Re: [Xen-devel] [PATCH RESEND 03/14] libxc: Add placeholders for ACPI tables blob and size

[Shannon Zhao <zhaoshenglong@xxxxxxxxxx>]

On 2016/6/7 20:02, Julien Grall wrote:
> Hello Shannon,
> 
> On 07/06/16 12:42, Julien Grall wrote:
>>>>>> Is there any particular reason to expose the list of the tables
>>>>>> outside
>>>>>> of the building code?
>>>>>>
>>>>>> I would provide a single buffer with all the tables inside.
>>>>>> Similar to
>>>>>> what you did for building the tables in the hypervisor.
>>>>> When it loads these tables to guest memory space, it needs to update
>>>>> the
>>>>> entries (pointing to other tables) of XSDT and also the
>>>>> xsdt_physical_address in RSDP.
>>>>
>>>> Why can't we load the ACPI tables at an hardcoded place in the memory
>>>> (for instance always at the beginning of the RAM)?
>>>>
>>> I think it's more reasonable to let the codes dynamically compute where
>>> it should put these tables at like what it does for the devicetree blob.
>>>
>>> And to an hardcoded place, can you make sure that kind of place is
>>> always available and not used by others?
>>
>> Yes, the toolstack is in charge of the memory layout. So it can ensure
>> that no-one else is using this region.
>>
>> My concern is, based on you patch #13, the ACPI tables are allocated
>> just after all the other modules. However, they cannot be relocated by
>> the kernel because they contain physical address. So they have to stay
>> in place for all the life of the domain.
>>
>> We should put them in a place where it will not impact the memory
>> allocation of the guest. The start of the RAM is a good place for that.
> 
> I though a bit more on this suggestion. If the ACPI tables are put at
> the beginning of the RAM, a guest may not be able to use super page.
> 
> I would suggest to move the ACPI table out of the real RAM to avoid any
> potential issue with the kernel memory allocation.
> 
> For instance we could define a IPA range to be use for ACPI (e.g
> 0x20000000 - 0x20200000) and expose to the guest using the ACPI_NVS type
> in the UEFI memory map.
> 
> Any opinion on this?
No, this will not work. UEFI will control the memory map by itself.

-- 
Shannon


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel