Re: [PATCH 0/2] don't use mapcount() to check large folio sharing

["Yin, Fengwei" <fengwei.yin@intel.com>]

On 8/2/2023 8:49 PM, Ryan Roberts wrote:
> On 02/08/2023 13:42, Yin, Fengwei wrote:
>>
>>
>> On 8/2/2023 8:40 PM, Ryan Roberts wrote:
>>> On 02/08/2023 13:35, Yin, Fengwei wrote:
>>>>
>>>>
>>>> On 8/2/2023 6:27 PM, Ryan Roberts wrote:
>>>>> On 28/07/2023 17:13, Yin Fengwei wrote:
>>>>>> In madvise_cold_or_pageout_pte_range() and madvise_free_pte_range(),
>>>>>> folio_mapcount() is used to check whether the folio is shared. But it's
>>>>>> not correct as folio_mapcount() returns total mapcount of large folio.
>>>>>>
>>>>>> Use folio_estimated_sharers() here as the estimated number is enough.
>>>>>>
>>>>>> Yin Fengwei (2):
>>>>>>   madvise: don't use mapcount() against large folio for sharing check
>>>>>>   madvise: don't use mapcount() against large folio for sharing check
>>>>>>
>>>>>>  mm/huge_memory.c | 2 +-
>>>>>>  mm/madvise.c     | 6 +++---
>>>>>>  2 files changed, 4 insertions(+), 4 deletions(-)
>>>>>>
>>>>>
>>>>> As a set of fixes, I agree this is definitely an improvement, so:
>>>>>
>>>>> Reviewed-By: Ryan Roberts
>>>> Thanks.
>>>>
>>>>>
>>>>>
>>>>> But I have a couple of comments around further improvements;
>>>>>
>>>>> Once we have the scheme that David is working on to be able to provide precise
>>>>> exclusive vs shared info, we will probably want to move to that. Although that
>>>>> scheme will need access to the mm_struct of a process known to be mapping the
>>>>> folio. We have that info, but its not passed to folio_estimated_sharers() so we
>>>>> can't just reimplement folio_estimated_sharers() - we will need to rework these
>>>>> call sites again.
>>>> Yes. This could be extra work. Maybe should delay till David's work is done.
>>>
>>> What you have is definitely an improvement over what was there before. And is
>>> probably the best we can do without David's scheme. So I wouldn't delay this.
>>> Just pointing out that we will be able to make it even better later on (if
>>> David's stuff goes in).
>> Yes. I agree that we should wait for David's work ready and do fix based on that.
> 
> I was suggesting the opposite - not waiting. Then we can do separate improvement
> later.
Let's wait for David's work ready. Otherwise, some other similar fix could be
submitted. Unfortunately, we can't build folio_estimated_sharers() based on
folio_maybe_mapped_shared() because latter one requires the extra parameter.


Regards
Yin, Fengwei

> 
>>
>>
>> Regards
>> Yin, Fengwei
>>
>>>
>>>>
>>>>>
>>>>> Given the aspiration for most of the memory to be large folios going forwards,
>>>>> wouldn't it be better to avoid splitting the large folio where the large folio
>>>>> is mapped entirely within the range of the madvise operation? Sorry if this has
>>>>> already been discussed and decided against - I didn't follow the RFC too
>>>>> closely. Or perhaps you plan to do this as a follow up?
>>>> Yes. We are on same page. RFC patchset did that. But there are some other opens
>>>> on the RFC. So I tried to submit this part of change which is bug fix. The other
>>>> thing left in RFC is optimization (avoid split large folio if we can).
>>>>
>>>>
>>>> Regards
>>>> Yin, Fengwei
>>>>
>>>>>
>>>>> Thanks,
>>>>> Ryan
>>>>>
>>>
>>>
>