Re: [PATCH v13 04/10] mm: split a folio in minimum folio order chunks

[Zi Yan <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 5966 bytes --]

On 30 Aug 2024, at 10:59, Pankaj Raghav wrote:

> On 30/08/2024 01:41, Luis Chamberlain wrote:
>> On Thu, Aug 29, 2024 at 06:12:26PM -0400, Zi Yan wrote:
>>> The issue is that the change to split_huge_page() makes split_huge_page_to_list_to_order()
>>> unlocks the wrong subpage. split_huge_page() used to pass the “page” pointer
>>> to split_huge_page_to_list_to_order(), which keeps that “page” still locked.
>>> But this patch changes the “page” passed into split_huge_page_to_list_to_order()
>>> always to the head page.
>>>
>>> This fixes the crash on my x86 VM, but it can be improved:
>>>
>>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
>>> index 7c50aeed0522..eff5d2fb5d4e 100644
>>> --- a/include/linux/huge_mm.h
>>> +++ b/include/linux/huge_mm.h
>>> @@ -320,10 +320,7 @@ bool can_split_folio(struct folio *folio, int *pextra_pins);
>>>  int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
>>>                 unsigned int new_order);
>>>  int split_folio_to_list(struct folio *folio, struct list_head *list);
>>> -static inline int split_huge_page(struct page *page)
>>> -{
>>> -       return split_folio(page_folio(page));
>>> -}
>>> +int split_huge_page(struct page *page);
>>>  void deferred_split_folio(struct folio *folio);
>>>
>>>  void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>> index c29af9451d92..4d723dab4336 100644
>>> --- a/mm/huge_memory.c
>>> +++ b/mm/huge_memory.c
>>> @@ -3297,6 +3297,25 @@ int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
>>>         return ret;
>>>  }
>>>
>>> +int split_huge_page(struct page *page)
>>> +{
>>> +       unsigned int min_order = 0;
>>> +       struct folio *folio = page_folio(page);
>>> +
>>> +       if (folio_test_anon(folio))
>>> +               goto out;
>>> +
>>> +       if (!folio->mapping) {
>>> +               if (folio_test_pmd_mappable(folio))
>>> +                       count_vm_event(THP_SPLIT_PAGE_FAILED);
>>> +               return -EBUSY;
>>> +       }
>>> +
>>> +       min_order = mapping_min_folio_order(folio->mapping);
>>> +out:
>>> +       return split_huge_page_to_list_to_order(page, NULL, min_order);
>>> +}
>>> +
>>>  int split_folio_to_list(struct folio *folio, struct list_head *list)
>>>  {
>>>         unsigned int min_order = 0;
>>
>>
>> Confirmed, and also although you suggest it can be improved, I thought
>> that we could do that by sharing more code and putting things in the
>> headers, the below also fixes this but tries to share more code, but
>> I think it is perhaps less easier to understand than your patch.
>>
> It feels a bit weird to pass both folio and the page in `split_page_folio_to_list()`.
>
> How about we extract the code that returns the min order so that we don't repeat.
>
> Something like this:
>
>
> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> index c275aa9cc105..d27febd5c639 100644
> --- a/include/linux/huge_mm.h
> +++ b/include/linux/huge_mm.h
> @@ -331,10 +331,24 @@ unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long add
>  bool can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins);
>  int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
>                 unsigned int new_order);
> +int min_order_for_split(struct folio *folio);
>  int split_folio_to_list(struct folio *folio, struct list_head *list);

Since split_folio() is no longer used below, this line can be removed.

>  static inline int split_huge_page(struct page *page)
>  {
> -       return split_folio(page_folio(page));
> +       struct folio *folio = page_folio(page);
> +       int ret = min_order_for_split(folio);
> +
> +       if (ret)
> +               return ret;

min_order_for_split() returns -EBUSY, 0, and a positive min_order. This if
statement should be "if (ret < 0)"?

> +
> +       /*
> +        * split_huge_page() locks the page before splitting and
> +        * expects the same page that has been split to be locked when
> +        * returned. split_folio_to_list() cannot be used here because
> +        * it converts the page to folio and passes the head page to be
> +        * split.
> +        */
> +       return split_huge_page_to_list_to_order(page, NULL, ret);
>  }
>  void deferred_split_folio(struct folio *folio);
>
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index 169f1a71c95d..b167e036d01b 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -3529,12 +3529,10 @@ int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
>         return ret;
>  }
>
> -int split_folio_to_list(struct folio *folio, struct list_head *list)
> +int min_order_for_split(struct folio *folio)
>  {
> -       unsigned int min_order = 0;
> -
>         if (folio_test_anon(folio))
> -               goto out;
> +               return 0;
>
>         if (!folio->mapping) {
>                 if (folio_test_pmd_mappable(folio))
> @@ -3542,10 +3540,17 @@ int split_folio_to_list(struct folio *folio, struct list_head *list)
>                 return -EBUSY;
>         }
>
> -       min_order = mapping_min_folio_order(folio->mapping);
> -out:
> -       return split_huge_page_to_list_to_order(&folio->page, list,
> -                                                       min_order);
> +       return mapping_min_folio_order(folio->mapping);
> +}
> +
> +int split_folio_to_list(struct folio *folio, struct list_head *list)
> +{
> +       int ret = min_order_for_split(folio);
> +
> +       if (ret)
> +               return ret;

Ditto.

> +
> +       return split_huge_page_to_list_to_order(&folio->page, list, ret);
>  }
>
>  void __folio_undo_large_rmappable(struct folio *folio)
>


--
Best Regards,
Yan, Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]