core/state,triedb/pathdb: wire bintrie leaves through stateUpdate

Drains the binaryHasher's LeafProducer side-channel in StateDB.commit and
threads the stem writes through stateUpdate.encodeBinary into the pathdb
state set as per-offset accountData entries (key = stem||offset, value =
32-byte leaf or nil for clears).

The flat-state codec gains a Flush method that owns the in-memory→disk
write path, replacing the codec-agnostic per-entry loop in writeStates.
The merkle codec preserves its historical per-entry behavior verbatim;
the bintrie codec aggregates per-offset writes by stem so each stem hits
disk via a single read-modify-write, satisfying the codec's pre-aggregation
requirement and updating the clean cache with the merged blob it just
produced (no extra disk read).

stateUpdate.encodeBinary returns empty origin maps for the bintrie path:
state-history rollback for bintrie is deferred to a follow-up PR (see
BINTRIE_FLAT_STATE_REORG_GAP.md), and the diskLayer.revert path will
panic before consuming origins anyway.

core/state/database_hasher_binary_test.go

@@ -382,3 +382,84 @@ func TestMerkleHasherNoLeafProducer(t *testing.T) {
 		t.Fatal("merkleHasher should NOT implement LeafProducer")
 	}
 }
+
+// TestStateUpdateEncodeBinaryFromLeaves verifies that stateUpdate.encodeBinary
+// turns a slice of StemWrite values into the per-offset accountData map that
+// pathdb's bintrie codec consumes. Three things matter:
+//
+//  1. Every leaf becomes one accountData entry, keyed by stem||offset.
+//  2. nil-value leaves (account/storage deletes) become nil entries.
+//  3. Non-nil leaves are deeply copied — encodeBinary must not retain
+//     pointers into the hasher's internal slab.
+//
+// storages/storageOrigin/accountOrigin remain empty: the bintrie path uses
+// only accountData (per the layered-read design) and does not yet support
+// state-history rollback.
+func TestStateUpdateEncodeBinaryFromLeaves(t *testing.T) {
+	// Build a small leaves slice covering each kind of write the binary
+	// hasher emits: account update (BasicData + CodeHash), storage write,
+	// and a delete (nil value).
+	var (
+		stemA [bintrie.StemSize]byte
+		stemB [bintrie.StemSize]byte
+	)
+	for i := range stemA {
+		stemA[i] = byte(0x10 + i)
+		stemB[i] = byte(0xA0 + i)
+	}
+	basicDataValue := bytes.Repeat([]byte{0xAA}, 32)
+	codeHashValue := bytes.Repeat([]byte{0xBB}, 32)
+	storageValue := bytes.Repeat([]byte{0xCC}, 32)
+
+	leaves := []StemWrite{
+		// Account update at stemA: BasicData + CodeHash.
+		{Stem: stemA, Offset: bintrie.BasicDataLeafKey, Value: basicDataValue},
+		{Stem: stemA, Offset: bintrie.CodeHashLeafKey, Value: codeHashValue},
+		// Storage write at stemB.
+		{Stem: stemB, Offset: 7, Value: storageValue},
+		// Account delete at a third stem (nil values clear offsets 0+1).
+		{Stem: [bintrie.StemSize]byte{0xFF, 0xFF}, Offset: bintrie.BasicDataLeafKey, Value: nil},
+		{Stem: [bintrie.StemSize]byte{0xFF, 0xFF}, Offset: bintrie.CodeHashLeafKey, Value: nil},
+	}
+
+	su := &stateUpdate{leaves: leaves}
+	accounts, accountOrigin, storages, storageOrigin, err := su.encodeBinary()
+	if err != nil {
+		t.Fatalf("encodeBinary: %v", err)
+	}
+
+	if len(accounts) != len(leaves) {
+		t.Fatalf("accounts len = %d, want %d", len(accounts), len(leaves))
+	}
+	if len(storages) != 0 {
+		t.Errorf("storages should be empty for bintrie, got %d entries", len(storages))
+	}
+	if len(accountOrigin) != 0 || len(storageOrigin) != 0 {
+		t.Errorf("origin maps should be empty for bintrie")
+	}
+
+	// Check each leaf round-trips through the map under its full key.
+	for i, w := range leaves {
+		var fullKey common.Hash
+		copy(fullKey[:bintrie.StemSize], w.Stem[:])
+		fullKey[bintrie.StemSize] = w.Offset
+		got, ok := accounts[fullKey]
+		if !ok {
+			t.Errorf("leaf %d: missing key %x", i, fullKey)
+			continue
+		}
+		if w.Value == nil {
+			if got != nil {
+				t.Errorf("leaf %d: nil leaf became %x", i, got)
+			}
+			continue
+		}
+		if !bytes.Equal(got, w.Value) {
+			t.Errorf("leaf %d: got %x, want %x", i, got, w.Value)
+		}
+		// Aliasing check: the encoder must own its bytes.
+		if len(got) > 0 && &got[0] == &w.Value[0] {
+			t.Errorf("leaf %d: encodeBinary aliased the input slice", i)
+		}
+	}
+}

