README.md

ics	title	stage	category	kind	requires	author	created	modified
4	Channel & Packet Semantics	draft	IBC/TAO	instantiation	2, 3, 5, 24	Christopher Goes <cwgoes@tendermint.com>	2019-03-07	2019-08-25

Synopsis

The "channel" abstraction provides message delivery semantics to the interblockchain communication protocol, in three categories: ordering, exactly-once delivery, and module permissioning. A channel serves as a conduit for packets passing between a module on one chain and a module on another, ensuring that packets are executed only once, delivered in the order in which they were sent (if necessary), and delivered only to the corresponding module owning the other end of the channel on the destination chain. Each channel is associated with a particular connection, and a connection may have any number of associated channels, allowing the use of common identifiers and amortising the cost of header verification across all the channels utilising a connection & light client.

Channels are payload-agnostic. The modules which send and receive IBC packets decide how to construct packet data and how to act upon the incoming packet data, and must utilise their own application logic to determine which state transactions to apply according to what data the packet contains.

Motivation

The interblockchain communication protocol uses a cross-chain message passing model. IBC packets are relayed from one blockchain to the other by external relayer processes. Chain A and chain B confirm new blocks independently, and packets from one chain to the other may be delayed, censored, or re-ordered arbitrarily. Packets are visible to relayers and can be read from a blockchain by any relayer process and submitted to any other blockchain.

The IBC protocol must provide ordering (for ordered channels) and exactly-once delivery guarantees to allow applications to reason about the combined state of connected modules on two chains.

Example: An application may wish to allow a single tokenized asset to be transferred between and held on multiple blockchains while preserving fungibility and conservation of supply. The application can mint asset vouchers on chain B when a particular IBC packet is committed to chain B, and require outgoing sends of that packet on chain A to escrow an equal amount of the asset on chain A until the vouchers are later redeemed back to chain A with an IBC packet in the reverse direction. This ordering guarantee along with correct application logic can ensure that total supply is preserved across both chains and that any vouchers minted on chain B can later be redeemed back to chain A.

In order to provide the desired ordering, exactly-once delivery, and module permissioning semantics to the application layer, the interblockchain communication protocol must implement an abstraction to enforce these semantics — channels are this abstraction.

Definitions

ConsensusState is as defined in ICS 2.

Connection is as defined in ICS 3.

Port and authenticateCapability are as defined in ICS 5.

hash is a generic collision-resistant hash function, the specifics of which must be agreed on by the modules utilising the channel. hash can be defined differently by different chains.

Identifier, get, set, delete, getCurrentHeight, and module-system related primitives are as defined in ICS 24.

A channel is a pipeline for exactly-once packet delivery between specific modules on separate blockchains, which has at least one end capable of sending packets and one end capable of receiving packets.

A bidirectional channel is a channel where packets can flow in both directions: from A to B and from B to A.

A unidirectional channel is a channel where packets can only flow in one direction: from A to B (or from B to A, the order of naming is arbitrary).

An ordered channel is a channel where packets are delivered exactly in the order which they were sent. This channel type offers a very strict guarantee of ordering. Either, the packets are received in the order they were sent, or if a packet in the sequence times out; then all future packets are also not receivable and the channel closes.

An ordered_allow_timeout channel is a less strict version of the ordered channel. Here, the channel logic will take a best effort approach to delivering the packets in order. In a stream of packets, the channel will relay all packets in order and if a packet in the stream times out, the timeout logic for that packet will execute and the rest of the later packets will continue processing in order. Thus, we do not close the channel on a timeout with this channel type.

An unordered channel is a channel where packets can be delivered in any order, which may differ from the order in which they were sent.

enum ChannelOrder {
  ORDERED,
  UNORDERED,
  ORDERED_ALLOW_TIMEOUT,
}

Directionality and ordering are independent, so one can speak of a bidirectional unordered channel, a unidirectional ordered channel, etc.

All channels provide exactly-once packet delivery, meaning that a packet sent on one end of a channel is delivered no more and no less than once, eventually, to the other end.

This specification only concerns itself with bidirectional channels. Unidirectional channels can use almost exactly the same protocol and will be outlined in a future ICS.

An end of a channel is a data structure on one chain storing channel metadata:

interface ChannelEnd {
  state: ChannelState
  ordering: ChannelOrder
  counterpartyPortIdentifier: Identifier
  counterpartyChannelIdentifier: Identifier
  connectionHops: [Identifier]
  version: string
}

• The state is the current state of the channel end.
• The ordering field indicates whether the channel is unordered, ordered, or ordered_allow_timeout.
• The counterpartyPortIdentifier identifies the port on the counterparty chain which owns the other end of the channel.
• The counterpartyChannelIdentifier identifies the channel end on the counterparty chain.
• The nextSequenceSend, stored separately, tracks the sequence number for the next packet to be sent.
• The nextSequenceRecv, stored separately, tracks the sequence number for the next packet to be received.
• The nextSequenceAck, stored separately, tracks the sequence number for the next packet to be acknowledged.
• The connectionHops stores the list of connection identifiers, in order, along which packets sent on this channel will travel. At the moment this list must be of length 1. In the future multi-hop channels may be supported.
• The version string stores an opaque channel version, which is agreed upon during the handshake. This can determine module-level configuration such as which packet encoding is used for the channel. This version is not used by the core IBC protocol. If the version string contains structured metadata for the application to parse and interpret, then it is considered best practice to encode all metadata in a JSON struct and include the marshalled string in the version field.

Channel ends have a state:

enum ChannelState {
  INIT,
  TRYOPEN,
  OPEN,
  CLOSED,
}

• A channel end in INIT state has just started the opening handshake.
• A channel end in TRYOPEN state has acknowledged the handshake step on the counterparty chain.
• A channel end in OPEN state has completed the handshake and is ready to send and receive packets.
• A channel end in CLOSED state has been closed and can no longer be used to send or receive packets.

