Aleksandr Bogomiakov
2166ba4800
AWG-only policy also selected a competing end-to-end plugin, so compatible external links could still have no compatible encrypted data channel. Carrier changes replaced tunnel state and status selected counters from different engines. Use one internal WireGuard plugin and a separate DataCarrier contract for external capabilities/lifecycle. Keep the logical link, keys, handshake, interface and counter epoch across policy, fallback and direct/relay changes. Control/discovery transport selection and membership remain independent. Share the bounded per-neighbour UDP WireGuard engine between ordinary WG and Amnezia masking. Encrypt the relay envelope inside the outer hop channel with fresh pair keys/PSK. Remove duplicate WG key/device code and the TCP/TLS inner cipher. TLS carries opaque internal WireGuard frames. Refuse missing native transports instead of letting default iroh impersonate their protocol names. Probe the actual authenticated carrier, independently of the shared IP tunnel; failed local probe submissions count as loss too. Apply the vendored iroh interface exclusion to actual direct candidates, not just public discovery records. Exclude overlay subnets from introductions and propagate the actual owned interface name, including kernel-assigned utun names, to every carrier. Production logs showed paths through the overlay itself. This removes a recursive transport path; it does not establish the historical cause of the reported 100 GB absolute counter. Exit-node traffic is also real peer IP traffic even when local applications are idle. Hide inner cipher choices/handshake clutter in CLI and tray. Report external routes, readiness and actionable policy/version failures. Count original IP packets only when their ciphertext is accepted by the logical transport, including queued packets' exact unpadded lengths. Exclude physical probes and keepalives. Export counter epochs; sum per-peer rate deltas and discard stale histories so joins/restarts cannot turn old totals into fresh traffic. Status also prints exact IP byte totals for comparison with the tray. Coordinated upgrade: data ALPN 6, internal wireguard-e2e capability v4, wg/awg v2, QUIC carrier v1, TLS v2, local IPC v23. Control ALPN 8, device/network identity derivation, signed records and saved network choices stay unchanged. Keep the historical wg-quic key store; transactionally import missing keys from wg then awg without replacing canonical keys. Forget removes legacy copies too, on the blocking pool, to prevent resurrection after restart. Preserve historic cipher options as shared settings, rejecting contradictory values. Validation on Linux in the Nix development shell: - cargo fmt --all -- --check - cargo clippy --locked --workspace --all-targets -- -D warnings - cargo test --locked --workspace --all-targets (562 passed, 3 ignored) The inherited service socket/secret/group environment overrides were removed for isolated CLI process tests. A new carrier test initially raced the first peer announcement; its deadline polling now waits for a peer row explicitly. Real loopback endpoints, membership, SQLite, WG/AWG/TLS encryption and memory TUNs cover mixed-carrier relaying and one smoltcp TCP connection surviving relay/direct/policy changes and a dropped transmission. Counter tests check exact plaintext lengths, refused sends, epochs, joins and stale histories. Public DHT, privileged kernel and manual benchmark tests remain ignored; macOS/Windows native TUNs, desktop GUI and WAN/DPI behavior were not run.
7.2 KiB
Userspace mesh routing
local OS → TUN → IP plugin (encrypt) → router → transport peer link
transport peer link → router → transport peer link (transit)
transport peer link → router → IP plugin (decrypt) → TUN → local OS
The TUN is the boundary of the local host. The kernel sees one overlay route; it never forwards transit traffic. WireGuard encryption remains end to end. The core router sees only an opaque payload and routing metadata. iroh is the current transport adapter, with fragmentation below the routing envelope so the default 1280-byte interface MTU also works on small QUIC paths.
Graph and lifecycle
Each authenticated control session advertises that peer's own direct data links. There is one table for each network, whatever protocols run over it: a link is a way to reach a neighbour, while the single internal protocol is an end-to-end tunnel to a peer. The envelope names the protocol, so the two are independent. A neighbour reached over several kinds of transport (UDP, TCP, QUIC datagrams) is sent to over the best open one, in the order the protocols were configured, and over the next if that one fails or cannot carry the frame. Every authenticated member enters the graph; a member in the middle needs no protocol in common with either end, only a link to each. A snapshot refreshes on the maintenance interval, expires after 90 seconds, and is withdrawn when the session ends. Link arrival, closure and changed announcements rebuild routes immediately. Unchanged announcements refresh their age without republishing the forwarding snapshot.
