Per review 011 §6 sequencing: the four amendment documents drafted before code. All pins, no options — the hedged alternatives are stated as rejected with rationale: - ADR-047 amendment 4: the marked-pub open contract (review 011 §1/§2's seven pins — marker semantics, publish_schema refusal, reserved input key + dispatch strip, wrapper-as-responder-of-record, ack-timing sentence, relay translation, client put-open capability) plus the registration invariants that make stubs unregistrable (§4.8). - ADR-052 (new): channel/control control-handle surface — establisher- contributed ChannelControl, wrapper-owned registry, channel:no_control_handle. - ADR-053 (new): channel/resources/subscribe live resource enumerators — snapshot-then-changes, scan-once aggregation, empty seed. - ADR-022 amendment: announced unmarked Sub/Pub forward through the real forwarding shapes; marked announcements refused (MARKED_SPEC_UNSUPPORTED).
6.8 KiB
ADR-052: The channel/control Control-Handle Surface
Status
Accepted — amends ADR-037 (channel/control's contract) and ADR-047
(amended §1: the generic ops stay). Resolves review 010 S-02 (the
channel/control stub shipped nine published versions; the stub was
pinned by tests and tracked as OQ-39) by implementing the surface the
ADR-037 text always promised. The stub's tracked design work (OQ-39)
is promoted here verbatim-ish: ALPN-crate-registered per-channel
control callbacks, a channels-layer callback registry keyed by
channel_id.
Context
ADR-037 specified channel/control as the out-of-band control op on
channel 0 (resize, signal, keepalive — anything that does not need
ordering relative to data), with the channels layer routing the
ALPN-specific message to a handler's control handle. The
implementation shipped a registered, discoverable stub that checked
channel existence and then returned channel:control_not_implemented
on every invocation — which perversely returned not found for unknown
channels and not implemented for real ones (review 010 S-02). The
implementation was tracked as OQ-39 with the shape recorded; nothing
blocked it except that no ALPN crate had asked yet. alktty's move of
its control surface to the channels path (ADR-049 §5's migration list)
and the general stub ruling (review 010) make that answer invalid.
The design constraint from ADR-039: ChannelManager is auth-blind,
ALPN-blind, transport-blind. The control-handle surface must not
reintroduce ALPN interpretation into the manager.
Decision
1. The control callback type and registry
pub type ChannelControl = Arc<
dyn Fn(Value) -> BoxFuture<'static, Result<Value, CallError>> + Send + Sync,
>;
The callback receives the message value (ALPN-specific JSON; the
channels layer does not interpret it) and returns the ALPN-specific
reply (or a CallError it wants surfaced verbatim — e.g.
target_gone). The live handle: HashMap<u32, ChannelControl> in the
ChannelManager's internals (the same map discipline as channels;
parking_lot mutex).
The manager stays blind to what a callback does — it stores opaque
closures, exactly as ChannelState already stores opaque handler_task
handles.
2. Registration: the establisher contributes the control handle
The ALPN crate registers a control callback by returning it from its
open-op establisher: Establishment::with_control(callback) (builder,
alongside with_reply_field — ADR-049 amendment 3). The wrapper
installs it under the allocated/adopted id
(manager.set_control_handle(id, callback)) and every teardown path
removes it (teardown_channel drops the entry; clear_all drains
them) — same closure-lifecycle discipline as the opener ledger.
Why the establisher, not the handler: the establisher already returns
typed payloads the wrapper routes (Establishment), it runs where the
ALPN crate has prepared its backend (the resize target is the backend
the establisher dialled/planned), and adding a registration path
through the OpenHandler would require handing the handler the
manager and the id — a signature break across every published consumer
for zero benefit. The wrapper-owned installation also makes the
callback's lifetime exactly the channel's lifetime by construction.
3. The op: routing and error surface
make_control_handler (the ADR-037 op, registered by
ChannelOps::register_on) becomes: validate channel_id (unchanged)
→ channel exists? (unchanged; channel:unknown_channel) → callback
registered for the id? →
| case | reply |
|---|---|
| unknown channel | channel:unknown_channel (existing class) |
| no callback registered | channel:no_control_handle — typed, retryable false, message: the channel's ALPN registered no control surface (details: { alpn }) |
| callback ok | { "control": <callback reply> } |
| callback err | the callback's CallError verbatim |
channel:no_control_handle is a real capability answer, not a stub:
a consumer that calls channel/control on a channel whose ALPN chose
not to expose control learns exactly that. The message body is
ALPN-specific JSON (unchanged; the channels layer does not read it).
The callback receives the message only — no identity, no
OperationContext. The authorization home is the op's own
access_control (checked by the registry before the handler, per
ADR-010/017 — the same op-level ACL every other surface runs); a
per-owner resize is expressed as ADR-011 resource_id_path on the
channel/control spec for ALPNs that want it, not by widening the
callback signature. Wire-visibility: channel:no_control_handle
joins the stable error set (ADR-016 — consumers branch on it
deliberately); the callback's reply shape is the producing crate's
schema concern.
4. What this does not decide
Ordering-sensitive control rides the data channel (unchanged — channel-operations.md's control-message division, DP-4 pinned). The manager keeps its blindness; the enumerators registry of ADR-053 is the analogous ALPN-level surface for discovery, not control.
Consequences
Positive: the advertised op serves; S-02 dies; ALPN crates get out-of-band control for free (resize/signal arrive as one routed call-op instead of per-crate in-band framing); teardown-bound callback lifetime is structural.
Negative: the manager's internals gain one map (opaque closures —
no blindness regression); a control callback racing teardown can
observe a torn-down backend through its own error (CallError) — the
same shape the handler EOF path already produces, not a new class.
Door type
One-way (wire-visible). channel:no_control_handle and the reply
envelope shape join the stable surface once consumers branch on them.
The callback type, the establisher carrier, and the map are two-way-door
implementation details inside it (the Establishment carrier is
#[non_exhaustive] — ADR-049 amendment 2's posture).
Verification gates
channel/controlroutes to the establisher-contributed callback and the reply is the callback's output (e2e over a real channels connection);channel:no_control_handleon a channel whose ALPN registered none;channel:unknown_channelpreserved (the pre-existing error);- teardown removes the handle (post-teardown control answers
channel:unknown_channel, never a stale callback); - callback
Errpasses through verbatim.
References
- Review 010 S-02 (the stub; the implement ruling), review 011 §4 (downstream-safe; implement preferred)
- OQ-39 (the promoted design note this ADR resolves)
- ADR-037 (the op's contract), ADR-047 (openable ALPNs; the wrapper
this registration rides), ADR-049 (the establisher carrier this
registers through), ADR-039 (manager blindness), ADR-011
(
resource_id_path— the per-owner control-gate lever)