Re: [PATCH 6.18.y] iommu/vt-d: Gather the unmapped range before freeing its page tables

[Jason Gunthorpe <jgg@ziepe.ca>]

On Thu, Aug 13, 2026 at 12:05:26AM +0000, Jose Fernandez (Anthropic) wrote:
> Under a userspace driver that maps and unmaps DMA buffers through VFIO
> type1 continuously, this shows up as wrong data in device reads and
> writes. An occasional DMAR fault on a mapped IOVA is the only thing in
> the logs. With an Intel DSA engine assigned through vfio-pci, remapping
> a 16 MiB buffer at a fixed IOVA and reading it through the device
> returned data from the wrong pages in 280 of 400 iterations. With a
> fresh IOVA per iteration it never did.
> 
> Add the range to the gather first and splice the freed tables into
> gather->freelist afterwards, so that they are only freed by a sync that
> also invalidates their range.
> 
> Mainline removed this code in v6.19 with commit d373449d8e97
> ("iommu/vt-d: Use the generic iommu page table") and is not affected.

Yeah, I noticed this bug category in iommupt at some point, I didn't
realize there was long existing errors in the drivers.

I don't know what the process is to put a unique patch into stable
(backporting the upstream iommupt based fix is not reasonable), but
this looks correct to me, it follows the pattern iommupt is using and
I have now extensively tested.

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason