AFAICT the sole purpose for the hideous x86_64 idle_notifier mess is
to support i7300_idle.  IMO this junk does not belong in IRQ handling,
etc.  Can we redo this to work in some kind of generic way?

I have no idea why it makes sense to twiddle I/O AT registers in the
beginning of whatever IRQ wakes up the CPU.

Note that, if absolutely necessary, the ECX bit 0 MWAIT extension can
be used to reliably execute code before handling interrupts that wake
us from idle.  That is, there could be a real cpuidle driver for that
chip that does:

cli;
poke ioat;
mwait(ecx = 1);
poke ioat;
sti;

Or we could delete the driver entirely.

--Andy