Re: vIOMMU - PCI pass through to Layer 2 VMs (Nested Virtualization)

[Eric Auger]

Hi,
On 10/9/23 12:25, Markus Frank wrote:
> Hi Eric,
> 
> thanks for the quick answer.
> 
> On 10/9/23 11:29, Eric Auger wrote:
>> Hi Markus,
>>
>> On 10/9/23 09:06, Markus Frank wrote:
>>> Hello,
>>>
>>> I have already sent this email to qemu-discuss but I did not get a
>>> reply.
>>> https://lists.nongnu.org/archive/html/qemu-discuss/2023-09/msg00034.html
>>> Maybe someone here could help me and reply to this email or the one on
>>> qemu-discuss?
>>>
>>> I would like to pass through PCI devices to Layer-2 VMs via Nested
>>> Virtualization.
>>>
>>> Is there current documentation for this topic somewhere?
>>>
>>> I used these parameters:
>>> -machine ...,kernel-irqchip=split
>>> -device intel-iommu
>>>
>>> With these parameters PCI pass through to L2-VMs worked fine.
>>>
>>>
>>> Now I come to the part where I get confused.
>>>
>>> https://wiki.qemu.org/Features/VT-d#With_Virtio_Devices
>>> Is this documentation relevant for PCI pass through? Do I need DMAR for
>>> virtio devices?
>> If you just want the host assigned devices to be protected by the
>> viommu, you don't need to add iommu_platform=on along with the
>> virtio-pci device>>
>>> And there is also the virtio-iommu device where I also could use the
>>> i440fx chipset.
>>> https://michael2012z.medium.com/virtio-iommu-789369049443
>>
>> you can use virtio-iommu with q35 machine.
> Yes I know. I meant that intel-iommu does not support i440fx and
> virtio-iommu does.

>>>
>>> When adding "-device virtio-iommu-pci" pci pass through also works
>>> but I get "kvm: virtio_iommu_translate no mapping for 0x1002030f000 for
>>> sid=240"
>>> when starting qemu. What could that mean?
>> Normally you shouldn't get any such error. This means there is no
>> mapping programmed by the iommu-driver for this requester id (0x240) and
>> this iova=0x1002030f000. But if I understand correctly this does not
>> prevent your device from working, correct?
> Yes. I didn't notice any problems. How could I find out what the
> requester id 0x240 refers to?
on your guest issue lspci and look at the end points BDF that matches
0x240.
>>>
>>> What do these parameters
>>> "disable-legacy=on,disable-modern=off,iommu_platform=on,ats=on"
>>> actually do? When do I need them and on which virtio devices?
>> you need them if you want your virtio devices to be protected by the
>> viommu. Otherwise the viommu is bypassed.
> Okay, so iommu_platform=on is more of a decision you should make per
> virtio-pci device.
> So simplified the advantage is more isolation and the disadvantage is
> less performance?
yes setting iommu_platform forces the driver to use the DMA API.

Eric
>>>
>>> And which device should I rather use: virtio-iommu or intel-iommu?
>> Both should be working. virtio-iommu is more recent and less used in
>> production than intel-iommu though.
>>
>> Thanks
>>
>> Eric
>>>
>>> Thanks in advance,
>>> Markus
>>>
>>>
>>
>>
>