core/state/statedb.go

@@ -1010,7 +1010,16 @@ func (s *StateDB) commit(deleteEmptyObjects bool, noStorageWiping bool, blockNum
 			builder.CollectWitness(s.witness)
 		}
 	}
-	return newStateUpdate(noStorageWiping, origin, root, blockNumber, deletes, updates, nodes, secondaryHashes), nil
+	// If the hasher tracks flat-state leaf production (currently only the
+	// binary hasher), drain the buffered stem writes so the downstream
+	// state update can carry them into the pathdb flat-state layer. Merkle
+	// hashers do not implement this interface and the call short-circuits
+	// to nil — newStateUpdate accepts nil as "no leaves".
+	var leaves []StemWrite
+	if producer, ok := s.hasher.(LeafProducer); ok {
+		leaves = producer.DrainStemWrites()
+	}
+	return newStateUpdate(noStorageWiping, origin, root, blockNumber, deletes, updates, nodes, secondaryHashes, leaves), nil
 }
 
 // commitAndFlush is a wrapper of commit which also commits the state mutations

core/state/stateupdate.go

@@ -91,6 +91,14 @@ type stateUpdate struct {
 	codes           map[common.Address]*contractCode // codes contains mutated contract codes, keyed by address.
 	nodes           *trienode.MergedNodeSet          // nodes aggregates all dirty trie nodes produced by the update.
 	secondaryHashes map[common.Address]Hashes        // hashes of secondary tries
+
+	// leaves is the ordered list of stem-offset writes harvested from a
+	// LeafProducer-capable hasher (the binary hasher). For merkle hashers
+	// it is always nil; for the binary hasher it is the bintrie's view of
+	// the same state mutations the trie just absorbed, in flat-state form.
+	// encodeBinary turns this into the per-offset accountData map that
+	// pathdb's bintrie codec consumes at flush time.
+	leaves []StemWrite
 }
 
 // empty returns a flag indicating the state transition is empty or not.
@@ -104,7 +112,11 @@ func (sc *stateUpdate) empty() bool {
 //
 // rawStorageKey is a flag indicating whether to use the raw storage slot key or
 // the hash of the slot key for constructing state update object.
-func newStateUpdate(rawStorageKey bool, originRoot common.Hash, root common.Hash, blockNumber uint64, deletes map[common.Hash]*accountDelete, updates map[common.Hash]*accountUpdate, nodes *trienode.MergedNodeSet, secondaryHashes map[common.Address]Hashes) *stateUpdate {
+//
+// leaves carries the per-offset stem writes produced by a LeafProducer-capable
+// hasher (the binary hasher). It is nil for merkle hashers and consumed by
+// encodeBinary to populate the bintrie flat-state map.
+func newStateUpdate(rawStorageKey bool, originRoot common.Hash, root common.Hash, blockNumber uint64, deletes map[common.Hash]*accountDelete, updates map[common.Hash]*accountUpdate, nodes *trienode.MergedNodeSet, secondaryHashes map[common.Address]Hashes, leaves []StemWrite) *stateUpdate {
 	var (
 		accounts       = make(map[common.Hash]*Account)
 		accountsOrigin = make(map[common.Address]*Account)
@@ -182,6 +194,7 @@ func newStateUpdate(rawStorageKey bool, originRoot common.Hash, root common.Hash
 		codes:           codes,
 		nodes:           nodes,
 		secondaryHashes: secondaryHashes,
+		leaves:          leaves,
 	}
 }
 
