// LastIndex implements the raft.Storage interface
func (s *raftStorage) LastIndex() (uint64, error) {
it := s.db.NewIterator(kv.BytesPrefix(s.entryKeyPrefix))
defer it.Release()
if !it.Last() {
return 0, it.Error()
}
indexPortion := it.Key()[len(s.entryKeyPrefix):]
return binary.BigEndian.Uint64(indexPortion), it.Error()
}
// lastSignedEpoch returns the last epoch for which we have any signature.
func (ks *Keyserver) lastSignedEpoch() uint64 {
iter := ks.db.NewIterator(kv.BytesPrefix([]byte{tableRatificationsPrefix}))
defer iter.Release()
if !iter.Last() {
return 0
}
ret := binary.BigEndian.Uint64(iter.Key()[1 : 1+8])
if iter.Error() != nil {
log.Printf("ERROR: db scan for last seh: %s", iter.Error())
return 0
}
return ret
}
// New initializes a replication.LogReplicator using an already open kv.DB.
func New(db kv.DB, prefix []byte) (replication.LogReplicator, error) {
nextIndex := uint64(0)
iter := db.NewIterator(kv.BytesPrefix(prefix))
if hasEntries := iter.Last(); hasEntries {
nextIndex = binary.BigEndian.Uint64(iter.Key()[len(prefix):]) + 1
}
iter.Release()
if err := iter.Error(); err != nil {
return nil, err
}
leaderHintSet := make(chan bool, 1)
leaderHintSet <- true
return &kvLog{
db: db,
prefix: prefix,
nextIndex: nextIndex,
propose: make(chan replication.LogEntry, 100),
leaderHintSet: leaderHintSet,
waitCommitted: make(chan replication.LogEntry),
stop: make(chan struct{}),
stopped: make(chan struct{}),
}, nil
}
// step is called by run and changes the in-memory state. No i/o allowed.
func (ks *Keyserver) step(step *proto.KeyserverStep, rs *proto.ReplicaState, wb kv.Batch) (deferredIO func()) {
// ks: &const
// step, rs, wb: &mut
switch step.Type.(type) {
case *proto.KeyserverStep_Update:
index := step.GetUpdate().Update.NewEntry.Index
prevUpdate, err := ks.getUpdate(index, math.MaxUint64)
if err != nil {
log.Printf("getUpdate: %s", err)
ks.wr.Notify(step.UID, updateOutput{Error: fmt.Errorf("internal error")})
return
}
if err := ks.verifyUpdateDeterministic(prevUpdate, step.GetUpdate()); err != nil {
ks.wr.Notify(step.UID, updateOutput{Error: err})
return
}
latestTree := ks.merkletree.GetSnapshot(rs.LatestTreeSnapshot)
// sanity check: compare previous version in Merkle tree vs in updates table
prevEntryHashTree, _, err := latestTree.Lookup(index)
if err != nil {
ks.wr.Notify(step.UID, updateOutput{Error: fmt.Errorf("internal error")})
return
}
var prevEntryHash []byte
if prevUpdate != nil {
prevEntryHash = make([]byte, 32)
sha3.ShakeSum256(prevEntryHash, prevUpdate.Update.NewEntry.Encoding)
}
if !bytes.Equal(prevEntryHashTree, prevEntryHash) {
log.Fatalf("ERROR: merkle tree and DB inconsistent for index %x: %x vs %x", index, prevEntryHashTree, prevEntryHash)
}
var entryHash [32]byte
sha3.ShakeSum256(entryHash[:], step.GetUpdate().Update.NewEntry.Encoding)
newTree, err := latestTree.BeginModification()
if err != nil {
ks.wr.Notify(step.UID, updateOutput{Error: fmt.Errorf("internal error")})
return
}
if err := newTree.Set(index, entryHash[:]); err != nil {
log.Printf("setting index '%x' gave error: %s", index, err)
ks.wr.Notify(step.UID, updateOutput{Error: fmt.Errorf("internal error")})
return
}
rs.LatestTreeSnapshot = newTree.Flush(wb).Nr
epochNr := rs.LastEpochDelimiter.EpochNumber + 1
wb.Put(tableUpdateRequests(index, epochNr), proto.MustMarshal(step.GetUpdate()))
ks.wr.Notify(step.UID, updateOutput{Epoch: epochNr})
rs.PendingUpdates = true
ks.updateEpochProposer()
if rs.LastEpochNeedsRatification {
// We need to wait for the last epoch to appear in the verifier log before
// inserting this update.
