eth: make traceChain avoid OOM on long-running tracing (#23736)

This PR changes long-running chain tracing, so that it at some points releases the memory trie db, and switch over to a fresh disk-backed trie.

eth/api_backend.go

@@ -352,8 +352,8 @@ func (b *EthAPIBackend) StartMining(threads int) error {
 	return b.eth.StartMining(threads)
 }
 
-func (b *EthAPIBackend) StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive bool) (*state.StateDB, error) {
-	return b.eth.stateAtBlock(block, reexec, base, checkLive)
+func (b *EthAPIBackend) StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive, preferDisk bool) (*state.StateDB, error) {
+	return b.eth.stateAtBlock(block, reexec, base, checkLive, preferDisk)
 }
 
 func (b *EthAPIBackend) StateAtTransaction(ctx context.Context, block *types.Block, txIndex int, reexec uint64) (core.Message, vm.BlockContext, *state.StateDB, error) {

eth/state_accessor.go

@@ -35,7 +35,17 @@ import (
 // are attempted to be reexecuted to generate the desired state. The optional
 // base layer statedb can be passed then it's regarded as the statedb of the
 // parent block.
-func (eth *Ethereum) stateAtBlock(block *types.Block, reexec uint64, base *state.StateDB, checkLive bool) (statedb *state.StateDB, err error) {
+// Parameters:
+// - block: The block for which we want the state (== state at the stateRoot of the parent)
+// - reexec: The maximum number of blocks to reprocess trying to obtain the desired state
+// - base: If the caller is tracing multiple blocks, the caller can provide the parent state
+//         continuously from the callsite.
+// - checklive: if true, then the live 'blockchain' state database is used. If the caller want to
+//        perform Commit or other 'save-to-disk' changes, this should be set to false to avoid
+//        storing trash persistently
+// - preferDisk: this arg can be used by the caller to signal that even though the 'base' is provided,
+//        it would be preferrable to start from a fresh state, if we have it on disk.
+func (eth *Ethereum) stateAtBlock(block *types.Block, reexec uint64, base *state.StateDB, checkLive bool, preferDisk bool) (statedb *state.StateDB, err error) {
 	var (
 		current  *types.Block
 		database state.Database
@@ -50,6 +60,15 @@ func (eth *Ethereum) stateAtBlock(block *types.Block, reexec uint64, base *state
 		}
 	}
 	if base != nil {
+		if preferDisk {
+			// Create an ephemeral trie.Database for isolating the live one. Otherwise
+			// the internal junks created by tracing will be persisted into the disk.
+			database = state.NewDatabaseWithConfig(eth.chainDb, &trie.Config{Cache: 16})
+			if statedb, err = state.New(block.Root(), database, nil); err == nil {
+				log.Info("Found disk backend for state trie", "root", block.Root(), "number", block.Number())
+				return statedb, nil
+			}
+		}
 		// The optional base statedb is given, mark the start point as parent block
 		statedb, database, report = base, base.Database(), false
 		current = eth.blockchain.GetBlock(block.ParentHash(), block.NumberU64()-1)
@@ -152,7 +171,7 @@ func (eth *Ethereum) stateAtTransaction(block *types.Block, txIndex int, reexec
 	}
 	// Lookup the statedb of parent block from the live database,
 	// otherwise regenerate it on the flight.
-	statedb, err := eth.stateAtBlock(parent, reexec, nil, true)
+	statedb, err := eth.stateAtBlock(parent, reexec, nil, true, false)
 	if err != nil {
 		return nil, vm.BlockContext{}, nil, err
 	}

eth/tracers/api.go

@@ -53,6 +53,13 @@ const (
 	// and reexecute to produce missing historical state necessary to run a specific
 	// trace.
 	defaultTraceReexec = uint64(128)
+
+	// defaultTracechainMemLimit is the size of the triedb, at which traceChain
+	// switches over and tries to use a disk-backed database instead of building
+	// on top of memory.
+	// For non-archive nodes, this limit _will_ be overblown, as disk-backed tries
+	// will only be found every ~15K blocks or so.
+	defaultTracechainMemLimit = common.StorageSize(500 * 1024 * 1024)
 )
 
 // Backend interface provides the common API services (that are provided by
@@ -67,7 +74,10 @@ type Backend interface {
 	ChainConfig() *params.ChainConfig
 	Engine() consensus.Engine
 	ChainDb() ethdb.Database
-	StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive bool) (*state.StateDB, error)
+	// StateAtBlock returns the state corresponding to the stateroot of the block.
+	// N.B: For executing transactions on block N, the required stateRoot is block N-1,
+	// so this method should be called with the parent.
+	StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive, preferDisk bool) (*state.StateDB, error)
 	StateAtTransaction(ctx context.Context, block *types.Block, txIndex int, reexec uint64) (core.Message, vm.BlockContext, *state.StateDB, error)
 }
 
