Add alknet-core architecture specs in docs/architecture/crates/core/ with focused subdocuments for core types, endpoint, auth, and config. Write ADR-010 (ALPN Router and Endpoint) defining AlknetEndpoint, HandlerRegistry, accept loop, and graceful shutdown. Write ADR-011 (AuthContext Structure) defining AuthContext fields, immutability in handle(), and IdentityProvider injection pattern. Resolve OQ-04 (static registration), OQ-12 (file paths only for v1). Add OQ-11 (auth observability). Fix remaining alknet-secret references to alknet-vault across ADRs 003/004/005/009.
7.4 KiB
ADR-010: ALPN Router and Endpoint
Status
Proposed
Context
ADR-001 establishes ALPN-based protocol dispatch: a single QUIC+TLS endpoint accepts connections, and the ALPN negotiated during the TLS handshake routes each connection to the correct ProtocolHandler. ADR-002 defines the ProtocolHandler trait. ADR-006 establishes one ALPN per connection. ADR-007 defines Connection and BiStream.
The question now is: how does the endpoint work? What accepts QUIC connections, negotiates ALPN, and hands connections to handlers? This is the central runtime piece of alknet-core — every handler depends on it.
The reference implementation (alknet-main) uses a Server struct that binds a TransportAcceptor, runs an accept loop, and dispatches to a ServerHandler based on transport type and interface kind. This has three problems that the ALPN model solves:
- Multiple listener types:
ListenerConfighas three variants (Stream, Http, Dns) with per-variant configuration and validation. ALPN eliminates this — one endpoint, one listener, ALPN does the routing. - Protocol detection by byte-peeking: The
stealthmodule reads the first bytes to detect SSH vs HTTP. ALPN negotiation makes this unnecessary — the TLS handshake tells you the protocol before any application bytes are read. - SSH-centric accept loop: The current
handle_connectionimmediately entersrussh::server::run_stream. In the new model, the accept loop is ALPN-agnostic — it doesn't know or care what protocol the handler speaks.
iroh's pattern
iroh's Router registers ProtocolHandler instances with ALPN strings, then calls endpoint.accept() in a loop. For each incoming connection, it reads the negotiated ALPN, looks up the handler, and calls handler.accept(connection). This is clean and proven.
Key design questions
- Handler registration: Static (at startup) or dynamic (at runtime)?
- TLS certificate management: How does the endpoint get TLS certs? Where does ACME fit?
- Connection lifecycle: Who owns the
quinn::Endpoint? How does graceful shutdown work? - Error handling: What happens when a handler panics? When ALPN negotiation fails?
Decision
Endpoint owns the QUIC endpoint
alknet-core owns the quinn::Endpoint directly. The endpoint binds to a single address, configures TLS with a rustls::ServerConfig that includes the ALPN strings from all registered handlers, and accepts connections in a loop.
pub struct AlknetEndpoint {
endpoint: quinn::Endpoint,
handlers: Arc<HandlerRegistry>,
dynamic: Arc<ArcSwap<DynamicConfig>>,
identity_provider: Arc<dyn IdentityProvider>,
shutdown: watch::Receiver<bool>,
}
There is no TransportAcceptor trait, no TransportKind enum, no ListenerConfig enum. QUIC+TLS+ALPN replaces all of that.
HandlerRegistry maps ALPN strings to ProtocolHandler instances
pub struct HandlerRegistry {
handlers: HashMap<&'static [u8], Arc<dyn ProtocolHandler>>,
}
Registration is static at startup. The CLI binary constructs a HandlerRegistry by inserting handlers for each ALPN, then passes it to AlknetEndpoint::new(). The ALPN strings in the TLS ServerConfig are derived from the registry's keys.
This is a two-way door (OQ-04): starting static is simple. If dynamic registration is needed later, the registry can be wrapped in ArcSwap<HandlerRegistry> and the TLS ServerConfig can be regenerated. But ALPN negotiation happens during the TLS handshake, so adding a handler at runtime requires the next connection to use the new ALPN — which the client already has to know about. Dynamic registration has limited value for v1.
Accept loop: connect, dispatch, spawn
loop {
incoming = endpoint.accept().await
connection = incoming.await // TLS handshake + ALPN negotiation
alpn = connection.alpn()
handler = registry.get(alpn)
match handler {
Some(h) => {
auth = resolve_endpoint_auth(connection) // TLS client cert, etc.
tokio::spawn(h.handle(connection, &auth))
}
None => connection.close()
}
}
Key behaviors:
- ALPN mismatch: The TLS handshake fails. This is correct — the client and server have no protocol in common.
- Handler not found: Should not happen — the
ServerConfigonly advertises ALPNs that have registered handlers. If somehow a connection negotiates an ALPN with no handler, the connection is closed with an error log. - Handler panic: The handler runs in a spawned tokio task. If it panics, the task is caught by tokio's panic handler. The connection is dropped. Other connections are unaffected.
- Graceful shutdown: A
watch::Sender<bool>signals the accept loop to stop accepting new connections. Existing connections are given a drain timeout (2 seconds default), then forcefully closed.
TLS certificate configuration
TLS certs come from StaticConfig:
- File paths (
tls_cert,tls_key) for manual provisioning - Self-signed for development
The rustls::ServerConfig is built from the cert + key + ALPN list at startup. The ALPN list is derived from HandlerRegistry::alpn_strings().
ACME auto-provisioning (Let's Encrypt) is not in scope for v1. It will be added as a feature later (see OQ-12).
Error taxonomy
pub enum EndpointError {
BindFailed(io::Error),
TlsConfig(io::Error),
HandlerNotFound(Vec<u8>), // ALPN string with no registered handler
}
pub enum HandlerError {
ConnectionClosed,
StreamError(io::Error),
AuthRequired,
Internal(Box<dyn std::error::Error + Send + Sync>),
}
EndpointError: Problems starting or running the endpoint. Fatal — the endpoint cannot accept connections.HandlerError: Problems within a handler'shandle()method. Non-fatal — the connection is closed, but the endpoint keeps running.
Consequences
Positive:
- Single accept loop replaces multiple listener types and byte-peeking
- ALPN negotiation happens at the TLS layer — no application-level protocol detection
- Adding a handler is registering an ALPN string — no endpoint code changes
- Handler panics are isolated — one bad handler can't take down the endpoint
quinn::Endpointis the only transport — no TransportAcceptor trait needed for v1- The endpoint is testable: give it mock handlers and a test ALPN, verify dispatch
Negative:
- Direct quinn dependency in alknet-core — WASM targets can't use quinn (mitigated: WASM clients don't run endpoints, they connect to them; the WASM door is for client-side handlers, not the endpoint itself)
- No runtime handler registration without regenerating the TLS config (mitigated: two-way door, start static, add ArcSwap later if needed)
- TLS cert provisioning is manual (file paths) for v1 — ACME auto-provisioning is a future feature (OQ-12)
- One address per endpoint — if you need to listen on multiple addresses, run multiple endpoints (acceptable for v1)
References
- ADR-001: ALPN-based protocol dispatch
- ADR-002: ProtocolHandler trait
- ADR-006: ALPN string convention and connection model
- ADR-007: BiStream type definition (Connection, SendStream, RecvStream)
- ADR-009: One-way door decision framework
- OQ-04: Dynamic handler registration (two-way door, start static)
- OQ-05: Multi-transport endpoint (two-way door, start with quinn)
- iroh Router pattern:
docs/research/references/iroh/ - Reference implementation:
alknet-main/crates/alknet-core/src/server/serve.rs