Re: [PATCH mm-unstable 1/1] mm/khugepaged: fix PMD collapse swap PTE accounting

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

On 2026/6/9 21:16, David Hildenbrand (Arm) wrote:
> On 6/9/26 14:04, Lance Yang wrote:
>> From: Lance Yang <lance.yang@linux.dev>
>>
>> mthp_collapse() uses mthp_present_ptes to decide whether a range has
>> enough occupied PTEs to try collapse. Swap PTEs accepted by
>> collapse_scan_pmd() are counted in unmapped, but are not represented in
>> mthp_present_ptes.
>>
>> When lower orders are enabled, collapse_scan_pmd() relaxes max_ptes_none
>> so the scan can cover the whole PMD and build the bitmap. mthp_collapse()
>> then checks the PMD-order candidate using the bitmap.
>>
>> With max_ptes_none set to 0, a range with 511 present PTEs and one swap
>> PTE no longer reaches collapse_huge_page(), even though PMD collapse can
>> handle swap PTEs up to max_ptes_swap.
>>
>> Account unmapped PTEs only for PMD order. PMD collapse supports swap PTEs
>> through max_ptes_swap, while lower-order mTHP collapse does not currently
>> support non-present PTEs. Keep non-present PTEs out of the lower-order
>> eligibility check.
>>
>> Signed-off-by: Lance Yang <lance.yang@linux.dev>
>> ---
>> Sent separately, as discussed in [1], to spell out the PMD-order swap PTE
>> case. Patch [2] is still only in mm-unstable, so no Fixes: tag.
>>
>> [1] https://lore.kernel.org/linux-mm/CAA1CXcD7WAiA1b9GTLAuNZ+kHaFx0SzZwpBkqAZ=s+RHsTUaow@mail.gmail.com/
>> [2] https://lore.kernel.org/linux-mm/20260605161422.213817-12-npache@redhat.com/
>>
>>   mm/khugepaged.c | 8 ++++++++
>>   1 file changed, 8 insertions(+)
>>
>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>> index b12187709f6d..617bca76db49 100644
>> --- a/mm/khugepaged.c
>> +++ b/mm/khugepaged.c
>> @@ -1508,6 +1508,14 @@ static enum scan_result mthp_collapse(struct mm_struct *mm,
>>   		nr_occupied_ptes = bitmap_weight_from(cc->mthp_present_ptes, offset,
>>   						      offset + nr_ptes);
>>   
>> +		/*
>> +		 * Swap PTEs accepted during the scan are counted in @unmapped,
>> +		 * not in the present-PTE bitmap. Account them for the PMD-order
>> +		 * candidate.
>> +		 */
>> +		if (is_pmd_order(order))
>> +			nr_occupied_ptes += unmapped;
>> +
> 
> LGTM, there is a bit of opportunity for cleanup in the future :)

Yes, follow-up cleanup material :)

> Acked-by: David Hildenbrand (Arm) <david@kernel.org>

Thanks!

> For example, as we no longer have the VMA here, collapse_max_ptes_none is
> imprecise in uffd VMAs. We might try collapsing where there sure is nothing to
> collapse.

Oh, good catch. We may end up trying a collapse that cannot really
go anywhere ... One for a follow-up.

> We could likely handle the userfaultfd_armed() part easier: some indication that
> we must not have any pte_none() would be sufficient.

Right. By the time we get to mthp_collapse(), we probably only need
to carry that as a small "no pte_none" constraint for the candidate
range.

> Also, I don't see a good reason why uffd would not be allowed to collapse with
> zeropages ... it's really just about missing faults due to pte_none().

Makes sense to me. I'll take a look when I get a chance. And yeah,
as Lorenzo said, better to clean up the khugepaged mess first
before piling more on top :)

Cheers, Lance