Hello!

This sounds reasonable. However as a start I was thinking at having
hardirq run only the softirq they posted actively, and local_bh_enable
run only the softirq that have been posted by the critical section (not
from hardirqs happening on top of it).

To all that I remember Ingo insisted that the run from local_bh_enable() 
is necessary for softirqs triggered by hardirqs while softirqs are disabled.
Otherwise he observed random delays and redundant scheduler activity
breaking at least tux fast path.

Alexey