Back to Bitchat

BLE Transport Architecture V3

docs/BLE-ARCHITECTURE-V3.md

1.7.113.6 KB
Original Source

BLE Transport Architecture V3

The plan of record for restructuring BLEService from an 8.3k-line god object into a layered mesh stack. ARCHITECTURE_V2 rebuilt the app layer above the transport and deliberately deferred the transport itself; this document covers that remainder: what already landed, the target shape, and the order for the rest.

Why the satellite strategy stalled

V2's transport approach was to peel pure policies and closure-driven handlers out of BLEService while the class kept coordinating. The ~30 pure policy structs were a clear win. The five big handler extractions were not: each needed an "environment" of 20–30 closures that weakly capture the service and hop queues back into its state. Logic left, but state ownership and synchronization never moved, so extraction paid a plumbing tax that grew as fast as the logic shrank — the five make*HandlerEnvironment() factories alone were ~1.5k lines. The file held ~60 mutable fields across four concurrency domains whose ownership lived in comments, and every new feature added Transport requirements, state maps, and switch cases to the same class.

Two chronic costs came straight from that structure: queue-order deadlocks (the July 9 main↔bleQueue ABBA freeze), and timing-dependent tests (correctness only observable through real queues and real time).

Target shape

A packet-radio stack with one rule per layer about state and threads:

  1. BLELinkLayer — the only CoreBluetooth import. Owns both managers, scanning/advertising, duty cycle, connection scheduling, MTU, write and notification backpressure buffers, state restoration. Speaks LinkEvent up (link up/down, bytes in, writable) and LinkCommand down (send bytes on link, scan/advertise policy). Knows nothing about packets, peers, or Noise. bleQueue-confined. A SimulatedLinkLayer implementing the same port gives multi-node tests real topologies with no radios and no wall-clock waits.
  2. Mesh engine — one serial queue owning all protocol state: wire codec, fragmentation, dedup, relay policy, peer registry, topology, gossip sync, Noise orchestration. Synchronous single-writer logic; the pure policy satellites slot in unchanged. Endgame: the engine core becomes handle(event, now) -> [Effect] (sans-I/O), which makes the whole mesh property-testable and fuzzable in simulation.
  3. Feature modules — courier, board, prekeys, private media, file transfer, voice, diagnostics, groups, verify/vouch each own their state and register for their message types. A new feature is a new module, not edits to the engine.
  4. App boundary — a small Transport core both transports genuinely implement, plus capability protocols discovered with as? (MeshBridgingTransport etc.), replacing the ~90-requirement god-protocol and its inert defaults.

Concurrency contract

State is owned one of three ways:

  • Engine-confined — mutated only on the serial engine queue (mesh.message). Cross-thread callers use onEngine.
  • bleQueue-confined — link-layer state next to CoreBluetooth objects (link store, write/notification buffers, link-auth maps).
  • Lock-backed store — state with legitimate cross-domain readers (peer registry, local identity/capabilities, traffic monitor). Writes still come from one domain; the lock exists so readers never block on a queue. Every mutation is a single whole-transition method, so readers never observe torn state.

Sync-edge order (deadlock freedom by construction, debug-enforced in onEngine):

main / test threads ──sync──▶ engine ──sync──▶ bleQueue
                                  └──sync──▶ noise / identity queues (leaves)

Nothing may sync-wait in the reverse direction: bleQueue and the crypto queues reach the engine only via async, and nothing sync-dispatches to main. Two subtleties worth knowing:

  • A closure executed inside a noise-manager critical section entered from an engine slot may touch engine state directly (the blocked slot makes it exclusive) but must never sync-re-enter the engine — that is a self-deadlock.
  • bleQueue critical sections (e.g. the verified-announce link rebind) must receive engine-derived values as arguments rather than fetching them through onEngine.

