triedb/pathdb, eth: use double-buffer mechanism in pathdb

core/state/snapshot/generate_test.go

@@ -242,6 +242,18 @@ func (t *testHelper) Commit() common.Hash {
 	}
 	t.triedb.Update(root, types.EmptyRootHash, 0, t.nodes, t.states)
 	t.triedb.Commit(root, false)
+
+	// re-open the trie database to ensure the frozen buffer
+	// is not referenced
+	//config := &triedb.Config{}
+	//if t.triedb.Scheme() == rawdb.PathScheme {
+	//	config.PathDB = &pathdb.Config{
+	//		SnapshotNoBuild: true,
+	//	} // disable caching
+	//} else {
+	//	config.HashDB = &hashdb.Config{} // disable caching
+	//}
+	//t.triedb = triedb.NewDatabase(t.triedb.Disk(), config)
 	return root
 }
 

core/state/statedb_test.go

@@ -974,20 +974,23 @@ func TestMissingTrieNodes(t *testing.T) {
 func testMissingTrieNodes(t *testing.T, scheme string) {
 	// Create an initial state with a few accounts
 	var (
-		tdb   *triedb.Database
+		tdb    *triedb.Database
-		memDb = rawdb.NewMemoryDatabase()
+		memDb  = rawdb.NewMemoryDatabase()
+		openDb = func() *triedb.Database {
+			if scheme == rawdb.PathScheme {
+				return triedb.NewDatabase(memDb, &triedb.Config{PathDB: &pathdb.Config{
+					TrieCleanSize:   0,
+					StateCleanSize:  0,
+					WriteBufferSize: 0,
+				}}) // disable caching
+			} else {
+				return triedb.NewDatabase(memDb, &triedb.Config{HashDB: &hashdb.Config{
+					CleanCacheSize: 0,
+				}}) // disable caching
+			}
+		}
 	)
-	if scheme == rawdb.PathScheme {
+	tdb = openDb()
-		tdb = triedb.NewDatabase(memDb, &triedb.Config{PathDB: &pathdb.Config{
-			TrieCleanSize:   0,
-			StateCleanSize:  0,
-			WriteBufferSize: 0,
-		}}) // disable caching
-	} else {
-		tdb = triedb.NewDatabase(memDb, &triedb.Config{HashDB: &hashdb.Config{
-			CleanCacheSize: 0,
-		}}) // disable caching
-	}
 	db := NewDatabase(tdb, nil)
 
 	var root common.Hash
@@ -1005,17 +1008,27 @@ func testMissingTrieNodes(t *testing.T, scheme string) {
 		tdb.Commit(root, false)
 	}
 	// Create a new state on the old root
-	state, _ = New(root, db)
 	// Now we clear out the memdb
 	it := memDb.NewIterator(nil, nil)
 	for it.Next() {
 		k := it.Key()
-		// Leave the root intact
+		if scheme == rawdb.HashScheme {
-		if !bytes.Equal(k, root[:]) {
+			if !bytes.Equal(k, root[:]) {
-			t.Logf("key: %x", k)
+				t.Logf("key: %x", k)
-			memDb.Delete(k)
+				memDb.Delete(k)
+			}
+		}
+		if scheme == rawdb.PathScheme {
+			rk := k[len(rawdb.TrieNodeAccountPrefix):]
+			if len(rk) != 0 {
+				t.Logf("key: %x", k)
+				memDb.Delete(k)
+			}
 		}
 	}
+	tdb = openDb()
+	db = NewDatabase(tdb, nil)
+	state, _ = New(root, db)
 	balance := state.GetBalance(addr)
 	// The removed elem should lead to it returning zero balance
 	if exp, got := uint64(0), balance.Uint64(); got != exp {

triedb/pathdb/buffer.go

@@ -17,6 +17,7 @@
 package pathdb
 
 import (
+	"errors"
 	"fmt"
 	"time"
 
@@ -37,6 +38,9 @@ type buffer struct {
 	limit  uint64    // The maximum memory allowance in bytes
 	nodes  *nodeSet  // Aggregated trie node set
 	states *stateSet // Aggregated state set
+
+	done     chan struct{} // notifier whether the content in buffer has been flushed or not
+	flushErr error         // error if any exception occurs during flushing
 }
 
