Re: [PATCH] mm/mmap: introduce vma_range_init()

[Yajun Deng <yajun.deng@linux.dev>]

On 2024/1/23 11:19, Liam R. Howlett wrote:
> * Yajun Deng <yajun.deng@linux.dev> [240122 21:23]:
>> On 2024/1/23 08:18, Liam R. Howlett wrote:
>>> * Andrew Morton <akpm@linux-foundation.org> [240122 18:40]:
>>>> On Mon, 22 Jan 2024 17:00:31 -0500 "Liam R. Howlett" <Liam.Howlett@Oracle.com> wrote:
>>>>
>>>>> * Yajun Deng <yajun.deng@linux.dev> [240110 21:15]:
>>>>>> There is a lot of code needs to set the range of vma, introduce
>>>>>> vma_range_init() to initialize the range of vma.
>>>>>>
>>>>>> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
>>>>>> ---
>>>>>>    include/linux/mm.h |  9 +++++++++
>>>>>>    mm/mmap.c          | 29 +++++++----------------------
>>>>>>    2 files changed, 16 insertions(+), 22 deletions(-)
>>>>> This isn't a whole lot of code, are there others?  We're losing code
>>>>> clarity in favour of saving 6 lines?
>>>>>
>>>> Oh.  I thought it was a nice cleanup which made things more clear.
>>> I'm not totally against it; that's why I suggested the changes below. I
>>> think a name change would go a long way for clarity. It's not as much as
>>> I though it would be though.
>>>
>>>>>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>>>>>> index f5a97dec5169..abb4534be3cc 100644
>>>>>> --- a/include/linux/mm.h
>>>>>> +++ b/include/linux/mm.h
>>>>>> @@ -3516,6 +3516,15 @@ static inline bool range_in_vma(struct vm_area_struct *vma,
>>>>>>    	return (vma && vma->vm_start <= start && end <= vma->vm_end);
>>>>>>    }
>>>>>> +static inline void vma_range_init(struct vm_area_struct *vma,
>>>>> Any reason this can't be in mm/internal.h ?
>>>> That would be good.
>>> One other thing, do we trust this to be inlined correctly by the
>>> compiler or should this be __always_inline?  I'd expect it to be okay as
>>> it is, but I've been proven wrong in a perf trace before..
>>>
>> Okay, I would take __always_inline and put it in mm/internal.h in v2.
> I'm not confident in this suggestion as the rest.
> Please rename the function when you move it.

inline is a suggestion, __always_inline is mandatory.
I think __always_inline is better if we're demanding.

>>>>> vma_range_set(), vma_set_range(), or just vma_range() might be a better
>>>>> name?  My thinking is that some of these are actually modifying the vma
>>>>> and not just initializing it, right?
>>>> I'd vote for vma_set_range().
>                        ^  This part, use vma_set_range() please.

Okay.


By the way, I sent another patch with ("mm/mmap: simplify vma_merge()") 
subject a few days ago.

Please comment if you would like.

>>> Using vma_set_range() leaves vma_range() or vma_size(), which could be
>>> added for the calculations of vma->vm_end - vma->vm_start.  Davidlohr
>>> suggested such a beast a few years ago, but that one would need to live
>>> in the include/linux/mm.h as it occurs a lot more.
>>>
>>> $ git grep "vma->vm_end - vma->vm_start" | wc -l
>>> 198
>>>
>>> .. for just those named vma.
>>>