[v3,1/2] kasan: catch invalid free before SLUB reinitializes the object

Commit Message

Jann Horn July 25, 2024, 3:31 p.m. UTC

Currently, when KASAN is combined with init-on-free behavior, the
initialization happens before KASAN's "invalid free" checks.

More importantly, a subsequent commit will want to use the object metadata
region to store an rcu_head, and we should let KASAN check that the object
pointer is valid before that. (Otherwise that change will make the existing
testcase kmem_cache_invalid_free fail.)

So add a new KASAN hook that allows KASAN to pre-validate a
kmem_cache_free() operation before SLUB actually starts modifying the
object or its metadata.

Acked-by: Vlastimil Babka <[e-mail]> #slub
Signed-off-by: Jann Horn <[e-mail]>
---
 include/linux/kasan.h | 16 ++++++++++++++++
 mm/kasan/common.c     | 51 +++++++++++++++++++++++++++++++++++++++------------
 mm/slub.c             |  7 +++++++
 3 files changed, 62 insertions(+), 12 deletions(-)

Comments

Andrey Konovalov July 26, 2024, 12:43 a.m. UTC | #1

On Thu, Jul 25, 2024 at 5:32 PM Jann Horn <[e-mail]> wrote:
>
> Currently, when KASAN is combined with init-on-free behavior, the
> initialization happens before KASAN's "invalid free" checks.
>
> More importantly, a subsequent commit will want to use the object metadata
> region to store an rcu_head, and we should let KASAN check that the object
> pointer is valid before that. (Otherwise that change will make the existing
> testcase kmem_cache_invalid_free fail.)

This is not the case since v3, right? Do we still need this patch?

If it's still needed, see the comment below.

Thank you!

> So add a new KASAN hook that allows KASAN to pre-validate a
> kmem_cache_free() operation before SLUB actually starts modifying the
> object or its metadata.
>
> Acked-by: Vlastimil Babka <[e-mail]> #slub
> Signed-off-by: Jann Horn <[e-mail]>
> ---
>  include/linux/kasan.h | 16 ++++++++++++++++
>  mm/kasan/common.c     | 51 +++++++++++++++++++++++++++++++++++++++------------
>  mm/slub.c             |  7 +++++++
>  3 files changed, 62 insertions(+), 12 deletions(-)
>
> diff --git a/include/linux/kasan.h b/include/linux/kasan.h
> index 70d6a8f6e25d..ebd93c843e78 100644
> --- a/include/linux/kasan.h
> +++ b/include/linux/kasan.h
> @@ -175,6 +175,16 @@ static __always_inline void * __must_check kasan_init_slab_obj(
>         return (void *)object;
>  }
>
> +bool __kasan_slab_pre_free(struct kmem_cache *s, void *object,
> +                       unsigned long ip);
> +static __always_inline bool kasan_slab_pre_free(struct kmem_cache *s,
> +                                               void *object)
> +{
> +       if (kasan_enabled())
> +               return __kasan_slab_pre_free(s, object, _RET_IP_);
> +       return false;
> +}

Please add a documentation comment for this new hook; something like
what we have for kasan_mempool_poison_pages() and some of the others.
(I've been meaning to add them for all of them, but still didn't get
around to that.)