 // newBuffer initializes the buffer with the provided states and trie nodes.
@@ -124,43 +128,74 @@ func (b *buffer) size() uint64 {
 
 // flush persists the in-memory dirty trie node into the disk if the configured
 // memory threshold is reached. Note, all data must be written atomically.
-func (b *buffer) flush(root common.Hash, db ethdb.KeyValueStore, freezer ethdb.AncientWriter, progress []byte, nodesCache, statesCache *fastcache.Cache, id uint64) error {
+func (b *buffer) flush(root common.Hash, db ethdb.KeyValueStore, freezer ethdb.AncientWriter, progress []byte, nodesCache, statesCache *fastcache.Cache, id uint64, postFlush func()) {
-	// Ensure the target state id is aligned with the internal counter.
+	if b.done != nil {
-	head := rawdb.ReadPersistentStateID(db)
+		panic("duplicated flush operation")
-	if head+b.layers != id {
-		return fmt.Errorf("buffer layers (%d) cannot be applied on top of persisted state id (%d) to reach requested state id (%d)", b.layers, head, id)
 	}
-	// Terminate the state snapshot generation if it's active
+	b.done = make(chan struct{}) // allocate the channel for notification
-	var (
-		start = time.Now()
-		batch = db.NewBatchWithSize((b.nodes.dbsize() + b.states.dbsize()) * 11 / 10) // extra 10% for potential pebble internal stuff
-	)
-	// Explicitly sync the state freezer to ensure all written data is persisted to disk
-	// before updating the key-value store.
-	//
-	// This step is crucial to guarantee that the corresponding state history remains
-	// available for state rollback.
-	if freezer != nil {
-		if err := freezer.SyncAncient(); err != nil {
-			return err
-		}
-	}
-	nodes := b.nodes.write(batch, nodesCache)
-	accounts, slots := b.states.write(batch, progress, statesCache)
-	rawdb.WritePersistentStateID(batch, id)
-	rawdb.WriteSnapshotRoot(batch, root)
 
-	// Flush all mutations in a single batch
+	go func() {
-	size := batch.ValueSize()
+		defer func() {
-	if err := batch.Write(); err != nil {
+			if postFlush != nil {
-		return err
+				postFlush()
-	}
+			}
-	commitBytesMeter.Mark(int64(size))
+			close(b.done)
-	commitNodesMeter.Mark(int64(nodes))
+		}()
-	commitAccountsMeter.Mark(int64(accounts))
+
-	commitStoragesMeter.Mark(int64(slots))
+		// Ensure the target state id is aligned with the internal counter.
-	commitTimeTimer.UpdateSince(start)
+		head := rawdb.ReadPersistentStateID(db)
-	b.reset()
+		if head+b.layers != id {
-	log.Debug("Persisted buffer content", "nodes", nodes, "accounts", accounts, "slots", slots, "bytes", common.StorageSize(size), "elapsed", common.PrettyDuration(time.Since(start)))
+			b.flushErr = fmt.Errorf("buffer layers (%d) cannot be applied on top of persisted state id (%d) to reach requested state id (%d)", b.layers, head, id)
-	return nil
+			return
+		}
+
+		// Terminate the state snapshot generation if it's active
+		var (
+			start = time.Now()
+			batch = db.NewBatchWithSize((b.nodes.dbsize() + b.states.dbsize()) * 11 / 10) // extra 10% for potential pebble internal stuff
+		)
+		// Explicitly sync the state freezer to ensure all written data is persisted to disk
+		// before updating the key-value store.
+		//
+		// This step is crucial to guarantee that the corresponding state history remains
+		// available for state rollback.
+		if freezer != nil {
+			if err := freezer.SyncAncient(); err != nil {
+				b.flushErr = err
+				return
+			}
+		}
+		nodes := b.nodes.write(batch, nodesCache)
+		accounts, slots := b.states.write(batch, progress, statesCache)
+		rawdb.WritePersistentStateID(batch, id)
+		rawdb.WriteSnapshotRoot(batch, root)
+
+		// Flush all mutations in a single batch
+		size := batch.ValueSize()
+		if err := batch.Write(); err != nil {
+			b.flushErr = err
+			return
+		}
+		commitBytesMeter.Mark(int64(size))
+		commitNodesMeter.Mark(int64(nodes))
+		commitAccountsMeter.Mark(int64(accounts))
+		commitStoragesMeter.Mark(int64(slots))
+		commitTimeTimer.UpdateSince(start)
+
+		// The content in the frozen buffer is kept for consequent state access,
+		// TODO (rjl493456442) measure the gc overhead for holding this struct.
+		// TODO (rjl493456442) can we somehow get rid of it after flushing??
+		b.reset()
+		log.Debug("Persisted buffer content", "nodes", nodes, "accounts", accounts, "slots", slots, "bytes", common.StorageSize(size), "elapsed", common.PrettyDuration(time.Since(start)))
+	}()
+}
+
+// waitFlush blocks until the buffer has been fully flushed and returns any
+// stored errors that occurred during the process.
+func (b *buffer) waitFlush() error {
+	if b.done == nil {
+		return errors.New("the buffer is not frozen")
+	}
+	<-b.done
+	return b.flushErr
 }

