In message <20060303211035.GA943@antioche.eu.org>Manuel Bouyer writes

>Ha OK, so it's in the driver itself. I though we were talking about the
>handler's return value ...

That's yet not entirely clear. Or at least, not to me.

I've run multiple bge devices in several systems, under sustained
overload conditions, for *days* on end, without seeing anything even
vaguely like Thor's problems.

I have a hard time buying that there's a bug like that in bge_intr()
which is exercised by Thor's workloads, but not by mine.