On Wed, 30 Mar 2011, Linus Torvalds wrote:

> On Wed, Mar 30, 2011 at 9:46 AM, Thomas Gleixner <tglx@linutronix.de> wrote:
> > On Wed, 30 Mar 2011, Thomas Gleixner wrote:
> >> On Wed, 30 Mar 2011, Linus Torvalds wrote:
> >> >
> >> > which just makes me go "Somebody is really really confused".
> >> >
> >> > The whole thing may mean that both handler_data and chip_data contains
> >> > the right thing, but it still makes me go "WTF?".
> >> >
> >> > Which way should I resolve it?
> >>
> >> get_irq_desc_data() maps to irq_desc_get_handler_data() so Richards
> >> resolution is correct even if the other might work as well.
> >
> > Older code which was not using any accessors has:
> >
> >      desc->handler_data
> >
> > But yeah, the init stuff is confusing as hell.
> 
> So I took the handler_data version, but I do think the chip_data one
> seems to make more conceptual sense. It would be good if somebody who
> can actually test that code and knows all the details of the
> particular irq controller could just take a look, and decide on using
> one or the other, and not both.

I think I figured it out. The cascade interrupts do not set chip_data,
they set handler data in qe_ic_init().

The interrupts which are demultiplexed by the cascade handlers set
chip_data on different irq_desc instances. So Ben's resolution would
have resulted in a NULL pointer dereference.

Thanks,

	tglx