Skip to content

The RAK7289CV2 power board

The RAK7289CV2 (WisGate Edge Pro) reaches its power board over Modbus RTU, through a CH340 USB-serial bridge inside the case. Two things sit on that bus:

  • the power board’s hardware watchdog. On the newer hardware revision, a unit on DC power whose OS does not feed it loses its power.
  • the optional RAK battery pack.

pulzr-rak-power feeds the one, reads the other, and reports both to Pulzar Cloud.

It is RAK-specific and deliberately small. It is not a general RS485 or Modbus service. The slaves, the registers and the line settings are RAK’s and fixed, and nothing about it is configurable. A gateway with real IO gets its own design, not this one grown.

Nothing on this bus has been seen answering. The design rests on three sources and one bench probe.

  • ChirpStack’s RAK support, the one known-working user of this bus: chirpstack/chirpstack-gateway-utils, src/rak/watchdog.rs and src/rak/battery.rs. They use Modbus RTU at 9600 8N1, with a 100 ms timeout per request.

    • Slave 0xC8 (200) is the watchdog. Writing 10 to holding register 0x0001 feeds it.
    • Slave 0x6E (110) is the battery pack. ChirpStack reads four holding registers, one at a time: 0x6000 voltage ×0.01 V, 0x6003 health %, 0x6005 remaining capacity, 0x6006 full-charge capacity.
  • chirpstack/chirpstack-gateway-os#155. RAK support says the newer RAK7289V2/CV2 revision has a hardware watchdog and “fully reboots every 10 minutes” unless WisGateOS feeds it. The reporter saw a unit on 12 V DC go off after about ten minutes and stay off, and saw no problem on PoE. ChirpStack feeds every 5 minutes from cron, and the reporter confirmed that keeps a DC unit up. The dmesg in that issue, from a RAK7289V2, shows the bridge: ch341 1-1.4:1.0: ch341-uart converter detected, so a CH340 (1a86:7523) at port 1-1.4 on the internal hub. RAK’s instructions there for disabling the watchdog name its WisGateOS daemon, spregread.

  • RAK’s own WisGateOS 2.3.1 (RAK636 image, read statically, not run).

    • /etc/hotplug.d/usb/11-symlinks maps 1-1.4:1.0 to /dev/ttyRS485, so the CH340 is RAK’s RS485 port. The same file has LoRa card 0 at 1-1.2, card 1 at 1-1.3, and the EG95’s AT port at 1-1.1:1.2.
    • The bus is driven by /usr/sbin/spregread, which uses libmodbus. Its init script runs it only when the unit’s factory capabilities (einfo.dev.caps) include solar_battery (battery) or hwwdg (watchdog). It first runs spregread -i /dev/ttyRS485 -e 0 -E 1 -F, whose help says “force off WDT”. So the watchdog can be switched off over Modbus as well as fed. RAK’s stop_wdt_service does the same, presumably so a long operation such as an upgrade is not cut off. That register is not known here: the binary has not been disassembled.
    • /etc/config/spregread: battery at device 110, address 0x6000, read every 10 s; wdtcfg timeout 10. That 10 matches the value ChirpStack writes to 0x0001, so the feed is most likely the watchdog’s timeout in minutes, rewritten.
  • The bench Pro (2026-10-26, PulzarOS 0.1.0, with kmod-usb-serial-ch341 from the 24.10.0 feed, on PoE). The CH340 came up as /dev/ttyUSB4 at 1-1.4, with the EG95 on ttyUSB0-3. At 9600 8N1, nothing answered:

    • a read of 0x0001 at slave 200
    • ChirpStack’s feed (function 06, 0x0001 = 10)
    • a read of 0x6000 at slave 110

    The line was silent for 5 s idle, and its own TX did not echo. No battery pack was fitted, so 110’s silence is expected. The unit’s factory descriptor (mtd factory at 0x2000) lists DEV_CAPS as dying_gasp solar_battery hwwdg lbt, so it has the watchdog and the battery hardware. Its watchdog was silent on PoE, so a unit that has one may not answer, or may not run it, off DC. A Lite’s descriptor lists only breathing_light.

Still open:

Question What depends on it
Does a newer-revision unit on DC answer at 200, and does the feed keep it up? Everything here. ChirpStack’s code and the reporter in #155 are the only evidence.
Does a fitted battery pack answer at 110 with those registers? The battery facts on the gateway page.
What does register 0x0001 = 10 mean? Most likely the timeout in minutes, matching RAK’s wdtcfg timeout 10. Feeding every two minutes is safe under any reading that keeps ChirpStack’s five working.
What does the watchdog do during a long sysupgrade, or in failsafe or recovery? Nothing feeds it in those states. A flash takes about four minutes on this board and the last feed is at most two minutes before it, so it should fit inside ten. Recovery on DC power will not fit. RAK switches the watchdog off over Modbus first (spregread -F), but which register that writes is not known.
Does the watchdog answer, or run at all, on PoE? Whether a silent watchdog can ever be a warning. The bench unit has one and was silent on PoE, so for now silence is never a fault.

