Programming Bitcoin Network, part 3

Previous parts:

  1. Part 1
  2. Part 2

DISCLAIMER. In these blog posts I describe only significant changes I made to the code since the last part. There are also insignificant refactorings and improvements I make along the way and don’t explain them here for brevity. Please, check these links to see all the changes:

Full code of this part: Jeiwan/tinybit/part_3

Changes since the previous part: part2…part3

Introduction

Aaand we’re back to continue our journey!

In the last part we finished binary subpackage that serializes and deserializes messages. Also, we finished the versions handshake, i.e. initial exchange of general information between nodes. As a result, our node can now connect to another node, send ‘version’ and ‘verack’ messages, and keep a connection alive.

But, there’s a problem with the latter…

Liveliness check

If you leave the node running for several minutes, it’ll get disconnected from the btcd node and you’ll see this in btcd logs:

[DBG] SRVR: New peer 127.0.0.1:57606 (inbound)
[DBG] PEER: Connected to 127.0.0.1:57606
[DBG] PEER: Sending verack to 127.0.0.1:57606 (inbound)
[DBG] PEER: Received verack from 127.0.0.1:57606 (inbound)
[DBG] PEER: Sending ping to 127.0.0.1:57606 (inbound)
[DBG] PEER: Sending ping to 127.0.0.1:57606 (inbound)
[WRN] PEER: Peer 127.0.0.1:57606 (inbound) no answer for 5m0s -- disconnecting
[INF] SYNC: Lost peer 127.0.0.1:57606 (inbound)
[DBG] SRVR: Removed peer 127.0.0.1:57606 (inbound)

It turns out that the btcd node sends ‘ping’ message to our node and expects an answer. If there’s no answer, it removes our node from the list of its peers. We don’t want that.

Pinging other nodes in a network is a common way of checking their liveliness, i.e. if the other node is still running and responding. If a node fails to reply in time and with proper message, it gets removed from the list of peers. So, we want to handle ‘ping’ message and send proper reply, which is a ‘pong’ message.

‘ping’ and ‘pong’ messages

Both ‘ping’ and ‘pong’ messages are identical and quite minimal:

type MsgPing struct {
	Nonce uint64
}

type MsgPong struct {
	Nonce uint64
}

The only field they have is Nonce, which is just a random number that’s used as an identifier of a ‘ping’ message. ‘pong’ messages simply returns the nonce that was sent in ‘ping’ message. This difference is reflected in their constructors:

func NewPingMsg(network string) (*Message, uint64, error) {
	nonce := rand.Uint64()
	payload := MsgPing{
		Nonce: nonce,
	}

	msg, err := NewMessage("ping", network, payload)

	return msg, nonce, nil
}

func NewPongMsg(network string, nonce uint64) (*Message, error) {
	payload := MsgPong{
		Nonce: nonce,
	}

	msg, err := NewMessage("pong", network, payload)

	return msg, nil
}

These messages are handled differently though. Whenever a node receives a ‘ping’ message, it should immediately built a ‘pong’ message and send it back:

func (n Node) handlePing(header *protocol.MessageHeader, conn io.ReadWriter) error {
	var ping protocol.MsgPing

	lr := io.LimitReader(conn, int64(header.Length))
	if err := binary.NewDecoder(lr).Decode(&ping); err != nil {
	}

	pong, err := protocol.NewPongMsg(n.Network, ping.Nonce)

	msg, err := binary.Marshal(pong)

	if _, err := conn.Write(msg); err != nil {
	}

	return nil
}

If a node receives a ‘pong’ message, it means it sent a ‘ping’ earlier. Sending pings is a part of a mechanism that we haven’t implemented yet, and before starting implementing it we should first build a list of peers our node is aware of.

List of peers

Nodes in the Bitcoin network shouldn’t depend only on one other node. Instead, each node should be connected to several nodes and shouldn’t trust one node more than any other node in the network. This means that nodes should be interchangeable. If one is removed it can be easily replaced by any other node.

We haven’t yet build a list of peers, and there’s no better time to fix this than now. First, let’s define a peer:

type Peer struct {
	Address    net.Addr
	Connection io.ReadWriteCloser
	PongCh     chan uint64
	Services   uint64
	UserAgent  string
	Version    int32
}

func (p Peer) ID() string {
	return p.Address.String()
}

Peers describes everything we want to know and remember about a peer node.

