// Can only be called from the handling goroutine for conn,
// inside a transaction.
func followUser(conn *userConn, followId uint64) {
conn.following = append(conn.following, followId)
user := store.GetEntity(followId)
// Send the user's username.
var dataMsg cliproto_down.UserData
dataMsg.UserId = new(uint64)
dataMsg.Key = new(string)
dataMsg.Value = new(string)
dataMsg.FirstUnapplied = new(uint64)
*dataMsg.UserId = followId
*dataMsg.Key = "name username"
*dataMsg.Value = user.Value(*dataMsg.Key)
*dataMsg.FirstUnapplied = store.InstructionFirstUnapplied()
conn.conn.SendProto(6, &dataMsg)
// Send information on all the user's sessions.
user.Attached(func(key string) {
*dataMsg.Key = key
*dataMsg.Value = "true"
conn.conn.SendProto(6, &dataMsg)
})
// Send a done message to indicate that we are finished.
var doneMsg cliproto_down.UserDataDone
doneMsg.UserId = dataMsg.UserId
doneMsg.FirstUnapplied = dataMsg.FirstUnapplied
conn.conn.SendProto(7, &doneMsg)
}
// Must be called inside transaction.
func sendFollowUserIdFail(conn *userConn, userId uint64, reason string) {
var msg cliproto_down.FollowUserIdFailed
msg.UserId = new(uint64)
msg.Reason = new(string)
msg.FirstUnapplied = new(uint64)
*msg.UserId = userId
*msg.Reason = reason
*msg.FirstUnapplied = store.InstructionFirstUnapplied()
conn.conn.SendProto(5, &msg)
}
// Must be called inside transaction.
func sendUserChange(conn *userConn, entityId uint64, key, value string) {
var msg cliproto_down.UserDataChange
msg.UserId = new(uint64)
msg.Key = new(string)
msg.Value = new(string)
msg.FirstUnapplied = new(uint64)
*msg.UserId = entityId
*msg.Key = key
*msg.Value = value
*msg.FirstUnapplied = store.InstructionFirstUnapplied()
conn.conn.SendProto(8, &msg)
}
func (f *followConn) firstUnappliedTimeout() {
store.StartTransaction()
defer store.EndTransaction()
f.lock.Lock()
defer f.lock.Unlock()
if f.closed {
return
}
msg := new(fproto.FirstUnapplied)
msg.FirstUnapplied = new(uint64)
*msg.FirstUnapplied = store.InstructionFirstUnapplied()
f.conn.SendProto(10, msg)
f.firstUnappliedTimer = time.AfterFunc(firstUnappliedTimerDuration,
func() { f.firstUnappliedTimeout() })
}
func process() {
for {
select {
// Try to make any needed outgoing connections.
case <-tryOutgoingCh:
store.StartTransaction()
connectionsLock.Lock()
coreNodes := config.CoreNodes()
if !store.Degraded() {
if config.IsCore() {
for _, node := range coreNodes {
if node == config.Id() {
continue
}
if connections[node] == nil {
go tryOutgoing(node)
}
}
} else {
// Settle for less than 2 core nodes,
// if there *aren't* 2 core nodes.
targetConns := 2
if targetConns > len(coreNodes) {
targetConns = len(coreNodes)
}
newOutgoing := targetConns -
len(connections)
used := -1
for newOutgoing > 0 {
r := processRand.Intn(
len(coreNodes))
// Don't do the same node
// twice.
if r == used {
continue
}
node := coreNodes[r]
if connections[node] == nil {
used = r
go tryOutgoing(node)
newOutgoing--
}
}
}
} else {
if len(connections) == 0 {
r := processRand.Intn(len(coreNodes))
node := coreNodes[r]
if connections[node] == nil {
go tryOutgoing(node)
}
}
}
connectionsLock.Unlock()
store.EndTransaction()
// After a random time, check again.
randDur := time.Duration(processRand.Intn(19)+1) * time.Second
tryOutgoingCh = time.NewTimer(randDur).C
// New received connection.
case conn := <-receivedConnCh:
conn.offerTimers = make(map[uint64]*time.Timer)
store.StartTransaction()
connectionsLock.Lock()
// If we are degraded, reject the connection,
// unless we have no other connection,
// and it is outbound.
if store.Degraded() &&
!(len(connections) == 0 && conn.outgoing) {
conn.lock.Lock()
conn.Close()
conn.lock.Unlock()
connectionsLock.Unlock()
store.EndTransaction()
break
}
// If we have an existing connection to this node,
// close it.
