Re: Folio mapcount

[Matthew Wilcox <willy@infradead.org>]

On Tue, Jul 04, 2023 at 09:22:07AM +0800, Yin, Fengwei wrote:
> 
> 
> On 7/3/2023 9:24 PM, Zi Yan wrote:
> > On 2 Jul 2023, at 21:09, Yin, Fengwei wrote:
> > 
> >> On 7/3/2023 3:51 AM, Zi Yan wrote:
> >>>> (3) Mapcount of first (or any other) subpage (compount+subpage): for
> >>>>     folio_estimated_sharers().
> >>> This is another estimation. I wonder if we can use a different estimation
> >>> like total_mapcount() > folio_nr_pages() instead.
> >> Considering the partial folio mapping, I suppose we should use
> >>    total_mapcount() > folio->_nr_pages_mapped
> >> as folio_estimated_sharers().
> > 
> > What you propose is to get a precise check instead of estimate, and you assume no PMD mapping and still require per-page mapcount.
> > 
> > What I am proposing is to get rid of per-page mapcount, which is my goal, and use a single mapcount to do a rough estimate.
> O. Sorry. I didn't notice your goal. So if the rough estimate is enough,
> total_mapcount() > folio_nr_pages() works.
> 
> Do we need to check all the cases to make sure the rough estimation is
> enough for all of them?

That's definitely not enough for deciding whether we need to COW a page or
not.  We need to be able to tell the difference between an order-4 folio
with page 1 mapped in two tasks and an order-4 folio with page 0 & page
1 mapped in a single task.