This is an experiment in implementing automatic peer-to-peer link setup in wireguard by:
- Automatically configuring IPv6 link-local ips for each peer, derived by hashing the peer's public key
- Using that to automatically share information about available peers and their endpoints
- Using that to try to automatically setup direct connections between peers (when possible) instead of routing through a central "server".
wirelink
will read basic configuration from:
- A config file
/etc/wireguard/wirelink.<interface>.json
- Some other extensions such as
.yaml
will be accepted too - Some settings can only be set in the config file for now
- Some other extensions such as
- Environment variables of the form
WIRELINK_<setting>
- Command line args (see
--help
)
Two systemd template units are provided:
[email protected]
, when enabled for an interface, will bind tightly to
[email protected]
. If you are using wg-quick
, this is the recommended
method. If you have wg-quick@wg0
enabled, then to activate the wirelink pair,
run systemctl enable wl-quick@wg0 && systemctl start wl-quick@wg0
. In the
future, the wirelink
instance will automatically start & stop along with the
wg-quick
instance.
[email protected]
is provided for more manual configurations, such as if you
configure your wireguard interface in /etc/network/interfaces
. Enable and
start it with e.g. systemctl enable wirelink@wg0 && systemctl start wirelink@wg0
similar to how wl-quick@
works. If the wireguard interface goes
down, the service will fail, but it is configured to auto-restart periodically
until the link comes back up.
Peers produce a list of local "facts" based on information from the wireguard device and the local network interfaces. Facts have:
- A subject
- Who is the fact about
- An attribute
- What attribute of the subject does the fact describe
- A value
- What is the value of that attribute
- A TTL
- For how long should this fact be considered valid
For now subjects are always a peer's public key. Attributes are commonly the peer's allowed ip value(s) and possible endpoints. Peers share endpoints of other peers if they have a live connection to that peer. Peers also share all their local IP addresses and their listening port in case they are on a public IP or other peers are on the same LAN.
Peers periodically send all their locally known facts to all the other peers, along with a generic placeholder "I'm here" fact that is used to detect link health. What facts are sent when is filtered to avoid sending facts they think the other peer already knows and is not going to forget before the next send time comes around.
Peers receive facts from other peers as they arrive, but filter them based on a trust model. For now the default trust configuration is simple:
- Peers are trusted to provide possible endpoints for themselves
- Peers are trusted to provide possible endpoints for other peers
- Peers that have an allowed ip value that implies they route packets for the network are trusted to provide AllowedIP values for other peers
- Nobody is trusted by default to provide information on new peers, i.e. all peers must have an externally configured list of the other peer public keys with which they are willing to communicate.
- Peers may have their default trust level overridden in the config file,
including marking peers that are trusted to tell us which peers are valid to
have in the network (
Membership
). If no trusted source (including the static config) says a peer should be a member, it gets removed.
Received facts are removed as they expire based on the given TTL value, or renewed as fresh versions come in from trusted sources.
To connect two peers that aren't directly connected, each end (independently) configures the remote peer in the local wireguard interface with that peer's automatic link local address. It then cycles through the known endpoints and attempts to contact the peer. This should work with simple NAT configurations, but may fail for more complex ones where a full STUN/ICE system would succeed, esp. since there is no coordination on which endpoints are being tried when.
If contact is successful, then the peer's other allowed IPs are added and traffic can start to flow directly (at least it can once both peers have reciprocated on this).
Once a live connection is established, it is monitored to see if it stays alive. If it goes down, and the local peer is not a router, then the allowed IPs other than the automatic link-local one are removed, so that traffic to that peer will be routed through a central router peer, and attempts to connect to that peer directly will resume. The removal of allowed IPs is not done for router nodes since they are the source of that information, and removing them from the router node would cause the network to forget that, and also obstruct that peer from reconnecting to the network.
Determining when there is a live connection to a peer is based on two things:
- Does the wireguard interface report a recent handshake?
- Recent is defined based on a combination of timeout values from the wireguard go implementation.
- Have we received an "I'm here" fact packet from the peer recently.
A couple key items from upstream inspired this:
- Automatic assigning of link-local IPv6 addresses:
https://lists.zx2c4.com/pipermail/wireguard/2017-April/001177.html
(and following messages, about the
wg set wg0 llv6 on
proposal) - The NAT hole-punching example that ships with the wireguard source: https://git.zx2c4.com/WireGuard/tree/contrib/examples/nat-hole-punching