[PATCH] x86/Kconfig: Fix CONFIG_CC_HAS_SANE_STACKPROTECTOR when cross compiling with clang

[PATCH] x86/Kconfig: Fix CONFIG_CC_HAS_SANE_STACKPROTECTOR when cross compiling with clang

[Nathan Chancellor]

Chimera Linux notes that CONFIG_CC_HAS_SANE_STACKPROTECTOR cannot be
enabled when cross compiling an x86_64 kernel with clang, even though it
does work when natively compiling.

When building on aarch64:

  $ make -sj"$(nproc)" ARCH=x86_64 LLVM=1 defconfig

  $ grep STACKPROTECTOR .config

When building on x86_64:

  $ make -sj"$(nproc)" ARCH=x86_64 LLVM=1 defconfig

  $ grep STACKPROTECTOR .config
  CONFIG_CC_HAS_SANE_STACKPROTECTOR=y
  CONFIG_HAVE_STACKPROTECTOR=y
  CONFIG_STACKPROTECTOR=y
  CONFIG_STACKPROTECTOR_STRONG=y

When clang is invoked without a '--target' flag, code is generated for
the default target, which is usually the host (it is configurable via
cmake). As a result, the has-stack-protector scripts will generate code
for the default target but check for x86 specific segment registers,
which cannot succeed if the default target is not x86.

$(CLANG_FLAGS) contains an explicit '--target' flag so pass that
variable along to the has-stack-protector scripts so that the stack
protector can be enabled when cross compiling with clang. The 32-bit
stack protector cannot currently be enabled with clang, as it does not
support '-mstack-protector-guard-symbol', so this results in no
functional change for ARCH=i386 when cross compiling.

Link: https://github.com/chimera-linux/cports/commit/0fb7e506d5f83fdf2104feb22cdac34934561226
Link: https://github.com/llvm/llvm-project/issues/48553
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
---

Fixes: 2a61f4747eea ("stack-protector: test compiler capability in Kconfig and drop AUTO mode")

might be appropriate; I am conflicted on fixes tags for problems that
that arise due to use cases that were not considered at the time of a
change, as it feels wrong to blame the commit for not looking far enough
into the future where it might be common for people to have workstations
running another architecture other than x86_64.

Chimera appears to use a 5.15 kernel so a

Cc: stable@vger.kernel.org

might be nice but some maintainers are picky about that so I leave it up
to you all.

 arch/x86/Kconfig | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index be0b95e51df6..076adde7ead9 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -391,8 +391,8 @@ config PGTABLE_LEVELS
 
 config CC_HAS_SANE_STACKPROTECTOR
 	bool
-	default $(success,$(srctree)/scripts/gcc-x86_64-has-stack-protector.sh $(CC)) if 64BIT
-	default $(success,$(srctree)/scripts/gcc-x86_32-has-stack-protector.sh $(CC))
+	default $(success,$(srctree)/scripts/gcc-x86_64-has-stack-protector.sh $(CC) $(CLANG_FLAGS)) if 64BIT
+	default $(success,$(srctree)/scripts/gcc-x86_32-has-stack-protector.sh $(CC) $(CLANG_FLAGS))
 	help
 	  We have to make sure stack protector is unconditionally disabled if
 	  the compiler produces broken code or if it does not let us control

base-commit: b13baccc3850ca8b8cccbf8ed9912dbaa0fdf7f3
-- 
2.36.1

Re: [PATCH] x86/Kconfig: Fix CONFIG_CC_HAS_SANE_STACKPROTECTOR when cross compiling with clang

[Nathan Chancellor]

Gentle ping for review.

