tsunagi/docs/wireguard.md

The WireGuard data plane

WireGuard is the first IP plugin. It carries user traffic between participants while the control plane keeps doing its own job: deciding who is in the network and carrying each participant's opaque announcement.

Module boundaries are in architecture.md, the control protocol in protocol.md, the security consequences in threat-model.md.

Userspace, not the kernel

WireGuard here is boringtun's protocol state machine running in this process. There is no kernel WireGuard module and no wg tool: the same code runs everywhere, and the protocol can be exercised in tests without any privileges at all.

The only privileged step left is creating a packet interface so the operating system can hand us IP packets, and even that is behind a trait (TunFactory) with an in-memory implementation.

	needs privileges	what it proves
MemoryTunFactory	no	handshake, encryption, routing, address ownership
MockProvisioner	no	the above, plus what would have been done to the host
ManagedTunFactory	CAP_NET_ADMIN	traffic actually reaches the OS

ManagedTunFactory creates the interface and configures it; see Provisioning the interface below and Privileges in ../README.md. With no capability the agent runs with --no-tun: everything but the last hop into the kernel still works.

Where the packets go

The plugin does not know and does not care. It is handed a PacketLink per peer by the agent and runs a WireGuard tunnel over it:

  TUN device (IP packets)                  PacketLink per peer
       |                                          |
       v                                          v
  destination address -> peer  --Tunn.encapsulate-->  ciphertext -> transport
  source address checked       <--Tunn.decapsulate--  ciphertext <- transport

Reachability — hole punching, relay fallback — belongs to the transport, which today is iroh. That is the whole reason the plugin's announcement says who it is and never where it is: there is no address for a peer to advertise, get wrong, or lie about.

Two peers behind NAT work exactly as well as iroh does. iroh hole punches a direct path when it can and falls back to a relay when it cannot; the tunnel rides on whichever it got. There is no separate STUN, no separate hole punching and no second set of NAT problems to solve for WireGuard.

Checking it from outside

tsunagi status asks a running agent over its local control socket and prints what it sees, including whether each tunnel has actually handshaken. See ../README.md.

Deterministic overlay addressing

A mesh with no coordinator cannot hand out addresses, so everyone derives their own. The result is an IPv6 unique local address (RFC 4193):

prefix (/64) = 0xfd || SHA-256( LP(domain) || LP("prefix") || LP(network_id) )[0..7]
iid    (64b) =         SHA-256( LP(domain) || LP("interface") || LP(network_id) || LP(wg_public_key) )[0..8]
address      = prefix || iid

with domain = "tsunagi-wireguard-overlay-v1" and LP(x) = u32_be(len(x)) || x, the same unambiguous encoding the rest of the project uses.

Two consequences matter:

• every member of a network derives the same /64, so the overlay is one subnet that nobody had to allocate;
• a member's address is bound to its WireGuard public key, so address ownership can be checked locally rather than believed.

IPv4: allocated, signed, and kept

IPv6 addresses are derived: a 64 bit interface identifier makes a collision impossible in practice, so nobody has to agree on anything. IPv4 has nothing like that room, so deriving would collide. Instead an address is allocated and then recorded as a signed fact, using the model in sync-model.md.

default range          10.13.37.0/24   (override with --ipv4-range)
who decides            the first member to claim; later ones adopt what they find
who signs              the claiming member, with its persistent device key
where it is kept       state.sqlite, and every replica that has seen it
what a return costs    nothing: the old address is reclaimed

How it works:

1. On joining, an agent reads back the records it already had and learns more from its peers.
2. If it already holds an address, it keeps it. That is the whole point: a participant that was away for a month comes back to the address it signed for, because the claim outlived the session.
3. Otherwise it picks a free one — starting from a position derived from its own identity, so two newcomers rarely start in the same place — signs the claim, commits it together with its version counter, and only then announces it.
4. Every replica merges what it receives into what it has. An author missing from a snapshot is left alone: absence is not deletion.

No vote is involved, deliberately. Anyone who knows the network secret can mint identities, so a majority proves nothing, and a quorum would stall with one participant online and diverge across a partition. Two members who claim the same address at the same moment are resolved by a rule both compute identically — the lower endpoint id keeps it — and the loser simply allocates again with a higher version.

