gaspersic/go-ethereum
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

core: improve snapshot journal recovery (#21594)

Browse Source 
* core/state/snapshot: introduce snapshot journal version

* core: update the disk layer in an atomic way

* core: persist the disk layer generator periodically

* core/state/snapshot: improve logging

* core/state/snapshot: forcibly ensure the legacy snapshot is matched

* core/state/snapshot: add debug logs

* core, tests: fix tests and special recovery case

* core: polish

* core: add more blockchain tests for snapshot recovery

* core/state: fix comment

* core: add recovery flag for snapshot

* core: add restart after start-after-crash tests

* core/rawdb: fix imports

* core: fix tests

* core: remove log

* core/state/snapshot: fix snapshot

* core: avoid callbacks in SetHead

* core: fix setHead cornercase where the threshold root has state

* core: small docs for the test cases

Co-authored-by: Péter Szilágyi <peterke@gmail.com>
This commit is contained in:
gary rong
2020-10-30 03:01:58 +08:00
committed by GitHub
parent 43c278cdf9
commit b63e3c37a6
11 changed files with 1792 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files

226
core/state/snapshot/journal.go
View File

@ -33,6 +33,8 @@ import (
"github.com/ethereum/go-ethereum/trie"
)
const journalVersion uint64 = 0
// journalGenerator is a disk layer entry containing the generator progress marker.
type journalGenerator struct {
Wiping bool // Whether the database was in progress of being wiped
@ -61,8 +63,87 @@ type journalStorage struct {
Vals [][]byte
}
// loadAndParseLegacyJournal tries to parse the snapshot journal in legacy format.
func loadAndParseLegacyJournal(db ethdb.KeyValueStore, base *diskLayer) (snapshot, journalGenerator, error) {
// Retrieve the journal, for legacy journal it must exist since even for
// 0 layer it stores whether we've already generated the snapshot or are
// in progress only.
journal := rawdb.ReadSnapshotJournal(db)
if len(journal) == 0 {
return nil, journalGenerator{}, errors.New("missing or corrupted snapshot journal")
}
r := rlp.NewStream(bytes.NewReader(journal), 0)
// Read the snapshot generation progress for the disk layer
var generator journalGenerator
if err := r.Decode(&generator); err != nil {
return nil, journalGenerator{}, fmt.Errorf("failed to load snapshot progress marker: %v", err)
}
// Load all the snapshot diffs from the journal
snapshot, err := loadDiffLayer(base, r)
if err != nil {
return nil, generator, err
}
return snapshot, generator, nil
}
// loadAndParseJournal tries to parse the snapshot journal in latest format.
func loadAndParseJournal(db ethdb.KeyValueStore, base *diskLayer) (snapshot, journalGenerator, error) {
// Retrieve the disk layer generator. It must exist, no matter the
// snapshot is fully generated or not. Otherwise the entire disk
// layer is invalid.
generatorBlob := rawdb.ReadSnapshotGenerator(db)
if len(generatorBlob) == 0 {
return nil, journalGenerator{}, errors.New("missing snapshot generator")
}
var generator journalGenerator
if err := rlp.DecodeBytes(generatorBlob, &generator); err != nil {
return nil, journalGenerator{}, fmt.Errorf("failed to decode snapshot generator: %v", err)
}
// Retrieve the diff layer journal. It's possible that the journal is
// not existent, e.g. the disk layer is generating while that the Geth
// crashes without persisting the diff journal.
// So if there is no journal, or the journal is not matched with disk
// layer, we just discard all diffs and try to recover them later.
journal := rawdb.ReadSnapshotJournal(db)
if len(journal) == 0 {
log.Warn("Loaded snapshot journal", "diskroot", base.root, "diffs", "missing")
return base, generator, nil
}
r := rlp.NewStream(bytes.NewReader(journal), 0)
// Firstly, resolve the first element as the journal version
version, err := r.Uint()
if err != nil {
