The root cause of this bug lies in the fact that the XICS interrupt controller
uses a radix tree for its reverse irq mapping and that we cannot allocate the tree
nodes (even GFP_ATOMIC) with preemption disabled.

Is that yet another caes of -rt changing some basic kernel semantics ?

  In fact, we have 2 nested preemption disabling when we want to allocate
a new node:

  - setup_irq() does a spin_lock_irqsave() before calling xics_startup() which
    then calls irq_radix_revmap() to insert a new node in the tree

  - irq_radix_revmap() also does a spin_lock_irqsave() (in irq_radix_wrlock())
    before the radix_tree_insert()

  The first patch moves the call to irq_radix_revmap() from xics_startup() out to
xics_host_map_direct() and xics_host_map_lpar() which are called with preemption
enabled.

I suppose that would work.

  The second patch is a little more involved in that it takes advantage of
the concurrent radix tree to simplify the locking requirements and allows
to allocate a new node outside a preemption disabled section.

  I just hope I've correctly understood the concurrent radix trees semantic
and got the (absence of) locking right.

Hrm, that will need some scrutinity.

Thanks for looking at this.

Cheers,
Ben.