On Wed, May 17, 2023 at 1:14 AM Dave Hansen <dave.hansen@intel.com> wrote:
>
> On 5/15/23 06:05, jeffxu@chromium.org wrote:
> > --- a/arch/x86/mm/pkeys.c
> > +++ b/arch/x86/mm/pkeys.c
> > @@ -20,7 +20,7 @@ int __execute_only_pkey(struct mm_struct *mm)
> >       /* Do we need to assign a pkey for mm's execute-only maps? */
> >       if (execute_only_pkey == -1) {
> >               /* Go allocate one to use, which might fail */
> > -             execute_only_pkey = mm_pkey_alloc(mm);
> > +             execute_only_pkey = mm_pkey_alloc(mm, 0);
> >               if (execute_only_pkey < 0)
> >                       return -1;
> >               need_to_set_mm_pkey = true;
>
> In your threat model, what mechanism prevents the attacker from
> modifying executable mappings?

There are different options how we can address this:
1) having a generic mseal() API as Jeff mentioned
2) tagging all code pages with the pkey we're using
    (would this affect memory sharing between processes?)
3) prevent this with seccomp + userspace validation
If we have pkey support, we will only create executable memory using
pkey_mprotect and everything else can be blocked with seccomp. This would still
allow turning R-X memory into RW- memory, but you can't change it back without
going through our codepath that has added validation.

There's a similar challenge with making RO memory writable. For this we'll need
to use approach 1) or 2) instead.

> I was trying to figure out if the implicit execute-only pkey should have
> the PKEY_ENFORCE_API bit set.  I think that in particular would probably
> cause some kind of ABI breakage, but it still reminded me that I have an
> incomplete picture of the threat model.