Now, the node can have a list of peers:

// node/node.go
type Node struct {
	...
	Peers        map[string]*Peer
	...
}

I decided to use a map instead of a list because it’s easier to get a node by its address, and removing one peer doesn’t affect other peers in the map (doesn’t change their indexes).

At this stage, we only have one peer node. And that OK for now.

Now, when do we want to add a peer? As soon we connect to a node? Actually, no. Before adding a peer we must first get basic information about it, i.e. finish a version handshake. Thus, we add new nodes as soon as we receive ‘version’ message from them:

// node/cmd_version.go
func (n Node) handleVersion(header *protocol.MessageHeader, conn net.Conn) error {
	var version protocol.MsgVersion

	lr := io.LimitReader(conn, int64(header.Length))
	if err := binary.NewDecoder(lr).Decode(&version); err != nil {
	}

	peer := Peer{
		Address:    conn.RemoteAddr(),
		Connection: conn,
		PongCh:     make(chan uint64),
		Services:   version.Services,
		UserAgent:  version.UserAgent.String,
		Version:    version.Version,
	}

	n.Peers[peer.ID()] = &peer
	...
}

And that’s it. Having a list of peers, we’re now ready to monitor their liveliness.

Monitoring peers’ liveliness

(This is where we implement handling of ‘pong’ messages.)

As soon as a peer is added, a peer liveliness monitor should start running. Let’s define how it should work:

  1. The monitor triggers once in a while and sends a ‘ping’ message to the peer.
  2. It waits for a ‘pong’ message containing the nonce from the ‘ping’ message.
  3. If no ‘pong’ message is received in a certain time span, then the peer is considered dead and is removed from the list.

In Golang terms, we want to have a goroutine running for each peer, that checks their liveliness. This means we want to have several concurrent processes that communicate with each other (or maybe not, what do you think?) and update/remove peers in a list.

Share Memory By Communicating

Traditionally, such concurrent tasks (that work on the same memory segments) are solved with the help of mutexes. Mutexes allow to temporarily lock certain memory segments to prevent multiple simultaneous reads and/or writes.

Golang, on the other hand, incentivizes a different approach, which is outlined in an official Golang blog post Share Memory By Communicating. I recommend you reading it and checking this wonderful code example before moving on.

So, instead of mutexes we’ll be using channels and goroutines. It took me some time to come up with this design:

  1. There will be a part that sends ping messages, waits for replies, and handles ’no reply’ case. We haven’t built anything related to this.
  2. Another part that receives ‘pong’ messages from other nodes. This will be a handler of ‘pong’ messages.
  3. Third part that connects the two above. This part knows what nonce was sent to what node and directs ‘pong’ replies.

Let’s begin with the latter part. This is a simple function like:

func (n Node) monitorPeers() {
	peerPings := make(map[uint64]string)

	for {
		select {
		case nonce := <-n.PongCh:
			peerID := peerPings[nonce]
			if peerID == "" {
				break
			}
			peer := n.Peers[peerID]
			if peer == nil {
				break
			}

			peer.PongCh <- nonce
			delete(peerPings, nonce)

		case pp := <-n.PingCh:
			peerPings[pp.nonce] = pp.peerID
		}
	}
}

We also need two new channels:

// node/node.go
type Node struct {
	...
	PingCh       chan peerPing
	PongCh       chan uint64
	...
}

peerPings is a state data structure that couples ping nonces and peer identifiers. n.PingCh is intended to pass ping notifications that are sent whenever a ‘ping’ message is sent (this will be implemented next). This notifications are as simple as:

type peerPing struct {
	nonce  uint64
	peerID string
}

This is where we share memory by communicating.

Whenever a peerPings notification is received, the state data structure (peerPings) is updated to be aware of a new ping message.

n.PongCh is intended to pass ‘pong’ messages from the handler to monitoring functions. Before doing this we should check the nonce and ensure the peer is stil in the list. After directing the nonce, it should be removed from peerPings to avoid memory leak.

I hope that part is clear.

Now, let’s implement this part:

There will be a part that sends ping messages, waits for replies, and handles ’no reply’ case.

