Files
alknet/docs/architecture/crates/tls/README.md
T
glm-5.2 192bd0a5a2 docs(arch): spec alknet-tls crate — shared TLS config across quinn + TCP+TLS + iroh (ADR-082, OQ-59)
The TLS setup in alknet-core is welded to quinn: build_rustls_server_config
produces a rustls::ServerConfig, then build_quinn_server_config_from_rustls
consumes it into quinn::ServerConfig — the rustls config is moved, not
shareable. ACME is worse: the AcmeState task is spawned inside the quinn
endpoint, so a TCP+TLS listener would need a second ACME state machine
(two orders for the same domain, two cert caches, Let's Encrypt
rate-limit risk).

alknet-tls extracts the TLS setup into a shareable TlsServerConfig:
- TlsServerConfig::new(identity, alpns) builds the rustls::ServerConfig
  once (including the ACME state machine if ACME)
- for_quinn() clones it for quinn (QuicServerConfig::try_from)
- for_tcp_tls() clones it for tokio-rustls (TlsAcceptor::from)
- rustls_config() borrows it for any other consumer
- One cert, one ACME state machine, N transports

TlsIdentity and Ed25519SecretKey stay in core (config types).
fingerprint.rs stays in core (shared by server endpoint + client
FingerprintPinVerifier in alknet-call — OQ-59 tracks whether it should
move; likely stays because moving would force alknet-call to depend on
alknet-tls, pulling TLS infra into client-only deployments).

iroh shares the key, not the rustls config — iroh has its own TLS built
into the Endpoint (takes iroh::SecretKey, not rustls::ServerConfig).
alknet-tls has no for_iroh() method; the assembly layer passes
Ed25519SecretKey to iroh directly.

Feature gates: quinn / tcp / acme as independent opt-ins. rustls always
present. tokio-rustls only when tcp. quinn only when quinn. rustls-acme
only when acme.

Files:
- crates/tls/README.md: crate spec (TlsServerConfig API, what moves/
  stays, feature gates, deps, behavior-preservation invariants)
- decisions/082-alknet-tls-extraction.md: the ADR (Proposed)
- questions/059-fingerprint-module-location.md: OQ-59 (should
  fingerprint.rs stay in core or move to alknet-tls? open, two-way,
  likely stays)
- open-questions.md: OQ-59 added to alknet-core table
- README.md: tls doc table row + ADR-082 row added

Review (architecture-reviewer): 3 critical (missing
build_quinn_server_config_from_rustls in README table, missing futures
dep, two TlsServerConfig code blocks disagreed on feature gates), 7
warnings (behavior-preservation invariants: max_early_data_size,
aws_lc_rs provider, verifier scheme list, acme-tls/1 ALPN;
rustls-pki-types dep; fingerprint.rs rustls-usage imprecision; ADR
missing acme-tls/1), 4 suggestions. All criticals and warnings addressed.
2026-07-13 10:58:45 +00:00

18 KiB

status, last_updated
status last_updated
draft 2026-07-13

alknet-tls

Shared TLS configuration and certificate management. Builds a rustls::ServerConfig (or an ACME state machine + cert resolver) once, and hands clones to multiple transports — quinn, tokio-rustls (TCP+TLS), and iroh — so one certificate identity serves QUIC and TCP endpoints simultaneously. One ACME state machine, one cert, N transports.

What

alknet-tls extracts the TLS setup that was welded to the quinn endpoint in alknet-core. The existing code (endpoint.rs) builds a rustls::ServerConfig from a TlsIdentity, then consumes it into a quinn::ServerConfig — making it impossible to reuse the same cert for a TCP+TLS listener. ACME is worse: the AcmeState task is spawned inside the quinn endpoint, so a TCP+TLS listener would need its own ACME state machine (two orders for the same domain, two cert caches, potential Let's Encrypt rate-limiting).

alknet-tls fixes this by making the TLS config shareable:

pub struct TlsServerConfig {
    config: rustls::ServerConfig,      // Clone — Arc internally
    acme_handle: Option<JoinHandle<()>>,  // one ACME task, shared
}

impl TlsServerConfig {
    pub async fn new(identity: &TlsIdentity, alpns: &[Vec<u8>]) -> Result<Self, TlsError>;

    /// Clone for quinn. Feature-gated on `quinn`.
    #[cfg(feature = "quinn")]
    pub fn for_quinn(&self) -> Result<quinn::ServerConfig, TlsError>;

    /// Clone for tokio-rustls (TCP+TLS). Feature-gated on `tcp`.
    #[cfg(feature = "tcp")]
    pub fn for_tcp_tls(&self) -> tokio_rustls::TlsAcceptor;

    /// The underlying rustls config, for any other consumer.
    pub fn rustls_config(&self) -> &rustls::ServerConfig;
}

The key property: rustls::ServerConfig is Clone (it holds Arcs to the cert resolver and verifier, not the raw key material). So one TlsServerConfig can feed quinn and TCP+TLS simultaneously — one cert, one ACME state machine, two transports.

Why

The cert-reuse problem is concrete

A hub that serves HTTP (TCP+TLS on 443) and channels (QUIC on 4433) with the same X.509 cert cannot do it with the current code. The rustls::ServerConfig is moved into quinn::ServerConfig and consumed. A TCP+TLS listener would have to build its own rustls::ServerConfig from the same TlsIdentity — re-loading the cert file, or re-deriving the raw key cert, or running a second ACME state machine for the same domains.

ACME is the worst case

Two ACME state machines for the same domain:

  • Two cert-order attempts (race condition on Let's Encrypt's rate limiter).
  • Two cert caches (divergent cache dirs or a shared dir with no coordination).
  • Two AcmeState tasks spawning duplicate resolver() instances.

One TlsServerConfig with one AcmeState task solves this: the ACME state machine runs once, the resolver() is Arc-shared across transports, and both quinn and TCP+TLS get the same cert from the same order.

The three use cases

Use case Identity Transports Browsers?
P2P / native clients RFC 7250 raw key (Ed25519) QUIC + TCP (fallback when UDP blocked) No (browsers can't do raw keys)
Domain-hosted / public service X.509 (manual or ACME) QUIC + TCP+TLS (same cert) Yes (via HTTPS / WebSocket; WebTransport when revived)
Development Self-signed Any No (untrusted)

In all cases, TLS + ALPNs "just works" — the TLS handshake negotiates the ALPN, the HandlerRegistry dispatches by ALPN, the transport is a parameter. The TLS crate's job is to make the cert available to whichever transports the deployment runs.

Why a separate crate

  1. Dependency isolation. tokio-rustls is a real dependency that shouldn't be forced on every alknet-core consumer. A node that only uses quinn doesn't need tokio-rustls. A node that uses TCP+TLS needs it. Feature-gating in core mixes concerns — core is types and traits, not transport-specific TLS setup.

  2. ACME is heavy. rustls-acme spawns a long-running async task, manages a cert cache, talks to Let's Encrypt. That's transport infra, not core types. It belongs in a crate the assembly layer pulls in when it needs real TLS setup, not in core.

  3. Quinn and iroh have their own TLS. Quinn wraps rustls::ServerConfig into quinn::ServerConfig. Iroh uses its own raw-key TLS built into the Endpoint. alknet-tls is the "I have a cert and I want to share it across transports" layer — it doesn't replace quinn's or iroh's TLS, it provides the shared cert source.

Architecture

What moves from alknet-core to alknet-tls

Component Current location New location
TlsIdentity enum alknet-core/config.rs stays in core (it's a config type)
Ed25519SecretKey alknet-core/config.rs stays in core (config type)
build_rustls_server_config() alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls (unconditional)
build_quinn_server_config_from_rustls() alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls (for_quinn() — wraps rustls config in QuicServerConfig)
TlsSetup (ACME state machine) alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls (the TlsServerConfig::new ACME path)
RawKeyCertResolver alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls
Ed25519SigningKey alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls
AcceptAnyCertVerifier alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls
SelfSignedCert / generate_self_signed_cert() alknet-core/endpoint.rs (#[cfg(feature = "quinn")]) alknet-tls
load_cert_chain() / load_private_key() alknet-core/endpoint.rs alknet-tls
fingerprint.rs alknet-core/fingerprint.rs stays in core (shared by server + client; client is in alknet-call; production code uses sha2 + manual DER only — rustls is test-only. See OQ-59.)

