erigon-pulse/trie/database.go

// Copyright 2018 The go-ethereum Authors
// This file is part of the go-ethereum library.
//
// The go-ethereum library is free software: you can redistribute it and/or modify
// it under the terms of the GNU Lesser General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// The go-ethereum library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.

package trie

import (
	"encoding/binary"
	"fmt"
	"io"
	"reflect"
	"sync"
	"time"

	"github.com/allegro/bigcache"
	"github.com/ledgerwatch/turbo-geth/common"
	"github.com/ledgerwatch/turbo-geth/ethdb"
	"github.com/ledgerwatch/turbo-geth/log"
	"github.com/ledgerwatch/turbo-geth/metrics"
	"github.com/ledgerwatch/turbo-geth/rlp"
)

var (
	memcacheCleanHitMeter   = metrics.NewRegisteredMeter("trie/memcache/clean/hit", nil)
	memcacheCleanMissMeter  = metrics.NewRegisteredMeter("trie/memcache/clean/miss", nil)
	memcacheCleanReadMeter  = metrics.NewRegisteredMeter("trie/memcache/clean/read", nil)
	memcacheCleanWriteMeter = metrics.NewRegisteredMeter("trie/memcache/clean/write", nil)

	memcacheFlushTimeTimer  = metrics.NewRegisteredResettingTimer("trie/memcache/flush/time", nil)
	memcacheFlushNodesMeter = metrics.NewRegisteredMeter("trie/memcache/flush/nodes", nil)
	memcacheFlushSizeMeter  = metrics.NewRegisteredMeter("trie/memcache/flush/size", nil)

	memcacheGCTimeTimer  = metrics.NewRegisteredResettingTimer("trie/memcache/gc/time", nil)
	memcacheGCNodesMeter = metrics.NewRegisteredMeter("trie/memcache/gc/nodes", nil)
	memcacheGCSizeMeter  = metrics.NewRegisteredMeter("trie/memcache/gc/size", nil)

	memcacheCommitTimeTimer  = metrics.NewRegisteredResettingTimer("trie/memcache/commit/time", nil)
	memcacheCommitNodesMeter = metrics.NewRegisteredMeter("trie/memcache/commit/nodes", nil)
	memcacheCommitSizeMeter  = metrics.NewRegisteredMeter("trie/memcache/commit/size", nil)
)

// secureKeyPrefix is the database key prefix used to store trie node preimages.
var secureKeyPrefix = []byte("secure-key-")

// secureKeyLength is the length of the above prefix + 32byte hash.
const secureKeyLength = 11 + 32

// DatabaseReader wraps the Get method of a backing store for the trie.
type DatabaseReader interface {
	// Get retrieves the value associated with key from the database.
	Get(bucket, key []byte) (value []byte, err error)

	GetAsOf(bucket, hBucket, key []byte, timestamp uint64) ([]byte, error)

	// Has retrieves whether a key is present in the database.
	Has(bucket, key []byte) (bool, error)

	Walk(bucket, key []byte, keybits uint, walker func([]byte, []byte) (bool, error)) error
}

// 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.Database // Persistent storage for matured trie nodes

	cleans  *bigcache.BigCache          // GC friendly memory cache of clean node RLPs
	dirties map[common.Hash]*cachedNode // Data and references relationships of dirty nodes
	oldest  common.Hash                 // Oldest tracked node, flush-list head
	newest  common.Hash                 // Newest tracked node, flush-list tail

	preimages map[common.Hash][]byte // Preimages of nodes from the secure trie

	gctime  time.Duration      // Time spent on garbage collection since last commit
	gcnodes uint64             // Nodes garbage collected since last commit
	gcsize  common.StorageSize // Data storage garbage collected since last commit

	flushtime  time.Duration      // Time spent on data flushing since last commit
	flushnodes uint64             // Nodes flushed since last commit
	flushsize  common.StorageSize // Data storage flushed since last commit

