Re: [PATCH 0/2] mm/userfaultfd: Fix uninitialized output field for -EAGAIN race

[James Houghton <jthoughton@google.com>]

On Fri, Apr 25, 2025 at 11:58 AM David Hildenbrand <david@redhat.com> wrote:
>
> On 25.04.25 17:45, James Houghton wrote:
> > On Thu, Apr 24, 2025 at 5:57 PM Peter Xu <peterx@redhat.com> wrote:
> >>
> >> When discussing some userfaultfd issues with Andrea, Andrea pointed out an
> >> ABI issue with userfaultfd that existed for years.  Luckily the issue
> >> should only be a very corner case one, and the fix (even if changing the
> >> kernel ABI) should only be in the good way, IOW there should have no risk
> >> breaking any userapp but only fixing.
> >
> > FWIW, my userspace basically looks like this:
> >
> > struct uffdio_continue uffdio_continue;
> > int64_t target_len = /* whatever */;
> > int64_t bytes_mapped = 0;
> > int ioctl_ret;
> > do {
> >    uffdio_continue.range = /* whatever */;
> >    uffdio_continue.mapped = 0;
> >    ioctl_ret = ioctl(uffd, UFFDIO_CONTINUE, &uffdio_continue);
> >    if (uffdio_continue.mapped < 0) { break; }
> >    bytes_mapped += uffdio_continue.mapped;
> > } while (bytes_mapped < target_len && errno == EAGAIN);
> >
> > I think your patch would indeed break this. (Perhaps I shouldn't be
> > reading from `mapped` without first checking that errno == EAGAIN.)
> >
> > Well, that's what I would say, except in practice I never actually hit
> > the mmap_changing case while invoking UFFDIO_CONTINUE. :)
>
> Hm, but what if mfill_atomic_continue() would already return -EAGAIN
> when checking mmap_changing etc?
>
> Wouldn't code already run into an issue there?

Ah, thanks David. You're right, my code is already broken! :(

So given that we already have a case where -EAGAIN is put in the
output field, I change my mind, let's keep putting -EAGAIN in the
output field, and I'll go fix my code.