Re: [PATCH] mm/gup: fix handling of zero page in follow_page_pte()

[David Hildenbrand (Red Hat)]

On 12.11.25 08:24, peng8420.li@xxxxxxxxx wrote:
> From: "peng8420.li" <peng8420.li@xxxxxxxxx>
>
> In some scenarios, `__vm_normal_page` might return NULL.
> In this case, if `is_zero_pfn(pte_pfn(pte))` is true, simply return `page = pte_page(pte)`;
> no further folio processing is needed.
>
> Signed-off-by: peng8420.li <peng8420.li@xxxxxxxxx>
> ---
>   mm/gup.c | 1 +
>   1 file changed, 1 insertion(+)
>
> diff --git a/mm/gup.c b/mm/gup.c
> index 95d948c8e86c..60ae5f0c20e0 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -838,6 +838,7 @@ static struct page *follow_page_pte(struct vm_area_struct *vma,
>   
>   		if (is_zero_pfn(pte_pfn(pte))) {
>   			page = pte_page(pte);
> +			goto out;
>   		} else {
>   			ret = follow_pfn_pte(vma, address, ptep, flags);
>   			page = ERR_PTR(ret);

I'm sorry, but that's just wrong. Observe how me might run into the 
gup_must_unshare() branch and how try_grab_folio() would grab references 
on the FOLL_GET path.

--
Cheers

David