TheMaxMur/RS-Key v0.2.3

RS-Key v0.2.3

on GitHub

[0.2.3] — 2026-06-18

Changed

• LED turns green (idle) as soon as the host configures the device, instead
  of staying on the red boot status until the first applet command arrives. A
  healthy, enumerated key that nothing is talking to yet — e.g. a Linux host with
  no PC/SC daemon running — used to look dead (red) even though it was ready. A
  device-level USB Handler::configured callback now flips the status on
  configuration. bcdDevice 0x0764 → 0x0765.

Fixed

• ~90 s boot stall (LED stuck on the red BOOT status) on some RP2350 boards.
  FidoRng::new seeds the HMAC-DRBG with 48 bytes from the hardware TRNG, and
  the embassy driver runs an autocorrelation health-check on every generated
  block — on a failed check it soft-resets and re-samples in a loop. At the
  default sample_count of 25, consecutive ROSC samples on a marginal unit are
  too correlated, so the check failed almost every time and seeding blocked a
  variable 30–105 s on every boot (init runs before the USB pull-up, so the
  device was simply absent from the bus that whole time — looked dead, worst on
  strict hosts). Raising sample_count to 1000 decorrelates the samples so the
  check passes first try: ~1.5 s boot, HW-verified on the affected board.
  Entropy quality is unchanged — the NIST health checks stay enabled and the
  source is the same; the seed is just gathered reliably. bcdDevice 0x0763
  → 0x0764.

• PIV tab still slow after the present-cache fix below: GET METADATA over
  empty key slots. That bitmap guarded read and size, but has_data — a
  third absent-probe method — still called the backend directly, so a missing
  FID scanned the whole partition. PIV GET METADATA checks has_data(slot)
  first, and ykman piv info / Yubico Authenticator's PIV tab read metadata for
  ~24 mostly-empty slots (9A/9C/9D/9E + 20 retired), so each tab switch paid
  ~24 full scans ≈ 4 s of green-blinking even though every individual APDU
  answered in ~30 ms. has_data now consults the same bitmap → O(1) for an
  absent slot; measured ykman piv info 4.16 s → 0.26 s (~16×) on hardware.
  bcdDevice 0x0762 → 0x0763.

• Slow applet listing (PIV especially), seen as long green-blinking when
  switching tabs in Yubico Authenticator. A backend read/size of an
  absent file scanned the entire ~1.4 MB KV partition to confirm absence, so
  enumerating a sparse object range was O(slots · flash) — opening the
  Certificates tab probes ~25 mostly-empty PIV certificate slots, each a full
  scan. (OATH had the same class of bug, fixed earlier; PIV/others did not.) The
  filesystem now keeps a fixed present/absent bitmap of all FIDs (rebuilt on
  boot, maintained on every write/remove), so an absent read/size returns
  without touching the backend — O(1) instead of a full scan. bcdDevice
0x0761 → 0x0762.

• USB enumeration race at boot (first field report). On a Waveshare RP2350
  the device would "blink red and not be recognised," recovering only after
  several replugs. builder.build() asserts the bus pull-up, so the host begins
  enumerating the moment the device attaches — but the task that answers control
  transfers (usb_task) was spawned only after a block of per-boot init (seed +
  attestation cert + OpenPGP DEK + flash writes, heaviest on a fresh device). The
  host enumerated into an attached-but-mute device and timed out the first
  descriptor request; a lenient host (macOS) usually won, a strict one often did
  not. Boot now completes all that init before attaching, and spawns
  usb_task immediately after build(), so enumeration is serviced with no
  blocking gap. bcdDevice 0x0760 → 0x0761.

Verify a download: https://github.com/TheMaxMur/RS-Key/blob/v0.2.3/docs/supply-chain.md