@@ -250,49 +263,50 @@ func (sc *stateUpdate) encodeMerkle() (map[common.Hash][]byte, map[common.Addres
 	return accounts, accountOrigin, storages, storageOrigin, nil
 }
 
+// encodeBinary produces the bintrie flat-state representation consumed by
+// pathdb. Unlike encodeMerkle (which keys accounts/storage by keccak hashes
+// and slim-RLP encodes the values), the bintrie path uses one entry per
+// EIP-7864 leaf:
+//
+//	key   = stem(31B) || offset(1B), zero-padded into a common.Hash
+//	value = the 32-byte leaf payload, or nil to clear the offset
+//
+// Account header writes (BasicData at offset 0, CodeHash at offset 1) and
+// storage slot / code chunk writes are uniform — the binary hasher emits
+// each as a stemWrite via DrainStemWrites and we route every one of them
+// into the accounts map. The storages map stays empty: bintrie has no
+// per-account storage grouping at the flat-state layer, and pathdb's
+// disklayer/lookup tree both work fine with a single accountData map of
+// 32-byte keys.
+//
+// accountOrigin and storageOrigin are returned empty because state-history
+// rollback for bintrie is deferred to a follow-up PR (see
+// BINTRIE_FLAT_STATE_REORG_GAP.md). The pathdb layer's revertTo path
+// panics for bintrie before it would observe these maps anyway.
 func (sc *stateUpdate) encodeBinary() (map[common.Hash][]byte, map[common.Address][]byte, map[common.Hash]map[common.Hash][]byte, map[common.Address]map[common.Hash][]byte, error) {
 	var (
-		accounts      = make(map[common.Hash][]byte)
+		accounts      = make(map[common.Hash][]byte, len(sc.leaves))
 		storages      = make(map[common.Hash]map[common.Hash][]byte)
 		accountOrigin = make(map[common.Address][]byte)
 		storageOrigin = make(map[common.Address]map[common.Hash][]byte)
 	)
-	for addr, prev := range sc.accountsOrigin {
+	for _, w := range sc.leaves {
-		if prev == nil {
+		var fullKey common.Hash
-			accountOrigin[addr] = nil
+		copy(fullKey[:len(w.Stem)], w.Stem[:])
-		} else {
+		fullKey[len(w.Stem)] = w.Offset
-			accountOrigin[addr] = types.SlimAccountRLP(types.StateAccount{
+		// nil Value means "clear this offset" (account delete or storage
-				Balance:  prev.Balance,
+		// slot wipe). The pathdb codec interprets a nil entry as a delete
-				Nonce:    prev.Nonce,
+		// during flush, matching merkle's nil-blob convention.
-				CodeHash: prev.CodeHash,
+		if w.Value == nil {
-			})
+			accounts[fullKey] = nil
+			continue
 		}
-
+		// Take an owning copy: the hasher reuses its underlying buffers
-		addrHash := crypto.Keccak256Hash(addr.Bytes())
+		// across blocks, so retaining its slices would create cross-block
-		data := sc.accounts[addrHash]
+		// aliasing bugs in the pathdb diff layer.
-		if data == nil {
+		v := make([]byte, len(w.Value))
-			accounts[addrHash] = nil
+		copy(v, w.Value)
-		} else {
+		accounts[fullKey] = v
-			accounts[addrHash] = types.SlimAccountRLP(types.StateAccount{
-				Balance:  data.Balance,
-				Nonce:    data.Nonce,
-				CodeHash: data.CodeHash,
-			})
-		}
-	}
-	for addr, slots := range sc.storagesOrigin {
-		subset := make(map[common.Hash][]byte)
-		for key, val := range slots {
-			subset[key] = encodeSlot(val)
-		}
-		storageOrigin[addr] = subset
-	}
-	for addrHash, slots := range sc.storages {
-		subset := make(map[common.Hash][]byte)
-		for key, val := range slots {
-			subset[key] = encodeSlot(val)
-		}
-		storages[addrHash] = subset
 	}
 	return accounts, accountOrigin, storages, storageOrigin, nil
 }

triedb/pathdb/flat_codec.go

@@ -19,6 +19,7 @@ package pathdb
 import (
 	"bytes"
 