A Packet, in the interblockchain communication protocol, is a particular interface defined as follows:

interface Packet {
  sequence: uint64
  timeoutHeight: Height
  timeoutTimestamp: uint64
  sourcePort: Identifier
  sourceChannel: Identifier
  destPort: Identifier
  destChannel: Identifier
  data: bytes
}

• The sequence number corresponds to the order of sends and receives, where a packet with an earlier sequence number must be sent and received before a packet with a later sequence number.
• The timeoutHeight indicates a consensus height on the destination chain after which the packet will no longer be processed, and will instead count as having timed-out.
• The timeoutTimestamp indicates a timestamp on the destination chain after which the packet will no longer be processed, and will instead count as having timed-out.
• The sourcePort identifies the port on the sending chain.
• The sourceChannel identifies the channel end on the sending chain.
• The destPort identifies the port on the receiving chain.
• The destChannel identifies the channel end on the receiving chain.
• The data is an opaque value which can be defined by the application logic of the associated modules.

Note that a Packet is never directly serialised. Rather it is an intermediary structure used in certain function calls that may need to be created or processed by modules calling the IBC handler.

An OpaquePacket is a packet, but cloaked in an obscuring data type by the host state machine, such that a module cannot act upon it other than to pass it to the IBC handler. The IBC handler can cast a Packet to an OpaquePacket and vice versa.

type OpaquePacket = object

In order to enable new channel types (e.g. ORDERED_ALLOW_TIMEOUT), the protocol introduces standardized packet receipts that will serve as sentinel values for the receiving chain to expliclity write to its store the outcome of a recvPacket.

enum PacketReceipt {
  SUCCESSFUL_RECEIPT,
  TIMEOUT_RECEIPT,
}

Desired Properties

Efficiency

• The speed of packet transmission and confirmation should be limited only by the speed of the underlying chains. Proofs should be batchable where possible.

Exactly-once delivery

• IBC packets sent on one end of a channel should be delivered exactly once to the other end.
• No network synchrony assumptions should be required for exactly-once safety. If one or both of the chains halt, packets may be delivered no more than once, and once the chains resume packets should be able to flow again.

Ordering

• On ordered channels, packets should be sent and received in the same order: if packet x is sent before packet y by a channel end on chain A, packet x must be received before packet y by the corresponding channel end on chain B. If packet x is sent before packet y by a channel and packet x is timed out; then packet y and any packet sent after x cannot be received.
• On ordered_allow_timeout channels, packets should be sent and received in the same order: if packet x is sent before packet y by a channel end on chain A, packet x must be received or timed out before packet y by the corresponding channel end on chain B.
• On unordered channels, packets may be sent and received in any order. Unordered packets, like ordered packets, have individual timeouts specified in terms of the destination chain's height.

Permissioning

• Channels should be permissioned to one module on each end, determined during the handshake and immutable afterwards (higher-level logic could tokenize channel ownership by tokenising ownership of the port). Only the module associated with a channel end should be able to send or receive on it.

Technical Specification

Dataflow visualisation

The architecture of clients, connections, channels and packets:

Preliminaries

Store paths

Channel structures are stored under a store path prefix unique to a combination of a port identifier and channel identifier:

function channelPath(portIdentifier: Identifier, channelIdentifier: Identifier): Path {
    return "channelEnds/ports/{portIdentifier}/channels/{channelIdentifier}"
}

The capability key associated with a channel is stored under the channelCapabilityPath:

function channelCapabilityPath(portIdentifier: Identifier, channelIdentifier: Identifier): Path {
  return "{channelPath(portIdentifier, channelIdentifier)}/key"
}

The nextSequenceSend, nextSequenceRecv, and nextSequenceAck unsigned integer counters are stored separately so they can be proved individually:

function nextSequenceSendPath(portIdentifier: Identifier, channelIdentifier: Identifier): Path {
    return "nextSequenceSend/ports/{portIdentifier}/channels/{channelIdentifier}"
}

function nextSequenceRecvPath(portIdentifier: Identifier, channelIdentifier: Identifier): Path {
    return "nextSequenceRecv/ports/{portIdentifier}/channels/{channelIdentifier}"
}

function nextSequenceAckPath(portIdentifier: Identifier, channelIdentifier: Identifier): Path {
    return "nextSequenceAck/ports/{portIdentifier}/channels/{channelIdentifier}"
}

Constant-size commitments to packet data fields are stored under the packet sequence number:

function packetCommitmentPath(portIdentifier: Identifier, channelIdentifier: Identifier, sequence: uint64): Path {
    return "commitments/ports/{portIdentifier}/channels/{channelIdentifier}/packets/{sequence}"
}

Absence of the path in the store is equivalent to a zero-bit.

Packet receipt data are stored under the packetReceiptPath. In the case of a successful receive, the destination chain writes a sentinel success value of SUCCESSFUL_RECEIPT. Some channel types MAY write a sentinel timeout value TIMEOUT_RECEIPT if the packet is received after the specified timeout.

function packetReceiptPath(portIdentifier: Identifier, channelIdentifier: Identifier, sequence: uint64): Path {
    return "receipts/ports/{portIdentifier}/channels/{channelIdentifier}/receipts/{sequence}"
}

Packet acknowledgement data are stored under the packetAcknowledgementPath:

function packetAcknowledgementPath(portIdentifier: Identifier, channelIdentifier: Identifier, sequence: uint64): Path {
    return "acks/ports/{portIdentifier}/channels/{channelIdentifier}/sequences/{sequence}"
}

Versioning

During the handshake process, two ends of a channel come to agreement on a version bytestring associated with that channel. The contents of this version bytestring are and will remain opaque to the IBC core protocol. Host state machines MAY utilise the version data to indicate supported IBC/APP protocols, agree on packet encoding formats, or negotiate other channel-related metadata related to custom logic on top of IBC.

