Re: [PATCH v2] mm: thp: support "THPeligible" semantics for mTHP with anonymous shmem

["Bang Li" <libang.li@antgroup.com>]

On 2024/7/3 18:25, Ryan Roberts wrote:
> On 03/07/2024 08:33, Bang Li wrote:
>> Hi Ryan,
>>
>> Thanks for the review!
>>
>> On 2024/7/2 16:18, Ryan Roberts wrote:
>>> On 02/07/2024 03:34, Bang Li wrote:
>>>> After the commit 7fb1b252afb5 ("mm: shmem: add mTHP support for
>>>> anonymous shmem"), we can configure different policies through
>>>> the multi-size THP sysfs interface for anonymous shmem. But
>>>> currently "THPeligible" indicates only whether the mapping is
>>>> eligible for allocating THP-pages as well as the THP is PMD
>>>> mappable or not for anonymous shmem, we need to support semantics
>>>> for mTHP with anonymous shmem similar to those for mTHP with
>>>> anonymous memory.
>>>>
>>>> Signed-off-by: Bang Li <libang.li@antgroup.com>
>>>> ---
>>>> Changes since v1 [1]:
>>>>    - Put anonymous shmem mthp related logic into
>>>>      thp_vma_allowable_orders() (per David)
>>>>
>>>> [1]
>>>> https://lore.kernel.org/linux-mm/20240628104926.34209-1-libang.li@antgroup.com/
>>>> ---
>>>>    include/linux/huge_mm.h | 11 +++++++++++
>>>>    mm/huge_memory.c        | 13 +++++++++----
>>>>    mm/shmem.c              |  9 +--------
>>>>    3 files changed, 21 insertions(+), 12 deletions(-)
>>>>
>>>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
>>>> index 212cca384d7e..f87136f38aa1 100644
>>>> --- a/include/linux/huge_mm.h
>>>> +++ b/include/linux/huge_mm.h
>>>> @@ -267,6 +267,10 @@ unsigned long thp_vma_allowable_orders(struct
>>>> vm_area_struct *vma,
>>>>        return __thp_vma_allowable_orders(vma, vm_flags, tva_flags, orders);
>>>>    }
>>>>    +unsigned long shmem_allowable_huge_orders(struct inode *inode,
>>>> +                struct vm_area_struct *vma, pgoff_t index,
>>>> +                bool global_huge);
>>>> +
>>>>    struct thpsize {
>>>>        struct kobject kobj;
>>>>        struct list_head node;
>>>> @@ -460,6 +464,13 @@ static inline unsigned long
>>>> thp_vma_allowable_orders(struct vm_area_struct *vma,
>>>>        return 0;
>>>>    }
>>>>    +static inline unsigned long shmem_allowable_huge_orders(struct inode *inode,
>>>> +                struct vm_area_struct *vma, pgoff_t index,
>>>> +                bool global_huge)
>>>> +{
>>>> +    return 0;
>>>> +}
>>>> +
>>>>    #define transparent_hugepage_flags 0UL
>>>>      #define thp_get_unmapped_area    NULL
>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>>> index c7ce28f6b7f3..ea377bb4af91 100644
>>>> --- a/mm/huge_memory.c
>>>> +++ b/mm/huge_memory.c
>>>> @@ -151,10 +151,15 @@ unsigned long __thp_vma_allowable_orders(struct
>>>> vm_area_struct *vma,
>>>>         * Must be done before hugepage flags check since shmem has its
>>>>         * own flags.
>>>>         */
>>>> -    if (!in_pf && shmem_file(vma->vm_file))
>>>> -        return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
>>>> -                     !enforce_sysfs, vma->vm_mm, vm_flags)
>>>> -            ? orders : 0;
>>>> +    if (!in_pf && shmem_file(vma->vm_file)) {
>>>> +        bool global_huge = shmem_is_huge(file_inode(vma->vm_file),
>>>> vma->vm_pgoff,
>>>> +                            !enforce_sysfs, vma->vm_mm, vm_flags);
>>>> +
>>>> +        if (!vma_is_anon_shmem(vma))
>>>> +            return global_huge? orders : 0;
>>>
>>> nit: missing space before '?'
>>
>> Yes, thanks.
>>
>>>
>>>> +        return shmem_allowable_huge_orders(file_inode(vma->vm_file),
>>>> +                            vma, vma->vm_pgoff, global_huge);
>>>
>>> What's the rationale for splitting these functions into shmem_is_huge() and
>>> shmem_allowable_huge_orders()? Why not just have a single
>>> shmem_allowable_huge_orders() that tells you the answer?
>>>
>>
>> Currently, shmem_is_huge() is used for all shmem implementations to determine
>> whether the conditions for using THP are met. And shmem_allowable_huge_orders()
>> is currently mainly used for anonymous shmem's mTHP to obtain all orders that
>> meet the conditions. In my opinion, there is no problem in separating these two
>> functions. In the future, when mTHP of other shmem types is also implemented,
>> will shmem_is_huge() be unnecessary?
> 
> Personally, I consider shmem_is_huge() to be superfluous; If you have one
> function, shmem_allowable_huge_orders(), that gives you all the information you
> need. If the inode only allows PMD-size, then only return that bit in the field.
> IMHO removing shmem_is_huge() would make things more readable.

Thank you very much for your opinion. I think there is no big problem in 
keeping the current interface status quo. In my opinion, when we 
implement mTHP for all shmems, it may be better to remove 
shmem_is_huge() and use shmem_allowable_huge_orders() uniformly. I 
believe that it won't be long before someone proposes a non-anonymous 
shared mTHP patchset. Anyway, thank you for your suggestion :).

Thanks,
Bang

> 
> Thanks,
> Ryan
> 
>>
>> Thanks,
>> Bang
>>
>>>> +    }
>>>>          if (!vma_is_anonymous(vma)) {
>>>>            /*
>>>> diff --git a/mm/shmem.c b/mm/shmem.c
>>>> index d495c0701a83..aa85df9c662a 100644
>>>> --- a/mm/shmem.c
>>>> +++ b/mm/shmem.c
>>>> @@ -1622,7 +1622,7 @@ static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t
>>>> limit_gfp)
>>>>    }
>>>>      #ifdef CONFIG_TRANSPARENT_HUGEPAGE
>>>> -static unsigned long shmem_allowable_huge_orders(struct inode *inode,
>>>> +unsigned long shmem_allowable_huge_orders(struct inode *inode,
>>>>                    struct vm_area_struct *vma, pgoff_t index,
>>>>                    bool global_huge)
>>>>    {
>>>> @@ -1707,13 +1707,6 @@ static unsigned long shmem_suitable_orders(struct
>>>> inode *inode, struct vm_fault
>>>>        return orders;
>>>>    }
>>>>    #else
>>>> -static unsigned long shmem_allowable_huge_orders(struct inode *inode,
>>>> -                struct vm_area_struct *vma, pgoff_t index,
>>>> -                bool global_huge)
>>>> -{
>>>> -    return 0;
>>>> -}
>>>> -
>>>>    static unsigned long shmem_suitable_orders(struct inode *inode, struct
>>>> vm_fault *vmf,
>>>>                           struct address_space *mapping, pgoff_t index,
>>>>                           unsigned long orders)