wb.Put(tableUpdatesPendingRatification(rs.NextIndexLog), proto.MustMarshal(step.GetUpdate().Update))
} else {
// We can deliver the update to verifiers right away.
return ks.verifierLogAppend(&proto.VerifierStep{Type: &proto.VerifierStep_Update{Update: step.GetUpdate().Update}}, rs, wb)
}
case *proto.KeyserverStep_EpochDelimiter:
if step.GetEpochDelimiter().EpochNumber <= rs.LastEpochDelimiter.EpochNumber {
return // a duplicate of this step has already been handled
}
rs.LastEpochDelimiter = *step.GetEpochDelimiter()
log.Printf("epoch %d", step.GetEpochDelimiter().EpochNumber)
rs.PendingUpdates = false
ks.resetEpochTimers(rs.LastEpochDelimiter.Timestamp.Time())
// rs.ThisReplicaNeedsToSignLastEpoch might already be true, if a majority
// signed that did not include us. This will make us skip signing the last
// epoch, but that's fine.
rs.ThisReplicaNeedsToSignLastEpoch = true
// However, it's not okay to see a new epoch delimiter before the previous
// epoch has been ratified.
if rs.LastEpochNeedsRatification {
log.Panicf("new epoch delimiter but last epoch not ratified")
}
rs.LastEpochNeedsRatification = true
ks.updateEpochProposer()
deferredIO = ks.updateSignatureProposer
snapshotNumberBytes := make([]byte, 8)
binary.BigEndian.PutUint64(snapshotNumberBytes, rs.LatestTreeSnapshot)
wb.Put(tableMerkleTreeSnapshot(step.GetEpochDelimiter().EpochNumber), snapshotNumberBytes)
latestTree := ks.merkletree.GetSnapshot(rs.LatestTreeSnapshot)
rootHash, err := latestTree.GetRootHash()
if err != nil {
log.Panicf("ks.latestTree.GetRootHash() failed: %s", err)
}
teh := &proto.EncodedTimestampedEpochHead{TimestampedEpochHead: proto.TimestampedEpochHead{
Head: proto.EncodedEpochHead{EpochHead: proto.EpochHead{
RootHash: rootHash,
PreviousSummaryHash: rs.PreviousSummaryHash,
Realm: ks.realm,
Epoch: step.GetEpochDelimiter().EpochNumber,
IssueTime: step.GetEpochDelimiter().Timestamp,
}, Encoding: nil},
Timestamp: step.GetEpochDelimiter().Timestamp,
}, Encoding: nil}
teh.Head.UpdateEncoding()
teh.UpdateEncoding()
if rs.PreviousSummaryHash == nil {
rs.PreviousSummaryHash = make([]byte, 64)
}
sha3.ShakeSum256(rs.PreviousSummaryHash[:], teh.Head.Encoding)
wb.Put(tableEpochHeads(step.GetEpochDelimiter().EpochNumber), proto.MustMarshal(teh))
case *proto.KeyserverStep_ReplicaSigned:
newSEH := step.GetReplicaSigned()
epochNr := newSEH.Head.Head.Epoch
// get epoch head
tehBytes, err := ks.db.Get(tableEpochHeads(epochNr))
if err != nil {
log.Panicf("get tableEpochHeads(%d): %s", epochNr, err)
}
// compare epoch head to signed epoch head
if got, want := tehBytes, newSEH.Head.Encoding; !bytes.Equal(got, want) {
log.Panicf("replica signed different head: wanted %x, got %x", want, got)
}
// insert all the new signatures into the ratifications table (there should
// actually only be one)
newSehBytes := proto.MustMarshal(newSEH)
for id := range newSEH.Signatures {
// the entry might already exist in the DB (if the proposals got
// duplicated), but it doesn't matter
wb.Put(tableRatifications(epochNr, id), newSehBytes)
}
deferredIO = func() {
// First write to DB, *then* notify subscribers. That way, if subscribers
// start listening before searching the DB, they're guaranteed to see the
// signature: either it's already in the DB, or they'll get notified. If
// the order was reversed, they could miss the notification but still not
// see anything in the DB.
