On Fri, Dec 30, 2011 at 8:30 AM, Rob Herring <firstname.lastname@example.org> wrote:
> On 12/14/2011 08:32 PM, David Daney wrote:
>> From: David Daney <email@example.com>
>> Back in early Nov. I send the first version of this patch set. Now
>> things are heating up again in the world of irq_domain, so I wanted to
>> try to get some closure on the issues I had. The Octeon patch is
>> included here to show how I am using irq_domain, but is part of a much
>> larger effort to merge Octeon device tree support.
>> The basic problem I am attempting to solve is using irq domains when
>> there is a 'non-linear' mapping of hwirq <--> irq within a domain.
>> Octeon has a single set of irq numbers that is used across two
>> different implementations of the interrupt controller as well as more
>> than 10 different SOCs all which use different subsets of the irq
>> number space. The result is that the hwirq to irq mapping function
>> contains many gaps and discontinuities, it is really quite random.
>> The existing irq domain infrastructure assumes a continuous linear
>> mapping of hwirq to irq that can be encapsulated by the irq_base,
>> hwirq_base and nr_irq elements of struct irq_domain. This is not
>> suitable for the Octeon implementation.
>> The gist of my change is to add an optional iterator function to
>> irq_domain_ops which knows how to iterate over the irq numbers in a
>> given domain. For simple linear domains (those currently supported),
>> we iterate using the current method based on irq_base, hwirq_base and
>> Summary of the patches:
>> 1) Get rid of some unused code to make subsequent changes simpler.
>> 2) Cleanup the data type used by various hwirq functions and users.
>> 3) Add the irq iterator, and fix up the ARM GIC code to use it instead
>> of the current irq_domain_for_each_irq().
>> 4) Add the Octeon users of the interface.
>> In an earlier exchange, Rob Herring had said:
>> ... Handling sparse irqs is a potentially common problem, so we
>> should address that in the core irqdomain code.
>> Which is what this patch set is doing.
>> There was a suggestion that perhaps having .to_irq() return a magic
>> value if there was no mapping would also work. However I prefer this
>> approach as it separates the concepts of iteration and mapping of irq
>> Please comment.
> Can we first have a patch that just allows irq domains to be enabled on
> MIPS. It collides because of multiple versions of irq_create_of_mapping.
I'm working on this. I made some poor decisions when first
implementing irq_domain which have made it difficult to bring into
sync with powerpc. Right now I'm working on a series to replace the
new irq_domain with the existing (working-for-years) powerpc code and
converting over all the users. Basically what I should have done in
the first place.
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.