From: Andrey Konovalov <andreyknvl@google.com>

KASAN's quarantine might save its metadata inside freed objects. As
this happens after the memory is zeroed by the slab allocator when
init_on_free is enabled, the memory coming out of quarantine is not
properly zeroed.

This causes lib/test_meminit.c tests to fail with Generic KASAN.

Zero the metadata when the object is removed from quarantine.

Fixes: 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options")
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
---
 mm/kasan/quarantine.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/mm/kasan/quarantine.c b/mm/kasan/quarantine.c
index 587da8995f2d..2e50869fd8e2 100644
--- a/mm/kasan/quarantine.c
+++ b/mm/kasan/quarantine.c
@@ -132,11 +132,22 @@ static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
 static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
 {
 	void *object = qlink_to_object(qlink, cache);
+	struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
 	unsigned long flags;
 
 	if (IS_ENABLED(CONFIG_SLAB))
 		local_irq_save(flags);
 
+	/*
+	 * If init_on_free is enabled and KASAN's free metadata is stored in
+	 * the object, zero the metadata. Otherwise, the object's memory will
+	 * not be properly zeroed, as KASAN saves the metadata after the slab
+	 * allocator zeroes the object.
+	 */
+	if (slab_want_init_on_free(cache) &&
+	    cache->kasan_info.free_meta_offset == 0)
+		memset(meta, 0, sizeof(*meta));
+
 	/*
 	 * As the object now gets freed from the quarantine, assume that its
 	 * free track is no longer valid.
-- 
2.25.1

On Mon, 20 Dec 2021 at 17:37, <andrey.konovalov@linux.dev> wrote:
>
> From: Andrey Konovalov <andreyknvl@google.com>
>
> KASAN's quarantine might save its metadata inside freed objects. As
> this happens after the memory is zeroed by the slab allocator when
> init_on_free is enabled, the memory coming out of quarantine is not
> properly zeroed.
>
> This causes lib/test_meminit.c tests to fail with Generic KASAN.
>
> Zero the metadata when the object is removed from quarantine.
>
> Fixes: 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options")
> Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
> ---
>  mm/kasan/quarantine.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
>
> diff --git a/mm/kasan/quarantine.c b/mm/kasan/quarantine.c
> index 587da8995f2d..2e50869fd8e2 100644
> --- a/mm/kasan/quarantine.c
> +++ b/mm/kasan/quarantine.c
> @@ -132,11 +132,22 @@ static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
>  static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
>  {
>         void *object = qlink_to_object(qlink, cache);
> +       struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
>         unsigned long flags;
>
>         if (IS_ENABLED(CONFIG_SLAB))
>                 local_irq_save(flags);
>
> +       /*
> +        * If init_on_free is enabled and KASAN's free metadata is stored in
> +        * the object, zero the metadata. Otherwise, the object's memory will
> +        * not be properly zeroed, as KASAN saves the metadata after the slab
> +        * allocator zeroes the object.
> +        */
> +       if (slab_want_init_on_free(cache) &&
> +           cache->kasan_info.free_meta_offset == 0)
> +               memset(meta, 0, sizeof(*meta));

memzero_explicit()

although in this case it probably doesn't matter much, because AFAIK
memzero_explicit() only exists to prevent the compiler from eliding
the zeroing. Up to you.

> +
>         /*
>          * As the object now gets freed from the quarantine, assume that its
>          * free track is no longer valid.
> --
> 2.25.1
>

