Re: [PATCH v3 4/9] x86/build: Warn on orphan section placement

[Marco Elver]

On Sat, 27 Jun 2020 at 17:44, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Thu, Jun 25, 2020 at 02:36:27AM +0800, kernel test robot wrote:
> > I love your patch! Perhaps something to improve:
> > [...]
> > config: x86_64-randconfig-a012-20200624 (attached as .config)
>
> CONFIG_KCSAN=y
>
> > compiler: clang version 11.0.0 (https://github.com/llvm/llvm-project 1d4c87335d5236ea1f35937e1014980ba961ae34)
> > [...]
> > All warnings (new ones prefixed by >>):
> >
> >    ld.lld: warning: drivers/built-in.a(mfd/mt6397-irq.o):(.init_array.0) is being placed in '.init_array.0'
>
> As far as I can tell, this is a Clang bug. But I don't know the
> internals here, so I've opened:
> https://bugs.llvm.org/show_bug.cgi?id=46478
>
> and created a work-around patch for the kernel:

Thanks, minor comments below.

With KCSAN this is:

Tested-by: Marco Elver <elver@xxxxxxxxxx>


> commit 915f2c343e59a14f00c68f4d7afcfdc621de0674
> Author: Kees Cook <keescook@xxxxxxxxxxxx>
> Date:   Sat Jun 27 08:07:54 2020 -0700
>
>     vmlinux.lds.h: Avoid KCSAN's unwanted sections

Since you found that it's also KASAN, this probably wants updating.

>     KCSAN (-fsanitize=thread) produces unwanted[1] .eh_frame and .init_array.*
>     sections. Add them to DISCARDS, except with CONFIG_CONSTRUCTORS, which
>     wants to keep .init_array.* sections.
>
>     [1] https://bugs.llvm.org/show_bug.cgi?id=46478
>
>     Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
>
> diff --git a/arch/x86/Makefile b/arch/x86/Makefile
> index f8a5b2333729..41c8c73de6c4 100644
> --- a/arch/x86/Makefile
> +++ b/arch/x86/Makefile
> @@ -195,7 +195,9 @@ endif
>  # Workaround for a gcc prelease that unfortunately was shipped in a suse release
>  KBUILD_CFLAGS += -Wno-sign-compare
>  #
> -KBUILD_CFLAGS += -fno-asynchronous-unwind-tables
> +KBUILD_AFLAGS += -fno-asynchronous-unwind-tables -fno-unwind-tables
> +KBUILD_CFLAGS += -fno-asynchronous-unwind-tables -fno-unwind-tables
> +KBUILD_LDFLAGS += $(call ld-option,--no-ld-generated-unwind-info)

Why are they needed? They are not mentioned in the commit message.

>  # Avoid indirect branches in kernel to deal with Spectre
>  ifdef CONFIG_RETPOLINE
> diff --git a/include/asm-generic/vmlinux.lds.h b/include/asm-generic/vmlinux.lds.h
> index b1dca0762fc5..a44ee16abc78 100644
> --- a/include/asm-generic/vmlinux.lds.h
> +++ b/include/asm-generic/vmlinux.lds.h
> @@ -934,10 +934,28 @@
>         EXIT_DATA
>  #endif
>
> +/*
> + * Clang's -fsanitize=thread produces unwanted sections (.eh_frame
> + * and .init_array.*), but CONFIG_CONSTRUCTORS wants to keep any
> + * .init_array.* sections.
> + * https://bugs.llvm.org/show_bug.cgi?id=46478
> + */
> +#if defined(CONFIG_KCSAN) && !defined(CONFIG_CONSTRUCTORS)

CONFIG_KASAN as well?

> +#define KCSAN_DISCARDS                                                 \
> +       *(.init_array) *(.init_array.*)                                 \
> +       *(.eh_frame)
> +#elif defined(CONFIG_KCSAN) && defined(CONFIG_CONSTRUCTORS)
> +#define KCSAN_DISCARDS                                                 \
> +       *(.eh_frame)
> +#else
> +#define KCSAN_DISCARDS
> +#endif
> +
>  #define DISCARDS                                                       \
>         /DISCARD/ : {                                                   \
>         EXIT_DISCARDS                                                   \
>         EXIT_CALL                                                       \
> +       KCSAN_DISCARDS                                                  \

Maybe just 'SANITIZER_DISCARDS'?

>         *(.discard)                                                     \
>         *(.discard.*)                                                   \
>         *(.modinfo)                                                     \
>
> --
> Kees Cook