Re: A mapcount riddle

[Peter Xu <peterx@redhat.com>]

On Wed, Jan 25, 2023 at 10:16:58AM -0800, Mike Kravetz wrote:
> On 01/25/23 11:46, Peter Xu wrote:
> > On Tue, Jan 24, 2023 at 03:35:38PM -0800, Mike Kravetz wrote:
> > > On 01/24/23 18:00, Peter Xu wrote:
> > > > On Tue, Jan 24, 2023 at 12:56:24PM -0800, Mike Kravetz wrote:
> > > > Besides, I'm also curious on the planned fix too regarding the two issues
> > > > mentioned.
> > > 
> > > My planned 'fix' is to simply check for shared a PMD
> > > (page_count(virt_to_page(pte))) to determine if page with mapcount == 1
> > > is shared.
> > 
> > I think having the current pte* won't easily work, we'll need to walk all
> > the pgtable that mapped this page.
> > 
> > To be explicit, one page can be mapped at pgtable1 which is shared by proc1
> > & proc2, and it can also be mapped at pgtable2 which is shared by proc3 &
> > proc4.  Then (assuming pte1* points to pgtable1):
> > 
> >   page_count(virt_to_page(pte1)) + page_mapcount(page)
> > 
> > Won't be the right mapcount we're looking for.
> 
> That assumes we want an accurate mapcount.  In the two code segments I pointed
> out,  we only need to know if more than one process maps the page.  We can
> get that with 'page_count(virt_to_page(pte)) + page_mapcount(page)'.

I see.  Yes it sounds working. We can check mapcount==1 first and only go
for the pgtable if mapcount>1, then we also know that's the only pmd it got
mapped at.

> 
> For now (and stable releases), I propose just fixing the improper behavior.
> If we really need an accurate mapcount in the shared PMD case (and I am not
> 100% sure we do), then we can add code to maintain or compute that.
> 
> Part of my reason for sending this email was to determine if there may
> be some places where we really do need an accurate mapcount in the shared
> PMD case.  I have not found any, but could have easily missed something.

Nothing I noticed either.

Thanks,

-- 
Peter Xu