Re: [PATCH] mm: fix COW mapping handing in generic_access_phys

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, May 28, 2025 at 08:00:34PM +0200, David Hildenbrand wrote:

> Having refcounted anon page in a VM_PFNMAP doesn't suddenly turn the whole
> thing into a MIXEDMAP where other things with a "struct page" are suddenly
> refcounted as well.

The special COW rules for PFNMAP require a single linear mapping so
that vma->vm_pgoff can encode the "special" range. So the only way to
get struct pages in a PFNMAP that also satisfy the COW rule is if you
have a single contiguous chunk of them.

Otherwise they turn into MIXEDMAP on non-special arches.

Which is a little detail I wonder how many DRM drivers trip up on :|

> ... and that should definitely be changed, such that MIXEDMAP is really only
> used when actually having non-refcounted things in there.

MIXEDMAP has to be used any time you can't meet the special rules that
remap_pfn_range() sets for PFNMAP.

Really I think we have three special cases, not just two:

VM_MIXEDMAP - The ptes can hold struct pages and if they do the refcount is
              managed by the core code

VM_PFNMAP|VM_SHARED - The ptes can never be converted into struct pages, nothing
            can ever refcount them and you can use fault/etc to store
	    things

VM_PFNMAP|!VM_SHARED - Must only be created by remap_pfn_range() and can store
            exactly 1 contiguous range of address that cannot be converted to
            struct pages and will never be refcounted. Probably can't
	    use fault

Most of my remarks are about the VM_PFNMAP|VM_SHARED case

For arches with the special bit VM_PFNMAP|!VM_SHARED == VM_MIXEDMAP

Jason