Re: mm: VM_BUG_ON_PAGE(PageTail(page)) in mbind

From: Kirill A. Shutemov
Date: Tue Jan 26 2016 - 16:07:51 EST


On Tue, Jan 26, 2016 at 12:48:23PM -0800, Andrew Morton wrote:
> On Tue, 26 Jan 2016 22:28:29 +0200 "Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> wrote:
>
> > The patch below fixes the issue for me, but this bug makes me wounder how
> > many bugs like this we have in kernel... :-/
> >
> > Looks like we are too permissive about which VMA is migratable:
> > vma_migratable() filters out VMA by VM_IO and VM_PFNMAP.
> > I think VM_DONTEXPAND also correlate with VMA which cannot be migrated.
> >
> > $ git grep VM_DONTEXPAND drivers | grep -v '\(VM_IO\|VM_PFNMAN\)' | wc -l
> > 33
> >
> > Hm.. :-|
> >
> > It worth looking on them closely... And I wouldn't be surprised if some
> > VMAs without all of these flags are not migratable too.
> >
> > Sigh.. Any thoughts?
>
> Sigh indeed. I think that both VM_DONTEXPAND and VM_DONTDUMP are
> pretty good signs that mbind() should not be mucking with this vma. If
> such a policy sometimes results in mbind failing to set a policy then
> that's not a huge loss - something runs a bit slower maybe.
>
> I mean, we only really expect mbind() to operate against regular old
> anon/pagecache memory, yes?

Well, it can work fine too if driver itself uses page tables to find out
which pages it should to operate on. I don't think it's a common case.

--
Kirill A. Shutemov