Re: [PATCH 1/2] mm/khugepaged: do synchronous writeback for MADV_COLLAPSE

["Garg, Shivank" <shivankg@amd.com>]

On 11/11/2025 2:46 AM, Lorenzo Stoakes wrote:
> OK ignore the past mail, I have managed to repro this locally and have a
> fix.
> 
> Turns out the swap code is doing something quite insane... I will send
> fix-patches to the series shortly.
> 
> Meanwhile I attach fix-patch! :)
> 
> Cheers, Lorenzo
> 
> ----8<----
> From c705fd85a806f53017df31e6b072c4bfa839e3a2 Mon Sep 17 00:00:00 2001
> From: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> Date: Mon, 10 Nov 2025 21:11:52 +0000
> Subject: [PATCH] fix
> 
> Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> ---
>  include/linux/leafops.h |  4 ++--
>  mm/swapfile.c           | 12 ++++++++++--
>  2 files changed, 12 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/leafops.h b/include/linux/leafops.h
> index a464a7e08c76..c4cd36760ea0 100644
> --- a/include/linux/leafops.h
> +++ b/include/linux/leafops.h
> @@ -56,7 +56,7 @@ static inline softleaf_t softleaf_from_pte(pte_t pte)
>  {
>  	softleaf_t arch_entry;
> 
> -	if (pte_present(pte))
> +	if (pte_present(pte) || pte_none(pte))
>  		return softleaf_mk_none();
> 
>  	pte = pte_swp_clear_flags(pte);
> @@ -95,7 +95,7 @@ static inline softleaf_t softleaf_from_pmd(pmd_t pmd)
>  {
>  	softleaf_t arch_entry;
> 
> -	if (pmd_present(pmd))
> +	if (pmd_present(pmd) || pmd_none(pmd))
>  		return softleaf_mk_none();
> 
>  	if (pmd_swp_soft_dirty(pmd))
> diff --git a/mm/swapfile.c b/mm/swapfile.c
> index fd23d9f7ae10..f0dcf261f652 100644
> --- a/mm/swapfile.c
> +++ b/mm/swapfile.c
> @@ -3202,9 +3202,17 @@ static int claim_swapfile(struct swap_info_struct *si, struct inode *inode)
>   */
>  unsigned long generic_max_swapfile_size(void)
>  {
> -	const softleaf_t entry = swp_entry(0, ~0UL);
> +	softleaf_t entry = swp_entry(0, ~0UL);
> +	const pte_t pte = softleaf_to_pte(entry);
> 
> -	return swp_offset(softleaf_from_pte(softleaf_to_pte(entry))) + 1;
> +	/*
> +	* Since the PTE can be an invalid swap entry (i.e. the none PTE), we do
> +	* this manually.
> +	*/
> +	entry = __pte_to_swp_entry(pte);
> +	entry = swp_entry(__swp_type(entry), __swp_offset(entry));
> +
> +	return swp_offset(entry) + 1;
>  }
> 
>  /* Can be overridden by an architecture for additional checks. */
> --
> 2.51.0

The fix works as expected.
The issue no longer reproduces on mm-new.

Thanks,
Shivank