core/rawdb, core/state/snapshot: runtime snapshot generation

cmd/geth/main.go

@@ -106,6 +106,7 @@ var (
 		utils.CacheDatabaseFlag,
 		utils.CacheTrieFlag,
 		utils.CacheGCFlag,
+		utils.CacheSnapshotFlag,
 		utils.CacheNoPrefetchFlag,
 		utils.ListenPortFlag,
 		utils.MaxPeersFlag,

cmd/geth/usage.go

@@ -137,6 +137,7 @@ var AppHelpFlagGroups = []flagGroup{
 			utils.CacheDatabaseFlag,
 			utils.CacheTrieFlag,
 			utils.CacheGCFlag,
+			utils.CacheSnapshotFlag,
 			utils.CacheNoPrefetchFlag,
 		},
 	},

cmd/utils/flags.go

@@ -383,14 +383,19 @@ var (
 	}
 	CacheTrieFlag = cli.IntFlag{
 		Name:  "cache.trie",
-		Usage: "Percentage of cache memory allowance to use for trie caching (default = 25% full mode, 50% archive mode)",
-		Value: 25,
+		Usage: "Percentage of cache memory allowance to use for trie caching (default = 15% full mode, 30% archive mode)",
+		Value: 15,
 	}
 	CacheGCFlag = cli.IntFlag{
 		Name:  "cache.gc",
 		Usage: "Percentage of cache memory allowance to use for trie pruning (default = 25% full mode, 0% archive mode)",
 		Value: 25,
 	}
+	CacheSnapshotFlag = cli.IntFlag{
+		Name:  "cache.snapshot",
+		Usage: "Percentage of cache memory allowance to use for snapshot caching (default = 10% full mode, 20% archive mode)",
+		Value: 10,
+	}
 	CacheNoPrefetchFlag = cli.BoolFlag{
 		Name:  "cache.noprefetch",
 		Usage: "Disable heuristic state prefetch during block import (less CPU and disk IO, more time waiting for data)",
@@ -1463,6 +1468,9 @@ func SetEthConfig(ctx *cli.Context, stack *node.Node, cfg *eth.Config) {
 	if ctx.GlobalIsSet(CacheFlag.Name) || ctx.GlobalIsSet(CacheGCFlag.Name) {
 		cfg.TrieDirtyCache = ctx.GlobalInt(CacheFlag.Name) * ctx.GlobalInt(CacheGCFlag.Name) / 100
 	}
+	if ctx.GlobalIsSet(CacheFlag.Name) || ctx.GlobalIsSet(CacheSnapshotFlag.Name) {
+		cfg.SnapshotCache = ctx.GlobalInt(CacheFlag.Name) * ctx.GlobalInt(CacheSnapshotFlag.Name) / 100
+	}
 	if ctx.GlobalIsSet(DocRootFlag.Name) {
 		cfg.DocRoot = ctx.GlobalString(DocRootFlag.Name)
 	}
@@ -1724,6 +1732,7 @@ func MakeChain(ctx *cli.Context, stack *node.Node) (chain *core.BlockChain, chai
 		TrieDirtyLimit:      eth.DefaultConfig.TrieDirtyCache,
 		TrieDirtyDisabled:   ctx.GlobalString(GCModeFlag.Name) == "archive",
 		TrieTimeLimit:       eth.DefaultConfig.TrieTimeout,
+		SnapshotLimit:       eth.DefaultConfig.SnapshotCache,
 	}
 	if ctx.GlobalIsSet(CacheFlag.Name) || ctx.GlobalIsSet(CacheTrieFlag.Name) {
 		cache.TrieCleanLimit = ctx.GlobalInt(CacheFlag.Name) * ctx.GlobalInt(CacheTrieFlag.Name) / 100

core/blockchain.go

@@ -62,8 +62,8 @@ var (
 	storageUpdateTimer = metrics.NewRegisteredTimer("chain/storage/updates", nil)
 	storageCommitTimer = metrics.NewRegisteredTimer("chain/storage/commits", nil)
 
-	snapshotAccountReadTimer = metrics.NewRegisteredTimer("chain/snapshot/accountreads", nil)
-	snapshotStorageReadTimer = metrics.NewRegisteredTimer("chain/snapshot/storagereads", nil)
+	snapshotAccountReadTimer = metrics.NewRegisteredTimer("chain/snapshot/account/reads", nil)
+	snapshotStorageReadTimer = metrics.NewRegisteredTimer("chain/snapshot/storage/reads", nil)
 	snapshotCommitTimer      = metrics.NewRegisteredTimer("chain/snapshot/commits", nil)
 
 	blockInsertTimer     = metrics.NewRegisteredTimer("chain/inserts", nil)
@@ -120,6 +120,7 @@ type CacheConfig struct {
 	TrieDirtyLimit      int           // Memory limit (MB) at which to start flushing dirty trie nodes to disk
 	TrieDirtyDisabled   bool          // Whether to disable trie write caching and GC altogether (archive node)
 	TrieTimeLimit       time.Duration // Time limit after which to flush the current in-memory trie to disk
+	SnapshotLimit       int           // Memory allowance (MB) to use for caching snapshot entries in memory
 }
 
 // BlockChain represents the canonical chain given a database with a genesis
@@ -194,6 +195,7 @@ func NewBlockChain(db ethdb.Database, cacheConfig *CacheConfig, chainConfig *par
 			TrieCleanLimit: 256,
 			TrieDirtyLimit: 256,
 			TrieTimeLimit:  5 * time.Minute,
+			SnapshotLimit:  256,
 		}
 	}
 	bodyCache, _ := lru.New(bodyCacheLimit)
@@ -300,10 +302,8 @@ func NewBlockChain(db ethdb.Database, cacheConfig *CacheConfig, chainConfig *par
 		}
 	}
 	// Load any existing snapshot, regenerating it if loading failed
-	head := bc.CurrentBlock()
-	if bc.snaps, err = snapshot.New(bc.db, "snapshot.rlp", head.Root()); err != nil {
-		return nil, err
-	}
+	bc.snaps = snapshot.New(bc.db, bc.stateCache.TrieDB(), "snapshot.rlp", bc.cacheConfig.SnapshotLimit, bc.CurrentBlock().Root())
+
 	// Take ownership of this particular state
 	go bc.update()
 	return bc, nil
@@ -497,6 +497,9 @@ func (bc *BlockChain) FastSyncCommitHead(hash common.Hash) error {
 	headBlockGauge.Update(int64(block.NumberU64()))
 	bc.chainmu.Unlock()
 
