Re: [PATCH 0/5] Add new headers for Hyper-V Dom0

[Nuno Das Neves <nunodasneves@linux.microsoft.com>]

On 10/31/2024 12:05 PM, Easwar Hariharan wrote:
> On 10/3/2024 12:50 PM, Nuno Das Neves wrote:
>> To support Hyper-V Dom0 (aka Linux as root partition), many new
>> definitions are required.
>>
>> The plan going forward is to directly import headers from
>> Hyper-V. This is a more maintainable way to import definitions
>> rather than via the TLFS doc. This patch series introduces
>> new headers (hvhdk.h, hvgdk.h, etc, see patch #3) directly
>> derived from Hyper-V code.
>>
>> This patch series replaces hyperv-tlfs.h with hvhdk.h, but only
>> in Microsoft-maintained Hyper-V code where they are needed. This
>> leaves the existing hyperv-tlfs.h in use elsewhere - notably for
>> Hyper-V enlightenments on KVM guests.
>>
>> An intermediary header "hv_defs.h" is introduced to conditionally
>> include either hyperv-tlfs.h or hvhdk.h. This is required because
>> several headers which today include hyperv-tlfs.h, are shared
>> between Hyper-V and KVM code (e.g. mshyperv.h).
>>
>> Summary:
>> Patch 1-2: Cleanup patches
>> Patch 3: Add the new headers (hvhdk.h, etc..) in include/hyperv/
>> Patch 4: Add hv_defs.h and use it in mshyperv.h, svm.h,
>>          hyperv_timer.h
>> Patch 5: Switch to the new headers, only in Hyper-V code
>>
>> Nuno Das Neves (5):
>>   hyperv: Move hv_connection_id to hyperv-tlfs.h
>>   hyperv: Remove unnecessary #includes
>>   hyperv: Add new Hyper-V headers
>>   hyperv: Add hv_defs.h to conditionally include hyperv-tlfs.h or
>>     hvhdk.h
>>   hyperv: Use hvhdk.h instead of hyperv-tlfs.h in Hyper-V code
>>

Hi Easwar, thanks for the questions. I will attempt to clarify.
> 
> What is the model for Hyper-V code that has both guest and host roles
> where the corresponding hypercalls are available for both? As I
> understand it, those are supposed to be in hvgdk*.h.
>

It's true that the naming of the files implies hvgdk*.h is for guests,
and hvhdk*.h (which includes hvgdk*.h), is for hosts/dom0. But I would
only take that as a rough guide.

The real reason for keeping these names is to make it a easier to copy
and maintain the definitions from the Windows code into Linux, by
keeping the provenance of exactly where they came from.

> For a specific example, IOMMU hypercalls can operate on stage 2 or stage
> 1 translations depending on the role of the (hyper) caller and the input
> values provided. Should a driver using these hypercalls import both
> hvhdk* and hvgdk*? What about hyperv-tlfs?
>

I'd recommend importing hvhdk.h since it contains everything you need
(including hvgdk*.h).

The goal of this patchset is to move away from hyperv-tlfs.h, because by
definition it should only contain definitions from the TLFS document.

> Patches 4 and 5 seem to draw a bright line between host and guest roles
> while the reality is more gray. Please do correct me if I'm wrong here,
> perhaps the picture would be clearer if Stas' suggestion of a new header
> file is implemented.
> 

Patches 4 and 5 introduce the new headers in a way that avoids any
potential impact on KVM and other non-Microsoft-maintained code.

The 'line' is not between guest and host, but between Microsoft-maintained
and non-Microsoft-maintained code.

Thanks,
Nuno

> Thanks,
> Easwar