When moving IRQs between CPUs to distribute the load of IRQ handling,
IRQ vectors are dynamically allocated and de-allocated on the relevant
CPUs.  De-allocation has to happen when certain constraints are met.
If these conditions are not met when first checked, the checking CPU
may send an interrupt to itself, in the expectation that this IRQ will
be delivered only after the condition preventing the cleanup has
cleared.  For two specific IRQ vectors this expectation was violated,
resulting in a continuous stream of self-interrupts, which renders the
CPU effectively unusable.