wchble

Embassy-style Bluetooth Low Energy layer for WCH MCUs, consumed by ch32-hal and ch5xx-hal — not a standalone firmware crate.

HALs own clocks, NVIC, Embassy executor, and board bring-up. This crate owns the radio → soft PHY → link layer → soft HCI path that Trouble Host drives, mirroring how Embassy splits bt-hci + trouble from chip HALs.

Trouble Host          (in HAL examples / apps)
    ↕  bt-hci (in-process soft HCI)
wchble::hci
    ↕
wchble::ll            # Rust link layer
    ↕
wchble::phy           # soft IFS / WFR
    ↕
wchble::rf            # pure Rust MMIO radio HAL
    ↕
ch32-hal / ch5xx-hal  # RCC BLE clock, HSE trim, IRQ install, time

RF backends (cfg-per-family under src/rf/family/):

Feature	Status
ch32v208	DevInit/cal + channel/shut/tx_basic/rx_basic/ISR (open wch_rf); HAL smoke next
ch572	CH5xx map + unique new-table constants (0x64 / 0x16000F); MMIO TBD
ch583	Bases + closed-disasm “old tables” (≈V208); MMIO apply TBD
ch585	DevInit/cal + channel/shut/tx_basic/rx_basic (carved CH585EVT lib) — verified on air (examples/ch585-rf-smoke)
ch592	DevInit/cal + channel/shut/tx_basic/rx_basic — verified on air (examples/ch592-rf-smoke)
ch595	Near-CH585 tables without nCO2522 (0x84 highs); MMIO TBD

Uncertain fields are tagged UNCERTAIN / Certainty::* for silicon debug.

RF/LL/HCI bring-up requires independent on-air verification. A stale sticky W1C TX_DONE can make TX report success without transmitting; clear completion status before a new TX. Respect the radio's DMA alignment requirements and seed a valid AdvA. Verify radiation with a second receiver, not merely a TX completion or timeout.

This repository used to wrap closed WCH .a BLE libraries via bindgen. That path is gone; the open register-level work continues from the CH32V208 NimBLE RF bring-up (wch_rf).

Workspace

Path	Role
src/rf	L0 radio HAL
src/phy	L1 schedule / IFS / WFR
src/ll	L2 link layer
src/hci	L3 soft HCI (bt-hci Controller)
examples/	Flashable RF smoke + Trouble ADV examples

On-target bring-up examples live in this repo while the HAL integration is in flight; production firmware still belongs in the HAL repos.

Features

# Radio / family only (HAL RF bring-up)
cargo check --features ch32v208

# Soft HCI + LL surface for Trouble
cargo check --features ch32v208,controller,phy

Enable at most one chip feature (ch32v208, ch572, ch583, ch585, ch592, ch595).

HAL integration sketch

// inside ch32-hal / application after BLE AHB clock + HSE trim (IRQs still off):
use wchble::hci::{self, SoftRunner};
use wchble::rf::family::ch32v208::Ch32v208Radio;
use wchble::rf::txrx::RX_DMA_MIN;
use wchble::rf::Radio;

static RX: StaticCell<[u8; RX_DMA_MIN]> = StaticCell::new();
let mut radio = Ch32v208Radio::new(RX.init([0; RX_DMA_MIN]))?;
radio.init(&hal_clock)?; // LLE → RFEND → BB → cal; HAL enables NVIC after

let controller = hci::new(); // ExternalController<HciTransport, 4>
let mut soft = SoftRunner::new();
// soft.set_bd_addr([...]);

// join Trouble's runner with SoftRunner::run — SoftRunner must reply
// Command Complete or Host exec hangs waiting on the in-process channel.
// let stack = trouble_host::new(controller, &mut resources);

For on-target RF smoke before Trouble, see examples/ch32v208-rf-smoke and the CH585 and CH592 RF smoke examples. Verify transmitted packets with a second receiver.

