On Wed, Jun 12, 2024 at 12:29 PM Dmitry Safonov via B4 Relay
<devnull+0x7f454c46.gmail.com@xxxxxxxxxx> wrote:
>
> From: Dmitry Safonov <0x7f454c46@xxxxxxxxx>
>
> At Arista some products use compatible 32-bit userspace running on x86.
> As a part of disto build for ia32 it also compiles the 64-bit kernel.
> While the toolchain for the kernel build is yet the same, with 64-bit gcc:
> > / @Bru-na-Boinne% file /usr/bin/gcc-11
> > /usr/bin/gcc-11: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2, BuildID[sha1]=6571ad50d8f12eece053f1bac7a95a2c767f32c9, for GNU/Linux 3.2.0, stripped
>
> It seems that gcc is being smart and detects that it's running in
> a 32-bit container (personality flag? 32-bit mmap base? something else
> inherited post-exec? haven't yet figured it out) and by default tries
> to build 32-bit binaries.
>
> That results in a failing toolchain check:
> > / @Bru-na-Boinne% printf "%b\n" "wrussq %rax, (%rbx)" | /usr/bin/gcc-11 -Wa,--fatal-warnings -c -x assembler-with-cpp -o /dev/null -
> > <stdin>: Assembler messages:
> > <stdin>:1: Error: `wrussq' is only supported in 64-bit mode
>
> Which passes when -m64 is directly specify for the build check:
> > / @Bru-na-Boinne% printf "%b\n" "wrussq %rax, (%rbx)" | /usr/bin/gcc-11 -m64 -Wa,--fatal-warnings -c -x assembler-with-cpp -o /dev/null -
> > / @Bru-na-Boinne% echo $?
> > 0
>
> As a result, kbuild produces different value for CONFIG_AS_WRUSS
> for native 64-bit containers and ia32 containers with 64-bit gcc,
> which produces different kernels with enabled/disabled
> CONFIG_X86_USER_SHADOW_STACK.
>
> arch/x86/Makefile already properly defines KBUILD_AFLAGS += -m64,
> which is luckly already available at the point of toolchain check
> in arch/x86/Kconfig.assembler
>
> By hacking around Kbuild variable the following way:
> > --- a/scripts/Kconfig.include
> > +++ b/scripts/Kconfig.include
> > @@ -13,7 +13,8 @@ left_paren := (
> >
> > # $(if-success,<command>,<then>,<else>)
> > # Return <then> if <command> exits with 0, <else> otherwise.
> > -if-success = $(shell,{ $(1); } >/dev/null 2>&1 && echo "$(2)" || echo "$(3)")
> > +if-success = $(shell,echo '$(1)' 1>&2;{ $(1); } >/dev/null 2>&1 && echo "$(2)" || echo "$(3)")
>
> I got the following output for the toolchain check, before:
> > linux @Bru-na-Boinne% make ARCH=x86_64 oldconfig V=1 2>&1 | grep wrus
> > printf "%b\n" "wrussq %rax,(%rbx)" | gcc -c -x assembler-with-cpp -o /dev/null -
>
> and after:
> > linux @Bru-na-Boinne% make ARCH=x86_64 oldconfig V=1 2>&1 | grep wrus
> > printf "%b\n" "wrussq %rax,(%rbx)" | gcc -D__ASSEMBLY__ -fno-PIE -m64 -c -x assembler-with-cpp -o /dev/null -
>
> Which seems appropriate to me.
> This also reflects the existing definition in scripts/Makefile.compiler
> for $(as-instr) that already has $(KBUILD_AFLAGS).
>
> In order to eliminate a possible circular dependency of
> Kconfig => arch/.../Makefile => Kconfig => ...
> which exist i.e. in arm64/Makefile for KASAN_SHADOW_SCALE_SHIFT that
> depends on CONFIG_KASAH_SW_TAGS and CONFIG_KASAN_GENERIC kconfigs,
> ignore KBUILD_AFLAGS difference in auto.conf.cmd as it is expected that
> the variable will differ between fist and later Makefile passes.
> Use that in Kconfig toolchain checks.
>
> Signed-off-by: Dmitry Safonov <0x7f454c46@xxxxxxxxx>
Perhaps, you could pass CROSS_COMPILE=x86_64-linux-gnu-
when building the 64-bit kernel.
x86_64-linux-gnu-gcc may understand 'wrussq %rax,(%rbx)'
even if the default gcc does not.
I am not sure if x86_64-linux-gnu-gcc is available
on your build machine.
Anyway, I think it should be possible to fix it with less hacky code.
Please test this patch:
https://lore.kernel.org/all/20240612050257.3670768-1-masahiroy@xxxxxxxxxx/T/#u
--
Best Regards
Masahiro Yamada
