eth/fetcher: don't spend too much time on transaction inclusion (#25524)

* eth/fetcher: introduce some lag in tx fetching

* eth/fetcher: change conditions a bit

* eth/fetcher: use per-batch quota check

* eth/fetcher: fix some comments

* eth/fetcher: address review concerns

* eth/fetcher: fix panic + add warn log

* eth/fetcher: fix log

* eth/fetcher: fix log

* cmd/devp2p/internal/ethtest: fix ignorign tx announcements from prev. tests

* cmd/devp2p/internal/ethtest: fix TestLargeTxRequest

This increases the number of tx relay messages the test waits for. Since
go-ethereum now processes incoming txs in smaller batches, the
announcement messages it sends are also smaller.

Co-authored-by: Felix Lange <fjl@twurst.com>

cmd/devp2p/internal/ethtest/helpers.go

@@ -357,9 +357,13 @@ func (s *Suite) waitAnnounce(conn *Conn, blockAnnouncement *NewBlock) error {
 				return fmt.Errorf("wrong block hash in announcement: expected %v, got %v", blockAnnouncement.Block.Hash(), hashes[0].Hash)
 			}
 			return nil
-		case *NewPooledTransactionHashes:
+
 		// ignore tx announcements from previous tests
+		case *NewPooledTransactionHashes:
 			continue
+		case *Transactions:
+			continue
+
 		default:
 			return fmt.Errorf("unexpected: %s", pretty.Sdump(msg))
 		}

cmd/devp2p/internal/ethtest/suite.go

@@ -544,9 +544,13 @@ func (s *Suite) TestNewPooledTxs(t *utesting.T) {
 				t.Fatalf("unexpected number of txs requested: wanted %d, got %d", len(hashes), len(msg.GetPooledTransactionsPacket))
 			}
 			return
+
 		// ignore propagated txs from previous tests
 		case *NewPooledTransactionHashes:
 			continue
+		case *Transactions:
+			continue
+
 		// ignore block announcements from previous tests
 		case *NewBlockHashes:
 			continue

cmd/devp2p/internal/ethtest/transaction.go

@@ -192,10 +192,10 @@ func sendMultipleSuccessfulTxs(t *utesting.T, s *Suite, txs []*types.Transaction
 	nonce = txs[len(txs)-1].Nonce()
 
 	// Wait for the transaction announcement(s) and make sure all sent txs are being propagated.
-	// all txs should be announced within 3 announcements.
+	// all txs should be announced within a couple announcements.
 	recvHashes := make([]common.Hash, 0)
 
-	for i := 0; i < 3; i++ {
+	for i := 0; i < 20; i++ {
 		switch msg := recvConn.readAndServe(s.chain, timeout).(type) {
 		case *Transactions:
 			for _, tx := range *msg {

eth/fetcher/tx_fetcher.go

@@ -262,22 +262,40 @@ func (f *TxFetcher) Notify(peer string, hashes []common.Hash) error {
 // direct request replies. The differentiation is important so the fetcher can
 // re-schedule missing transactions as soon as possible.
 func (f *TxFetcher) Enqueue(peer string, txs []*types.Transaction, direct bool) error {
-	// Keep track of all the propagated transactions
-	if direct {
-		txReplyInMeter.Mark(int64(len(txs)))
-	} else {
-		txBroadcastInMeter.Mark(int64(len(txs)))
+	var (
+		inMeter          = txReplyInMeter
+		knownMeter       = txReplyKnownMeter
+		underpricedMeter = txReplyUnderpricedMeter
+		otherRejectMeter = txReplyOtherRejectMeter
+	)
+	if !direct {
+		inMeter = txBroadcastInMeter
+		knownMeter = txBroadcastKnownMeter
+		underpricedMeter = txBroadcastUnderpricedMeter
+		otherRejectMeter = txBroadcastOtherRejectMeter
 	}
+	// Keep track of all the propagated transactions
+	inMeter.Mark(int64(len(txs)))
+
 	// Push all the transactions into the pool, tracking underpriced ones to avoid
 	// re-requesting them and dropping the peer in case of malicious transfers.
 	var (
 		added = make([]common.Hash, 0, len(txs))
+		delay time.Duration
+	)
+	// proceed in batches
+	for i := 0; i < len(txs); i += 128 {
+		end := i + 128
+		if end > len(txs) {
+			end = len(txs)
+		}
+		var (
 			duplicate   int64
 			underpriced int64
 			otherreject int64
 		)
-	errs := f.addTxs(txs)
-	for i, err := range errs {
+		batch := txs[i:end]
+		for j, err := range f.addTxs(batch) {
 			// Track the transaction hash if the price is too low for us.
 			// Avoid re-request this transaction when we receive another
 			// announcement.
@@ -285,7 +303,7 @@ func (f *TxFetcher) Enqueue(peer string, txs []*types.Transaction, direct bool)
 				for f.underpriced.Cardinality() >= maxTxUnderpricedSetSize {
 					f.underpriced.Pop()
 				}
-			f.underpriced.Add(txs[i].Hash())
+				f.underpriced.Add(batch[j].Hash())
 			}
 			// Track a few interesting failure types
 			switch {
@@ -300,19 +318,23 @@ func (f *TxFetcher) Enqueue(peer string, txs []*types.Transaction, direct bool)
 			default:
 				otherreject++
 			}
-		added = append(added, txs[i].Hash())
+			added = append(added, batch[j].Hash())
+		}
+		knownMeter.Mark(duplicate)
+		underpricedMeter.Mark(underpriced)
+		otherRejectMeter.Mark(otherreject)
+
+		// If 'other reject' is >25% of the deliveries in any batch, abort. Either we are
+		// out of sync with the chain or the peer is griefing us.
+		if otherreject > 128/4 {
+			delay = 200 * time.Millisecond
+			log.Warn("Peer delivering useless transactions", "peer", peer, "ignored", len(txs)-end)
+			break
 		}
-	if direct {
-		txReplyKnownMeter.Mark(duplicate)
-		txReplyUnderpricedMeter.Mark(underpriced)
-		txReplyOtherRejectMeter.Mark(otherreject)
-	} else {
-		txBroadcastKnownMeter.Mark(duplicate)
-		txBroadcastUnderpricedMeter.Mark(underpriced)
-		txBroadcastOtherRejectMeter.Mark(otherreject)
 	}
 	select {
 	case f.cleanup <- &txDelivery{origin: peer, hashes: added, direct: direct}:
+		time.Sleep(delay)
 		return nil
 	case <-f.quit:
 		return errTerminated