Re: [PATCH v3] mempolicy: Optimize queue_folios_pte_range by PTE batching

[David Hildenbrand <david@redhat.com>]

On 16.04.25 10:41, Baolin Wang wrote:
> 
> 
> On 2025/4/16 16:22, David Hildenbrand wrote:
>> On 16.04.25 08:32, Baolin Wang wrote:
>>>
>>>
>>> On 2025/4/16 13:30, Dev Jain wrote:
>>>> After the check for queue_folio_required(), the code only cares about
>>>> the
>>>> folio in the for loop, i.e the PTEs are redundant. Therefore, optimize
>>>> this loop by skipping over a PTE batch mapping the same folio.
>>>>
>>>> With a test program migrating pages of the calling process, which
>>>> includes
>>>> a mapped VMA of size 4GB with pte-mapped large folios of order-9, and
>>>> migrating once back and forth node-0 and node-1, the average execution
>>>> time reduces from 7.5 to 4 seconds, giving an approx 47% speedup.
>>>>
>>>> v2->v3:
>>>>     - Don't use assignment in if condition
>>>>
>>>> v1->v2:
>>>>     - Follow reverse xmas tree declarations
>>>>     - Don't initialize nr
>>>>     - Move folio_pte_batch() immediately after retrieving a normal folio
>>>>     - increment nr_failed in one shot
>>>>
>>>> Acked-by: David Hildenbrand <david@redhat.com>
>>>> Signed-off-by: Dev Jain <dev.jain@arm.com>
>>>> ---
>>>>     mm/mempolicy.c | 12 ++++++++++--
>>>>     1 file changed, 10 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
>>>> index b28a1e6ae096..4d2dc8b63965 100644
>>>> --- a/mm/mempolicy.c
>>>> +++ b/mm/mempolicy.c
>>>> @@ -566,6 +566,7 @@ static void queue_folios_pmd(pmd_t *pmd, struct
>>>> mm_walk *walk)
>>>>     static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
>>>>                 unsigned long end, struct mm_walk *walk)
>>>>     {
>>>> +    const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
>>>>         struct vm_area_struct *vma = walk->vma;
>>>>         struct folio *folio;
>>>>         struct queue_pages *qp = walk->private;
>>>> @@ -573,6 +574,7 @@ static int queue_folios_pte_range(pmd_t *pmd,
>>>> unsigned long addr,
>>>>         pte_t *pte, *mapped_pte;
>>>>         pte_t ptent;
>>>>         spinlock_t *ptl;
>>>> +    int max_nr, nr;
>>>>         ptl = pmd_trans_huge_lock(pmd, vma);
>>>>         if (ptl) {
>>>> @@ -586,7 +588,9 @@ static int queue_folios_pte_range(pmd_t *pmd,
>>>> unsigned long addr,
>>>>             walk->action = ACTION_AGAIN;
>>>>             return 0;
>>>>         }
>>>> -    for (; addr != end; pte++, addr += PAGE_SIZE) {
>>>> +    for (; addr != end; pte += nr, addr += nr * PAGE_SIZE) {
>>>> +        max_nr = (end - addr) >> PAGE_SHIFT;
>>>> +        nr = 1;
>>>>             ptent = ptep_get(pte);
>>>>             if (pte_none(ptent))
>>>>                 continue;
>>>> @@ -598,6 +602,10 @@ static int queue_folios_pte_range(pmd_t *pmd,
>>>> unsigned long addr,
>>>>             folio = vm_normal_folio(vma, addr, ptent);
>>>>             if (!folio || folio_is_zone_device(folio))
>>>>                 continue;
>>>> +        if (folio_test_large(folio) && max_nr != 1)
>>>> +            nr = folio_pte_batch(folio, addr, pte, ptent,
>>>> +                         max_nr, fpb_flags,
>>>> +                         NULL, NULL, NULL);
>>>>             /*
>>>>              * vm_normal_folio() filters out zero pages, but there might
>>>>              * still be reserved folios to skip, perhaps in a VDSO.
>>>> @@ -630,7 +638,7 @@ static int queue_folios_pte_range(pmd_t *pmd,
>>>> unsigned long addr,
>>>>             if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
>>>>                 !vma_migratable(vma) ||
>>>>                 !migrate_folio_add(folio, qp->pagelist, flags)) {
>>>> -            qp->nr_failed++;
>>>> +            qp->nr_failed += nr;
>>>
>>> Sorry for chiming in late, but I am not convinced that 'qp->nr_failed'
>>> should add 'nr' when isolation fails.
>>
>> This patch does not change the existing behavior. But I stumbled over
>> that as well ... and scratched my head.
>>
>>>
>>>    From the comments of queue_pages_range():
>>> "
>>> * >0 - this number of misplaced folios could not be queued for moving
>>>     *      (a hugetlbfs page or a transparent huge page being counted
>>> as 1).
>>> "
>>>
>>> That means if a large folio is failed to isolate, we should only add '1'
>>> for qp->nr_failed instead of the number of pages in this large folio.
>>> Right?
>>
>> I think what the doc really meant is "PMD-mapped THP". PTE-mapped THPs
>> always had the same behavior: per PTE of the THP we would increment
>> nr_failed by 1.
> 
> No? For pte-mapped THPs, it only adds 1 for the large folio, since we
> have below check in queue_folios_pte_range().
> 
> if (folio == qp->large)
> 	continue;
> 
> Or I missed anything else?

Ah, I got confused by that and thought it would only be for LRU 
isolation purposes.

Yeah, it will kind-of work for now and I think you are correct that we 
would only increment nr_failed by 1.

I still think that counting nr_failed that way is dubious. We should be 
counting pages, which is something that user space from migrate_pages() 
could understand. Having it count arbitrary THPs/large folio sizes is 
really questionable.

But that is indeed a separate issue to resolve.

-- 
Cheers,

David / dhildenb