	dirtiesSize   common.StorageSize // Storage size of the dirty node cache (exc. metadata)
	childrenSize  common.StorageSize // Storage size of the external children tracking
	preimagesSize common.StorageSize // Storage size of the preimages cache

	lock sync.RWMutex
}

// rawNode is a simple binary blob used to differentiate between collapsed trie
// nodes and already encoded RLP binary blobs (while at the same time store them
// in the same cache fields).
type rawNode []byte

func (n rawNode) canUnload(uint16, uint16) bool { panic("this should never end up in a live trie") }
func (n rawNode) cache() (hashNode, bool)       { panic("this should never end up in a live trie") }
func (n rawNode) fstring(ind string) string     { panic("this should never end up in a live trie") }

// rawFullNode represents only the useful data content of a full node, with the
// caches and flags stripped out to minimize its data storage. This type honors
// the same RLP encoding as the original parent.
type rawFullNode [17]node

func (n rawFullNode) canUnload(uint16, uint16) bool { panic("this should never end up in a live trie") }
func (n rawFullNode) cache() (hashNode, bool)       { panic("this should never end up in a live trie") }
func (n rawFullNode) fstring(ind string) string     { panic("this should never end up in a live trie") }

func (n rawFullNode) EncodeRLP(w io.Writer) error {
	var nodes [17]node

	for i, child := range n {
		if child != nil {
			nodes[i] = child
		} else {
			nodes[i] = nilValueNode
		}
	}
	return rlp.Encode(w, nodes)
}

// rawShortNode represents only the useful data content of a short node, with the
// caches and flags stripped out to minimize its data storage. This type honors
// the same RLP encoding as the original parent.
type rawShortNode struct {
	Key []byte
	Val node
}

func (n rawShortNode) canUnload(uint16, uint16) bool { panic("this should never end up in a live trie") }
func (n rawShortNode) cache() (hashNode, bool)       { panic("this should never end up in a live trie") }
func (n rawShortNode) fstring(ind string) string     { panic("this should never end up in a live trie") }

// cachedNode is all the information we know about a single cached node in the
// memory database write layer.
type cachedNode struct {
	node node   // Cached collapsed trie node, or raw rlp data
	size uint16 // Byte size of the useful cached data

	parents  uint32                 // Number of live nodes referencing this one
	children map[common.Hash]uint16 // External children referenced by this node

	flushPrev common.Hash // Previous node in the flush-list
	flushNext common.Hash // Next node in the flush-list
}

// cachedNodeSize is the raw size of a cachedNode data structure without any
// node data included. It's an approximate size, but should be a lot better
// than not counting them.
var cachedNodeSize = int(reflect.TypeOf(cachedNode{}).Size())

// cachedNodeChildrenSize is the raw size of an initialized but empty external
// reference map.
const cachedNodeChildrenSize = 48

// rlp returns the raw rlp encoded blob of the cached node, either directly from
// the cache, or by regenerating it from the collapsed node.
func (n *cachedNode) rlp() []byte {
	return nil
}

// obj returns the decoded and expanded trie node, either directly from the cache,
// or by regenerating it from the rlp encoded blob.
func (n *cachedNode) obj(hash common.Hash, cachegen uint16) node {
	// FIXME: whu is this not implemented?
	return nil
}

// childs returns all the tracked children of this node, both the implicit ones
// from inside the node as well as the explicit ones from outside the node.
func (n *cachedNode) childs() []common.Hash {
	return nil
}

// gatherChildren traverses the node hierarchy of a collapsed storage node and
// retrieves all the hashnode children.
func gatherChildren(n node, children *[]common.Hash) {
}

// simplifyNode traverses the hierarchy of an expanded memory node and discards
// all the internal caches, returning a node that only contains the raw data.
func simplifyNode(n node) node {
	return nil
}

