[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20241018224502.GC2635543@thelio-3990X>
Date: Fri, 18 Oct 2024 15:45:02 -0700
From: Nathan Chancellor <nathan@...nel.org>
To: Arnd Bergmann <arnd@...nel.org>
Cc: Tiffany Lin <tiffany.lin@...iatek.com>,
Andrew-CT Chen <andrew-ct.chen@...iatek.com>,
Yunfei Dong <yunfei.dong@...iatek.com>,
Mauro Carvalho Chehab <mchehab@...nel.org>,
Matthias Brugger <matthias.bgg@...il.com>,
AngeloGioacchino Del Regno <angelogioacchino.delregno@...labora.com>,
Arnd Bergmann <arnd@...db.de>,
Nick Desaulniers <ndesaulniers@...gle.com>,
Bill Wendling <morbo@...gle.com>,
Justin Stitt <justinstitt@...gle.com>,
Hans Verkuil <hverkuil@...all.nl>,
Sebastian Fricke <sebastian.fricke@...labora.com>,
Benjamin Gaignard <benjamin.gaignard@...labora.com>,
linux-media@...r.kernel.org, linux-kernel@...r.kernel.org,
linux-arm-kernel@...ts.infradead.org,
linux-mediatek@...ts.infradead.org, llvm@...ts.linux.dev
Subject: Re: [PATCH] media: mediatek: vcodec: mark
vdec_vp9_slice_map_counts_eob_coef noinline
On Fri, Oct 18, 2024 at 03:14:42PM +0000, Arnd Bergmann wrote:
> From: Arnd Bergmann <arnd@...db.de>
>
> With KASAN enabled, clang fails to optimize the inline version of
> vdec_vp9_slice_map_counts_eob_coef() properly, leading to kilobytes
> of temporary values spilled to the stack:
>
> drivers/media/platform/mediatek/vcodec/decoder/vdec/vdec_vp9_req_lat_if.c:1526:12: error: stack frame size (2160) exceeds limit (2048) in 'vdec_vp9_slice_update_prob' [-Werror,-Wframe-larger-than]
>
> This seems to affect all versions of clang including the latest (clang-20),
> but the degree of stack overhead is different per release.
>
> Marking the function as noinline_for_stack is harmless here and avoids
> the problem completely.
>
> Signed-off-by: Arnd Bergmann <arnd@...db.de>
> ---
> I have not come to a conclusion on how exactly clang fails to do this
> right, but can provide the .config and/or preprocessed source files
> and command line if we think this should be fixed in clang.
I think this might be related to the issue I reported to upstream LLVM,
as a regression within the past couple of weeks:
https://github.com/llvm/llvm-project/issues/111903
If this is a reasonable workaround, it might be worth doing but I will
probably wait until after the LLVM Developers Meeting next week to ping
the thread to have a better chance of visibility. If we want to work
around this in the kernel, we should Cc stable, as this warning is
present there too.
> ---
> .../mediatek/vcodec/decoder/vdec/vdec_vp9_req_lat_if.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/media/platform/mediatek/vcodec/decoder/vdec/vdec_vp9_req_lat_if.c b/drivers/media/platform/mediatek/vcodec/decoder/vdec/vdec_vp9_req_lat_if.c
> index eea709d93820..47c302745c1d 100644
> --- a/drivers/media/platform/mediatek/vcodec/decoder/vdec/vdec_vp9_req_lat_if.c
> +++ b/drivers/media/platform/mediatek/vcodec/decoder/vdec/vdec_vp9_req_lat_if.c
> @@ -1188,7 +1188,8 @@ static int vdec_vp9_slice_setup_lat(struct vdec_vp9_slice_instance *instance,
> return ret;
> }
>
> -static
> +/* clang stack usage explodes if this is inlined */
> +static noinline_for_stack
> void vdec_vp9_slice_map_counts_eob_coef(unsigned int i, unsigned int j, unsigned int k,
> struct vdec_vp9_slice_frame_counts *counts,
> struct v4l2_vp9_frame_symbol_counts *counts_helper)
> --
> 2.39.5
>
Powered by blists - more mailing lists