if connections[conn.node] != nil {
other := connections[conn.node]
other.lock.Lock()
other.Close()
if other.firstUnappliedTimer != nil {
other.firstUnappliedTimer.Stop()
}
other.lock.Unlock()
}
// Add to our connections.
connections[conn.node] = conn
// Send initial position and leader messages.
posMsg := new(fproto.Position)
posMsg.FirstUnapplied = new(uint64)
posMsg.Degraded = new(bool)
*posMsg.FirstUnapplied =
store.InstructionFirstUnapplied()
*posMsg.Degraded = store.Degraded()
conn.conn.SendProto(2, posMsg)
proposal, leader := store.Proposal()
leaderMsg := new(fproto.Leader)
leaderMsg.Proposal = new(uint64)
leaderMsg.Leader = new(uint32)
*leaderMsg.Proposal = proposal
*leaderMsg.Leader = uint32(leader)
conn.conn.SendProto(11, leaderMsg)
connectionsLock.Unlock()
store.EndTransaction()
// Start timer for sending first unapplied
// instruction updates.
if config.IsCore() && conn.node <= 0x2000 {
conn.lock.Lock()
conn.firstUnappliedTimer = time.AfterFunc(
firstUnappliedTimerDuration,
func() { conn.firstUnappliedTimeout() })
conn.lock.Unlock()
}
// Start handling received messages from the connection.
go handleConn(conn)
// Terminated connection.
case conn := <-terminatedConnCh:
connectionsLock.Lock()
if connections[conn.node] == conn {
delete(connections, conn.node)
conn.lock.Lock()
conn.closed = true
if conn.firstUnappliedTimer != nil {
conn.firstUnappliedTimer.Stop()
}
conn.lock.Unlock()
}
connectionsLock.Unlock()
}
}
}
// Must be called from the processing goroutine, inside a transaction.
func addPromise(node uint16, msg *coproto.Promise) {
receivedPromises[node] = msg
// If we have promises from a majority of core nodes,
// become leader.
if len(receivedPromises) > len(config.CoreNodes())/2 {
log.Print("core/consensus: became leader")
stopLeaderTimeout()
amLeader = true
proposal, leader := store.Proposal()
// Find a slot number above all those in promise messages,
// and above our first unapplied.
firstUnapplied := store.InstructionFirstUnapplied()
limit := firstUnapplied
for _, msg := range receivedPromises {
for _, accepted := range msg.Accepted {
if *accepted.Slot >= limit {
limit = *accepted.Slot + 1
}
}
}
// Start our next slot after the limit.
nextProposalSlot = limit
// For all slots between this and our first unapplied,
// submit a previously accepted instruction unless we
// know an instruction was already chosen.
// Fills these slots with proposals.
// This is O(n^2) in the number of instructions between our
// first unapplied and limit.
// TODO: Improve worst-case complexity.
start := store.InstructionStart()
slots := store.InstructionSlots()
for i := firstUnapplied; i < limit; i++ {
// If we already have a chosen instruction, skip.
rel := int(i - start)
if len(slots[rel]) == 1 && slots[rel][0].IsChosen() {
continue
}
// Find the previously accepted instruction
// accepted with the highest proposal number.
var bestInst *coproto.Instruction
var bp uint64 // Best proposal
var bl uint16 // Best leader
for _, msg := range receivedPromises {
for _, accepted := range msg.Accepted {
if *accepted.Slot != i {
continue
}
if bestInst == nil {
bestInst = accepted
bp = *accepted.Proposal
bl = uint16(*accepted.Leader)
continue
}
// TODO: This indent is just absurd.
p := *accepted.Proposal
l := uint16(*accepted.Leader)
if store.CompareProposals(p, l,
bp, bl) {
bestInst = accepted
bp = *accepted.Proposal
bl = uint16(*accepted.Leader)
}
}
}
// If we didn't find an instruction, make an empty one.
if bestInst == nil {
empty := new(coproto.ChangeRequest)
empty.RequestEntity = new(uint64)
empty.RequestNode = new(uint32)
*empty.RequestEntity = uint64(config.Id())
*empty.RequestNode = uint32(config.Id())
bestInst := new(coproto.Instruction)
bestInst.Slot = new(uint64)
*bestInst.Slot = i
bestInst.Request = empty
}
// Add proposal timeout.