On Mon, Dec 20, 2021 at 6:07 PM Marco Elver <elver@google.com> wrote:
>
> On Mon, 20 Dec 2021 at 17:37, <andrey.konovalov@linux.dev> wrote:
> >
> > From: Andrey Konovalov <andreyknvl@google.com>
> >
> > KASAN's quarantine might save its metadata inside freed objects. As
> > this happens after the memory is zeroed by the slab allocator when
> > init_on_free is enabled, the memory coming out of quarantine is not
> > properly zeroed.
> >
> > This causes lib/test_meminit.c tests to fail with Generic KASAN.
> >
> > Zero the metadata when the object is removed from quarantine.
> >
> > Fixes: 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options")
> > Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
> > ---
> >  mm/kasan/quarantine.c | 11 +++++++++++
> >  1 file changed, 11 insertions(+)
> >
> > diff --git a/mm/kasan/quarantine.c b/mm/kasan/quarantine.c
> > index 587da8995f2d..2e50869fd8e2 100644
> > --- a/mm/kasan/quarantine.c
> > +++ b/mm/kasan/quarantine.c
> > @@ -132,11 +132,22 @@ static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
> >  static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
> >  {
> >         void *object = qlink_to_object(qlink, cache);
> > +       struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
> >         unsigned long flags;
> >
> >         if (IS_ENABLED(CONFIG_SLAB))
> >                 local_irq_save(flags);
> >
> > +       /*
> > +        * If init_on_free is enabled and KASAN's free metadata is stored in
> > +        * the object, zero the metadata. Otherwise, the object's memory will
> > +        * not be properly zeroed, as KASAN saves the metadata after the slab
> > +        * allocator zeroes the object.
> > +        */
> > +       if (slab_want_init_on_free(cache) &&
> > +           cache->kasan_info.free_meta_offset == 0)
> > +               memset(meta, 0, sizeof(*meta));
>
> memzero_explicit()
>
> although in this case it probably doesn't matter much, because AFAIK
> memzero_explicit() only exists to prevent the compiler from eliding
> the zeroing. Up to you.

I've thought about using memzero_explicit(), but the rest of
init_on_alloc/free code uses memset(0) so I decided to use it as well.
If we decide to switch to memzero_explicit(), it makes sense to do it
everywhere.

Thanks!

On Mon, 20 Dec 2021 at 18:16, Andrey Konovalov <andreyknvl@gmail.com> wrote:
>
> On Mon, Dec 20, 2021 at 6:07 PM Marco Elver <elver@google.com> wrote:
> >
> > On Mon, 20 Dec 2021 at 17:37, <andrey.konovalov@linux.dev> wrote:
> > >
> > > From: Andrey Konovalov <andreyknvl@google.com>
> > >
> > > KASAN's quarantine might save its metadata inside freed objects. As
> > > this happens after the memory is zeroed by the slab allocator when
> > > init_on_free is enabled, the memory coming out of quarantine is not
> > > properly zeroed.
> > >
> > > This causes lib/test_meminit.c tests to fail with Generic KASAN.
> > >
> > > Zero the metadata when the object is removed from quarantine.
> > >
> > > Fixes: 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options")
> > > Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
> > > ---
> > >  mm/kasan/quarantine.c | 11 +++++++++++
> > >  1 file changed, 11 insertions(+)
> > >
> > > diff --git a/mm/kasan/quarantine.c b/mm/kasan/quarantine.c
> > > index 587da8995f2d..2e50869fd8e2 100644
> > > --- a/mm/kasan/quarantine.c
> > > +++ b/mm/kasan/quarantine.c
> > > @@ -132,11 +132,22 @@ static void *qlink_to_object(struct qlist_node *qlink, struct kmem_cache *cache)
> > >  static void qlink_free(struct qlist_node *qlink, struct kmem_cache *cache)
> > >  {
> > >         void *object = qlink_to_object(qlink, cache);
> > > +       struct kasan_free_meta *meta = kasan_get_free_meta(cache, object);
> > >         unsigned long flags;
> > >
> > >         if (IS_ENABLED(CONFIG_SLAB))
> > >                 local_irq_save(flags);
> > >
> > > +       /*
> > > +        * If init_on_free is enabled and KASAN's free metadata is stored in
> > > +        * the object, zero the metadata. Otherwise, the object's memory will
> > > +        * not be properly zeroed, as KASAN saves the metadata after the slab
> > > +        * allocator zeroes the object.
> > > +        */
> > > +       if (slab_want_init_on_free(cache) &&
> > > +           cache->kasan_info.free_meta_offset == 0)
> > > +               memset(meta, 0, sizeof(*meta));
> >
> > memzero_explicit()
> >
> > although in this case it probably doesn't matter much, because AFAIK
> > memzero_explicit() only exists to prevent the compiler from eliding
> > the zeroing. Up to you.
>
> I've thought about using memzero_explicit(), but the rest of
> init_on_alloc/free code uses memset(0) so I decided to use it as well.
> If we decide to switch to memzero_explicit(), it makes sense to do it
> everywhere.

memzero_explicit() is newer than those existing memset(0) -- new code
should probably start using it.

So I'd opt for just using it here. Who knows what other optimizations
future compilers may come up with.