github.com/sonm-io/go-ethereum@v1.8.18/trie/database.go (about)

     1  // Copyright 2018 The go-ethereum Authors
     2  // This file is part of the go-ethereum library.
     3  //
     4  // The go-ethereum library is free software: you can redistribute it and/or modify
     5  // it under the terms of the GNU Lesser General Public License as published by
     6  // the Free Software Foundation, either version 3 of the License, or
     7  // (at your option) any later version.
     8  //
     9  // The go-ethereum library is distributed in the hope that it will be useful,
    10  // but WITHOUT ANY WARRANTY; without even the implied warranty of
    11  // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
    12  // GNU Lesser General Public License for more details.
    13  //
    14  // You should have received a copy of the GNU Lesser General Public License
    15  // along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
    16  
    17  package trie
    18  
    19  import (
    20  	"fmt"
    21  	"io"
    22  	"sync"
    23  	"time"
    24  
    25  	"github.com/ethereum/go-ethereum/common"
    26  	"github.com/ethereum/go-ethereum/ethdb"
    27  	"github.com/ethereum/go-ethereum/log"
    28  	"github.com/ethereum/go-ethereum/metrics"
    29  	"github.com/ethereum/go-ethereum/rlp"
    30  )
    31  
    32  var (
    33  	memcacheFlushTimeTimer  = metrics.NewRegisteredResettingTimer("trie/memcache/flush/time", nil)
    34  	memcacheFlushNodesMeter = metrics.NewRegisteredMeter("trie/memcache/flush/nodes", nil)
    35  	memcacheFlushSizeMeter  = metrics.NewRegisteredMeter("trie/memcache/flush/size", nil)
    36  
    37  	memcacheGCTimeTimer  = metrics.NewRegisteredResettingTimer("trie/memcache/gc/time", nil)
    38  	memcacheGCNodesMeter = metrics.NewRegisteredMeter("trie/memcache/gc/nodes", nil)
    39  	memcacheGCSizeMeter  = metrics.NewRegisteredMeter("trie/memcache/gc/size", nil)
    40  
    41  	memcacheCommitTimeTimer  = metrics.NewRegisteredResettingTimer("trie/memcache/commit/time", nil)
    42  	memcacheCommitNodesMeter = metrics.NewRegisteredMeter("trie/memcache/commit/nodes", nil)
    43  	memcacheCommitSizeMeter  = metrics.NewRegisteredMeter("trie/memcache/commit/size", nil)
    44  )
    45  
    46  // secureKeyPrefix is the database key prefix used to store trie node preimages.
    47  var secureKeyPrefix = []byte("secure-key-")
    48  
    49  // secureKeyLength is the length of the above prefix + 32byte hash.
    50  const secureKeyLength = 11 + 32
    51  
    52  // DatabaseReader wraps the Get and Has method of a backing store for the trie.
    53  type DatabaseReader interface {
    54  	// Get retrieves the value associated with key from the database.
    55  	Get(key []byte) (value []byte, err error)
    56  
    57  	// Has retrieves whether a key is present in the database.
    58  	Has(key []byte) (bool, error)
    59  }
    60  
    61  // Database is an intermediate write layer between the trie data structures and
    62  // the disk database. The aim is to accumulate trie writes in-memory and only
    63  // periodically flush a couple tries to disk, garbage collecting the remainder.
    64  type Database struct {
    65  	diskdb ethdb.Database // Persistent storage for matured trie nodes
    66  
    67  	nodes  map[common.Hash]*cachedNode // Data and references relationships of a node
    68  	oldest common.Hash                 // Oldest tracked node, flush-list head
    69  	newest common.Hash                 // Newest tracked node, flush-list tail
    70  
    71  	preimages map[common.Hash][]byte // Preimages of nodes from the secure trie
    72  	seckeybuf [secureKeyLength]byte  // Ephemeral buffer for calculating preimage keys
    73  
    74  	gctime  time.Duration      // Time spent on garbage collection since last commit
    75  	gcnodes uint64             // Nodes garbage collected since last commit
    76  	gcsize  common.StorageSize // Data storage garbage collected since last commit
    77  
    78  	flushtime  time.Duration      // Time spent on data flushing since last commit
    79  	flushnodes uint64             // Nodes flushed since last commit
    80  	flushsize  common.StorageSize // Data storage flushed since last commit
    81  
    82  	nodesSize     common.StorageSize // Storage size of the nodes cache (exc. flushlist)
    83  	preimagesSize common.StorageSize // Storage size of the preimages cache
    84  
    85  	lock sync.RWMutex
    86  }
    87  
    88  // rawNode is a simple binary blob used to differentiate between collapsed trie
    89  // nodes and already encoded RLP binary blobs (while at the same time store them
    90  // in the same cache fields).