What landed in this pass

  • Lock-backed peer state (BLEPeerRegistryStore): every main-actor Transport read (isPeerConnected, nicknames, snapshots, capability queries) reads a lock, not a queue. Runtime capability bits moved into BLELocalIdentityStateStore beside the identity they ride announces with.
  • bleQueue owns the link buffers: pendingPeripheralWrites, pendingNotifications, pendingWriteBuffers are bleQueue-confined (their producers and drains already ran there); the notification drain no longer invokes CoreBluetooth from a transport queue.
  • One serial engine queue: the concurrent message queue and the collections queue it guarded state with are one serial domain; every barrier flag and per-field ownership comment deleted; ~98 cross-queue hops removed. onEngine documents and debug-enforces the sync-edge order — and its trap caught two latent inversions during migration (the announce-rebind path and the noise session-generation closures).
  • Capability ports: gateway/bridge/courier wiring, the panic lifecycle, and radio-state reads go through MeshBridgingTransport, PanicResettingTransport, and BluetoothStateReporting; no app code casts to BLEService anymore.
  • Feature-owned state: BLEMeshPingTracker (the /ping probe map and per-link response budget) and BLEPrivateMediaSessionStore (the six generation-keyed private-media maps plus the convergence-deferral set, as whole-transition methods under a leaf lock), both with direct unit tests. The private-media store also took the last routine main-actor sync reads off the engine and turned the noise-critical-section transitions into ordinary leaf-lock calls. Remaining feature state (courier, board, prekeys) already lives in injected stores.
  • Transport split: the mesh-only surface left the god-protocol. Transport is core only (lifecycle, identity, snapshots, basic messaging, noise wrappers); files/private media, voice, courier, groups, board, diagnostics, verification, and the public archive are eight capability protocols discovered with as?, alongside the bridging/panic/radio-state ports. The inert-defaults extension is gone; consumers that relied on a default keep its safe floor explicitly at the call site.
  • Contract pinning: BLEQueueContractTests greps the transport sources — only onEngine may sync-enter the engine, transport code never sync-dispatches to main, and the collections queue stays deleted (waivable per line with queue-contract-ok: plus a reason).

Full suite green throughout (1,964 tests), identical wall-clock — BLE throughput is nowhere near what one serial queue sustains.

Remaining roadmap (in order)

  1. Link-layer extraction. Move the CB delegates, scheduling, duty cycle, and buffers behind LinkEvent/LinkCommand ports.

    Link-auth boundary (decided): bindings become engine-owned. Today noiseAuthenticatedLinkOwners, the rebind containment rules, and the peer↔link binding maps live on bleQueue so that "check binding + auth, then act" is one critical section (the rebind path and the authenticated-send commit point in notifyOrEnqueueIfAccepted). That atomicity exists to stop a binding from changing between a security check and its action — and the engine's serial slot provides exactly the same guarantee once every rebind is an engine operation. The residual stolen-link risk is unchanged: directed payloads are Noise ciphertext, useless on a link that changed hands after the decision. Making bindings engine state also puts the receive path in its sans-I/O shape: the link layer reports received(bytes, linkID) and the engine resolves the sender binding, instead of the CB delegate resolving peers before handoff. The link layer keeps only physical link state (CB objects, connect/subscribe lifecycles, backpressure buffers) keyed by opaque link IDs.

    Extraction order: (a) the binding-free radio half — scanning, advertising, duty cycle, connection budget/scheduling — moves first (it makes no peer decisions); (b) bindings + link-auth migrate to the engine, converting readLinkState callers; (c) the delegates shrink to event emission and move behind the port.