For every new peer our node will do the following:

func (n *Node) monitorPeer(peer *Peer) {
	for {
		time.Sleep(pingIntervalSec * time.Second)

		ping, nonce, err := protocol.NewPingMsg(n.Network)

		msg, err := binary.Marshal(ping)

		if _, err := peer.Connection.Write(msg); err != nil {
			n.disconnectPeer(peer.ID())
		}

First, wait a few minutes. It’s very likely that a newly added peer is alive, so we don’t need to fire a ping message right away. Then, build a ‘ping’ message and send it.

After a ‘ping’ was sent, we need to notify monitorPeers function:

		n.PingCh <- peerPing{
			nonce:  nonce,
			peerID: peer.ID(),
		}

And now we can start waiting for a reply:

		t := time.NewTimer(pingTimeoutSec * time.Second)

		select {
		case pn := <-peer.PongCh:
			if pn != nonce {
				n.disconnectPeer(peer.ID())
				return
			}
			logrus.Debugf("got 'pong' from %s", peer)
		case <-t.C:
			n.disconnectPeer(peer.ID())
			return
		}

        t.Stop()
    }
}

If the timer is triggered before we receive a ‘pong’, the peer must be disconnected. If a ‘pong’ is received, validate the nonce and that’s it – we don’t need to do anything else.

This whole cycle should run indefinitely, as long as the peer is in the list. And it should be started as soon as a peer is added:

// node/cmd_version.go
func (n Node) handleVersion(header *protocol.MessageHeader, conn net.Conn) error {
	...
	n.Peers[peer.ID()] = &peer
	go n.monitorPeer(&peer)
	...
}

Now, the final part:

Another part that receives ‘pong’ messages from other nodes.

We already know how to build these:

func (n Node) handlePong(header *protocol.MessageHeader, conn io.ReadWriter) error {
	var pong protocol.MsgPing

	lr := io.LimitReader(conn, int64(header.Length))
	if err := binary.NewDecoder(lr).Decode(&pong); err != nil {
		return err
	}

	n.PongCh <- pong.Nonce

	return nil
}

Yes, that’s simple: read a ‘pong’ message, decode it, and send to n.PongCh. The other guys will do all the work.

And that’s it! At this point we’re having a node that:

  1. Can stay alive by replying to ‘ping’ messages sent by other nodes.
  2. Maintain a list of peers, check their liveliness, and remove dead ones.

Sweet! Time to add new features!

Mempool Monitor

After so many lines of code we can finally start thinking about usefulness. For the next several blog posts, let’s set a goal of building a Mempool Monitor.

Mempool is simply a list of transactions that haven’t been mined yet. The Bitcoin network is organized in such way that there’re no centralized nodes. As a consequence, all transactions and all blocks are delivered to every node in the network. This means that our node too can received every transaction and every block! But we won’t go for the blocks for now.

Let’s begin with receiving transactions.

‘inv’ message

As I said above, every node in the network receives every transaction. But there’s an optimization was made to reduce the bandwidth and to not literally send every transaction to every node as soon as a new transaction is submitted.

Transactions and blocks transferring happens this way:

  1. When a node gets a new transaction, it sends ‘inv’ message to its peers. ‘inv’ means inventory and it literally says “Hey! I have these…”. But ‘inv’ doesn’t contain full data, only hashes.
  2. Any peer that receives the message can decide where it wants to get full data or not.
  3. If a peer wants full data, it sends a ‘getdata’ reply specifying a list of hashes it want to get full data for.
  4. A node that receives ‘getdata’ checks what objects were requested (transactions or blocks) and sends them in related messages: ’tx’ for transaction and ‘block’ for block (one transaction/block per message).

In code, ‘inv’ message looks like so:

type MsgInv struct {
	Count     uint8
	Inventory []InvVector
}

type InvVector struct {
	Type uint32
	Hash [32]byte
}
  1. Count specifies the number of objects it describes.
  2. Type (besides other values) can be 1 for transactions or 2 for bocks.
  3. Hash is a hash of a transaction or a block.

Serialization and deserialization of this and other new types is omitted to save space. They’re a bit different but not hard. See full changes.

‘getdata’ is absolutely identical.

