Re: [PATCH 01/10] irq: move some interrupt arch_* functions into struct irq_chip.

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, 24 Mar 2010, Ian Campbell wrote:
> On Wed, 2010-03-24 at 17:44 +0000, Thomas Gleixner wrote: 
> > On Wed, 24 Mar 2010, Ian Campbell wrote:
> > 
> > > On Mon, 2010-03-22 at 10:19 +0000, Thomas Gleixner wrote:
> > > > On Sun, 21 Mar 2010, Yinghai Lu wrote:
> > > > 
> > > > > From: Ian Campbell <ian.campbell@citrix.com>
> > > > > 
> > > > > Move arch_init_copy_chip_data and arch_free_chip_data into function
> > > > > pointers in struct irq_chip since they operate on irq_desc->chip_data.
> > > > 
> > > > Not sure about that. These functions are solely used by x86 and there
> > > > is really no need to generalize them.
> > > 
> > > I thought the idea of struct irq_chip was to allow the potential for
> > > multiple IRQ controllers in a system? Given that it seems that struct
> > > irq_desc->chip_data ought to be available for use by whichever struct
> > > irq_chip is managing a given interrupt. At the moment this is not
> > > possible because we step around the abstraction using these arch_*
> > > methods.
> > 
> > Right, but you have exactly _ONE_ irq_chip associated to an irq_desc,
> > but that same irq_chip can be associated to several irq_descs. So
> > irq_desc->data is there to provide data for the irq_chip functions
> > depending on what irq they handle (e.g. base_address ...). 
> > 
> > irq_desc->chip_data is set when the irq_chip is assigned to the
> > irq_desc.
> > 
> > So there is no point in having functions in irq_chip to set
> > irq_desc->chip_data.
> 
> So how do you know which is the appropriate irq_chip specific function
> to call given an irq_desc that you want to copy/free/migrate? The
> contents of the chip_data pointer will take different forms for
> different irq_chips. The way the generic code is currently structured it
> appears you can't (or at least don't) just do a shallow copy by copying
> the irq_desc->chip_data pointer itself -- you need to do a deep copy
> using a function which understands that type of chip_data.

The design of sparse_irq or to be honest the lack of design grew that
crap and it's not only this detail which is a nightmare. That pointer
should probably be simply copied. Either that or if the chip data
require to be node bound we need something along the line:

struct sparse_irq_chip_data {
       void *data;
       void (*copy)(...);
       void (*free)(...);
};

and a corresponding field in irq_desc.

I'm looking into sparse_irq right now anyway because it has other way
more serious short comings.

> How is this operation different to having pointers in irq_chip for
> enabling/disabling/masking interrupts for each irq_chip?

Because that's the purpose of the irq_chip perhaps ?

Thanks,

	tglx