On Tue, Sep 8, 2026 at 11:00 AM Andreas Hindborg <[email protected]> wrote:
>
> Alice Ryhl <[email protected]> writes:
>
> > On Mon, Sep 7, 2026 at 3:58 PM Andreas Hindborg <[email protected]> 
> > wrote:
> >>
> >> "Alice Ryhl" <[email protected]> writes:
> >>
> >> > On Mon, Sep 7, 2026 at 2:38 PM Andreas Hindborg <[email protected]> 
> >> > wrote:
> >> >>
> >> >> Alice Ryhl <[email protected]> writes:
> >> >>
> >> >> > On Sun, Sep 6, 2026 at 3:02 PM Gary Guo <[email protected]> wrote:
> >> >> >>
> >> >> >> On Tue Aug 25, 2026 at 2:20 PM BST, Alice Ryhl wrote:
> >> >> >> > On Mon, Aug 24, 2026 at 01:17:56PM +0200, Andreas Hindborg wrote:
> >> >> >> >> +        // SAFETY: We just successfully allocated a page, so we 
> >> >> >> >> now have ownership of the newly
> >> >> >> >> +        // allocated page. We transfer that ownership to the new 
> >> >> >> >> `Owned<Page>` object.
> >> >> >> >> +        // Since `Page` is transparent, we can cast the pointer 
> >> >> >> >> directly.
> >> >> >> >> +        Ok(unsafe { Owned::from_raw(page.cast()) })
> >> >> >> >
> >> >> >> > This doesn't satisfy the safety requirements of Owned::from_raw()
> >> >> >> > because the page may be used with vm_insert_page(), which 
> >> >> >> > increments its
> >> >> >> > refcount and causes it to be shared the vma system, and this occurs
> >> >> >> > before Page::release() is called.
> >> >> >>
> >> >> >> I suppose the existing vm_insert_page() abstraction we have is 
> >> >> >> already
> >> >> >> problematic, because it uses `&Page`?
> >> >> >>
> >> >> >> Maybe we want to change the API to use `ARef<Page>` so it already 
> >> >> >> has to be
> >> >> >> shared? Conceptually it takes a reference count from a `&Page`, 
> >> >> >> which isn't
> >> >> >> possible because `Page` is not `AlwaysRefCounted`, so it needs a 
> >> >> >> `&ARef<Page>`
> >> >> >> to be able to do that op.
> >> >> >
> >> >> > Honestly, the problem is the safety requirements of Owned::from_raw().
> >> >> > Pages have a "special" main reference, and free_page() does more than
> >> >> > put_page(). It even does something when the refcount does not hit
> >> >> > zero.
> >> >> >
> >> >> > The correct behavior for Page is to allow the user to hold one
> >> >> > Owned<Page> whose drop calls free_page(), *plus* any number of
> >> >> > ARef<Page> references that invoke put_page() on drop. This way, the
> >> >> > owned page controls the special drop codepath.
> >> >>
> >> >> This does not mesh well with the model of `Owned` behaving like
> >> >> `UniqueArc` to `ARef` behaving like an `Arc`.
> >> >
> >> > It's a different case, I agree.
> >> >
> >> >> Please help me understand; with page having a main ref and an auxiliary
> >> >> refcount, if we model that with a single `Owned<Page>` and a number of
> >> >> `ARef<Page>`, what would happen in the case where the main ref (`Owned`)
> >> >> is dropped first? Is this legal?
> >> >
> >> > Yes, it's legal.
> >> >
> >> >> My intuition here would be to follow Garry's suggestion and have
> >> >> `vm_insert_page` take an `ARef<Page>`. Can you elaborate why this is not
> >> >> an option?
> >> >
> >> > That would be the wrong ownership semantics. The C side increments the
> >> > refcount rather than take ownership of a passed-in refcount, so the
> >> > correct argument type is &Page.
> >>
> >> Or `&ARef<Page>`? Because with `Page` being `Ownable` it would not be
> >> OK to increment the refcount from just a `&Page`.
> >
> > That's just a limitation in your traits. There is no problem with
> > creating a refcount from just a &Page.
>
> Yes, I agree. The limitation is by design. My point is, `vm_insert_page`
> can work fine with the `&ARef<Page>` type. So if that would be OK with
> you, I would patch that as part of this patch.

IMO, Page should implement AlwaysRefcounted and this method should
take &Page. Anything else just adds limitations that are not needed.
But I'm ok not implementing Ownable for it.

> >> > Now, for the use-case in Binder I believe we could switch to
> >> > put_page(), as we don't need the extra stuff from free_page(), in
> >> > which case we only want ARef<Page> and do not require Owned support.
> >> > But as long as we are calling free_page(), it needs to not be
> >> > clonable.
> >>
> >> You are calling `__free_page` directly in binder? I could not find this
> >> with grep except in the C binder. Is this code you are calling into from
> >> Rust binder?
> >
> > Via the Drop of Page, not directly of course.
> >
> >> From `__free_pages()` kernel-doc (`mm/page_alloc.c`):
> >>
> >> > If the last reference to this page is speculative, it will be
> >> > released by put_page() which only frees the first page of a
> >> > non-compound allocation. [...] **If you want to use the page's
> >> > reference count to decide when to free the allocation, you should
> >> > allocate a compound page, and use put_page() instead of
> >> > __free_pages().**
> >>
> >> I guess this is what you are referring to? We could change `Page` to
> >> this style as well and all would be fine I think?
> >>
> >> At any rate, today I learned about Pfn walkers, compaction,
> >> memory-failure, etc. that take transient refcounts
> >> pages. With that in mind we need to reformulate the invariant of
> >> `Ownable` to exclude optimistic transient references.
> >
> > Yes we can switch Rust Binder to use put_page() and have Page not
> > implement Ownable at all. That would be okay, because Binder's pages
> > are not higher order.
>
> As far as I can tell, we can also switch `Owned<Page>` to allocate the
> page with `__GFP_COMP` and use `put_page` in drop.

I'm not entirely sure if __GFP_COMP is needed, since the page is order
zero to begin with, so I don't think it makes a difference.

> > But this case still does highlight that
> > Ownable<_> currently does not address the use-case where the
> > Ownable<T> is a special reference there can only be one of, rather
> > than a completely unique reference.
>
> This is not the use case `Ownable` aims to solve.

Maybe not, but then it doesn't apply to Page.

> > For instance, the GPUVM abstraction would be another case where
> > Ownable<_> could be used if we have this other model where Ownable<_>
> > and ARef<_> are legal at the same time. Currently GPUVM uses a special
> > UniqueRefGpuVm type to solve this problem.
>
> I am wondering if we can make case I intend to solve (Ownable is a
> unique reference) coexist with the "at most one of this ref can exist".
> The "at most one of this kind" sounds similar to the `ListArc` use case
> to me, which is a different problem.

I think they can co-exist. If AlwaysRefcounted is implemented, it is
the ListArc use-case. If Refcounted but not AlwaysRefcounted is
implemented, it is UniqueArc use-case.

Alice

Reply via email to