Re: [PATCH] THP: need is_vma_temporary_stack() when reference transparent_hugepage_enabled

[Alex Shi <alex.shi@intel.com>]

On 05/03/2012 07:25 PM, Andrea Arcangeli wrote:

> Hi,
> 
> On Thu, May 03, 2012 at 08:56:57AM +0800, Alex Shi wrote:
>>
>>> My preference would still be to remove the is_vma_temporary_stack and
>>> use two vmas during mremap of execve, that would remove the "vma"
>>> parameter from transparent_hugepage_enabled() but others prefers to
>>> skip a vma allocation in execve and stick to is_vma_temporary_stack,
>>> which is fair enough argument.
>>
>>
>> Actually, current transparent_hugepage_enabled just means the vma is in
>> THP enable ENV, the vma is just possibly has some large page, no grantee
>> really has.  But in lots situations, user wants to know if a vma or a
>> part of memory really include a large page. not the possibility.
>>
>> So, it will be great to see a real large page checking function appearing.
> 
> Well, to know if a VMA (or a memory range) really includes a THP, it'd
> require to hold the page_table_lock and a loop on all pmds in the
> range, but by the time you relase the lock things may have already
> changed as split_huge_page can run at any time, madvise(MADV_DONTNEED)
> too if you only hold the mmap_sem in read mode and the THP page
> fault. In fact while holding the mmap_sem in read mode (the usual read
> lock you need to take to lookup and stabilize the vma) a THP can be
> freed and reallocated under it, and that's what pmd_trans_unstable is
> about.


Appreciate for explanation!