> +
>  bool __kasan_slab_free(struct kmem_cache *s, void *object,
>                         unsigned long ip, bool init);
>  static __always_inline bool kasan_slab_free(struct kmem_cache *s,
> @@ -371,6 +381,12 @@ static inline void *kasan_init_slab_obj(struct kmem_cache *cache,
>  {
>         return (void *)object;
>  }
> +
> +static inline bool kasan_slab_pre_free(struct kmem_cache *s, void *object)
> +{
> +       return false;
> +}
> +
>  static inline bool kasan_slab_free(struct kmem_cache *s, void *object, bool init)
>  {
>         return false;
> diff --git a/mm/kasan/common.c b/mm/kasan/common.c
> index 85e7c6b4575c..7c7fc6ce7eb7 100644
> --- a/mm/kasan/common.c
> +++ b/mm/kasan/common.c
> @@ -208,31 +208,52 @@ void * __must_check __kasan_init_slab_obj(struct kmem_cache *cache,
>         return (void *)object;
>  }
>
> -static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
> -                                     unsigned long ip, bool init)
> +enum free_validation_result {
> +       KASAN_FREE_IS_IGNORED,
> +       KASAN_FREE_IS_VALID,
> +       KASAN_FREE_IS_INVALID
> +};
> +
> +static enum free_validation_result check_slab_free(struct kmem_cache *cache,
> +                                               void *object, unsigned long ip)
>  {
> -       void *tagged_object;
> +       void *tagged_object = object;
>
> -       if (!kasan_arch_is_ready())
> -               return false;
> +       if (is_kfence_address(object) || !kasan_arch_is_ready())
> +               return KASAN_FREE_IS_IGNORED;
>
> -       tagged_object = object;
>         object = kasan_reset_tag(object);
>
>         if (unlikely(nearest_obj(cache, virt_to_slab(object), object) != object)) {
>                 kasan_report_invalid_free(tagged_object, ip, KASAN_REPORT_INVALID_FREE);
> -               return true;
> +               return KASAN_FREE_IS_INVALID;
>         }
>
> -       /* RCU slabs could be legally used after free within the RCU period. */
> -       if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
> -               return false;
> -
>         if (!kasan_byte_accessible(tagged_object)) {
>                 kasan_report_invalid_free(tagged_object, ip, KASAN_REPORT_DOUBLE_FREE);
> -               return true;
> +               return KASAN_FREE_IS_INVALID;
>         }
>
> +       return KASAN_FREE_IS_VALID;
> +}
> +
> +static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
> +                                     unsigned long ip, bool init)
> +{
> +       void *tagged_object = object;
> +       enum free_validation_result valid = check_slab_free(cache, object, ip);

I believe we don't need check_slab_free() here, as it was already done
in kasan_slab_pre_free()? Checking just kasan_arch_is_ready() and
is_kfence_address() should save a bit on performance impact.

Though if we remove check_slab_free() from here, we do need to add it
to __kasan_mempool_poison_object().

> +
> +       if (valid == KASAN_FREE_IS_IGNORED)
> +               return false;
> +       if (valid == KASAN_FREE_IS_INVALID)
> +               return true;
> +
> +       object = kasan_reset_tag(object);
> +
> +       /* RCU slabs could be legally used after free within the RCU period. */
> +       if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
> +               return false;

I vaguely recall there was some reason why this check was done before
the kasan_byte_accessible() check, but I might be wrong. Could you try
booting the kernel with only this patch applied to see if anything
breaks?




> +
>         kasan_poison(object, round_up(cache->object_size, KASAN_GRANULE_SIZE),
>                         KASAN_SLAB_FREE, init);
>
> @@ -242,6 +263,12 @@ static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
>         return false;
>  }
>
> +bool __kasan_slab_pre_free(struct kmem_cache *cache, void *object,
> +                               unsigned long ip)
> +{
> +       return check_slab_free(cache, object, ip) == KASAN_FREE_IS_INVALID;
> +}
> +
>  bool __kasan_slab_free(struct kmem_cache *cache, void *object,
>                                 unsigned long ip, bool init)
>  {
> diff --git a/mm/slub.c b/mm/slub.c
> index 4927edec6a8c..34724704c52d 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -2170,6 +2170,13 @@ bool slab_free_hook(struct kmem_cache *s, void *x, bool init)
>         if (kfence_free(x))
>                 return false;
>
> +       /*
> +        * Give KASAN a chance to notice an invalid free operation before we
> +        * modify the object.
> +        */
> +       if (kasan_slab_pre_free(s, x))
> +               return false;
> +
>         /*
>          * As memory initialization might be integrated into KASAN,
>          * kasan_slab_free and initialization memset's must be
>
> --
> 2.45.2.1089.g2a221341d9-goog
>

Jann Horn July 26, 2024, 1:51 p.m. UTC | #2

On Fri, Jul 26, 2024 at 2:43 AM Andrey Konovalov <[e-mail]> wrote:
> On Thu, Jul 25, 2024 at 5:32 PM Jann Horn <[e-mail]> wrote:
> >
> > Currently, when KASAN is combined with init-on-free behavior, the
> > initialization happens before KASAN's "invalid free" checks.
> >
> > More importantly, a subsequent commit will want to use the object metadata
> > region to store an rcu_head, and we should let KASAN check that the object
> > pointer is valid before that. (Otherwise that change will make the existing
> > testcase kmem_cache_invalid_free fail.)
>
> This is not the case since v3, right?

Oh, you're right, this text is now wrong.

> Do we still need this patch?

I just tried removing this patch from the series; without it, the
kmem_cache_invalid_free kunit test fails because the kmem_cache_free()
no longer synchronously notices that the pointer is misaligned. I
guess I could change the testcase like this to make the tests pass
without this patch, but I'd like to hear from you or another KASAN
person whether you think that's a reasonable change:

diff --git a/mm/kasan/kasan_test.c b/mm/kasan/kasan_test.c
index cba782a4b072..f44b0dcb0e84 100644
--- a/mm/kasan/kasan_test.c
+++ b/mm/kasan/kasan_test.c
@@ -981,14 +981,21 @@ static void kmem_cache_invalid_free(struct kunit *test)
        if (!p) {
                kunit_err(test, "Allocation failed: %s\n", __func__);
                kmem_cache_destroy(cache);
                return;
        }

-       /* Trigger invalid free, the object doesn't get freed. */
-       KUNIT_EXPECT_KASAN_FAIL(test, kmem_cache_free(cache, p + 1));
+       /*
+        * Trigger invalid free, the object doesn't get freed.
+        * Note that the invalid free detection may happen asynchronously
+        * under CONFIG_SLUB_RCU_DEBUG.
+        */
+       KUNIT_EXPECT_KASAN_FAIL(test, ({
+               kmem_cache_free(cache, p + 1);
+               rcu_barrier();
+       }));

Being able to get rid of this patch would be a nice simplification, so
if you think asynchronous invalid-free detection for TYPESAFE_BY_RCU
slabs is fine, I'll happily throw it out.

Andrey Konovalov July 27, 2024, 12:47 a.m. UTC | #3

On Fri, Jul 26, 2024 at 3:52 PM Jann Horn <[e-mail]> wrote:
>
> > Do we still need this patch?
>
> I just tried removing this patch from the series; without it, the
> kmem_cache_invalid_free kunit test fails because the kmem_cache_free()
> no longer synchronously notices that the pointer is misaligned. I
> guess I could change the testcase like this to make the tests pass
> without this patch, but I'd like to hear from you or another KASAN
> person whether you think that's a reasonable change:

Ah, I see. I think detecting a bug earlier if we can is better. So I
don't mind keeping this patch, was just confused by the commit
message.

Adding on top of my comments from before: I think if you move
check_slab_free() out of poison_slab_object() (but add to
__kasan_mempool_poison_object()), and move is_kfence_address() and
kasan_arch_is_ready() to poison_slab_object()'s callers, you won't
even need the free_validation_result enum, so the patch should become
simpler.

You can also rename check_slab_free() to check_slab_allocation() to
make it be named similarly to the already existing
check_page_allocation(). (I think we should also later move
kasan_arch_is_ready() out of check_page_allocation() into the
high-level hooks for consistency; it also seems cleaner to have all of
these ignore checks in the high-level functions instead of lower-level
inlined ones.)

Thanks!