return nil, journalGenerator{}, err
}
if version != journalVersion {
return nil, journalGenerator{}, fmt.Errorf("journal version mismatch, want %d got %v", journalVersion, version)
}
// Secondly, resolve the disk layer root, ensure it's continuous
// with disk layer.
var root common.Hash
if err := r.Decode(&root); err != nil {
return nil, journalGenerator{}, errors.New("missing disk layer root")
}
// The diff journal is not matched with disk, discard them.
// It can happen that Geth crashes without persisting the latest
// diff journal.
if !bytes.Equal(root.Bytes(), base.root.Bytes()) {
log.Warn("Loaded snapshot journal", "diskroot", base.root, "diffs", "unmatched")
return base, generator, nil
}
// Load all the snapshot diffs from the journal
snapshot, err := loadDiffLayer(base, r)
if err != nil {
return nil, journalGenerator{}, err
}
log.Debug("Loaded snapshot journal", "diskroot", base.root, "diffhead", snapshot.Root())
return snapshot, generator, nil
}
// loadSnapshot loads a pre-existing state snapshot backed by a key-value store.
func loadSnapshot(diskdb ethdb.KeyValueStore, triedb *trie.Database, cache int, root common.Hash) (snapshot, error) {
func loadSnapshot(diskdb ethdb.KeyValueStore, triedb *trie.Database, cache int, root common.Hash, recovery bool) (snapshot, error) {
// Retrieve the block number and hash of the snapshot, failing if no snapshot
// is present in the database (or crashed mid-update).
baseRoot := rawdb.ReadSnapshotRoot(diskdb)
@ -75,28 +156,36 @@ func loadSnapshot(diskdb ethdb.KeyValueStore, triedb *trie.Database, cache int,
cache: fastcache.New(cache * 1024 * 1024),
root: baseRoot,
}
// Retrieve the journal, it must exist since even for 0 layer it stores whether
// we've already generated the snapshot or are in progress only
journal := rawdb.ReadSnapshotJournal(diskdb)
if len(journal) == 0 {
return nil, errors.New("missing or corrupted snapshot journal")
var legacy bool
snapshot, generator, err := loadAndParseJournal(diskdb, base)
if err != nil {
log.Debug("Failed to load new-format journal", "error", err)
snapshot, generator, err = loadAndParseLegacyJournal(diskdb, base)
legacy = true
}
r := rlp.NewStream(bytes.NewReader(journal), 0)
// Read the snapshot generation progress for the disk layer
var generator journalGenerator
if err := r.Decode(&generator); err != nil {
return nil, fmt.Errorf("failed to load snapshot progress marker: %v", err)
}
// Load all the snapshot diffs from the journal
snapshot, err := loadDiffLayer(base, r)
if err != nil {
return nil, err
}
// Entire snapshot journal loaded, sanity check the head and return
// Journal doesn't exist, don't worry if it's not supposed to
// Entire snapshot journal loaded, sanity check the head. If the loaded
// snapshot is not matched with current state root, print a warning log
// or discard the entire snapshot it's legacy snapshot.
//
// Possible scenario: Geth was crashed without persisting journal and then
// restart, the head is rewound to the point with available state(trie)
// which is below the snapshot. In this case the snapshot can be recovered
// by re-executing blocks but right now it's unavailable.
if head := snapshot.Root(); head != root {
return nil, fmt.Errorf("head doesn't match snapshot: have %#x, want %#x", head, root)
// If it's legacy snapshot, or it's new-format snapshot but
// it's not in recovery mode, returns the error here for
// rebuilding the entire snapshot forcibly.
if legacy || !recovery {
return nil, fmt.Errorf("head doesn't match snapshot: have %#x, want %#x", head, root)
}
// It's in snapshot recovery, the assumption is held that
// the disk layer is always higher than chain head. It can
// be eventually recovered when the chain head beyonds the
// disk layer.
log.Warn("Snapshot is not continuous with chain", "snaproot", head, "chainroot", root)
}
// Everything loaded correctly, resume any suspended operations
if !generator.Done {
@ -203,7 +292,9 @@ func (dl *diskLayer) Journal(buffer *bytes.Buffer) (common.Hash, error) {
if dl.stale {
return common.Hash{}, ErrSnapshotStale
}
// Write out the generator marker
// Write out the generator marker. Note it's a standalone disk layer generator
// which is not mixed with journal. It's ok if the generator is persisted while
// journal is not.
entry := journalGenerator{
Done: dl.genMarker == nil,
Marker: dl.genMarker,
@ -214,9 +305,12 @@ func (dl *diskLayer) Journal(buffer *bytes.Buffer) (common.Hash, error) {
entry.Slots = stats.slots
entry.Storage = uint64(stats.storage)
}
if err := rlp.Encode(buffer, entry); err != nil {
blob, err := rlp.EncodeToBytes(entry)
if err != nil {
return common.Hash{}, err
}
log.Debug("Journalled disk layer", "root", dl.root, "complete", dl.genMarker == nil)
rawdb.WriteSnapshotGenerator(dl.diskdb, blob)
return dl.root, nil
}
@ -266,5 +360,97 @@ func (dl *diffLayer) Journal(buffer *bytes.Buffer) (common.Hash, error) {
if err := rlp.Encode(buffer, storage); err != nil {
return common.Hash{}, err
}
log.Debug("Journalled diff layer", "root", dl.root, "parent", dl.parent.Root())
return base, nil
}
// LegacyJournal writes the persistent layer generator stats into a buffer
// to be stored in the database as the snapshot journal.
//
// Note it's the legacy version which is only used in testing right now.
func (dl *diskLayer) LegacyJournal(buffer *bytes.Buffer) (common.Hash, error) {
// If the snapshot is currently being generated, abort it
var stats *generatorStats
if dl.genAbort != nil {
abort := make(chan *generatorStats)
dl.genAbort <- abort
if stats = <-abort; stats != nil {
stats.Log("Journalling in-progress snapshot", dl.root, dl.genMarker)
}
}
// Ensure the layer didn't get stale
dl.lock.RLock()
defer dl.lock.RUnlock()
if dl.stale {
return common.Hash{}, ErrSnapshotStale
}
// Write out the generator marker
entry := journalGenerator{
Done: dl.genMarker == nil,
Marker: dl.genMarker,
}
if stats != nil {
entry.Wiping = (stats.wiping != nil)
entry.Accounts = stats.accounts
entry.Slots = stats.slots
entry.Storage = uint64(stats.storage)
}
if err := rlp.Encode(buffer, entry); err != nil {
return common.Hash{}, err
}
return dl.root, nil
}
// Journal writes the memory layer contents into a buffer to be stored in the
// database as the snapshot journal.
//
// Note it's the legacy version which is only used in testing right now.
func (dl *diffLayer) LegacyJournal(buffer *bytes.Buffer) (common.Hash, error) {
// Journal the parent first
base, err := dl.parent.LegacyJournal(buffer)
if err != nil {
return common.Hash{}, err
}
// Ensure the layer didn't get stale
dl.lock.RLock()
defer dl.lock.RUnlock()
if dl.Stale() {
return common.Hash{}, ErrSnapshotStale
}
// Everything below was journalled, persist this layer too
if err := rlp.Encode(buffer, dl.root); err != nil {
return common.Hash{}, err
}
destructs := make([]journalDestruct, 0, len(dl.destructSet))
for hash := range dl.destructSet {
destructs = append(destructs, journalDestruct{Hash: hash})
}
if err := rlp.Encode(buffer, destructs); err != nil {
return common.Hash{}, err
}
accounts := make([]journalAccount, 0, len(dl.accountData))
for hash, blob := range dl.accountData {
accounts = append(accounts, journalAccount{Hash: hash, Blob: blob})
}
if err := rlp.Encode(buffer, accounts); err != nil {
return common.Hash{}, err
}
storage := make([]journalStorage, 0, len(dl.storageData))
for hash, slots := range dl.storageData {
keys := make([]common.Hash, 0, len(slots))
vals := make([][]byte, 0, len(slots))
for key, val := range slots {
keys = append(keys, key)
vals = append(vals, val)
}
storage = append(storage, journalStorage{Hash: hash, Keys: keys, Vals: vals})
}
if err := rlp.Encode(buffer, storage); err != nil {
return common.Hash{}, err
}
log.Debug("Journalled diff layer", "root", dl.root, "parent", dl.parent.Root())
return base, nil
}