Re: [PATCH] mm: hugetlb: kill set_huge_swap_pte_at()

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2022/6/27 15:14, Anshuman Khandual wrote:
> 
> 
> On 6/27/22 12:25, Qi Zheng wrote:
>>
>>
>> On 2022/6/27 14:18, Anshuman Khandual wrote:
>>>
>>>
>>> On 6/26/22 20:27, Qi Zheng wrote:
>>>> The commit e5251fd43007 ("mm/hugetlb: introduce set_huge_swap_pte_at()
>>>> helper") add set_huge_swap_pte_at() to handle swap entries on
>>>> architectures that support hugepages consisting of contiguous ptes.
>>>> And currently the set_huge_swap_pte_at() is only overridden by arm64.
>>>>
>>>> The set_huge_swap_pte_at() provide a sz parameter to help determine
>>>> the number of entries to be updated. But in fact, all hugetlb swap
>>>> entries contain pfn information, so we can find the corresponding
>>>> folio through the pfn recorded in the swap entry, then the folio_size()
>>>> is the number of entries that need to be updated.
>>>>
>>>> And considering that users will easily cause bugs by ignoring the
>>>> difference between set_huge_swap_pte_at() and set_huge_pte_at().
>>>> Let's handle swap entries in set_huge_pte_at() and remove the
>>>> set_huge_swap_pte_at(), then we can call set_huge_pte_at()
>>>> anywhere, which simplifies our coding.
>>>>
>>>> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
>>>> ---
>>>>    arch/arm64/include/asm/hugetlb.h |  3 ---
>>>>    arch/arm64/mm/hugetlbpage.c      | 34 ++++++++++++++++----------------
>>>>    include/linux/hugetlb.h          | 13 ------------
>>>>    mm/hugetlb.c                     |  8 +++-----
>>>>    mm/rmap.c                        | 11 +++--------
>>>>    5 files changed, 23 insertions(+), 46 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/include/asm/hugetlb.h b/arch/arm64/include/asm/hugetlb.h
>>>> index 1fd2846dbefe..d20f5da2d76f 100644
>>>> --- a/arch/arm64/include/asm/hugetlb.h
>>>> +++ b/arch/arm64/include/asm/hugetlb.h
>>>> @@ -46,9 +46,6 @@ extern void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
>>>>                   pte_t *ptep, unsigned long sz);
>>>>    #define __HAVE_ARCH_HUGE_PTEP_GET
>>>>    extern pte_t huge_ptep_get(pte_t *ptep);
>>>> -extern void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
>>>> -                 pte_t *ptep, pte_t pte, unsigned long sz);
>>>> -#define set_huge_swap_pte_at set_huge_swap_pte_at
>>>>      void __init arm64_hugetlb_cma_reserve(void);
>>>>    diff --git a/arch/arm64/mm/hugetlbpage.c b/arch/arm64/mm/hugetlbpage.c
>>>> index c9e076683e5d..58b89b9d13e0 100644
>>>> --- a/arch/arm64/mm/hugetlbpage.c
>>>> +++ b/arch/arm64/mm/hugetlbpage.c
>>>> @@ -238,6 +238,13 @@ static void clear_flush(struct mm_struct *mm,
>>>>        flush_tlb_range(&vma, saddr, addr);
>>>>    }
>>>>    +static inline struct folio *hugetlb_swap_entry_to_folio(swp_entry_t entry)
>>>> +{
>>>> +    VM_BUG_ON(!is_migration_entry(entry) && !is_hwpoison_entry(entry));
>>>> +
>>>> +    return page_folio(pfn_to_page(swp_offset(entry)));
>>>> +}
>>>
>>> Extracting this huge page size from swap entry is an additional operation which
>>> will increase the over all cost for set_huge_swap_pte_at(). At present the size
>>
>> Hmm, I think this cost is very small. And replacing
>> set_huge_swap_pte_at() by transparently handling swap entries helps
>> reduce possible bugs, which is worthwhile.
> 
> Possible bugs ? There are just six call sites for this function.

For the swap entry of hugetlb, we need to remember that we should
call set_huge_swap_pte_at() instead of set_huge_pte_at() every
time. And the difference between the two is only on arm64, other
architectures should not need to pay attention to this.

> Although this proposed patch is functionally correct, I dont see
> a valid enough reason to increase the overall cost in the path.

-- 
Thanks,
Qi