lnd.xprv/autopilot/agent.go

594 lines
18 KiB
Go
Raw Normal View History

2017-08-11 07:14:41 +03:00
package autopilot
import (
"net"
2017-08-11 07:14:41 +03:00
"sync"
"sync/atomic"
"github.com/btcsuite/btcd/btcec"
"github.com/btcsuite/btcutil"
2018-07-31 10:17:17 +03:00
"github.com/davecgh/go-spew/spew"
"github.com/lightningnetwork/lnd/lnwire"
2017-08-11 07:14:41 +03:00
)
2018-01-30 03:07:26 +03:00
// Config couples all the items that an autopilot agent needs to function.
2017-08-11 07:14:41 +03:00
// All items within the struct MUST be populated for the Agent to be able to
// carry out its duties.
type Config struct {
// Self is the identity public key of the Lightning Network node that
// is being driven by the agent. This is used to ensure that we don't
// accidentally attempt to open a channel with ourselves.
Self *btcec.PublicKey
// Heuristic is an attachment heuristic which will govern to whom we
// open channels to, and also what those channels look like in terms of
// desired capacity. The Heuristic will take into account the current
// state of the graph, our set of open channels, and the amount of
// available funds when determining how channels are to be opened.
// Additionally, a heuristic make also factor in extra-graph
// information in order to make more pertinent recommendations.
Heuristic AttachmentHeuristic
// ChanController is an interface that is able to directly manage the
// creation, closing and update of channels within the network.
ChanController ChannelController
// ConnectToPeer attempts to connect to the peer using one of its
// advertised addresses. The boolean returned signals whether the peer
// was already connected.
ConnectToPeer func(*btcec.PublicKey, []net.Addr) (bool, error)
// DisconnectPeer attempts to disconnect the peer with the given public
// key.
DisconnectPeer func(*btcec.PublicKey) error
2017-08-11 07:14:41 +03:00
// WalletBalance is a function closure that should return the current
// available balance o the backing wallet.
WalletBalance func() (btcutil.Amount, error)
// Graph is an abstract channel graph that the Heuristic and the Agent
// will use to make decisions w.r.t channel allocation and placement
// within the graph.
Graph ChannelGraph
// MaxPendingOpens is the maximum number of pending channel
// establishment goroutines that can be lingering. We cap this value in
// order to control the level of parallelism caused by the autopiloit
// agent.
MaxPendingOpens uint16
2017-08-11 07:14:41 +03:00
// TODO(roasbeef): add additional signals from fee rates and revenue of
// currently opened channels
}
// channelState is a type that represents the set of active channels of the
// backing LN node that the Agent should be ware of. This type contains a few
// helper utility methods.
type channelState map[lnwire.ShortChannelID]Channel
// Channels returns a slice of all the active channels.
2017-08-11 07:14:41 +03:00
func (c channelState) Channels() []Channel {
chans := make([]Channel, 0, len(c))
for _, channel := range c {
chans = append(chans, channel)
}
return chans
}
// ConnectedNodes returns the set of nodes we currently have a channel with.
// This information is needed as we want to avoid making repeated channels with
// any node.
func (c channelState) ConnectedNodes() map[NodeID]struct{} {
nodes := make(map[NodeID]struct{})
for _, channels := range c {
nodes[channels.Node] = struct{}{}
}
// TODO(roasbeef): add outgoing, nodes, allow incoming and outgoing to
// per node
// * only add node is chan as funding amt set
2017-08-11 07:14:41 +03:00
return nodes
}
// Agent implements a closed-loop control system which seeks to autonomously
// optimize the allocation of satoshis within channels throughput the network's
// channel graph. An agent is configurable by swapping out different
// AttachmentHeuristic strategies. The agent uses external signals such as the
// wallet balance changing, or new channels being opened/closed for the local
// node as an indicator to re-examine its internal state, and the amount of
// available funds in order to make updated decisions w.r.t the channel graph.
// The Agent will automatically open, close, and splice in/out channel as
// necessary for it to step closer to its optimal state.
