Re: [PATCH mm-unstable v1] mm: add a total mapcount for large folios

[Yin Fengwei]

On 8/10/23 03:26, David Hildenbrand wrote:
> On 09.08.23 21:21, Matthew Wilcox wrote:
>> On Wed, Aug 09, 2023 at 08:07:43PM +0100, Ryan Roberts wrote:
>>>> +++ b/mm/hugetlb.c
>>>> @@ -1479,7 +1479,7 @@ static void __destroy_compound_gigantic_folio(struct folio *folio,
>>>>       struct page *p;
>>>>         atomic_set(&folio->_entire_mapcount, 0);
>>>> -    atomic_set(&folio->_nr_pages_mapped, 0);
>>>> +    atomic_set(&folio->_total_mapcount, 0);
>>>
>>> Just checking this is definitely what you intended? _total_mapcount is -1 when
>>> it means "no pages mapped", so 0 means 1 page mapped?
>>
>> We're destroying the page here, so rather than setting the meaning of
>> this, we're setting the contents of this memory to 0.
>>
>>
>> Other thoughts that ran through my mind ... can we wrap?  I don't think
>> we can; we always increment total_mapcount by 1, no matter whether we're
>> incrementing entire_mapcount or an individual page's mapcount, and we
>> always call folio_get() first, so we can't increment total_mapcount
>> past 2^32 because folio_get() will die first.  We might be able to
>> wrap past 2^31, but I don't think so.
> 
> From my understanding, if we wrap the total mapcount, we already wrapped the refcount -- as you say, grabbing a reference ahead of time for each mapping is mandatory. Both are 31bit values. We could treat the total mapcount as an unsigned int, but that's rather future work.
> 
> Also, even folio_mapcount() and total_mapcount() return an "int" as of now.
> 
> But yes, I also thought about that. In the future we might want (at least) for bigger folios refcount+total_mapcount to be 64bit. Or we manage to decouple both and only have the total_mapcount be 64bit only.
This means pvmw may need to check more than 2^32 pte entries. :). I hope
we have other way to get bigger folio and keep mapcount not too big.


Regards
Yin, Fengwei

>