triedb/pathdb/database.go

@@ -119,7 +119,10 @@ type Config struct {
 	StateCleanSize  int    // Maximum memory allowance (in bytes) for caching clean state data
 	WriteBufferSize int    // Maximum memory allowance (in bytes) for write buffer
 	ReadOnly        bool   // Flag whether the database is opened in read only mode
-	SnapshotNoBuild bool   // Flag Whether the background generation is allowed
+
+	// Testing configurations
+	SnapshotNoBuild bool // Flag Whether the background generation is allowed
+	NoAsyncFlush    bool // Flag whether the background generation is allowed
 }
 
 // sanitize checks the provided user configurations and changes anything that's
@@ -434,6 +437,9 @@ func (db *Database) Disable() error {
 	// Terminate the state generator if it's active and mark the disk layer
 	// as stale to prevent access to persistent state.
 	disk := db.tree.bottom()
+	if err := disk.waitFlush(); err != nil {
+		return err
+	}
 	if disk.generator != nil {
 		disk.generator.stop()
 	}
@@ -592,12 +598,18 @@ func (db *Database) Close() error {
 	// following mutations.
 	db.readOnly = true
 
-	// Terminate the background generation if it's active
+	// Block until the background flushing is finished. It must
-	disk := db.tree.bottom()
+	// be done before terminating the potential background snapshot
-	if disk.generator != nil {
+	// generator.
-		disk.generator.stop()
+	dl := db.tree.bottom()
+	if err := dl.waitFlush(); err != nil {
+		return err
 	}
-	disk.resetCache() // release the memory held by clean cache
+	// Terminate the background generation if it's active
+	if dl.generator != nil {
+		dl.generator.stop()
+	}
+	dl.resetCache() // release the memory held by clean cache
 
 	// Close the attached state history freezer.
 	if db.freezer == nil {

triedb/pathdb/database_test.go

@@ -129,6 +129,7 @@ func newTester(t *testing.T, historyLimit uint64, isVerkle bool, layers int) *te
 			TrieCleanSize:   256 * 1024,
 			StateCleanSize:  256 * 1024,
 			WriteBufferSize: 256 * 1024,
+			NoAsyncFlush:    true,
 		}, isVerkle)
 
 		obj = &tester{

triedb/pathdb/disklayer.go

@@ -41,9 +41,11 @@ type diskLayer struct {
 	nodes  *fastcache.Cache // GC friendly memory cache of clean nodes
 	states *fastcache.Cache // GC friendly memory cache of clean states
 
-	buffer *buffer      // Dirty buffer to aggregate writes of nodes and states
+	buffer *buffer // Live buffer to aggregate writes
-	stale  bool         // Signals that the layer became stale (state progressed)
+	frozen *buffer // Frozen node buffer waiting for flushing
-	lock   sync.RWMutex // Lock used to protect stale flag and genMarker
+
+	stale bool         // Signals that the layer became stale (state progressed)
+	lock  sync.RWMutex // Lock used to protect stale flag and genMarker
 
