Where is the code that the subject describes?
I have two problems here.
- I don't see anything handling the hotplug case, and forcing us to
physical mode.
- Ingo's other patch asserts that hotplug should be made to handle
logical deliver mode.
With logical deliver mode the experimental evidence is that the
destination cpu is a hint, and you can arrive at a cpu that is not
in your cpu mask. Now I only saw that problem on hyperthreaded cpus
but we didn't have the code enabled long enough to see it in other
cases.
Maybe if I can finish getting irq migration back into process context,
and someone verifies that the cpu disable in the hotplug path actually
disables the cpu/hyperthread instead of sitting in a hlt loop. We
won't have a problem. But broadcast ipis and irqs that don't go where
you tell them to are things we need to be very careful with.
Eric
-