+	// Destroy any existing state snapshot and regenerate it in the background
+	bc.snaps.Rebuild(block.Root())
+
 	log.Info("Committed new head block", "number", block.Number(), "hash", hash)
 	return nil
 }
@@ -851,7 +854,8 @@ func (bc *BlockChain) Stop() {
 	bc.wg.Wait()
 
 	// Ensure that the entirety of the state snapshot is journalled to disk.
-	if err := bc.snaps.Journal(bc.CurrentBlock().Root()); err != nil {
+	snapBase, err := bc.snaps.Journal(bc.CurrentBlock().Root(), "snapshot.rlp")
+	if err != nil {
 		log.Error("Failed to journal state snapshot", "err", err)
 	}
 	// Ensure the state of a recent block is also stored to disk before exiting.
@@ -872,6 +876,12 @@ func (bc *BlockChain) Stop() {
 				}
 			}
 		}
+		if snapBase != (common.Hash{}) {
+			log.Info("Writing snapshot state to disk", "root", snapBase)
+			if err := triedb.Commit(snapBase, true); err != nil {
+				log.Error("Failed to commit recent state trie", "err", err)
+			}
+		}
 		for !bc.triegc.Empty() {
 			triedb.Dereference(bc.triegc.PopItem().(common.Hash))
 		}

core/rawdb/database.go

@@ -282,9 +282,9 @@ func InspectDatabase(db ethdb.Database) error {
 			receiptSize += size
 		case bytes.HasPrefix(key, txLookupPrefix) && len(key) == (len(txLookupPrefix)+common.HashLength):
 			txlookupSize += size
-		case bytes.HasPrefix(key, StateSnapshotPrefix) && len(key) == (len(StateSnapshotPrefix)+common.HashLength):
+		case bytes.HasPrefix(key, SnapshotAccountPrefix) && len(key) == (len(SnapshotAccountPrefix)+common.HashLength):
 			accountSnapSize += size
-		case bytes.HasPrefix(key, StateSnapshotPrefix) && len(key) == (len(StateSnapshotPrefix)+2*common.HashLength):
+		case bytes.HasPrefix(key, SnapshotStoragePrefix) && len(key) == (len(SnapshotStoragePrefix)+2*common.HashLength):
 			storageSnapSize += size
 		case bytes.HasPrefix(key, preimagePrefix) && len(key) == (len(preimagePrefix)+common.HashLength):
 			preimageSize += size

core/rawdb/schema.go

@@ -53,9 +53,10 @@ var (
 	blockBodyPrefix     = []byte("b") // blockBodyPrefix + num (uint64 big endian) + hash -> block body
 	blockReceiptsPrefix = []byte("r") // blockReceiptsPrefix + num (uint64 big endian) + hash -> block receipts
 
-	txLookupPrefix      = []byte("l") // txLookupPrefix + hash -> transaction/receipt lookup metadata
-	bloomBitsPrefix     = []byte("B") // bloomBitsPrefix + bit (uint16 big endian) + section (uint64 big endian) + hash -> bloom bits
-	StateSnapshotPrefix = []byte("s") // StateSnapshotPrefix + account hash [+ storage hash] -> account/storage trie value
+	txLookupPrefix        = []byte("l") // txLookupPrefix + hash -> transaction/receipt lookup metadata
+	bloomBitsPrefix       = []byte("B") // bloomBitsPrefix + bit (uint16 big endian) + section (uint64 big endian) + hash -> bloom bits
+	SnapshotAccountPrefix = []byte("a") // SnapshotAccountPrefix + account hash -> account trie value
+	SnapshotStoragePrefix = []byte("s") // SnapshotStoragePrefix + account hash + storage hash -> storage trie value
 
 	preimagePrefix = []byte("secure-key-")      // preimagePrefix + hash -> preimage
 	configPrefix   = []byte("ethereum-config-") // config prefix for the db
@@ -149,19 +150,19 @@ func txLookupKey(hash common.Hash) []byte {
 	return append(txLookupPrefix, hash.Bytes()...)
 }
 
-// accountSnapshotKey = StateSnapshotPrefix + hash
+// accountSnapshotKey = SnapshotAccountPrefix + hash
 func accountSnapshotKey(hash common.Hash) []byte {
-	return append(StateSnapshotPrefix, hash.Bytes()...)
+	return append(SnapshotAccountPrefix, hash.Bytes()...)
 }
 
-// storageSnapshotKey = StateSnapshotPrefix + account hash + storage hash
+// storageSnapshotKey = SnapshotStoragePrefix + account hash + storage hash
 func storageSnapshotKey(accountHash, storageHash common.Hash) []byte {
-	return append(append(StateSnapshotPrefix, accountHash.Bytes()...), storageHash.Bytes()...)
+	return append(append(SnapshotStoragePrefix, accountHash.Bytes()...), storageHash.Bytes()...)
 }
 
-// storageSnapshotsKey = StateSnapshotPrefix + account hash + storage hash
+// storageSnapshotsKey = SnapshotStoragePrefix + account hash + storage hash
 func storageSnapshotsKey(accountHash common.Hash) []byte {
-	return append(StateSnapshotPrefix, accountHash.Bytes()...)
+	return append(SnapshotStoragePrefix, accountHash.Bytes()...)
 }
 
 // bloomBitsKey = bloomBitsPrefix + bit (uint16 big endian) + section (uint64 big endian) + hash

core/state/snapshot/difflayer_test.go

@@ -24,7 +24,9 @@ import (
 	"path"
 	"testing"
 
+	"github.com/VictoriaMetrics/fastcache"
 	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/ethdb/memorydb"
 	"github.com/ethereum/go-ethereum/rlp"
 )
 
