[libvirt] [qemu RFC] qapi: add "firmware.json"

[Laszlo Ersek]

On 04/10/18 09:33, Thomas Huth wrote:
> On 09.04.2018 18:50, Laszlo Ersek wrote:
>> On 04/09/18 10:19, Gerd Hoffmann wrote:
>>>>> +{ 'enum' : 'SystemFirmwareType',
>>>>> +  'data' : [ 'bios', 'slof', 'uboot', 'uefi' ] }
>>>>
>>>> The naming here is quite a bad mixture between firmware interface
>>>> ('bios', 'uefi') and firmware implementations ('slof', 'uboot'). There
>>>> could be other implementations of BIOS or UEFI than SeaBIOS and EDK2 ...
>>>> so I'd suggest to rather name them 'seabios' and 'edk2' here instead.
>>>
>>> uboot for example implements uefi unterfaces too (dunno how complete,
>>> but reportly recent versions can run uefi shell and grub just fine).
>>
>> Indeed: when I was struggling with this enum type and tried to look for
>> more firmware types to add, my googling turned up the "UEFI on Top of
>> U-Boot" whitepaper, from Alex and Andreas :)
>>
>> Again, this reaches to the root of the problem: when a user creates a
>> new domain, using high-level tools, they just want to tick "UEFI". (Dan
>> has emphasized this to me several times, so I think I get the idea by
>> now, if not the full environment.) We cannot ask the user, "please be
>> more specific, do you want UEFI from edk2, or UEFI on top of U-Boot?"
> 
> But you are designing a rather low-level interface here, which should
> IMHO rather be precise than fuzzy. So should this "just want to tick
> UEFI" rather be handled in the high-level tools instead?
> 
> Alternatively, what about providing some kind of "alias" or "nickname"
> setting here, too? So the EDK2 builds would get
> SystemFirmwareType="edk2" and "SystemFirmwareAlias="uefi" for example.

I hope I understand you right -- I think your suggestion ties in with my
other email I just sent in this thread. So, we could tell libvirtd,
"this firmware is of type 'UEFI', and you must use the 'ovmf_smm'
mapping method to run it, with this file or that file as varstore template".

We could even describe the parameters for this or that mapping method
structurally in the schema (in a discriminated union in QAPI JSON, or in
an XSD choice element). For example, "ovmf" and "ovmf_smm" would both
take "OvmfSplitFileOptions" -- a list of single varstore template files
with feature enum contants attached  --, while "SeaBiosOptions" would be
an empty structure.

I feel the key question here is whether we are allowed to directly
reference a mapping method we know libvirt implements. If we are, that
makes things a lot clearer (and easier, I should hope).

Thanks
Laszlo