Re: [PATCH 5/5] xen/arm: Only enable physical IRQs when the guest asks

[Julien Grall <julien.grall@linaro.org>]

On 06/26/2013 11:55 AM, Ian Campbell wrote:

> On Tue, 2013-06-25 at 18:38 +0100, Julien Grall wrote:
>>>> @@ -719,11 +731,18 @@ int gic_route_irq_to_guest(struct domain *d, const struct dt_irq *irq,
>>>>      unsigned long flags;
>>>>      int retval;
>>>>      bool_t level;
>>>> +    struct pending_irq *p;
>>>> +    /* XXX: handler other VCPU than 0 */
>>>
>>> That should be something like "XXX: handle VCPUs other than 0".
>>>
>>> This only matters if we can route SGIs or PPIs to the guest though I
>>> think, since they are the only banked interrupts? For SPIs we actually
>>> want to actively avoid doing this multiple times, don't we?
>>
>>
>> Yes. Here the VCPU is only used to retrieved the struct pending_irq.
> 
> Which is per-CPU for PPIs and SGIs. Do we not care about PPIs here?

I don't see reason to route physical PPIs. Is it possible to have a
device (other than the timer and the GIC) with PPIs?

>>>
>>> For the banked interrupts I think we just need a loop here, or for
>>> p->desc to not be part of the pending_irq struct but actually part of
>>> some separate per-domain datastructure, since it would be very weird to
>>> have a domain where the PPIs differed between CPUs. (I'm not sure if
>>> that is allowed by the hardware, I bet it is, but it would be a
>>> pathological case IMHO...).
>>
>>> I think a perdomain irq_desc * array is probably the right answer,
>>> unless someone can convincingly argue that PPI routing differing between
>>> VCPUs in a guest is a useful thing...
>>
>>
>> Until now, I didn't see PPIs on other devices than the arch timers and
>> the GIC. I don't know if it's possible, but pending_irq are also banked
>> for PPIs, so it's not an issue.
>>
>> The issue is how do we link the physical PPI to the virtual PPI? Is a
>> 1:1 mapping. How does Xen handle PPI when a it is coming on VCPUs which
>> doesn't handle it (for instance a domU)?
> 
> How do you mean?


My sentence wasn't clear.

As for the arch timer, which is using PPIs, routing theses interrupts
require some code to support the device in Xen, mainly to save/restore
the context when the VCPU is moved.

Can we assume that Xen will never route PPIs to a guest?

-- 
Julien