// expandNode traverses the node hierarchy of a collapsed storage node and converts
// all fields and keys into expanded memory form.
func expandNode(hash hashNode, n node, cachegen uint16) node {
	// FIXME: Intentionally? does nothing in TurboGeth
	return nil
}

// trienodeHasher is a struct to be used with BigCache, which uses a Hasher to
// determine which shard to place an entry into. It's not a cryptographic hash,
// just to provide a bit of anti-collision (default is FNV64a).
//
// Since trie keys are already hashes, we can just use the key directly to
// map shard id.
type trienodeHasher struct{}

// Sum64 implements the bigcache.Hasher interface.
func (t trienodeHasher) Sum64(key string) uint64 {
	return binary.BigEndian.Uint64([]byte(key))
}

// NewDatabase creates a new trie database to store ephemeral trie content before
// its written out to disk or garbage collected. No read cache is created, so all
// data retrievals will hit the underlying disk database.
func NewDatabase(diskdb ethdb.Database) *Database {
	return NewDatabaseWithCache(diskdb, 0)
}

// NewDatabaseWithCache creates a new trie database to store ephemeral trie content
// before its written out to disk or garbage collected. It also acts as a read cache
// for nodes loaded from disk.
func NewDatabaseWithCache(diskdb ethdb.Database, cache int) *Database {
	var cleans *bigcache.BigCache
	if cache > 0 {
		cleans, _ = bigcache.NewBigCache(bigcache.Config{
			Shards:             1024,
			LifeWindow:         time.Hour,
			MaxEntriesInWindow: cache * 1024,
			MaxEntrySize:       512,
			HardMaxCacheSize:   cache,
			Hasher:             trienodeHasher{},
		})
	}
	return &Database{
		diskdb: diskdb,
		cleans: cleans,
		dirties: map[common.Hash]*cachedNode{{}: {
			children: make(map[common.Hash]uint16),
		}},
		preimages: make(map[common.Hash][]byte),
	}
}

// DiskDB retrieves the persistent storage backing the trie database.
// FIXME: maybe a more restrictive interface?
func (db *Database) DiskDB() ethdb.Database {
	return db.diskdb
}

// InsertBlob writes a new reference tracked blob to the memory database if it's
// yet unknown. This method should only be used for non-trie nodes that require
// reference counting, since trie nodes are garbage collected directly through
// their embedded children.
func (db *Database) InsertBlob(hash common.Hash, blob []byte) {
}

// insert inserts a collapsed trie node into the memory database. This method is
// a more generic version of InsertBlob, supporting both raw blob insertions as
// well ex trie node insertions. The blob must always be specified to allow proper
// size tracking.
func (db *Database) insert(hash common.Hash, blob []byte, node node) {
	// If the node's already cached, skip
	if _, ok := db.dirties[hash]; ok {
		return
	}
	// Create the cached entry for this node
	entry := &cachedNode{
		node:      simplifyNode(node),
		size:      uint16(len(blob)),
		flushPrev: db.newest,
	}
	for _, child := range entry.childs() {
		if c := db.dirties[child]; c != nil {
			c.parents++
		}
	}
	db.dirties[hash] = entry

	// Update the flush-list endpoints
	if db.oldest == (common.Hash{}) {
		db.oldest, db.newest = hash, hash
	} else {
		db.dirties[db.newest].flushNext, db.newest = hash, hash
	}
	db.dirtiesSize += common.StorageSize(common.HashLength + entry.size)
}

// insertPreimage writes a new trie node pre-image to the memory database if it's
// yet unknown. The method will make a copy of the slice.
//
// Note, this method assumes that the database's lock is held!
func (db *Database) insertPreimage(hash common.Hash, preimage []byte) {
	if _, ok := db.preimages[hash]; ok {
		return
	}
	db.preimages[hash] = common.CopyBytes(preimage)
	db.preimagesSize += common.StorageSize(common.HashLength + len(preimage))
}

// node retrieves a cached trie node from memory, or returns nil if none can be
// found in the memory cache.
func (db *Database) node(hash common.Hash, cachegen uint16) node {
	// FIXME: Intentionally? does nothing in TurboGeth
	return nil
}

// Node retrieves an encoded cached trie node from memory. If it cannot be found
// cached, the method queries the persistent database for the content.
func (db *Database) Node(hash common.Hash) ([]byte, error) {
	return db.diskdb.Get(nil, hash[:])
}

// Nodes retrieves the hashes of all the nodes cached within the memory database.
// This method is extremely expensive and should only be used to validate internal
// states in test code.
func (db *Database) Nodes() []common.Hash {
	db.lock.RLock()
	defer db.lock.RUnlock()

	var hashes = make([]common.Hash, 0, len(db.dirties))
	for hash := range db.dirties {
		if hash != (common.Hash{}) { // Special case for "root" references/nodes
			hashes = append(hashes, hash)
		}
	}
	return hashes
}

// 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.Lock()
	defer db.lock.Unlock()

	db.reference(child, parent)
}

// reference is the private locked version of Reference.
func (db *Database) reference(child common.Hash, parent common.Hash) {
	// If the node does not exist, it's a node pulled from disk, skip
	node, ok := db.dirties[child]
	if !ok {
		return
	}
	// If the reference already exists, only duplicate for roots
	if db.dirties[parent].children == nil {
		db.dirties[parent].children = make(map[common.Hash]uint16)
		db.childrenSize += cachedNodeChildrenSize
	} else if _, ok = db.dirties[parent].children[child]; ok && parent != (common.Hash{}) {
		return
	}
	node.parents++
	db.dirties[parent].children[child]++
	if db.dirties[parent].children[child] == 1 {
		db.childrenSize += common.HashLength + 2 // uint16 counter
	}
}

// Dereference removes an existing reference from a root node.
func (db *Database) Dereference(root common.Hash) {
	// Sanity check to ensure that the meta-root is not removed
	if root == (common.Hash{}) {
		log.Error("Attempted to dereference the trie cache meta root")
		return
	}
	db.lock.Lock()
	defer db.lock.Unlock()

	nodes, storage, start := len(db.dirties), db.dirtiesSize, time.Now()
	db.dereference(root, common.Hash{})

	db.gcnodes += uint64(nodes - len(db.dirties))
	db.gcsize += storage - db.dirtiesSize
	db.gctime += time.Since(start)

	memcacheGCTimeTimer.Update(time.Since(start))
	memcacheGCSizeMeter.Mark(int64(storage - db.dirtiesSize))
	memcacheGCNodesMeter.Mark(int64(nodes - len(db.dirties)))

	log.Debug("Dereferenced trie from memory database", "nodes", nodes-len(db.dirties), "size", storage-db.dirtiesSize, "time", time.Since(start),
		"gcnodes", db.gcnodes, "gcsize", db.gcsize, "gctime", db.gctime, "livenodes", len(db.dirties), "livesize", db.dirtiesSize)
}

// dereference is the private locked version of Dereference.
func (db *Database) dereference(child common.Hash, parent common.Hash) {
	// Dereference the parent-child
	node := db.dirties[parent]

	if node.children != nil && node.children[child] > 0 {
		node.children[child]--
		if node.children[child] == 0 {
			delete(node.children, child)
			db.childrenSize -= (common.HashLength + 2) // uint16 counter
		}
	}
	// If the child does not exist, it's a previously committed node.
	node, ok := db.dirties[child]
	if !ok {
		return
	}
	// If there are no more references to the child, delete it and cascade
	if node.parents > 0 {
		// This is a special cornercase where a node loaded from disk (i.e. not in the
		// memcache any more) gets reinjected as a new node (short node split into full,
		// then reverted into short), causing a cached node to have no parents. That is
		// no problem in itself, but don't make maxint parents out of it.
		node.parents--
	}
	if node.parents == 0 {
		// Remove the node from the flush-list
		switch child {
		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{}
		default:
			db.dirties[node.flushPrev].flushNext = node.flushNext
			db.dirties[node.flushNext].flushPrev = node.flushPrev
		}
		// Dereference all children and delete the node
		for _, hash := range node.childs() {
			db.dereference(hash, child)
		}
		delete(db.dirties, child)
		db.dirtiesSize -= common.StorageSize(common.HashLength + int(node.size))
		if node.children != nil {
			db.childrenSize -= cachedNodeChildrenSize
		}
	}
}

