Re: [PATCH 1/6] mm: ksm: use more folio api in ksm_might_need_to_copy()

[Kefeng Wang <wangkefeng.wang@huawei.com>]

On 2023/11/13 16:32, David Hildenbrand wrote:
> On 09.11.23 08:09, Kefeng Wang wrote:
>>
>>
>> On 2023/11/8 21:59, Matthew Wilcox wrote:
>>> On Wed, Nov 08, 2023 at 09:40:09AM +0800, Kefeng Wang wrote:
>>>>
>>>>
>>>> On 2023/11/7 22:24, Matthew Wilcox wrote:
>>>>> On Tue, Nov 07, 2023 at 09:52:11PM +0800, Kefeng Wang wrote:
>>>>>>     struct page *ksm_might_need_to_copy(struct page *page,
>>>>>> -            struct vm_area_struct *vma, unsigned long address)
>>>>>> +            struct vm_area_struct *vma, unsigned long addr)
>>>>>>     {
>>>>>>         struct folio *folio = page_folio(page);
>>>>>>         struct anon_vma *anon_vma = folio_anon_vma(folio);
>>>>>> -    struct page *new_page;
>>>>>> +    struct folio *new_folio;
>>>>>> -    if (PageKsm(page)) {
>>>>>> -        if (page_stable_node(page) &&
>>>>>> +    if (folio_test_ksm(folio)) {
>>>>>> +        if (folio_stable_node(folio) &&
>>>>>>                 !(ksm_run & KSM_RUN_UNMERGE))
>>>>>>                 return page;    /* no need to copy it */
>>>>>>         } else if (!anon_vma) {
>>>>>>             return page;        /* no need to copy it */
>>>>>> -    } else if (page->index == linear_page_index(vma, address) &&
>>>>>> +    } else if (page->index == linear_page_index(vma, addr) &&
>>>>>
>>>>> Hmm.  page->index is going away.  What should we do here instead?
>>>>
>>>> Do you mean to replace page->index to folio->index, or kill index from
>>>> struct page?
>>>
>>> I'm asking you what we should do.
>>>
>>> Tail pages already don't have a valid ->index (or ->mapping).
>>> So presumably we can't see a tail page here today.  But will we in 
>>> future?
>>
>> I think we could replace page->index to page_to_pgoff(page).
> 
> What the second part of that code does is check whether a page might 
> have been a KSM page before swapout.
> 
> Once a KSM page is swapped out, we lose the KSM marker. To recover, we 
> have to check whether the new page logically "fits" into the VMA.
> 
> Large folios are never KSM folios, and we only swap in small folios (and 
> in the future, once we would swap in large folios, they couldn't have 
> been KSM folios before).
> 
> So you could return early in the function if we have a large folio and 
> make all operations based on the (small) folio.

Sure, I will add folio_test_large check ahead and convert page->index to 
folio->index, and adjust the logical if ksm and swapin support large 
folio, thanks.