Host state machines MAY also safely ignore the version data or specify an empty string.

Sub-protocols

Note: If the host state machine is utilising object capability authentication (see ICS 005), all functions utilising ports take an additional capability parameter.

Identifier validation

Channels are stored under a unique (portIdentifier, channelIdentifier) prefix. The validation function validatePortIdentifier MAY be provided.

type validateChannelIdentifier = (portIdentifier: Identifier, channelIdentifier: Identifier) => boolean

If not provided, the default validateChannelIdentifier function will always return true.

Channel lifecycle management

Initiator	Datagram	Chain acted upon	Prior state (A, B)	Posterior state (A, B)
Actor	ChanOpenInit	A	(none, none)	(INIT, none)
Relayer	ChanOpenTry	B	(INIT, none)	(INIT, TRYOPEN)
Relayer	ChanOpenAck	A	(INIT, TRYOPEN)	(OPEN, TRYOPEN)
Relayer	ChanOpenConfirm	B	(OPEN, TRYOPEN)	(OPEN, OPEN)

Initiator	Datagram	Chain acted upon	Prior state (A, B)	Posterior state (A, B)
Actor	ChanCloseInit	A	(OPEN, OPEN)	(CLOSED, OPEN)
Relayer	ChanCloseConfirm	B	(CLOSED, OPEN)	(CLOSED, CLOSED)

Opening handshake

The chanOpenInit function is called by a module to initiate a channel opening handshake with a module on another chain.

The opening channel must provide the identifiers of the local channel identifier, local port, remote port, and remote channel identifier.

When the opening handshake is complete, the module which initiates the handshake will own the end of the created channel on the host ledger, and the counterparty module which it specifies will own the other end of the created channel on the counterparty chain. Once a channel is created, ownership cannot be changed (although higher-level abstractions could be implemented to provide this).

Chains MUST implement a function generateIdentifier which chooses an identifier, e.g. by incrementing a counter:

type generateIdentifier = () -> Identifier

function chanOpenInit(
  order: ChannelOrder,
  connectionHops: [Identifier],
  portIdentifier: Identifier,
  counterpartyPortIdentifier: Identifier,
  version: string): CapabilityKey {
    channelIdentifier = generateIdentifier()
    abortTransactionUnless(validateChannelIdentifier(portIdentifier, channelIdentifier))

    abortTransactionUnless(connectionHops.length === 1) // for v1 of the IBC protocol

    abortTransactionUnless(provableStore.get(channelPath(portIdentifier, channelIdentifier)) === null)
    connection = provableStore.get(connectionPath(connectionHops[0]))

    // optimistic channel handshakes are allowed
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(authenticateCapability(portPath(portIdentifier), portCapability))
    channel = ChannelEnd{INIT, order, counterpartyPortIdentifier,
                         "", connectionHops, version}
    provableStore.set(channelPath(portIdentifier, channelIdentifier), channel)
    channelCapability = newCapability(channelCapabilityPath(portIdentifier, channelIdentifier))
    provableStore.set(nextSequenceSendPath(portIdentifier, channelIdentifier), 1)
    provableStore.set(nextSequenceRecvPath(portIdentifier, channelIdentifier), 1)
    provableStore.set(nextSequenceAckPath(portIdentifier, channelIdentifier), 1)
    return channelCapability
}

The chanOpenTry function is called by a module to accept the first step of a channel opening handshake initiated by a module on another chain.

function chanOpenTry(
  order: ChannelOrder,
  connectionHops: [Identifier],
  portIdentifier: Identifier,
  counterpartyChosenChannelIdentifer: Identifier,
  counterpartyPortIdentifier: Identifier,
  counterpartyChannelIdentifier: Identifier,
  version: string, // deprecated
  counterpartyVersion: string,
  proofInit: CommitmentProof,
  proofHeight: Height): CapabilityKey {
    channelIdentifier = generateIdentifier()

    abortTransactionUnless(validateChannelIdentifier(portIdentifier, channelIdentifier))
    abortTransactionUnless(connectionHops.length === 1) // for v1 of the IBC protocol
    abortTransactionUnless(authenticateCapability(portPath(portIdentifier), portCapability))
    connection = provableStore.get(connectionPath(connectionHops[0]))
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)
    expected = ChannelEnd{INIT, order, portIdentifier,
                          "", [connection.counterpartyConnectionIdentifier], counterpartyVersion}
    abortTransactionUnless(connection.verifyChannelState(
      proofHeight,
      proofInit,
      counterpartyPortIdentifier,
      counterpartyChannelIdentifier,
      expected
    ))
    channel = ChannelEnd{TRYOPEN, order, counterpartyPortIdentifier,
                         counterpartyChannelIdentifier, connectionHops, version}
    provableStore.set(channelPath(portIdentifier, channelIdentifier), channel)
    channelCapability = newCapability(channelCapabilityPath(portIdentifier, channelIdentifier))

    // initialize channel sequences
    provableStore.set(nextSequenceSendPath(portIdentifier, channelIdentifier), 1)
    provableStore.set(nextSequenceRecvPath(portIdentifier, channelIdentifier), 1)
    provableStore.set(nextSequenceAckPath(portIdentifier, channelIdentifier), 1)

    return channelCapability
}

The chanOpenAck is called by the handshake-originating module to acknowledge the acceptance of the initial request by the counterparty module on the other chain.

