[PATCH] media: vivid: work around high stack usage with clang

[Arnd Bergmann]

Building a KASAN-enabled kernel with clang ends up in a case where too
much is inlined into vivid_thread_vid_cap() and the stack usage grows
a lot, possibly when the register allocation fails to produce efficient
code and spills a lot of temporaries to the stack. This uses more
than twice the amount of stack than the sum of the individual functions
when they are not inlined:

drivers/media/platform/vivid/vivid-kthread-cap.c:766:12: error: stack frame size of 2208 bytes in function 'vivid_thread_vid_cap' [-Werror,-Wframe-larger-than=]

Marking two of the key functions in here as 'noinline_for_stack' avoids
the pathological case in clang without any apparent downside for gcc.

Signed-off-by: Arnd Bergmann <ar...@arndb.de>
---
Not sure how much effort we want to put into fixing clang to not
get into this case. I could open an llvm bug report if something
thinks this has a chance of getting fixed there.
---
drivers/media/platform/vivid/vivid-kthread-cap.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/media/platform/vivid/vivid-kthread-cap.c b/drivers/media/platform/vivid/vivid-kthread-cap.c
index 6cf495a7d5cc..4f94897e6303 100644
--- a/drivers/media/platform/vivid/vivid-kthread-cap.c
+++ b/drivers/media/platform/vivid/vivid-kthread-cap.c
@@ -232,8 +232,8 @@ static void *plane_vaddr(struct tpg_data *tpg, struct vivid_buffer *buf,
return vbuf;
}

-static int vivid_copy_buffer(struct vivid_dev *dev, unsigned p, u8 *vcapbuf,
- struct vivid_buffer *vid_cap_buf)
+static noinline_for_stack int vivid_copy_buffer(struct vivid_dev *dev, unsigned p,
+ u8 *vcapbuf, struct vivid_buffer *vid_cap_buf)
{
bool blank = dev->must_blank[vid_cap_buf->vb.vb2_buf.index];
struct tpg_data *tpg = &dev->tpg;
@@ -670,7 +670,8 @@ static void vivid_cap_update_frame_period(struct vivid_dev *dev)
dev->cap_frame_period = f_period;
}

-static void vivid_thread_vid_cap_tick(struct vivid_dev *dev, int dropped_bufs)
+static noinline_for_stack void vivid_thread_vid_cap_tick(struct vivid_dev *dev,
+ int dropped_bufs)
{
struct vivid_buffer *vid_cap_buf = NULL;
struct vivid_buffer *vbi_cap_buf = NULL;
--
2.20.0

[Nick Desaulniers]

On Thu, Jul 18, 2019 at 7:16 AM Arnd Bergmann <ar...@arndb.de> wrote:
>
> Building a KASAN-enabled kernel with clang ends up in a case where too
> much is inlined into vivid_thread_vid_cap() and the stack usage grows
> a lot, possibly when the register allocation fails to produce efficient
> code and spills a lot of temporaries to the stack. This uses more
> than twice the amount of stack than the sum of the individual functions
> when they are not inlined:
>
> drivers/media/platform/vivid/vivid-kthread-cap.c:766:12: error: stack frame size of 2208 bytes in function 'vivid_thread_vid_cap' [-Werror,-Wframe-larger-than=]
>
> Marking two of the key functions in here as 'noinline_for_stack' avoids
> the pathological case in clang without any apparent downside for gcc.

Thanks for the patch, should be one less warning for CI.
Acked-by: Nick Desaulniers <ndesau...@google.com>

> --
> You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to clang-built-li...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/20190718141652.3323402-1-arnd%40arndb.de.



--
Thanks,
~Nick Desaulniers