Re: pvops-2.6.32 - Interrupt routing problem

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 03/21/2010 02:47 PM, Bastian Blank wrote:
> On Fri, Mar 19, 2010 at 01:13:41PM +0100, Bastian Blank wrote:
>    
>> The real fix could be:
>> - Allow the hypervisor to lock interrupts it uses. Zero would be in it by
>>    default. The interrupt for the used serial interface would be added.
>>    All other pins are free to be programmed by the kernel once.
>> - Don't do register_gsi calls from the initial setup in the kernel if no
>>    ACPI override is present, only setup the rest.
>>      
> Okay, I think I found another problem. Currently the setup looks like
> this:
> - PHYSDEVOP_setup_gsi: set trigger and polarity, unmask pin
> - PHYSDEVOP_map_pirq: map to pirq, set irq handler to guest
>
> If an interrupt fires between this two calles, what happens?
>    

Doesn't sound like a good thing to be doing.  The unmask should probably 
be deferred until everything has been set up to receive an interrupt.

> Usually I would do the setup the other way around. This would also make
> it possible to use the irq table to allow the kernel to only set IO-APIC
> pins mapped to already properly registered interrupts.
>    

Sounds reasonable.

     J