tornadocontrib/bsc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

core/state/snapshot: faster account iteration, CLI integration

Browse Source
This commit is contained in:
Martin Holst Swende 2020-01-19 20:57:56 +01:00 committed by Péter Szilágyi
parent 6ddb92a089
commit 19099421dc
No known key found for this signature in database
GPG Key ID: E9AE538CEDF8293D
9 changed files with 69 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
cmd/geth/chaincmd.go
View File

@ -79,6 +79,7 @@ The dumpgenesis command dumps the genesis block configuration in JSON format to
utils.CacheFlag,
utils.SyncModeFlag,
utils.GCModeFlag,
utils.SnapshotFlag,
utils.CacheDatabaseFlag,
utils.CacheGCFlag,
},

1
cmd/geth/main.go
View File

@ -91,6 +91,7 @@ var (
utils.SyncModeFlag,
utils.ExitWhenSyncedFlag,
utils.GCModeFlag,
utils.SnapshotFlag,
utils.LightServeFlag,
utils.LightLegacyServFlag,
utils.LightIngressFlag,

10
cmd/utils/flags.go
View File

@ -225,6 +225,10 @@ var (
Usage: `Blockchain garbage collection mode ("full", "archive")`,
Value: "full",
}
SnapshotFlag = cli.BoolFlag{
Name: "snapshot",
Usage: `Enables snapshot-database mode -- experimental work in progress feature`,
}
LightKDFFlag = cli.BoolFlag{
Name: "lightkdf",
Usage: "Reduce key-derivation RAM & CPU usage at some expense of KDF strength",
@ -1471,6 +1475,9 @@ func SetEthConfig(ctx *cli.Context, stack *node.Node, cfg *eth.Config) {
if ctx.GlobalIsSet(CacheFlag.Name) || ctx.GlobalIsSet(CacheSnapshotFlag.Name) {
cfg.SnapshotCache = ctx.GlobalInt(CacheFlag.Name) * ctx.GlobalInt(CacheSnapshotFlag.Name) / 100
}
if !ctx.GlobalIsSet(SnapshotFlag.Name) {
cfg.SnapshotCache = 0 // Disabled
}
if ctx.GlobalIsSet(DocRootFlag.Name) {
cfg.DocRoot = ctx.GlobalString(DocRootFlag.Name)
}
@ -1734,6 +1741,9 @@ func MakeChain(ctx *cli.Context, stack *node.Node) (chain *core.BlockChain, chai
TrieTimeLimit: eth.DefaultConfig.TrieTimeout,
SnapshotLimit: eth.DefaultConfig.SnapshotCache,
}
if !ctx.GlobalIsSet(SnapshotFlag.Name) {
cache.SnapshotLimit = 0 // Disabled
}
if ctx.GlobalIsSet(CacheFlag.Name) || ctx.GlobalIsSet(CacheTrieFlag.Name) {
cache.TrieCleanLimit = ctx.GlobalInt(CacheFlag.Name) * ctx.GlobalInt(CacheTrieFlag.Name) / 100
}

13
core/blockchain.go
View File

@ -302,8 +302,9 @@ func NewBlockChain(db ethdb.Database, cacheConfig *CacheConfig, chainConfig *par
}
}
// Load any existing snapshot, regenerating it if loading failed
if bc.cacheConfig.SnapshotLimit > 0 {
bc.snaps = snapshot.New(bc.db, bc.stateCache.TrieDB(), bc.cacheConfig.SnapshotLimit, bc.CurrentBlock().Root())
}
// Take ownership of this particular state
go bc.update()
return bc, nil
@ -498,8 +499,9 @@ func (bc *BlockChain) FastSyncCommitHead(hash common.Hash) error {
bc.chainmu.Unlock()
// Destroy any existing state snapshot and regenerate it in the background
if bc.snaps != nil {
bc.snaps.Rebuild(block.Root())
}
log.Info("Committed new head block", "number", block.Number(), "hash", hash)
return nil
}
@ -854,10 +856,13 @@ func (bc *BlockChain) Stop() {
bc.wg.Wait()
// Ensure that the entirety of the state snapshot is journalled to disk.
snapBase, err := bc.snaps.Journal(bc.CurrentBlock().Root())
if err != nil {
var snapBase common.Hash
if bc.snaps != nil {
var err error
if snapBase, err = bc.snaps.Journal(bc.CurrentBlock().Root()); err != nil {
log.Error("Failed to journal state snapshot", "err", err)
}
}
// Ensure the state of a recent block is also stored to disk before exiting.
// We're writing three different states to catch different restart scenarios:
// - HEAD: So we don't need to reprocess any blocks in the general case

