Re: [PATCH v2 1/1] iommu/sva: Invalidate KVA range on kernel TLB flush

[Baolu Lu]

On 7/11/25 16:17, Yu Zhang wrote:
> On Thu, Jul 10, 2025 at 08:26:06AM -0700, Dave Hansen wrote:
> > On 7/10/25 06:22, Jason Gunthorpe wrote:
> > > > Why does this matter? We flush the CPU TLB in a bunch of different ways,
> > > > _especially_ when it's being done for kernel mappings. For example,
> > > > __flush_tlb_all() is a non-ranged kernel flush which has a completely
> > > > parallel implementation with flush_tlb_kernel_range(). Call sites that
> > > > use_it_ are unaffected by the patch here.
> > > >
> > > > Basically, if we're only worried about vmalloc/vfree freeing page
> > > > tables, then this patch is OK. If the problem is bigger than that, then
> > > > we need a more comprehensive patch.
> > > I think we are worried about any place that frees page tables.
> > The two places that come to mind are the remove_memory() code and
> > __change_page_attr().
> >
> > The remove_memory() gunk is in arch/x86/mm/init_64.c. It has a few sites
> > that do flush_tlb_all(). Now that I'm looking at it, there look to be
> > some races between freeing page tables pages and flushing the TLB. But,
> > basically, if you stick to the sites in there that do flush_tlb_all()
> > after free_pagetable(), you should be good.
> >
> > As for the __change_page_attr() code, I think the only spot you need to
> > hit is cpa_collapse_large_pages() and maybe the one in
> > __split_large_page() as well.
> >
> > This is all disturbingly ad-hoc, though. The remove_memory() code needs
> > fixing and I'll probably go try to bring some order to the chaos in the
> > process of fixing it up. But that's a separate problem than this IOMMU fun.
> Could we consider to split the flush_tlb_kernel_range() into 2 different
> versions:
> - the one which only flushes the CPU TLB
> - the one which flushes the CPU paging structure cache and then notifies
>    IOMMU to do the same(e.g., in pud_free_pmd_page()/pmd_free_pte_page())?

From the perspective of an IOMMU, there is no need to split. IOMMU SVA
only allows the device to access user-space memory with user
permission. Access to kernel address space with privileged permission
is not allowed. Therefore, the IOMMU subsystem only needs a callback to
invalidate the paging structure cache.

Thanks,
baolu