TlsIdentity and Ed25519SecretKey stay in core because they're config types — StaticConfig holds a TlsIdentity, and config types belong in core. alknet-tls re-exports them for convenience. fingerprint.rs stays in core because it's shared by both the server path (endpoint extracts fingerprint from the client cert) and the client path (FingerprintPinVerifier in alknet-call matches the server's cert against a pinned fingerprint). The production code in fingerprint.rs uses only sha2 and manual DER parsing — the rustls::sign usage is in the test helper only. See OQ-59 for the full trade-off.

TlsServerConfig

The central type. Built once from a TlsIdentity + ALPN list, shared across transports.

pub struct TlsServerConfig {
    /// The rustls server config. Clone-safe (holds Arcs to cert resolver
    /// and verifier, not raw key material).
    config: rustls::ServerConfig,
    /// The ACME state machine task, if ACME is active. One task, shared
    /// — dropping this handle does NOT stop the ACME state machine
    /// (it's owned by the config, not the endpoint).
    acme_handle: Option<tokio::task::JoinHandle<()>>,
}

Construction:

impl TlsServerConfig {
    /// Build a TLS server config from the given identity and ALPN list.
    /// For ACME, spawns the ACME state machine task and wires its
    /// resolver into the rustls config. For X509/RawKey/SelfSigned,
    /// loads the cert and builds the resolver directly.
    pub async fn new(
        identity: &TlsIdentity,
        alpns: &[Vec<u8>],
    ) -> Result<Self, TlsError>;
}

The ALPN list is the set of ALPNs the deployment wants to advertise (alknet/call, alknet/channels, h2, http/1.1, etc.). For ACME, the acme-tls/1 ALPN is appended automatically (for the TLS-ALPN-01 challenge, ADR-027 §7).

Behavior-preservation invariants

The extraction must preserve these load-bearing TLS behaviors from the current code. An implementer who omits any of these produces a crate that compiles and passes type-checks but silently changes TLS behavior:

  • max_early_data_size = u32::MAX on all server config paths (X509, RawKey, SelfSigned, ACME). Enables 0-RTT / early data. Omitting it disables 0-RTT, silently breaking clients that use it.
  • rustls::crypto::aws_lc_rs::default_provider() as the crypto provider on all paths. Matches iroh's tls-aws-lc-rs feature. Do not switch to ring or the process-default provider without an ADR — different FIPS status, different platform support.
  • AcceptAnyCertVerifier's supported_verify_schemes() returns ED25519 + ECDSA P-256/P-384 + RSA PSS/PKCS1 (SHA256/384/512). This list determines which client cert signature algorithms the server accepts. Must be preserved verbatim.
  • acme-tls/1 ALPN append for the ACME path only. The TLS-ALPN-01 challenge requires the server to advertise acme-tls/1 in its ALPN list. Appended in TlsServerConfig::new's ACME branch, not by the caller.

Transport-specific accessors:

impl TlsServerConfig {
    /// Produce a `quinn::ServerConfig` for a QUIC listener. Clones the
    /// rustls config (cheap — Arc-shared cert resolver), wraps it in
    /// `QuicServerConfig`. Feature-gated on `quinn`.
    #[cfg(feature = "quinn")]
    pub fn for_quinn(&self) -> Result<quinn::ServerConfig, TlsError>;

    /// Produce a `tokio_rustls::TlsAcceptor` for a TCP+TLS listener.
    /// Clones the rustls config. Feature-gated on `tcp` (pulls
    /// `tokio-rustls`).
    #[cfg(feature = "tcp")]
    pub fn for_tcp_tls(&self) -> tokio_rustls::TlsAcceptor;

