On Mon, May 03, 2010 at 02:40:44AM +0900, Minchan Kim wrote:
> On Fri, Apr 30, 2010 at 1:21 AM, Andrea Arcangeli <aarcange@redhat.com> wrote:
> > Hi Mel,
> >
> > did you see my proposed fix? I'm running with it applied, I'd be
> > interested if you can test it. Surely it will also work for new
> > anon-vma code in upstream, because at that point there's just 1
> > anon-vma and nothing else attached to the vma.
> >
> > http://git.kernel.org/?p=linux/kernel/git/andrea/aa.git;a=commit;h=6efa1dfa5152ef8d7f26beb188d6877525a9dd03
> >
> > I think it's wrong to try to handle the race in rmap walk by making
> > magic checks on vm_flags VM_GROWSDOWN|GROWSUP and
> > vma->vm_mm->map_count == 1, when we can fix it fully and simply in
> > exec.c by indexing two vmas in the same anon-vma with a different
> > vm_start so the pages will be found at all times by the rmap_walk.
> >
> 
> I like this approach than exclude temporal stack while migration.
> 
> If we look it through viewpoint of performance, Mel and Kame's one
> look good and simple. But If I look it through viewpoint of
> correctness, Andrea's one looks good.
> I mean Mel's approach is that problem is here but let us solve it with
> there. it makes dependency between here and there. And In future, if
> temporal stack and rmap code might be problem, we also should solve it
> in there. :)

That explains exactly why I wanted to solve it locally to exec.c and
by using the same method of mremap. And it fixes all users not just
migrate (split_huge_page may also need it in the future if we ever
allow the user stack to be born huge instead of growing huge with
khugepaged if needed).