[PATCH] synchronize_irq needs a barrier

[Benjamin Herrenschmidt]

> I probably need to add this synchronize_irq() logic in dmfe.c too, but I probably do it later,
> I think I am overestimating this race, since most drivers don't do dev->insuspend checks in IRQ handler.
> Maybe even just use free_irq() after all....

Most drivers are probably underestimating the race :-)

free_irq() would work provided that you did the masking on chip before
(and unmask only after request_irq on the way back in). But it's a bit
like using a 10 tons truck to crush an ant... 

Ben.