Re: [PATCH v2] khugepaged: remove redundant index check for pmd-folios

[Lance Yang <lance.yang@linux.dev>]

On 2026/2/27 22:35, Dev Jain wrote:
> Claim: folio_order(folio) == HPAGE_PMD_ORDER => folio->index == start.
> 
> Proof: Both loops in hpage_collapse_scan_file and collapse_file, which
> iterate on the xarray, have the invariant that
> start <= folio->index < start + HPAGE_PMD_NR ... (i)
> 
> A folio is always naturally aligned in the pagecache, therefore
> folio_order == HPAGE_PMD_ORDER => IS_ALIGNED(folio->index, HPAGE_PMD_NR) == true ... (ii)
> 
> thp_vma_allowable_order -> thp_vma_suitable_order requires that the virtual
> offsets in the VMA are aligned to the order,
> => IS_ALIGNED(start, HPAGE_PMD_NR) == true ... (iii)
> 
> Combining (i), (ii) and (iii), the claim is proven.
> 
> Therefore, remove this check.
> While at it, simplify the comments.
> 
> Signed-off-by: Dev Jain <dev.jain@arm.com>
> ---
> v1->v2:
>   - Remove the check instead of converting to VM_WARN_ON
>   - While at it, simplify the comments
> 
> Based on mm-new (8982358e1c87).
> 
>   mm/khugepaged.c | 14 ++++----------
>   1 file changed, 4 insertions(+), 10 deletions(-)
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 5f668c1dd0fe4..b7b4680d27ab1 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -2015,9 +2015,7 @@ static enum scan_result collapse_file(struct mm_struct *mm, unsigned long addr,
>   		 * we locked the first folio, then a THP might be there already.
>   		 * This will be discovered on the first iteration.
>   		 */
> -		if (folio_order(folio) == HPAGE_PMD_ORDER &&
> -		    folio->index == start) {
> -			/* Maybe PMD-mapped */
> +		if (folio_order(folio) == HPAGE_PMD_ORDER) {
>   			result = SCAN_PTE_MAPPED_HUGEPAGE;
>   			goto out_unlock;
>   		}
> @@ -2345,15 +2343,11 @@ static enum scan_result hpage_collapse_scan_file(struct mm_struct *mm,
>   			continue;
>   		}
>   
> -		if (folio_order(folio) == HPAGE_PMD_ORDER &&
> -		    folio->index == start) {
> -			/* Maybe PMD-mapped */
> +		if (folio_order(folio) == HPAGE_PMD_ORDER) {
>   			result = SCAN_PTE_MAPPED_HUGEPAGE;
>   			/*
> -			 * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
> -			 * by the caller won't touch the page cache, and so
> -			 * it's safe to skip LRU and refcount checks before
> -			 * returning.
> +			 * PMD-sized THP implies that we can only try
> +			 * retracting the PTE table.
>   			 */
>   			folio_put(folio);
>   			break;

LGTM!

The proof is sound, the combination of the loop invariant, natural
alignment, and VMA alignment requirements indeed makes the index
check redundant :D

Reviewed-by: Lance Yang <lance.yang@linux.dev>