core, ethdb, trie: mode dirty data to clean cache on flush (#19307)

This PR is a more advanced form of the dirty-to-clean cacher (#18995),
where we reuse previous database write batches as datasets to uncache,
saving a dirty-trie-iteration and a dirty-trie-rlp-reencoding per block.

1 files changed, 72 insertions, 48 deletions

diff --git a/trie/database.go b/trie/database.go
index 6df1a7f79..3bbcb6ade 100644
--- a/trie/database.go
+++ b/trie/database.go
@@ -59,6 +59,11 @@ const secureKeyLength = 11 + 32
 // Database is an intermediate write layer between the trie data structures and
 // the disk database. The aim is to accumulate trie writes in-memory and only
 // periodically flush a couple tries to disk, garbage collecting the remainder.
+//
+// Note, the trie Database is **not** thread safe in its mutations, but it **is**
+// thread safe in providing individual, independent node access. The rationale
+// behind this split design is to provide read access to RPC handlers and sync
+// servers even while the trie is executing expensive garbage collection.
 type Database struct {
 	diskdb ethdb.KeyValueStore // Persistent storage for matured trie nodes
 
@@ -465,8 +470,8 @@ func (db *Database) Nodes() []common.Hash {
 
 // Reference adds a new reference from a parent node to a child node.
 func (db *Database) Reference(child common.Hash, parent common.Hash) {
-	db.lock.RLock()
-	defer db.lock.RUnlock()
+	db.lock.Lock()
+	defer db.lock.Unlock()
 
 	db.reference(child, parent)
 }
@@ -561,13 +566,14 @@ func (db *Database) dereference(child common.Hash, parent common.Hash) {
 
 // Cap iteratively flushes old but still referenced trie nodes until the total
 // memory usage goes below the given threshold.
+//
+// Note, this method is a non-synchronized mutator. It is unsafe to call this
+// concurrently with other mutators.
 func (db *Database) Cap(limit common.StorageSize) error {
 	// Create a database batch to flush persistent data out. It is important that
 	// outside code doesn't see an inconsistent state (referenced data removed from
 	// memory cache during commit but not yet in persistent storage). This is ensured
 	// by only uncaching existing data when the database write finalizes.
-	db.lock.RLock()
-
 	nodes, storage, start := len(db.dirties), db.dirtiesSize, time.Now()
 	batch := db.diskdb.NewBatch()
 
@@ -583,12 +589,10 @@ func (db *Database) Cap(limit common.StorageSize) error {
 		for hash, preimage := range db.preimages {
 			if err := batch.Put(db.secureKey(hash[:]), preimage); err != nil {
 				log.Error("Failed to commit preimage from trie database", "err", err)
-				db.lock.RUnlock()
 				return err
 			}
 			if batch.ValueSize() > ethdb.IdealBatchSize {
 				if err := batch.Write(); err != nil {
-					db.lock.RUnlock()
 					return err
 				}
 				batch.Reset()
@@ -601,14 +605,12 @@ func (db *Database) Cap(limit common.StorageSize) error {
 		// Fetch the oldest referenced node and push into the batch
 		node := db.dirties[oldest]
 		if err := batch.Put(oldest[:], node.rlp()); err != nil {
-			db.lock.RUnlock()
 			return err
 		}
 		// If we exceeded the ideal batch size, commit and reset
 		if batch.ValueSize() >= ethdb.IdealBatchSize {
 			if err := batch.Write(); err != nil {
 				log.Error("Failed to write flush list to disk", "err", err)
-				db.lock.RUnlock()
 				return err
 			}
 			batch.Reset()
@@ -623,11 +625,8 @@ func (db *Database) Cap(limit common.StorageSize) error {
 	// Flush out any remainder data from the last batch
 	if err := batch.Write(); err != nil {
 		log.Error("Failed to write flush list to disk", "err", err)
-		db.lock.RUnlock()
 		return err
 	}
-	db.lock.RUnlock()
-
 	// Write successful, clear out the flushed data
 	db.lock.Lock()
 	defer db.lock.Unlock()
@@ -661,16 +660,16 @@ func (db *Database) Cap(limit common.StorageSize) error {
 }
 
 // Commit iterates over all the children of a particular node, writes them out
-// to disk, forcefully tearing down all references in both directions.
+// to disk, forcefully tearing down all references in both directions. As a side
+// effect, all pre-images accumulated up to this point are also written.
 //
-// As a side effect, all pre-images accumulated up to this point are also written.
+// Note, this method is a non-synchronized mutator. It is unsafe to call this
+// concurrently with other mutators.
 func (db *Database) Commit(node common.Hash, report bool) error {
 	// Create a database batch to flush persistent data out. It is important that
 	// outside code doesn't see an inconsistent state (referenced data removed from
 	// memory cache during commit but not yet in persistent storage). This is ensured
 	// by only uncaching existing data when the database write finalizes.
-	db.lock.RLock()
-
 	start := time.Now()
 	batch := db.diskdb.NewBatch()
 
