Re: [PATCH] mm/thp: support mTHP in thp_underused()

[Baolin Wang <baolin.wang@linux.alibaba.com>]

CC Joanne.

On 7/15/26 2:42 PM, Nico Pache wrote:
> On Wed, Jul 15, 2026 at 12:28 AM Ye Liu <ye.liu@linux.dev> wrote:
>>
>> From: Ye Liu <liuye@kylinos.cn>
>>
>> When khugepaged_max_ptes_none is configured to a non-default value,
>> scale it for non-PMD-mappable folios to match the logic in
>> collapse_max_ptes_none(), so mTHP can be classified as underused
>> instead of being silently skipped due to unscaled PMD-sized thresholds.
> 
> IIUC we don't actually add non-PMD folios to the deferred split list
> and only operate on fully mapped folios, so consequently, we never try
> to shrink them with the underused shrinker.

Right. Joanne sent a similar patch set[1] to split underused mTHP 
folios, but David said "needs more proper thought".

[1] 
https://lore.kernel.org/all/20260707201735.4113107-1-joannelkoong@gmail.com/

>> Signed-off-by: Ye Liu <liuye@kylinos.cn>
>> ---
>>   mm/huge_memory.c | 21 +++++++++++++--------
>>   1 file changed, 13 insertions(+), 8 deletions(-)
>>
>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>> index 25612af72dca..c642bec967fa 100644
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -4489,24 +4489,29 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
>>   static bool thp_underused(struct folio *folio)
>>   {
>>          int num_zero_pages = 0, num_filled_pages = 0;
>> +       unsigned int max_ptes_none = khugepaged_max_ptes_none;
>> +       unsigned int folio_nr = folio_nr_pages(folio);
>>          int i;
>>
>> -       if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1)
>> +       if (!folio_test_pmd_mappable(folio)) {
>> +               if (max_ptes_none == HPAGE_PMD_NR - 1)
>> +                       max_ptes_none = folio_nr - 1;
>> +               else if (max_ptes_none)
>> +                       return false;
>> +       }
>> +
>> +       if (max_ptes_none == folio_nr - 1)
>>                  return false;
>>
>>          if (folio_contain_hwpoisoned_page(folio))
>>                  return false;
>>
>> -       for (i = 0; i < folio_nr_pages(folio); i++) {
>> +       for (i = 0; i < folio_nr; i++) {
>>                  if (pages_identical(folio_page(folio, i), ZERO_PAGE(0))) {
>> -                       if (++num_zero_pages > khugepaged_max_ptes_none)
>> +                       if (++num_zero_pages > max_ptes_none)
>>                                  return true;
>>                  } else {
>> -                       /*
>> -                        * Another path for early exit once the number
>> -                        * of non-zero filled pages exceeds threshold.
>> -                        */
>> -                       if (++num_filled_pages >= HPAGE_PMD_NR - khugepaged_max_ptes_none)
>> +                       if (++num_filled_pages >= folio_nr - max_ptes_none)
>>                                  return false;
>>                  }
>>          }
>> --
>> 2.43.0
>>