kasan: fix redzone overwritten issue under SLUB debug
Subsystems (4)
v3 cover kasan: fix redzone overwritten issue under SLUB debug
From: Yee Lee <[email protected]>
Issue: In SLUB debug, hwtag kasan_unpoison() would overwrite the redzone
in those objects with unaligned size.
This patch Adds memzero_explict() to separate the initialization for such condition. The new code path is executed about 1.1% during nromal booting process. The penalty is acceptable since it only works in debug mode.
=============
Exp: QEMUv5.2(+mte)/SLUB_debug mode
code path exec : 941/80854 (1.1%)
---
Changed since v2:
- Use if-statement to refine the condition block instead of preprocessor #if
- Add comment block about this new code path in source code.
---
Changed since v1:
- Apply IS_ENABLED to wrap codes under SLUB debug mode.
- Replace memset() by memzero_explict().
---
Yee Lee (1):
kasan: Add memzero init for unaligned size under SLUB debug
mm/kasan/kasan.h | 10 ++++++++++ 1 file changed, 10 insertions(+)
Signature · 5 lines
2.18.0 _______________________________________________ Linux-mediatek mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-mediatek
PATCH v3 1/1 kasan: Add memzero init for unaligned size under SLUB debug
From: Yee Lee <[email protected]>
Issue: when SLUB debug is on, hwtag kasan_unpoison() would overwrite the redzone of object with unaligned size.
An additional memzero_explicit() path is added to replacing init by hwtag instruction for those unaligned size at SLUB debug mode.
The penalty is acceptable since they are only enabled in debug mode, not production builds. A block of comment is added for explanation.
--- mm/kasan/kasan.h | 10 ++++++++++ 1 file changed, 10 insertions(+)
mm/kasan/kasan.h +10 -0
@@ -387,6 +387,16 @@ static inline void kasan_unpoison(const void *addr, size_t size, bool init)if (WARN_ON ((unsigned long )addr &KASAN_GRANULE_MASK ))return ; +/* +* Explicitly initialize the memory with the precise object size +* to avoid overwriting the SLAB redzone. This disables initialization +* in the arch code and may thus lead to performance penalty. +* The penalty is accepted since SLAB redzones aren't enabled in production builds. +*/ +if (IS_ENABLED (CONFIG_SLUB_DEBUG ) && init && ((unsigned long )size &KASAN_GRANULE_MASK )) { + init = false; +memzero_explicit ((void *)addr, size); + } size =round_up (size,KASAN_GRANULE_SIZE );hw_set_mem_tag_range ((void *)addr, size, tag, init);
Signature · 1 line
2.18.0
Re: [PATCH v3 1/1] kasan: Add memzero init for unaligned size under SLUB debug
Quoted 11 lines · On Wed, Jun 30, 2021 at 09:49PM +0800, [email protected] wrote:
From: Yee Lee <[email protected]>
Issue: when SLUB debug is on, hwtag
kasan_unpoison()would overwrite the redzone of object with unaligned size.An additional
memzero_explicit()path is added to replacing init by hwtag instruction for those unaligned size at SLUB debug mode.The penalty is acceptable since they are only enabled in debug mode, not production builds. A block of comment is added for explanation.
Quoted 7 lines
Signed-off-by Yee Lee <[email protected]>Suggested-by Marco Elver <[email protected]>Suggested-by Andrey Konovalov <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Alexander Potapenko <[email protected]>Cc Dmitry Vyukov <[email protected]>Cc Andrew Morton <[email protected]>
In future, please add changes to each version after an additional '---'. Example:
---
v2:
* Use IS_ENABLED(CONFIG_SLUB_DEBUG) in if-statement.
Quoted 9 lines
--- mm/kasan/kasan.h | 10 ++++++++++ 1 file changed, 10 insertions(+)mm/kasan/kasan.h +0
-0
mm/kasan/kasan.h static inline void kasan_unpoison(const void *addr, size_t size, bool init) · line 395
Quoted 9 lines
mm/kasan/kasan.h +6
-0@@ -387,6 +387,16 @@ static inline void kasan_unpoison(const void *addr, size_t size, bool init)if (WARN_ON ((unsigned long )addr &KASAN_GRANULE_MASK ))return ; +/* +* Explicitly initialize the memory with the precise object size +* to avoid overwriting the SLAB redzone. This disables initialization +* in the arch code and may thus lead to performance penalty. +* The penalty is accepted since SLAB redzones aren't enabled in production builds. +*/
Can we please format the comment properly:
mm/kasan/kasan.h +4 -4
@@ -388,10 +388,10 @@ static inline void kasan_unpoison(const void *addr, size_t size, bool init)if (WARN_ON ((unsigned long )addr &KASAN_GRANULE_MASK ))return ;/* -* Explicitly initialize the memory with the precise object size -* to avoid overwriting the SLAB redzone. This disables initialization -* in the arch code and may thus lead to performance penalty. -* The penalty is accepted since SLAB redzones aren't enabled in production builds. +* Explicitly initialize the memory with the precise object size to +* avoid overwriting the SLAB redzone. This disables initialization in +* the arch code and may thus lead to performance penalty. The penalty +* is accepted since SLAB redzones aren't enabled in production builds. */ if (IS_ENABLED (CONFIG_SLUB_DEBUG ) && init && ((unsigned long )size &KASAN_GRANULE_MASK )) { init = false;
mm/kasan/kasan.h static inline void kasan_unpoison(const void *addr, size_t size, bool init) · line 402
Quoted 7 lines
mm/kasan/kasan.h +4
-0+if (IS_ENABLED (CONFIG_SLUB_DEBUG ) && init && ((unsigned long )size &KASAN_GRANULE_MASK )) { + init = false; +memzero_explicit ((void *)addr, size); + } size =round_up (size,KASAN_GRANULE_SIZE );hw_set_mem_tag_range ((void *)addr, size, tag, init);
I think this solution might be fine for now, as I don't see an easy way to do this without some major refactor to use kmem_cache_debug_flags().