@@ -678,41 +677,47 @@ func (db *Database) Commit(node common.Hash, report bool) error {
 	for hash, preimage := range db.preimages {
 		if err := batch.Put(db.secureKey(hash[:]), preimage); err != nil {
 			log.Error("Failed to commit preimage from trie database", "err", err)
-			db.lock.RUnlock()
 			return err
 		}
+		// If the batch is too large, flush to disk
 		if batch.ValueSize() > ethdb.IdealBatchSize {
 			if err := batch.Write(); err != nil {
-				db.lock.RUnlock()
 				return err
 			}
 			batch.Reset()
 		}
 	}
+	// Since we're going to replay trie node writes into the clean cache, flush out
+	// any batched pre-images before continuing.
+	if err := batch.Write(); err != nil {
+		return err
+	}
+	batch.Reset()
+
 	// Move the trie itself into the batch, flushing if enough data is accumulated
 	nodes, storage := len(db.dirties), db.dirtiesSize
-	if err := db.commit(node, batch); err != nil {
+
+	uncacher := &cleaner{db}
+	if err := db.commit(node, batch, uncacher); err != nil {
 		log.Error("Failed to commit trie from trie database", "err", err)
-		db.lock.RUnlock()
 		return err
 	}
-	// Write batch ready, unlock for readers during persistence
+	// Trie mostly committed to disk, flush any batch leftovers
 	if err := batch.Write(); err != nil {
 		log.Error("Failed to write trie to disk", "err", err)
-		db.lock.RUnlock()
 		return err
 	}
-	db.lock.RUnlock()
-
-	// Write successful, clear out the flushed data
+	// Uncache any leftovers in the last batch
 	db.lock.Lock()
 	defer db.lock.Unlock()
 
+	batch.Replay(uncacher)
+	batch.Reset()
+
+	// Reset the storage counters and bumpd metrics
 	db.preimages = make(map[common.Hash][]byte)
 	db.preimagesSize = 0
 
-	db.uncache(node)
-
 	memcacheCommitTimeTimer.Update(time.Since(start))
 	memcacheCommitSizeMeter.Mark(int64(storage - db.dirtiesSize))
 	memcacheCommitNodesMeter.Mark(int64(nodes - len(db.dirties)))
@@ -732,14 +737,14 @@ func (db *Database) Commit(node common.Hash, report bool) error {
 }
 
 // commit is the private locked version of Commit.
-func (db *Database) commit(hash common.Hash, batch ethdb.Batch) error {
+func (db *Database) commit(hash common.Hash, batch ethdb.Batch, uncacher *cleaner) error {
 	// If the node does not exist, it's a previously committed node
 	node, ok := db.dirties[hash]
 	if !ok {
 		return nil
 	}
 	for _, child := range node.childs() {
-		if err := db.commit(child, batch); err != nil {
+		if err := db.commit(child, batch, uncacher); err != nil {
 			return err
 		}
 	}
@@ -751,39 +756,58 @@ func (db *Database) commit(hash common.Hash, batch ethdb.Batch) error {
 		if err := batch.Write(); err != nil {
 			return err
 		}
+		db.lock.Lock()
+		batch.Replay(uncacher)
 		batch.Reset()
+		db.lock.Unlock()
 	}
 	return nil
 }
 
-// uncache is the post-processing step of a commit operation where the already
-// persisted trie is removed from the cache. The reason behind the two-phase
-// commit is to ensure consistent data availability while moving from memory
-// to disk.
-func (db *Database) uncache(hash common.Hash) {
+// cleaner is a database batch replayer that takes a batch of write operations
+// and cleans up the trie database from anything written to disk.
+type cleaner struct {
+	db *Database
+}
+
+// Put reacts to database writes and implements dirty data uncaching. This is the
+// post-processing step of a commit operation where the already persisted trie is
+// removed from the dirty cache and moved into the clean cache. The reason behind
+// the two-phase commit is to ensure ensure data availability while moving from
+// memory to disk.
+func (c *cleaner) Put(key []byte, rlp []byte) error {
+	hash := common.BytesToHash(key)
+
 	// If the node does not exist, we're done on this path
-	node, ok := db.dirties[hash]
+	node, ok := c.db.dirties[hash]
 	if !ok {
-		return
+		return nil
 	}
 	// Node still exists, remove it from the flush-list
 	switch hash {
-	case db.oldest:
-		db.oldest = node.flushNext
-		db.dirties[node.flushNext].flushPrev = common.Hash{}
-	case db.newest:
-		db.newest = node.flushPrev
-		db.dirties[node.flushPrev].flushNext = common.Hash{}
+	case c.db.oldest:
+		c.db.oldest = node.flushNext
+		c.db.dirties[node.flushNext].flushPrev = common.Hash{}
+	case c.db.newest:
+		c.db.newest = node.flushPrev
+		c.db.dirties[node.flushPrev].flushNext = common.Hash{}
 	default:
-		db.dirties[node.flushPrev].flushNext = node.flushNext
-		db.dirties[node.flushNext].flushPrev = node.flushPrev
+		c.db.dirties[node.flushPrev].flushNext = node.flushNext
+		c.db.dirties[node.flushNext].flushPrev = node.flushPrev
 	}
-	// Uncache the node's subtries and remove the node itself too
-	for _, child := range node.childs() {
-		db.uncache(child)
+	// Remove the node from the dirty cache
+	delete(c.db.dirties, hash)
+	c.db.dirtiesSize -= common.StorageSize(common.HashLength + int(node.size))
+
+	// Move the flushed node into the clean cache to prevent insta-reloads
+	if c.db.cleans != nil {
+		c.db.cleans.Set(string(hash[:]), rlp)
 	}
-	delete(db.dirties, hash)
-	db.dirtiesSize -= common.StorageSize(common.HashLength + int(node.size))
+	return nil
+}
+
+func (c *cleaner) Delete(key []byte) error {
+	panic("Not implemented")
 }
 
 // Size returns the current storage size of the memory cache in front of the