function chanOpenAck(
  portIdentifier: Identifier,
  channelIdentifier: Identifier,
  counterpartyChannelIdentifier: Identifier,
  counterpartyVersion: string,
  proofTry: CommitmentProof,
  proofHeight: Height) {
    channel = provableStore.get(channelPath(portIdentifier, channelIdentifier))
    abortTransactionUnless(channel.state === INIT)
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(portIdentifier, channelIdentifier), capability))
    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)
    expected = ChannelEnd{TRYOPEN, channel.order, portIdentifier,
                          channelIdentifier, [connection.counterpartyConnectionIdentifier], counterpartyVersion}
    abortTransactionUnless(connection.verifyChannelState(
      proofHeight,
      proofTry,
      channel.counterpartyPortIdentifier,
      counterpartyChannelIdentifier,
      expected
    ))
    channel.state = OPEN
    channel.version = counterpartyVersion
    channel.counterpartyChannelIdentifier = counterpartyChannelIdentifier
    provableStore.set(channelPath(portIdentifier, channelIdentifier), channel)
}

The chanOpenConfirm function is called by the handshake-accepting module to acknowledge the acknowledgement of the handshake-originating module on the other chain and finish the channel opening handshake.

function chanOpenConfirm(
  portIdentifier: Identifier,
  channelIdentifier: Identifier,
  proofAck: CommitmentProof,
  proofHeight: Height) {
    channel = provableStore.get(channelPath(portIdentifier, channelIdentifier))
    abortTransactionUnless(channel !== null)
    abortTransactionUnless(channel.state === TRYOPEN)
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(portIdentifier, channelIdentifier), capability))
    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)
    expected = ChannelEnd{OPEN, channel.order, portIdentifier,
                          channelIdentifier, [connection.counterpartyConnectionIdentifier], channel.version}
    abortTransactionUnless(connection.verifyChannelState(
      proofHeight,
      proofAck,
      channel.counterpartyPortIdentifier,
      channel.counterpartyChannelIdentifier,
      expected
    ))
    channel.state = OPEN
    provableStore.set(channelPath(portIdentifier, channelIdentifier), channel)
}

Closing handshake

The chanCloseInit function is called by either module to close their end of the channel. Once closed, channels cannot be reopened.

Calling modules MAY atomically execute appropriate application logic in conjunction with calling chanCloseInit.

Any in-flight packets can be timed-out as soon as a channel is closed.

function chanCloseInit(
  portIdentifier: Identifier,
  channelIdentifier: Identifier) {
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(portIdentifier, channelIdentifier), capability))
    channel = provableStore.get(channelPath(portIdentifier, channelIdentifier))
    abortTransactionUnless(channel !== null)
    abortTransactionUnless(channel.state !== CLOSED)
    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)
    channel.state = CLOSED
    provableStore.set(channelPath(portIdentifier, channelIdentifier), channel)
}

The chanCloseConfirm function is called by the counterparty module to close their end of the channel, since the other end has been closed.

Calling modules MAY atomically execute appropriate application logic in conjunction with calling chanCloseConfirm.

Once closed, channels cannot be reopened and identifiers cannot be reused. Identifier reuse is prevented because we want to prevent potential replay of previously sent packets. The replay problem is analogous to using sequence numbers with signed messages, except where the light client algorithm "signs" the messages (IBC packets), and the replay prevention sequence is the combination of port identifier, channel identifier, and packet sequence - hence we cannot allow the same port identifier & channel identifier to be reused again with a sequence reset to zero, since this might allow packets to be replayed. It would be possible to safely reuse identifiers if timeouts of a particular maximum height/time were mandated & tracked, and future specification versions may incorporate this feature.

function chanCloseConfirm(
  portIdentifier: Identifier,
  channelIdentifier: Identifier,
  proofInit: CommitmentProof,
  proofHeight: Height) {
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(portIdentifier, channelIdentifier), capability))
    channel = provableStore.get(channelPath(portIdentifier, channelIdentifier))
    abortTransactionUnless(channel !== null)
    abortTransactionUnless(channel.state !== CLOSED)
    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)
    expected = ChannelEnd{CLOSED, channel.order, portIdentifier,
                          channelIdentifier, [connection.counterpartyConnectionIdentifier], channel.version}
    abortTransactionUnless(connection.verifyChannelState(
      proofHeight,
      proofInit,
      channel.counterpartyPortIdentifier,
      channel.counterpartyChannelIdentifier,
      expected
    ))
    channel.state = CLOSED
    provableStore.set(channelPath(portIdentifier, channelIdentifier), channel)
}

Packet flow & handling

A day in the life of a packet

The following sequence of steps must occur for a packet to be sent from module 1 on machine A to module 2 on machine B, starting from scratch.

The module can interface with the IBC handler through ICS 25 or ICS 26.

1. Initial client & port setup, in any order
   1. Client created on A for B (see ICS 2)
   2. Client created on B for A (see ICS 2)
   3. Module 1 binds to a port (see ICS 5)
   4. Module 2 binds to a port (see ICS 5), which is communicated out-of-band to module 1
2. Establishment of a connection & channel, optimistic send, in order
   1. Connection opening handshake started from A to B by module 1 (see ICS 3)
   2. Channel opening handshake started from 1 to 2 using the newly created connection (this ICS)
   3. Packet sent over the newly created channel from 1 to 2 (this ICS)
3. Successful completion of handshakes (if either handshake fails, the connection/channel can be closed & the packet timed-out)
   1. Connection opening handshake completes successfully (see ICS 3) (this will require participation of a relayer process)
   2. Channel opening handshake completes successfully (this ICS) (this will require participation of a relayer process)
4. Packet confirmation on machine B, module 2 (or packet timeout if the timeout height has passed) (this will require participation of a relayer process)
5. Acknowledgement (possibly) relayed back from module 2 on machine B to module 1 on machine A

Represented spatially, packet transit between two machines can be rendered as follows:

Sending packets

The sendPacket function is called by a module in order to send data (in the form of an IBC packet) on a channel end owned by the calling module.