16
core/state/snapshot/difflayer.go
View File

@ -23,6 +23,7 @@ import (
"math/rand"
"sort"
"sync"
"sync/atomic"
"time"
"github.com/ethereum/go-ethereum/common"
@ -92,7 +93,7 @@ type diffLayer struct {
memory uint64 // Approximate guess as to how much memory we use
root common.Hash // Root hash to which this snapshot diff belongs to
stale bool // Signals that the layer became stale (state progressed)
stale uint32 // Signals that the layer became stale (state progressed)
accountList []common.Hash // List of account for iteration. If it exists, it's sorted, otherwise it's nil
accountData map[common.Hash][]byte // Keyed accounts for direct retrival (nil means deleted)
@ -237,10 +238,7 @@ func (dl *diffLayer) Parent() snapshot {
// Stale return whether this layer has become stale (was flattened across) or if
// it's still live.
func (dl *diffLayer) Stale() bool {
dl.lock.RLock()
defer dl.lock.RUnlock()
return dl.stale
return atomic.LoadUint32(&dl.stale) != 0
}
// Account directly retrieves the account associated with a particular hash in
@ -288,7 +286,7 @@ func (dl *diffLayer) accountRLP(hash common.Hash, depth int) ([]byte, error) {
// If the layer was flattened into, consider it invalid (any live reference to
// the original should be marked as unusable).
if dl.stale {
if dl.Stale() {
return nil, ErrSnapshotStale
}
// If the account is known locally, return it. Note, a nil account means it was
@ -342,7 +340,7 @@ func (dl *diffLayer) storage(accountHash, storageHash common.Hash, depth int) ([
// If the layer was flattened into, consider it invalid (any live reference to
// the original should be marked as unusable).
if dl.stale {
if dl.Stale() {
return nil, ErrSnapshotStale
}
// If the account is known locally, try to resolve the slot locally. Note, a nil
@ -401,11 +399,9 @@ func (dl *diffLayer) flatten() snapshot {
// Before actually writing all our data to the parent, first ensure that the
// parent hasn't been 'corrupted' by someone else already flattening into it
if parent.stale {
if atomic.SwapUint32(&parent.stale, 1) != 0 {
panic("parent diff layer is stale") // we've flattened into the same parent from two children, boo
}
parent.stale = true
// Overwrite all the updated accounts blindly, merge the sorted list
for hash, data := range dl.accountData {
parent.accountData[hash] = data