Every pair uses one internal WireGuard channel, matched by its fixed opaque capability and independent of the selected external protocols. The tunnel is created when a member and its key are known; its logical link survives temporary loss of all physical paths. Carrier selection changes, automatic fallback and shorter routes keep the same keys, interface and counters. The lower endpoint id still chooses the mutually enabled external carrier in its preference order. A relayed AWG first hop can lead to a QUIC next hop: the user's choice governs this device's adjacent links, while the inner ciphertext stays unchanged. An existing TCP connection continues after transient loss if its own timeout has not expired; a sufficiently long outage can still terminate it.
Breadth-first search computes the shortest directed paths and every equal-cost first hop in sorted order. An actual direct link always wins. The first row is always supplied by local transport state, never by a remote claim. Routes stop at 16 links. A packet also carries a decreasing hop limit, bounding temporary loops if different nodes have not yet received the same topology update.
The existing control plane still forms authenticated pairwise sessions. Routing provides multihop data paths among these members; it does not add control-plane flooding or carry control sessions through the data plane, so a member that has no control session with another has no data path to it either. Announcements are volatile, not durable membership or availability guarantees.
What "online" means
One definition, computed by the agent and reported identically by tsng status
and the tray: a member is online when this agent can reach it by any path — a
live authenticated control session (QUIC keep-alive every 5 seconds, so a
vanished peer is noticed within about 15), or a data path to it through other
members. It is never inferred from ping, which a firewall may block, or from a
tunnel's last handshake, which says only when traffic last flowed.
How it is reached is reported separately: direct over a named transport,
relayed through a named member over N hops, or no data path yet while one is
being set up. Being relayed is not being degraded in any way but distance, and
is shown so it is not a surprise.
Packet path
The control loop binds next hops to transport handles in an immutable table
and publishes it through ArcSwap. Readers do not acquire the topology mutex.
The transit routine validates the fixed 82-byte envelope, looks up the source
and destination, selects a cached next hop, decrements one byte, and calls the
transport's synchronous datagram send. It never searches the graph, formats an
endpoint string, validates an Ed25519 key, decrypts, or touches TUN.
An exclusively owned receive buffer is reused when decrementing the hop limit.
Shared buffers require a copy through the safe bytes API. There is no added
transit queue or timer; every raw transport has its own reader. Readers yield
after 64 ready packets to avoid starving other runtime tasks. Local delivery
uses a bounded 256-datagram inbox and drops a new packet when full. Transport
queues and fragment reassembly retain their own limits.
PacketLink::send_flow accepts an opaque 64-bit flow identifier. IP plugins
derive it before encryption from source/destination addresses, protocol and
TCP/UDP ports. WireGuard keeps tags alongside its bounded pending queue, so
the first application packets retain their flow identity after a handshake.
Transit preserves the tag; it never hashes changing ciphertext. Handshake and
keepalive frames use flow zero. Fragmented IP traffic uses a coarse address/
protocol hash because later fragments lack ports; fragmented flows between
the same addresses coalesce. A transition between fragmented and unfragmented
traffic can change its path. The default overlay MTU avoids needing IP
fragmentation for ordinary host TCP traffic.
ECMP is deterministic for a source and flow while the table is unchanged. Closed next hops are skipped until the control loop replaces the table. A topology change can move a flow; datagrams remain unreliable and unordered, and there is no promise to preserve order across a physical path failure. Logical peer links and end-to-end encryption state survive these changes.
Carrier health and switching
Carrier setup failures and persistent encrypted delivery loss advance the initiator's ordered fallback cycle. Logical end-to-end links survive while physical links are retired explicitly. The AWG carrier negotiates a fresh private pair profile on every reentry, without changing topology's durable state. See amnezia.md for thresholds and bounds.
Verification and measurement
The default offline suite exercises four real agents, iroh transports and WireGuard tunnels with memory TUNs, plus graph/forwarder unit tests. It needs neither administrative rights nor public relays/DHT.
Run the forwarding microbenchmark explicitly:
cargo test --release -p tsunagi-core --lib forwarding_benchmark -- --ignored --nocapture
It processes one million 64-byte frames and one million 1280-byte frames with two equal next hops. Frame construction is outside the timed section; the actual ingress routine, header validation, snapshot lookup, ECMP, TTL update, buffer release and mock transport submission are inside. The mock transport counts sends without storing frames. Results are CPU forwarding cost, not end-to-end network latency or VPN throughput; encryption, fragmentation, sockets, congestion and scheduling contribute separately.
Wire compatibility: control ALPN tsunagi/ctrl/8, data ALPN tsunagi/data/6.
Upgrade every participant together; saved identities and network state persist.