Re: [musl] Initial xtensa/fdpic port review

[Max Filippov <jcmvbkbc@gmail.com>]

On Wed, Feb 28, 2024 at 10:30 AM Rich Felker <dalias@libc.org> wrote:
> On Wed, Feb 28, 2024 at 09:20:33AM -0800, Max Filippov wrote:
> > > > diff --git a/ldso/dynlink.c b/ldso/dynlink.c
> > > > index ceca3c98..25563af3 100644
> > > > --- a/ldso/dynlink.c
> > > > +++ b/ldso/dynlink.c
> > > > @@ -1420,6 +1420,7 @@ static void reloc_all(struct dso *p)
> > > >               if (!DL_FDPIC)
> > > >                       do_relr_relocs(p, laddr(p, dyn[DT_RELR]), dyn[DT_RELRSZ]);
> > > >
> > > > +#if 0
> > > >               if (head != &ldso && p->relro_start != p->relro_end) {
> > > >                       long ret = __syscall(SYS_mprotect, laddr(p, p->relro_start),
> > > >                               p->relro_end-p->relro_start, PROT_READ);
> > > > @@ -1429,6 +1430,7 @@ static void reloc_all(struct dso *p)
> > > >                               if (runtime) longjmp(*rtld_fail, 1);
> > > >                       }
> > > >               }
> > > > +#endif
> > >
> > > Was this breaking something? Relro linking probably should have been
> > > disabled on fdpic, and we ignore ENOSYS anyway for nommu.
> >
> > Yeah, I do some of the testing in linux-user QEMU, it has MMU
> > and mprotect calls actually succeed. Failures were coming from the
> > relocation code, but my impression was that relro should be applied
> > after the relocation completion and it should just work, hence the
> > WIP pile.
>
> Yes, relro is only supposed to be applied after all relocations were
> done.

So I've found two issues with this. First is that loadmap entries generated
by the QEMU (and by the linux kernel AFAICS) are not page-aligned,
but the relro segment addresses fetched by the loader are. Since the
laddr() doesn't check that a translation for the address it was given
was actually found the results are not always correct, mprotect fails
and it all terminates early.

With a workaround for that part in place I see that relro protection is
applied to the executable image before its __fdpic_fixup had a chance
to run, there are rofixups for the .init_array and .fini_array, but they both
are a part of the relro segment.

-- 
Thanks.
-- Max