Re: [RFC PATCH v2 02/11] mm: khugepaged: generalize collapse_file() for shmem mTHP support

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

On 6/12/26 6:16 PM, Nico Pache wrote:
> 
> 
> On 6/10/26 4:29 AM, Baolin Wang wrote:
>> Generalize the order of the collapse_file() function to support future
>> shmem mTHP collapse.
>>
>> No functional changes in this patch.
>>
>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>> ---
>>   mm/khugepaged.c | 27 +++++++++++++++------------
>>   1 file changed, 15 insertions(+), 12 deletions(-)
>>
>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>> index 631459172e19..4adc8c6de062 100644
>> --- a/mm/khugepaged.c
>> +++ b/mm/khugepaged.c
>> @@ -2214,6 +2214,7 @@ static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
>>    * @file: file that collapse on
>>    * @start: collapse start address
>>    * @cc: collapse context and scratchpad
>> + * @order: folio order being collapsed to
>>    *
>>    * Basic scheme is simple, details are more complex:
>>    *  - allocate and lock a new huge page;
>> @@ -2232,15 +2233,17 @@ static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
>>    *    + unlock and free huge page;
>>    */
>>   static enum scan_result collapse_file(struct mm_struct *mm, unsigned long addr,
>> -		struct file *file, pgoff_t start, struct collapse_control *cc)
>> +		struct file *file, pgoff_t start, struct collapse_control *cc,
>> +		int order)
>>   {
>> -	const unsigned int max_ptes_none = collapse_max_ptes_none(cc, NULL, HPAGE_PMD_ORDER);
>> +	const unsigned int max_ptes_none = collapse_max_ptes_none(cc, NULL, order);
>>   	struct address_space *mapping = file->f_mapping;
>> +	const unsigned long nr_pages = 1UL << order;
>>   	struct page *dst;
>>   	struct folio *folio, *tmp, *new_folio;
>> -	pgoff_t index = 0, end = start + HPAGE_PMD_NR;
>> +	pgoff_t index = 0, end = start + nr_pages;
>>   	LIST_HEAD(pagelist);
>> -	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
>> +	XA_STATE_ORDER(xas, &mapping->i_pages, start, order);
>>   	enum scan_result result = SCAN_SUCCEED;
>>   	int nr_none = 0;
>>   	bool is_shmem = shmem_file(file);
>> @@ -2252,9 +2255,9 @@ static enum scan_result collapse_file(struct mm_struct *mm, unsigned long addr,
>>   	 * mapping, the shmem check can be removed.
>>   	 */
>>   	VM_WARN_ON_ONCE(!is_shmem && !mapping_pmd_folio_support(mapping));
>> -	VM_WARN_ON_ONCE(start & (HPAGE_PMD_NR - 1));
>> +	VM_WARN_ON_ONCE(start & (nr_pages - 1));
>>   
>> -	result = alloc_charge_folio(&new_folio, mm, cc, HPAGE_PMD_ORDER);
>> +	result = alloc_charge_folio(&new_folio, mm, cc, order);
>>   	if (result != SCAN_SUCCEED)
>>   		goto out;
>>   
>> @@ -2591,12 +2594,12 @@ static enum scan_result collapse_file(struct mm_struct *mm, unsigned long addr,
>>   	}
>>   
>>   	if (is_shmem) {
>> -		lruvec_stat_mod_folio(new_folio, NR_SHMEM, HPAGE_PMD_NR);
>> +		lruvec_stat_mod_folio(new_folio, NR_SHMEM, nr_pages);
>>   		lruvec_stat_mod_folio(new_folio, NR_SHMEM_THPS, HPAGE_PMD_NR);
> 
> Is this a accounting bug? (not your changes) but
> 
> lruvec_stat_mod_folio(new_folio, NR_SHMEM_THPS, HPAGE_PMD_NR);
> 
> If this stat is in THPs why are we iterating it by 512? shouldnt it just be +- 1

This is not a bug and this is intended behavior. Please refer to commit 
57b2847d3c1d ("mm: memcontrol: convert NR_SHMEM_THPS account to pages") 
for details.