On Fri, Jun 17, 2022 at 11:08:46AM -0700, Nathan Chancellor wrote:
> Chimera Linux notes that CONFIG_CC_HAS_SANE_STACKPROTECTOR cannot be
> enabled when cross compiling an x86_64 kernel with clang, even though it
> does work when natively compiling.
> 
> When building on aarch64:
> 
>   $ make -sj"$(nproc)" ARCH=x86_64 LLVM=1 defconfig
> 
>   $ grep STACKPROTECTOR .config
> 
> When building on x86_64:
> 
>   $ make -sj"$(nproc)" ARCH=x86_64 LLVM=1 defconfig
> 
>   $ grep STACKPROTECTOR .config
>   CONFIG_CC_HAS_SANE_STACKPROTECTOR=y
>   CONFIG_HAVE_STACKPROTECTOR=y
>   CONFIG_STACKPROTECTOR=y
>   CONFIG_STACKPROTECTOR_STRONG=y
> 
> When clang is invoked without a '--target' flag, code is generated for
> the default target, which is usually the host (it is configurable via
> cmake). As a result, the has-stack-protector scripts will generate code
> for the default target but check for x86 specific segment registers,
> which cannot succeed if the default target is not x86.
> 
> $(CLANG_FLAGS) contains an explicit '--target' flag so pass that
> variable along to the has-stack-protector scripts so that the stack
> protector can be enabled when cross compiling with clang. The 32-bit
> stack protector cannot currently be enabled with clang, as it does not
> support '-mstack-protector-guard-symbol', so this results in no
> functional change for ARCH=i386 when cross compiling.
> 
> Link: https://github.com/chimera-linux/cports/commit/0fb7e506d5f83fdf2104feb22cdac34934561226
> Link: https://github.com/llvm/llvm-project/issues/48553
> Signed-off-by: Nathan Chancellor <nathan@kernel.org>
> ---
> 
> Fixes: 2a61f4747eea ("stack-protector: test compiler capability in Kconfig and drop AUTO mode")
> 
> might be appropriate; I am conflicted on fixes tags for problems that
> that arise due to use cases that were not considered at the time of a
> change, as it feels wrong to blame the commit for not looking far enough
> into the future where it might be common for people to have workstations
> running another architecture other than x86_64.
> 
> Chimera appears to use a 5.15 kernel so a
> 
> Cc: stable@vger.kernel.org
> 
> might be nice but some maintainers are picky about that so I leave it up
> to you all.
> 
>  arch/x86/Kconfig | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> index be0b95e51df6..076adde7ead9 100644
> --- a/arch/x86/Kconfig
> +++ b/arch/x86/Kconfig
> @@ -391,8 +391,8 @@ config PGTABLE_LEVELS
>  
>  config CC_HAS_SANE_STACKPROTECTOR
>  	bool
> -	default $(success,$(srctree)/scripts/gcc-x86_64-has-stack-protector.sh $(CC)) if 64BIT
> -	default $(success,$(srctree)/scripts/gcc-x86_32-has-stack-protector.sh $(CC))
> +	default $(success,$(srctree)/scripts/gcc-x86_64-has-stack-protector.sh $(CC) $(CLANG_FLAGS)) if 64BIT
> +	default $(success,$(srctree)/scripts/gcc-x86_32-has-stack-protector.sh $(CC) $(CLANG_FLAGS))
>  	help
>  	  We have to make sure stack protector is unconditionally disabled if
>  	  the compiler produces broken code or if it does not let us control
> 
> base-commit: b13baccc3850ca8b8cccbf8ed9912dbaa0fdf7f3
> -- 
> 2.36.1
> 
> 

Re: [PATCH] x86/Kconfig: Fix CONFIG_CC_HAS_SANE_STACKPROTECTOR when cross compiling with clang

[Dave Hansen]

On 6/17/22 11:08, Nathan Chancellor wrote:
> When clang is invoked without a '--target' flag, code is generated for
> the default target, which is usually the host (it is configurable via
> cmake). As a result, the has-stack-protector scripts will generate code
> for the default target but check for x86 specific segment registers,
> which cannot succeed if the default target is not x86.

I guess the real root cause here is the direct use of '$(CC)' without
any other flags.  Adding '$(CLANG_FLAGS)' seems like a pretty normal
fix, like in scripts/Kconfig.include.

I suspect there's another one of these here:

	arch/x86/um/vdso/Makefile:      cmd_vdso = $(CC) -nostdlib -o $@

but I wouldn't be surprised if UML doesn't work with clang in the first
place.

Re: [PATCH] x86/Kconfig: Fix CONFIG_CC_HAS_SANE_STACKPROTECTOR when cross compiling with clang

[Nathan Chancellor]

On Wed, Jul 06, 2022 at 03:25:51PM -0700, Dave Hansen wrote:
> On 6/17/22 11:08, Nathan Chancellor wrote:
> > When clang is invoked without a '--target' flag, code is generated for
> > the default target, which is usually the host (it is configurable via
> > cmake). As a result, the has-stack-protector scripts will generate code
> > for the default target but check for x86 specific segment registers,
> > which cannot succeed if the default target is not x86.
> 
> I guess the real root cause here is the direct use of '$(CC)' without
> any other flags.  Adding '$(CLANG_FLAGS)' seems like a pretty normal
> fix, like in scripts/Kconfig.include.

Right, also see the following commits for other areas where this was
addressed.

58d746c119df ("efi/libstub: Add $(CLANG_FLAGS) to x86 flags")
d5cbd80e302d ("x86/boot: Add $(CLANG_FLAGS) to compressed KBUILD_CFLAGS")
8abe7fc26ad8 ("x86/build: Propagate $(CLANG_FLAGS) to $(REALMODE_FLAGS)")

> I suspect there's another one of these here:
> 
> 	arch/x86/um/vdso/Makefile:      cmd_vdso = $(CC) -nostdlib -o $@
> 
> but I wouldn't be surprised if UML doesn't work with clang in the first
> place.

We have started testing UML with clang and it does work but I suspect
there is little value to cross compiling a UML kernel, as it has to run
in an x86 userland anyways, rather than through QEMU or other
virtualization solutions. That is not something I plan to do anyways.
If someone does and a fix similar to this one is needed, it can be done
at that time.

Thank you for picking up this change!

Cheers,
Nathan