However, I think there's an intermediate solution where we only check the static-key 'slub_debug_enabled' though. Because I've checked, and various major distros _do_ enabled CONFIG_SLUB_DEBUG. But the static branch just makes sure there's no performance overhead.
Checking the static branch requires including mm/slab.h into mm/kasan/kasan.h, which we currently don't do and perhaps wanted to avoid. Although I don't see a reason there, because there's no circular dependency even if we did.
Andrey, any opinion?
In case you guys think checking static key is the better solution, I think the below would work together with the pre-requisite patch at the end:
mm/kasan/kasan.h +4 -1
@@ -6,6 +6,8 @@#include <linux/kfence.h> #include <linux/stackdepot.h> +#include "../slab.h" +#ifdef CONFIG_KASAN_HW_TAGS #include <linux/static_key.h>
@@ -393,7 +395,8 @@ static inline void kasan_unpoison(const void *addr, size_t size, bool init)* the arch code and may thus lead to performance penalty. The penalty * is accepted since SLAB redzones aren't enabled in production builds. */ -if (IS_ENABLED (CONFIG_SLUB_DEBUG ) && init && ((unsigned long )size &KASAN_GRANULE_MASK )) { +if (slub_debug_enabled_unlikely () && + init && ((unsigned long )size &KASAN_GRANULE_MASK )) { init = false;memzero_explicit ((void *)addr, size); }
[ Note: You can pick the below patch up by extracting it from the email
and running 'git am -s <file>'. You could then use it as part of a patch series together with your original patch. ]
From: Marco Elver <[email protected]>
Date: Wed, 30 Jun 2021 20:56:57 +0200
Subject: [PATCH] mm: introduce helper to check slub_debug_enabled
Introduce a helper to check slub_debug_enabled, so that we can confine the use of #ifdef to the definition of the slub_debug_enabled_unlikely() helper.
--- mm/slab.h | 15 +++++++++++---- 1 file changed, 11 insertions(+), 4 deletions(-)
mm/slab.h +11 -4
@@ -215,10 +215,18 @@ DECLARE_STATIC_KEY_TRUE(slub_debug_enabled);DECLARE_STATIC_KEY_FALSE (slub_debug_enabled);#endif extern void print_tracking (struct kmem_cache *s,void *object); +static inline bool slub_debug_enabled_unlikely (void ) +{ +return static_branch_unlikely (&slub_debug_enabled); +}#else static inline void print_tracking (struct kmem_cache *s,void *object) { } +static inline bool slub_debug_enabled_unlikely (void ) +{ +return false; +}#endif /*
@@ -228,11 +236,10 @@ static inline void print_tracking(struct kmem_cache *s, void *object)*/ static inline bool kmem_cache_debug_flags (struct kmem_cache *s,slab_flags_t flags) { -#ifdef CONFIG_SLUB_DEBUG -VM_WARN_ON_ONCE (!(flags &SLAB_DEBUG_FLAGS )); -if (static_branch_unlikely (&slub_debug_enabled)) +if (IS_ENABLED (CONFIG_SLUB_DEBUG )) +VM_WARN_ON_ONCE (!(flags &SLAB_DEBUG_FLAGS )); +if (slub_debug_enabled_unlikely ())return s->flags & flags; -#endif return false; }
Signature · 1 line
2.32.0.93.g670b81a890-goog
Andrey Konovalov
reviewerRe: [PATCH v3 1/1] kasan: Add memzero init for unaligned size under SLUB debug
Quoted 17 lines · Marco Elver
Review
Quoted 8 lines
(quoted diff) +4
-0+ if (IS_ENABLED(CONFIG_SLUB_DEBUG) && init && ((unsigned long)size & KASAN_GRANULE_MASK)) { + init = false; + memzero_explicit((void *)addr, size); + } size = round_up(size, KASAN_GRANULE_SIZE); hw_set_mem_tag_range((void *)addr, size, tag, init);I think this solution might be fine for now, as I don't see an easy way to do this without some major refactor to use
kmem_cache_debug_flags().However, I think there's an intermediate solution where we only check the static-key '
slub_debug_enabled' though. Because I've checked, and various major distros _do_ enabledCONFIG_SLUB_DEBUG. But the static branch just makes sure there's no performance overhead.
Quoted 4 lines
Checking the static branch requires including
mm/slab.hintomm/kasan/kasan.h, which we currently don't do and perhaps wanted to avoid. Although I don't see a reason there, because there's no circular dependency even if we did.
Most likely this won't be a problem. We already include ../slab.h into many mm/kasan/*.c files.
Quoted 1 line
Andrey, any opinion?
I like this approach. Easy to implement and is better than checking only CONFIG_SLUB_DEBUG.
Thanks, Marco!