Host requirements and non-KVM compatibility

[TSnake41]

Hello,

In the XCP-ng Project (Xen), we are looking to support at some point MxGPU.
I would like to know a bit more on how MxGPU works, and how it expects VFIO/IOMMU subsystems to behave, as things are a bit more complicated in our case with Dom0 being a service VM, and we are able to passthrough individual PCI VF but performing specific VFIO DMA mappings is much trickier (*).

Some of the questions I have :

• Which VFIO features are used ? Is VFIO_IOMMU_TYPE1 support expected (i.e are arbitrary iova->physical mappings created) ?
• Are vGPU's DMA requests emitted through SR-IOV VF (hence any form of PCI Passthrough of the VF should work including things happening on behalf of Linux) ?

Thanks

(*) Yet there is a giant WIP topic on support for IOMMU and VFIO in Dom0, at least for non-passthrough use cases.

[sekhatir]

Hello,
I apologize for the delay in my response.

Regarding your questions about MxGPU and its interaction with VFIO/IOMMU subsystems, I can confirm that VFIO_IOMMU_TYPE1 support is indeed supported. As for your second question, vGPU's DMA requests are indeed emitted through SR-IOV VF, so any form of PCI passthrough of the VF should function correctly, including operations happening on behalf of Linux.

Thank you for your patience and understanding.
Best regards,