One owner, one request at a time. The watchdog and the battery pack are slaves of one master, and a second master on the line corrupts both. So one process opens the port, and nothing else does:

  • The MQTT agent’s rak-power op prints the daemon’s record. It does not read the bus.
  • The ModemManager tty hotplug override (boards/common/files/etc/hotplug.d/tty/25-modemmanager-tty) skips the bridge’s tty, so ModemManager never writes AT probes onto the bus.

The port is found by USB path, not by number. On a CV2 the EG95 modem takes four ttyUSB nodes. Which of the modem and the CH340 is ttyUSB0 depends on which driver binds first, and ChirpStack hardcodes /dev/ttyUSB0. The daemon looks under /sys/bus/usb/devices/1-1.4/1-1.4:1.0/ for the tty, checks the vid:pid, and looks again whenever the node goes away. On the bench Pro the CH340 was ttyUSB4.

Only boards that declare it. PULZR_RAK_POWER_USB and PULZR_RAK_POWER_ID live in boards/common/rakwireless.conf, empty by default, and only boards/rak7289cv2/board.conf sets them and adds PACKAGES_RAK_POWER. No other board carries the package, and the init script starts nothing on a board that does not set the port.

The watchdog comes first.

  • It is fed every 120 s. Once it has answered, a failure is retried every 15 s. The watchdog allows ten minutes, so this survives four missed feeds where ChirpStack’s cron survives one.
  • A tick always tries the watchdog before the battery pack.
  • The init script starts the daemon at START=15, with no gate on adoption, network or broker, and respawns it with no limit.

Watchdog states.

State Meaning
fed The last feed succeeded.
failing It answered earlier in this boot and has stopped. On DC power the unit is ten minutes from a power cut. Retried every 15 s, logged as an error, reported at once.
absent It has never answered this boot. An older revision has no watchdog, and one that has it may not answer on PoE (the bench unit does not), so this is not a fault: logged once at info, never reported on its own, and still fed every 2 minutes in case it is there. While the port itself is missing, as it is for the first seconds of boot, it is retried every 10 s instead.

The battery pack is an optional accessory, so one that has never answered is absent and is asked every 30 minutes, in case one is fitted later, instead of every 5. Pulzar Cloud shows nothing for a pack that is absent.

Telling Pulzar Cloud. The record rides pulzr-claim report as facts.rakPower, the same report that carries the rest of the gateway’s health. That report goes every six hours. The daemon also asks for one when the summary changes:

  • the port state
  • the watchdog state
  • the battery pack’s state, or its charge crossing 50, 20 or 10 %

Those reports are at most one every five minutes. The first settled state is not one of them, because the report loop’s own first report goes 30 seconds after boot. A voltage that drifts waits for the six-hourly report, or for Refresh on the gateway page, which asks over MQTT.

Times are ages. The record says fedAgo, not a timestamp. The gateway’s clock is 1970 until NTP answers, and the watchdog is fed from the first seconds of boot. Pulzar Cloud counts back from when the record arrived.

The task was a maintained Modbus library rather than hand-rolled RTU framing and CRC. The one OpenWrt packages is libmodbus 3.1.8, and its bindings are:

Binding Size on mipsel Usable here
lua-libmodbus 0.7 8 KB, plus liblua (64 KB) Yes
python3-libmodbus needs python3-light, megabytes No: does not fit a 16 MB board
ucode none exists No: writing one means compiling a package per architecture, which this feed does not do

So the daemon is Lua 5.1. libmodbus does the framing, CRC, timeouts and error recovery (ERROR_RECOVERY_PROTOCOL: flush and wait out the timeout after a bad frame). The rest is also OpenWrt packages: lua-cjson writes the record, and libubox-lua’s uloop runs the timer. With kmod-usb-serial-ch341, which the package depends on, it adds about 100 KB compressed, to the RAK7289CV2 image only.

Release images strip whole-line -- comments from Lua files (scripts/strip-lua-comments.py, called from harden-payload.sh). They do not compile Lua to bytecode. That would need luac from OpenWrt’s exact Lua build, whose LNUM patch changes the bytecode format.

Terminal window
pulzr-rak-power status # the record, ages in seconds
logread -e rak-power # port, watchdog and battery state changes

scripts/test-rak-power.sh runs two halves.

Unit half. It drives the scheduler and the battery arithmetic against a fake bus whose answers and clock the test chooses. It covers:

  • feed timing and retries
  • failing versus absent, and a silent watchdog fed without an alarm
  • a port that is missing, breaks or renumbers
  • the battery pack’s register order, a missing pack, and report bands
  • report rate-limiting
  • finding the tty in a fake sysfs

Simulated half. It builds lua-libmodbus 0.7 from the tarball whose hash is OpenWrt’s PKG_HASH, and links it to the host’s libmodbus. It then talks over a pty pair to a pymodbus power board (scripts/rak-power-test/simulator.py), an independent framer and decoder. It checks:

  • the watchdog write the simulated power board receives
  • the battery values over the wire
  • a watchdog that falls silent mid-run
  • the daemon’s own entry point feeding the watchdog

RAK_POWER_UNIT_ONLY=1 runs only the unit half, on a machine without a C compiler or pymodbus.

What neither half covers is the hardware itself: the CH340 under this kernel, real line timing, and the power board’s own behaviour.