req := makeExtChangeRequest(bestInst.Request)
addProposalTimeout(i, req)
// Send proposal.
bestInst.Proposal = new(uint64)
bestInst.Leader = new(uint32)
*bestInst.Proposal = proposal
*bestInst.Leader = uint32(leader)
sendProposal(bestInst)
}
// Discard received promise messages.
receivedPromises = nil
// Make an instruction proposal for each waiting change.
for _, req := range waitingRequests {
slot := nextProposalSlot
nextProposalSlot++
addProposalTimeout(slot, req)
inst := makeInst(slot, proposal, leader, req)
sendProposal(inst)
}
// Clear waiting changes.
waitingRequests = nil
}
}
// Top-level function of the processing goroutine.
func process() {
connections = make(map[uint16][]*connect.BaseConn)
// On startup, make an outgoing connection attempt to all other
// core nodes, before continuing.
for _, node := range config.CoreNodes() {
if node == config.Id() {
continue
}
conn, err := connect.Dial(
connect.CONSENSUS_PROTOCOL, node)
if err != nil {
// Can't reach the other node.
continue
}
log.Print("core/consensus: made outgoing connection to ", node)
connections[node] = append(connections[node], conn)
go handleConn(node, conn)
}
// Retry connections once per config.CHANGE_TIMEOUT_PERIOD.
// Largely arbitrary.
reconnectTicker := time.Tick(config.CHANGE_TIMEOUT_PERIOD)
for {
select {
// Connection retry tick.
// We should try to make an outgoing connection to any node
// that we do not have at least one connection to.
// We need to make these asynchronously, because connections
// are slow.
case <-reconnectTicker:
for _, node := range config.CoreNodes() {
if node == config.Id() {
continue
}
if len(connections[node]) > 0 {
continue
}
go outgoingConn(node)
}
// New change request, for us to propose as leader.
case req := <-newChangeCh:
store.StartTransaction()
if !amLeader && receivedPromises != nil {
waitingRequests = append(waitingRequests, req)
} else if !amLeader {
waitingRequests = append(waitingRequests, req)
// Start attempting to be leader.
m := make(map[uint16]*coproto.Promise)
receivedPromises = m
proposal, _ := store.Proposal()
proposal++
store.SetProposal(proposal, config.Id())
firstUn := store.InstructionFirstUnapplied()
// Send prepare messages to all other nodes.
var prepare coproto.Prepare
prepare.Proposal = new(uint64)
prepare.FirstUnapplied = new(uint64)
*prepare.Proposal = proposal
*prepare.FirstUnapplied = firstUn
for _, node := range config.CoreNodes() {
if node == config.Id() {
continue
}
if len(connections[node]) != 0 {
c := connections[node][0]
c.SendProto(2, &prepare)
}
}
// Behave as if we got a promise message
// from ourselves.
var promise coproto.Promise
promise.Proposal = prepare.Proposal
promise.PrevProposal = promise.Proposal
promise.Leader = new(uint32)
*promise.Leader = uint32(config.Id())
promise.PrevLeader = promise.Leader
addPromise(config.Id(), &promise)
} else {
newSlot := nextProposalSlot
nextProposalSlot++
addProposalTimeout(newSlot, req)
proposal, leader := store.Proposal()
inst := makeInst(newSlot, proposal, leader,
req)
sendProposal(inst)
}
store.EndTransaction()
// Leadership attempt timed out.
case timedOutProposal := <-leaderTimeoutCh:
store.StartTransaction()
proposal, leader := store.Proposal()
// If the proposal has changed since that
// leadership attempt, ignore it.
if leader != config.Id() {
return
}
if proposal != timedOutProposal {
return
}
// If we successfully became leader, ignore it.
if amLeader {
return
}
// Otherwise, stop our attempt to become leader.
stopBeingLeader()
store.EndTransaction()
// Proposal timed out.
case timeout := <-proposalTimeoutCh:
store.StartTransaction()
// If this timeout was not canceled, a proposal failed.