@@ -61,7 +63,7 @@ func TestMergeBasics(t *testing.T) {
 		}
 	}
 	// Add some (identical) layers on top
-	parent := newDiffLayer(emptyLayer{}, common.Hash{}, accounts, storage)
+	parent := newDiffLayer(emptyLayer(), common.Hash{}, accounts, storage)
 	child := newDiffLayer(parent, common.Hash{}, accounts, storage)
 	child = newDiffLayer(child, common.Hash{}, accounts, storage)
 	child = newDiffLayer(child, common.Hash{}, accounts, storage)
@@ -122,7 +124,7 @@ func TestMergeDelete(t *testing.T) {
 	}
 
 	// Add some flip-flopping layers on top
-	parent := newDiffLayer(emptyLayer{}, common.Hash{}, flip(), storage)
+	parent := newDiffLayer(emptyLayer(), common.Hash{}, flip(), storage)
 	child := parent.Update(common.Hash{}, flop(), storage)
 	child = child.Update(common.Hash{}, flip(), storage)
 	child = child.Update(common.Hash{}, flop(), storage)
@@ -165,7 +167,7 @@ func TestInsertAndMerge(t *testing.T) {
 	{
 		var accounts = make(map[common.Hash][]byte)
 		var storage = make(map[common.Hash]map[common.Hash][]byte)
-		parent = newDiffLayer(emptyLayer{}, common.Hash{}, accounts, storage)
+		parent = newDiffLayer(emptyLayer(), common.Hash{}, accounts, storage)
 	}
 	{
 		var accounts = make(map[common.Hash][]byte)
@@ -186,34 +188,11 @@ func TestInsertAndMerge(t *testing.T) {
 	}
 }
 
-type emptyLayer struct{}
-
-func (emptyLayer) Update(blockRoot common.Hash, accounts map[common.Hash][]byte, storage map[common.Hash]map[common.Hash][]byte) *diffLayer {
-	panic("implement me")
-}
-
-func (emptyLayer) Journal() error {
-	panic("implement me")
-}
-
-func (emptyLayer) Stale() bool {
-	panic("implement me")
-}
-
-func (emptyLayer) Root() common.Hash {
-	return common.Hash{}
-}
-
-func (emptyLayer) Account(hash common.Hash) (*Account, error) {
-	return nil, nil
-}
-
-func (emptyLayer) AccountRLP(hash common.Hash) ([]byte, error) {
-	return nil, nil
-}
-
-func (emptyLayer) Storage(accountHash, storageHash common.Hash) ([]byte, error) {
-	return nil, nil
+func emptyLayer() *diskLayer {
+	return &diskLayer{
+		diskdb: memorydb.New(),
+		cache:  fastcache.New(500 * 1024),
+	}
 }
 
 // BenchmarkSearch checks how long it takes to find a non-existing key
@@ -234,7 +213,7 @@ func BenchmarkSearch(b *testing.B) {
 		return newDiffLayer(parent, common.Hash{}, accounts, storage)
 	}
 	var layer snapshot
-	layer = emptyLayer{}
+	layer = emptyLayer()
 	for i := 0; i < 128; i++ {
 		layer = fill(layer)
 	}
@@ -272,7 +251,7 @@ func BenchmarkSearchSlot(b *testing.B) {
 		return newDiffLayer(parent, common.Hash{}, accounts, storage)
 	}
 	var layer snapshot
-	layer = emptyLayer{}
+	layer = emptyLayer()
 	for i := 0; i < 128; i++ {
 		layer = fill(layer)
 	}
@@ -313,7 +292,7 @@ func BenchmarkFlatten(b *testing.B) {
 	for i := 0; i < b.N; i++ {
 		b.StopTimer()
 		var layer snapshot
-		layer = emptyLayer{}
+		layer = emptyLayer()
 		for i := 1; i < 128; i++ {
 			layer = fill(layer)
 		}
@@ -357,17 +336,14 @@ func BenchmarkJournal(b *testing.B) {
 		}
 		return newDiffLayer(parent, common.Hash{}, accounts, storage)
 	}
-	var layer snapshot
-	layer = &diskLayer{
-		journal: path.Join(os.TempDir(), "difflayer_journal.tmp"),
-	}
+	layer := snapshot(new(diskLayer))
 	for i := 1; i < 128; i++ {
 		layer = fill(layer)
 	}
 	b.ResetTimer()
 
 	for i := 0; i < b.N; i++ {
-		f, _ := layer.(*diffLayer).journal()
+		f, _, _ := layer.Journal(path.Join(os.TempDir(), "difflayer_journal.tmp"))
 		f.Close()
 	}
 }

core/state/snapshot/disklayer.go

@@ -17,6 +17,7 @@
 package snapshot
 
 import (
+	"bytes"
 	"sync"
 
 	"github.com/VictoriaMetrics/fastcache"
@@ -24,17 +25,21 @@ import (
 	"github.com/ethereum/go-ethereum/core/rawdb"
 	"github.com/ethereum/go-ethereum/ethdb"
 	"github.com/ethereum/go-ethereum/rlp"
+	"github.com/ethereum/go-ethereum/trie"
 )
 
 // diskLayer is a low level persistent snapshot built on top of a key-value store.
 type diskLayer struct {
-	journal string              // Path of the snapshot journal to use on shutdown
-	db      ethdb.KeyValueStore // Key-value store containing the base snapshot
-	cache   *fastcache.Cache    // Cache to avoid hitting the disk for direct access
+	diskdb ethdb.KeyValueStore // Key-value store containing the base snapshot
+	triedb *trie.Database      // Trie node cache for reconstuction purposes
+	cache  *fastcache.Cache    // Cache to avoid hitting the disk for direct access
 
 	root  common.Hash // Root hash of the base snapshot
 	stale bool        // Signals that the layer became stale (state progressed)
 
+	genMarker []byte                    // Marker for the state that's indexed during initial layer generation
+	genAbort  chan chan *generatorStats // Notification channel to abort generating the snapshot in this layer
+
 	lock sync.RWMutex
 }
 
