Re: [PATCH] vhost: Unbreak SMMU and virtio-iommu on dev-iotlb support

[Auger Eric]

Hi,

[Adding David and Greg in CC]


On 2/8/21 7:37 PM, Peter Xu wrote:
> On Mon, Feb 08, 2021 at 11:21:23AM +0800, Jason Wang wrote:
> 
> [...]
> 
>>> I'm not sure I remember it right, but we seem to have similar discussion
>>> previously on "what if the user didn't specify ats=on" - I think at that 
>>> time
>>> the conclusion was that we ignore the failure since that's not a valid
>>> configuration for qemu.
>>
>>
>> Yes, but I think I was wrong at that time.
> 
> I can't say you're wrong - I actually still agree with you that at least
> there's a priority of things we'd do, and this one is not extremely important
> if that's not a major use case (say, if you will 100% always suggest an user 
> to
> use ats=on for a viommu enabled vhost).
> 
>>>
>>> The other issue I'm worried is (I think I mentioned it somewhere, but just 
>>> to
>>> double confirm): I'd like to make sure SMMU and virtio-iommu are the only 
>>> IOMMU
>>> platform that will use vhost.
>>
>>
>> For upstream, it won't be easy :)
> 
> Sorry I definitely didn't make myself clear... :)
> 
> To be explicit, does ppc use vhost kernel too?  Since I know at least ppc has
> its own translation unit and its iommu notifier in qemu, so I'm unsure whether
> the same patch would break ppc too, because vhost could also ignore all UNMAP
> sent by the ppc vIOMMU.
> 
>>
>>
>>>    Otherwise IIUC we need to fix those vIOMMUs too.
>>
>>
>> Right, last time I check AMD IOMMU emulation, it simply trigger device IOTLB
>> invalidation during IOTLB invalidation which looks wrong.
> 
> I did quickly grep IOMMU_NOTIFIER_UNMAP in amd_iommu.c and saw nothing. It
> seems amd iommu is not ready for any kind of IOMMU notifiers yet.

for context, we experienced a regression with vsmmuv3/vhost and
virtio-iommu/vhost integration. We wondered whether the ppc viommu is
able to protect vhost devices and if this relies on legacy
IOMMU_NOTIFIER_UNMAP notifiers. ie. vhost does not register this
notifier anymore but instead register dev-iotlb unmap notifier.

Thanks

Eric
> 
> Thanks,
>