Re: [PATCH] mm/hugetlb.c: fix unnecessary address expansion of pmd sharing

[Li Xinhai <lixinhai.lxh@gmail.com>]

On 1/5/21 2:59 AM, Mike Kravetz wrote:
> On 1/3/21 11:10 PM, Li Xinhai wrote:
>>
>>
>> On 1/4/21 11:55 AM, Mike Kravetz wrote:
>>> I believe the only case where your patch produced incorrect results is
>>> when the range was within a vma that was smaller than PUD_SIZE.  Do you
>>> agree?
>>>
>> Not exactly. We need to consider for vma which span at least one
>> PUD_SIZE after align its vm_start and vm_end.
>>
> 
> I know that I provided an incorrect example again.  Sorry (again)!
> 
> Can you provide an example where adding the simple check for vma size less
> than PUD_SIZE to your original patch will not work.  The logic in your V2
> patch is correct.  However, I am having a hard time finding a problem with
> this simpler approach.
> 


Thanks for checking. An example like this:
(vm_start, vm_end) = (2G - 6M, 2G+8M), so this vma is bigger than
PUD_SIZE, and the check for vma size bigger than PUD_SIZE will pass.

With (start, end) = (2G-4M, 2G-2M), the previous patch will not adjust
start(because adjust it from (2G-4M) to 1G will exceeding vm_start), but
it will adjust end from (2G-2M) to 2G (because 2G still below vm_end).

The adjustment of end is incorrect, because (2G-6M, 2G) range of vma
is not allowed for PMD sharing(i.e., that range do not fully occupy
PUD_SIZE). If we make the wrong adjustment, then we will unnecessarily
impact to range (2G-2M, 2G).