 	// The generator is set if the state snapshot was not fully completed,
 	// regardless of whether the background generation is running or not.
@@ -51,7 +53,7 @@ type diskLayer struct {
 }
 
 // newDiskLayer creates a new disk layer based on the passing arguments.
-func newDiskLayer(root common.Hash, id uint64, db *Database, nodes *fastcache.Cache, states *fastcache.Cache, buffer *buffer) *diskLayer {
+func newDiskLayer(root common.Hash, id uint64, db *Database, nodes *fastcache.Cache, states *fastcache.Cache, buffer *buffer, frozen *buffer) *diskLayer {
 	// Initialize the clean caches if the memory allowance is not zero
 	// or reuse the provided caches if they are not nil (inherited from
 	// the original disk layer).
@@ -68,6 +70,7 @@ func newDiskLayer(root common.Hash, id uint64, db *Database, nodes *fastcache.Ca
 		nodes:  nodes,
 		states: states,
 		buffer: buffer,
+		frozen: frozen,
 	}
 }
 
@@ -114,16 +117,19 @@ func (dl *diskLayer) node(owner common.Hash, path []byte, depth int) ([]byte, co
 	if dl.stale {
 		return nil, common.Hash{}, nil, errSnapshotStale
 	}
-	// Try to retrieve the trie node from the not-yet-written
+	// Try to retrieve the trie node from the not-yet-written node buffer first
-	// node buffer first. Note the buffer is lock free since
+	// (both the live one and the frozen one). Note the buffer is lock free since
-	// it's impossible to mutate the buffer before tagging the
+	// it's impossible to mutate the buffer before tagging the layer as stale.
-	// layer as stale.
+	for _, buffer := range []*buffer{dl.buffer, dl.frozen} {
-	n, found := dl.buffer.node(owner, path)
+		if buffer != nil && !buffer.empty() {
-	if found {
+			n, found := buffer.node(owner, path)
-		dirtyNodeHitMeter.Mark(1)
+			if found {
-		dirtyNodeReadMeter.Mark(int64(len(n.Blob)))
+				dirtyNodeHitMeter.Mark(1)
-		dirtyNodeHitDepthHist.Update(int64(depth))
+				dirtyNodeReadMeter.Mark(int64(len(n.Blob)))
-		return n.Blob, n.Hash, &nodeLoc{loc: locDirtyCache, depth: depth}, nil
+				dirtyNodeHitDepthHist.Update(int64(depth))
+				return n.Blob, n.Hash, &nodeLoc{loc: locDirtyCache, depth: depth}, nil
+			}
+		}
 	}
 	dirtyNodeMissMeter.Mark(1)
 
@@ -144,6 +150,11 @@ func (dl *diskLayer) node(owner common.Hash, path []byte, depth int) ([]byte, co
 	} else {
 		blob = rawdb.ReadStorageTrieNode(dl.db.diskdb, owner, path)
 	}
+	// Store the resolved data in the clean cache. The background buffer flusher
+	// may also write to the clean cache concurrently, but two writers cannot
+	// write the same item with different content. If the item already exists,
+	// it will be found in the frozen buffer, eliminating the need to check the
+	// database.
 	if dl.nodes != nil && len(blob) > 0 {
 		dl.nodes.Set(key, blob)
 		cleanNodeWriteMeter.Mark(int64(len(blob)))
@@ -162,22 +173,25 @@ func (dl *diskLayer) account(hash common.Hash, depth int) ([]byte, error) {
 	if dl.stale {
 		return nil, errSnapshotStale
 	}
-	// Try to retrieve the account from the not-yet-written
+	// Try to retrieve the trie node from the not-yet-written node buffer first
-	// node buffer first. Note the buffer is lock free since
+	// (both the live one and the frozen one). Note the buffer is lock free since
-	// it's impossible to mutate the buffer before tagging the
+	// it's impossible to mutate the buffer before tagging the layer as stale.
-	// layer as stale.
+	for _, buffer := range []*buffer{dl.buffer, dl.frozen} {
-	blob, found := dl.buffer.account(hash)
+		if buffer != nil && !buffer.empty() {
-	if found {
+			blob, found := buffer.account(hash)
-		dirtyStateHitMeter.Mark(1)
+			if found {
-		dirtyStateReadMeter.Mark(int64(len(blob)))
+				dirtyStateHitMeter.Mark(1)
-		dirtyStateHitDepthHist.Update(int64(depth))
+				dirtyStateReadMeter.Mark(int64(len(blob)))
+				dirtyStateHitDepthHist.Update(int64(depth))
 
