Re: [PATCH] Exclude zero size section for RELRO segment match

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

On Fri, Feb 16, 2018 at 3:44 PM, Alan Modra <amodra@gmail.com> wrote:
> On Fri, Feb 16, 2018 at 04:13:04AM -0800, H.J. Lu wrote:
>> On Fri, Feb 16, 2018 at 4:08 AM, Alan Modra <amodra@gmail.com> wrote:
>> > On Thu, Feb 15, 2018 at 03:58:16AM -0800, H.J. Lu wrote:
>> >> When seaching a LOAD segment which overlaps with the RELRO segment, we
>> >> need to check the size of the last non-zero size section in the LOAD
>> >> segment.
>> >
>> > Why?  Do you have a case where the current algorithm goes wrong?
>> >
>>
>> Yes, apply this patch and I get assert in
>> assign_file_positions_for_non_load_sections
>> on x86-64:
>
> On looking at the tbss1 fail, I see a PT_LOAD with a single
> zero length .eh_frame section.  So we have a degenerate PT_LOAD.
> I think the following simpler patch is sufficient.
>
> Your testsuite change is OK, but please check a large set of targets
> before committing in case some targets won't support -z separate-code.

These tests are only run on Linux:

# Make sure that binutils can correctly handle ld output in ELF.

if { ![istarget *-*-linux*]
     && ![istarget *-*-nacl*]
     && ![istarget *-*-gnu*]} {
    return
}

if { [istarget *-*-linux*aout*]
     || [istarget *-*-linux*oldld*] } {
    return
}

I tested a couple Linux targets and checked it in.

Thanks.

>         PR 22845
>         * elf.c (assign_file_positions_for_non_load_sections): Ignore
>         degenerate zero size PT_LOAD segments when finding one overlapping
>         the PT_GNU_RELRO segment.
>
> diff --git a/bfd/elf.c b/bfd/elf.c
> index 397fa62..b069b59 100644
> --- a/bfd/elf.c
> +++ b/bfd/elf.c
> @@ -5897,6 +5897,7 @@ assign_file_positions_for_non_load_sections (bfd *abfd,
>                    lm = lm->next, lp++)
>                 {
>                   if (lp->p_type == PT_LOAD
> +                     && lp->p_memsz != 0
>                       && lm->count != 0
>                       && lm->sections[lm->count - 1]->vma >= start
>                       && lm->sections[0]->vma < end)
>
> --
> Alan Modra
> Australia Development Lab, IBM



-- 
H.J.