    /// Borrow the underlying rustls config, for consumers that need to
    /// build their own transport-specific wrapper (e.g. iroh, which
    /// has its own TLS built in but shares the Ed25519 key).
    pub fn rustls_config(&self) -> &rustls::ServerConfig;
}

Iroh: shares the key, not the rustls config

Iroh is different from quinn and TCP+TLS: it has its own TLS built into the Endpoint, using RFC 7250 raw keys. It does not consume a rustls::ServerConfig — it takes an iroh::SecretKey and handles TLS internally. So alknet-tls does not have a for_iroh() method. Instead, the assembly layer reads the Ed25519SecretKey from StaticConfig (stays in core) and passes it to iroh's Endpoint::builder().secret_key() directly. alknet-tls is involved only when iroh is not the sole transport — in that case, the same Ed25519SecretKey feeds both TlsServerConfig::new(TlsIdentity::RawKey(key), ...) (for quinn/TCP) and iroh::SecretKey::from_bytes(key.as_bytes()) (for iroh).

The fingerprint is normalized across all three paths (ADR-030 §6): ed25519:<hex> for raw keys, whether the cert came from quinn's RawKeyCertResolver, iroh's built-in TLS, or a future TCP+TLS raw-key path. fingerprint.rs (in core) handles this.

Feature gates

[features]
default = []
quinn = ["dep:quinn"]      # for_quinn() — wraps rustls config for quinn
tcp = ["dep:tokio-rustls"] # for_tcp_tls() — wraps rustls config for TCP+TLS
acme = ["dep:rustls-acme"] # ACME state machine

A deployment that only uses quinn enables quinn. A deployment that uses TCP+TLS enables tcp. A deployment that uses both enables both. ACME is opt-in (heavy dep, long-running task). The rustls dep is always present (it's the core TLS library).

Dependencies

alknet-tls
├── alknet-core       (TlsIdentity, Ed25519SecretKey, fingerprint — re-exported)
├── rustls            (ServerConfig, cert types — always present)
├── rustls-pki-types  (CertificateDer, PrivateKeyDer, etc. — via rustls re-export
│                     or direct dep; core lists it directly)
├── rustls-pemfile    (cert/key file loading — always present)
├── rcgen             (self-signed cert generation — always present)
├── ed25519-dalek     (Ed25519 signing key — always present, via core)
├── sha2              (fingerprint computation — always present, via core)
├── tokio             (spawn for ACME task — always present)
├── futures           (StreamExt for ACME event loop — present; only used
│                     in the ACME path, can be acme-gated or always present)
├── tracing           (logging)
├── quinn             (optional — for_quinn())
├── tokio-rustls      (optional — for_tcp_tls())
└── rustls-acme       (optional — ACME state machine)

alknet-core loses the quinn, iroh, rustls-pemfile, rcgen, ed25519-dalek, and rustls-acme deps from its [features] section — they move to alknet-tls. Core keeps a narrow rustls / rustls-pki-types dep only if fingerprint.rs stays (OQ-59): the production fingerprint code uses sha2 + manual DER only, but the test helper build_ed25519_spki_der uses rustls::sign::public_key_to_spki. If fingerprint.rs moves to alknet-tls, core becomes rustls-free.

What AlknetEndpoint does after the refactor

AlknetEndpoint::new() currently builds TlsSetup internally. After the refactor, it takes an Arc<TlsServerConfig>:

impl AlknetEndpoint {
    pub async fn new(
        static_config: &StaticConfig,
        tls_config: Arc<TlsServerConfig>,  // ← new param
        handlers: HandlerRegistry,
        dynamic: Arc<ArcSwap<DynamicConfig>>,
        identity_provider: Arc<dyn IdentityProvider>,
    ) -> Result<Self, EndpointError>;
}

The endpoint calls tls_config.for_quinn() to get its quinn server config. The ACME handle is on the TlsServerConfig, not the endpoint — the endpoint doesn't own the ACME task. The assembly layer builds the TlsServerConfig once, passes Arc::clone() to the endpoint, and passes another Arc::clone() to the TCP+TLS accept loop (which lives in the hub or a future TcpTlsAcceptor).

