Linux linking stage broken on ARM with binutils 2.33.1 (BFD)

[Nathan Chancellor]

Hi Martin,

On Tue, Jul 08, 2025 at 09:44:44PM +0200, Martin Wetterwald wrote:
> When I upgraded from 6.12.22 to 6.12.23 (without changing anything else), the
> linking stage of vmlinux broke (arch ARM).
> 
> > /home/.../linux-6.12.23/scripts/link-vmlinux.sh: line 49: 136961 Segmentation fault      (core dumped) ${ld} ${ldflags} -o ${output} ${wl}--whole-archive ${objs} ${wl}--no-whole-archive ${wl}--start-group ${libs} ${wl}--end-group ${kallsymso} ${btf_vmlinux_bin_o} ${ldlibs}
> > make[2]: *** [/home/.../linux-6.12.23/scripts/Makefile.vmlinux:34: vmlinux] Error 139
> > make[1]: *** [/home/.../linux-6.12.23/Makefile:1179: vmlinux] Error 2
> > make: *** [/home/.../linux-6.12.23/Makefile:224: __sub-make] Error 2
> 
> This patch seems related:
> 
> > commit 59fc42318305cb38efb4f5565408150419be8451
> > Author: Nathan Chancellor <nathan at kernel.org>
> > Date:   Thu Mar 20 22:33:49 2025 +0100
> >
> >     ARM: 9443/1: Require linker to support KEEP within OVERLAY for DCE
> >
> >     commit e7607f7d6d81af71dcc5171278aadccc94d277cd upstream.
> >
> >     ld.lld prior to 21.0.0 does not support using the KEEP keyword within an
> >     overlay description, which may be needed to avoid discarding necessary
> >     sections within an overlay with '--gc-sections', which can be enabled
> >     for the kernel via CONFIG_LD_DEAD_CODE_DATA_ELIMINATION.
> >
> >     Disallow CONFIG_LD_DEAD_CODE_DATA_ELIMINATION without support for KEEP
> >     within OVERLAY and introduce a macro, OVERLAY_KEEP, that can be used to
> >     conditionally add KEEP when it is properly supported to avoid breaking
> >     old versions of ld.lld.
> >
> >     Cc: stable at vger.kernel.org
> >     Link: https://github.com/llvm/llvm-project/commit/381599f1fe973afad3094e55ec99b1620dba7d8c
> >     Reviewed-by: Linus Walleij <linus.walleij at linaro.org>
> >     Signed-off-by: Nathan Chancellor <nathan at kernel.org>
> >     Signed-off-by: Russell King (Oracle) <rmk+kernel at armlinux.org.uk>
> >     [nathan: Fix conflict in init/Kconfig due to lack of RUSTC symbols]
> >     Signed-off-by: Nathan Chancellor <nathan at kernel.org>
> >     Signed-off-by: Greg Kroah-Hartman <gregkh at linuxfoundation.org>
> 
> I'm using BFD linker from GNU binutils 2.33.1 (built using musl-cross-make with
> musl 1.2.5 and GCC 14).
> 
> > scripts/ld-version.sh /tmp/arm-linux-musleabihf-ld
> > BFD 23301
> 
> I don't know if it was the intention of the patch, but it is not only changing
> a behavior for LLD, but also for BFD:
> - Before the patch, ld BFD 2.33.1 was not selecting
>   HAVE_LD_DEAD_CODE_DATA_ELIMINATION, and my build worked.
> - After the patch, ld BFD 2.33.1 is activating
>   HAVE_LD_DEAD_CODE_DATA_ELIMINATION, and it is segfaulting.
> Shouldn't we keep the previous behavior and only select this option when BFD is
> >= 2.36.0 (and not just when we have a BFD linker) ?
> 
> The issue is the interaction between LD_CAN_USE_KEEP_IN_OVERLAY and
> HAVE_LD_DEAD_CODE_DATA_ELIMINATION.
> 
> Nathan, I've seen your proposal here: https://lkml.org/lkml/2025/7/5/431.
> I confirm it also fixes the issue on my side. Should I propose the patch?

Yes, your analysis is spot on, sorry again for the breakage :(

I sent that diff to the list formally yesterday:

https://lore.kernel.org/all/20250707-arm-fix-dce-older-binutils-v1-1-3b5e59dc3b26@kernel.org/

I plan to put it in Russell's patch tracker for application in the next
couple of days, I wanted to wait a few days for the opportunity for
review.

Cheers,
Nathan