Unreliable HW watchdog?

Hi,

I'm having issues with an unreliable HW watchdog on the Orange Pi Zero. I am using the standard Armbian kernels (tried 4.19.62, 5.3.x, 5.4.x - all the same) and have nowayout=1:

[    4.328286] sunxi-wdt 1c20ca0.watchdog: Watchdog enabled (timeout=16 sec, nowayout=1)

I'm feeding /dev/watchdog every 2 seconds. Superficially it works, for example, if I do the following the WD triggers after 16 sec and the board reboots:

echo c > /proc/sysrq-trigger

However I am finding that in other crash scenarios, the WD reboot sometimes happens and sometimes does not. For example, with a forkbomb :(){ :|:& };: about 70% of the time the WD reboots the board, but 30% of the time the board stays crashed and the WD never reboots it. (I've confirmed that the forkbomb does stop the watchdog feeder, so the WD should trigger a reset.) I have to do a hard power cycle to get it to boot again.

Has anyone else had experience with an unreliable watchdog on these boards? Any more insight or fixes?

-Adrian