Re: [pve-devel] [PATCH docs v4 4/5] added vIOMMU documentation

[Markus Frank <m.frank@proxmox.com>]

On 1/16/23 11:00, Wolfgang Bumiller wrote:
> On Fri, Jan 13, 2023 at 02:31:36PM +0100, Markus Frank wrote:
>>
>>
>> On 1/13/23 11:09, Wolfgang Bumiller wrote:
>>> On Fri, Nov 25, 2022 at 03:08:56PM +0100, Markus Frank wrote:
>>>> Signed-off-by: Markus Frank <m.frank@proxmox.com>
>>>> ---
>>>>    qm-pci-passthrough.adoc | 25 +++++++++++++++++++++++++
>>>>    1 file changed, 25 insertions(+)
>>>>
>>>> diff --git a/qm-pci-passthrough.adoc b/qm-pci-passthrough.adoc
>>>> index fa6ba35..7ed4d49 100644
>>>> --- a/qm-pci-passthrough.adoc
>>>> +++ b/qm-pci-passthrough.adoc
>>>> @@ -389,6 +389,31 @@ Example configuration with an `Intel GVT-g vGPU` (`Intel Skylake 6700k`):
>>>>    With this set, {pve} automatically creates such a device on VM start, and
>>>>    cleans it up again when the VM stops.
>>>> +[[qm_pci_viommu]]
>>>> +vIOMMU
>>>> +~~~~~~
>>>> +
>>>> +vIOMMU enables the option to passthrough pci devices to Level-2 VMs
>>>> +in Level-1 VMs via Nested Virtualisation.
>>>> +
>>>> +Host-Requirement: Set `intel_iommu=on` or `amd_iommu=on` depending on your
>>>> +CPU.
>>>
>>> And by "CPU" you mean kernel command line? ;-)
>>
>> Host-Requirement: Add `intel_iommu=on` or `amd_iommu=on`
>> depending on your CPU to your kernel command line.
>>
>> like this?
>>>
>>>> +
>>>> +VM-Requirement: For both Intel and AMD CPUs you will have to set
>>>> +`intel_iommu=on` as a Linux boot parameter in the vIOMMU-enabled-VM, because
>>>> +Qemu implements the Intel variant.
>>>
>>> ^ As mentioned, there does appear to be an amd_iommu device in the qemu
>>> code, so would the amd variant work?
>>>
>>> In my reply to the code patch I mentioned checking the host arch. But if
>>> you say we can use intel_iommu on AMD as well, I'd say, if both work,
>>> give the user a choice, otherwise we can of course just stick to the one
>>> that works ;-)
>>
>> intel_iommu works better on my AMD CPU than amd_iommu ;)
> 
> Can you define "better"?
> My main concern is that if we don't give users the option to choose, the
> only data point we have is yours ;-)
> If we explicitly mention that you can use one on the other in the docs,
> people can try it themselves and maybe we'll see some feedback on the
> forums etc.
> 
> However, I'm fine with a patch for only the intel version for now as we
> can always add an option later.
> 
>> Moreover it adds an extra AMDVI-PCI device that is using the first pci address.
>> `kvm: -device VGA,id=vga,bus=pcie.0,addr=0x1: PCI: slot 1 function 0 not available for VGA, in use by AMDVI-PCI,id=(null)`
> 
> For that I'd say, try to add the AMDVI-PCI device manually to an
> explicitly chosen slot. We need to avoid automatically added devices
> like the plague, because moving them later can break live snapshots (and
> windows).
> 
I manually added a AMDVI-PCI device with a explicitly chosen slot but
amd-iommu still adds an extra AMDVI-PCI device.
I do not see any way to prevent this or to change the address
of the additionally added AMDVI-PCI device (from amd-iommu).

I think amd-iommu is very impractical if we cannot set this slot manually.

>>
>> I cannot find any good documentation for amd_iommu but it also seems like
>> it has less features.
> 
> Less, or just not configurable? ;-)
> I mean, if it works it works ;-)
> 
>>
>> $ qemu-system-x86_64 -device 'amd-iommu,help'
>> amd-iommu options:
>>    device-iotlb=<bool>    -  (default: false)
>>    intremap=<OnOffAuto>   - on/off/auto (default: "auto")
>>    pt=<bool>              -  (default: true)
>> $ qemu-system-x86_64 -device 'intel-iommu,help'
>> intel-iommu options:
>>    aw-bits=<uint8>        -  (default: 39)
>>    caching-mode=<bool>    -  (default: false)
>>    device-iotlb=<bool>    -  (default: false)
>>    dma-drain=<bool>       -  (default: true)
>>    dma-translation=<bool> -  (default: true)
>>    eim=<OnOffAuto>        - on/off/auto (default: "auto")
>>    intremap=<OnOffAuto>   - on/off/auto (default: "auto")
>>    pt=<bool>              -  (default: true)
>>    snoop-control=<bool>   -  (default: false)
>>    version=<uint32>       -  (default: 0)
>>    x-buggy-eim=<bool>     -  (default: false)
>>    x-pasid-mode=<bool>    -  (default: false)
>>    x-scalable-mode=<bool> -  (default: false)