Re: KASAN: use-after-free Read in percpu_ref_switch_to_atomic_rcu

[Jann Horn <jannh@google.com>]

On Fri, Mar 6, 2020 at 4:34 PM Jens Axboe <axboe@kernel.dk> wrote:
> On 3/6/20 7:57 AM, Jann Horn wrote:
> > +paulmck
> >
> > On Wed, Mar 4, 2020 at 3:40 PM Jens Axboe <axboe@kernel.dk> wrote:
> >> On 3/4/20 12:59 AM, Dmitry Vyukov wrote:
> >>> On Fri, Feb 7, 2020 at 9:14 AM syzbot
> >>> <syzbot+e017e49c39ab484ac87a@syzkaller.appspotmail.com> wrote:
> >>>>
> >>>> Hello,
> >>>>
> >>>> syzbot found the following crash on:
> >>>>
> >>>> HEAD commit:    4c7d00cc Merge tag 'pwm/for-5.6-rc1' of git://git.kernel.o..
> >>>> git tree:       upstream
> >>>> console output: https://syzkaller.appspot.com/x/log.txt?x=12fec785e00000
> >>>> kernel config:  https://syzkaller.appspot.com/x/.config?x=e162021ddededa72
> >>>> dashboard link: https://syzkaller.appspot.com/bug?extid=e017e49c39ab484ac87a
> >>>> compiler:       clang version 10.0.0 (https://github.com/llvm/llvm-project/ c2443155a0fb245c8f17f2c1c72b6ea391e86e81)
> >>>>
> >>>> Unfortunately, I don't have any reproducer for this crash yet.
> >>>>
> >>>> IMPORTANT: if you fix the bug, please add the following tag to the commit:
> >>>> Reported-by: syzbot+e017e49c39ab484ac87a@syzkaller.appspotmail.com
> >>>
> >>> +io_uring maintainers
> >>>
> >>> Here is a repro:
> >>> https://gist.githubusercontent.com/dvyukov/6b340beab6483a036f4186e7378882ce/raw/cd1922185516453c201df8eded1d4b006a6d6a3a/gistfile1.txt
> >>
> >> I've queued up a fix for this:
> >>
> >> https://git.kernel.dk/cgit/linux-block/commit/?h=io_uring-5.6&id=9875fe3dc4b8cff1f1b440fb925054a5124403c3
> >
> > I believe that this fix relies on call_rcu() having FIFO ordering; but
> > <https://www.kernel.org/doc/Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.html#Callback%20Registry>
> > says:
> >
> > | call_rcu() normally acts only on CPU-local state[...] It simply
> > enqueues the rcu_head structure on a per-CPU list,
> >
> > Is this fix really correct?
>
> That's a good point, there's a potentially stronger guarantee we need
> here that isn't "nobody is inside an RCU critical section", but rather
> that we're depending on a previous call_rcu() to have happened. Hence I
> think you are right - it'll shrink the window drastically, since the
> previous callback is already queued up, but it's not a full close.
>
> Hmm...

You could potentially hack up the semantics you want by doing a
call_rcu() whose callback does another call_rcu(), or something like
that - but I'd like to hear paulmck's opinion on this first.