ks.signatureBroadcast.Publish(epochNr, newSEH)
}
if epochNr != rs.LastEpochDelimiter.EpochNumber {
break
}
if rs.ThisReplicaNeedsToSignLastEpoch && newSEH.Signatures[ks.replicaID] != nil {
rs.ThisReplicaNeedsToSignLastEpoch = false
ks.updateEpochProposer()
// updateSignatureProposer should in general be called after writes
// have been flushed to db, but given ThisReplicaNeedsToSignLast =
// false we know that updateSignatureProposer will not access the db.
ks.updateSignatureProposer()
}
// get all existing ratifications for this epoch
allSignatures := make(map[uint64][]byte)
existingRatifications, err := ks.allRatificationsForEpoch(epochNr)
if err != nil {
log.Panicf("allRatificationsForEpoch(%d): %s", epochNr, err)
}
for _, seh := range existingRatifications {
for id, sig := range seh.Signatures {
allSignatures[id] = sig
}
}
// check whether the epoch was already ratified
wasRatified := coname.VerifyPolicy(ks.serverAuthorized, tehBytes, allSignatures)
if wasRatified {
break
}
for id, sig := range newSEH.Signatures {
allSignatures[id] = sig
}
// check whether the epoch has now become ratified
nowRatified := coname.VerifyPolicy(ks.serverAuthorized, tehBytes, allSignatures)
if !nowRatified {
break
}
if !rs.LastEpochNeedsRatification {
log.Panicf("%x: thought last epoch was not already ratified, but it was", ks.replicaID)
}
rs.LastEpochNeedsRatification = false
ks.updateEpochProposer()
var teh proto.EncodedTimestampedEpochHead
err = teh.Unmarshal(tehBytes)
if err != nil {
log.Panicf("invalid epoch head %d (%x): %s", epochNr, tehBytes, err)
}
allSignaturesSEH := &proto.SignedEpochHead{
Head: teh,
Signatures: allSignatures,
}
oldDeferredIO := deferredIO
deferredSendEpoch := ks.verifierLogAppend(&proto.VerifierStep{Type: &proto.VerifierStep_Epoch{Epoch: allSignaturesSEH}}, rs, wb)
deferredSendUpdates := []func(){}
iter := ks.db.NewIterator(kv.BytesPrefix([]byte{tableUpdatesPendingRatificationPrefix}))
defer iter.Release()
for iter.Next() {
update := &proto.SignedEntryUpdate{}
err := update.Unmarshal(iter.Value())
if err != nil {
log.Panicf("invalid pending update %x: %s", iter.Value(), err)
}
deferredSendUpdates = append(deferredSendUpdates, ks.verifierLogAppend(&proto.VerifierStep{Type: &proto.VerifierStep_Update{Update: update}}, rs, wb))
wb.Delete(iter.Key())
}
deferredIO = func() {
oldDeferredIO()
// First, send the ratified epoch to verifiers
deferredSendEpoch()
// Then send updates that were waiting for that epoch to go out
for _, f := range deferredSendUpdates {
f()
}
}
case *proto.KeyserverStep_VerifierSigned:
rNew := step.GetVerifierSigned()
for id := range rNew.Signatures {
// Note: The signature *must* have been authenticated before being inserted
// into the log, or else verifiers could just trample over everyone else's
// signatures, including our own.
dbkey := tableRatifications(rNew.Head.Head.Epoch, id)
wb.Put(dbkey, proto.MustMarshal(rNew))
}
ks.wr.Notify(step.UID, nil)
return func() {
// As above, first write to DB, *then* notify subscribers.
ks.signatureBroadcast.Publish(rNew.Head.Head.Epoch, rNew)
}
default:
log.Panicf("unknown step pb in replicated log: %#v", step)
}
return
}