The hostapd backend never populated IPs: NewDHCPCorrelator was defined
but never instantiated, and even when it was, the parser only handled
ISC dhcpd's text format. On a BusyBox-based router using udhcpd, every
device showed up with an empty IP.
Two fixes:
- Add a udhcpd binary lease parser. The format is documented in
busybox/networking/udhcp/dhcpd.{h,c}: an 8-byte big-endian unix-time
header followed by 36-byte dyn_lease records (expires, IP, MAC,
20-byte hostname, 2-byte pad). ParseLeases auto-detects the format
by inspecting the header so the same code path handles both udhcpd
and ISC text leases.
- Wire the DHCPCorrelator into Backend.Initialize and have it merge
two sources: ARP first (universal IP fallback for any station that
has been talked to) and DHCP leases on top (authoritative, carries
the hostname). ARP fills the gap when leases are missing or the
station uses a static IP; DHCP wins on conflict.
Default DHCPLeasesPath updated to /var/lib/udhcpd/udhcpd.leases — the
common BusyBox path. Configurable as before.
Embed the IEEE OUI registry (~1MB pre-processed text file) and resolve
the vendor for every station MAC. Locally administered MACs (U/L bit
set, used by iOS/Android private addresses and virtual interfaces) are
skipped so we don't return spurious matches against randomized prefixes.
The vendor name shows up in the device card as a secondary line, and
falls back to the title position when no DHCP hostname is available —
"Apple" with the IP and MAC is far more useful than "Sans nom".
The lookup table loads lazily (sync.Once) on the first call so the
~40k-entry parse only runs when the station discovery code is exercised.
Hostapd already parsed signal strength and rx/tx counters but the
station -> ConnectedDevice conversion threw them away. Add signalDbm,
rxBytes, txBytes and connectedAt to the OpenAPI schema and the
ConnectedDevice model, and centralise the conversion in
station.ToConnectedDevice so handlers, the periodic refresh and the
event callbacks all serialise the same shape.
Two follow-on bugs surfaced while wiring this up:
- The hostapd backend only stored station entries on first contact.
Subsequent polls were dropped, so signal and byte counters never
refreshed. Reconcile updates in checkStationChanges.
- ConnectedAt was reset to time.Now() on every conversion. Track
FirstSeen on HostapdStation when the station joins, and preserve
the timestamp across periodic refreshes in app.go so the UI's
"connected since" badge is stable.
Frontend gains a metrics row on each device card with signal bars,
total traffic and a live duration. Falls back gracefully when a
backend (DHCP, ARP) doesn't expose these fields.
GuessDeviceType silently returned "mobile" for any device whose hostname
or MAC OUI didn't match a known pattern, so the UI labelled every
unidentified device as a phone. Default to "unknown" instead and broaden
hostname matching (pixel/galaxy, thinkpad, imac/-pc, QEMU OUI).
The hostapd backend was also dropping DHCP hostnames on the floor: the
correlator only forwarded MAC->IP, and convertStation hard-coded the
hostname to "". Replace UpdateIPMapping with UpdateLeaseInfo that carries
both maps so hostnames flow through to ConnectedDevice.Name.
Frontend gains a "Sans nom" fallback when no hostname is available and
French labels for the device-type badge.
Bind to localhost by default and stop echoing backend errors (which can
embed credentials or low-level details) back over the API and log
broadcast. Validate hotspot SSID/passphrase/channel before writing
hostapd.conf and tighten its mode to 0600 since it stores the WPA PSK.
Restrict WebSocket upgrades to same-origin so a LAN browser can't be
turned into a proxy for the API.
Guard shared state: status reads/writes go through StatusMutex (the
periodic updater races with the toggle and status handlers otherwise),
broadcastToWebSockets no longer mutates the client map under RLock, and
station-event callbacks now run under SafeGo so a panic in app code can't
take down the daemon. Stop channels in hostapd, dhcp, and iwd signal
monitors are now closed under sync.Once to survive concurrent Stop calls.
App.Shutdown is idempotent and waits for the periodic loops before
closing backends, so signal-driven and deferred shutdowns no longer race.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