    91  type rawNode []byte
    92  
    93  func (n rawNode) canUnload(uint16, uint16) bool { panic("this should never end up in a live trie") }
    94  func (n rawNode) cache() (hashNode, bool)       { panic("this should never end up in a live trie") }
    95  func (n rawNode) fstring(ind string) string     { panic("this should never end up in a live trie") }
    96  
    97  // rawFullNode represents only the useful data content of a full node, with the
    98  // caches and flags stripped out to minimize its data storage. This type honors
    99  // the same RLP encoding as the original parent.
   100  type rawFullNode [17]node
   101  
   102  func (n rawFullNode) canUnload(uint16, uint16) bool { panic("this should never end up in a live trie") }
   103  func (n rawFullNode) cache() (hashNode, bool)       { panic("this should never end up in a live trie") }
   104  func (n rawFullNode) fstring(ind string) string     { panic("this should never end up in a live trie") }
   105  
   106  func (n rawFullNode) EncodeRLP(w io.Writer) error {
   107  	var nodes [17]node
   108  
   109  	for i, child := range n {
   110  		if child != nil {
   111  			nodes[i] = child
   112  		} else {
   113  			nodes[i] = nilValueNode
   114  		}
   115  	}
   116  	return rlp.Encode(w, nodes)
   117  }
   118  
   119  // rawShortNode represents only the useful data content of a short node, with the
   120  // caches and flags stripped out to minimize its data storage. This type honors
   121  // the same RLP encoding as the original parent.
   122  type rawShortNode struct {
   123  	Key []byte
   124  	Val node
   125  }
   126  
   127  func (n rawShortNode) canUnload(uint16, uint16) bool { panic("this should never end up in a live trie") }
   128  func (n rawShortNode) cache() (hashNode, bool)       { panic("this should never end up in a live trie") }
   129  func (n rawShortNode) fstring(ind string) string     { panic("this should never end up in a live trie") }
   130  
   131  // cachedNode is all the information we know about a single cached node in the
   132  // memory database write layer.
   133  type cachedNode struct {
   134  	node node   // Cached collapsed trie node, or raw rlp data
   135  	size uint16 // Byte size of the useful cached data
   136  
   137  	parents  uint16                 // Number of live nodes referencing this one
   138  	children map[common.Hash]uint16 // External children referenced by this node
   139  
   140  	flushPrev common.Hash // Previous node in the flush-list
   141  	flushNext common.Hash // Next node in the flush-list
   142  }
   143  
   144  // rlp returns the raw rlp encoded blob of the cached node, either directly from
   145  // the cache, or by regenerating it from the collapsed node.
   146  func (n *cachedNode) rlp() []byte {
   147  	if node, ok := n.node.(rawNode); ok {
   148  		return node
   149  	}
   150  	blob, err := rlp.EncodeToBytes(n.node)
   151  	if err != nil {
   152  		panic(err)
   153  	}
   154  	return blob
   155  }
   156  
   157  // obj returns the decoded and expanded trie node, either directly from the cache,
   158  // or by regenerating it from the rlp encoded blob.
   159  func (n *cachedNode) obj(hash common.Hash, cachegen uint16) node {
   160  	if node, ok := n.node.(rawNode); ok {
   161  		return mustDecodeNode(hash[:], node, cachegen)
   162  	}
   163  	return expandNode(hash[:], n.node, cachegen)
   164  }
   165  
   166  // childs returns all the tracked children of this node, both the implicit ones
   167  // from inside the node as well as the explicit ones from outside the node.
   168  func (n *cachedNode) childs() []common.Hash {
   169  	children := make([]common.Hash, 0, 16)
   170  	for child := range n.children {
   171  		children = append(children, child)
   172  	}
   173  	if _, ok := n.node.(rawNode); !ok {
   174  		gatherChildren(n.node, &children)
   175  	}
   176  	return children
   177  }
   178  
   179  // gatherChildren traverses the node hierarchy of a collapsed storage node and
   180  // retrieves all the hashnode children.
