# Network API
Suitably-empowered code inside a vat can access a "network API" that works vaguely like the BSD socket API. This code can:
- Open a listening port on various networking stacks.
- Initiate connections to remote ports.
- Send and receive data over these connections
- Finally close the connection and/or ports.
The type of connection is limited by the host in which the vat is running. Chain-based machines must operate in a platonic realm of replicated consensus, so their network options are limited to protocols like IBC, which allow one gestalt chain to talk to other chain-like entities. Each such entity is defined by an evolving set of consensus rules, which typically include a current set of validator public keys and a particular history of hashed block identifiers.
[CAVEAT: IBC uses "Connection" to mean a chain-to-chain "hop", and "Channel" to mean a Port-to-Port pathway through a series of hops. https://github.com/cosmos/ics/blob/master/ibc/1_IBC_TERMINOLOGY.md#connection . This is unfortunate, because IBC "Channels" correspond most precisely to TCP "connections", and most discussions of network APIs (including this one, below) will talk about "connections" extensively.
For now, our IBC implementation can only use pre-established hops, and provides no means for user-level code to create new hops (IBC Connections) at runtime. But user-level code can create new IBC Channels at any time. The terminology confusion will be most obvious in the section on "Accepting an Inbound Connection", where the user code is really accepting an inbound IBC Channel.]
A channel via these IBC hops will terminate in IBC-aware code on either end.
These endpoints might be traditional (static) IBC handlers (such as an ICS-20
token transfer module), or dynamic IBC handlers (e.g. running in a SwingSet vat).
SwingSet vat code that wants to speak to vat code in a different SwingSet machine
would not use the IBC connection directly: instead it would simply perform normal
eventual-send operations (target~.foo(args)
) and let the "CapTP" promise-pipelining
layer handle the details. But vat code which wants to speak to an ICS-20 handler
in some other chain would need to use this layer.
Vats which live inside a solo machine are able to use traditional networking layers, like TCP, HTTP, and WebSockets. This enables them to communicate with e.g. browser-side UI frontends that use WebSockets to query the vat for order status. These connections do not have to follow normal ocap rules, so vat code which accept them must provide their own authentication and access control protections.
Solo machines may be able to talk to chains and vice versa using specialized protocols. This will be used by CapTP to provide ocap messaging between these heterogeneous machine types.
# The agoric-sdk User Local Port
Each user of the Agoric testnet gets a few personal IBC listening ports. You
can access these Port
objects in the home.ibcport
array, and you can learn
their local address by calling something like home.ibcport[0]~.getLocalAddress()
,
which will give you a local address like /ibc-port/portbvmnfb
.
This is currently the only way for user code to get an IBC Port
, though non-IBC
ports can be allocated using the local home.network
object. This is an advanced
use case, to be documented later.
# Connecting to a Remote Port
To establish a connection, you must start with a local Port
object, and you must
know the name of the remote endpoint. The remote endpoint will have a name like
/ibc-hop/$HOPNAME/ibc-port/$PORTNAME/ordered/$VERSION
(where ibc-hop
, ibc-port
and ordered
are literal strings, spelled just like that, but $HOPNAME
, $PORTNAME
,
and $VERSION
are placeholders for arbitrary values that will vary from one endpoint
to another).
You must also prepare a ConnectionHandler
object to manage the connection you're
about to create. This has a number of methods which will be called when the things
happen to the connection, including packets arriving. This is described below.
Then you will call the connect()
method on your local Port
. This will return
a Promise
that will fire with a new Connection
object, on which you can send
data. Your ConnectionHandler
will be notified about the new channel, and will
receive inbound data from the other side.
home.ibcport[0]~.connect(endpoint, connectionHandler)
.then(conn => doSomethingWithConnection(conn));
# Opening a Listening Port and Accepting an Inbound Connection
The other side of connect()
is a "listening port". These ports are waiting for
inbound connections to be established.
To get a listening port, you need a NetworkInterface
object (such as the one on
your ag-solo
under home.network
) and ask it to bind()
to an endpoint. You
can either provide a specific port name, or allow the API to allocate a random
one for you. The endpoint specifies the type of connection that this port will
be able to accept (IBC, TCP, etc), and some properties of that connection. bind()
uses a "multiaddress" to encode this information.
// ask for a random allocation - ends with a slash
home.network~.bind('/ibc-port/`)
.then(port => usePort(port));
//
// or ask for a specific port name
home.network~.bind('/ibc-port/my-cool-port-name`)
.then(port => usePort(port));
IBC has named "hops" (what they call "Connections" in the IBC spec) which each
carry data between two specific chains. These hops are different from the
connections described in this document. When you bind a port like /ibc-port/$PORT
without specifying the "hop", any IBC chain can initiate a connection to this port.
You can ask the Port
object this returns for its local address, which is especially
useful if you had asked for a random allocation (since otherwise you have no way
to know what address you got):
port~.getLocalAddress().then(localAddress => useIt(localAddress))
home.ibcport[0]~.addListener()
Once the port is bound, you must call addListener
to mark it as ready for inbound
connections. You must provide this with a ListenHandler
object, which has methods
to react to listening events. As with ConnectionHandler
, these methods are all
optional.
onListen(port, handler)
: called when the port starts listeningonAccept(port, remote, handler)
: called when a new channel has been acceptedonError(port, rejection, handler)
: called if the port is no longer able to accept channels, such as if the Connection to the remote chain has failed, perhaps because a consensus failure was observedonRemove(port, handler)
: called when theListenHandler
is being removed
Once your ChannelHandler
is prepared, call addListener
:
port.addListener(handler).then(() => console.log('listener is active'))
Of all the methods, onAccept
is the interesting one. It is called with a remote
endpoint, which tells you the address of the Port
at the other end, where someone
else called .connect
. You can use this to decide if you want to accept the
connection, or what sort of authority to exercise in response to messages arriving
therein.
If you choose to accept, your onAccept
method must return a Promise
that fires
with a ConnectionHandler
. This will be used just like the one you would pass into
connect()
. To decline, throw an error.
# Sending Data
The Networking API (at least for IBC) provides a "record pipe", in which each packet is transmitted intact over the network, requiring no additional framing to distinguish where one packet ends and the next one begins. This is in contrast to the "byte pipe" provided by a TCP connection, where you must typically prepend length headers to establish framing boundaries.
Once you have a Connection
object, you send data by calling its send
method:
connection.send('data');
send
actually returns a Promise (for more Bytes
), which contains the ACK data
for this message. For IBC, if anything but an empty ''
ACK is needed, you must
resolve the onReceive
returned Promise within the same block as the received
packet message (i.e. it must not depend upon further input from the chain in order
to resolve). This restriction may be enforced differently or lifted for other
network implementations.
NOTE: The type of this data is currently a string. Ideally we would also accept
Node.js Buffer
objects, or Javascript ArrayBuffer
and TypedArray
objects,
but unfortunately neither can be serialized by our current inter-vat marshalling code.
# Receiving Data: The ConnectionHandler
You must provide each open connection with a ConnectionHandler
object, where
you write methods that will be called when various things happen to the connection.
You can share a single handler object between multiple connections, if you like,
or you can make a separate one for each.
You can omit any of the methods and those events will simply be ignored. All these
methods include the Connection object as the first argument, and the ConnectionHandler
itself as the last argument, which might help if you want to share a common handler
function among multiple connections.
onOpen(connection, handler)
: this is called when the connection is established, which tells you that the remote end has successfully accepted the connection requestonReceive(connection, packetBytes, handler)
: this is called each time the remote end sends a packet full of dataonClose(connection, reason, handler)
: this is called when the connection is closed, either because one side wanted it to close, or because an error occurred
The reason
in onclose
is optional, as in it may be undefined
.
onReceive
is the most important method. Each time the remote end sends a packet,
your onReceive
method will be called with the data inside that packet (currently
as a String, but ideally as an ArrayBuffer with a custom .toString()
method with
an optional encoding
argument (default 'latin1'
), so that it can contain
arbitrary bytes).
The return value of onReceive
is nominally a Promise for the ACK data of the
message (and should thus appear as the eventual resolution of the Promise returned
by connection.send()
on the other side). However the ACK data can only appear in
the block which includes the transaction that delivered the message, so there is a
limited time window during which this data can be successfully delivered, and there
is no guarantee that onReceive
will return a Promise that resolves in time. If
the promise does not resolve in time, the implementation will automatically send
an empty ''
ACK. Because ACKs with data in them are difficult for your handler
to get right, it is better to avoid them, where possible.
# Closing the Connection
When a given Connection ceases to be useful, you should close it:
connection.close();
This initiates a shutdown. The ConnectionHandler
on both sides will eventually
see their onClose()
methods be called, with a reason
. It allows them to
distinguish an intentional onClose()
(reason
is undefined
) from some error
condition.
# Removing a Listener
When you no longer wish to receive connections on a port, you can remove the listener:
port.removeListener(handler).then(() => console.log('removed'));
You must provide the handler you added, to enable the future ability to have multiple listeners on the same port.
Note that if you want to listen on this port again, you can just
call port.addListener(...)
, as before. If you want to start a new connection,
you can always call port.connect(...)
.
# Closing the Port Entirely
Removing a listener doesn't release the port address to make it available
for other bind
requests. You can call:
port.revoke();
to completely deallocate the port, remove all listeners, close all pending connections, and release its address.
CAUTION: Be aware that if you call home.ibcport[0]~.revoke()
, it will be
useless for new .connect
or .addListener
attempts. You will need to provision
a new Agoric client via https://testnet.agoric.com/ to obtain a new setup with a
functioning home.ibcport
.