33
core/state/snapshot/iterator.go
View File

@ -64,7 +64,7 @@ type diffAccountIterator struct {
// is explicitly tracked since the referenced diff layer might go stale after
// the iterator was positioned and we don't want to fail accessing the old
// value as long as the iterator is not touched any more.
curAccount []byte
//curAccount []byte
layer *diffLayer // Live layer to retrieve values from
keys []common.Hash // Keys left in the layer to iterate
@ -98,22 +98,13 @@ func (it *diffAccountIterator) Next() bool {
if len(it.keys) == 0 {
return false
}
// Iterator seems to be still alive, retrieve and cache the live hash and
// account value, or fail now if layer became stale
it.layer.lock.RLock()
defer it.layer.lock.RUnlock()
if it.layer.stale {
if it.layer.Stale() {
it.fail, it.keys = ErrSnapshotStale, nil
return false
}
// Iterator seems to be still alive, retrieve and cache the live hash
it.curHash = it.keys[0]
if blob, ok := it.layer.accountData[it.curHash]; !ok {
panic(fmt.Sprintf("iterator referenced non-existent account: %x", it.curHash))
} else {
it.curAccount = blob
}
// Values cached, shift the iterator and notify the user of success
// key cached, shift the iterator and notify the user of success
it.keys = it.keys[1:]
return true
}
@ -130,8 +121,22 @@ func (it *diffAccountIterator) Hash() common.Hash {
}
// Account returns the RLP encoded slim account the iterator is currently at.
// This method may _fail_, if the underlying layer has been flattened between
// the call to Next and Acccount. That type of error will set it.Err.
// This method assumes that flattening does not delete elements from
// the accountdata mapping (writing nil into it is fine though), and will panic
// if elements have been deleted.
func (it *diffAccountIterator) Account() []byte {
return it.curAccount
it.layer.lock.RLock()
blob, ok := it.layer.accountData[it.curHash]
if !ok {
panic(fmt.Sprintf("iterator referenced non-existent account: %x", it.curHash))
}
it.layer.lock.RUnlock()
if it.layer.Stale() {
it.fail, it.keys = ErrSnapshotStale, nil
}
return blob
}
// Release is a noop for diff account iterators as there are no held resources.

18
core/state/snapshot/iterator_fast.go
View File

@ -65,6 +65,7 @@ func (its weightedAccountIterators) Swap(i, j int) {
type fastAccountIterator struct {
tree *Tree // Snapshot tree to reinitialize stale sub-iterators with
root common.Hash // Root hash to reinitialize stale sub-iterators through
curAccount []byte
iterators weightedAccountIterators
initiated bool
@ -160,9 +161,20 @@ func (fi *fastAccountIterator) Next() bool {
// Don't forward first time -- we had to 'Next' once in order to
// do the sorting already
fi.initiated = true
return true
fi.curAccount = fi.iterators[0].it.Account()
if innerErr := fi.iterators[0].it.Error(); innerErr != nil {
fi.fail = innerErr
}
return fi.next(0)
return fi.Error() == nil
}
if !fi.next(0) {
return false
}
fi.curAccount = fi.iterators[0].it.Account()
if innerErr := fi.iterators[0].it.Error(); innerErr != nil {
fi.fail = innerErr
}
return fi.Error() == nil
}
// next handles the next operation internally and should be invoked when we know
@ -259,7 +271,7 @@ func (fi *fastAccountIterator) Hash() common.Hash {
// Account returns the current key
func (fi *fastAccountIterator) Account() []byte {
return fi.iterators[0].it.Account()
return fi.curAccount
}
// Release iterates over all the remaining live layer iterators and releases each

2
core/state/snapshot/journal.go
View File

@ -210,7 +210,7 @@ func (dl *diffLayer) Journal(buffer *bytes.Buffer) (common.Hash, error) {
dl.lock.RLock()
defer dl.lock.RUnlock()
if dl.stale {
if dl.Stale() {
return common.Hash{}, ErrSnapshotStale
}
// Everything below was journalled, persist this layer too

3
core/state/snapshot/snapshot.go
View File

@ -22,6 +22,7 @@ import (
"errors"
"fmt"
"sync"
"sync/atomic"
"github.com/ethereum/go-ethereum/common"
"github.com/ethereum/go-ethereum/core/rawdb"
@ -552,7 +553,7 @@ func (t *Tree) Rebuild(root common.Hash) {
case *diffLayer:
// If the layer is a simple diff, simply mark as stale
layer.lock.Lock()
layer.stale = true
atomic.StoreUint32(&layer.stale, 1)
layer.lock.Unlock()
default: