Re: [PATCH v2 1/3] rcu/kasan: record and print call_rcu() call stack

From: Dmitry Vyukov
Date: Wed May 13 2020 - 05:16:16 EST


On Wed, May 13, 2020 at 11:05 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
>
> On Wed, 2020-05-13 at 08:51 +0200, 'Dmitry Vyukov' via kasan-dev wrote:
> > On Wed, May 13, 2020 at 3:48 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > > > > > Are you sure it will increase object size?
> > > > > > I think we overlap kasan_free_meta with the object as well. The only
> > > > > > case we don't overlap kasan_free_meta with the object are
> > > > > > SLAB_TYPESAFE_BY_RCU || cache->ctor. But these are rare and it should
> > > > > > only affect small objects with small redzones.
> > > > > > And I think now we simply have a bug for these objects, we check
> > > > > > KASAN_KMALLOC_FREE and then assume object contains free stack, but for
> > > > > > objects with ctor, they still contain live object data, we don't store
> > > > > > free stack in them.
> > > > > > Such objects can be both free and still contain user data.
> > > > > >
> > > > >
> > > > > Overlay kasan_free_meta. I see. but overlay it only when the object was
> > > > > freed. kasan_free_meta will be used until free object.
> > > > > 1). When put object into quarantine, it need kasan_free_meta.
> > > > > 2). When the object exit from quarantine, it need kasan_free_meta
> > > > >
> > > > > If we choose to overlay kasan_free_meta, then the free stack will be
> > > > > stored very late. It may has no free stack in report.
> > > >
> > > > Sorry, I don't understand what you mean.
> > > >
> > > > Why will it be stored too late?
> > > > In __kasan_slab_free() putting into quarantine and recording free
> > > > stack are literally adjacent lines of code:
> > > >
> > > > static bool __kasan_slab_free(struct kmem_cache *cache, void *object,
> > > > unsigned long ip, bool quarantine)
> > > > {
> > > > ...
> > > > kasan_set_free_info(cache, object, tag);
> > > > quarantine_put(get_free_info(cache, object), cache);
> > > >
> > > >
> > > > Just to make sure, what I meant is that we add free_track to kasan_free_meta:
> > > >
> > > > struct kasan_free_meta {
> > > > struct qlist_node quarantine_link;
> > > > + struct kasan_track free_track;
> > > > };
> > > >
> > >
> > > When I see above struct kasan_free_meta, I know why you don't understand
> > > my meaning, because I thought you were going to overlay the
> > > quarantine_link by free_track, but it seems like to add free_track to
> > > kasan_free_meta. Does it enlarge meta-data size?
> >
> > I would assume it should not increase meta-data size. In both cases we
> > store exactly the same information inside of the object: quarantine
> > link and free track.
> > I see it more as a question of code organization. We already have a
> > concept of "this data is placed inside of the freed object", we
> > already have a name for it (kasan_free_meta), we already have code to
> > choose where to place it, we already have helper functions to access
> > it. And your change effectively duplicates all of this to place the
> > free track.
> >
>
> I want to make a summary. Which of the following is the approach we
> want? or if I have some misunderstandings, please help me to correct.
> Thanks.
>
> 1) For different object, then it will has two ways.
> 1.a) When object are LAB_TYPESAFE_BY_RCU || cache->ctor, then store free
> stack into free track of struct kasan_free_meta.
> 2.b) Except 1.a), store free stack into freed object.
>
> or
>
> 2) We always store free stack into free track of struct kasan_free_meta

I meant 2): We always store free stack into free track of struct
kasan_free_meta.
I think it will do the same as other options but just with less code
(and simpler code).

Maybe I am missing something here?




> > > > And I think its life-time and everything should be exactly what we need.
> > > >
> > > > Also it should help to fix the problem with ctors: kasan_free_meta is
> > > > already allocated on the side for such objects, and that's exactly
> > > > what we need for objects with ctor's.
> > >
> > > I see.