Re: [PATCH v5 1/4] kasan: Emit different calls for instrumentable memintrinsics

[Catalin Marinas]

Hi Marco,

On Fri, Feb 24, 2023 at 09:59:39AM +0100, Marco Elver wrote:
> Clang 15 provides an option to prefix memcpy/memset/memmove calls with
> __asan_/__hwasan_ in instrumented functions: https://reviews.llvm.org/D122724
> 
> GCC will add support in future:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108777
> 
> Use it to regain KASAN instrumentation of memcpy/memset/memmove on
> architectures that require noinstr to be really free from instrumented
> mem*() functions (all GENERIC_ENTRY architectures).
> 
> Fixes: 69d4c0d32186 ("entry, kasan, x86: Disallow overriding mem*() functions")
> Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
> Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> Reviewed-by: Andrey Konovalov <andreyknvl@xxxxxxxxx>
> Tested-by: Linux Kernel Functional Testing <lkft@xxxxxxxxxx>
> Tested-by: Naresh Kamboju <naresh.kamboju@xxxxxxxxxx>
[...]
> diff --git a/scripts/Makefile.kasan b/scripts/Makefile.kasan
> index b9e94c5e7097..fa9f836f8039 100644
> --- a/scripts/Makefile.kasan
> +++ b/scripts/Makefile.kasan
> @@ -38,6 +38,11 @@ endif
>  
>  CFLAGS_KASAN += $(call cc-param,asan-stack=$(stack_enable))
>  
> +# Instrument memcpy/memset/memmove calls by using instrumented __asan_mem*()
> +# instead. With compilers that don't support this option, compiler-inserted
> +# memintrinsics won't be checked by KASAN on GENERIC_ENTRY architectures.
> +CFLAGS_KASAN += $(call cc-param,asan-kernel-mem-intrinsic-prefix=1)
> +
>  endif # CONFIG_KASAN_GENERIC
>  
>  ifdef CONFIG_KASAN_SW_TAGS
> @@ -54,6 +59,9 @@ CFLAGS_KASAN := -fsanitize=kernel-hwaddress \
>  		$(call cc-param,hwasan-inline-all-checks=0) \
>  		$(instrumentation_flags)
>  
> +# Instrument memcpy/memset/memmove calls by using instrumented __hwasan_mem*().
> +CFLAGS_KASAN += $(call cc-param,hwasan-kernel-mem-intrinsic-prefix=1)

This patch breaks the arm64 kernel builds with KASAN_SW_TAGS enabled and
clang prior to version 15. Those prior clang versions don't like the
'-mllvm -hwasan-kernel-mem-intrinsic-prefix=1' option, end up printing
the help text instead of generating the object.

Do we need some combination of cc-option and cc-param? Or at least
disable this instrumentation if earlier clang versions are used.

It's already in mainline as commit
51287dcb00cc715c27bf6a6b4dbd431621c5b65a.

Thanks.

-- 
Catalin