Status

CH585 direct-radio bring-up (2026-11-09): ch585-scannable has verified scan responses; ch585-peripheral adds an unencrypted LE 1M, CSA #1 connection prototype. macOS connect/hold/disconnect has passed ten successive 10-second cycles with firmware data RX/TX evidence and a clean 360-second finite run. A separate 120-second client hold completed, but joint firmware acceptance failed on a duration discrepancy and a later extra connection with supervision timeout; it is not reported as passing. This example bypasses the scaffold HCI runner: it is not yet a Trouble/GATT peripheral, and air IFS is unmeasured.

CH585 P4 real Host/GATT integration (2026-11-10): ch585-trouble-gatt connects the event-level radio pump to a bounded single-connection soft HCI/ACL bridge and real Trouble Host. The native real-Host fixture passes discovery, read/write, notification, credit recycling and reconnect tests. With nominal HSE/PLL timing and a corrected RX completion deadline, two separate five-cycle macOS runs passed service/ characteristic discovery, reads, writes, verified sequence notifications, unsubscribe and normal disconnect/reconnect. Each run completed 25 reads, 65 writes and 50 notifications, with clean finite RF finish and smoke recovery. Earlier failed attempts are retained. The original single-cycle capture had a CRC-bad CONNECT and required forced shutdown. A separate new capture verified all three connections, bidirectional data and notifications, with clean SIGINT shutdown in that run. Packet evidence then exposed a receive start/completion deadline mix-up; its correction reduced missed events from 168/339 to 1/172 and restored 1:1 data exchange in the captured links. Whole-run sniffer following and orderly shutdown remain intermittent, not guaranteed.

The first manual P4 phone check succeeded on 2026-11-10: the user reports all five requested steps (read, Write Request/readback, notifications, disabled- notification write/read, disconnect/reconnect persistence) completed on Android, and a later iOS attempt also succeeded, in one directly flashed 360-second window. Firmware recorded four accepted connections and disconnects, six reads, five writes, five queued notifications, 0xAAAAAAAA final value, no Host/bridge/RF stop/TX/RX/supervision error, and a clean 360.000021 s finish; smoke recovery and discovery passed. Both platforms' counters share that window, so platform attribution relies on the user's separate UI observations. This is one manual session per platform, not the planned three isolated acceptance rounds or wireless capture evidence; sustained reliability remains unverified.

The preflight defect is fixed in tools/ch585-preflight: one retained probe session halts, stops the engine, requires two zero acknowledgements, verifies the download and starts the image. Two clean three-cycle macOS acceptance runs then passed with exactly three connections/disconnects, fifteen reads, thirty- nine writes, thirty notifications and no missed events, retries, supervision or bridge errors; smoke recovery passed both times. Earlier attempts are retained: one had an extra connection while a phone was still present, and one hit a transient download failure. Sustained reliability, supervision-timeout recovery, queue-pressure stress and calibrated air timing remain unverified. Open gaps are tracked in BACKLOG.md. See the example for the finite runner and safety/recovery policy.

CH585 P5 service expansion (2026-11-10): ch585-trouble-gatt now also exposes the standard Device Information Service (manufacturer/model/revision) and a Nordic UART-style write/notify pipe. The native real-Host fixture covers DIS discovery and reads plus a UART write/echo notification; a three-cycle macOS run read all three information fields and echoed three distinct strings without RF, Host or bridge errors, then restored smoke. ATT MTU remains 23, so long-data/MTU negotiation and UART flow stress are not yet validated. See BACKLOG.md.

Soft HCI Phase B verified on air (2026-11-08): hci::new() is bt_hci::ExternalController over bounded embassy channels; SoftRunner answers Reset / ReadBdAddr / event masks / LE buffer size / legacy adv cmds. examples/ch32v208-trouble-adv joins SoftRunner with a Trouble Host and advertises ADV_IND on 37/38/39; macOS CoreBluetooth sees wchble (RSSI ≈ -54) via examples/ch32v208-trouble-adv/tools/scan_wchble.swift. Next: SCAN_RSP / connections (Phase C).

License

Apache-2.0. See LICENSE-APACHE.