-		if len(blob) == 0 {
+				if len(blob) == 0 {
-			stateAccountInexMeter.Mark(1)
+					stateAccountInexMeter.Mark(1)
-		} else {
+				} else {
-			stateAccountExistMeter.Mark(1)
+					stateAccountExistMeter.Mark(1)
+				}
+				return blob, nil
+			}
 		}
-		return blob, nil
 	}
 	dirtyStateMissMeter.Mark(1)
 
@@ -203,7 +217,13 @@ func (dl *diskLayer) account(hash common.Hash, depth int) ([]byte, error) {
 		cleanStateMissMeter.Mark(1)
 	}
 	// Try to retrieve the account from the disk.
-	blob = rawdb.ReadAccountSnapshot(dl.db.diskdb, hash)
+	blob := rawdb.ReadAccountSnapshot(dl.db.diskdb, hash)
+
+	// Store the resolved data in the clean cache. The background buffer flusher
+	// may also write to the clean cache concurrently, but two writers cannot
+	// write the same item with different content. If the item already exists,
+	// it will be found in the frozen buffer, eliminating the need to check the
+	// database.
 	if dl.states != nil {
 		dl.states.Set(hash[:], blob)
 		cleanStateWriteMeter.Mark(int64(len(blob)))
@@ -231,21 +251,24 @@ func (dl *diskLayer) storage(accountHash, storageHash common.Hash, depth int) ([
 	if dl.stale {
 		return nil, errSnapshotStale
 	}
-	// Try to retrieve the storage slot from the not-yet-written
+	// Try to retrieve the trie node from the not-yet-written node buffer first
-	// node buffer first. Note the buffer is lock free since
+	// (both the live one and the frozen one). Note the buffer is lock free since
-	// it's impossible to mutate the buffer before tagging the
+	// it's impossible to mutate the buffer before tagging the layer as stale.
-	// layer as stale.
+	for _, buffer := range []*buffer{dl.buffer, dl.frozen} {
-	if blob, found := dl.buffer.storage(accountHash, storageHash); found {
+		if buffer != nil && !buffer.empty() {
-		dirtyStateHitMeter.Mark(1)
+			if blob, found := buffer.storage(accountHash, storageHash); found {
-		dirtyStateReadMeter.Mark(int64(len(blob)))
+				dirtyStateHitMeter.Mark(1)
-		dirtyStateHitDepthHist.Update(int64(depth))
+				dirtyStateReadMeter.Mark(int64(len(blob)))
+				dirtyStateHitDepthHist.Update(int64(depth))
 
-		if len(blob) == 0 {
+				if len(blob) == 0 {
-			stateStorageInexMeter.Mark(1)
+					stateStorageInexMeter.Mark(1)
-		} else {
+				} else {
-			stateStorageExistMeter.Mark(1)
+					stateStorageExistMeter.Mark(1)
+				}
+				return blob, nil
+			}
 		}
-		return blob, nil
 	}
 	dirtyStateMissMeter.Mark(1)
 