Jann Horn July 30, 2024, 10:30 a.m. UTC | #4

On Fri, Jul 26, 2024 at 2:43 AM Andrey Konovalov <[e-mail]> wrote:
> On Thu, Jul 25, 2024 at 5:32 PM Jann Horn <[e-mail]> wrote:
> > Currently, when KASAN is combined with init-on-free behavior, the
> > initialization happens before KASAN's "invalid free" checks.
[...]
> > So add a new KASAN hook that allows KASAN to pre-validate a
> > kmem_cache_free() operation before SLUB actually starts modifying the
> > object or its metadata.
> >
> > Acked-by: Vlastimil Babka <[e-mail]> #slub
> > Signed-off-by: Jann Horn <[e-mail]>
> > ---
> >  include/linux/kasan.h | 16 ++++++++++++++++
> >  mm/kasan/common.c     | 51 +++++++++++++++++++++++++++++++++++++++------------
> >  mm/slub.c             |  7 +++++++
> >  3 files changed, 62 insertions(+), 12 deletions(-)
> >
> > diff --git a/include/linux/kasan.h b/include/linux/kasan.h
> > index 70d6a8f6e25d..ebd93c843e78 100644
> > --- a/include/linux/kasan.h
> > +++ b/include/linux/kasan.h
> > @@ -175,6 +175,16 @@ static __always_inline void * __must_check kasan_init_slab_obj(
> >         return (void *)object;
> >  }
> >
> > +bool __kasan_slab_pre_free(struct kmem_cache *s, void *object,
> > +                       unsigned long ip);
> > +static __always_inline bool kasan_slab_pre_free(struct kmem_cache *s,
> > +                                               void *object)
> > +{
> > +       if (kasan_enabled())
> > +               return __kasan_slab_pre_free(s, object, _RET_IP_);
> > +       return false;
> > +}
>
> Please add a documentation comment for this new hook; something like
> what we have for kasan_mempool_poison_pages() and some of the others.
> (I've been meaning to add them for all of them, but still didn't get
> around to that.)

Ack, done in v4.

> > +static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
> > +                                     unsigned long ip, bool init)
> > +{
> > +       void *tagged_object = object;
> > +       enum free_validation_result valid = check_slab_free(cache, object, ip);
>
> I believe we don't need check_slab_free() here, as it was already done
> in kasan_slab_pre_free()? Checking just kasan_arch_is_ready() and
> is_kfence_address() should save a bit on performance impact.
>
> Though if we remove check_slab_free() from here, we do need to add it
> to __kasan_mempool_poison_object().

Ack, changed in v4.

> > +
> > +       if (valid == KASAN_FREE_IS_IGNORED)
> > +               return false;
> > +       if (valid == KASAN_FREE_IS_INVALID)
> > +               return true;
> > +
> > +       object = kasan_reset_tag(object);
> > +
> > +       /* RCU slabs could be legally used after free within the RCU period. */
> > +       if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
> > +               return false;
>
> I vaguely recall there was some reason why this check was done before
> the kasan_byte_accessible() check, but I might be wrong. Could you try
> booting the kernel with only this patch applied to see if anything
> breaks?

I tried booting it to a graphical environment and running the kunit
tests, nothing immediately broke from what I can tell...

Jann Horn July 30, 2024, 10:30 a.m. UTC | #5

On Sat, Jul 27, 2024 at 2:47 AM Andrey Konovalov <[e-mail]> wrote:
> On Fri, Jul 26, 2024 at 3:52 PM Jann Horn <[e-mail]> wrote:
> >
> > > Do we still need this patch?
> >
> > I just tried removing this patch from the series; without it, the
> > kmem_cache_invalid_free kunit test fails because the kmem_cache_free()
> > no longer synchronously notices that the pointer is misaligned. I
> > guess I could change the testcase like this to make the tests pass
> > without this patch, but I'd like to hear from you or another KASAN
> > person whether you think that's a reasonable change:
>
> Ah, I see. I think detecting a bug earlier if we can is better. So I
> don't mind keeping this patch, was just confused by the commit
> message.

ack, changed it in v4

> Adding on top of my comments from before: I think if you move
> check_slab_free() out of poison_slab_object() (but add to
> __kasan_mempool_poison_object()), and move is_kfence_address() and
> kasan_arch_is_ready() to poison_slab_object()'s callers, you won't
> even need the free_validation_result enum, so the patch should become
> simpler.

right, makes sense, changed in v4

> You can also rename check_slab_free() to check_slab_allocation() to
> make it be named similarly to the already existing
> check_page_allocation().

done in v4

Patch

diff --git a/include/linux/kasan.h b/include/linux/kasan.h
index 70d6a8f6e25d..ebd93c843e78 100644
--- a/include/linux/kasan.h
+++ b/include/linux/kasan.h
@@ -175,6 +175,16 @@  static __always_inline void * __must_check kasan_init_slab_obj(
 	return (void *)object;
 }
 
