Re: [Qemu-devel] [V3 3/4] hw/i386: ACPI table for AMD IO MMU

[David kiarie]

On Thu, Jan 14, 2016 at 7:19 PM, Jan Kiszka <address@hidden> wrote:
> On 2016-01-14 17:09, David kiarie wrote:
>> On Thu, Jan 14, 2016 at 6:42 PM, Jan Kiszka <address@hidden> wrote:
>>> On 2016-01-14 16:39, Michael S. Tsirkin wrote:
>>>> On Thu, Jan 14, 2016 at 03:15:38PM +0300, David kiarie wrote:
>>>>> On Thu, Jan 14, 2016 at 1:09 PM, Michael S. Tsirkin <address@hidden> 
>>>>> wrote:
>>>>>> On Thu, Jan 14, 2016 at 11:04:27AM +0300, David Kiarie wrote:
>>>>>>> Add IVRS table for AMD IO MMU. Also reverve MMIO
>>>>>>
>>>>>> reserve?
>>>>>
>>>>> Yeah, typo.
>>>>>
>>>>>>
>>>>>>> region for IO MMU via ACPI
>>>>>>
>>>>>>
>>>>>> It does not look like you reserve anything.
>>>>>>
>>>>>> Pls add a link to hardware spec (in
>>>>>> the device implementation) so we can check
>>>>>> what does real hardware do.
>>>>>>
>>>>>> If this is it:
>>>>>> http://developer.amd.com/wordpress/media/2012/10/488821.pdf
>>>>>>
>>>>>> then the way that works seems to be by guest
>>>>>> programming the MMIO base.
>>>>>> We should do the same: patch seabios and EFI to do this.
>>>>>
>>>>> Yes, that's the spec.
>>>>>
>>>>> We thought this could be possible via ACPI (without patching BIOS ), no ?
>>>>
>>>> I don't see how. We should do it the way it happens on real hardware.
>>>>
>>>
>>> Doesn't Seabios retrieve certain ACPI fragments from QEMU via a
>>> pv-interface by now?
>>>
>>> Anyway, the question remains where this address comes from: The BIOS,
>>> which then writes it into some hw config register and reports it in
>>> addition via ACPI or the hardware (hard-wired).
>>
>> Will look at patching BIOS.
>
> Scanning through the spec again: As the address is not hard-wired but
> configured via registers in the extended capabilities of the
> corresponding PCI functions of an IOMMU (and also enabled that way!),
> it's up to the BIOS to allocate an appropriate address for a system.

Is there a way to modify IVRS/ACPI table from seabios because the same
address is supposed to be the IVRS table otherwise I'd have to
hard-code it :/

Anyway, will check.

>
> Jan
>
>