On 12/17/2025 8:18 AM, Wei Yang wrote:
> On Tue, Dec 16, 2025 at 11:11:38AM +0000, Shivank Garg wrote:
>> Replace 'goto skip' with actual logic for better code readability.
>>
>> No functional change.
>>
>> Signed-off-by: Shivank Garg <[email protected]>
>> ---
>> mm/khugepaged.c | 7 ++++---
>> 1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>> index 6c8c35d3e0c9..107146f012b1 100644
>> --- a/mm/khugepaged.c
>> +++ b/mm/khugepaged.c
>> @@ -2442,14 +2442,15 @@ static unsigned int khugepaged_scan_mm_slot(unsigned
>> int pages, int *result,
>> break;
>> }
>> if (!thp_vma_allowable_order(vma, vma->vm_flags,
>> TVA_KHUGEPAGED, PMD_ORDER)) {
>> -skip:
>> progress++;
>> continue;
>> }
>> hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
>> hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
>> - if (khugepaged_scan.address > hend)
>> - goto skip;
>> + if (khugepaged_scan.address > hend) {
>> + progress++;
>> + continue;
>> + }
>
> Hi, Shivank
>
> The change here looks good, while I come up with an question.
>
> The @progress here seems record two things:
>
> * number of pages scaned
> * number of vma skipped
>
Three things: number of mm. It's incremented 1 for whole
khugepaged_scan_mm_slot().
> While in very rare case, we may miss to count the second case.
>
> For example, we have following vmas in a process:
>
> vma1 vma2
> +----------------+------------+
> |2M |1M |
> +----------------+------------+
>
> Let's assume vma1 is exactly HPAGE_PMD_SIZE and also HPAGE_PMD_SIZE aligned.
> But vma2 is only half of HPAGE_PMD_SIZE.
>
> When scan finish vma1 and start on vma2, we would have hstart = hend =
> address. So we continue here but would not do real scan, since address ==
> hend.
>
> I am thinking whether this could handle it:
>
> if (khugepaged_scan.address > hend || hend <= hstart) {
> progress++;
> continue;
> }
>
> Do you thinks I am correct on this?
I think you're correct.
IIUC, @progress acts as rate limiter here.
It is increasing +1 for whole, and then increases by +1 per VMA (if skipped),
or by +HPAGE_PMD_NR (if actually scanned).
So, progress ensuring the hugepaged_do_scan run only until (progress >= pages)
at which point it yields and sleeps (wait_event_freezable).
Without your suggested fix, if a process contains a large number of small VMAs
(where
round_up hstart >= round_down(hend), it will unfairly consume more CPU cycles
before
yielding compared to a process with fewer or aligned VMAs.
I think your suggestion is ensuring fairness by charging 'progress' count
correctly.
Thanks,
Shivank