//
// TODO(roasbeef): prob re-word
type Agent struct {
// Only to be used atomically.
started uint32
stopped uint32
// cfg houses the configuration state of the Ant.
cfg Config
// chanState tracks the current set of open channels.
chanState channelState
// stateUpdates is a channel that any external state updates that may
// affect the heuristics of the agent will be sent over.
stateUpdates chan interface{}
// totalBalance is the total number of satoshis the backing wallet is
// known to control at any given instance. This value will be updated
// when the agent receives external balance update signals.
totalBalance btcutil.Amount
quit chan struct{}
wg sync.WaitGroup
}
// New creates a new instance of the Agent instantiated using the passed
// configuration and initial channel state. The initial channel state slice
// should be populated with the set of Channels that are currently opened by
// the backing Lightning Node.
func New(cfg Config, initialState []Channel) (*Agent, error) {
a := &Agent{
cfg: cfg,
chanState: make(map[lnwire.ShortChannelID]Channel),
quit: make(chan struct{}),
stateUpdates: make(chan interface{}),
}
for _, c := range initialState {
a.chanState[c.ChanID] = c
}
return a, nil
}
// Start starts the agent along with any goroutines it needs to perform its
// normal duties.
func (a *Agent) Start() error {
if !atomic.CompareAndSwapUint32(&a.started, 0, 1) {
return nil
}
log.Infof("Autopilot Agent starting")
startingBalance, err := a.cfg.WalletBalance()
if err != nil {
return err
}
a.wg.Add(1)
go a.controller(startingBalance)
return nil
}
// Stop signals the Agent to gracefully shutdown. This function will block
// until all goroutines have exited.
func (a *Agent) Stop() error {
if !atomic.CompareAndSwapUint32(&a.stopped, 0, 1) {
return nil
}
log.Infof("Autopilot Agent stopping")
close(a.quit)
a.wg.Wait()
return nil
}
// balanceUpdate is a type of external state update that reflects an
// increase/decrease in the funds currently available to the wallet.
type balanceUpdate struct {
balanceDelta btcutil.Amount
}
// chanOpenUpdate is a type of external state update the indicates a new
// channel has been opened, either by the Agent itself (within the main
// controller loop), or by an external user to the system.
type chanOpenUpdate struct {
newChan Channel
}
// chanOpenFailureUpdate is a type of external state update that indicates
// a previous channel open failed, and that it might be possible to try again.
type chanOpenFailureUpdate struct{}
2017-08-11 07:14:41 +03:00
// chanCloseUpdate is a type of external state update that indicates that the
// backing Lightning Node has closed a previously open channel.
type chanCloseUpdate struct {
closedChans []lnwire.ShortChannelID
}
// OnBalanceChange is a callback that should be executed each time the balance of
2017-08-11 07:14:41 +03:00
// the backing wallet changes.
func (a *Agent) OnBalanceChange(delta btcutil.Amount) {
go func() {
a.stateUpdates <- &balanceUpdate{
balanceDelta: delta,
}
}()
}
// OnChannelOpen is a callback that should be executed each time a new channel
// is manually opened by the user or any system outside the autopilot agent.
func (a *Agent) OnChannelOpen(c Channel) {
go func() {
a.stateUpdates <- &chanOpenUpdate{
newChan: c,
}
}()
}
// OnChannelOpenFailure is a callback that should be executed when the
// autopilot has attempted to open a channel, but failed. In this case we can
// retry channel creation with a different node.
func (a *Agent) OnChannelOpenFailure() {
go func() {
a.stateUpdates <- &chanOpenFailureUpdate{}
}()
}
2017-08-11 07:14:41 +03:00
// OnChannelClose is a callback that should be executed each time a prior
// channel has been closed for any reason. This includes regular
// closes, force closes, and channel breaches.
func (a *Agent) OnChannelClose(closedChans ...lnwire.ShortChannelID) {
go func() {
a.stateUpdates <- &chanCloseUpdate{
closedChans: closedChans,
}
}()
}
// mergeNodeMaps merges the Agent's set of nodes that it already has active
// channels open to, with the set of nodes that are pending new channels. This
// ensures that the Agent doesn't attempt to open any "duplicate" channels to
// the same node.
func mergeNodeMaps(a map[NodeID]struct{}, b map[NodeID]struct{},
c map[NodeID]Channel) map[NodeID]struct{} {
res := make(map[NodeID]struct{}, len(a)+len(b)+len(c))
for nodeID := range a {
res[nodeID] = struct{}{}
}
for nodeID := range b {
res[nodeID] = struct{}{}
}
for nodeID := range c {
res[nodeID] = struct{}{}
}
return res
}
// mergeChanState merges the Agent's set of active channels, with the set of
// channels awaiting confirmation. This ensures that the agent doesn't go over
// the prescribed channel limit or fund allocation limit.
func mergeChanState(pendingChans map[NodeID]Channel,
activeChans channelState) []Channel {
numChans := len(pendingChans) + len(activeChans)
totalChans := make([]Channel, 0, numChans)
for _, activeChan := range activeChans.Channels() {
totalChans = append(totalChans, activeChan)
}
for _, pendingChan := range pendingChans {
totalChans = append(totalChans, pendingChan)
}
return totalChans
}
2017-08-11 07:14:41 +03:00
// controller implements the closed-loop control system of the Agent. The
// controller will make a decision w.r.t channel placement within the graph
// based on: its current internal state of the set of active channels open,
2017-08-11 07:14:41 +03:00
// and external state changes as a result of decisions it makes w.r.t channel
// allocation, or attributes affecting its control loop being updated by the
// backing Lightning Node.
func (a *Agent) controller(startingBalance btcutil.Amount) {
defer a.wg.Done()
// We'll start off by assigning our starting balance, and injecting
// that amount as an initial wake up to the main controller goroutine.
a.OnBalanceChange(startingBalance)
// TODO(roasbeef): do we in fact need to maintain order?