The range is agreed, not configured per member. --ipv4-range says what this agent would use; a network that has already settled on something else wins, and the agent adopts it. So the flag matters for whoever starts the network and is harmless afterwards. Pass --ipv4-range none for an IPv6-only overlay.

The agent puts the address on the interface itself, as soon as the network has agreed on it — no restart, and the interface is not recreated, which would drop every tunnel riding on it.

It then checks that it is really there, by binding a UDP socket to it, which needs no privileges. That check is deliberately independent of the code that did the assigning: the failure it guards against is a silent one. With the wrong address on the interface, packets leave with the wrong source and every peer drops them as not belonging to us, which looks like a broken network rather than a broken assumption.

A release tombstone exists in the record type and merges correctly, but nothing emits one yet, so an address stays claimed until the network is forgotten.

Address ownership is enforced, not announced

Kernel WireGuard enforces AllowedIPs. In userspace that is our job, and device does it on both sides:

• outbound, a packet is routed to the peer that owns its destination address; a destination nobody owns is counted as unroutable and dropped;
• inbound, a decrypted packet is dropped unless its source is exactly the address derived for the peer whose tunnel decrypted it.

Both apply to IPv4 and IPv6 alike.

So a participant cannot receive traffic addressed to somebody else and cannot forge traffic that appears to come from somebody else. A participant who knows the network secret can mint many keys and therefore occupy many addresses, but it cannot choose to collide with an existing member without finding a hash preimage.

The announcement also carries the address the peer believes it has. It is never used — only cross-checked — so a version skew produces a clear rejection rather than silent non-connectivity.

MTU

Two constraints pull against each other.

IPv6 sets a floor of 1280 bytes (RFC 8200), and Linux enforces it brutally: an interface whose MTU drops below 1280 loses IPv6 entirely — its /proc/sys/net/ipv6/conf/<dev> directory disappears and ip -6 address add answers Invalid argument. So the overlay MTU cannot go below 1280, and the plugin refuses a smaller one at startup instead of letting it fail obscurely.

The transport sets a ceiling. Every packet rides in one datagram and WireGuard adds 32 bytes, so a link must carry mtu + 32 = 1312 bytes. A direct QUIC path typically offers around 1380, which fits. A relayed path can offer less, and then full-size packets do not fit: they are dropped and counted as dropped_oversize, never truncated, and the plugin reports the exact numbers when the tunnel is set up.

There is no room left to trade, so the default MTU is exactly 1280. Fragmenting a packet across several datagrams would lift the ceiling and is not implemented.

Lifecycle

• A network is activated → the plugin loads or creates its key for that network, derives the interface name, and creates the packet interface. If that fails — no privileges, for instance — the key and the announcement still work and the interface is retried on the next reconcile.
• A peer announces its key → recorded.
• A data link to that peer arrives → recorded.
• Reconciliation starts a tunnel for every peer that has both, and removes tunnels for peers that lost either.
• A network is deactivated, or the agent shuts down → the interface and every tunnel go away. The key stays, so coming back keeps the same overlay address.

There is no external configuration file and no command line tool, so unlike a kernel-WireGuard setup there is nothing outside this process for anybody to edit. Reconciliation is purely "do the running tunnels match what is known".

Using it

# On both machines: the agent, then the network
tsunagi up
tsunagi join --network lab --secret "$SECRET"

See the two-machine walkthrough in ../README.md.

From the library:

use std::sync::Arc;
use tsunagi::config::{AgentConfig, StoragePaths, TransportPolicy};
use tsunagi::dataplane::IpPlugin;
use tsunagi::dataplane::wireguard::{MemoryTunFactory, WireguardConfig, WireguardPlugin};
use tsunagi::identity::{NetworkName, NetworkSecret};
use tsunagi::{Agent, Result};

#[tokio::main]
async fn main() -> Result<()> {
    let paths = StoragePaths::user_default()?;

