Re: [PATCH v1] mm/gup: disallow FOLL_FORCE|FOLL_WRITE on hugetlb mappings

[David Hildenbrand]

On 31.10.22 17:14, Jason Gunthorpe wrote:
> On Mon, Oct 31, 2022 at 04:25:24PM +0100, David Hildenbrand wrote:
> > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > Cc: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> > Cc: Peter Xu <peterx@xxxxxxxxxx>
> > Cc: John Hubbard <jhubbard@xxxxxxxxxx>
> > Cc: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > Reported-by: syzbot+f0b97304ef90f0d0b1dc@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> > ---
> >
> > I assume this has been broken at least since 2014, when mm/gup.c came to
> > life. I failed to come up with a suitable Fixes tag quickly.
>
> I'm worried this would break RDMA over hugetlbfs maps - which is a
> real thing people do.
>
> MikeK do you have test cases?

This patch here only silences the warning. The warning+failing is 
already in 6.0, and so far nobody (besides syzbot) complained.

RDMA (due to FOLL_FORCE) would now fail (instead of doing something 
wrong) on MAP_PRIVATE hugetlb mappings that are R/O. Do we have any 
actual examples of such RDMA usage? I was able to understand why this 
case (MAP_PRIVATE, PROT_READ) is important for !hugetlb, but I don't 
immediately see under which situations this would apply to hugetlb.

While we could implement FOLL_FORCE for hugetlb, at least for RDMA we 
will be moving away from FOLL_FORCE instead --- I'll be posting these 
patches shortly.

So considering upcoming changes, at least RDMA is rather a bad excuse 
for more widespread FOLL_FORCE support.

--
Thanks,

David / dhildenb