				failed = err
 				break
 			}
-			if err := statedb.Reset(root); err != nil {
+			statedb, err = state.New(root, database, nil)
+			if err != nil {
 				failed = err
 				break
 			}
@@ -699,7 +700,8 @@ func (api *PrivateDebugAPI) computeStateDB(block *types.Block, reexec uint64) (*
 		if err != nil {
 			return nil, err
 		}
-		if err := statedb.Reset(root); err != nil {
+		statedb, err = state.New(root, database, nil)
+		if err != nil {
 			return nil, fmt.Errorf("state reset after block %d failed: %v", block.NumberU64(), err)
 		}
 		database.TrieDB().Reference(root, common.Hash{})

miner/worker.go

@@ -303,6 +303,9 @@ func (w *worker) isRunning() bool {
 // close terminates all background threads maintained by the worker.
 // Note the worker does not support being closed multiple times.
 func (w *worker) close() {
+	if w.current != nil && w.current.state != nil {
+		w.current.state.StopPrefetcher()
+	}
 	atomic.StoreInt32(&w.running, 0)
 	close(w.exitCh)
 }
@@ -642,10 +645,14 @@ func (w *worker) resultLoop() {
 
 // makeCurrent creates a new environment for the current cycle.
 func (w *worker) makeCurrent(parent *types.Block, header *types.Header) error {
+	// Retrieve the parent state to execute on top and start a prefetcher for
+	// the miner to speed block sealing up a bit
 	state, err := w.chain.StateAt(parent.Root())
 	if err != nil {
 		return err
 	}
+	state.StartPrefetcher("miner")
+
 	env := &environment{
 		signer:    types.NewEIP155Signer(w.chainConfig.ChainID),
 		state:     state,
@@ -654,7 +661,6 @@ func (w *worker) makeCurrent(parent *types.Block, header *types.Header) error {
 		uncles:    mapset.NewSet(),
 		header:    header,
 	}
-
 	// when 08 is processed ancestors contain 07 (quick block)
 	for _, ancestor := range w.chain.GetBlocksFromHash(parent.Hash(), 7) {
 		for _, uncle := range ancestor.Uncles() {
@@ -663,9 +669,14 @@ func (w *worker) makeCurrent(parent *types.Block, header *types.Header) error {
 		env.family.Add(ancestor.Hash())
 		env.ancestors.Add(ancestor.Hash())
 	}
-
 	// Keep track of transactions which return errors so they can be removed
 	env.tcount = 0
+
+	// Swap out the old work with the new one, terminating any leftover prefetcher
+	// processes in the mean time and starting a new one.
+	if w.current != nil && w.current.state != nil {
+		w.current.state.StopPrefetcher()
+	}
 	w.current = env
 	return nil
 }
@@ -719,7 +730,6 @@ func (w *worker) updateSnapshot() {
 		w.current.receipts,
 		new(trie.Trie),
 	)
-
 	w.snapshotState = w.current.state.Copy()
 }