>-----Original Message-----
>From: Joao Martins <[email protected]>
>Subject: Re: [PATCH v4 08/10] vfio/migration: Fix a check on
>vbasedev->iommu_dirty_tracking
>
>On 30/10/2025 09:10, Duan, Zhenzhong wrote:
>>> From: Avihai Horon <[email protected]>
>>> On 29/10/2025 11:53, Zhenzhong Duan wrote:
>>>
>>> BTW, do you have a real production use case for migration with VFIO
>>> IOMMU type1 dirty tracking? I mean, is the scenario you described in
>>> patch #7 a real use case or you just fixed it for completeness?
>>> If there is no use case, maybe patch #9 is not really needed?
>>
>> patch7 is a real use case, in guest, we switch device from IOMMU domain
>to block domain and see the issue.
>> We need to send accurate unmap notification with actual mapping during
>migration, for both backend.
>>
>
>I think the real question is why you are using type1 backend overall for
>purposes of dirty tracking.
Just because we had set bcontainer->dirty_pages_supported = true, the dirty
tracking
by query through kernel VFIO type1 interface is used during migration, even
though it's 'perpectual'.
>
>type1 dirty tracking just returns everything in the bitmap as 1s. There's no
>actual dirty tracking and we usually call 'perpectual' dirty tracking because
>everything DMA mapped as write is always returned as dirty no matter what
>you
>do. It doesn't look at pagetable neither for the unmap get dirty.
>
>It's only (known) use has been for testing (in the lack of IOMMU HW +
>IOMMUFD)
You mean testing live migration or the legacy VFIO type1 dirty tracking
interface?
If you mean the first, we can force enable it with 'enable-migration=on' and set
bcontainer->dirty_pages_supported = false.
>
>But reading your statement in a different way: you are saying that you use
>*two*
>backends at the same time? Why would you do that?
I mean patch7 is needed no matter legacy backend or IOMMUFD backend.
btw: we do support a qemu cmdline with mixed backend, e.g., two devices,
one backed by legacy backend and the other backed by IOMMUFD. But I'm not mean
that.
Thanks
Zhenzhong