Calling modules MUST execute application logic atomically in conjunction with calling sendPacket.

The IBC handler performs the following steps in order:

• Checks that the channel & connection are open to send packets
• Checks that the calling module owns the sending port (see ICS 5)
• Checks that the timeout height specified has not already passed on the destination chain
• Increments the send sequence counter associated with the channel
• Stores a constant-size commitment to the packet data & packet timeout

Note that the full packet is not stored in the state of the chain - merely a short hash-commitment to the data & timeout value. The packet data can be calculated from the transaction execution and possibly returned as log output which relayers can index.

function sendPacket(
  capability: CapabilityKey,
  sourcePort: Identifier,
  sourceChannel: Identifier,
  timeoutHeight: Height,
  timeoutTimestamp: uint64,
  data: bytes) {
    channel = provableStore.get(channelPath(sourcePort, sourceChannel))

    // check that the channel & connection are open to send packets; 
    // note: optimistic sends are permitted once the handshake has started
    abortTransactionUnless(channel !== null)
    abortTransactionUnless(channel.state !== CLOSED)
    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    abortTransactionUnless(connection !== null)

    // check if the calling module owns the sending port
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(sourcePort, sourceChannel), capability))
    
    // check that the timeout height hasn't already passed in the local client tracking the receiving chain
    latestClientHeight = provableStore.get(clientPath(connection.clientIdentifier)).latestClientHeight()
    abortTransactionUnless(packet.timeoutHeight === 0 || latestClientHeight < packet.timeoutHeight)

    // increment the send sequence counter
    sequence = provableStore.get(nextSequenceSendPath(packet.sourcePort, packet.sourceChannel))
    provableStore.set(nextSequenceSendPath(packet.sourcePort, packet.sourceChannel), sequence+1)

    // store commitment to the packet data & packet timeout
    provableStore.set(
      packetCommitmentPath(sourcePort, sourceChannel, sequence),
      hash(data, timeoutHeight, timeoutTimestamp)
    )

    // log that a packet can be safely sent
    emitLogEntry("sendPacket", {sequence: sequence, data: data, timeoutHeight: timeoutHeight, timeoutTimestamp: timeoutTimestamp})
}

Receiving packets

The recvPacket function is called by a module in order to receive an IBC packet sent on the corresponding channel end on the counterparty chain.

Atomically in conjunction with calling recvPacket, calling modules MUST either execute application logic or queue the packet for future execution.

The IBC handler performs the following steps in order:

• Checks that the channel & connection are open to receive packets
• Checks that the calling module owns the receiving port
• Checks that the packet metadata matches the channel & connection information
• Checks that the packet sequence is the next sequence the channel end expects to receive (for ordered and ordered_allow_timeout channels)
• Checks that the timeout height and timestamp have not yet passed
• Checks the inclusion proof of packet data commitment in the outgoing chain's state
• Sets a store path to indicate that the packet has been received (unordered channels only)
• Increments the packet receive sequence associated with the channel end (ordered and ordered_allow_timeout channels only)

We pass the address of the relayer that signed and submitted the packet to enable a module to optionally provide some rewards. This provides a foundation for fee payment, but can be used for other techniques as well (like calculating a leaderboard).

function recvPacket(
  packet: OpaquePacket,
  proof: CommitmentProof,
  proofHeight: Height,
  relayer: string): Packet {

    channel = provableStore.get(channelPath(packet.destPort, packet.destChannel))
    abortTransactionUnless(channel !== null)
    abortTransactionUnless(channel.state === OPEN)
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(packet.destPort, packet.destChannel), capability))
    abortTransactionUnless(packet.sourcePort === channel.counterpartyPortIdentifier)
    abortTransactionUnless(packet.sourceChannel === channel.counterpartyChannelIdentifier)

    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)

    abortTransactionUnless(connection.verifyPacketData(
      proofHeight,
      proof,
      packet.sourcePort,
      packet.sourceChannel,
      packet.sequence,
      concat(packet.data, packet.timeoutHeight, packet.timeoutTimestamp)
    ))

    // do sequence check before any state changes
    if channel.order == ORDERED || channel.order == ORDERED_ALLOW_TIMEOUT {
        nextSequenceRecv = provableStore.get(nextSequenceRecvPath(packet.destPort, packet.destChannel))
        abortTransactionUnless(packet.sequence === nextSequenceRecv)
    }

    switch channel.order {
      case ORDERED:
      case UNORDERED:
        abortTransactionUnless(packet.timeoutHeight === 0 || getConsensusHeight() < packet.timeoutHeight)
        abortTransactionUnless(packet.timeoutTimestamp === 0 || currentTimestamp() < packet.timeoutTimestamp)
        break;

      case ORDERED_ALLOW_TIMEOUT:
        // for ORDERED_ALLOW_TIMEOUT, we do not abort on timeout
        // instead increment next sequence recv and write the sentinel timeout value in packet receipt
        // then return
        if (getConsensusHeight() >= packet.timeoutHeight && packet.timeoutHeight != 0) || (currentTimestamp() >= packet.timeoutTimestamp && packet.timeoutTimestamp != 0) {
          nextSequenceRecv = nextSequenceRecv + 1
          provableStore.set(nextSequenceRecvPath(packet.destPort, packet.destChannel), nextSequenceRecv)
          provableStore.set(
          packetReceiptPath(packet.destPort, packet.destChannel, packet.sequence),
            TIMEOUT_RECEIPT
          )
        }
        return;

      default:
        // unsupported channel type
        abortTransactionUnless(false)
    }