The TCP+TLS accept loop (out of scope for this crate)

alknet-tls provides for_tcp_tls() -> TlsAcceptor. The actual TCP accept loop (TcpListener::accept → TlsAcceptor::accept → Connection::from_bidi → HandlerRegistry::dispatch) lives elsewhere — in alknet-hub (the hub is the primary multi-transport consumer) or in a future alknet-core TcpTlsAcceptor module. alknet-tls is the cert provider, not the accept loop. This keeps alknet-tls focused on TLS setup and cert sharing, not transport accept logic.

Crate dependencies (in the dep graph)

alknet-tls
├── alknet-core (TlsIdentity, Ed25519SecretKey, fingerprint)

alknet-core (loses TLS setup code)
├── (rustls — only for fingerprint.rs types, if kept)

alknet-call (client-side verifier — unchanged)
├── alknet-core (fingerprint.rs)

alknet-http (future: TCP+TLS accept loop)
├── alknet-tls (TlsServerConfig::for_tcp_tls())
├── alknet-core (Connection::from_bidi, HandlerRegistry)

alknet-hub (multi-transport endpoint)
├── alknet-tls (TlsServerConfig — shared across quinn + TCP)
├── alknet-channels-call (ChannelClient)
├── alknet-call (CallAdapter, Dispatcher)
├── alknet-http (HttpAdapter)
├── alknet-core (AlknetEndpoint, HandlerRegistry, Connection)

alknet-tls depends on alknet-core only. No handler crate depends on alknet-tls — they depend on alknet-core for types and on alknet-tls only indirectly through the assembly layer. The assembly layer (the deployment binary) builds the TlsServerConfig and passes it to the endpoint and the TCP+TLS accept loop.

Design Decisions

All design decisions are documented as ADRs in decisions/.

ADR Decision Summary
082 alknet-tls crate extraction Extract TLS setup from alknet-core/endpoint.rs; TlsServerConfig shareable across quinn + TCP+TLS + iroh; one ACME state machine

Open Questions

See open-questions.md for full details.

  • OQ-59 (open): Should fingerprint.rs stay in alknet-core or move to alknet-tls? It uses rustls::pki_types and rustls::sign types, which creates a rustls dep in core. If it moves to alknet-tls, the client-side FingerprintPinVerifier (in alknet-call) would depend on alknet-tls — a new dep edge. If it stays, core keeps a narrow rustls dep. Decision-ready — the answer depends on whether we want core to be rustls-free.

References

  • docs/architecture/decisions/027-tls-identity-redesign-acme-rawkey-decoupling.md — TlsIdentity (RawKey / X509 / Acme), RFC 7250, browser limitation
  • docs/architecture/decisions/030-peerentry-and-identity-id-decoupling.md §6 — fingerprint normalization (ed25519:<hex> across quinn/iroh)
  • docs/architecture/decisions/034-outgoing-only-x509-and-three-peer-roles.md — client-side verifier selection (CA vs fingerprint pin)
  • docs/architecture/decisions/065-connection-from-stream-generic-single-stream.md — Connection::from_stream/from_bidi (TCP+TLS path)
  • docs/architecture/decisions/010-alpn-router-and-endpoint.md Amendment 1 — TCP+TLS dispatch via from_stream (accept loop outside the endpoint struct)
  • docs/architecture/crates/core/endpoint.md — current endpoint design (TLS section will be amended to point to alknet-tls)
  • docs/architecture/crates/core/config.md — TlsIdentity, StaticConfig
  • crates/alknet-core/src/endpoint.rs — the code being extracted (build_rustls_server_config, TlsSetup, RawKeyCertResolver, Ed25519SigningKey, AcceptAnyCertVerifier, generate_self_signed_cert, load_cert_chain, load_private_key)
  • crates/alknet-core/src/fingerprint.rs — fingerprint extraction (shared by server endpoint and client verifier)