+	"github.com/VictoriaMetrics/fastcache"
 	"github.com/ethereum/go-ethereum/common"
 	"github.com/ethereum/go-ethereum/core/rawdb"
 	"github.com/ethereum/go-ethereum/crypto"
@@ -127,6 +128,23 @@ type flatStateCodec interface {
 	// marker. Returns the same semantics as bytes.Compare. Used by the
 	// disklayer.account/storage gating logic and by writeStates.
 	MarkerCompare(key []byte, marker []byte) int
+
+	// Flush drains all pending mutations from the in-memory accountData and
+	// storageData maps into the supplied batch and updates the clean cache
+	// in lockstep. The codec controls iteration order, key derivation, and
+	// any aggregation that may be required (e.g. the bintrie codec must
+	// merge per-offset writes into per-stem read-modify-writes to avoid
+	// quadratic disk reads).
+	//
+	// Entries strictly past genMarker (per the codec's MarkerCompare
+	// semantics) are skipped because they will be regenerated by the
+	// background snapshot generator.
+	//
+	// Returns (account-entry count, storage-entry count) for metric
+	// reporting; the merkle codec reports one per map entry, while the
+	// bintrie codec reports one per logical offset write (so the metrics
+	// remain comparable across schemes).
+	Flush(batch ethdb.Batch, genMarker []byte, accountData map[common.Hash][]byte, storageData map[common.Hash]map[common.Hash][]byte, clean *fastcache.Cache) (int, int)
 }
 
 // merkleFlatCodec implements flatStateCodec for the keccak-keyed MPT flat
@@ -214,3 +232,72 @@ func (c *merkleFlatCodec) SplitMarker(marker []byte) ([]byte, []byte) {
 func (c *merkleFlatCodec) MarkerCompare(key []byte, marker []byte) int {
 	return bytes.Compare(key, marker)
 }
+
+// Flush drains the supplied account/storage maps into the batch using the
+// historical merkle per-entry layout: one rawdb write per accountData entry
+// and one per storage slot. Entries past the genMarker are skipped (the
+// generator will fill them in). The clean cache is kept in sync with each
+// write so subsequent reads do not stale.
+//
+// This is the implementation that previously lived directly in writeStates.
+// It has been moved into the codec so the bintrie codec can supply its own
+// per-stem aggregating implementation alongside this one.
+func (c *merkleFlatCodec) Flush(batch ethdb.Batch, genMarker []byte, accountData map[common.Hash][]byte, storageData map[common.Hash]map[common.Hash][]byte, clean *fastcache.Cache) (int, int) {
+	var (
+		accounts int
+		slots    int
+	)
+	for addrHash, blob := range accountData {
+		// Skip any account not yet covered by the snapshot. The account
+		// at the generation marker position (addrHash == genMarker[:common.HashLength])
+		// should still be updated, as it would be skipped in the next
+		// generation cycle.
+		if genMarker != nil && bytes.Compare(addrHash[:], genMarker) > 0 {
+			continue
+		}
+		accounts++
+		cacheKey := c.AccountCacheKey(addrHash)
+		if len(blob) == 0 {
+			c.DeleteAccount(batch, addrHash)
+			if clean != nil {
+				clean.Set(cacheKey, nil)
+			}
+		} else {
+			c.WriteAccount(batch, addrHash, blob)
+			if clean != nil {
+				clean.Set(cacheKey, blob)
+			}
+		}
+	}
+	for addrHash, storages := range storageData {
+		// Skip any account not covered yet by the snapshot
+		if genMarker != nil && bytes.Compare(addrHash[:], genMarker) > 0 {
+			continue
+		}
+		midAccount := genMarker != nil && bytes.Equal(addrHash[:], genMarker[:common.HashLength])
+
+		for storageHash, blob := range storages {
+			// Skip any storage slot not yet covered by the snapshot. The storage slot
+			// at the generation marker position (addrHash == genMarker[:common.HashLength]
+			// and storageHash == genMarker[common.HashLength:]) should still be updated,
+			// as it would be skipped in the next generation cycle.
+			if midAccount && bytes.Compare(storageHash[:], genMarker[common.HashLength:]) > 0 {
+				continue
+			}
+			slots++
+			cacheKey := c.StorageCacheKey(addrHash, storageHash)
+			if len(blob) == 0 {
+				c.DeleteStorage(batch, addrHash, storageHash)
+				if clean != nil {
+					clean.Set(cacheKey, nil)
+				}
+			} else {
+				c.WriteStorage(batch, addrHash, storageHash, blob)
+				if clean != nil {
+					clean.Set(cacheKey, blob)
+				}
+			}
+		}
+	}
+	return accounts, slots
+}