    // all assertions passed (except sequence check), we can alter state

    switch channel.order {
      case ORDERED:
      case ORDERED_ALLOW_TIMEOUT:
        nextSequenceRecv = nextSequenceRecv + 1
        provableStore.set(nextSequenceRecvPath(packet.destPort, packet.destChannel), nextSequenceRecv)
        break;

      case UNORDERED:
        // for unordered channels we must set the receipt so it can be verified on the other side
        // this receipt does not contain any data, since the packet has not yet been processed
        // it's the sentinel success receipt: []byte{0x01}
        abortTransactionUnless(provableStore.get(packetReceiptPath(packet.destPort, packet.destChannel, packet.sequence) === null))
        provableStore.set(
          packetReceiptPath(packet.destPort, packet.destChannel, packet.sequence),
          SUCCESFUL_RECEIPT
        )
      break;
    }
    
    // log that a packet has been received
    emitLogEntry("recvPacket", {sequence: packet.sequence, timeoutHeight: packet.timeoutHeight, port: packet.destPort, channel: packet.destChannel,
                                timeoutTimestamp: packet.timeoutTimestamp, data: packet.data})

    // return transparent packet
    return packet
}

Writing acknowledgements

The writeAcknowledgement function is called by a module in order to write data which resulted from processing an IBC packet that the sending chain can then verify, a sort of "execution receipt" or "RPC call response".

Calling modules MUST execute application logic atomically in conjunction with calling writeAcknowledgement.

This is an asynchronous acknowledgement, the contents of which do not need to be determined when the packet is received, only when processing is complete. In the synchronous case, writeAcknowledgement can be called in the same transaction (atomically) with recvPacket.

Acknowledging packets is not required; however, if an ordered channel uses acknowledgements, either all or no packets must be acknowledged (since the acknowledgements are processed in order). Note that if packets are not acknowledged, packet commitments cannot be deleted on the source chain. Future versions of IBC may include ways for modules to specify whether or not they will be acknowledging packets in order to allow for cleanup.

writeAcknowledgement does not check if the packet being acknowledged was actually received, because this would result in proofs being verified twice for acknowledged packets. This aspect of correctness is the responsibility of the calling module. The calling module MUST only call writeAcknowledgement with a packet previously received from recvPacket.

The IBC handler performs the following steps in order:

• Checks that an acknowledgement for this packet has not yet been written
• Sets the opaque acknowledgement value at a store path unique to the packet

function writeAcknowledgement(
  packet: Packet,
  acknowledgement: bytes) {

    // cannot already have written the acknowledgement
    abortTransactionUnless(provableStore.get(packetAcknowledgementPath(packet.destPort, packet.destChannel, packet.sequence) === null))

    // write the acknowledgement
    provableStore.set(
      packetAcknowledgementPath(packet.destPort, packet.destChannel, packet.sequence),
      hash(acknowledgement)
    )

    // log that a packet has been acknowledged
    emitLogEntry("writeAcknowledgement", {sequence: packet.sequence, timeoutHeight: packet.timeoutHeight, port: packet.destPort, channel: packet.destChannel,
                                timeoutTimestamp: packet.timeoutTimestamp, data: packet.data, acknowledgement})
}

Processing acknowledgements

The acknowledgePacket function is called by a module to process the acknowledgement of a packet previously sent by the calling module on a channel to a counterparty module on the counterparty chain. acknowledgePacket also cleans up the packet commitment, which is no longer necessary since the packet has been received and acted upon.

Calling modules MAY atomically execute appropriate application acknowledgement-handling logic in conjunction with calling acknowledgePacket.

We pass the relayer address just as in Receiving packets to allow for possible incentivization here as well.

function acknowledgePacket(
  packet: OpaquePacket,
  acknowledgement: bytes,
  proof: CommitmentProof,
  proofHeight: Height,
  relayer: string): Packet {

    // abort transaction unless that channel is open, calling module owns the associated port, and the packet fields match
    channel = provableStore.get(channelPath(packet.sourcePort, packet.sourceChannel))
    abortTransactionUnless(channel !== null)
    abortTransactionUnless(channel.state === OPEN)
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(packet.sourcePort, packet.sourceChannel), capability))
    abortTransactionUnless(packet.destPort === channel.counterpartyPortIdentifier)
    abortTransactionUnless(packet.destChannel === channel.counterpartyChannelIdentifier)

    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    abortTransactionUnless(connection !== null)
    abortTransactionUnless(connection.state === OPEN)

    // verify we sent the packet and haven't cleared it out yet
    abortTransactionUnless(provableStore.get(packetCommitmentPath(packet.sourcePort, packet.sourceChannel, packet.sequence))
           === hash(packet.data, packet.timeoutHeight, packet.timeoutTimestamp))

    // abort transaction unless correct acknowledgement on counterparty chain
    abortTransactionUnless(connection.verifyPacketAcknowledgement(
      proofHeight,
      proof,
      packet.destPort,
      packet.destChannel,
      packet.sequence,
      acknowledgement
    ))

    // abort transaction unless acknowledgement is processed in order
    if (channel.order === ORDERED || channel.order == ORDERED_ALLOW_TIMEOUT) {
      nextSequenceAck = provableStore.get(nextSequenceAckPath(packet.sourcePort, packet.sourceChannel))
      abortTransactionUnless(packet.sequence === nextSequenceAck)
      nextSequenceAck = nextSequenceAck + 1
      provableStore.set(nextSequenceAckPath(packet.sourcePort, packet.sourceChannel), nextSequenceAck)
    }

    // all assertions passed, we can alter state

    // delete our commitment so we can't "acknowledge" again
    provableStore.delete(packetCommitmentPath(packet.sourcePort, packet.sourceChannel, packet.sequence))

    // return transparent packet
    return packet
}

Acknowledgement Envelope

The acknowledgement returned from the remote chain is defined as arbitrary bytes in the IBC protocol. This data may either encode a successful execution or a failure (anything besides a timeout). There is no generic way to distinguish the two cases, which requires that any client-side packet visualiser understands every app-specific protocol in order to distinguish the case of successful or failed relay. In order to reduce this issue, we offer an additional specification for acknowledgement formats, which SHOULD be used by the app-specific protocols.

message Acknowledgement {
  oneof response {
    bytes result = 21;
    string error = 22;
  }
}

If an application uses a different format for acknowledgement bytes, it MUST not deserialise to a valid protobuf message of this format. Note that all packets contain exactly one non-empty field, and it must be result or error. The field numbers 21 and 22 were explicitly chosen to avoid accidental conflicts with other protobuf message formats used for acknowledgements. The first byte of any message with this format will be the non-ASCII values 0xaa (result) or 0xb2 (error).