   181  func gatherChildren(n node, children *[]common.Hash) {
   182  	switch n := n.(type) {
   183  	case *rawShortNode:
   184  		gatherChildren(n.Val, children)
   185  
   186  	case rawFullNode:
   187  		for i := 0; i < 16; i++ {
   188  			gatherChildren(n[i], children)
   189  		}
   190  	case hashNode:
   191  		*children = append(*children, common.BytesToHash(n))
   192  
   193  	case valueNode, nil:
   194  
   195  	default:
   196  		panic(fmt.Sprintf("unknown node type: %T", n))
   197  	}
   198  }
   199  
   200  // simplifyNode traverses the hierarchy of an expanded memory node and discards
   201  // all the internal caches, returning a node that only contains the raw data.
   202  func simplifyNode(n node) node {
   203  	switch n := n.(type) {
   204  	case *shortNode:
   205  		// Short nodes discard the flags and cascade
   206  		return &rawShortNode{Key: n.Key, Val: simplifyNode(n.Val)}
   207  
   208  	case *fullNode:
   209  		// Full nodes discard the flags and cascade
   210  		node := rawFullNode(n.Children)
   211  		for i := 0; i < len(node); i++ {
   212  			if node[i] != nil {
   213  				node[i] = simplifyNode(node[i])
   214  			}
   215  		}
   216  		return node
   217  
   218  	case valueNode, hashNode, rawNode:
   219  		return n
   220  
   221  	default:
   222  		panic(fmt.Sprintf("unknown node type: %T", n))
   223  	}
   224  }
   225  
   226  // expandNode traverses the node hierarchy of a collapsed storage node and converts
   227  // all fields and keys into expanded memory form.
   228  func expandNode(hash hashNode, n node, cachegen uint16) node {
   229  	switch n := n.(type) {
   230  	case *rawShortNode:
   231  		// Short nodes need key and child expansion
   232  		return &shortNode{
   233  			Key: compactToHex(n.Key),
   234  			Val: expandNode(nil, n.Val, cachegen),
   235  			flags: nodeFlag{
   236  				hash: hash,
   237  				gen:  cachegen,
   238  			},
   239  		}
   240  
   241  	case rawFullNode:
   242  		// Full nodes need child expansion
   243  		node := &fullNode{
   244  			flags: nodeFlag{
   245  				hash: hash,
   246  				gen:  cachegen,
   247  			},
   248  		}
   249  		for i := 0; i < len(node.Children); i++ {
   250  			if n[i] != nil {
   251  				node.Children[i] = expandNode(nil, n[i], cachegen)
   252  			}
   253  		}
   254  		return node
   255  
   256  	case valueNode, hashNode:
   257  		return n
   258  
   259  	default:
   260  		panic(fmt.Sprintf("unknown node type: %T", n))
   261  	}
   262  }
   263  
   264  // NewDatabase creates a new trie database to store ephemeral trie content before
   265  // its written out to disk or garbage collected.
   266  func NewDatabase(diskdb ethdb.Database) *Database {
   267  	return &Database{
   268  		diskdb:    diskdb,
   269  		nodes:     map[common.Hash]*cachedNode{{}: {}},
   270  		preimages: make(map[common.Hash][]byte),
   271  	}
   272  }
   273  
   274  // DiskDB retrieves the persistent storage backing the trie database.
   275  func (db *Database) DiskDB() DatabaseReader {
   276  	return db.diskdb
   277  }
   278  
   279  // InsertBlob writes a new reference tracked blob to the memory database if it's
   280  // yet unknown. This method should only be used for non-trie nodes that require
   281  // reference counting, since trie nodes are garbage collected directly through
   282  // their embedded children.
   283  func (db *Database) InsertBlob(hash common.Hash, blob []byte) {
   284  	db.lock.Lock()
   285  	defer db.lock.Unlock()
   286  
   287  	db.insert(hash, blob, rawNode(blob))
   288  }
   289  
   290  // insert inserts a collapsed trie node into the memory database. This method is
   291  // a more generic version of InsertBlob, supporting both raw blob insertions as
   292  // well ex trie node insertions. The blob must always be specified to allow proper
   293  // size tracking.
   294  func (db *Database) insert(hash common.Hash, blob []byte, node node) {
   295  	// If the node's already cached, skip
   296  	if _, ok := db.nodes[hash]; ok {
   297  		return
   298  	}
   299  	// Create the cached entry for this node
   300  	entry := &cachedNode{
   301  		node:      simplifyNode(node),
   302  		size:      uint16(len(blob)),
   303  		flushPrev: db.newest,
   304  	}
   305  	for _, child := range entry.childs() {
   306  		if c := db.nodes[child]; c != nil {
   307  			c.parents++
   308  		}
   309  	}
   310  	db.nodes[hash] = entry
   311  
   312  	// Update the flush-list endpoints
   313  	if db.oldest == (common.Hash{}) {
   314  		db.oldest, db.newest = hash, hash
   315  	} else {
   316  		db.nodes[db.newest].flushNext, db.newest = hash, hash
   317  	}
   318  	db.nodesSize += common.StorageSize(common.HashLength + entry.size)
   319  }
   320  
   321  // insertPreimage writes a new trie node pre-image to the memory database if it's
   322  // yet unknown. The method will make a copy of the slice.
   323  //
   324  // Note, this method assumes that the database's lock is held!
   325  func (db *Database) insertPreimage(hash common.Hash, preimage []byte) {
   326  	if _, ok := db.preimages[hash]; ok {
   327  		return
   328  	}
   329  	db.preimages[hash] = common.CopyBytes(preimage)
   330  	db.preimagesSize += common.StorageSize(common.HashLength + len(preimage))
   331  }
   332  
   333  // node retrieves a cached trie node from memory, or returns nil if none can be
   334  // found in the memory cache.
   335  func (db *Database) node(hash common.Hash, cachegen uint16) node {
   336  	// Retrieve the node from cache if available
   337  	db.lock.RLock()
   338  	node := db.nodes[hash]
   339  	db.lock.RUnlock()
   340  
   341  	if node != nil {
   342  		return node.obj(hash, cachegen)
   343  	}
   344  	// Content unavailable in memory, attempt to retrieve from disk
   345  	enc, err := db.diskdb.Get(hash[:])
   346  	if err != nil || enc == nil {
   347  		return nil
   348  	}
   349  	return mustDecodeNode(hash[:], enc, cachegen)
   350  }
   351  
   352  // Node retrieves an encoded cached trie node from memory. If it cannot be found
   353  // cached, the method queries the persistent database for the content.
   354  func (db *Database) Node(hash common.Hash) ([]byte, error) {
   355  	// Retrieve the node from cache if available
   356  	db.lock.RLock()
   357  	node := db.nodes[hash]
   358  	db.lock.RUnlock()
   359  
   360  	if node != nil {
   361  		return node.rlp(), nil
   362  	}
   363  	// Content unavailable in memory, attempt to retrieve from disk
   364  	return db.diskdb.Get(hash[:])
   365  }
   366  
   367  // preimage retrieves a cached trie node pre-image from memory. If it cannot be
   368  // found cached, the method queries the persistent database for the content.
   369  func (db *Database) preimage(hash common.Hash) ([]byte, error) {
   370  	// Retrieve the node from cache if available
   371  	db.lock.RLock()
   372  	preimage := db.preimages[hash]
   373  	db.lock.RUnlock()
   374  
   375  	if preimage != nil {
   376  		return preimage, nil
   377  	}
   378  	// Content unavailable in memory, attempt to retrieve from disk
   379  	return db.diskdb.Get(db.secureKey(hash[:]))
   380  }
   381  
   382  // secureKey returns the database key for the preimage of key, as an ephemeral
   383  // buffer. The caller must not hold onto the return value because it will become
   384  // invalid on the next call.
   385  func (db *Database) secureKey(key []byte) []byte {
   386  	buf := append(db.seckeybuf[:0], secureKeyPrefix...)
   387  	buf = append(buf, key...)
   388  	return buf
   389  }
   390  
   391  // Nodes retrieves the hashes of all the nodes cached within the memory database.
   392  // This method is extremely expensive and should only be used to validate internal
   393  // states in test code.
