On 20/12/17 20:04, Paolo Bonzini wrote: > On 20/12/2017 02:47, Alexey Kardashevskiy wrote: >>> - if we foresee having more IOMMU devices in KVM, let's rename >>> KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE to KVM_DEV_VFIO_GROUP_ATTACH_IOMMU and >>> add a new function >>> >>> int iommu_memory_region_get_kvm_fd(IOMMUMemoryRegion *iommu); >>> >>> that requires no object-type check in VFIO. >> >> This is how it started and the comment was that this KVM fd needs to have a >> well defined semantic which I struggle to provide. > > I don't think the definition should be anything more than "it can be > passed to KVM_DEV_VFIO_GROUP_ATTACH_IOMMU". I'll discuss it with Alex > later today.
That was more David's concern :) -- Alexey