Re: [PATCH] ld.so: Handle read-only dynamic section gracefully [BZ #28340]

["H.J. Lu" <hjl.tools@gmail.com>]

On Thu, Sep 16, 2021 at 10:59 AM Siddhesh Poyarekar
<siddhesh@sourceware.org> wrote:
>
> On 9/16/21 11:08 PM, H.J. Lu wrote:
> > On Thu, Sep 16, 2021 at 9:45 AM Siddhesh Poyarekar
> > <siddhesh@sourceware.org> wrote:
> >>
> >> On 9/16/21 8:48 PM, H.J. Lu wrote:
> >>> On Thu, Sep 16, 2021 at 7:11 AM Carlos O'Donell <carlos@redhat.com> wrote:
> >>>>
> >>>> On 9/16/21 00:48, Florian Weimer via Libc-alpha wrote:
> >>>>> * H. J. Lu:
> >>>>>
> >>>>>> There is nothing wrong with read-only dynamic segment.
> >>>>>
> >>>>> A relocated DYNAMIC array is part of the ABI for !DL_RO_DYN_SECTION.
> >>>>> ELF requires that DT_STRTAB is present.  DT_STRTAB needs relocation.
> >>>>> This means that for !DL_RO_DYN_SECTION, the dynamic segment cannot be in
> >>>>> a read-ony LOAD segment for a valid ELF file.
> >>>>
> >>>> I agree strongly with this position.
> >>>>
> >>>> Even with PT_GNU_RELRO, we must only go in one direction from RW -> RO (to avoid
> >>>> other security issues e.g. hardening not loosening the restrictions).
> >>>>
> >>>> In theory the vDSO is invalid.
> >>>>
> >>>> In practice it is a DL_RO_DYN_SECTION DSO but selected dynamically at runtime
> >>>> rather than statically at compile time for the target.
> >>
> >> Actually it isn't.  The dynamic section in the vdso is already relocated
> >> by the kernel when it's mapped in.  DL_RO_DYN_SECTION DSOs are not
> >> relocated because of which any references to pointers written in the
> >> .dynamic section need to be relocated.
> >
> > No.  Relocation of vDSO dynamic section is done by elf_get_dynamic_info.
> > Here is a patch to remove the hack for vDSO.
> >
>
> Ahh, I misunderstood the comment in setup_vdso, sorry.  I verified by
> running under the debugger that the kernel doesn't adjust .dynamic
> entries before mapping the vdso.
>
> This updated patch is definitely better IMO but it still doesn't resolve
> the two outstanding questions posed so far.  I'm on the fence about the
> first one (it's imprecise but the cost of imprecision doesn't seem high
> enough to warrant the extra check) but slanted slightly towards
> allocating memory and writing out relocated addresses in the interest of
> keeping the user experience with dl_iterate_phdr and friends consistent.
>
> 1. Should the readonly decision be based solely on DYNAMIC flags or also
> consider flags on the encompassing LOAD segment?

It should purely depend on PT_DYNAMIC segment.

> 2. Do we want to leave .dynamic unrelocated for read-only DYNAMIC or
> should we instead allocate an array to write relocated addresses in
> there, like we did for vdso?

My patch removed the hack for vDSO to leave the read-only
PT_DYNAMIC segment unrelocated.

-- 
H.J.