Re: [PATCH] mm: remove an avoidable load of page refcount in page_ref_add_unless

[Mateusz Guzik]

On Mon, Dec 9, 2024 at 3:22 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 09.12.24 13:33, Mateusz Guzik wrote:
> > That is to say I think this thread just about exhausted the time
> > warranted by this patch. No hard feelz if it gets dropped, but then I
> > do strongly suggest adding a justification to the extra load.
>
> Maybe it's sufficient for now to simply do your change with a comment:
>
> diff --git a/include/linux/page_ref.h b/include/linux/page_ref.h
> index 8c236c651d1d6..1efc992ad5687 100644
> --- a/include/linux/page_ref.h
> +++ b/include/linux/page_ref.h
> @@ -234,7 +234,13 @@ static inline bool page_ref_add_unless(struct page *page, int nr, int u)
>
>          rcu_read_lock();
>          /* avoid writing to the vmemmap area being remapped */
> -       if (!page_is_fake_head(page) && page_ref_count(page) != u)
> +       if (!page_is_fake_head(page))
> +               /*
> +                * atomic_add_unless() will currently never modify the value
> +                * if it already is u. If that ever changes, we'd have to have
> +                * a separate check here, such that we won't be writing to
> +                * write-protected vmemmap areas.
> +                */
>                  ret = atomic_add_unless(&page->_refcount, nr, u);
>          rcu_read_unlock();
>
>
> It would bail out during testing ... hopefully, such that we can detect any such change.
>

Not my call to make, but looks good. ;)

fwiw I don't need any credit and I would be more than happy if you
just submitted the thing as your own without me being mentioned. *No*
cc would also be appreciated.

-- 
Mateusz Guzik <mjguzik gmail.com>