Re: [PATCH RFC 05/19] mm: add early FAULT_FLAG_WRITE consistency checks

[David Hildenbrand]

On 07.11.22 20:03, Nadav Amit wrote:
> On Nov 7, 2022, at 8:17 AM, David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> > !! External Email
> >
> > Let's catch abuse of FAULT_FLAG_WRITE early, such that we don't have to
> > care in all other handlers and might get "surprises" if we forget to do
> > so.
> >
> > Write faults without VM_MAYWRITE don't make any sense, and our
> > maybe_mkwrite() logic could have hidden such abuse for now.
> >
> > Write faults without VM_WRITE on something that is not a COW mapping is
> > similarly broken, and e.g., do_wp_page() could end up placing an
> > anonymous page into a shared mapping, which would be bad.
> >
> > This is a preparation for reliable R/O long-term pinning of pages in
> > private mappings, whereby we want to make sure that we will never break
> > COW in a read-only private mapping.
> >
> > Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> > ---
> > mm/memory.c | 8 ++++++++
> > 1 file changed, 8 insertions(+)
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index fe131273217a..826353da7b23 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -5159,6 +5159,14 @@ static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
> >                  */
> >                 if (!is_cow_mapping(vma->vm_flags))
> >                         *flags &= ~FAULT_FLAG_UNSHARE;
> > +       } else if (*flags & FAULT_FLAG_WRITE) {
> > +               /* Write faults on read-only mappings are impossible ... */
> > +               if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
> > +                       return VM_FAULT_SIGSEGV;
> > +               /* ... and FOLL_FORCE only applies to COW mappings. */
> > +               if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
> > +                                !is_cow_mapping(vma->vm_flags)))
> > +                       return VM_FAULT_SIGSEGV;
>
> Not sure about the WARN_*(). Seems as if it might trigger in benign even if
> rare scenarios, e.g., mprotect() racing with page-fault.

We most certainly would want to catch any such broken/racy cases. There 
are no benign cases I could possibly think of.

Page faults need the mmap lock in read. mprotect() / VMA changes need 
the mmap lock in write. Whoever calls handle_mm_fault() is supposed to 
properly check VMA permissions.


--
Thanks,

David / dhildenb