Re: [PATCH 09/12] mm/khugepaged: retract_page_tables() without mmap or vma lock

[Jason Gunthorpe]

On Thu, Jun 01, 2023 at 12:18:43AM +0200, Jann Horn wrote:

> 3. We have to *serialize* with page table walks performed by the
> IOMMU. We're doing an RCU barrier to synchronize against page table
> walks from the MMU, but without an appropriate mmu_notifier call, we
> have nothing to ensure that we aren't yanking a page table out from
> under an IOMMU page table walker while it's in the middle of its walk.
> Sure, this isn't very likely in practice, the IOMMU page table walker
> is probably pretty fast, but still we need some kind of explicit
> synchronization to make this robust, I think.

There is another thread talking about this..

Broadly we are saying that we need to call mmu ops invalidate_range at
any time the normal CPU TLB would be invalidated.

invalidate_range will not return until the iommu HW is coherent with
the current state of the page table.

Jason