[PATCH] [RFC] ubsan: disallow bounds checking with gcov on broken gcc

[Kees Cook]

On Thu, Jun 01, 2023 at 05:18:11PM +0200, Arnd Bergmann wrote:
> From: Arnd Bergmann <arnd at arndb.de>
> 
> Combining UBSAN and GCOV in randconfig builds results in a number of
> stack frame size warnings, such as:
> 
> crypto/twofish_common.c:683:1: error: the frame size of 2040 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c:1589:1: error: the frame size of 1696 bytes is larger than 1400 bytes [-Werror=frame-larger-than=]
> drivers/media/platform/verisilicon/hantro_g2_vp9_dec.c:754:1: error: the frame size of 1260 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> drivers/staging/media/ipu3/ipu3-css-params.c:1206:1: error: the frame size of 1080 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> drivers/staging/media/rkvdec/rkvdec-vp9.c:1042:1: error: the frame size of 2176 bytes is larger than 2048 bytes [-Werror=frame-larger-than=]
> drivers/staging/media/rkvdec/rkvdec-vp9.c:995:1: error: the frame size of 1656 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> 
> I managed to track this down to the -fsanitize=bounds option clashing
> with the -fprofile-arcs option, which leads a lot of spilled temporary
> variables in generated instrumentation code.
> 
> Hopefully this can be addressed in future gcc releases the same way
> that clang handles the combination, but for existing compiler releases,
> it seems best to disable one of the two flags. This can be done either
> globally by just not passing both at the same time, or locally using
> the no_sanitize or no_instrument_function attributes in the affected
> functions.
> 
> Try the simplest approach here, and turn off -fsanitize=bounds on
> gcc when GCOV is enabled, leaving the rest of UBSAN working. Doing
> this globally also helps avoid inefficient code from the same
> problem that did not push the build over the warning limit.
> 
> Reported-by: Dan Carpenter <dan.carpenter at linaro.org>
> Link: https://lore.kernel.org/stable/6b1a0ee6-c78b-4873-bfd5-89798fce9899@kili.mountain/
> Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110074
> Link: https://godbolt.org/z/zvf7YqK5K
> Signed-off-by: Arnd Bergmann <arnd at arndb.de>

I think more production systems will have CONFIG_UBSAN_BOUNDS enabled
(e.g. Ubuntu has had it enabled for more than a year now) than GCOV,
so I'd prefer we maintain all*config coverage for the more commonly
used config.

> ---
>  lib/Kconfig.ubsan | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/lib/Kconfig.ubsan b/lib/Kconfig.ubsan
> index f7cbbad2bb2f4..8f71ff8f27576 100644
> --- a/lib/Kconfig.ubsan
> +++ b/lib/Kconfig.ubsan
> @@ -29,6 +29,8 @@ config UBSAN_TRAP
>  
>  config CC_HAS_UBSAN_BOUNDS_STRICT
>  	def_bool $(cc-option,-fsanitize=bounds-strict)
> +	# work around https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110074
> +	depends on GCC_VERSION > 140000 || !GCOV_PROFILE_ALL
>  	help
>  	  The -fsanitize=bounds-strict option is only available on GCC,
>  	  but uses the more strict handling of arrays that includes knowledge

Alternatively, how about falling back to -fsanitize=bounds instead, as
that (which has less coverage) wasn't triggering the stack frame
warnings?

i.e. fall back through these:
	-fsanitize=array-bounds (Clang)
	-fsanitize=bounds-strict (!GCOV || bug fixed in GCC)
	-fsanitize=bounds

-- 
Kees Cook