Spin locks and completions are expensive in hard IRQ context and cause
problems with RT kernels.  In RT kernels, both spin locks and
completions can schedule(), so we can't use them in hard irq context.

Move handling code into the irq thread function to avoid that.
Reported-by: Gregoire Gentil <gregoire@alwaysinnovating.com>
Signed-off-by: Luciano Coelho <coelho@ti.com>