>When I tried enabling interrupts in these functions (which one _should_
>be able to), the machine fell over running lmbench (which it shouldn't).

Well, I just `#if 0'd all the code that turns off interrupts during
TLB and cache flushes, and ran lmbench, and it didn't even hiccup.
It's conceivable that this had something to do with the
splimp-splclock inversion that I fixed when I added splaudio, but I'm
not sure.  I definitely don't see any bug here, though.