@@ -273,6 +296,12 @@ func (dl *diskLayer) storage(accountHash, storageHash common.Hash, depth int) ([
 	}
 	// Try to retrieve the account from the disk
 	blob := rawdb.ReadStorageSnapshot(dl.db.diskdb, accountHash, storageHash)
+
+	// Store the resolved data in the clean cache. The background buffer flusher
+	// may also write to the clean cache concurrently, but two writers cannot
+	// write the same item with different content. If the item already exists,
+	// it will be found in the frozen buffer, eliminating the need to check the
+	// database.
 	if dl.states != nil {
 		dl.states.Set(key, blob)
 		cleanStateWriteMeter.Mark(int64(len(blob)))
@@ -341,7 +370,8 @@ func (dl *diskLayer) commit(bottom *diffLayer, force bool) (*diskLayer, error) {
 	// truncation) surpasses the persisted state ID, we take the necessary action
 	// of forcibly committing the cached dirty states to ensure that the persisted
 	// state ID remains higher.
-	if !force && rawdb.ReadPersistentStateID(dl.db.diskdb) < oldest {
+	persistedID := rawdb.ReadPersistentStateID(dl.db.diskdb)
+	if !force && persistedID < oldest {
 		force = true
 	}
 	// Merge the trie nodes and flat states of the bottom-most diff layer into the
@@ -351,32 +381,67 @@ func (dl *diskLayer) commit(bottom *diffLayer, force bool) (*diskLayer, error) {
 	// Terminate the background state snapshot generation before mutating the
 	// persistent state.
 	if combined.full() || force {
+		// Wait until the previous frozen buffer is fully flushed
+		if dl.frozen != nil {
+			if err := dl.frozen.waitFlush(); err != nil {
+				return nil, err
+			}
+		}
+		// Release the frozen buffer and the internally referenced maps will
+		// be reclaimed by GC.
+		dl.frozen = nil
+
 		// Terminate the background state snapshot generator before flushing
 		// to prevent data race.
-		var progress []byte
+		var (
-		if dl.generator != nil {
+			progress []byte
-			dl.generator.stop()
+			gen      = dl.generator
-			progress = dl.generator.progressMarker()
+		)
+		if gen != nil {
+			gen.stop()
+			progress = gen.progressMarker()
 
 			// If the snapshot has been fully generated, unset the generator
 			if progress == nil {
+				gen = nil
 				dl.setGenerator(nil)
 			} else {
 				log.Info("Paused snapshot generation")
 			}
 		}
-		// Flush the content in combined buffer. Any state data after the progress
+
-		// marker will be ignored, as the generator will pick it up later.
+		// Freeze the live buffer and schedule background flushing
-		if err := combined.flush(bottom.root, dl.db.diskdb, dl.db.freezer, progress, dl.nodes, dl.states, bottom.stateID()); err != nil {
+		dl.frozen = combined
-			return nil, err
+		dl.frozen.flush(bottom.root, dl.db.diskdb, dl.db.freezer, progress, dl.nodes, dl.states, bottom.stateID(), func() {
+			// Resume the background generation if it's not completed yet.
+			// The generator is assumed to be available if the progress is
+			// not nil.
+			//
+			// Notably, the generator will be shared and linked by all the
+			// disk layer instances, regardless of the generation is terminated
+			// or not.
+			if progress != nil {
+				gen.run(bottom.root)
+			}
+		})
+		// Block until the frozen buffer is fully flushed out if the async flushing
+		// is not allowed.
+		if dl.db.config.NoAsyncFlush {
+			if err := dl.frozen.waitFlush(); err != nil {
+				return nil, err
+			}
 		}
-		// Resume the background generation if it's not completed yet
+		// Block until the frozen buffer is fully flushed out if the oldest history
-		if progress != nil {
+		// surpasses the persisted state ID.
-			dl.generator.run(bottom.root)
+		if persistedID < oldest {
+			if err := dl.frozen.waitFlush(); err != nil {
+				return nil, err
+			}
 		}
+		combined = newBuffer(dl.db.config.WriteBufferSize, nil, nil, 0)
 	}
 	// Link the generator if snapshot is not yet completed
-	ndl := newDiskLayer(bottom.root, bottom.stateID(), dl.db, dl.nodes, dl.states, combined)
+	ndl := newDiskLayer(bottom.root, bottom.stateID(), dl.db, dl.nodes, dl.states, combined, dl.frozen)
 	if dl.generator != nil {
 		ndl.setGenerator(dl.generator)
 	}
@@ -428,7 +493,7 @@ func (dl *diskLayer) revert(h *history) (*diskLayer, error) {
 		if err != nil {
 			return nil, err
 		}
-		ndl := newDiskLayer(h.meta.parent, dl.id-1, dl.db, dl.nodes, dl.states, dl.buffer)
+		ndl := newDiskLayer(h.meta.parent, dl.id-1, dl.db, dl.nodes, dl.states, dl.buffer, dl.frozen)
 