@@ -80,18 +85,26 @@ func (dl *diskLayer) AccountRLP(hash common.Hash) ([]byte, error) {
 	if dl.stale {
 		return nil, ErrSnapshotStale
 	}
+	// If the layer is being generated, ensure the requested hash has already been
+	// covered by the generator.
+	if dl.genMarker != nil && bytes.Compare(hash[:], dl.genMarker) > 0 {
+		return nil, ErrNotCoveredYet
+	}
+	// If we're in the disk layer, all diff layers missed
+	snapshotDirtyAccountMissMeter.Mark(1)
+
 	// Try to retrieve the account from the memory cache
-	if blob := dl.cache.Get(nil, hash[:]); blob != nil {
-		snapshotCleanHitMeter.Mark(1)
-		snapshotCleanReadMeter.Mark(int64(len(blob)))
+	if blob, found := dl.cache.HasGet(nil, hash[:]); found {
+		snapshotCleanAccountHitMeter.Mark(1)
+		snapshotCleanAccountReadMeter.Mark(int64(len(blob)))
 		return blob, nil
 	}
 	// Cache doesn't contain account, pull from disk and cache for later
-	blob := rawdb.ReadAccountSnapshot(dl.db, hash)
+	blob := rawdb.ReadAccountSnapshot(dl.diskdb, hash)
 	dl.cache.Set(hash[:], blob)
 
-	snapshotCleanMissMeter.Mark(1)
-	snapshotCleanWriteMeter.Mark(int64(len(blob)))
+	snapshotCleanAccountMissMeter.Mark(1)
+	snapshotCleanAccountWriteMeter.Mark(int64(len(blob)))
 
 	return blob, nil
 }
@@ -109,18 +122,26 @@ func (dl *diskLayer) Storage(accountHash, storageHash common.Hash) ([]byte, erro
 	}
 	key := append(accountHash[:], storageHash[:]...)
 
+	// If the layer is being generated, ensure the requested hash has already been
+	// covered by the generator.
+	if dl.genMarker != nil && bytes.Compare(key, dl.genMarker) > 0 {
+		return nil, ErrNotCoveredYet
+	}
+	// If we're in the disk layer, all diff layers missed
+	snapshotDirtyStorageMissMeter.Mark(1)
+
 	// Try to retrieve the storage slot from the memory cache
-	if blob := dl.cache.Get(nil, key); blob != nil {
-		snapshotCleanHitMeter.Mark(1)
-		snapshotCleanReadMeter.Mark(int64(len(blob)))
+	if blob, found := dl.cache.HasGet(nil, key); found {
+		snapshotCleanStorageHitMeter.Mark(1)
+		snapshotCleanStorageReadMeter.Mark(int64(len(blob)))
 		return blob, nil
 	}
 	// Cache doesn't contain storage slot, pull from disk and cache for later
-	blob := rawdb.ReadStorageSnapshot(dl.db, accountHash, storageHash)
+	blob := rawdb.ReadStorageSnapshot(dl.diskdb, accountHash, storageHash)
 	dl.cache.Set(key, blob)
 
-	snapshotCleanMissMeter.Mark(1)
-	snapshotCleanWriteMeter.Mark(int64(len(blob)))
+	snapshotCleanStorageMissMeter.Mark(1)
+	snapshotCleanStorageWriteMeter.Mark(int64(len(blob)))
 
 	return blob, nil
 }
@@ -131,9 +152,3 @@ func (dl *diskLayer) Storage(accountHash, storageHash common.Hash) ([]byte, erro
 func (dl *diskLayer) Update(blockHash common.Hash, accounts map[common.Hash][]byte, storage map[common.Hash]map[common.Hash][]byte) *diffLayer {
 	return newDiffLayer(dl, blockHash, accounts, storage)
 }
-
-// Journal commits an entire diff hierarchy to disk into a single journal file.
-func (dl *diskLayer) Journal() error {
-	// There's no journalling a disk layer
-	return nil
-}

core/state/snapshot/difflayer_journal.go → core/state/snapshot/journal.go

@@ -18,14 +18,32 @@ package snapshot
 
 import (
 	"bufio"
+	"encoding/binary"
+	"errors"
 	"fmt"
 	"io"
 	"os"
+	"time"
 
+	"github.com/VictoriaMetrics/fastcache"
 	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/core/rawdb"
+	"github.com/ethereum/go-ethereum/ethdb"
+	"github.com/ethereum/go-ethereum/log"
 	"github.com/ethereum/go-ethereum/rlp"
+	"github.com/ethereum/go-ethereum/trie"
 )
 
+// journalGenerator is a disk layer entry containing the generator progress marker.
+type journalGenerator struct {
+	Wiping   bool // Whether the database was in progress of being wiped
+	Done     bool // Whether the generator finished creating the snapshot
+	Marker   []byte
+	Accounts uint64
+	Slots    uint64
+	Storage  uint64
+}
+
 // journalAccount is an account entry in a diffLayer's disk journal.
 type journalAccount struct {
 	Hash common.Hash
@@ -39,6 +57,76 @@ type journalStorage struct {
 	Vals [][]byte
 }
 
+// loadSnapshot loads a pre-existing state snapshot backed by a key-value store.
+func loadSnapshot(diskdb ethdb.KeyValueStore, triedb *trie.Database, journal string, cache int, root common.Hash) (snapshot, error) {
+	// Retrieve the block number and hash of the snapshot, failing if no snapshot
+	// is present in the database (or crashed mid-update).
+	baseRoot := rawdb.ReadSnapshotRoot(diskdb)
+	if baseRoot == (common.Hash{}) {
+		return nil, errors.New("missing or corrupted snapshot")
+	}
+	base := &diskLayer{
+		diskdb: diskdb,
+		triedb: triedb,
+		cache:  fastcache.New(cache * 1024 * 1024),
+		root:   baseRoot,
+	}
+	// Open the journal, it must exist since even for 0 layer it stores whether
+	// we've already generated the snapshot or are in progress only
+	file, err := os.Open(journal)
+	if err != nil {
+		return nil, err
+	}
+	r := rlp.NewStream(file, 0)
+
+	// Read the snapshot generation progress for the disk layer
+	var generator journalGenerator
+	if err := r.Decode(&generator); err != nil {
+		return nil, fmt.Errorf("failed to load snapshot progress marker: %v", err)
+	}
+	// Load all the snapshot diffs from the journal
+	snapshot, err := loadDiffLayer(base, r)
+	if err != nil {
+		return nil, err
+	}
+	// Entire snapshot journal loaded, sanity check the head and return
+	// Journal doesn't exist, don't worry if it's not supposed to
+	if head := snapshot.Root(); head != root {
+		return nil, fmt.Errorf("head doesn't match snapshot: have %#x, want %#x", head, root)
+	}
+	// Everything loaded correctly, resume any suspended operations
+	if !generator.Done {
+		// If the generator was still wiping, restart one from scratch (fine for
+		// now as it's rare and the wiper deletes the stuff it touches anyway, so
+		// restarting won't incur a lot of extra database hops.
+		var wiper chan struct{}
+		if generator.Wiping {
+			log.Info("Resuming previous snapshot wipe")
+			wiper = wipeSnapshot(diskdb, false)
+		}
+		// Whether or not wiping was in progress, load any generator progress too
+		base.genMarker = generator.Marker
+		if base.genMarker == nil {
+			base.genMarker = []byte{}
+		}
+		base.genAbort = make(chan chan *generatorStats)
+
+		var origin uint64
+		if len(generator.Marker) >= 8 {
+			origin = binary.BigEndian.Uint64(generator.Marker)
+		}
+		go base.generate(&generatorStats{
+			wiping:   wiper,
+			origin:   origin,
+			start:    time.Now(),
+			accounts: generator.Accounts,
+			slots:    generator.Slots,
+			storage:  common.StorageSize(generator.Storage),
+		})
+	}
+	return snapshot, nil
+}
+
 // loadDiffLayer reads the next sections of a snapshot journal, reconstructing a new
 // diff and verifying that it can be linked to the requested parent.
 func loadDiffLayer(parent snapshot, r *rlp.Stream) (snapshot, error) {
@@ -74,39 +162,71 @@ func loadDiffLayer(parent snapshot, r *rlp.Stream) (snapshot, error) {
 	return loadDiffLayer(newDiffLayer(parent, root, accountData, storageData), r)
 }
 