triedb/pathdb/flat_codec_bintrie.go

@@ -20,6 +20,7 @@ import (
 	"bytes"
 	"fmt"
 
+	"github.com/VictoriaMetrics/fastcache"
 	"github.com/ethereum/go-ethereum/common"
 	"github.com/ethereum/go-ethereum/core/rawdb"
 	"github.com/ethereum/go-ethereum/ethdb"
@@ -236,22 +237,27 @@ func (c *bintrieFlatCodec) DeleteStorage(batch ethdb.Batch, _ common.Hash, stora
 // Write/Delete methods to ensure the policy (nil value clears, empty
 // blob deletes) is consistent.
 //
+// Returns the merged blob (or nil if the stem was deleted) so callers
+// such as Flush can repopulate the clean cache without an extra disk
+// read. The returned slice is freshly allocated and owned by the caller.
+//
 // Important: the read comes from c.db, NOT from the batch. A second
 // call for the same stem within a flush would re-read the pre-flush
 // state; see the pre-aggregation requirement documented on
 // bintrieFlatCodec.
-func (c *bintrieFlatCodec) applyWrites(batch ethdb.Batch, stem []byte, writes []stemOffsetValue) {
+func (c *bintrieFlatCodec) applyWrites(batch ethdb.Batch, stem []byte, writes []stemOffsetValue) []byte {
 	existing := rawdb.ReadBinTrieStem(c.db, stem)
 	merged, err := mergeStemBlob(existing, writes)
 	if err != nil {
 		crit("bintrie applyWrites: %v", err)
-		return
+		return nil
 	}
 	if merged == nil {
 		rawdb.DeleteBinTrieStem(batch, stem)
-		return
+		return nil
 	}
 	rawdb.WriteBinTrieStem(batch, stem, merged)
+	return merged
 }
 
 // splitAccountBlob validates and splits the two-slot account payload
@@ -375,6 +381,94 @@ func (c *bintrieFlatCodec) MarkerCompare(key []byte, marker []byte) int {
 	return bytes.Compare(key, marker)
 }
 