// * use sync.Cond if so
// failedNodes lists nodes that we've previously attempted to initiate
// channels with, but didn't succeed.
failedNodes := make(map[NodeID]struct{})
// pendingOpens tracks the channels that we've requested to be
// initiated, but haven't yet been confirmed as being fully opened.
// This state is required as otherwise, we may go over our allotted
// channel limit, or open multiple channels to the same node.
pendingOpens := make(map[NodeID]Channel)
var pendingMtx sync.Mutex
updateBalance := func() {
newBalance, err := a.cfg.WalletBalance()
if err != nil {
log.Warnf("unable to update wallet balance: %v", err)
return
}
a.totalBalance = newBalance
}
2017-08-11 07:14:41 +03:00
// TODO(roasbeef): add 10-minute wake up timer
for {
select {
// A new external signal has arrived. We'll use this to update
// our internal state, then determine if we should trigger a
// channel state modification (open/close, splice in/out).
case signal := <-a.stateUpdates:
log.Infof("Processing new external signal")
switch update := signal.(type) {
// The balance of the backing wallet has changed, if
// more funds are now available, we may attempt to open
// up an additional channel, or splice in funds to an
// existing one.
case *balanceUpdate:
log.Debugf("Applying external balance state "+
"update of: %v", update.balanceDelta)
a.totalBalance += update.balanceDelta
// The channel we tried to open previously failed for
// whatever reason.
case *chanOpenFailureUpdate:
log.Debug("Retrying after previous channel open failure.")
updateBalance()
2017-08-11 07:14:41 +03:00
// A new channel has been opened successfully. This was
// either opened by the Agent, or an external system
// that is able to drive the Lightning Node.
case *chanOpenUpdate:
log.Debugf("New channel successfully opened, "+
"updating state with: %v",
spew.Sdump(update.newChan))
newChan := update.newChan
a.chanState[newChan.ChanID] = newChan
pendingMtx.Lock()
delete(pendingOpens, newChan.Node)
pendingMtx.Unlock()
updateBalance()
2017-08-11 07:14:41 +03:00
// A channel has been closed, this may free up an
// available slot, triggering a new channel update.
case *chanCloseUpdate:
log.Debugf("Applying closed channel "+
"updates: %v",
spew.Sdump(update.closedChans))
for _, closedChan := range update.closedChans {
delete(a.chanState, closedChan)
}
updateBalance()
2017-08-11 07:14:41 +03:00
}
pendingMtx.Lock()
log.Debugf("Pending channels: %v", spew.Sdump(pendingOpens))
pendingMtx.Unlock()
2017-08-11 07:14:41 +03:00
// With all the updates applied, we'll obtain a set of
// the current active channels (confirmed channels),
// and also factor in our set of unconfirmed channels.
confirmedChans := a.chanState
pendingMtx.Lock()
totalChans := mergeChanState(pendingOpens, confirmedChans)
pendingMtx.Unlock()
2017-08-11 07:14:41 +03:00
// Now that we've updated our internal state, we'll
// consult our channel attachment heuristic to
// determine if we should open up any additional
// channels or modify existing channels.
availableFunds, numChans, needMore := a.cfg.Heuristic.NeedMoreChans(
totalChans, a.totalBalance,
2017-08-11 07:14:41 +03:00
)
if !needMore {
continue
}
log.Infof("Triggering attachment directive dispatch, "+
"total_funds=%v", a.totalBalance)
2017-08-11 07:14:41 +03:00
// We're to attempt an attachment so we'll o obtain the
// set of nodes that we currently have channels with so
// we avoid duplicate edges.
connectedNodes := a.chanState.ConnectedNodes()
pendingMtx.Lock()
nodesToSkip := mergeNodeMaps(connectedNodes, failedNodes, pendingOpens)
pendingMtx.Unlock()
2017-08-11 07:14:41 +03:00
// If we reach this point, then according to our
// heuristic we should modify our channel state to tend
// towards what it determines to the optimal state. So
// we'll call Select to get a fresh batch of attachment
// directives, passing in the amount of funds available
// for us to use.
