Re: [PATCH v4 5/5] uprobe: collapse THP pmd after removing all uprobes

[Song Liu <songliubraving@fb.com>]

> On Jun 21, 2019, at 9:30 AM, Song Liu <songliubraving@fb.com> wrote:
> 
> 
> 
>> On Jun 21, 2019, at 6:45 AM, Song Liu <songliubraving@fb.com> wrote:
>> 
>> 
>> 
>>> On Jun 21, 2019, at 6:36 AM, Kirill A. Shutemov <kirill@shutemov.name> wrote:
>>> 
>>> On Fri, Jun 21, 2019 at 01:17:05PM +0000, Song Liu wrote:
>>>> 
>>>> 
>>>>> On Jun 21, 2019, at 5:48 AM, Kirill A. Shutemov <kirill@shutemov.name> wrote:
>>>>> 
>>>>> On Thu, Jun 13, 2019 at 10:57:47AM -0700, Song Liu wrote:
>>>>>> After all uprobes are removed from the huge page (with PTE pgtable), it
>>>>>> is possible to collapse the pmd and benefit from THP again. This patch
>>>>>> does the collapse.
>>>>>> 
>>>>>> An issue on earlier version was discovered by kbuild test robot.
>>>>>> 
>>>>>> Reported-by: kbuild test robot <lkp@intel.com>
>>>>>> Signed-off-by: Song Liu <songliubraving@fb.com>
>>>>>> ---
>>>>>> include/linux/huge_mm.h |  7 +++++
>>>>>> kernel/events/uprobes.c |  5 ++-
>>>>>> mm/huge_memory.c        | 69 +++++++++++++++++++++++++++++++++++++++++
>>>>> 
>>>>> I still sync it's duplication of khugepaged functinallity. We need to fix
>>>>> khugepaged to handle SCAN_PAGE_COMPOUND and probably refactor the code to
>>>>> be able to call for collapse of particular range if we have all locks
>>>>> taken (as we do in uprobe case).
>>>>> 
>>>> 
>>>> I see the point now. I misunderstood it for a while. 
>>>> 
>>>> If we add this to khugepaged, it will have some conflicts with my other 
>>>> patchset. How about we move the functionality to khugepaged after these
>>>> two sets get in? 
>>> 
>>> Is the last patch of the patchset essential? I think this part can be done
>>> a bit later in a proper way, no?
>> 
>> Technically, we need this patch to regroup pmd mapped page, and thus get 
>> the performance benefit after the uprobe is detached. 
>> 
>> On the other hand, if we get the first 4 patches of the this set and the 
>> other set in soonish. I will work on improving this patch right after that..
> 
> Actually, it might be pretty easy. We can just call try_collapse_huge_pmd() 
> in khugepaged.c (in khugepaged_scan_shmem() or khugepaged_scan_file() after 
> my other set). 
> 
> Let me fold that in and send v5. 

On a second thought, if we would have khugepaged to do collapse, we need a
dedicated bit to tell khugepaged which pmd to collapse. Otherwise, it may 
accidentally collapse pmd that are split by other split_huge_pmd. 

I could not think of a good solution here. In this case, we really need a
flag for this specific pmd. Flags for the compound_page or for the vma 
would not work, as those could be shared by multiple pmds. 

If the analysis above is correct, we probably need uprobe to specifically
call try_collapse_huge_pmd() for now.

Please share your suggestions. 

Thanks!
Song