   394  func (db *Database) Nodes() []common.Hash {
   395  	db.lock.RLock()
   396  	defer db.lock.RUnlock()
   397  
   398  	var hashes = make([]common.Hash, 0, len(db.nodes))
   399  	for hash := range db.nodes {
   400  		if hash != (common.Hash{}) { // Special case for "root" references/nodes
   401  			hashes = append(hashes, hash)
   402  		}
   403  	}
   404  	return hashes
   405  }
   406  
   407  // Reference adds a new reference from a parent node to a child node.
   408  func (db *Database) Reference(child common.Hash, parent common.Hash) {
   409  	db.lock.RLock()
   410  	defer db.lock.RUnlock()
   411  
   412  	db.reference(child, parent)
   413  }
   414  
   415  // reference is the private locked version of Reference.
   416  func (db *Database) reference(child common.Hash, parent common.Hash) {
   417  	// If the node does not exist, it's a node pulled from disk, skip
   418  	node, ok := db.nodes[child]
   419  	if !ok {
   420  		return
   421  	}
   422  	// If the reference already exists, only duplicate for roots
   423  	if db.nodes[parent].children == nil {
   424  		db.nodes[parent].children = make(map[common.Hash]uint16)
   425  	} else if _, ok = db.nodes[parent].children[child]; ok && parent != (common.Hash{}) {
   426  		return
   427  	}
   428  	node.parents++
   429  	db.nodes[parent].children[child]++
   430  }
   431  
   432  // Dereference removes an existing reference from a root node.
   433  func (db *Database) Dereference(root common.Hash) {
   434  	// Sanity check to ensure that the meta-root is not removed
   435  	if root == (common.Hash{}) {
   436  		log.Error("Attempted to dereference the trie cache meta root")
   437  		return
   438  	}
   439  	db.lock.Lock()
   440  	defer db.lock.Unlock()
   441  
   442  	nodes, storage, start := len(db.nodes), db.nodesSize, time.Now()
   443  	db.dereference(root, common.Hash{})
   444  
   445  	db.gcnodes += uint64(nodes - len(db.nodes))
   446  	db.gcsize += storage - db.nodesSize
   447  	db.gctime += time.Since(start)
   448  
   449  	memcacheGCTimeTimer.Update(time.Since(start))
   450  	memcacheGCSizeMeter.Mark(int64(storage - db.nodesSize))
   451  	memcacheGCNodesMeter.Mark(int64(nodes - len(db.nodes)))
   452  
   453  	log.Debug("Dereferenced trie from memory database", "nodes", nodes-len(db.nodes), "size", storage-db.nodesSize, "time", time.Since(start),
   454  		"gcnodes", db.gcnodes, "gcsize", db.gcsize, "gctime", db.gctime, "livenodes", len(db.nodes), "livesize", db.nodesSize)
   455  }
   456  
   457  // dereference is the private locked version of Dereference.
   458  func (db *Database) dereference(child common.Hash, parent common.Hash) {
   459  	// Dereference the parent-child
   460  	node := db.nodes[parent]
   461  
   462  	if node.children != nil && node.children[child] > 0 {
   463  		node.children[child]--
   464  		if node.children[child] == 0 {
   465  			delete(node.children, child)
   466  		}
   467  	}
   468  	// If the child does not exist, it's a previously committed node.
   469  	node, ok := db.nodes[child]
   470  	if !ok {
   471  		return
   472  	}
   473  	// If there are no more references to the child, delete it and cascade
   474  	if node.parents > 0 {
   475  		// This is a special cornercase where a node loaded from disk (i.e. not in the
   476  		// memcache any more) gets reinjected as a new node (short node split into full,
   477  		// then reverted into short), causing a cached node to have no parents. That is
   478  		// no problem in itself, but don't make maxint parents out of it.
   479  		node.parents--
   480  	}
   481  	if node.parents == 0 {
   482  		// Remove the node from the flush-list
   483  		switch child {
   484  		case db.oldest:
   485  			db.oldest = node.flushNext
   486  			db.nodes[node.flushNext].flushPrev = common.Hash{}
   487  		case db.newest:
   488  			db.newest = node.flushPrev
   489  			db.nodes[node.flushPrev].flushNext = common.Hash{}
   490  		default:
   491  			db.nodes[node.flushPrev].flushNext = node.flushNext
   492  			db.nodes[node.flushNext].flushPrev = node.flushPrev
   493  		}
   494  		// Dereference all children and delete the node
   495  		for _, hash := range node.childs() {
   496  			db.dereference(hash, child)
   497  		}
   498  		delete(db.nodes, child)
   499  		db.nodesSize -= common.StorageSize(common.HashLength + int(node.size))
   500  	}
   501  }
   502  
   503  // Cap iteratively flushes old but still referenced trie nodes until the total
   504  // memory usage goes below the given threshold.