’tx’ message

And here comes the most difficult part: ’tx’ message.

type MsgTx struct {
	Version    int32
	Flag       uint16
	TxInCount  uint8
	TxIn       []TxInput
	TxOutCount uint8
	TxOut      []TxOutput
	TxWitness  TxWitnessData
	LockTime   uint32
}
  1. Version specifies transaction format version. It’s always 1 as of now.
  2. Flag is very tricky to deserialize and serialize because it can be missing or can be set to 1. When it’s set to 1, the transaction uses Segwit data to store signature script. When it’s missing, the transaction is a legacy one. As of December 2019 there are slightly more than 50% of Segwit transactions in the network, which means that our node must support them too.
  3. TxInCount is the number of inputs.
  4. TxIn is a list of inputs.
  5. TxOutCount is the number of outputs.
  6. TxOut is a list of outputs.
  7. TxWitness is only set when Flag is set to 1. It holds a list of witness data elements.
  8. LockTime is used to delay transactions.

Next, transaction inputs:

type TxInput struct {
	PreviousOutput  OutPoint
	ScriptLength    uint8
	SignatureScript []byte
	Sequence        uint32
}

Inputs reference previous outputs and contain unlocking scripts, which are called signature scripts. OutPoint is a structure that points to a previous output. To find an output, we need to know a transaction hash and output’s index – and these are the only two fields OutPoint has.

Now, let’s look at output:

type TxOutput struct {
	Value          int64
	PkScriptLength uint8
	PkScript       []byte
}

Transaction outputs hold Bitcoin amount and locking script, which is called public key script. Not much interesting here.

But here’s what’s interesting:

type TxWitnessData struct {
	Count   uint8
	Witness []TxWitness
}

type TxWitness struct {
	Length uint8
	Data   []byte
}

Witness data is an alternative way of specifying signature scripts. These are basically the same thing.

Before Segwit soft fork happened there were only signature scripts. The soft fork moved them outside of inputs so they’re not calculated in the transaction merkle tree. To make the change backwards compatible, old SignatureScript was saved and the new field got name “witness”. If a transaction has the witness flag set (Flag field of MsgTx), it must contain witness data and it’s witness data that is used to unlock Bitcoins.

New handlers

Alright, we’re now having all the structures we need to start receiving transactions. Let’s add handlers for the newly implemented messages.

First, ‘inv’:

func (no Node) handleInv(header *protocol.MessageHeader, conn io.ReadWriter) error {
	var inv protocol.MsgInv

	lr := io.LimitReader(conn, int64(header.Length))
	if err := binary.NewDecoder(lr).Decode(&inv); err != nil {
	}

	var getData protocol.MsgGetData
	getData.Count = inv.Count
	getData.Inventory = inv.Inventory

	getDataMsg, err := protocol.NewMessage("getdata", no.Network, getData)
	if err != nil {
	}

	msg, err := binary.Marshal(getDataMsg)
	if err != nil {
	}

	if _, err := conn.Write(msg); err != nil {
	}

	return nil
}

The logic behind this is simple: whenever our node gets an ‘inv’ message, it replies with an identical ‘getdata’ message because we want to get all new transactions (and blocks).

Now, ’tx’ handler:

func (no Node) handleTx(header *protocol.MessageHeader, conn io.ReadWriter) error {
	var tx protocol.MsgTx

	lr := io.LimitReader(conn, int64(header.Length))
	if err := binary.NewDecoder(lr).Decode(&tx); err != nil {
		return err
	}

	logrus.Debugf("transaction: %+v", tx)

	return nil
}

For now, let’s just print a new transaction out and not save it to the mempool.

Setting up a wallet

The time has come to set up a wallet, because we want to send a test transaction to ensure that our node does indeed receive it.

btcd doesn’t have wallet functionality included. Actually, it was extracted into a separate application called btcwallet. So, we’re going to us it.

As of December 2019, there’s a bug in btcwallet that doesn’t allow to use it. We’ll need to fix it, luckily it’s just one line.

Installing btcwallet

Repeat these steps to install btcwallet:

  1. cd ~/.tmp

  2. git clone https://github.com/btcsuite/btcwallet

  3. cd btcwallet

  4. Open file walletsetup.go and, on line 222, change:

    db, err := walletdb.Create("bdb", dbPath)
    

    `` To:

    db, err := walletdb.Create("bdb", dbPath, true)
    

    `` (add a true as the third argument)

  5. Save the file and run:

    $ go install -v . ./cmd/...
    

