| Message ID | 20240612-as-instr-opt-wrussq-v2-1-bd950f7eead7@gmail.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC,v2] kbuild/x86: Use $(KBUILD_AFLAGS) in Kbuild's version of $(as-instr) | expand |
On Wed, Jun 12, 2024 at 12:29 PM 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[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@gmail.com> 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@kernel.org/T/#u
diff --git a/scripts/Kconfig.include b/scripts/Kconfig.include index 3ee8ecfb8c04..d8574c1faf2d 100644 --- a/scripts/Kconfig.include +++ b/scripts/Kconfig.include @@ -33,7 +33,7 @@ ld-option = $(success,$(LD) -v $(1)) # $(as-instr,<instr>) # Return y if the assembler supports <instr>, n otherwise -as-instr = $(success,printf "%b\n" "$(1)" | $(CC) $(CLANG_FLAGS) -Wa$(comma)--fatal-warnings -c -x assembler-with-cpp -o /dev/null -) +as-instr = $(success,printf "%b\n" "$(1)" | $(CC) $(CLANG_FLAGS) $(KBUILD_AFLAGS) -Wa$(comma)--fatal-warnings -c -x assembler-with-cpp -o /dev/null -) # check if $(CC) and $(LD) exist $(error-if,$(failure,command -v $(CC)),C compiler '$(CC)' not found) diff --git a/scripts/kconfig/preprocess.c b/scripts/kconfig/preprocess.c index f0a4a218c4a5..cf7389f13f00 100644 --- a/scripts/kconfig/preprocess.c +++ b/scripts/kconfig/preprocess.c @@ -61,6 +61,17 @@ static void env_del(struct env *e) free(e); } +static bool env_ignore(const char *name) +{ + /* + * Break the circular dependency Kconfig => arch/.../Makefile => Kconfig + * On the first Makefile pass only AFLAGS that do not depend on + * CONFIG_* set will be defined. Those are used by Kconfig for + * toolchain checks. + */ + return !strcmp(name, "KBUILD_AFLAGS"); +} + /* The returned pointer must be freed when done */ static char *env_expand(const char *name) { @@ -83,7 +94,8 @@ static char *env_expand(const char *name) * We need to remember all referenced environment variables. * They will be written out to include/config/auto.conf.cmd */ - env_add(name, value); + if (!env_ignore(name)) + env_add(name, value); return xstrdup(value); }