 		// Link the generator if it exists
 		if dl.generator != nil {
@@ -437,6 +502,16 @@ func (dl *diskLayer) revert(h *history) (*diskLayer, error) {
 		log.Debug("Reverted data in write buffer", "oldroot", h.meta.root, "newroot", h.meta.parent, "elapsed", common.PrettyDuration(time.Since(start)))
 		return ndl, nil
 	}
+	// Block until the frozen buffer is fully flushed
+	if dl.frozen != nil {
+		if err := dl.frozen.waitFlush(); err != nil {
+			return nil, err
+		}
+		// Unset the frozen buffer if it exists, otherwise these "reverted"
+		// states will still be accessible after revert in frozen buffer.
+		dl.frozen = nil
+	}
+
 	// Terminate the generation before writing any data into database
 	var progress []byte
 	if dl.generator != nil {
@@ -455,7 +530,7 @@ func (dl *diskLayer) revert(h *history) (*diskLayer, error) {
 	}
 	// Link the generator and resume generation if the snapshot is not yet
 	// fully completed.
-	ndl := newDiskLayer(h.meta.parent, dl.id-1, dl.db, dl.nodes, dl.states, dl.buffer)
+	ndl := newDiskLayer(h.meta.parent, dl.id-1, dl.db, dl.nodes, dl.states, dl.buffer, dl.frozen)
 	if dl.generator != nil && !dl.generator.completed() {
 		ndl.generator = dl.generator
 		ndl.generator.run(h.meta.parent)
@@ -500,3 +575,12 @@ func (dl *diskLayer) genMarker() []byte {
 	}
 	return dl.generator.progressMarker()
 }
+
+// waitFlush blocks until the buffer has been fully flushed and returns any
+// stored errors that occurred during the process.
+func (dl *diskLayer) waitFlush() error {
+	if dl.frozen == nil {
+		return nil
+	}
+	return dl.frozen.waitFlush()
+}

triedb/pathdb/generate.go

@@ -186,7 +186,7 @@ func generateSnapshot(triedb *Database, root common.Hash, noBuild bool) *diskLay
 		stats     = &generatorStats{start: time.Now()}
 		genMarker = []byte{} // Initialized but empty!
 	)
-	dl := newDiskLayer(root, 0, triedb, nil, nil, newBuffer(triedb.config.WriteBufferSize, nil, nil, 0))
+	dl := newDiskLayer(root, 0, triedb, nil, nil, newBuffer(triedb.config.WriteBufferSize, nil, nil, 0), nil)
 	dl.setGenerator(newGenerator(triedb.diskdb, noBuild, genMarker, stats))
 
 	if !noBuild {

triedb/pathdb/generate_test.go

@@ -49,6 +49,7 @@ func newGenTester() *genTester {
 	disk := rawdb.NewMemoryDatabase()
 	config := *Defaults
 	config.SnapshotNoBuild = true // no background generation
+	config.NoAsyncFlush = true    // no async flush
 	db := New(disk, &config, false)
 	tr, _ := trie.New(trie.StateTrieID(types.EmptyRootHash), db)
 	return &genTester{

triedb/pathdb/iterator_test.go

@@ -255,6 +255,7 @@ func TestFastIteratorBasics(t *testing.T) {
 func TestAccountIteratorTraversal(t *testing.T) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -299,6 +300,7 @@ func TestAccountIteratorTraversal(t *testing.T) {
 func TestStorageIteratorTraversal(t *testing.T) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -343,6 +345,7 @@ func TestStorageIteratorTraversal(t *testing.T) {
 func TestAccountIteratorTraversalValues(t *testing.T) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -459,6 +462,7 @@ func TestAccountIteratorTraversalValues(t *testing.T) {
 func TestStorageIteratorTraversalValues(t *testing.T) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -592,6 +596,7 @@ func TestAccountIteratorLargeTraversal(t *testing.T) {
 	// Build up a large stack of snapshots
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -664,6 +669,7 @@ func TestAccountIteratorFlattening(t *testing.T) {
 func TestAccountIteratorSeek(t *testing.T) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -736,6 +742,7 @@ func TestStorageIteratorSeek(t *testing.T) {
 func testStorageIteratorSeek(t *testing.T, newIterator func(db *Database, root, account, seek common.Hash) StorageIterator) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -808,6 +815,7 @@ func TestAccountIteratorDeletions(t *testing.T) {
 func testAccountIteratorDeletions(t *testing.T, newIterator func(db *Database, root, seek common.Hash) AccountIterator) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -848,6 +856,7 @@ func testAccountIteratorDeletions(t *testing.T, newIterator func(db *Database, r
 func TestStorageIteratorDeletions(t *testing.T) {
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -916,6 +925,7 @@ func TestStaleIterator(t *testing.T) {
 func testStaleIterator(t *testing.T, newIter func(db *Database, hash common.Hash) Iterator) {
 	config := &Config{
 		WriteBufferSize: 16 * 1024 * 1024,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -971,6 +981,7 @@ func BenchmarkAccountIteratorTraversal(b *testing.B) {
 	}
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()
@@ -1066,6 +1077,7 @@ func BenchmarkAccountIteratorLargeBaselayer(b *testing.B) {
 	}
 	config := &Config{
 		WriteBufferSize: 0,
+		NoAsyncFlush:    true,
 	}
 	db := New(rawdb.NewMemoryDatabase(), config, false)
 	db.waitGeneration()

triedb/pathdb/journal.go

@@ -160,7 +160,7 @@ func (db *Database) loadLayers() layer {
 		log.Info("Failed to load journal, discard it", "err", err)
 	}
 	// Return single layer with persistent state.
-	return newDiskLayer(root, rawdb.ReadPersistentStateID(db.diskdb), db, nil, nil, newBuffer(db.config.WriteBufferSize, nil, nil, 0))
+	return newDiskLayer(root, rawdb.ReadPersistentStateID(db.diskdb), db, nil, nil, newBuffer(db.config.WriteBufferSize, nil, nil, 0), nil)
 }
 
 // loadDiskLayer reads the binary blob from the layer journal, reconstructing
@@ -192,7 +192,7 @@ func (db *Database) loadDiskLayer(r *rlp.Stream) (layer, error) {
 	if err := states.decode(r); err != nil {
 		return nil, err
 	}
-	return newDiskLayer(root, id, db, nil, nil, newBuffer(db.config.WriteBufferSize, &nodes, &states, id-stored)), nil
+	return newDiskLayer(root, id, db, nil, nil, newBuffer(db.config.WriteBufferSize, &nodes, &states, id-stored), nil), nil
 }
 
 // loadDiffLayer reads the next sections of a layer journal, reconstructing a new
@@ -301,6 +301,12 @@ func (db *Database) Journal(root common.Hash) error {
 	} else { // disk layer only on noop runs (likely) or deep reorgs (unlikely)
 		log.Info("Persisting dirty state to disk", "root", root, "layers", disk.buffer.layers)
 	}
+	// Block until the background flushing is finished. It must
+	// be done before terminating the potential background snapshot
+	// generator.
+	if err := disk.waitFlush(); err != nil {
+		return err
+	}
 	// Terminate the background state generation if it's active
 	if disk.generator != nil {
 		disk.generator.stop()

triedb/pathdb/layertree.go

@@ -195,6 +195,10 @@ func (tree *layerTree) cap(root common.Hash, layers int) error {
 		}
 		tree.base = base
 
+		// Block until the frozen buffer is fully flushed
+		if err := base.waitFlush(); err != nil {
+			return err
+		}
 		// Reset the layer tree with the single new disk layer
 		tree.layers = map[common.Hash]layer{
 			base.rootHash(): base,

triedb/pathdb/layertree_test.go

@@ -27,7 +27,7 @@ import (
 
 func newTestLayerTree() *layerTree {
 	db := New(rawdb.NewMemoryDatabase(), nil, false)
-	l := newDiskLayer(common.Hash{0x1}, 0, db, nil, nil, newBuffer(0, nil, nil, 0))
+	l := newDiskLayer(common.Hash{0x1}, 0, db, nil, nil, newBuffer(0, nil, nil, 0), nil)
 	t := newLayerTree(l)
 	return t
 }