> this is impossible if we do hardirq_endlock()+softirq_endlock() with local > IRQs disabled [like my second quick-patch did] ... No need to add > additional expensive spinlocks, bottom-half performance sucks anyway.