    (a) and (b) are done. (a) landed as BLERadioController (#1539). (b) landed in two steps: #1540 cohered the loose maps into BLELinkAuthState + BLELinkBindings (still bleQueue-owned, behavior-identical), and the option-B flip then moved ownership to the engine. Since the flip:

    • linkAuth/linkBindings are engine-owned behind a DEBUG dispatchPrecondition trap; bleQueue code cannot touch them.
    • The receive path is in its sans-I/O shape: bleQueue decodes frames and hands (packet, linkID) up through ingestDecodedPacket (which captures the panic lifecycle at the handoff); attributeAndHandlePacket resolves the sender binding, admits or rejects the claimed sender, applies raw-announce binding, and records ingress — all on the engine. Per-link frame order is preserved end to end (both queues are serial), which supersedes the old batch-local TOCTOU binding.
    • The rotation rebind is one engine slot (rebindLinkAfterVerifiedDirectAnnounce): containment checks, proof retirement, binding flip, reconnect decision, and rotated-identity retirement, with only CoreBluetooth cancels hopping to bleQueue.
    • Authenticated-send eligibility (notifyOrEnqueueIfAccepted, writeOrEnqueueIfAccepted) is checked on the engine — serialized against rebinds by construction — and only the physical admission (updateValue / write / backpressure queues) runs on bleQueue.
    • Teardown splits: bleQueue delegates do physical work inline (discardPeripheralLinkPhysical) and queue the identity half (retirePeripheralLinkIdentity, binding survivor repair) to the engine. A binding can briefly outlive its physical link; queries that need liveness join against the physical store via readLinkState (the engine→bleQueue sync direction), and the queued retirement converges the two.
  2. Sans-I/O engine core + simulator. Make the engine formally handle(event) -> [Effect], feed it from a SimulatedLinkLayer, and move the multi-node E2E suite onto deterministic simulation (no waitUntil, no timing hygiene battles). Property tests become possible: relay-storm bounds, partition-heal convergence, dedup soundness under duplicate floods. The remaining feature code moves (courier, board, prekey, voice, file, group handlers out of the packet switch) ride this seam as handler-registered modules instead of getting closure-environment extractions now.

    The simulator half is done — simulator-first. Because the B2 receive path already hands (packet, linkID) up through one choke point, SimulatedMesh (bitchatTests/Simulation/) wires real CB-free BLEService engines edge-to-edge through the outbound tap and _test_ingestFrame (the production attribution path), with per-edge synthetic link IDs and manual-scheduler time. Five deterministic multi-node tests run in ~40ms: announce/bind convergence, end-to-end Noise establishment, line-topology relay within a TTL/frame budget, duplicate-flood dedup, and the panic rotation single-slot rebind + containment — the scenario that previously required two phones. Fidelity boundary: no physical links, so fanout planning/backpressure is not exercised; protocol behavior is. On its first day the simulator found a real bug: the forced-announce throttle survived panic, so a rotation within bleForceAnnounceMinIntervalSeconds of the last announce left the new identity invisible until the next maintenance cycle (BLEAnnounceThrottle.reset() now runs in the panic slot). The upward port is named and the delegates live behind it. BLELinkEvent (frameDecoded + the four physical lifecycle transitions) is the enumerable bleQueue→engine surface; every crossing goes through emitLinkEvent into one engine consumer (handleLinkEvent), and the simulated mesh drives lifecycle events through the identical enum a radio does (see linkDropEventRetiresBindingAndReconnectHeals). The CoreBluetooth delegate extensions moved to their own files — BLEService+LinkLayerCentralRole.swift / BLEService+LinkLayerPeripheralRole.swift — as physical bookkeeping plus event emission; the physical-domain members they share are internal with the queue contract enforced by the existing traps and grep guards rather than access control.

    Deliberately not done: a formal handle(event) -> [Effect] effect system, and splitting the engine-domain feature handlers into more files. Both would flip the engine's private state (noiseService, peerRegistry, the identity domain) to internal for purely cosmetic file counts — the domains are already uniform (one queue, one rule set) and mechanically guarded. The effect formalization should ride actual feature-module extractions when a feature earns its own module, not precede them.

What this is not

No wire changes: packet formats, signing (padding is signed), the peerID identity binding, and courier tag construction are untouched — see the wire-landmines notes before assuming any of that is local.