I think the mips problem is that the mips switch_exit() is clobbering
the proc0 stack pointer. Instead of running on top of any existing
proc0 kernel stack, the mips switch_exit() is re-setting the kernel
stack pointer as if proc0 had no stack.

When switch_exit()'s stack puking grows beyond the frames into which
uvm_scheduler never returns (main()?), that clobbers uvm_scheduler's
stack.

if we fix mips switch_exit() to slurp the current sp from proc0's
regs, it should Just Work, SW tlb misses and all.

Sure,, we'll need multiple idle threads for one-big-lock {a}symmetric
MP.  Thinking of it as `multiple interrupt stacks' with idle() running
on the per-CPU interrupt stack works for me.