On Tue, Apr 07, 2020 at 10:27:15AM +0200, Dmitry Vyukov wrote:
> On Tue, Apr 7, 2020 at 4:43 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
> >
> > On Mon, Apr 06, 2020 at 07:15:34PM -0700, Andrew Morton wrote:
> > > On Mon, 6 Apr 2020 21:55:35 -0400 Peter Xu <peterx@xxxxxxxxxx> wrote:
> > >
> > > > On Mon, Apr 06, 2020 at 06:39:41PM -0700, Andrew Morton wrote:
> > > > > On Mon, 6 Apr 2020 20:47:45 -0400 Peter Xu <peterx@xxxxxxxxxx> wrote:
> > > > >
> > > > > > >From 23800bff6fa346a4e9b3806dc0cfeb74498df757 Mon Sep 17 00:00:00 2001
> > > > > > From: Peter Xu <peterx@xxxxxxxxxx>
> > > > > > Date: Mon, 6 Apr 2020 20:40:13 -0400
> > > > > > Subject: [PATCH] mm/mempolicy: Allow lookup_node() to handle fatal signal
> > > > > >
> > > > > > lookup_node() uses gup to pin the page and get node information. It
> > > > > > checks against ret>=0 assuming the page will be filled in. However
> > > > > > it's also possible that gup will return zero, for example, when the
> > > > > > thread is quickly killed with a fatal signal. Teach lookup_node() to
> > > > > > gracefully return an error -EFAULT if it happens.
> > > > > >
> > > > > > ...
> > > > > >
> > > > > > --- a/mm/mempolicy.c
> > > > > > +++ b/mm/mempolicy.c
> > > > > > @@ -902,7 +902,10 @@ static int lookup_node(struct mm_struct *mm, unsigned long addr)
> > > > > >
> > > > > > int locked = 1;
> > > > > > err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
> > > > > > - if (err >= 0) {
> > > > > > + if (err == 0) {
> > > > > > + /* E.g. GUP interupted by fatal signal */
> > > > > > + err = -EFAULT;
> > > > > > + } else if (err > 0) {
> > > > > > err = page_to_nid(p);
> > > > > > put_page(p);
> > > > > > }
> > > > >
> > > > > Doh. Thanks.
> > > > >
> > > > > Should it have been -EINTR?
> > > >
> > > > It looks ok to me too. I was returning -EFAULT to follow the same
> > > > value as get_vaddr_frames() (which is the other caller of
> > > > get_user_pages_locked()). So far the only path that I found can
> > > > trigger this is when there's a fatal signal pending right after the
> > > > gup. If so, the userspace won't have a chance to see the -EINTR (or
> > > > whatever we return) anyways.
> > >
> > > Yup. I guess we're a victim of get_user_pages()'s screwy return value
> > > conventions - the caller cannot distinguish between invalid-addr and
> > > fatal-signal.
> >
> > Indeed.
> >
> > >
> > > Which makes one wonder why lookup_node() ever worked. What happens if
> > > get_mempolicy(MPOL_F_NODE) is passed a wild userspace address?
> > >
> >
> > I'm not familiar with mempolicy at all, but do you mean MPOL_F_NODE
> > with MPOL_F_ADDR? Asked since iiuc if only MPOL_F_NODE is specified,
> > the kernel should not use the userspace addr at all (which seems to be
> > the thing we do now). get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) seems to
> > return -EFAULT as expected, though I agree maybe it would still be
> > nicer to differentiate the two cases.
>
> Am I reading this correctly that we put an initialized struct page* in
> this case? If so, with stack spraying this looks like an "interesting"
> bug.
Yeah, so far it should be fine, but... ideally I guess we should init
page==NULL in lookup_node() too to avoid potential risk on exploiting.
Maybe we could squash this into the fix if still possible.
Thanks,
--
Peter Xu
