Re: [PATCH RFC 00/14] mm: vm_normal_page*() + CoW PFNMAP improvements

[David Hildenbrand]

On 17.06.25 18:18, David Hildenbrand wrote:
> On 17.06.25 17:43, David Hildenbrand wrote:
> > RFC because it's based on mm-new where some things might still change
> > around the devmap removal stuff.
> >
> > While removing support for CoW PFNMAPs is a noble goal, I am not even sure
> > if we can remove said support for e.g., /dev/mem that easily.
> >
> > In the end, Cow PFNMAPs are pretty simple: everything is "special" except
> > CoW'ed anon folios, that are "normal".
> >
> > The only complication is: how to identify such pages without pte_special().
> > Because with pte_special(), it's easy.
> >
> > Well, of course, one day all architectures might support pte_special() ...
> > either because we added support for pte_special() or removed support for
> > ... these architectures from Linux.
> >
> > No need to wait for that day. Let's do some cleanups around
> > vm_normal_page()/vm_normal_page_pmd() and handling of the huge zero folio,
> > and remove the "horrible special case to handle copy-on-write behaviour"
> > that does questionable things in remap_pfn_range() with a VMA, simply by
> >
> > ... looking for anonymous folios in CoW PFNMAPs to identify anonymous
> > folios? I know, sounds crazy ;)
>
> I'll mention one corner case that just occurred to me: assume someone
> maps arbitrary /dev/mem that is actually used by the kernel for user
> space, and then some of that memory gets allocated as anonymous memory,
> it would probably be a problem.
>
> Hmm, I'll have to think about that, and the interaction with
> CONFIG_STRICT_DEVMEM.

The /dev/mem mapping of arbitrary memory is indeed the hard case. To 
handle all that, patch #11 is too simplistic.

I have some idea how to make it work, but have to think about a couple 
of corner cases.

--
Cheers,

David / dhildenb