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

Author: rjl493456442

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
-		memDb = rawdb.NewMemoryDatabase()
+		tdb    *triedb.Database
+		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 = 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
-	}
+	tdb = openDb()
 	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 !bytes.Equal(k, root[:]) {
-			t.Logf("key: %x", k)
-			memDb.Delete(k)
+		if scheme == rawdb.HashScheme {
+			if !bytes.Equal(k, root[:]) {
+				t.Logf("key: %x", 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 {
-	// Ensure the target state id is aligned with the internal counter.
-	head := rawdb.ReadPersistentStateID(db)
-	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)
+func (b *buffer) flush(root common.Hash, db ethdb.KeyValueStore, freezer ethdb.AncientWriter, progress []byte, nodesCache, statesCache *fastcache.Cache, id uint64, postFlush func()) {
+	if b.done != nil {
+		panic("duplicated flush operation")
 	}
-	// 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 {
-			return err
+	b.done = make(chan struct{}) // allocate the channel for notification
+
+	go func() {
+		defer func() {
+			if postFlush != nil {
+				postFlush()
+			}
+			close(b.done)
+		}()
+
+		// Ensure the target state id is aligned with the internal counter.
+		head := rawdb.ReadPersistentStateID(db)
+		if head+b.layers != id {
+			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
 		}
+
+		// 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")
 	}
-	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 {
-		return err
-	}
-	commitBytesMeter.Mark(int64(size))
-	commitNodesMeter.Mark(int64(nodes))
-	commitAccountsMeter.Mark(int64(accounts))
-	commitStoragesMeter.Mark(int64(slots))
-	commitTimeTimer.UpdateSince(start)
-	b.reset()
-	log.Debug("Persisted buffer content", "nodes", nodes, "accounts", accounts, "slots", slots, "bytes", common.StorageSize(size), "elapsed", common.PrettyDuration(time.Since(start)))
-	return nil
+	<-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
 
+	// Block until the background flushing is finished. It must
+	// be done before terminating the potential background snapshot
+	// generator.
+	dl := db.tree.bottom()
+	if err := dl.waitFlush(); err != nil {
+		return err
+	}
 	// Terminate the background generation if it's active
-	disk := db.tree.bottom()
-	if disk.generator != nil {
-		disk.generator.stop()
+	if dl.generator != nil {
+		dl.generator.stop()
 	}
-	disk.resetCache() // release the memory held by clean cache
+	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{