2023-02-14 10:30:46

by Arnd Bergmann

[permalink] [raw]
Subject: [PATCH] [RFC] maple_tree: reduce stack usage with gcc-9 and earlier

From: Arnd Bergmann <[email protected]>

gcc-10 changed the way inlining works to be less aggressive, but
older versions run into an oversized stack frame warning whenever
CONFIG_KASAN_STACK is enabled, as that forces variables from
inlined callees to be non-overlapping:

lib/maple_tree.c: In function 'mas_wr_bnode':
lib/maple_tree.c:4320:1: error: the frame size of 1424 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]

Change the annotations on mas_store_b_node() and mas_commit_b_node()
to explicitly forbid inlining in this configuration, which is
the same behavior that newer versions already have.

Cc: Andrey Ryabinin <[email protected]>
Cc: Alexander Potapenko <[email protected]>
Cc: Andrey Konovalov <[email protected]>
Cc: Dmitry Vyukov <[email protected]>
Cc: Vincenzo Frascino <[email protected]>
Cc: [email protected]
Signed-off-by: Arnd Bergmann <[email protected]>
---
lib/maple_tree.c | 11 +++++++++--
1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/lib/maple_tree.c b/lib/maple_tree.c
index 5e9703189259..646297cae5d1 100644
--- a/lib/maple_tree.c
+++ b/lib/maple_tree.c
@@ -146,6 +146,13 @@ struct maple_subtree_state {
struct maple_big_node *bn;
};

+#ifdef CONFIG_KASAN_STACK
+/* Prevent mas_wr_bnode() from exceeding the stack frame limit */
+#define noinline_for_kasan noinline_for_stack
+#else
+#define noinline_for_kasan inline
+#endif
+
/* Functions */
static inline struct maple_node *mt_alloc_one(gfp_t gfp)
{
@@ -2107,7 +2114,7 @@ static inline void mas_bulk_rebalance(struct ma_state *mas, unsigned char end,
*
* Return: The actual end of the data stored in @b_node
*/
-static inline void mas_store_b_node(struct ma_wr_state *wr_mas,
+static noinline_for_kasan void mas_store_b_node(struct ma_wr_state *wr_mas,
struct maple_big_node *b_node, unsigned char offset_end)
{
unsigned char slot;
@@ -3579,7 +3586,7 @@ static inline bool mas_reuse_node(struct ma_wr_state *wr_mas,
* @b_node: The maple big node
* @end: The end of the data.
*/
-static inline int mas_commit_b_node(struct ma_wr_state *wr_mas,
+static noinline_for_kasan int mas_commit_b_node(struct ma_wr_state *wr_mas,
struct maple_big_node *b_node, unsigned char end)
{
struct maple_node *node;
--
2.39.1



2023-02-14 12:08:42

by David Hildenbrand

[permalink] [raw]
Subject: Re: [PATCH] [RFC] maple_tree: reduce stack usage with gcc-9 and earlier

On 14.02.23 11:30, Arnd Bergmann wrote:
> From: Arnd Bergmann <[email protected]>
>
> gcc-10 changed the way inlining works to be less aggressive, but
> older versions run into an oversized stack frame warning whenever
> CONFIG_KASAN_STACK is enabled, as that forces variables from
> inlined callees to be non-overlapping:
>
> lib/maple_tree.c: In function 'mas_wr_bnode':
> lib/maple_tree.c:4320:1: error: the frame size of 1424 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
>
> Change the annotations on mas_store_b_node() and mas_commit_b_node()
> to explicitly forbid inlining in this configuration, which is
> the same behavior that newer versions already have.
>
> Cc: Andrey Ryabinin <[email protected]>
> Cc: Alexander Potapenko <[email protected]>
> Cc: Andrey Konovalov <[email protected]>
> Cc: Dmitry Vyukov <[email protected]>
> Cc: Vincenzo Frascino <[email protected]>
> Cc: [email protected]
> Signed-off-by: Arnd Bergmann <[email protected]>
> ---

Reviewed-by: David Hildenbrand <[email protected]>

--
Thanks,

David / dhildenb


2023-02-14 19:40:22

by Andrew Morton

[permalink] [raw]
Subject: Re: [PATCH] [RFC] maple_tree: reduce stack usage with gcc-9 and earlier

On Tue, 14 Feb 2023 11:30:24 +0100 Arnd Bergmann <[email protected]> wrote:

> From: Arnd Bergmann <[email protected]>
>
> gcc-10 changed the way inlining works to be less aggressive, but
> older versions run into an oversized stack frame warning whenever
> CONFIG_KASAN_STACK is enabled, as that forces variables from
> inlined callees to be non-overlapping:
>
> lib/maple_tree.c: In function 'mas_wr_bnode':
> lib/maple_tree.c:4320:1: error: the frame size of 1424 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
>
> Change the annotations on mas_store_b_node() and mas_commit_b_node()
> to explicitly forbid inlining in this configuration, which is
> the same behavior that newer versions already have.
>
> ...
>
> --- a/lib/maple_tree.c
> +++ b/lib/maple_tree.c
> @@ -146,6 +146,13 @@ struct maple_subtree_state {
> struct maple_big_node *bn;
> };
>
> +#ifdef CONFIG_KASAN_STACK
> +/* Prevent mas_wr_bnode() from exceeding the stack frame limit */
> +#define noinline_for_kasan noinline_for_stack
> +#else
> +#define noinline_for_kasan inline
> +#endif

Should noinline_for_kasan be defined in kasan.h? maple_tree.c is
unlikely to be the only place in the kernel which could use this
treatment?

