Re: [PATCH] slub: Introduce CONFIG_SLUB_RCU_DEBUG

[Jann Horn]

On Sat, Aug 26, 2023 at 5:32 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> On Fri, 25 Aug 2023 at 23:15, Jann Horn <jannh@xxxxxxxxxx> wrote:
> > Currently, KASAN is unable to catch use-after-free in SLAB_TYPESAFE_BY_RCU
> > slabs because use-after-free is allowed within the RCU grace period by
> > design.
> >
> > Add a SLUB debugging feature which RCU-delays every individual
> > kmem_cache_free() before either actually freeing the object or handing it
> > off to KASAN, and change KASAN to poison freed objects as normal when this
> > option is enabled.
> >
> > Note that this creates a 16-byte unpoisoned area in the middle of the
> > slab metadata area, which kinda sucks but seems to be necessary in order
> > to be able to store an rcu_head in there without triggering an ASAN
> > splat during RCU callback processing.
>
> Nice!
>
> Can't we unpoision this rcu_head right before call_rcu() and repoison
> after receiving the callback?

Yeah, I think that should work. It looks like currently
kasan_unpoison() is exposed in include/linux/kasan.h but
kasan_poison() is not, and its inline definition probably means I
can't just move it out of mm/kasan/kasan.h into include/linux/kasan.h;
do you have a preference for how I should handle this? Hmm, and it
also looks like code outside of mm/kasan/ anyway wouldn't know what
are valid values for the "value" argument to kasan_poison().
I also have another feature idea that would also benefit from having
something like kasan_poison() available in include/linux/kasan.h, so I
would prefer that over adding another special-case function inside
KASAN for poisoning this piece of slab metadata...

I guess I could define a wrapper around kasan_poison() in
mm/kasan/generic.c that uses a new poison value for "some other part
of the kernel told us to poison this area", and then expose that
wrapper with a declaration in include/mm/kasan.h? Something like:

void kasan_poison_outline(const void *addr, size_t size, bool init)
{
  kasan_poison(addr, size, KASAN_CUSTOM, init);
}

> What happens on cache destruction?
> Currently we purge quarantine on cache destruction to be able to
> safely destroy the cache. I suspect we may need to somehow purge rcu
> callbacks as well, or do something else.

Ooh, good point, I hadn't thought about that... currently
shutdown_cache() assumes that all the objects have already been freed,
then puts the kmem_cache on a list for
slab_caches_to_rcu_destroy_workfn(), which then waits with an
rcu_barrier() until the slab's pages are all gone.

Luckily kmem_cache_destroy() is already a sleepable operation, so
maybe I should just slap another rcu_barrier() in there for builds
with this config option enabled... I think that should be fine for an
option mostly intended for debugging.