[PATCH] irq: convert generic-chip to use irq_domain

[Rob Herring]

On 12/11/2011 07:11 PM, Rob Herring wrote:
> Shawn,
> 
> On 12/11/2011 10:00 AM, Shawn Guo wrote:
>> Hi Rob,
>>
>> I installed the patch and played it a little bit, and I some feedback
>> below.
> 
> Thanks!
> 
>> * It breaks some existing irq_chip_generic users like imx5, which
>>   is currently numbering irq from 0.  For such users, irq_alloc_descs()
>>   will fail because it's asked to search irq #0 starting from irq #1.
>>   Yes, I know numbering irq from 0 is a bad idea and really needs to
>>   get fixed.  But is it your intention to break such users and force
>>   them get fixed with this patch?
> 
> I was trying not to break things and allow the existing irq base.
> 
> My problem is irq 0-15 are unallocated, 16-159 are the GIC. So when I
> try to allocate 8 irqdescs for gpio, I don't really want irq_alloc_descs
> to pick 0-15. Ideally we would just globally reserve 0-15 on ARM, but
> this will cause problems. So starting at 1 was at least slightly better
> than 0.
> 
> I think I'll just skip irq_alloc_descs call when the irq_base is >= 0.
> 
>> * I thought your patch is taking care of all irqdomain stuff inside
>>   irq_chip_generic in a way transparent to its users.  But it really
>>   took me some time to figure out that users still need to populate
>>   the .of_node for irq_domain encapsulated in irq_chip_generic.
> 
> Sorry about that. It would have been evident in my follow on patches for
> pl061 gpio.
> 
>> * If I understand irq_chip_generic correctly, it only supports up to
>>   32 irqs in a chip.  The imx5 tzic supports 128 interrupt lines, and
>>   the current driver implements it as 4 generic irq chips.  But from
>>   hardware and device tree point of view, it's really just one
>>   controller, so we have only one tzic node in dts, and hence we only
>>   have the same one .of_node for 4 irq domains.  I'm afraid such
>>   implementation will break irq_create_of_mapping()?
>>   (For gpio interrupt controller, it should fit perfectly.)
>>
> 
> Yeah, that's a problem that needs addressing. You can override the
> domain ops with a custom dt_translate function. Perhaps this can be done
> generically:
> 
> 	struct irq_chip_generic *gc;
> 
> 	list_for_each_entry(gc, &gc_list, list) {
> 		if (gc->domain != d || d->of_node != controller)
> 			continue;
> 		
> 		// Got a match, fill in data.
> 
> 		return 0;
> 	}
> 
> 	return -EINVAL;

Actually, there's a much simpler fix:

diff --git a/kernel/irq/irqdomain.c b/kernel/irq/irqdomain.c
index 200ce83..3832ac6 100644
--- a/kernel/irq/irqdomain.c
+++ b/kernel/irq/irqdomain.c
@@ -135,6 +135,9 @@ int irq_domain_simple_dt_translate(struct irq_domain *d,
                return -EINVAL;
        if (intsize < 1)
                return -EINVAL;
+       if ((intspec[0] < d->hwirq_base) ||
+           (intspec[0] >= d->hwirq_base + d->nr_irq))
+               return -EINVAL;

        *out_hwirq = intspec[0];
        *out_type = IRQ_TYPE_NONE;


Rob

> 
> There's still the problem that hwirq is not going to be 0-127 for the
> AVIC as you probably want. For that you can initialize hwirq_base for
> each domain. The AVIC code is still free to override parts of the domain
> setup.
> 
> Rob