-// journal is the internal version of Journal that also returns the journal file
+// Journal is the internal version of Journal that also returns the journal file
 // so subsequent layers know where to write to.
-func (dl *diffLayer) journal() (io.WriteCloser, error) {
-	// If we've reached the bottom, open the journal
-	var writer io.WriteCloser
-	if parent, ok := dl.parent.(*diskLayer); ok {
-		file, err := os.Create(parent.journal)
-		if err != nil {
-			return nil, err
+func (dl *diskLayer) Journal(path string) (io.WriteCloser, common.Hash, error) {
+	// If the snapshot is currenty being generated, abort it
+	var stats *generatorStats
+	if dl.genAbort != nil {
+		abort := make(chan *generatorStats)
+		dl.genAbort <- abort
+
+		if stats = <-abort; stats != nil {
+			stats.Log("Journalling in-progress snapshot", dl.genMarker)
 		}
-		writer = file
 	}
-	// If we haven't reached the bottom yet, journal the parent first
-	if writer == nil {
-		file, err := dl.parent.(*diffLayer).journal()
-		if err != nil {
-			return nil, err
-		}
-		writer = file
+	// Ensure the layer didn't get stale
+	dl.lock.RLock()
+	defer dl.lock.RUnlock()
+
+	if dl.stale {
+		return nil, common.Hash{}, ErrSnapshotStale
+	}
+	// We've reached the bottom, open the journal
+	file, err := os.Create(path)
+	if err != nil {
+		return nil, common.Hash{}, err
+	}
+	// Write out the generator marker
+	entry := journalGenerator{
+		Done:   dl.genMarker == nil,
+		Marker: dl.genMarker,
+	}
+	if stats != nil {
+		entry.Wiping = (stats.wiping != nil)
+		entry.Accounts = stats.accounts
+		entry.Slots = stats.slots
+		entry.Storage = uint64(stats.storage)
 	}
+	if err := rlp.Encode(file, entry); err != nil {
+		file.Close()
+		return nil, common.Hash{}, err
+	}
+	return file, dl.root, nil
+}
+
+// Journal is the internal version of Journal that also returns the journal file
+// so subsequent layers know where to write to.
+func (dl *diffLayer) Journal(path string) (io.WriteCloser, common.Hash, error) {
+	// Journal the parent first
+	writer, base, err := dl.parent.Journal(path)
+	if err != nil {
+		return nil, common.Hash{}, err
+	}
+	// Ensure the layer didn't get stale
 	dl.lock.RLock()
 	defer dl.lock.RUnlock()
 
 	if dl.stale {
 		writer.Close()
-		return nil, ErrSnapshotStale
+		return nil, common.Hash{}, ErrSnapshotStale
 	}
 	// Everything below was journalled, persist this layer too
 	buf := bufio.NewWriter(writer)
 	if err := rlp.Encode(buf, dl.root); err != nil {
 		buf.Flush()
 		writer.Close()
-		return nil, err
+		return nil, common.Hash{}, err
 	}
 	accounts := make([]journalAccount, 0, len(dl.accountData))
 	for hash, blob := range dl.accountData {
@@ -115,7 +235,7 @@ func (dl *diffLayer) journal() (io.WriteCloser, error) {
 	if err := rlp.Encode(buf, accounts); err != nil {
 		buf.Flush()
 		writer.Close()
-		return nil, err
+		return nil, common.Hash{}, err
 	}
 	storage := make([]journalStorage, 0, len(dl.storageData))
 	for hash, slots := range dl.storageData {
@@ -130,8 +250,8 @@ func (dl *diffLayer) journal() (io.WriteCloser, error) {
 	if err := rlp.Encode(buf, storage); err != nil {
 		buf.Flush()
 		writer.Close()
-		return nil, err
+		return nil, common.Hash{}, err
 	}
 	buf.Flush()
-	return writer, nil
+	return writer, base, nil
 }