Timeouts

Application semantics may require some timeout: an upper limit to how long the chain will wait for a transaction to be processed before considering it an error. Since the two chains have different local clocks, this is an obvious attack vector for a double spend - an attacker may delay the relay of the receipt or wait to send the packet until right after the timeout - so applications cannot safely implement naive timeout logic themselves.

Note that in order to avoid any possible "double-spend" attacks, the timeout algorithm requires that the destination chain is running and reachable. One can prove nothing in a complete network partition, and must wait to connect; the timeout must be proven on the recipient chain, not simply the absence of a response on the sending chain.

Sending end

The timeoutPacket function is called by a module which originally attempted to send a packet to a counterparty module, where the timeout height or timeout timestamp has passed on the counterparty chain without the packet being committed, to prove that the packet can no longer be executed and to allow the calling module to safely perform appropriate state transitions.

Calling modules MAY atomically execute appropriate application timeout-handling logic in conjunction with calling timeoutPacket.

In the case of an ordered channel, timeoutPacket checks the recvSequence of the receiving channel end and closes the channel if a packet has timed out.

In the case of an unordered channel, timeoutPacket checks the absence of the receipt key (which will have been written if the packet was received). Unordered channels are expected to continue in the face of timed-out packets.

If relations are enforced between timeout heights of subsequent packets, safe bulk timeouts of all packets prior to a timed-out packet can be performed. This specification omits details for now.

Since we allow optimistic sending of packets (i.e. sending a packet before a channel opens), we must also allow optimistic timing out of packets. With optimistic sends, the packet may be sent on a channel that eventually opens or a channel that will never open. If the channel does open after the packet has timed out, then the packet will never be received on the counterparty so we can safely timeout optimistically. If the channel never opens, then we MUST timeout optimistically so that any state changes made during the optimistic send by the application can be safely reverted.

We pass the relayer address just as in Receiving packets to allow for possible incentivization here as well.

function timeoutPacket(
  packet: OpaquePacket,
  proof: CommitmentProof,
  proofHeight: Height,
  nextSequenceRecv: Maybe<uint64>,
  relayer: string): Packet {

    channel = provableStore.get(channelPath(packet.sourcePort, packet.sourceChannel))
    abortTransactionUnless(channel !== null)

    abortTransactionUnless(authenticateCapability(channelCapabilityPath(packet.sourcePort, packet.sourceChannel), capability))
    abortTransactionUnless(packet.destChannel === channel.counterpartyChannelIdentifier)

    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    // note: the connection may have been closed
    abortTransactionUnless(packet.destPort === channel.counterpartyPortIdentifier)

    // check that timeout height or timeout timestamp has passed on the other end
    abortTransactionUnless(
      (packet.timeoutHeight > 0 && proofHeight >= packet.timeoutHeight) ||
      (packet.timeoutTimestamp > 0 && connection.getTimestampAtHeight(proofHeight) > packet.timeoutTimestamp))

    // verify we actually sent this packet, check the store
    abortTransactionUnless(provableStore.get(packetCommitmentPath(packet.sourcePort, packet.sourceChannel, packet.sequence))
           === hash(packet.data, packet.timeoutHeight, packet.timeoutTimestamp))

    switch channel.order {
      case ORDERED:
        // ordered channel: check that packet has not been received
        // only allow timeout on next sequence so all sequences before the timed out packet are processed (received/timed out)
        // before this packet times out
        abortTransactionUnless(nextSequenceRecv == packet.sequence)
        // ordered channel: check that the recv sequence is as claimed
        abortTransactionUnless(connection.verifyNextSequenceRecv(
          proofHeight,
          proof,
          packet.destPort,
          packet.destChannel,
          nextSequenceRecv
        ))
        break;

      case UNORDERED:
        // unordered channel: verify absence of receipt at packet index
        abortTransactionUnless(connection.verifyPacketReceiptAbsence(
          proofHeight,
          proof,
          packet.destPort,
          packet.destChannel,
          packet.sequence
        ))
        break;

      // NOTE: For ORDERED_ALLOW_TIMEOUT, the relayer must first attempt the receive on the destination chain
      // before the timeout receipt can be written and subsequently proven on the sender chain in timeoutPacket
      case ORDERED_ALLOW_TIMEOUT:
        // ordered channel: check that packet has not been received
        // only allow timeout on next sequence so all sequences before the timed out packet are processed (received/timed out)
        // before this packet times out
        abortTransactionUnless(nextSequenceRecv == packet.sequence)
        abortTransactionUnless(connection.verifyPacketReceipt(
          proofHeight,
          proof,
          packet.destPort,
          packet.destChannel,
          packet.sequence
          TIMEOUT_RECEIPT,
        ))
        break;

      default:
        // unsupported channel type
        abortTransactionUnless(true)
    } 

    // all assertions passed, we can alter state

    // delete our commitment
    provableStore.delete(packetCommitmentPath(packet.sourcePort, packet.sourceChannel, packet.sequence))

    // only close on strictly ORDERED channels
    if channel.order === ORDERED {
      // ordered channel: close the channel
      channel.state = CLOSED
      provableStore.set(channelPath(packet.sourcePort, packet.sourceChannel), channel)
    }
    // on ORDERED_ALLOW_TIMEOUT, increment NextSequenceAck so that next packet can be acknowledged after this packet timed out.
    if channel.order === ORDERED_ALLOW_TIMEOUT {
      nextSequenceAck = nextSequenceAck + 1
      provableStore.set(nextSequenceAckPath(packet.srcPort, packet.srcChannel), nextSequenceAck)
    }

    // return transparent packet
    return packet
}

Timing-out on close

The timeoutOnClose function is called by a module in order to prove that the channel to which an unreceived packet was addressed has been closed, so the packet will never be received (even if the timeoutHeight or timeoutTimestamp has not yet been reached).

