[PATCH] x86-64: Increase output copy relocation section alignment

[H.J. Lu]

On Mon, Feb 17, 2025 at 9:01 PM Alan Modra <amodra@gmail.com> wrote:
>
> On Mon, Feb 17, 2025 at 06:14:29PM +0800, H.J. Lu wrote:
> >       PR ld/32690
> >       * elfxx-x86.c (_bfd_x86_elf_adjust_dynamic_symbol): Adjust the
> >       output copy relocation section alignment if needed.
>
> This is a tricky situation, isn't it?  Good job tracking down exactly

I will try to create a testcase.

> what was happening!  The alignment would propagate to the output
> section on the next sizing pass, but that's too late for the relro
> code on the current pass.
>
> > --- a/bfd/elfxx-x86.c
> > +++ b/bfd/elfxx-x86.c
> > @@ -3390,6 +3390,7 @@ _bfd_x86_elf_adjust_dynamic_symbol (struct bfd_link_info *info,
> >    struct elf_dyn_relocs *p;
> >    const struct elf_backend_data *bed
> >      = get_elf_backend_data (info->output_bfd);
> > +  bool ret;
> >
> >    eh = (struct elf_x86_link_hash_entry *) h;
> >
> > @@ -3608,7 +3609,23 @@ _bfd_x86_elf_adjust_dynamic_symbol (struct bfd_link_info *info,
> >        h->needs_copy = 1;
> >      }
> >
> > -  return _bfd_elf_adjust_dynamic_copy (info, h, s);
> > +  ret = _bfd_elf_adjust_dynamic_copy (info, h, s);
> > +  if (bed->target_id == X86_64_ELF_DATA)
>
> Why just for x86_64?

I only have a test for x86_64.

>
> > +    {
> > +      /* NB: Increase the output copy relocation section alignment so
> > +      that relro base and end values are properly computed.  */
> > +      asection *output_sec
> > +     = bfd_get_linker_section (info->output_bfd, s->name);
>
> Also, can't you use output_sec = s->output_section here?

For x86_64, s->output_section is NULL since one_lang_size_sections_pass
is run before allocation.

> > +      if (output_sec)
> > +     {
> > +       unsigned int power_of_two = bfd_section_alignment (s);
> > +       /* Adjust the section alignment if needed.  */
> > +       if (power_of_two > bfd_section_alignment (output_sec)
> > +           && !bfd_set_section_alignment (output_sec, power_of_two))
> > +         return false;
> > +     }
> > +    }
> > +  return ret;
> >  }
> >
> >  void
>
> In fact, can't you bump the output section alignment in
> _bfd_elf_adjust_dynamic_copy for all targets?
>
> diff --git a/bfd/elflink.c b/bfd/elflink.c
> index c9468ff42ac..61750e325c5 100644
> --- a/bfd/elflink.c
> +++ b/bfd/elflink.c
> @@ -3389,6 +3389,14 @@ _bfd_elf_adjust_dynamic_copy (struct bfd_link_info *info,
>        /* Adjust the section alignment if needed.  */
>        if (!bfd_set_section_alignment (dynbss, power_of_two))
>         return false;
> +      /* Adjust output section alignment too, so that layout adjusts
> +        for alignment on the current lang_size_sections pass.  This
> +        is important for lang_size_relro_segment which will use the
> +        input section alignment.  */
> +      asection *osec = dynbss->output_section;

This won't work for x86-64 for the reason I mentioned above.

> +      if (osec && power_of_two > bfd_section_alignment (osec)
> +         && !bfd_set_section_alignment (osec, power_of_two))
> +       return false;
>      }
>
>    /* We make sure that the symbol will be aligned properly.  */
>
>
> --
> Alan Modra



-- 
H.J.