On Tue, Mar 11, 2025 at 03:06:30PM +0100, Heiko Carstens wrote:
> On Tue, Mar 11, 2025 at 01:42:01PM +0000, Lorenzo Stoakes wrote:
> > > Just like for arm64, and x86_64 the s390 part is just adding the new
> > > vm flag to the _install_mappings() call in vdso code. Otherwise there
> > > is nothing to be considered.
> >
> > No, they are not just adding a flag, they are enabling the sealing of
> > system mappings, if a user chooses to make use of it, which already breaks
> > a number of useland applications that rely on remapping these.
> >
> > if the architecture code ever needs to unmap/remap these, then this breaks
> > your architecture.
> >
> > I think it would be sensible to clearly indicate that enabling this feature
> > does not break the s390 architecture and you've confirmed that by checking
> > the code and ensuring that nowhere does it rely upon doing this.
> >
> > Likely that's the case, but I'd suggest you ought to make sure...
> >
> > x86-64 and arm64 were checked for this and confirmed to not ever need this.
> >
> > This is why we're restricting by architecture.
>
> Ok, I was assuming more that whoever enables that config option knows
> what he or she is doing. However as far as I know there is no s390
> user space which relies on remapping vdso mappings.
Yeah, we allow for the 'user knows what they're doing' stuff when it comes to
_userland_, but we obviously don't want to ship a known-broken kernel :)
>
> When it comes to unmapping vdso: this would break user space since
> commit df29a7440c4b ("s390/signal: switch to using vdso for sigreturn
> and syscall restart") - there haven't been any reports.
OK that seems to implicitly suggest that you're fine with sealing then?
I had a quick glance and it seemed fine for s390. I mean it's _weird_ to remap
any of this stuff so it'd be the odd one out that does it (ppc _seems_ to in
_some_ circumstances need it, for instance).
So I think we're good? :)
