[RFC V2] mm:add zero_page _mapcount when mapped into user space

[kirill@shutemov.name (Kirill A. Shutemov)]

On Thu, Dec 04, 2014 at 02:10:53PM +0800, Wang, Yalin wrote:
> > -----Original Message-----
> > From: Kirill A. Shutemov [mailto:kirill at shutemov.name]
> > Sent: Tuesday, December 02, 2014 7:30 PM
> > To: Wang, Yalin
> > Cc: 'linux-kernel at vger.kernel.org'; 'linux-mm at kvack.org'; 'linux-arm-
> > kernel at lists.infradead.org'
> > Subject: Re: [RFC V2] mm:add zero_page _mapcount when mapped into user
> > space
> > 
> > On Tue, Dec 02, 2014 at 05:27:36PM +0800, Wang, Yalin wrote:
> > > This patch add/dec zero_page's _mapcount to make sure the mapcount is
> > > correct for zero_page, so that when read from /proc/kpagecount,
> > > zero_page's mapcount is also correct, userspace process like procrank
> > > can calculate PSS correctly.
> > 
> > I don't have specific code path to point to, but I would expect zero page
> > with non-zero mapcount would cause a problem with rmap.
> > 
> > How do you test the change?
> > 
> I just test it to see the mapcount from /proc/pid/pagemap  and /proc/kpagecount ,
> It works well,

I took a closer look and your patch is broken in multiple places:
 - on zap_pte_range() you don't decrement mapcount;
 - you don't update rss counters for mm;
 - copy_one_pte() doesn't increase mapcount;
 - ...

Basically, each and every vm_normal_page() call must be audited. As first
step. And you totally skip huge zero page.

Proper mapcount handling for zero page would require a lot more work and I
don't think it worth it. Gain is too small.

NAK.

> The problem is that when I see /proc/pid/smaps ,
> The Rss / Pss don't calculate zero_page map,
> Because smaps_pte_entry() --> vm_normal_page( ),
> Will return NULL for zero_page,
> 
> But when userspace process cat /proc/pid/pagemap  ,
> It will see zero_page mapped,
> And will treat as Rss ,  
> This is weird, should we also omit zero_page in /proc/pid/pagemap ?
> Or add zero_page as Rss in /proc/pid/smaps ? 
> 
> I think we should add zero_page into Rss ,
> Because it is really mapped into userspace address space.
> And will let userspace memory analysis more accurate .

It would be easier for userspace to find out pfn of zero page and take it
into account.

Note: some architectures have multiple zero page due to coloring.

-- 
 Kirill A. Shutemov