Re: [PATCH v2] elf: Support DT_RELR relative relocation format [BZ #27924]

[Fangrui Song <maskray@google.com>]

On 2021-10-29, H.J. Lu wrote:
>On Fri, Oct 29, 2021 at 11:22 AM Carlos O'Donell via Binutils
><binutils@sourceware.org> wrote:
>>
>> On 10/16/21 20:50, Fangrui Song via Binutils wrote:
>> > PIE and shared objects usually have many relative relocations. In
>> > 2017/2018, SHT_RELR/DT_RELR was proposed on
>> > https://groups.google.com/g/generic-abi/c/bX460iggiKg/m/GxjM0L-PBAAJ
>> > ("Proposal for a new section type SHT_RELR") and is a pre-standard. RELR
>> > usually takes 3% or smaller space than R_*_RELATIVE relocations. The
>> > virtual memory size of a mostly statically linked PIE is typically 5~10%
>> > smaller.
>>
>> We've been going over this patch on the weekly Monday patch queue review.
>>
>> I took a note to point out that one of the blockers here is that it is difficult
>> to immediately test this work because it requires a working glibc build using
>> ldd (which has support for DT_RELR).
>>
>> What is the status of the lld support patches for glibc?
>>
>> If we made progress on the lld support then we'd be able to more easily review
>> a testable configuration and keep the review going forward.
>>
>
>I raised the mysterious crash issue:
>
>https://groups.google.com/g/generic-abi/c/bX460iggiKg
>
>We need an updated proposal without mysterious crashes.  One option is
>that the linker should bump EI_ABIVERSION when generating DT_RELR.

The generic ABI says :  "EI_ABIVERSION: ... The interpretation of this
version number is dependent on the ABI identified by the EI_OSABI
field."

Operating systems decide their EI_ABIVERSION.  Having a linker option
not bumping the ABI version can benefit some operating systems. We know
that FreeBSD (ELFOSABI_FREEBSD)/Fuchsia/ChromeOS don't find it necessary
(or don't want) to bump the ABI version.

See https://groups.google.com/g/generic-abi/c/vdG_G4l3N-Y for Solaris'
development model. They simply don't support (new object, old system) (I
believe that means a diagnostic is good but not required). "it's been
pretty drama free over many decades of SunOS."

---

So many folks on binutils/glibc are more interested on Linux and let's
discuss Linux:

Even within ELFOSABI_GNU, different architectures may have different
EI_ABIVERSION values.
I know that mips may use EI_ABIVERSION==1 for
(e_eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC
position-dependent executables.


If you think a linker option like -z relr=glibc or --pack-dyn-relocs=relr-glibc
is useful to rule out (DT_RELR object, glibc not supporting DT_RELR)
loudly, you may choose the `_dl_have_relr` in .dynsym scheme
(https://sourceware.org/pipermail/binutils/2021-October/118347.html).

I will probably just make such an option an alias for
--pack-dyn-relocs=relr because LLD has (1) some users who don't need this
checking (2) --pack-dyn-relocs=relr semantic change would be unnecessary
(3) I don't want them to migrate away from --pack-dyn-relocs=relr just
because glibc has a different development model.

>After this issue is resolved, I can look into the bfd linker support.

Will the above said, I grealy appreciate that you can take a stab on the
GNU ld support. This will benefit so many Linux distributions.

I learned that my /usr/bin/perf_* can be 20% smaller if switching to
DT_RELR :) Perhaps 5~10% is more typical for other executables.