[vfio-users] IOMMU page table of QEMU vfio device passthrough

[Alex Williamson]

On Tue, 5 Jun 2018 18:52:43 -0700
Huaicheng Li <huaicheng at cs.uchicago.edu> wrote:

> Hi Alex,
> 
> VFIO does not setup
> > mappings per DMA buffer, it maps all of guest physical memory through
> > the IOMMU so that the device can DMA to any GPA and it's already
> > mapped.  We call this transparent assignment, the guest can perform DMA
> > without any knowledge of or interaction with the host IOMMU.  
> 
> 
> Could you briefly explain how all of guest physical memory is mapped
> through IOMMU.
> I do see a ``vfio_dma_map`` call with `size` almost equal to my guest
> memory size, is
> this the place where IOMMU is instructed to enable such transparent
> assignment you
> mentioned?

I don't know that I have a brief explanation for this, you'll need to
look at the code.  The entry point to vfio is the MemoryListener (which
calls vfio_dma_map) attached to the device AddressSpace, which is the
system memory address space when a guest IOMMU is not present.
RAMBlocks are mapped through MemoryRegions and sections of those come
in through the MemoryListner to be populated through the IOMMU.

> Since the guest driver is issuing IOs with GPA, how does the
> device do
> DMA to the corresponding physical DMA address?

This is why we have an IOMMU.  The guest drivers and the device operate
in the guest physical address space, the IOMMU handles the translation
to the host physical address space.

> Is the GPA->HPA a static
> mapping
> so the HPA can be calculated instead of walking the page tables (which
> those DMA APIs
> do)?

Without a guest IOMMU, the translations are almost entirely static, for
instance the RAM in your computer doesn't spontaneously decide to
relocate itself in the physical address space of your system, neither
does RAM in a VM.  All of guest physical memory is pinned in the host
and mapped through the IOMMU.  Regardless of this, there is no page
table walking to create these mappings.  Each RAMBlock is just a range
of memory with a start and end offset in the guest physical address
space.  Thanks,

Alex