// We stop being leader.
if proposalTimeouts[timeout.slot] == timeout {
stopBeingLeader()
}
store.EndTransaction()
// New received connection.
case receivedConn := <-receivedConnCh:
node := receivedConn.node
conn := receivedConn.conn
connections[node] = append(connections[node], conn)
// Received message.
case recvMsg := <-receivedMsgCh:
node := recvMsg.node
conn := recvMsg.conn
msg := recvMsg.msg
switch *msg.MsgType {
case 2:
processPrepare(node, conn, msg.Content)
case 3:
processPromise(node, conn, msg.Content)
case 4:
processAccept(node, conn, msg.Content)
case 5:
processAccepted(node, conn, msg.Content)
case 6:
processNack(node, conn, msg.Content)
default:
// Unknown message.
conn.Close()
}
// Terminate received connection.
case terminatedConn := <-terminatedConnCh:
node := terminatedConn.node
conn := terminatedConn.conn
for i, other := range connections[node] {
if other != conn {
continue
}
conns := connections[node]
conns = append(conns[:i], conns[i+1:]...)
connections[node] = conns
break
}
}
}
}
func handlePosition(f *followConn, content []byte) {
store.StartTransaction()
defer store.EndTransaction()
f.lock.Lock()
defer f.lock.Unlock()
if f.closed {
return
}
var msg fproto.Position
if err := proto.Unmarshal(content, &msg); err != nil {
f.Close()
return
}
if config.IsCore() && f.node <= 0x2000 {
store.SetNodeFirstUnapplied(f.node, *msg.FirstUnapplied)
}
if store.Degraded() && *msg.Degraded {
f.Close()
return
}
start := store.InstructionStart()
if *msg.FirstUnapplied < start {
f.sendingBurst = true
burstMsg := new(baseproto.Message)
burstMsg.MsgType = new(uint32)
burstMsg.Content = []byte{}
*burstMsg.MsgType = 3
f.conn.Send(burstMsg)
go sendBurst(f)
return
} else if *msg.FirstUnapplied <= store.InstructionFirstUnapplied() &&
*msg.Degraded {
f.sendingBurst = true
burstMsg := new(baseproto.Message)
burstMsg.MsgType = new(uint32)
*burstMsg.MsgType = 3
f.conn.Send(burstMsg)
go sendBurst(f)
return
} else if *msg.Degraded {
f.Close()
return
}
// Send all chosen instructions above the first unapplied point.
instructions := store.InstructionSlots()
relativeSlot := int(*msg.FirstUnapplied - store.InstructionStart())
for ; relativeSlot < len(instructions); relativeSlot++ {
slot := store.InstructionStart() + uint64(relativeSlot)
slotValues := instructions[relativeSlot]
if len(slotValues) != 1 || !slotValues[0].IsChosen() {
continue
}
// Convert the change request to our internal format.
sendInstructionData(f, slot, slotValues[0].ChangeRequest())
}
}
// Sends a burst to the given connection.
// The Bursting message should already have been sent,
// and f.sendingBurst set to true, before calling this asynchronously.
func sendBurst(f *followConn) {
log.Print("shared/follow: sending burst to ", f.node)
globalProp := new(fproto.GlobalProperty)
globalProp.Key = new(string)
globalProp.Value = new(string)
entityProp := new(fproto.EntityProperty)
entityProp.Entity = new(uint64)
entityProp.Key = new(string)
entityProp.Value = new(string)
aborted := false
store.Burst(func(key, value string) bool {
*globalProp.Key = key
*globalProp.Value = value
if !f.conn.SendProto(4, globalProp) {
aborted = true
}
return !aborted
}, func(entity uint64, key, value string) bool {
*entityProp.Entity = entity
*entityProp.Key = key
*entityProp.Value = value
if !f.conn.SendProto(5, entityProp) {
aborted = true
}
return !aborted
})
if aborted {
return
}
store.StartTransaction()
defer store.EndTransaction()
f.lock.Lock()
defer f.lock.Unlock()
f.sendingBurst = false
burstDone := new(fproto.BurstDone)
burstDone.FirstUnapplied = new(uint64)
*burstDone.FirstUnapplied = store.InstructionFirstUnapplied()
f.conn.SendProto(6, burstDone)
// Send an instruction chosen message for all chosen instructions above
// our first unapplied point.
instructions := store.InstructionSlots()
firstUnapplied := store.InstructionFirstUnapplied()
relativeSlot := int(firstUnapplied - store.InstructionStart())
for ; relativeSlot < len(instructions); relativeSlot++ {
slot := store.InstructionStart() + uint64(relativeSlot)
slotValues := instructions[relativeSlot]
if len(slotValues) != 1 || !slotValues[0].IsChosen() {
continue
}
// Convert the change request to our internal format.
chosen := new(fproto.InstructionChosen)
chosen.Slot = new(uint64)
*chosen.Slot = slot
f.conn.SendProto(7, chosen)
}
}