+// Flush drains the in-memory accountData and storageData maps into the
+// batch using the bintrie per-stem layout. The maps are expected to hold
+// per-offset entries — each key is a 32-byte (stem || offset) tuple
+// produced by AccountKey/StorageKey, and each value is a 32-byte leaf
+// (or nil to clear that offset).
+//
+// All entries are first grouped by stem, then a single
+// read-modify-write is issued per stem so the codec touches each stem
+// at most once during a flush. This is what allows the per-call
+// pre-aggregation requirement documented on bintrieFlatCodec to be
+// satisfied even when many writes target the same stem.
+//
+// storageData is also walked because higher-level callers may emit
+// storage entries that the codec routes through the storage map for
+// historical reasons; for the bintrie path, entries should normally
+// arrive on accountData but we accept either layout.
+//
+// Returns (offset count from accountData, offset count from storageData)
+// so the metric reporting in writeStates remains comparable to the
+// merkle path. The clean cache is updated with the merged stem blob
+// (one cache entry per stem, not per offset) — readers extract the
+// requested offset on hit.
+func (c *bintrieFlatCodec) Flush(batch ethdb.Batch, genMarker []byte, accountData map[common.Hash][]byte, storageData map[common.Hash]map[common.Hash][]byte, clean *fastcache.Cache) (int, int) {
+	// Aggregate per-offset writes into per-stem batches. We use [31]byte
+	// as the map key because bytes slices aren't hashable in Go and the
+	// stem itself is fixed size; the alternative (using common.Hash with
+	// a zero pad) would waste a byte per entry.
+	type aggregator struct {
+		writes []stemOffsetValue
+	}
+	aggregated := make(map[[bintrie.StemSize]byte]*aggregator)
+
+	addWrite := func(fullKey common.Hash, value []byte) {
+		var stem [bintrie.StemSize]byte
+		copy(stem[:], fullKey[:bintrie.StemSize])
+		offset := fullKey[bintrie.StemSize]
+		ag, exists := aggregated[stem]
+		if !exists {
+			ag = &aggregator{}
+			aggregated[stem] = ag
+		}
+		ag.writes = append(ag.writes, stemOffsetValue{Offset: offset, Value: value})
+	}
+
+	var (
+		accountWrites int
+		storageWrites int
+	)
+	for fullKey, value := range accountData {
+		// genMarker filtering: skip stems that the generator hasn't
+		// reached yet. We compare against the FULL key (stem || offset)
+		// because the bintrie marker is itself a 32-byte key.
+		if genMarker != nil && bytes.Compare(fullKey[:], genMarker) > 0 {
+			continue
+		}
+		accountWrites++
+		addWrite(fullKey, value)
+	}
+	for _, slots := range storageData {
+		for fullKey, value := range slots {
+			if genMarker != nil && bytes.Compare(fullKey[:], genMarker) > 0 {
+				continue
+			}
+			storageWrites++
+			addWrite(fullKey, value)
+		}
+	}
+	// Issue one RMW per stem and update the clean cache with the merged
+	// blob (or invalidate it if the stem was deleted).
+	for stem, ag := range aggregated {
+		merged := c.applyWrites(batch, stem[:], ag.writes)
+		if clean != nil {
+			// Reuse AccountCacheKey to derive the cache key — for
+			// bintrie this only depends on the stem so the trailing
+			// offset byte in the synthetic full key is irrelevant.
+			var fullKey common.Hash
+			copy(fullKey[:bintrie.StemSize], stem[:])
+			cacheKey := c.AccountCacheKey(fullKey)
+			if merged == nil {
+				clean.Set(cacheKey, nil)
+			} else {
+				clean.Set(cacheKey, merged)
+			}
+		}
+	}
+	return accountWrites, storageWrites
+}
+
 // crit is a shim around log.Crit that allows tests to replace the fatal
 // behavior with a panic if needed. Defined at the package level to match
 // the single-call-per-error style used by the merkle codec.

