While testing the genirq code on ARM, a condition was found whereby
the Neponset IRQ handler was being re-entered, causing the system
to deadlock.

Under the ARM IRQ code, this would not have been a visible problem
because the "simple" IRQ handling had no re-entrancy protection.

Resolve this by acknowledging the parent interrupt after we mask it
when we are going to handle one of our "special" level-based sources
(from ethernet or USAR chip.)
Signed-off-by: NRussell King <rmk+kernel@arm.linux.org.uk>