chanCandidates, err := a.cfg.Heuristic.Select(
a.cfg.Self, a.cfg.Graph, availableFunds,
numChans, nodesToSkip,
2017-08-11 07:14:41 +03:00
)
if err != nil {
log.Errorf("Unable to select candidates for "+
"attachment: %v", err)
continue
}
if len(chanCandidates) == 0 {
log.Infof("No eligible candidates to connect to")
continue
}
log.Infof("Attempting to execute channel attachment "+
"directives: %v", spew.Sdump(chanCandidates))
// For each recommended attachment directive, we'll
// launch a new goroutine to attempt to carry out the
// directive. If any of these succeed, then we'll
// receive a new state update, taking us back to the
// top of our controller loop.
pendingMtx.Lock()
2017-08-11 07:14:41 +03:00
for _, chanCandidate := range chanCandidates {
// Before we proceed, we'll check to see if
// this attempt would take us past the total
// number of allowed pending opens. If so, then
// we'll skip this round and wait for an
// attempt to either fail or succeed.
if uint16(len(pendingOpens))+1 >
a.cfg.MaxPendingOpens {
log.Debugf("Reached cap of %v pending "+
"channel opens, will retry "+
"after success/failure",
a.cfg.MaxPendingOpens)
continue
}
2017-08-11 07:14:41 +03:00
go func(directive AttachmentDirective) {
// We'll start out by attempting to
// connect to the peer in order to begin
// the funding workflow.
2017-08-11 07:14:41 +03:00
pub := directive.PeerKey
alreadyConnected, err := a.cfg.ConnectToPeer(
pub, directive.Addrs,
)
if err != nil {
log.Warnf("Unable to connect "+
"to %x: %v",
pub.SerializeCompressed(),
err)
// Since we failed to connect to
// them, we'll mark them as
// failed so that we don't
// attempt to connect to them
// again.
nodeID := NewNodeID(pub)
pendingMtx.Lock()
failedNodes[nodeID] = struct{}{}
pendingMtx.Unlock()
// Finally, we'll trigger the
// agent to select new peers to
// connect to.
a.OnChannelOpenFailure()
return
}
// If we were succesful, we'll track
// this peer in our set of pending
// opens. We do this here to ensure we
// don't stall on selecting new peers if
// the connection attempt happens to
// take too long.
pendingMtx.Lock()
if uint16(len(pendingOpens))+1 >
a.cfg.MaxPendingOpens {
pendingMtx.Unlock()
// Since we've reached our max
// number of pending opens,
// we'll disconnect this peer
// and exit. However, if we were
// previously connected to them,
// then we'll make sure to
// maintain the connection
// alive.
if alreadyConnected {
return
}
err = a.cfg.DisconnectPeer(
pub,
)
if err != nil {
log.Warnf("Unable to "+
"disconnect peer "+
"%x: %v",
pub.SerializeCompressed(),
err)
}
return
}
nodeID := NewNodeID(directive.PeerKey)
pendingOpens[nodeID] = Channel{
Capacity: directive.ChanAmt,
Node: nodeID,
}
pendingMtx.Unlock()
// We can then begin the funding
// workflow with this peer.
err = a.cfg.ChanController.OpenChannel(
pub, directive.ChanAmt,
2017-08-11 07:14:41 +03:00
)
if err != nil {
log.Warnf("Unable to open "+
"channel to %x of %v: %v",
pub.SerializeCompressed(),
directive.ChanAmt, err)
// As the attempt failed, we'll
// clear the peer from the set of
// pending opens and mark them
// as failed so we don't attempt
// to open a channel to them
// again.
pendingMtx.Lock()
delete(pendingOpens, nodeID)
failedNodes[nodeID] = struct{}{}
pendingMtx.Unlock()
// Trigger the agent to
// re-evaluate everything and
// possibly retry with a
// different node.
a.OnChannelOpenFailure()
2017-08-11 07:14:41 +03:00
// Finally, we should also
// disconnect the peer if we
// weren't already connected to
// them beforehand by an
// external subsystem.
if alreadyConnected {
return
}
err = a.cfg.DisconnectPeer(pub)
if err != nil {
log.Warnf("Unable to "+
"disconnect peer "+
"%x: %v",
pub.SerializeCompressed(),
err)
}
}
2017-08-11 07:14:41 +03:00
}(chanCandidate)
}
pendingMtx.Unlock()
2017-08-11 07:14:41 +03:00
// The agent has been signalled to exit, so we'll bail out
// immediately.
case <-a.quit:
return
}
}
}