   505  func (db *Database) Cap(limit common.StorageSize) error {
   506  	// Create a database batch to flush persistent data out. It is important that
   507  	// outside code doesn't see an inconsistent state (referenced data removed from
   508  	// memory cache during commit but not yet in persistent storage). This is ensured
   509  	// by only uncaching existing data when the database write finalizes.
   510  	db.lock.RLock()
   511  
   512  	nodes, storage, start := len(db.nodes), db.nodesSize, time.Now()
   513  	batch := db.diskdb.NewBatch()
   514  
   515  	// db.nodesSize only contains the useful data in the cache, but when reporting
   516  	// the total memory consumption, the maintenance metadata is also needed to be
   517  	// counted. For every useful node, we track 2 extra hashes as the flushlist.
   518  	size := db.nodesSize + common.StorageSize((len(db.nodes)-1)*2*common.HashLength)
   519  
   520  	// If the preimage cache got large enough, push to disk. If it's still small
   521  	// leave for later to deduplicate writes.
   522  	flushPreimages := db.preimagesSize > 4*1024*1024
   523  	if flushPreimages {
   524  		for hash, preimage := range db.preimages {
   525  			if err := batch.Put(db.secureKey(hash[:]), preimage); err != nil {
   526  				log.Error("Failed to commit preimage from trie database", "err", err)
   527  				db.lock.RUnlock()
   528  				return err
   529  			}
   530  			if batch.ValueSize() > ethdb.IdealBatchSize {
   531  				if err := batch.Write(); err != nil {
   532  					db.lock.RUnlock()
   533  					return err
   534  				}
   535  				batch.Reset()
   536  			}
   537  		}
   538  	}
   539  	// Keep committing nodes from the flush-list until we're below allowance
   540  	oldest := db.oldest
   541  	for size > limit && oldest != (common.Hash{}) {
   542  		// Fetch the oldest referenced node and push into the batch
   543  		node := db.nodes[oldest]
   544  		if err := batch.Put(oldest[:], node.rlp()); err != nil {
   545  			db.lock.RUnlock()
   546  			return err
   547  		}
   548  		// If we exceeded the ideal batch size, commit and reset
   549  		if batch.ValueSize() >= ethdb.IdealBatchSize {
   550  			if err := batch.Write(); err != nil {
   551  				log.Error("Failed to write flush list to disk", "err", err)
   552  				db.lock.RUnlock()
   553  				return err
   554  			}
   555  			batch.Reset()
   556  		}
   557  		// Iterate to the next flush item, or abort if the size cap was achieved. Size
   558  		// is the total size, including both the useful cached data (hash -> blob), as
   559  		// well as the flushlist metadata (2*hash). When flushing items from the cache,
   560  		// we need to reduce both.
   561  		size -= common.StorageSize(3*common.HashLength + int(node.size))
   562  		oldest = node.flushNext
   563  	}
   564  	// Flush out any remainder data from the last batch
   565  	if err := batch.Write(); err != nil {
   566  		log.Error("Failed to write flush list to disk", "err", err)
   567  		db.lock.RUnlock()
   568  		return err
   569  	}
   570  	db.lock.RUnlock()
   571  
   572  	// Write successful, clear out the flushed data
   573  	db.lock.Lock()
   574  	defer db.lock.Unlock()
   575  
   576  	if flushPreimages {
   577  		db.preimages = make(map[common.Hash][]byte)
   578  		db.preimagesSize = 0
   579  	}
   580  	for db.oldest != oldest {
   581  		node := db.nodes[db.oldest]
   582  		delete(db.nodes, db.oldest)
   583  		db.oldest = node.flushNext
   584  
   585  		db.nodesSize -= common.StorageSize(common.HashLength + int(node.size))
   586  	}
   587  	if db.oldest != (common.Hash{}) {
   588  		db.nodes[db.oldest].flushPrev = common.Hash{}
   589  	}
   590  	db.flushnodes += uint64(nodes - len(db.nodes))
   591  	db.flushsize += storage - db.nodesSize
   592  	db.flushtime += time.Since(start)
   593  
   594  	memcacheFlushTimeTimer.Update(time.Since(start))
   595  	memcacheFlushSizeMeter.Mark(int64(storage - db.nodesSize))
   596  	memcacheFlushNodesMeter.Mark(int64(nodes - len(db.nodes)))
   597  
   598  	log.Debug("Persisted nodes from memory database", "nodes", nodes-len(db.nodes), "size", storage-db.nodesSize, "time", time.Since(start),
   599  		"flushnodes", db.flushnodes, "flushsize", db.flushsize, "flushtime", db.flushtime, "livenodes", len(db.nodes), "livesize", db.nodesSize)
   600  
   601  	return nil
   602  }
   603  
   604  // Commit iterates over all the children of a particular node, writes them out
   605  // to disk, forcefully tearing down all references in both directions.
