I made some updates to my patch:
- the cpu bitmaps are removed, the outstanding cpus are tracked by an
- bugfixes for offlining cpus: some irq on/off's were wrong.
- bugfix for offlining a nohz cpu [untested].
- a bugfix the main state machine function: if interrupts arrived after
reading the global state, then the state machine would oops. Now the
local interrupts are disabled early enough. Additionally, BUG_ON's test
for invalid state changes.
- some debug outputs removed.
Offlining a cpu doesn't oops any more, although further testing is needed.