    // MemoryTunFactory needs no privileges; swap in ManagedTunFactory for a
    // real interface.
    let plugin = WireguardPlugin::open(
        WireguardConfig::new(paths.state_dir.join("wireguard")),
        Arc::new(MemoryTunFactory::new()),
    )
    .await
    .expect("wireguard plugin");

    let agent = Agent::spawn(
        AgentConfig::new(paths)
            .with_transport(TransportPolicy::N0Defaults)
            .with_plugin(plugin.clone() as Arc<dyn IpPlugin>),
    )
    .await?;

    let network = agent
        .join_network(&NetworkName::new("lab")?, &NetworkSecret::generate())
        .await?;

    if let Some(view) = plugin.overview(network) {
        println!("{} on {}", view.interface, view.overlay_address);
    }
    agent.shutdown().await;
    Ok(())
}

Limits and future work

• Full mesh only. Every member runs a tunnel to every other member. Routing through an intermediate participant is not implemented.
• Nothing frees an address yet. The release record exists and merges, but no command emits one.
• A snapshot grows with the number of members ever seen, and is capped per message rather than compacted.
• No routes, DNS or firewall rules. The plugin creates its interface and nothing else. Anything beyond the overlay /64 is the operator's business.
• Membership is session-scoped. A peer leaves the overlay when its control session ends; surviving a long absence is the same future work.
• Userspace costs CPU. Kernel WireGuard is faster. A kernel backend could return behind the same boundary, but it would give up transport-provided NAT traversal unless paired with a local proxy.
• The interface is managed, not prepared. On Linux the agent creates the TUN interface and configures it over netlink, in process. See Provisioning below.
• No provisioner exists for macOS or Windows yet. There the agent attaches to an interface prepared by hand and says so.
• The system interface path is not exercised by the default suite, because it needs privileges. Everything else about the data plane is.

Provisioning the interface

Everything the printed ip recipe used to do happens in process now, over netlink. The shape of it is a reconciliation rather than a sequence of commands: the agent is handed a plan — name, MTU, the addresses the interface should carry and no others — observes what is actually on the host, and applies the difference. Running it twice changes nothing the second time.

The decision of what to change is dataplane::wireguard::provision::plan_changes: pure, platform-independent and unit-tested on every platform. Only the execution is behind InterfaceProvisioner, which has three implementations — netlink on Linux, a MockProvisioner over a pretend host for the tests, and one that refuses with an explanation everywhere else.

Cleanup is the default, not an action

The interface is created by opening /dev/net/tun and is not made persistent, so the kernel destroys it when the last descriptor closes. A clean shutdown, a panic, SIGKILL and a power cut all leave the same amount behind: nothing. There is no path by which a dead agent leaves an interface, because keeping one alive is what requires a live process.

This also removes the two settings the manual recipe needed. keep_addr_on_down existed only because an interface nobody held open lost carrier and had its IPv6 addresses flushed; nodad only because duplicate address detection cannot finish without carrier. An interface held open for its whole life has carrier for its whole life.

Repairing what an older run left

Two things can still be sitting on the name: an interface created persistent by the old recipe, and — narrowly — one from a run killed between TUNSETIFF and the agent recording it. Both are replaced, which discards their stale addresses with them.

The two refusals are the interesting part:

• A link that is not a TUN is never touched. The name is derived from the network id, so colliding with a real device is unlikely rather than impossible, and deleting somebody's bridge is not a recoverable mistake.
• A TUN another process holds open is never deleted. Carrier is the signal: a TUN has it exactly while something is attached. An attached one is a working overlay, almost certainly a second agent on this host, and it is told to use a different --wg-prefix instead.

Privilege

CAP_NET_ADMIN is required and is kept out of the effective set except around the netlink calls that need it. setcap cap_net_admin+p leaves it permitted but not effective at exec, which is the resting state; the agent raises it for a few milliseconds at startup and again when its address allocation changes.

Two facts shape how that is done. Capabilities on Linux are per thread, and netlink checks the credentials of whichever thread calls sendmsg — which, with an async client, is the connection task and not the caller. So raising a capability around an await would be wrong in the way that works until the scheduler moves the task.

Therefore: all netlink work runs on one dedicated thread with a current-thread runtime, where nothing is polled outside a block_on, and the capability is raised immediately before that call and lowered immediately after. Opening the TUN descriptor is the other privileged act; it is a synchronous call with no await between the guard and the release, so it stays on its own thread by construction.

What it cannot be told to do

Nothing here takes a name, an address or a command from the network. The interface name is derived from the network id, the addresses come from the local plugin and the signed allocation records, and no external program is executed at any point — there is no ip, no shell and no PATH involved.