Calling modules MAY atomically execute appropriate application timeout-handling logic in conjunction with calling timeoutOnClose.

We pass the relayer address just as in Receiving packets to allow for possible incentivization here as well.

function timeoutOnClose(
  packet: Packet,
  proof: CommitmentProof,
  proofClosed: CommitmentProof,
  proofHeight: Height,
  nextSequenceRecv: Maybe<uint64>,
  relayer: string): Packet {

    channel = provableStore.get(channelPath(packet.sourcePort, packet.sourceChannel))
    // note: the channel may have been closed
    abortTransactionUnless(authenticateCapability(channelCapabilityPath(packet.sourcePort, packet.sourceChannel), capability))
    abortTransactionUnless(packet.destChannel === channel.counterpartyChannelIdentifier)

    connection = provableStore.get(connectionPath(channel.connectionHops[0]))
    // note: the connection may have been closed
    abortTransactionUnless(packet.destPort === channel.counterpartyPortIdentifier)

    // verify we actually sent this packet, check the store
    abortTransactionUnless(provableStore.get(packetCommitmentPath(packet.sourcePort, packet.sourceChannel, packet.sequence))
           === hash(packet.data, packet.timeoutHeight, packet.timeoutTimestamp))

    // check that the opposing channel end has closed
    expected = ChannelEnd{CLOSED, channel.order, channel.portIdentifier,
                          channel.channelIdentifier, channel.connectionHops.reverse(), channel.version}
    abortTransactionUnless(connection.verifyChannelState(
      proofHeight,
      proofClosed,
      channel.counterpartyPortIdentifier,
      channel.counterpartyChannelIdentifier,
      expected
    ))

    if channel.order === ORDERED || channel.order == ORDERED_ALLOW_TIMEOUT {
      // ordered channel: check that the recv sequence is as claimed
      abortTransactionUnless(connection.verifyNextSequenceRecv(
        proofHeight,
        proof,
        packet.destPort,
        packet.destChannel,
        nextSequenceRecv
      ))
      // ordered channel: check that packet has not been received
      abortTransactionUnless(nextSequenceRecv <= packet.sequence)
    } else
      // unordered channel: verify absence of receipt at packet index
      abortTransactionUnless(connection.verifyPacketReceiptAbsence(
        proofHeight,
        proof,
        packet.destPort,
        packet.destChannel,
        packet.sequence
      ))

    // all assertions passed, we can alter state

    // delete our commitment
    provableStore.delete(packetCommitmentPath(packet.sourcePort, packet.sourceChannel, packet.sequence))

    // return transparent packet
    return packet
}

Cleaning up state

Packets must be acknowledged in order to be cleaned-up.

Reasoning about race conditions

Simultaneous handshake attempts

If two machines simultaneously initiate channel opening handshakes with each other, attempting to use the same identifiers, both will fail and new identifiers must be used.

Identifier allocation

There is an unavoidable race condition on identifier allocation on the destination chain. Modules would be well-advised to utilise pseudo-random, non-valuable identifiers. Managing to claim the identifier that another module wishes to use, however, while annoying, cannot man-in-the-middle a handshake since the receiving module must already own the port to which the handshake was targeted.

Timeouts / packet confirmation

There is no race condition between a packet timeout and packet confirmation, as the packet will either have passed the timeout height prior to receipt or not.

Man-in-the-middle attacks during handshakes

Verification of cross-chain state prevents man-in-the-middle attacks for both connection handshakes & channel handshakes since all information (source, destination client, channel, etc.) is known by the module which starts the handshake and confirmed prior to handshake completion.

Connection / channel closure with in-flight packets

If a connection or channel is closed while packets are in-flight, the packets can no longer be received on the destination chain and can be timed-out on the source chain.

Querying channels

Channels can be queried with queryChannel:

function queryChannel(connId: Identifier, chanId: Identifier): ChannelEnd | void {
    return provableStore.get(channelPath(connId, chanId))
}

Properties & Invariants

• The unique combinations of channel & port identifiers are first-come-first-serve: once a pair has been allocated, only the modules owning the ports in question can send or receive on that channel.
• Packets are delivered exactly once, assuming that the chains are live within the timeout window, and in case of timeout can be timed-out exactly once on the sending chain.
• The channel handshake cannot be man-in-the-middle attacked by another module on either blockchain or another blockchain's IBC handler.

Backwards Compatibility

Not applicable.

Forwards Compatibility

Data structures & encoding can be versioned at the connection or channel level. Channel logic is completely agnostic to packet data formats, which can be changed by the modules any way they like at any time.

Example Implementation

Coming soon.

Other Implementations

Coming soon.

History

Jun 5, 2019 - Draft submitted

Jul 4, 2019 - Modifications for unordered channels & acknowledgements

Jul 16, 2019 - Alterations for multi-hop routing future compatibility

Jul 29, 2019 - Revisions to handle timeouts after connection closure

Aug 13, 2019 - Various edits

Aug 25, 2019 - Cleanup

Jan 10, 2022 - Add ORDERED_ALLOW_TIMEOUT channel type and appropriate logic

Copyright

All content herein is licensed under Apache 2.0.