trie/triedb/pathdb: improve dirty node flushing trigger

trie/triedb/pathdb/disklayer.go

@@ -172,37 +172,65 @@ func (dl *diskLayer) commit(bottom *diffLayer, force bool) (*diskLayer, error) {
 	dl.lock.Lock()
 	defer dl.lock.Unlock()
 
-	// Construct and store the state history first. If crash happens
+	// Construct and store the state history first. If crash happens after storing
-	// after storing the state history but without flushing the
+	// the state history but without flushing the corresponding states(journal),
-	// corresponding states(journal), the stored state history will
+	// the stored state history will be truncated from head in the next restart.
-	// be truncated in the next restart.
+	var (
+		overflow bool
+		oldest   uint64
+	)
 	if dl.db.freezer != nil {
-		err := writeHistory(dl.db.diskdb, dl.db.freezer, bottom, dl.db.config.StateHistory)
+		err := writeHistory(dl.db.freezer, bottom)
 		if err != nil {
 			return nil, err
 		}
+		// Determine if the persisted history object has exceeded the configured
+		// limitation, set the overflow as true if so.
+		tail, err := dl.db.freezer.Tail()
+		if err != nil {
+			return nil, err
+		}
+		limit := dl.db.config.StateHistory
+		if limit != 0 && bottom.stateID()-tail > limit {
+			overflow = true
+			oldest = bottom.stateID() - limit + 1 // track the id of history **after truncation**
+		}
 	}
 	// Mark the diskLayer as stale before applying any mutations on top.
 	dl.stale = true
 
-	// Store the root->id lookup afterwards. All stored lookups are
+	// Store the root->id lookup afterwards. All stored lookups are identified
-	// identified by the **unique** state root. It's impossible that
+	// by the **unique** state root. It's impossible that in the same chain
-	// in the same chain blocks are not adjacent but have the same
+	// blocks are not adjacent but have the same root.
-	// root.
 	if dl.id == 0 {
 		rawdb.WriteStateID(dl.db.diskdb, dl.root, 0)
 	}
 	rawdb.WriteStateID(dl.db.diskdb, bottom.rootHash(), bottom.stateID())
 
-	// Construct a new disk layer by merging the nodes from the provided
+	// Construct a new disk layer by merging the nodes from the provided diff
-	// diff layer, and flush the content in disk layer if there are too
+	// layer, and flush the content in disk layer if there are too many nodes
-	// many nodes cached. The clean cache is inherited from the original
+	// cached. The clean cache is inherited from the original disk layer.
-	// disk layer for reusing.
 	ndl := newDiskLayer(bottom.root, bottom.stateID(), dl.db, dl.cleans, dl.buffer.commit(bottom.nodes))
-	err := ndl.buffer.flush(ndl.db.diskdb, ndl.cleans, ndl.id, force)
+
-	if err != nil {
+	// In a unique scenario where the ID of the oldest history object (after tail
+	// truncation) surpasses the persisted state ID, we take the necessary action
+	// of forcibly committing the cached dirty nodes to ensure that the persisted
+	// state ID remains higher.
+	if !force && rawdb.ReadPersistentStateID(dl.db.diskdb) < oldest {
+		force = true
+	}
+	if err := ndl.buffer.flush(ndl.db.diskdb, ndl.cleans, ndl.id, force); err != nil {
 		return nil, err
 	}
+	// To remove outdated history objects from the end, we set the 'tail' parameter
+	// to 'oldest-1' due to the offset between the freezer index and the history ID.
+	if overflow {
+		pruned, err := truncateFromTail(ndl.db.diskdb, ndl.db.freezer, oldest-1)
+		if err != nil {
+			return nil, err
+		}
+		log.Debug("Prune state history", "number", pruned)
+	}
 	return ndl, nil
 }
 

trie/triedb/pathdb/history.go

@@ -512,38 +512,28 @@ func readHistory(freezer *rawdb.ResettableFreezer, id uint64) (*history, error)
 	return &dec, nil
 }
 
-// writeHistory writes the state history with provided state set. After
+// writeHistory persists the state history with the provided state set.
-// storing the corresponding state history, it will also prune the stale
+func writeHistory(freezer *rawdb.ResettableFreezer, dl *diffLayer) error {
-// histories from the disk with the given threshold.
-func writeHistory(db ethdb.KeyValueStore, freezer *rawdb.ResettableFreezer, dl *diffLayer, limit uint64) error {
 	// Short circuit if state set is not available.
 	if dl.states == nil {
 		return errors.New("state change set is not available")
 	}
 	var (
-		err   error
+		start   = time.Now()
-		n     int
+		history = newHistory(dl.rootHash(), dl.parentLayer().rootHash(), dl.block, dl.states)
-		start = time.Now()
-		h     = newHistory(dl.rootHash(), dl.parentLayer().rootHash(), dl.block, dl.states)
 	)
-	accountData, storageData, accountIndex, storageIndex := h.encode()
+	accountData, storageData, accountIndex, storageIndex := history.encode()
 	dataSize := common.StorageSize(len(accountData) + len(storageData))
 	indexSize := common.StorageSize(len(accountIndex) + len(storageIndex))
 
 	// Write history data into five freezer table respectively.
-	rawdb.WriteStateHistory(freezer, dl.stateID(), h.meta.encode(), accountIndex, storageIndex, accountData, storageData)
+	rawdb.WriteStateHistory(freezer, dl.stateID(), history.meta.encode(), accountIndex, storageIndex, accountData, storageData)
 
-	// Prune stale state histories based on the config.
-	if limit != 0 && dl.stateID() > limit {
-		n, err = truncateFromTail(db, freezer, dl.stateID()-limit)
-		if err != nil {
-			return err
-		}
-	}
 	historyDataBytesMeter.Mark(int64(dataSize))
 	historyIndexBytesMeter.Mark(int64(indexSize))
 	historyBuildTimeMeter.UpdateSince(start)
-	log.Debug("Stored state history", "id", dl.stateID(), "block", dl.block, "data", dataSize, "index", indexSize, "pruned", n, "elapsed", common.PrettyDuration(time.Since(start)))
+	log.Debug("Stored state history", "id", dl.stateID(), "block", dl.block, "data", dataSize, "index", indexSize, "elapsed", common.PrettyDuration(time.Since(start)))
+
 	return nil
 }