I suppose we can do that when the need arises.

2023-02-15 17:22:58

by Liam R. Howlett

[permalink] [raw]
Subject: Re: [PATCH] [RFC] maple_tree: reduce stack usage with gcc-9 and earlier

* Arnd Bergmann <[email protected]> [230214 05:30]:
> From: Arnd Bergmann <[email protected]>
>
> gcc-10 changed the way inlining works to be less aggressive, but
> older versions run into an oversized stack frame warning whenever
> CONFIG_KASAN_STACK is enabled, as that forces variables from
> inlined callees to be non-overlapping:
>
> lib/maple_tree.c: In function 'mas_wr_bnode':
> lib/maple_tree.c:4320:1: error: the frame size of 1424 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
>
> Change the annotations on mas_store_b_node() and mas_commit_b_node()
> to explicitly forbid inlining in this configuration, which is
> the same behavior that newer versions already have.
>
> Cc: Andrey Ryabinin <[email protected]>
> Cc: Alexander Potapenko <[email protected]>
> Cc: Andrey Konovalov <[email protected]>
> Cc: Dmitry Vyukov <[email protected]>
> Cc: Vincenzo Frascino <[email protected]>
> Cc: [email protected]
> Signed-off-by: Arnd Bergmann <[email protected]>
> ---
> lib/maple_tree.c | 11 +++++++++--
> 1 file changed, 9 insertions(+), 2 deletions(-)
>
> diff --git a/lib/maple_tree.c b/lib/maple_tree.c
> index 5e9703189259..646297cae5d1 100644
> --- a/lib/maple_tree.c
> +++ b/lib/maple_tree.c
> @@ -146,6 +146,13 @@ struct maple_subtree_state {
> struct maple_big_node *bn;
> };
>
> +#ifdef CONFIG_KASAN_STACK
> +/* Prevent mas_wr_bnode() from exceeding the stack frame limit */

nit: Should there be more info in here? You did add it to two functions
and it's a problem for frame-larger-than=1024

Reviewed-by: Liam R. Howlett <[email protected]>


> +#define noinline_for_kasan noinline_for_stack
> +#else
> +#define noinline_for_kasan inline
> +#endif
> +
> /* Functions */
> static inline struct maple_node *mt_alloc_one(gfp_t gfp)
> {
> @@ -2107,7 +2114,7 @@ static inline void mas_bulk_rebalance(struct ma_state *mas, unsigned char end,
> *
> * Return: The actual end of the data stored in @b_node
> */
> -static inline void mas_store_b_node(struct ma_wr_state *wr_mas,
> +static noinline_for_kasan void mas_store_b_node(struct ma_wr_state *wr_mas,
> struct maple_big_node *b_node, unsigned char offset_end)
> {
> unsigned char slot;
> @@ -3579,7 +3586,7 @@ static inline bool mas_reuse_node(struct ma_wr_state *wr_mas,
> * @b_node: The maple big node
> * @end: The end of the data.
> */
> -static inline int mas_commit_b_node(struct ma_wr_state *wr_mas,
> +static noinline_for_kasan int mas_commit_b_node(struct ma_wr_state *wr_mas,
> struct maple_big_node *b_node, unsigned char end)
> {
> struct maple_node *node;
> --
> 2.39.1
>

2023-02-17 10:01:58

by Andrey Konovalov

[permalink] [raw]
Subject: Re: [PATCH] [RFC] maple_tree: reduce stack usage with gcc-9 and earlier

On Tue, Feb 14, 2023 at 8:40 PM Andrew Morton <[email protected]> wrote:
>
> On Tue, 14 Feb 2023 11:30:24 +0100 Arnd Bergmann <[email protected]> wrote:
>
> > From: Arnd Bergmann <[email protected]>
> >
> > gcc-10 changed the way inlining works to be less aggressive, but
> > older versions run into an oversized stack frame warning whenever
> > CONFIG_KASAN_STACK is enabled, as that forces variables from
> > inlined callees to be non-overlapping:
> >
> > lib/maple_tree.c: In function 'mas_wr_bnode':
> > lib/maple_tree.c:4320:1: error: the frame size of 1424 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> >
> > Change the annotations on mas_store_b_node() and mas_commit_b_node()
> > to explicitly forbid inlining in this configuration, which is
> > the same behavior that newer versions already have.
> >
> > ...
> >
> > --- a/lib/maple_tree.c
> > +++ b/lib/maple_tree.c
> > @@ -146,6 +146,13 @@ struct maple_subtree_state {
> > struct maple_big_node *bn;
> > };
> >
> > +#ifdef CONFIG_KASAN_STACK
> > +/* Prevent mas_wr_bnode() from exceeding the stack frame limit */
> > +#define noinline_for_kasan noinline_for_stack
> > +#else
> > +#define noinline_for_kasan inline
> > +#endif
>
> Should noinline_for_kasan be defined in kasan.h? maple_tree.c is
> unlikely to be the only place in the kernel which could use this
> treatment?

We could also define it in include/linux/compiler_types.h along with
other KASAN attributes.