[Intel-gfx] [Announcement] 2015-Q3 release of XenGT - a Mediated Graphics Passthrough Solution from Intel

[Jike Song]

On 11/21/2015 01:25 AM, Alex Williamson wrote:
> On Fri, 2015-11-20 at 08:10 +0000, Tian, Kevin wrote:
>>
>> Here is a more concrete example:
>>
>> KVMGT doesn't require IOMMU. All DMA targets are already replaced with
>> HPA thru shadow GTT. So DMA requests from GPU all contain HPAs.
>>
>> When IOMMU is enabled, one simple approach is to have vGPU IOMMU
>> driver configure system IOMMU with identity mapping (HPA->HPA). We
>> can't use (GPA->HPA) since GPAs from multiple VMs are conflicting.
>>
>> However, we still have host gfx driver running. When IOMMU is enabled,
>> dma_alloc_*** will return IOVA (drvers/iommu/iova.c) in host gfx driver,
>> which will have IOVA->HPA programmed to system IOMMU.
>>
>> One IOMMU device entry can only translate one address space, so here
>> comes a conflict (HPA->HPA vs. IOVA->HPA). To solve this, vGPU IOMMU
>> driver needs to allocate IOVA from iova.c for each VM w/ vGPU assigned,
>> and then KVMGT will program IOVA in shadow GTT accordingly. It adds
>> one additional mapping layer (GPA->IOVA->HPA). In this way two
>> requirements can be unified together since only IOVA->HPA mapping
>> needs to be built.
>>
>> So unlike existing type1 IOMMU driver which controls IOMMU alone, vGPU
>> IOMMU driver needs to cooperate with other agent (iova.c here) to
>> co-manage system IOMMU. This may not impact existing VFIO framework.
>> Just want to highlight additional work here when implementing the vGPU
>> IOMMU driver.
>
> Right, so the existing i915 driver needs to use the DMA API and calls
> like dma_map_page() to enable translations through the IOMMU.  With
> dma_map_page(), the caller provides a page address (~HPA) and is
> returned an IOVA.  So unfortunately you don't get to take the shortcut
> of having an identity mapping through the IOMMU unless you want to
> convert i915 entirely to using the IOMMU API, because we also can't have
> the conflict that an HPA could overlap an IOVA for a previously mapped
> page.
>
> The double translation, once through the GPU MMU and once through the
> system IOMMU is going to happen regardless of whether we can identity
> map through the IOMMU.  The only solution to this would be for the GPU
> to participate in ATS and provide pre-translated transactions from the
> GPU.  All of this is internal to the i915 driver (or vfio extension of
> that driver) and needs to be done regardless of what sort of interface
> we're using to expose the vGPU to QEMU.  It just seems like VFIO
> provides a convenient way of doing this since you'll have ready access
> to the HVA-GPA mappings for the user.
>
> I think the key points though are:
>
>        * the VFIO type1 IOMMU stores GPA to HVA translations
>        * get_user_pages() on the HVA will pin the page and give you a
>          page
>        * dma_map_page() receives that page, programs the system IOMMU and
>          provides an IOVA
>        * the GPU MMU can then be programmed with the GPA to IOVA
>          translations

Thanks for such a nice example! I'll do my home work and get back to you
shortly :)

>
> Thanks,
> Alex
>

--
Thanks,
Jike