+bool __kasan_slab_pre_free(struct kmem_cache *s, void *object,
+			unsigned long ip);
+static __always_inline bool kasan_slab_pre_free(struct kmem_cache *s,
+						void *object)
+{
+	if (kasan_enabled())
+		return __kasan_slab_pre_free(s, object, _RET_IP_);
+	return false;
+}
+
 bool __kasan_slab_free(struct kmem_cache *s, void *object,
 			unsigned long ip, bool init);
 static __always_inline bool kasan_slab_free(struct kmem_cache *s,
@@ -371,6 +381,12 @@  static inline void *kasan_init_slab_obj(struct kmem_cache *cache,
 {
 	return (void *)object;
 }
+
+static inline bool kasan_slab_pre_free(struct kmem_cache *s, void *object)
+{
+	return false;
+}
+
 static inline bool kasan_slab_free(struct kmem_cache *s, void *object, bool init)
 {
 	return false;
diff --git a/mm/kasan/common.c b/mm/kasan/common.c
index 85e7c6b4575c..7c7fc6ce7eb7 100644
--- a/mm/kasan/common.c
+++ b/mm/kasan/common.c
@@ -208,31 +208,52 @@  void * __must_check __kasan_init_slab_obj(struct kmem_cache *cache,
 	return (void *)object;
 }
 
-static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
-				      unsigned long ip, bool init)
+enum free_validation_result {
+	KASAN_FREE_IS_IGNORED,
+	KASAN_FREE_IS_VALID,
+	KASAN_FREE_IS_INVALID
+};
+
+static enum free_validation_result check_slab_free(struct kmem_cache *cache,
+						void *object, unsigned long ip)
 {
-	void *tagged_object;
+	void *tagged_object = object;
 
-	if (!kasan_arch_is_ready())
-		return false;
+	if (is_kfence_address(object) || !kasan_arch_is_ready())
+		return KASAN_FREE_IS_IGNORED;
 
-	tagged_object = object;
 	object = kasan_reset_tag(object);
 
 	if (unlikely(nearest_obj(cache, virt_to_slab(object), object) != object)) {
 		kasan_report_invalid_free(tagged_object, ip, KASAN_REPORT_INVALID_FREE);
-		return true;
+		return KASAN_FREE_IS_INVALID;
 	}
 
-	/* RCU slabs could be legally used after free within the RCU period. */
-	if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
-		return false;
-
 	if (!kasan_byte_accessible(tagged_object)) {
 		kasan_report_invalid_free(tagged_object, ip, KASAN_REPORT_DOUBLE_FREE);
-		return true;
+		return KASAN_FREE_IS_INVALID;
 	}
 
+	return KASAN_FREE_IS_VALID;
+}
+
+static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
+				      unsigned long ip, bool init)
+{
+	void *tagged_object = object;
+	enum free_validation_result valid = check_slab_free(cache, object, ip);
+
+	if (valid == KASAN_FREE_IS_IGNORED)
+		return false;
+	if (valid == KASAN_FREE_IS_INVALID)
+		return true;
+
+	object = kasan_reset_tag(object);
+
+	/* RCU slabs could be legally used after free within the RCU period. */
+	if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
+		return false;
+
 	kasan_poison(object, round_up(cache->object_size, KASAN_GRANULE_SIZE),
 			KASAN_SLAB_FREE, init);
 
@@ -242,6 +263,12 @@  static inline bool poison_slab_object(struct kmem_cache *cache, void *object,
 	return false;
 }
 
+bool __kasan_slab_pre_free(struct kmem_cache *cache, void *object,
+				unsigned long ip)
+{
+	return check_slab_free(cache, object, ip) == KASAN_FREE_IS_INVALID;
+}
+
 bool __kasan_slab_free(struct kmem_cache *cache, void *object,
 				unsigned long ip, bool init)
 {
diff --git a/mm/slub.c b/mm/slub.c
index 4927edec6a8c..34724704c52d 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2170,6 +2170,13 @@  bool slab_free_hook(struct kmem_cache *s, void *x, bool init)
 	if (kfence_free(x))
 		return false;
 
+	/*
+	 * Give KASAN a chance to notice an invalid free operation before we
+	 * modify the object.
+	 */
+	if (kasan_slab_pre_free(s, x))
+		return false;
+
 	/*
 	 * As memory initialization might be integrated into KASAN,
 	 * kasan_slab_free and initialization memset's must be