triedb/pathdb/flat_codec_bintrie_test.go

@@ -265,3 +265,126 @@ func TestBintrieCodecSplitMarker(t *testing.T) {
 		t.Fatalf("SplitMarker: acc=%x full=%x, want both %x", acc, full, marker)
 	}
 }
+
+// TestBintrieCodecFlushAggregates verifies the per-stem aggregation that
+// the codec's Flush method performs. Two distinct offsets at the SAME stem
+// should produce a single on-disk stem blob containing both offsets after
+// one Flush call — proving the codec collapses what would have been N
+// read-modify-writes into one.
+//
+// Three offsets are written across two stems (2 + 1) so we exercise both
+// the multi-offset and single-offset paths in a single test.
+func TestBintrieCodecFlushAggregates(t *testing.T) {
+	codec, db := newTestBintrieCodec(t)
+
+	// Build a per-offset accountData map mimicking what encodeBinary
+	// produces from a binaryHasher.DrainStemWrites: the keys are full
+	// 32-byte (stem || offset) tuples and the values are 32-byte leaves.
+	addr := common.HexToAddress("0xCafeBabeDeadBeef00112233445566778899aabb")
+	stem := bintrie.GetBinaryTreeKey(addr, make([]byte, 32))[:bintrie.StemSize]
+
+	basicData := bytes.Repeat([]byte{0xAA}, stemBlobValueSize)
+	codeHash := bytes.Repeat([]byte{0xBB}, stemBlobValueSize)
+	storageVal := bytes.Repeat([]byte{0xCC}, stemBlobValueSize)
+	otherStem := bytes.Repeat([]byte{0x42}, bintrie.StemSize)
+	otherVal := bytes.Repeat([]byte{0xDD}, stemBlobValueSize)
+
+	mkKey := func(stem []byte, offset byte) common.Hash {
+		var k common.Hash
+		copy(k[:bintrie.StemSize], stem)
+		k[bintrie.StemSize] = offset
+		return k
+	}
+	accountData := map[common.Hash][]byte{
+		mkKey(stem, bintrie.BasicDataLeafKey):       basicData,
+		mkKey(stem, bintrie.CodeHashLeafKey):        codeHash,
+		mkKey(stem, 64):                             storageVal, // header storage slot
+		mkKey(otherStem, bintrie.BasicDataLeafKey):  otherVal,
+	}
+
+	batch := db.NewBatch()
+	accW, stoW := codec.Flush(batch, nil, accountData, nil, nil)
+	flushBatch(t, batch)
+
+	if accW != 4 {
+		t.Errorf("account write count: got %d, want 4", accW)
+	}
+	if stoW != 0 {
+		t.Errorf("storage write count: got %d, want 0 (no storage map)", stoW)
+	}
+
+	// All three offsets at `stem` should be readable from a single on-disk
+	// blob; aggregation worked iff the second/third writes did not clobber
+	// the first.
+	blob := rawdb.ReadBinTrieStem(db, stem)
+	if len(blob) == 0 {
+		t.Fatal("stem blob missing after Flush")
+	}
+	for offset, want := range map[byte][]byte{
+		bintrie.BasicDataLeafKey: basicData,
+		bintrie.CodeHashLeafKey:  codeHash,
+		64:                       storageVal,
+	} {
+		got, err := extractStemOffset(blob, offset)
+		if err != nil {
+			t.Fatalf("extract offset %d: %v", offset, err)
+		}
+		if !bytes.Equal(got, want) {
+			t.Errorf("offset %d: got %x, want %x", offset, got, want)
+		}
+	}
+
+	// The other stem should also have its single offset.
+	otherBlob := rawdb.ReadBinTrieStem(db, otherStem)
+	if got, _ := extractStemOffset(otherBlob, bintrie.BasicDataLeafKey); !bytes.Equal(got, otherVal) {
+		t.Errorf("other stem BasicData: got %x, want %x", got, otherVal)
+	}
+}
+
+// TestBintrieCodecFlushDelete verifies that nil-valued entries in the
+// accountData map clear the corresponding offset, and that clearing every
+// populated offset at a stem removes the on-disk key entirely (matching
+// the per-call DeleteStorage semantics tested elsewhere).
+func TestBintrieCodecFlushDelete(t *testing.T) {
+	codec, db := newTestBintrieCodec(t)
+
+	// Seed: write two offsets at one stem.
+	stem := bytes.Repeat([]byte{0x77}, bintrie.StemSize)
+	v0 := bytes.Repeat([]byte{0x01}, stemBlobValueSize)
+	v1 := bytes.Repeat([]byte{0x02}, stemBlobValueSize)
+
+	mkKey := func(offset byte) common.Hash {
+		var k common.Hash
+		copy(k[:bintrie.StemSize], stem)
+		k[bintrie.StemSize] = offset
+		return k
+	}
+	batch := db.NewBatch()
+	codec.Flush(batch, nil, map[common.Hash][]byte{
+		mkKey(0): v0,
+		mkKey(1): v1,
+	}, nil, nil)
+	flushBatch(t, batch)
+
+	// Now flush a nil for offset 0 — only offset 1 should remain.
+	batch = db.NewBatch()
+	codec.Flush(batch, nil, map[common.Hash][]byte{mkKey(0): nil}, nil, nil)
+	flushBatch(t, batch)
+
+	blob := rawdb.ReadBinTrieStem(db, stem)
+	if got, _ := extractStemOffset(blob, 0); got != nil {
+		t.Errorf("offset 0 should be cleared, got %x", got)
+	}
+	if got, _ := extractStemOffset(blob, 1); !bytes.Equal(got, v1) {
+		t.Errorf("offset 1 should survive, got %x want %x", got, v1)
+	}
+
+	// Clear the last remaining offset; the on-disk key should disappear.
+	batch = db.NewBatch()
+	codec.Flush(batch, nil, map[common.Hash][]byte{mkKey(1): nil}, nil, nil)
+	flushBatch(t, batch)
+
+	if raw := rawdb.ReadBinTrieStem(db, stem); raw != nil {
+		t.Errorf("stem should be deleted, got %x", raw)
+	}
+}