core/state/snapshot/snapshot_test.go

@@ -31,9 +31,9 @@ import (
 func TestDiskLayerExternalInvalidationFullFlatten(t *testing.T) {
 	// Create an empty base layer and a snapshot tree out of it
 	base := &diskLayer{
-		db:    rawdb.NewMemoryDatabase(),
-		root:  common.HexToHash("0x01"),
-		cache: fastcache.New(1024 * 500),
+		diskdb: rawdb.NewMemoryDatabase(),
+		root:   common.HexToHash("0x01"),
+		cache:  fastcache.New(1024 * 500),
 	}
 	snaps := &Tree{
 		layers: map[common.Hash]snapshot{
@@ -54,7 +54,7 @@ func TestDiskLayerExternalInvalidationFullFlatten(t *testing.T) {
 		t.Errorf("pre-cap layer count mismatch: have %d, want %d", n, 2)
 	}
 	// Commit the diff layer onto the disk and ensure it's persisted
-	if err := snaps.Cap(common.HexToHash("0x02"), 0, 0); err != nil {
+	if err := snaps.Cap(common.HexToHash("0x02"), 0); err != nil {
 		t.Fatalf("failed to merge diff layer onto disk: %v", err)
 	}
 	// Since the base layer was modified, ensure that data retrievald on the external reference fail
@@ -76,9 +76,9 @@ func TestDiskLayerExternalInvalidationFullFlatten(t *testing.T) {
 func TestDiskLayerExternalInvalidationPartialFlatten(t *testing.T) {
 	// Create an empty base layer and a snapshot tree out of it
 	base := &diskLayer{
-		db:    rawdb.NewMemoryDatabase(),
-		root:  common.HexToHash("0x01"),
-		cache: fastcache.New(1024 * 500),
+		diskdb: rawdb.NewMemoryDatabase(),
+		root:   common.HexToHash("0x01"),
+		cache:  fastcache.New(1024 * 500),
 	}
 	snaps := &Tree{
 		layers: map[common.Hash]snapshot{
@@ -102,7 +102,10 @@ func TestDiskLayerExternalInvalidationPartialFlatten(t *testing.T) {
 		t.Errorf("pre-cap layer count mismatch: have %d, want %d", n, 3)
 	}
 	// Commit the diff layer onto the disk and ensure it's persisted
-	if err := snaps.Cap(common.HexToHash("0x03"), 2, 0); err != nil {
+	defer func(memcap uint64) { aggregatorMemoryLimit = memcap }(aggregatorMemoryLimit)
+	aggregatorMemoryLimit = 0
+
+	if err := snaps.Cap(common.HexToHash("0x03"), 2); err != nil {
 		t.Fatalf("failed to merge diff layer onto disk: %v", err)
 	}
 	// Since the base layer was modified, ensure that data retrievald on the external reference fail
@@ -124,9 +127,9 @@ func TestDiskLayerExternalInvalidationPartialFlatten(t *testing.T) {
 func TestDiffLayerExternalInvalidationFullFlatten(t *testing.T) {
 	// Create an empty base layer and a snapshot tree out of it
 	base := &diskLayer{
-		db:    rawdb.NewMemoryDatabase(),
-		root:  common.HexToHash("0x01"),
-		cache: fastcache.New(1024 * 500),
+		diskdb: rawdb.NewMemoryDatabase(),
+		root:   common.HexToHash("0x01"),
+		cache:  fastcache.New(1024 * 500),
 	}
 	snaps := &Tree{
 		layers: map[common.Hash]snapshot{
@@ -150,7 +153,7 @@ func TestDiffLayerExternalInvalidationFullFlatten(t *testing.T) {
 	ref := snaps.Snapshot(common.HexToHash("0x02"))
 
 	// Flatten the diff layer into the bottom accumulator
-	if err := snaps.Cap(common.HexToHash("0x03"), 1, 1024*1024); err != nil {
+	if err := snaps.Cap(common.HexToHash("0x03"), 1); err != nil {
 		t.Fatalf("failed to flatten diff layer into accumulator: %v", err)
 	}
 	// Since the accumulator diff layer was modified, ensure that data retrievald on the external reference fail
@@ -172,9 +175,9 @@ func TestDiffLayerExternalInvalidationFullFlatten(t *testing.T) {
 func TestDiffLayerExternalInvalidationPartialFlatten(t *testing.T) {
 	// Create an empty base layer and a snapshot tree out of it
 	base := &diskLayer{
-		db:    rawdb.NewMemoryDatabase(),
-		root:  common.HexToHash("0x01"),
-		cache: fastcache.New(1024 * 500),
+		diskdb: rawdb.NewMemoryDatabase(),
+		root:   common.HexToHash("0x01"),
+		cache:  fastcache.New(1024 * 500),
 	}
 	snaps := &Tree{
 		layers: map[common.Hash]snapshot{
@@ -202,14 +205,14 @@ func TestDiffLayerExternalInvalidationPartialFlatten(t *testing.T) {
 
 	// Doing a Cap operation with many allowed layers should be a no-op
 	exp := len(snaps.layers)
-	if err := snaps.Cap(common.HexToHash("0x04"), 2000, 1024*1024); err != nil {
+	if err := snaps.Cap(common.HexToHash("0x04"), 2000); err != nil {
 		t.Fatalf("failed to flatten diff layer into accumulator: %v", err)
 	}
 	if got := len(snaps.layers); got != exp {
 		t.Errorf("layers modified, got %d exp %d", got, exp)
 	}
 	// Flatten the diff layer into the bottom accumulator
-	if err := snaps.Cap(common.HexToHash("0x04"), 2, 1024*1024); err != nil {
+	if err := snaps.Cap(common.HexToHash("0x04"), 2); err != nil {
 		t.Fatalf("failed to flatten diff layer into accumulator: %v", err)
 	}
 	// Since the accumulator diff layer was modified, ensure that data retrievald on the external reference fail
@@ -236,9 +239,9 @@ func TestPostCapBasicDataAccess(t *testing.T) {
 	}
 	// Create a starting base layer and a snapshot tree out of it
 	base := &diskLayer{
-		db:    rawdb.NewMemoryDatabase(),
-		root:  common.HexToHash("0x01"),
-		cache: fastcache.New(1024 * 500),
+		diskdb: rawdb.NewMemoryDatabase(),
+		root:   common.HexToHash("0x01"),
+		cache:  fastcache.New(1024 * 500),
 	}
 	snaps := &Tree{
 		layers: map[common.Hash]snapshot{
@@ -280,11 +283,11 @@ func TestPostCapBasicDataAccess(t *testing.T) {
 		t.Error(err)
 	}
 	// Cap to a bad root should fail
-	if err := snaps.Cap(common.HexToHash("0x1337"), 0, 1024); err == nil {
+	if err := snaps.Cap(common.HexToHash("0x1337"), 0); err == nil {
 		t.Errorf("expected error, got none")
 	}
 	// Now, merge the a-chain
-	snaps.Cap(common.HexToHash("0xa3"), 0, 1024)
+	snaps.Cap(common.HexToHash("0xa3"), 0)
 
