On Wed, 16 Sep 2015, Marc Zyngier wrote:

On 16/09/15 02:53, Rafael J. Wysocki wrote:

quoted

I've just seen quite a few bugs where a pointer to something completely invalid
have been silently passed via (void *) which often results in very interesting
breakage (that is really hard to debug for that matter).

I actually tried to prototype this yesterday, and ended up in hell. The
main issue is the point where the generic irqdomain code meets the DT
subsystem (which is basically any interrupt controller, including those
being ACPI driven). The domain_token to of_node path is dead easy, but
you cannot do the reverse conversion, so this would have to spread
around like a cancer. I gave up.

At this stage, I see two options: sticking with (void *) with the risk
of breakage and subtle bugs which we all love to track down (not!), or
do a major U-turn and make device_node a strict requirement for domain
lookup.

After the above experiment, I now see some value in actually keeping
device_node around, using it as the token, and allocating it when
required (the typical case being those interrupt controllers that are
both DT and ACPI). I'll play with it a bit more.

Right. We already have the fwnode_type in struct device_node, so we
can reuse it. If you care about the size of the struct, then you can
make some of the struct members conditional on CONFIG_OF.

Thanks,

	tglx