triedb/pathdb/flush.go

@@ -17,8 +17,6 @@
 package pathdb
 
 import (
-	"bytes"
-
 	"github.com/VictoriaMetrics/fastcache"
 	"github.com/ethereum/go-ethereum/common"
 	"github.com/ethereum/go-ethereum/core/rawdb"
@@ -71,71 +69,16 @@ func writeNodes(batch ethdb.Batch, nodes map[common.Hash]map[string]*trienode.No
 // This function assumes the background generator is already terminated and states
 // before the supplied marker has been correctly generated.
 //
-// The codec parameter abstracts the trie-specific persistence and cache key
+// The codec parameter abstracts the trie-specific persistence: merkleFlatCodec
-// derivation. The marker comparisons retain merkle-specific shape (two-tier
+// performs a per-entry rawdb write for each accountData/storageData entry,
-// account+storage marker) because the bintrie path uses a separate writer
+// while bintrieFlatCodec aggregates per-offset writes into per-stem
-// (writeStems, added in a later commit) that operates on a single-tier
+// read-modify-writes. Either way, the genMarker filtering, cache update, and
-// marker over stems rather than (account, storage) pairs.
+// metric reporting all happen inside the codec — writeStates is just a thin
+// dispatcher.
 //
 // TODO(rjl493456442) do we really need this generation marker? The state updates
 // after the marker can also be written and will be fixed by generator later if
 // it's outdated.
 func writeStates(batch ethdb.Batch, codec flatStateCodec, genMarker []byte, accountData map[common.Hash][]byte, storageData map[common.Hash]map[common.Hash][]byte, clean *fastcache.Cache) (int, int) {
-	var (
+	return codec.Flush(batch, genMarker, accountData, storageData, clean)
-		accounts int
-		slots    int
-	)
-	for addrHash, blob := range accountData {
-		// Skip any account not yet covered by the snapshot. The account
-		// at the generation marker position (addrHash == genMarker[:common.HashLength])
-		// should still be updated, as it would be skipped in the next
-		// generation cycle.
-		if genMarker != nil && bytes.Compare(addrHash[:], genMarker) > 0 {
-			continue
-		}
-		accounts += 1
-		cacheKey := codec.AccountCacheKey(addrHash)
-		if len(blob) == 0 {
-			codec.DeleteAccount(batch, addrHash)
-			if clean != nil {
-				clean.Set(cacheKey, nil)
-			}
-		} else {
-			codec.WriteAccount(batch, addrHash, blob)
-			if clean != nil {
-				clean.Set(cacheKey, blob)
-			}
-		}
-	}
-	for addrHash, storages := range storageData {
-		// Skip any account not covered yet by the snapshot
-		if genMarker != nil && bytes.Compare(addrHash[:], genMarker) > 0 {
-			continue
-		}
-		midAccount := genMarker != nil && bytes.Equal(addrHash[:], genMarker[:common.HashLength])
-
-		for storageHash, blob := range storages {
-			// Skip any storage slot not yet covered by the snapshot. The storage slot
-			// at the generation marker position (addrHash == genMarker[:common.HashLength]
-			// and storageHash == genMarker[common.HashLength:]) should still be updated,
-			// as it would be skipped in the next generation cycle.
-			if midAccount && bytes.Compare(storageHash[:], genMarker[common.HashLength:]) > 0 {
-				continue
-			}
-			slots += 1
-			cacheKey := codec.StorageCacheKey(addrHash, storageHash)
-			if len(blob) == 0 {
-				codec.DeleteStorage(batch, addrHash, storageHash)
-				if clean != nil {
-					clean.Set(cacheKey, nil)
-				}
-			} else {
-				codec.WriteStorage(batch, addrHash, storageHash, blob)
-				if clean != nil {
-					clean.Set(cacheKey, blob)
-				}
-			}
-		}
-	}
-	return accounts, slots
 }