   606  //
   607  // As a side effect, all pre-images accumulated up to this point are also written.
   608  func (db *Database) Commit(node common.Hash, report bool) error {
   609  	// Create a database batch to flush persistent data out. It is important that
   610  	// outside code doesn't see an inconsistent state (referenced data removed from
   611  	// memory cache during commit but not yet in persistent storage). This is ensured
   612  	// by only uncaching existing data when the database write finalizes.
   613  	db.lock.RLock()
   614  
   615  	start := time.Now()
   616  	batch := db.diskdb.NewBatch()
   617  
   618  	// Move all of the accumulated preimages into a write batch
   619  	for hash, preimage := range db.preimages {
   620  		if err := batch.Put(db.secureKey(hash[:]), preimage); err != nil {
   621  			log.Error("Failed to commit preimage from trie database", "err", err)
   622  			db.lock.RUnlock()
   623  			return err
   624  		}
   625  		if batch.ValueSize() > ethdb.IdealBatchSize {
   626  			if err := batch.Write(); err != nil {
   627  				return err
   628  			}
   629  			batch.Reset()
   630  		}
   631  	}
   632  	// Move the trie itself into the batch, flushing if enough data is accumulated
   633  	nodes, storage := len(db.nodes), db.nodesSize
   634  	if err := db.commit(node, batch); err != nil {
   635  		log.Error("Failed to commit trie from trie database", "err", err)
   636  		db.lock.RUnlock()
   637  		return err
   638  	}
   639  	// Write batch ready, unlock for readers during persistence
   640  	if err := batch.Write(); err != nil {
   641  		log.Error("Failed to write trie to disk", "err", err)
   642  		db.lock.RUnlock()
   643  		return err
   644  	}
   645  	db.lock.RUnlock()
   646  
   647  	// Write successful, clear out the flushed data
   648  	db.lock.Lock()
   649  	defer db.lock.Unlock()
   650  
   651  	db.preimages = make(map[common.Hash][]byte)
   652  	db.preimagesSize = 0
   653  
   654  	db.uncache(node)
   655  
   656  	memcacheCommitTimeTimer.Update(time.Since(start))
   657  	memcacheCommitSizeMeter.Mark(int64(storage - db.nodesSize))
   658  	memcacheCommitNodesMeter.Mark(int64(nodes - len(db.nodes)))
   659  
   660  	logger := log.Info
   661  	if !report {
   662  		logger = log.Debug
   663  	}
   664  	logger("Persisted trie from memory database", "nodes", nodes-len(db.nodes)+int(db.flushnodes), "size", storage-db.nodesSize+db.flushsize, "time", time.Since(start)+db.flushtime,
   665  		"gcnodes", db.gcnodes, "gcsize", db.gcsize, "gctime", db.gctime, "livenodes", len(db.nodes), "livesize", db.nodesSize)
   666  
   667  	// Reset the garbage collection statistics
   668  	db.gcnodes, db.gcsize, db.gctime = 0, 0, 0
   669  	db.flushnodes, db.flushsize, db.flushtime = 0, 0, 0
   670  
   671  	return nil
   672  }
   673  
   674  // commit is the private locked version of Commit.
