[PATCH] drm/selftests/mm: reduce per-function stack usage

[PATCH] drm/selftests/mm: reduce per-function stack usage

[Arnd Bergmann]

The check_reserve_boundaries() function has a large array on the stack,
over 500 bytes. It gets inlined into __igt_reserve, which has multiple
other large structures as well but stayed just under the stack size
warning limit of 1024 bytes until one more member got added to struct
drm_mm_node, causing a warning:

drivers/gpu/drm/selftests/test-drm_mm.c:371:12: error:
stack frame size of 1032 bytes in function '__igt_reserve' [-Werror,-Wframe-larger-than=]

As far as I can tell, this is not nice but will not be called from
a context that is already low for the kernel stack, so just annotate
the inner function as noinline_for_stack to ensure that each function
by itself stays under the warning limit.

Fixes: 0cdea4455acd ("drm/mm: optimize rb_hole_addr rbtree search")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 drivers/gpu/drm/selftests/test-drm_mm.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/selftests/test-drm_mm.c b/drivers/gpu/drm/selftests/test-drm_mm.c
index 9aabe82dcd3a..30108c330db8 100644
--- a/drivers/gpu/drm/selftests/test-drm_mm.c
+++ b/drivers/gpu/drm/selftests/test-drm_mm.c
@@ -323,9 +323,8 @@ static bool expect_reserve_fail(struct drm_mm *mm, struct drm_mm_node *node)
 	return false;
 }
 
-static bool check_reserve_boundaries(struct drm_mm *mm,
-				     unsigned int count,
-				     u64 size)
+static noinline_for_stack bool
+check_reserve_boundaries(struct drm_mm *mm, unsigned int count, u64 size)
 {
 	const struct boundary {
 		u64 start, size;
-- 
2.26.2

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH] drm/selftests/mm: reduce per-function stack usage

[Chris Wilson]

Quoting Arnd Bergmann (2020-05-29 21:15:26)
> The check_reserve_boundaries() function has a large array on the stack,
> over 500 bytes. It gets inlined into __igt_reserve, which has multiple
> other large structures as well but stayed just under the stack size
> warning limit of 1024 bytes until one more member got added to struct
> drm_mm_node, causing a warning:
> 
> drivers/gpu/drm/selftests/test-drm_mm.c:371:12: error:
> stack frame size of 1032 bytes in function '__igt_reserve' [-Werror,-Wframe-larger-than=]
> 
> As far as I can tell, this is not nice but will not be called from
> a context that is already low for the kernel stack, so just annotate
> the inner function as noinline_for_stack to ensure that each function
> by itself stays under the warning limit.
> 
> Fixes: 0cdea4455acd ("drm/mm: optimize rb_hole_addr rbtree search")
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> ---
>  drivers/gpu/drm/selftests/test-drm_mm.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/selftests/test-drm_mm.c b/drivers/gpu/drm/selftests/test-drm_mm.c
> index 9aabe82dcd3a..30108c330db8 100644
> --- a/drivers/gpu/drm/selftests/test-drm_mm.c
> +++ b/drivers/gpu/drm/selftests/test-drm_mm.c
> @@ -323,9 +323,8 @@ static bool expect_reserve_fail(struct drm_mm *mm, struct drm_mm_node *node)
>         return false;
>  }
>  
> -static bool check_reserve_boundaries(struct drm_mm *mm,
> -                                    unsigned int count,
> -                                    u64 size)
> +static noinline_for_stack bool
> +check_reserve_boundaries(struct drm_mm *mm, unsigned int count, u64 size)
>  {
>         const struct boundary {

It's this const [] right? Hmm, if we felt adventurous that could be a
small set of multiplication factors (0, -1, 1, count, count+1, ...) and
made static.
-Chris
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH] drm/selftests/mm: reduce per-function stack usage

[Arnd Bergmann]

On Fri, May 29, 2020 at 10:26 PM Chris Wilson <chris@chris-wilson.co.uk> wrote:
> Quoting Arnd Bergmann (2020-05-29 21:15:26)

> >
> > diff --git a/drivers/gpu/drm/selftests/test-drm_mm.c b/drivers/gpu/drm/selftests/test-drm_mm.c
> > index 9aabe82dcd3a..30108c330db8 100644
> > --- a/drivers/gpu/drm/selftests/test-drm_mm.c
> > +++ b/drivers/gpu/drm/selftests/test-drm_mm.c
> > @@ -323,9 +323,8 @@ static bool expect_reserve_fail(struct drm_mm *mm, struct drm_mm_node *node)
> >         return false;
> >  }
> >
> > -static bool check_reserve_boundaries(struct drm_mm *mm,
> > -                                    unsigned int count,
> > -                                    u64 size)
> > +static noinline_for_stack bool
> > +check_reserve_boundaries(struct drm_mm *mm, unsigned int count, u64 size)
> >  {
> >         const struct boundary {
>
> It's this const [] right? Hmm, if we felt adventurous that could be a
> small set of multiplication factors (0, -1, 1, count, count+1, ...) and
> made static.

That was my first thought, but I couldn't figure out whether 'count'
could be replaced by any compile-time constant.

      Arnd
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH] drm/selftests/mm: reduce per-function stack usage

[Chris Wilson]

Quoting Arnd Bergmann (2020-05-29 21:43:47)
> On Fri, May 29, 2020 at 10:26 PM Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > Quoting Arnd Bergmann (2020-05-29 21:15:26)
> 
> > >
> > > diff --git a/drivers/gpu/drm/selftests/test-drm_mm.c b/drivers/gpu/drm/selftests/test-drm_mm.c
> > > index 9aabe82dcd3a..30108c330db8 100644
> > > --- a/drivers/gpu/drm/selftests/test-drm_mm.c
> > > +++ b/drivers/gpu/drm/selftests/test-drm_mm.c
> > > @@ -323,9 +323,8 @@ static bool expect_reserve_fail(struct drm_mm *mm, struct drm_mm_node *node)
> > >         return false;
> > >  }
> > >
> > > -static bool check_reserve_boundaries(struct drm_mm *mm,
> > > -                                    unsigned int count,
> > > -                                    u64 size)
> > > +static noinline_for_stack bool
> > > +check_reserve_boundaries(struct drm_mm *mm, unsigned int count, u64 size)
> > >  {
> > >         const struct boundary {
> >
> > It's this const [] right? Hmm, if we felt adventurous that could be a
> > small set of multiplication factors (0, -1, 1, count, count+1, ...) and
> > made static.
> 
> That was my first thought, but I couldn't figure out whether 'count'
> could be replaced by any compile-time constant.

I just realised I sent a sketch of a patch to the wrong place. If we
replace struct boundary with { int start; int size; const char *name; }
that should reduce it from 408 to 272. (Where start, size are the
multiples.)

Probably not worth the hassle, the saving is too small overall leaving
it uncomfortably close to a future warning.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
-Chris
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel