Re: [PATCH 1/4] mm: Introduce vm_uffd_ops API

[David Hildenbrand]

On 23.06.25 15:59, Peter Xu wrote:
> On Mon, Jun 23, 2025 at 10:25:33AM +0200, David Hildenbrand wrote:
> > On 20.06.25 21:03, Peter Xu wrote:
> >
> > Hi Peter,
>
> Hey David,
>
> > > Introduce a generic userfaultfd API for vm_operations_struct, so that one
> > > vma, especially when as a module, can support userfaults without modifying
> >
> > The sentence is confusing ("vma ... as a module").
> >
> > Did you mean something like ".. so that a vma that is backed by a
> > special-purpose in-memory filesystem like shmem or hugetlb can support
> > userfaultfd without modifying the uffd core; this is required when the
> > in-memory filesystem is built as a module."
>
> I wanted to avoid mentioning of "in-memory file systems" here.

I thought one of the challenges of supporting guest_memfd on anything 
that is not a special in-memory file system is also related to how the 
pagecache handles readahead.

So ...

> How about an updated commit like this?
>
>    Currently, most of the userfaultfd features are implemented directly in the
>    core mm.  It will invoke VMA specific functions whenever necessary.  So far
>    it is fine because it almost only interacts with shmem and hugetlbfs.
>
>    This patch introduces a generic userfaultfd API for vm_operations_struct,
>    so that any type of file (including kernel modules that can be compiled
>    separately from the kernel core) can support userfaults without modifying
>    the core files.

.... is it really "any file" ? I doubt it, but you likely have a better 
idea on how it all could just work with "any file".

>    After this API applied, if a module wants to support userfaultfd, the
>    module should only need to touch its own file and properly define
>    vm_uffd_ops, instead of changing anything in core mm.
>
>    ...

Talking about files and modules is still confusing I'm afraid. It's 
really a special-purpose file (really, not any ordinary files on 
ordinary filesystems), no?

> > > the core files.  More importantly, when the module can be compiled out of
> > > the kernel.
> > >
> > > So, instead of having core mm referencing modules that may not ever exist,
> > > we need to have modules opt-in on core mm hooks instead.
> > >
> > > After this API applied, if a module wants to support userfaultfd, the
> > > module should only need to touch its own file and properly define
> > > vm_uffd_ops, instead of changing anything in core mm.
> >
> > Talking about modules that much is a bit confusing. I think this is more
> > about cleanly supporting in-memory filesystems, without the need to
> > special-case each and every one of them; can be viewed a cleanup independent
> > of the module requirement from guest_memfd.
>
> Yes.  But if we don't need to support kernel modules actually we don't need
> this.. IMHO it's so far really about cleanly support kernel modules, which
> can even be out-of-tree (though that's not my purpose of the change..).
>
> Please help check if above updated commit message would be better.

I agree that another special-purpose file (like implemented by 
guest_memfd) would need that. But if we could get rid of 
"hugetlb"/"shmem" special-casing in userfaultfd, it would be a rasonable 
independent cleanup.

But I can spot in patch #3 now:

"Hugetlbfs still has its own hard-coded handler in userfaultfd, due to 
limitations similar to vm_operations_struct.fault(). TODO: generalize it 
to use the API function."

I would have hoped that we clean that up in one go instead.

> > > Note that such API will not work for anonymous. Core mm will process
> > > anonymous memory separately for userfault operations like before.
> > >
> > > This patch only introduces the API alone so that we can start to move
> > > existing users over but without breaking them.
> > >
> > > Currently the uffd_copy() API is almost designed to be the simplistic with
> > > minimum mm changes to move over to the API.
> >
> > Is there a way to move part of the actual implementation (how this is all
> > wired up) from patch #4 into this patch, to then only remove the old
> > shmem/hugetlb hooks (that are effectively unused) in patch #4?
>
> Not much I really removed on the hooks, but I was trying to reuse almost
> existing functions.  Here hugetlb is almost untouched on hooks, then I
> reused the shmem existing function for uffd_copy() rather than removing it
> (I did need to remove the definition in the shmem header though becuse it's
> not needed to be exported).
>
> The major thing got removed in patch 4 was some random checks over uffd ops
> and vma flags.  I intentionally made them all in patch 4 to make review
> possible.  Otherwise it can be slightly awkward to reason what got removed
> without knowing what is protecting those checks.

Agreed. It's a shame the new API is not a proper replacement for hugetlb 
special casing just yet ...

--
Cheers,

David / dhildenb