   675  func (db *Database) commit(hash common.Hash, batch ethdb.Batch) error {
   676  	// If the node does not exist, it's a previously committed node
   677  	node, ok := db.nodes[hash]
   678  	if !ok {
   679  		return nil
   680  	}
   681  	for _, child := range node.childs() {
   682  		if err := db.commit(child, batch); err != nil {
   683  			return err
   684  		}
   685  	}
   686  	if err := batch.Put(hash[:], node.rlp()); err != nil {
   687  		return err
   688  	}
   689  	// If we've reached an optimal batch size, commit and start over
   690  	if batch.ValueSize() >= ethdb.IdealBatchSize {
   691  		if err := batch.Write(); err != nil {
   692  			return err
   693  		}
   694  		batch.Reset()
   695  	}
   696  	return nil
   697  }
   698  
   699  // uncache is the post-processing step of a commit operation where the already
   700  // persisted trie is removed from the cache. The reason behind the two-phase
   701  // commit is to ensure consistent data availability while moving from memory
   702  // to disk.
   703  func (db *Database) uncache(hash common.Hash) {
   704  	// If the node does not exist, we're done on this path
   705  	node, ok := db.nodes[hash]
   706  	if !ok {
   707  		return
   708  	}
   709  	// Node still exists, remove it from the flush-list
   710  	switch hash {
   711  	case db.oldest:
   712  		db.oldest = node.flushNext
   713  		db.nodes[node.flushNext].flushPrev = common.Hash{}
   714  	case db.newest:
   715  		db.newest = node.flushPrev
   716  		db.nodes[node.flushPrev].flushNext = common.Hash{}
   717  	default:
   718  		db.nodes[node.flushPrev].flushNext = node.flushNext
   719  		db.nodes[node.flushNext].flushPrev = node.flushPrev
   720  	}
   721  	// Uncache the node's subtries and remove the node itself too
   722  	for _, child := range node.childs() {
   723  		db.uncache(child)
   724  	}
   725  	delete(db.nodes, hash)
   726  	db.nodesSize -= common.StorageSize(common.HashLength + int(node.size))
   727  }
   728  
   729  // Size returns the current storage size of the memory cache in front of the
   730  // persistent database layer.
   731  func (db *Database) Size() (common.StorageSize, common.StorageSize) {
   732  	db.lock.RLock()
   733  	defer db.lock.RUnlock()
   734  
   735  	// db.nodesSize only contains the useful data in the cache, but when reporting
   736  	// the total memory consumption, the maintenance metadata is also needed to be
   737  	// counted. For every useful node, we track 2 extra hashes as the flushlist.
   738  	var flushlistSize = common.StorageSize((len(db.nodes) - 1) * 2 * common.HashLength)
   739  	return db.nodesSize + flushlistSize, db.preimagesSize
   740  }
   741  
   742  // verifyIntegrity is a debug method to iterate over the entire trie stored in
   743  // memory and check whether every node is reachable from the meta root. The goal
   744  // is to find any errors that might cause memory leaks and or trie nodes to go
   745  // missing.
   746  //
   747  // This method is extremely CPU and memory intensive, only use when must.
   748  func (db *Database) verifyIntegrity() {
   749  	// Iterate over all the cached nodes and accumulate them into a set
   750  	reachable := map[common.Hash]struct{}{{}: {}}
   751  
   752  	for child := range db.nodes[common.Hash{}].children {
   753  		db.accumulate(child, reachable)
   754  	}
   755  	// Find any unreachable but cached nodes
   756  	unreachable := []string{}
   757  	for hash, node := range db.nodes {
   758  		if _, ok := reachable[hash]; !ok {
   759  			unreachable = append(unreachable, fmt.Sprintf("%x: {Node: %v, Parents: %d, Prev: %x, Next: %x}",
   760  				hash, node.node, node.parents, node.flushPrev, node.flushNext))
   761  		}
   762  	}
   763  	if len(unreachable) != 0 {
   764  		panic(fmt.Sprintf("trie cache memory leak: %v", unreachable))
   765  	}
   766  }
   767  
   768  // accumulate iterates over the trie defined by hash and accumulates all the
   769  // cached children found in memory.
   770  func (db *Database) accumulate(hash common.Hash, reachable map[common.Hash]struct{}) {
   771  	// Mark the node reachable if present in the memory cache
   772  	node, ok := db.nodes[hash]
   773  	if !ok {
   774  		return
   775  	}
   776  	reachable[hash] = struct{}{}
   777  
   778  	// Iterate over all the children and accumulate them too
   779  	for _, child := range node.childs() {
   780  		db.accumulate(child, reachable)
   781  	}
   782  }