// 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 {
	// Intentionally disabled in Turbo-Geth
	// FIXME: Add more detailed explanation or remove the methid completely
	return nil
}

// Commit iterates over all the children of a particular node, writes them out
// 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.
//
// 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 {
	// Intentionally disabled in Turbo-Geth
	// FIXME: Add more detailed explanation or remove the methid completely
	return nil
}

// 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 := c.db.dirties[hash]
	if !ok {
		return nil
	}
	// Node still exists, remove it from the flush-list
	switch 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:
		c.db.dirties[node.flushPrev].flushNext = node.flushNext
		c.db.dirties[node.flushNext].flushPrev = node.flushPrev
	}
	// Remove the node from the dirty cache
	delete(c.db.dirties, hash)
	c.db.dirtiesSize -= common.StorageSize(common.HashLength + int(node.size))
	if node.children != nil {
		c.db.dirtiesSize -= common.StorageSize(cachedNodeChildrenSize + len(node.children)*(common.HashLength+2))
	}
	// Move the flushed node into the clean cache to prevent insta-reloads
	if c.db.cleans != nil {
		c.db.cleans.Set(string(hash[:]), rlp)
	}
	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
// persistent database layer.
func (db *Database) Size() (common.StorageSize, common.StorageSize) {
	db.lock.RLock()
	defer db.lock.RUnlock()

	// db.dirtiesSize only contains the useful data in the cache, but when reporting
	// the total memory consumption, the maintenance metadata is also needed to be
	// counted.
	var metadataSize = common.StorageSize((len(db.dirties) - 1) * cachedNodeSize)
	var metarootRefs = common.StorageSize(len(db.dirties[common.Hash{}].children) * (common.HashLength + 2))
	return db.dirtiesSize + db.childrenSize + metadataSize - metarootRefs, db.preimagesSize
}

// verifyIntegrity is a debug method to iterate over the entire trie stored in
// memory and check whether every node is reachable from the meta root. The goal
// is to find any errors that might cause memory leaks and or trie nodes to go
// missing.
//
// This method is extremely CPU and memory intensive, only use when must.
func (db *Database) verifyIntegrity() {
	// Iterate over all the cached nodes and accumulate them into a set
	reachable := map[common.Hash]struct{}{{}: {}}

	for child := range db.dirties[common.Hash{}].children {
		db.accumulate(child, reachable)
	}
	// Find any unreachable but cached nodes
	var unreachable []string
	for hash, node := range db.dirties {
		if _, ok := reachable[hash]; !ok {
			unreachable = append(unreachable, fmt.Sprintf("%x: {Node: %v, Parents: %d, Prev: %x, Next: %x}",
				hash, node.node, node.parents, node.flushPrev, node.flushNext))
		}
	}
	if len(unreachable) != 0 {
		panic(fmt.Sprintf("trie cache memory leak: %v", unreachable))
	}
}

// accumulate iterates over the trie defined by hash and accumulates all the
// cached children found in memory.
func (db *Database) accumulate(hash common.Hash, reachable map[common.Hash]struct{}) {
	// Mark the node reachable if present in the memory cache
	node, ok := db.dirties[hash]
	if !ok {
		return
	}
	reachable[hash] = struct{}{}

	// Iterate over all the children and accumulate them too
	for _, child := range node.childs() {
		db.accumulate(child, reachable)
	}
}