@@ -320,6 +330,7 @@ func (api *API) traceChain(ctx context.Context, start, end *types.Block, config
 			}
 			close(results)
 		}()
+		var preferDisk bool
 		// Feed all the blocks both into the tracer, as well as fast process concurrently
 		for number = start.NumberU64(); number < end.NumberU64(); number++ {
 			// Stop tracing if interruption was requested
@@ -349,18 +360,24 @@ func (api *API) traceChain(ctx context.Context, start, end *types.Block, config
 			}
 			// Prepare the statedb for tracing. Don't use the live database for
 			// tracing to avoid persisting state junks into the database.
-			statedb, err = api.backend.StateAtBlock(localctx, block, reexec, statedb, false)
+			statedb, err = api.backend.StateAtBlock(localctx, block, reexec, statedb, false, preferDisk)
 			if err != nil {
 				failed = err
 				break
 			}
-			if statedb.Database().TrieDB() != nil {
+			if trieDb := statedb.Database().TrieDB(); trieDb != nil {
 				// Hold the reference for tracer, will be released at the final stage
-				statedb.Database().TrieDB().Reference(block.Root(), common.Hash{})
+				trieDb.Reference(block.Root(), common.Hash{})
 
 				// Release the parent state because it's already held by the tracer
 				if parent != (common.Hash{}) {
-					statedb.Database().TrieDB().Dereference(parent)
+					trieDb.Dereference(parent)
+				}
+				// Prefer disk if the trie db memory grows too much
+				s1, s2 := trieDb.Size()
+				if !preferDisk && (s1+s2) > defaultTracechainMemLimit {
+					log.Info("Switching to prefer-disk mode for tracing", "size", s1+s2)
+					preferDisk = true
 				}
 			}
 			parent = block.Root()
@@ -496,7 +513,7 @@ func (api *API) IntermediateRoots(ctx context.Context, hash common.Hash, config
 	if config != nil && config.Reexec != nil {
 		reexec = *config.Reexec
 	}
-	statedb, err := api.backend.StateAtBlock(ctx, parent, reexec, nil, true)
+	statedb, err := api.backend.StateAtBlock(ctx, parent, reexec, nil, true, false)
 	if err != nil {
 		return nil, err
 	}
@@ -557,7 +574,7 @@ func (api *API) traceBlock(ctx context.Context, block *types.Block, config *Trac
 	if config != nil && config.Reexec != nil {
 		reexec = *config.Reexec
 	}
-	statedb, err := api.backend.StateAtBlock(ctx, parent, reexec, nil, true)
+	statedb, err := api.backend.StateAtBlock(ctx, parent, reexec, nil, true, false)
 	if err != nil {
 		return nil, err
 	}
@@ -646,7 +663,7 @@ func (api *API) standardTraceBlockToFile(ctx context.Context, block *types.Block
 	if config != nil && config.Reexec != nil {
 		reexec = *config.Reexec
 	}
-	statedb, err := api.backend.StateAtBlock(ctx, parent, reexec, nil, true)
+	statedb, err := api.backend.StateAtBlock(ctx, parent, reexec, nil, true, false)
 	if err != nil {
 		return nil, err
 	}
@@ -810,7 +827,7 @@ func (api *API) TraceCall(ctx context.Context, args ethapi.TransactionArgs, bloc
 	if config != nil && config.Reexec != nil {
 		reexec = *config.Reexec
 	}
-	statedb, err := api.backend.StateAtBlock(ctx, block, reexec, nil, true)
+	statedb, err := api.backend.StateAtBlock(ctx, block, reexec, nil, true, false)
 	if err != nil {
 		return nil, err
 	}

eth/tracers/api_test.go

@@ -138,7 +138,7 @@ func (b *testBackend) ChainDb() ethdb.Database {
 	return b.chaindb
 }
 
-func (b *testBackend) StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive bool) (*state.StateDB, error) {
+func (b *testBackend) StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive bool, preferDisk bool) (*state.StateDB, error) {
 	statedb, err := b.chain.StateAt(block.Root())
 	if err != nil {
 		return nil, errStateNotFound

eth/tracers/tracer.go

@@ -553,17 +553,10 @@ func New(code string, ctx *Context) (*Tracer, error) {
 	tracer.vm.Pop()
 	hasExit := tracer.vm.GetPropString(tracer.tracerObject, "exit")
 	tracer.vm.Pop()
-
 	if hasEnter != hasExit {
 		return nil, fmt.Errorf("trace object must expose either both or none of enter() and exit()")
 	}
-	if !hasStep {
-		// If there's no step function, the enter and exit must be present
-		if !hasEnter {
-			return nil, fmt.Errorf("trace object must expose either step() or both enter() and exit()")
-		}
-	}
-	tracer.traceCallFrames = hasEnter
+	tracer.traceCallFrames = hasEnter && hasExit
 	tracer.traceSteps = hasStep
 
 	// Tracer is valid, inject the big int library to access large numbers

les/api_backend.go

@@ -324,7 +324,7 @@ func (b *LesApiBackend) CurrentHeader() *types.Header {
 	return b.eth.blockchain.CurrentHeader()
 }
 
-func (b *LesApiBackend) StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive bool) (*state.StateDB, error) {
+func (b *LesApiBackend) StateAtBlock(ctx context.Context, block *types.Block, reexec uint64, base *state.StateDB, checkLive bool, preferDisk bool) (*state.StateDB, error) {
 	return b.eth.stateAtBlock(ctx, block, reexec)
 }