Re: possible deadlock in rfcomm_sk_state_change

From: Luiz Augusto von Dentz
Date: Tue Aug 30 2022 - 13:45:24 EST


Hi Desmond,

On Mon, Aug 29, 2022 at 11:48 PM Desmond Cheong Zhi Xi
<desmondcheongzx@xxxxxxxxx> wrote:
>
> +cc Bluetooth and Networking maintainers
>
> Hi Jiacheng,
>
> On 28/8/22 04:03, Jiacheng Xu wrote:
> > Hi,
> >
> > I believe the deadlock is more than possible but actually real.
> > I got a poc that could stably trigger the deadlock.
> >
> > poc: https://drive.google.com/file/d/1PjqvMtHsrrGM1MIRGKl_zJGR-teAMMQy/view?usp=sharing
> >
> > Description/Root cause:
> > In rfcomm_sock_shutdown(), lock_sock() is called when releasing and
> > shutting down socket.
> > However, lock_sock() has to be called once more when the sk_state is
> > changed because the
> > lock is not always held when rfcomm_sk_state_change() is called. One
> > such call stack is:
> >
> > rfcomm_sock_shutdown():
> > lock_sock();
> > __rfcomm_sock_close():
> > rfcomm_dlc_close():
> > __rfcomm_dlc_close():
> > rfcomm_dlc_lock();
> > rfcomm_sk_state_change():
> > lock_sock();
> >
> > Besides the recursive deadlock, there is also an
> > issue of a lock hierarchy inversion between rfcomm_dlc_lock() and
> > lock_sock() if the socket is locked in rfcomm_sk_state_change().
>
>
> Thanks for the poc and for following the trail all the way to the root
> cause - this was a known issue and I didn't realize the patch wasn't
> applied.
>
> > > Reference:
> https://lore.kernel.org/all/20211004180734.434511-1-desmondcheongzx@xxxxxxxxx/
> >
>
> Fwiw, I tested the patch again with syzbot. It still applies cleanly to
> the head of bluetooth-next and seems to address the root cause.
>
> Any thoughts from the maintainers on this issue and the proposed fix?

We probably need to introduce a test to rfcomm-tester to reproduce
this sort of problem, I also would like to avoid introducing a work
just to trigger a state change since we don't have such problem on the
likes of L2CAP socket so perhaps we need to rework the code a little
bit to avoid the locking problems.

> Best,
> Desmond



--
Luiz Augusto von Dentz