Received-SPF: pass (google.com: domain of linux-kernel+bounces-210933-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Precedence: bulk
MIME-Version: 1.0
References: <20240612-as-instr-opt-wrussq-v2-1-bd950f7eead7@gmail.com>
In-Reply-To: <20240612-as-instr-opt-wrussq-v2-1-bd950f7eead7@gmail.com>
From: Masahiro Yamada <masahiroy@kernel.org>
Date: Wed, 12 Jun 2024 14:06:18 +0900
Message-ID: <CAK7LNARe9ijaNT_FSfM2M5pJh-+89REmU1xZxXO-eM_nZnfkiA@mail.gmail.com>
Subject: Re: [PATCH RFC v2] kbuild/x86: Use $(KBUILD_AFLAGS) in Kbuild's
 version of $(as-instr)
To: 0x7f454c46@gmail.com
Cc: Borislav Petkov <bp@alien8.de>, Dave Hansen <dave.hansen@linux.intel.com>, 
	Ingo Molnar <mingo@redhat.com>, "H. Peter Anvin" <hpa@zytor.com>, Thomas Gleixner <tglx@linutronix.de>, x86@kernel.org, 
	linux-kernel@vger.kernel.org, linux-kbuild@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Wed, Jun 12, 2024 at 12:29=E2=80=AFPM Dmitry Safonov via B4 Relay
<devnull+0x7f454c46.gmail.com@kernel.org> wrote:
>
> From: Dmitry Safonov <0x7f454c46@gmail.com>
>
> 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[sha=
1]=3D6571ad50d8f12eece053f1bac7a95a2c767f32c9, for GNU/Linux 3.2.0, strippe=
d
>
> 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 +=3D -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  :=3D (
> >
> >  # $(if-success,<command>,<then>,<else>)
> >  # Return <then> if <command> exits with 0, <else> otherwise.
> > -if-success =3D $(shell,{ $(1); } >/dev/null 2>&1 && echo "$(2)" || ech=
o "$(3)")
> > +if-success =3D $(shell,echo '$(1)' 1>&2;{ $(1); } >/dev/null 2>&1 && e=
cho "$(2)" || echo "$(3)")
>
> I got the following output for the toolchain check, before:
> > linux @Bru-na-Boinne% make ARCH=3Dx86_64 oldconfig V=3D1 2>&1 | grep wr=
us
> > printf "%b\n" "wrussq %rax,(%rbx)" | gcc  -c -x assembler-with-cpp -o /=
dev/null -
>
> and after:
> > linux @Bru-na-Boinne% make ARCH=3Dx86_64 oldconfig V=3D1 2>&1 | grep wr=
us
> > 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 =3D> arch/.../Makefile =3D> Kconfig =3D> ...
> 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@gmail.com>


Perhaps, you could pass CROSS_COMPILE=3Dx86_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@kernel.org/T=
/#u


--=20
Best Regards
Masahiro Yamada