 	// At this point, a2 got merged into a1. Thus, a1 is now modified, and as a1 is
 	// the parent of b2, b2 should no longer be able to iterate into parent.
@@ -308,7 +311,7 @@ func TestPostCapBasicDataAccess(t *testing.T) {
 	}
 	// Now, merge it again, just for fun. It should now error, since a3
 	// is a disk layer
-	if err := snaps.Cap(common.HexToHash("0xa3"), 0, 1024); err == nil {
+	if err := snaps.Cap(common.HexToHash("0xa3"), 0); err == nil {
 		t.Error("expected error capping the disk layer, got none")
 	}
 }

core/state/snapshot/wipe.go

+// Copyright 2019 The go-ethereum Authors
+// This file is part of the go-ethereum library.
+//
+// The go-ethereum library is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// The go-ethereum library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Lesser General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public License
+// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
+
+package snapshot
+
+import (
+	"bytes"
+	"time"
+
+	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/core/rawdb"
+	"github.com/ethereum/go-ethereum/ethdb"
+	"github.com/ethereum/go-ethereum/log"
+)
+
+// wipeSnapshot starts a goroutine to iterate over the entire key-value database
+// and delete all the  data associated with the snapshot (accounts, storage,
+// metadata). After all is done, the snapshot range of the database is compacted
+// to free up unused data blocks.
+func wipeSnapshot(db ethdb.KeyValueStore, full bool) chan struct{} {
+	// Wipe the snapshot root marker synchronously
+	if full {
+		rawdb.DeleteSnapshotRoot(db)
+	}
+	// Wipe everything else asynchronously
+	wiper := make(chan struct{}, 1)
+	go func() {
+		if err := wipeContent(db); err != nil {
+			log.Error("Failed to wipe state snapshot", "err", err) // Database close will trigger this
+			return
+		}
+		close(wiper)
+	}()
+	return wiper
+}
+
+// wipeContent iterates over the entire key-value database and deletes all the
+// data associated with the snapshot (accounts, storage), but not the root hash
+// as the wiper is meant to run on a background thread but the root needs to be
+// removed in sync to avoid data races. After all is done, the snapshot range of
+// the database is compacted to free up unused data blocks.
+func wipeContent(db ethdb.KeyValueStore) error {
+	if err := wipeKeyRange(db, "accounts", rawdb.SnapshotAccountPrefix, len(rawdb.SnapshotAccountPrefix)+common.HashLength); err != nil {
+		return err
+	}
+	if err := wipeKeyRange(db, "storage", rawdb.SnapshotStoragePrefix, len(rawdb.SnapshotStoragePrefix)+2*common.HashLength); err != nil {
+		return err
+	}
+	// Compact the snapshot section of the database to get rid of unused space
+	start := time.Now()
+
+	log.Info("Compacting snapshot account area ")
+	end := common.CopyBytes(rawdb.SnapshotAccountPrefix)
+	end[len(end)-1]++
+
+	if err := db.Compact(rawdb.SnapshotAccountPrefix, end); err != nil {
+		return err
+	}
+	log.Info("Compacting snapshot storage area ")
+	end = common.CopyBytes(rawdb.SnapshotStoragePrefix)
+	end[len(end)-1]++
+
+	if err := db.Compact(rawdb.SnapshotStoragePrefix, end); err != nil {
+		return err
+	}
+	log.Info("Compacted snapshot area in database", "elapsed", common.PrettyDuration(time.Since(start)))
+
+	return nil
+}
+
+// wipeKeyRange deletes a range of keys from the database starting with prefix
+// and having a specific total key length.
+func wipeKeyRange(db ethdb.KeyValueStore, kind string, prefix []byte, keylen int) error {
+	// Batch deletions together to avoid holding an iterator for too long
+	var (
+		batch = db.NewBatch()
+		items int
+	)
+	// Iterate over the key-range and delete all of them
+	start, logged := time.Now(), time.Now()
+
+	it := db.NewIteratorWithStart(prefix)
+	for it.Next() {
+		// Skip any keys with the correct prefix but wrong lenth (trie nodes)
+		key := it.Key()
+		if !bytes.HasPrefix(key, prefix) {
+			break
+		}
+		if len(key) != keylen {
+			continue
+		}
+		// Delete the key and periodically recreate the batch and iterator
+		batch.Delete(key)
+		items++
+
+		if items%10000 == 0 {
+			// Batch too large (or iterator too long lived, flush and recreate)
+			it.Release()
+			if err := batch.Write(); err != nil {
+				return err
+			}
+			batch.Reset()
+			it = db.NewIteratorWithStart(key)
+
+			if time.Since(logged) > 8*time.Second {
+				log.Info("Deleting state snapshot leftovers", "kind", kind, "wiped", items, "elapsed", common.PrettyDuration(time.Since(start)))
+				logged = time.Now()
+			}
+		}
+	}
+	it.Release()
+	if err := batch.Write(); err != nil {
+		return err
+	}
+	log.Info("Deleted state snapshot leftovers", "kind", kind, "wiped", items, "elapsed", common.PrettyDuration(time.Since(start)))
+	return nil
+}

core/state/snapshot/generate_test.go → core/state/snapshot/wipe_test.go

@@ -59,17 +59,31 @@ func TestWipe(t *testing.T) {
 		// Randomize the suffix, dedup and inject it under the snapshot namespace
 		keysuffix := make([]byte, keysize)
 		rand.Read(keysuffix)
-		db.Put(append(rawdb.StateSnapshotPrefix, keysuffix...), randomHash().Bytes())
+
+		if rand.Int31n(2) == 0 {
+			db.Put(append(rawdb.SnapshotAccountPrefix, keysuffix...), randomHash().Bytes())
+		} else {
+			db.Put(append(rawdb.SnapshotStoragePrefix, keysuffix...), randomHash().Bytes())
+		}
 	}
 	// Sanity check that all the keys are present
 	var items int
 