    ``

  6. btcwallet should now be installed.

I’ve prepared a btcwallet config, we’ll be always using it. The file is called btcwallet.conf.

Installing btcctl

Both btcd and btcwallet are servers, and we need somehow to control them. We need to tell btcd to mine new blocks and we need to use btcwallet to send transactions. There’s one tool to control both of them, it’s called btcctl and it should already be installed in your system if you have btcd installed. Otherwise, install btcd.

I’ve also prepared two configs for btcctl. One of them, btcctl-wallet.conf, is used to interact with the wallet, and the other, btcctl.conf, is used to control the btcd node.

Setting up wallets

We’ll need two wallets: one for the miner and one for the user (we’ll call her Alice).

Before running btcwallet we must create a default wallet. Since, this is purely for development and testing, you can use whatever passphrase you want.

$ btcwallet -C ./btcwallet.conf --create
Enter the private passphrase for your new wallet:
Confirm passphrase:
Do you want to add an additional layer of encryption for public data? (n/no/y/yes) [no]: no
Do you have an existing wallet seed you want to use? (n/no/y/yes) [no]: no
Your wallet generation seed is:
...
Once you have stored the seed in a safe and secure location, enter "OK" to continue: OK
Creating the wallet...
[INF] WLLT: Opened wallet

btcwallet is also a server, so we need to run it alongside btcd:

// Console window 1
$ btcd --configfile ./btcd.conf

// Console window 2
$ btcwallet -C ./btcwallet.conf

With both of the servers running, we can create an account for Alice:

$ btcctl -C ./btcctl-wallet.conf createnewaccount alice
$ btcctl -C ./btcctl-wallet.conf listaccounts
{
  "alice": 0,
  "default": 0,
  "imported": 0
}

Good! Now we need to generate two addresses: one for the miner and one for Alice.

// Unlock your wallet first
$ btcctl -C ./btcctl-wallet.conf walletpassphrase PASSPHRASE 3600
$ btcctl -C ./btcctl-wallet.conf getnewaddress
MINER_ADDRESS
$ btcctl -C ./btcctl-wallet.conf getnewaddress alice
ALICE_ADDRESS

Next step is to setup miner’s address. Stop btcd and start it like that:

$ btcd --configfile ./btcd.conf --miningaddr=MINER_ADDRESS

MINER_ADDRESS will now receive all block rewards!

Let’s mine a few blocks:

$ btcctl -C ./btcctl.conf generate 100
[...a hundred of hashes...]
$ btcctl -C ./btcctl-wallet.conf getbalance
50

Awesome! The balance is 50 BTC because coinbase transactions need 100 confirmations to become spendable. 50 is quite enough for our purposes

Final testing

Now, run or node to see what messages it receives.

$ DEBUG=1 tinybit

Finally, let’s send a few bitcoins to Alice:

$ btcctl -C ./btcctl-wallet.conf sendtoaddress ALICE_ADDRESS 0.00001

And check logs of tinybit:

DEBU[0000] received message: version
DEBU[0000] new peer /btcwire:0.5.0/btcd:0.12.0/ (127.0.0.1:9333)
DEBU[0000] received message: verack
DEBU[0050] received message: inv
DEBU[0050] received message: tx
DEBU[0050] transaction: {Version:1 Flag:0 TxInCount:1 TxIn:[{PreviousOutput:{Hash:[...] Index:0} ScriptLength:106 SignatureScript:[...] Sequence:4294967295}] TxOutCount:2 TxOut:[{Value:1000 PkScriptLength:25 PkScript:[...]} {Value:4999998776 PkScriptLength:22 PkScript:[...]}] TxWitness:{Count:0 Witness:[]} LockTime:0}

Congratulations!

Conclusion

That’s it for today. Our node now can monitor peers and receive transactions! It’s too early to save transactions in the mempool because we don’t know yet how to remove them from there. For this, we first need to learn to handle new blocks.

Also, our node can only connect to a single peer and we definitely want to fix this before running the node in the testnet 😉

See you!

  1. Full code of this part: Jeiwan/tinybit/part_3
  2. Changes since the previous part: part2…part3
  3. A Golang blog post: Share Memory By Communicating
  4. BIP 141, Segregated witness
  5. Bitcoin Protocol documentation