On Tue, 2021-11-16 at 09:41 +0100, Petr Mladek wrote:
[...]

If I wanted to implement a super-reliable panic() I would
use some external device that would cause power-reset when
the watched device is not responding.

They're called watchdog timers.  We have a whole subsystem full of
them:

drivers/watchdog

We used them in old cluster HA systems to guarantee successful recovery
of shared state from contaminated cluster members, but I think they'd
serve the reliable panic need equally well.  Most server class systems
today have them built in (on the BMC if they don't have a separate
mechanism), they're just not usually activated.

James