-	it := db.NewIteratorWithPrefix(rawdb.StateSnapshotPrefix)
+	it := db.NewIteratorWithPrefix(rawdb.SnapshotAccountPrefix)
+	defer it.Release()
+
+	for it.Next() {
+		key := it.Key()
+		if len(key) == len(rawdb.SnapshotAccountPrefix)+common.HashLength {
+			items++
+		}
+	}
+	it = db.NewIteratorWithPrefix(rawdb.SnapshotStoragePrefix)
 	defer it.Release()
 
 	for it.Next() {
 		key := it.Key()
-		if len(key) == len(rawdb.StateSnapshotPrefix)+32 || len(key) == len(rawdb.StateSnapshotPrefix)+64 {
+		if len(key) == len(rawdb.SnapshotStoragePrefix)+2*common.HashLength {
 			items++
 		}
 	}
@@ -80,16 +94,24 @@ func TestWipe(t *testing.T) {
 		t.Errorf("snapshot block marker mismatch: have %#x, want <not-nil>", hash)
 	}
 	// Wipe all snapshot entries from the database
-	if err := wipeSnapshot(db); err != nil {
-		t.Fatalf("failed to wipe snapshot: %v", err)
-	}
+	<-wipeSnapshot(db, true)
+
 	// Iterate over the database end ensure no snapshot information remains
-	it = db.NewIteratorWithPrefix(rawdb.StateSnapshotPrefix)
+	it = db.NewIteratorWithPrefix(rawdb.SnapshotAccountPrefix)
+	defer it.Release()
+
+	for it.Next() {
+		key := it.Key()
+		if len(key) == len(rawdb.SnapshotAccountPrefix)+common.HashLength {
+			t.Errorf("snapshot entry remained after wipe: %x", key)
+		}
+	}
+	it = db.NewIteratorWithPrefix(rawdb.SnapshotStoragePrefix)
 	defer it.Release()
 
 	for it.Next() {
 		key := it.Key()
-		if len(key) == len(rawdb.StateSnapshotPrefix)+32 || len(key) == len(rawdb.StateSnapshotPrefix)+64 {
+		if len(key) == len(rawdb.SnapshotStoragePrefix)+2*common.HashLength {
 			t.Errorf("snapshot entry remained after wipe: %x", key)
 		}
 	}

core/state/statedb.go

@@ -845,8 +845,8 @@ func (s *StateDB) Commit(deleteEmptyObjects bool) (common.Hash, error) {
 			if err := s.snaps.Update(root, parent, s.snapAccounts, s.snapStorage); err != nil {
 				log.Warn("Failed to update snapshot tree", "from", parent, "to", root, "err", err)
 			}
-			if err := s.snaps.Cap(root, 16, 4*1024*1024); err != nil {
-				log.Warn("Failed to cap snapshot tree", "root", root, "layers", 16, "memory", 4*1024*1024, "err", err)
+			if err := s.snaps.Cap(root, 128); err != nil {
+				log.Warn("Failed to cap snapshot tree", "root", root, "layers", 128, "err", err)
 			}
 		}
 		s.snap, s.snapAccounts, s.snapStorage = nil, nil, nil

eth/backend.go

@@ -127,7 +127,8 @@ func New(ctx *node.ServiceContext, config *Config) (*Ethereum, error) {
 		config.Miner.GasPrice = new(big.Int).Set(DefaultConfig.Miner.GasPrice)
 	}
 	if config.NoPruning && config.TrieDirtyCache > 0 {
-		config.TrieCleanCache += config.TrieDirtyCache
+		config.TrieCleanCache += config.TrieDirtyCache * 3 / 5
+		config.SnapshotCache += config.TrieDirtyCache * 3 / 5
 		config.TrieDirtyCache = 0
 	}
 	log.Info("Allocated trie memory caches", "clean", common.StorageSize(config.TrieCleanCache)*1024*1024, "dirty", common.StorageSize(config.TrieDirtyCache)*1024*1024)
@@ -184,6 +185,7 @@ func New(ctx *node.ServiceContext, config *Config) (*Ethereum, error) {
 			TrieDirtyLimit:      config.TrieDirtyCache,
 			TrieDirtyDisabled:   config.NoPruning,
 			TrieTimeLimit:       config.TrieTimeout,
+			SnapshotLimit:       config.SnapshotCache,
 		}
 	)
 	eth.blockchain, err = core.NewBlockChain(chainDb, cacheConfig, chainConfig, eth.engine, vmConfig, eth.shouldPreserve)
@@ -204,7 +206,7 @@ func New(ctx *node.ServiceContext, config *Config) (*Ethereum, error) {
 	eth.txPool = core.NewTxPool(config.TxPool, chainConfig, eth.blockchain)
 
 	// Permit the downloader to use the trie cache allowance during fast sync
-	cacheLimit := cacheConfig.TrieCleanLimit + cacheConfig.TrieDirtyLimit
+	cacheLimit := cacheConfig.TrieCleanLimit + cacheConfig.TrieDirtyLimit + cacheConfig.SnapshotLimit
 	checkpoint := config.Checkpoint
 	if checkpoint == nil {
 		checkpoint = params.TrustedCheckpoints[genesisHash]

eth/config.go

@@ -50,6 +50,7 @@ var DefaultConfig = Config{
 	TrieCleanCache:     256,
 	TrieDirtyCache:     256,
 	TrieTimeout:        60 * time.Minute,
+	SnapshotCache:      256,
 	Miner: miner.Config{
 		GasFloor: 8000000,
 		GasCeil:  8000000,
@@ -125,6 +126,7 @@ type Config struct {
 	TrieCleanCache int
 	TrieDirtyCache int
 	TrieTimeout    time.Duration
+	SnapshotCache  int
 
 	// Mining options
 	Miner miner.Config

trie/iterator.go

@@ -29,7 +29,6 @@ import (
 type Iterator struct {
 	nodeIt NodeIterator
 
-	Nodes int    // Number of nodes iterated over
 	Key   []byte // Current data key on which the iterator is positioned on
 	Value []byte // Current data value on which the iterator is positioned on
 	Err   error
@@ -47,7 +46,6 @@ func NewIterator(it NodeIterator) *Iterator {
 // Next moves the iterator forward one key-value entry.
 func (it *Iterator) Next() bool {
 	for it.nodeIt.Next(true) {
-		it.Nodes++
 		if it.nodeIt.Leaf() {
 			it.Key = it.nodeIt.LeafKey()
 			it.Value = it.nodeIt.LeafBlob()