Re: [PATCH v2 1/4] mm: Introduce vm_uffd_ops API

[Peter Xu <peterx@redhat.com>]

On Fri, Sep 19, 2025 at 03:34:39PM +0100, Lorenzo Stoakes wrote:
> Peter -
> 
> I've been staying out of this discussion as I'm about to go to Kernel
> Recipes and then off on a (well-needed!) holiday, and I simply lack the
> bandwidth right now.
> 
> But I think we should all calm down a little here :)
> 
> Liam and I (more so Liam recently for above reasons) have pushed back
> because we have both personally experienced the consequences of giving
> drivers too much flexibility wrt core mm functionality.
> 
> This is the sole reason we have done so.
> 
> We are both eager to find a way forward that is constructive and works well
> for everybody involved. We WANT this series to land.
> 
> So I think perhaps we should take a step back and identify clearly what the
> issues are and how we might best address them.
> 
> I spoke to Mike off-list who suggested perhaps things aren't quite
> egregious as they seem with uffd_get_folio() so perhaps this is a means of
> moving forward.
> 
> But I think in broad terms - let's identify what the sensible options are,
> and then drill down into whichever one we agree is best to move forwards
> with.
> 
> Again, apologies for not being able to be more involved here,
> workload/other engagements dictate that I am unable to be.

That's totally fine, Lorenzo.  I appreciate your help on figuring things
out.

I do agree the discussion actually went nowhere.

I think so far the "issues" is very much clear, about exporting
uffd_get_folio(), as you correctly pointed out and I'm glad you discussed
with Mike.

My point is that hook is totally fine, and we need that exactly because we
want to keep ->fault() semantic clean.

Just to mention, if this series cannot land, I prefer landing Nikita's very
old version (a).  That'll make mm fault() ugly, I pointed that out, but if
all the people prefer that and all the people like to sign-off with it, I'm
OK from userfaultfd perspective.  I don't make judgement there.

Then this series can drop uffd_get_folio() and keep the rest in one way or
another, describing memory type attributes only, and need to cooperate only
a driver with a ->fault() that works for the new flag.  But then this
series will be a pure cleanup.  I'll likely then put this series aside as
it stops blocking things, and I also have a queue to flush myself elsewhere.

I wished we can just go with this series with uffd_get_folio() only.  Feel
free to discuss with more people, and let me know how this series should
move on.

Thanks a lot,

-- 
Peter Xu