Re: [PATCH v4 3/4] mm: Support batched unmap for lazyfree large folios during reclamation

[David Hildenbrand <david@redhat.com>]

On 25.06.25 12:49, Barry Song wrote:
> On Wed, Jun 25, 2025 at 10:43 PM David Hildenbrand <david@redhat.com> wrote:
>>
>> On 25.06.25 12:38, Barry Song wrote:
>>>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>>>> index fb63d9256f09..241d55a92a47 100644
>>>>> --- a/mm/rmap.c
>>>>> +++ b/mm/rmap.c
>>>>> @@ -1847,12 +1847,25 @@ void folio_remove_rmap_pud(struct folio *folio, struct page *page,
>>>>>
>>>>>     /* We support batch unmapping of PTEs for lazyfree large folios */
>>>>>     static inline bool can_batch_unmap_folio_ptes(unsigned long addr,
>>>>> -                     struct folio *folio, pte_t *ptep)
>>>>> +                                           struct folio *folio, pte_t *ptep,
>>>>> +                                           struct vm_area_struct *vma)
>>>>>     {
>>>>>         const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
>>>>> +     unsigned long next_pmd, vma_end, end_addr;
>>>>>         int max_nr = folio_nr_pages(folio);
>>>>>         pte_t pte = ptep_get(ptep);
>>>>>
>>>>> +     /*
>>>>> +      * Limit the batch scan within a single VMA and within a single
>>>>> +      * page table.
>>>>> +      */
>>>>> +     vma_end = vma->vm_end;
>>>>> +     next_pmd = ALIGN(addr + 1, PMD_SIZE);
>>>>> +     end_addr = addr + (unsigned long)max_nr * PAGE_SIZE;
>>>>> +
>>>>> +     if (end_addr > min(next_pmd, vma_end))
>>>>> +             return false;
>>>>
>>>> May I suggest that we clean all that up as we fix it?
>>>>
>>>> Maybe something like this:
>>>>
>>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>>> index 3b74bb19c11dd..11fbddc6ad8d6 100644
>>>> --- a/mm/rmap.c
>>>> +++ b/mm/rmap.c
>>>> @@ -1845,23 +1845,38 @@ void folio_remove_rmap_pud(struct folio *folio, struct page *page,
>>>>     #endif
>>>>     }
>>>>
>>>> -/* We support batch unmapping of PTEs for lazyfree large folios */
>>>> -static inline bool can_batch_unmap_folio_ptes(unsigned long addr,
>>>> -                       struct folio *folio, pte_t *ptep)
>>>> +static inline unsigned int folio_unmap_pte_batch(struct folio *folio,
>>>> +               struct page_vma_mapped_walk *pvmw, enum ttu_flags flags,
>>>> +               pte_t pte)
>>>>     {
>>>>            const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
>>>> -       int max_nr = folio_nr_pages(folio);
>>>> -       pte_t pte = ptep_get(ptep);
>>>> +       struct vm_area_struct *vma = pvmw->vma;
>>>> +       unsigned long end_addr, addr = pvmw->address;
>>>> +       unsigned int max_nr;
>>>> +
>>>> +       if (flags & TTU_HWPOISON)
>>>> +               return 1;
>>>> +       if (!folio_test_large(folio))
>>>> +               return 1;
>>>> +
>>>> +       /* We may only batch within a single VMA and a single page table. */
>>>> +       end_addr = min_t(unsigned long, ALIGN(addr + 1, PMD_SIZE), vma->vm_end);
>>>
>>> Is this pmd_addr_end()?
>>>
>>
>> Yes, that could be reused as well here.
>>
>>>> +       max_nr = (end_addr - addr) >> PAGE_SHIFT;
>>>>
>>>> +       /* We only support lazyfree batching for now ... */
>>>>            if (!folio_test_anon(folio) || folio_test_swapbacked(folio))
>>>> -               return false;
>>>> +               return 1;
>>>>            if (pte_unused(pte))
>>>> -               return false;
>>>> -       if (pte_pfn(pte) != folio_pfn(folio))
>>>> -               return false;
>>>> +               return 1;
>>>> +       /* ... where we must be able to batch the whole folio. */
>>>> +       if (pte_pfn(pte) != folio_pfn(folio) || max_nr != folio_nr_pages(folio))
>>>> +               return 1;
>>>> +       max_nr = folio_pte_batch(folio, addr, pvmw->pte, pte, max_nr, fpb_flags,
>>>> +                                NULL, NULL, NULL);
>>>>
>>>> -       return folio_pte_batch(folio, addr, ptep, pte, max_nr, fpb_flags, NULL,
>>>> -                              NULL, NULL) == max_nr;
>>>> +       if (max_nr != folio_nr_pages(folio))
>>>> +               return 1;
>>>> +       return max_nr;
>>>>     }
>>>>
>>>>     /*
>>>> @@ -2024,9 +2039,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
>>>>                            if (pte_dirty(pteval))
>>>>                                    folio_mark_dirty(folio);
>>>>                    } else if (likely(pte_present(pteval))) {
>>>> -                       if (folio_test_large(folio) && !(flags & TTU_HWPOISON) &&
>>>> -                           can_batch_unmap_folio_ptes(address, folio, pvmw.pte))
>>>> -                               nr_pages = folio_nr_pages(folio);
>>>> +                       nr_pages = folio_unmap_pte_batch(folio, &pvmw, flags, pteval);
>>>>                            end_addr = address + nr_pages * PAGE_SIZE;
>>>>                            flush_cache_range(vma, address, end_addr);
>>>>
>>>>
>>>> Note that I don't quite understand why we have to batch the whole thing or fallback to
>>>> individual pages. Why can't we perform other batches that span only some PTEs? What's special
>>>> about 1 PTE vs. 2 PTEs vs. all PTEs?
>>>>
>>>>
>>>> Can someone enlighten me why that is required?
>>>
>>> It's probably not a strict requirement — I thought cases where the
>>> count is greater than 1 but less than nr_pages might not provide much
>>> practical benefit, except perhaps in very rare edge cases, since
>>> madv_free() already calls split_folio().
>>
>> Okay, but it makes the code more complicated. If there is no reason to
>> prevent the batching, we should drop it.
> 
> It's not necessarily more complex, since page_vma_mapped_walk() still
> has to check each PTE individually and can't skip ahead based on nr.
> With nr_pages batched, we can exit the loop early in one go.

I said "complicated", not "complex". The code is more complicated than 
necessary.

-- 
Cheers,

David / dhildenb