[lvc-project] [PATCH 2/2] can: j1939: avoid possible use-after-free when j1939_can_rx_register fails

[Oleksij Rempel]

Hi Fedor,

On Fri, May 26, 2023 at 09:50:26PM +0300, Fedor Pchelkin wrote:
> Hi Oleksij,
> 
> thanks for the reply!
> 
> On Fri, May 26, 2023 at 08:15:00PM +0200, Oleksij Rempel wrote:
> > Hi Fedor,
> > 
> > On Fri, May 26, 2023 at 08:19:10PM +0300, Fedor Pchelkin wrote:
> > 
> > 
> > Thank you for your investigation. How about this change?
> > --- a/net/can/j1939/main.c
> > +++ b/net/can/j1939/main.c
> > @@ -285,8 +285,7 @@ struct j1939_priv *j1939_netdev_start(struct net_device *ndev)
> >                  */
> >                 kref_get(&priv_new->rx_kref);
> >                 spin_unlock(&j1939_netdev_lock);
> > -               dev_put(ndev);
> > -               kfree(priv);
> > +               j1939_priv_put(priv);
> 
> I don't think that's good because the priv which is directly freed here is
> still local to the thread, and parallel threads don't have any access to
> it. j1939_priv_create() has allocated a fresh priv and called dev_hold()
> so dev_put() and kfree() here are okay.
> 
> >                 return priv_new;
> >         }
> >         j1939_priv_set(ndev, priv);
> > @@ -300,8 +299,7 @@ struct j1939_priv *j1939_netdev_start(struct net_device *ndev)
> >  
> >   out_priv_put:
> >         j1939_priv_set(ndev, NULL);
> > -       dev_put(ndev);
> > -       kfree(priv);
> > +       j1939_priv_put(priv);
> >  
> >         return ERR_PTR(ret);
> >  }
> > 
> > If I see it correctly, the problem is kfree() which is called without respecting
> > the ref counting. If CPU1 has priv_new, refcounting is increased. The priv will
> > not be freed on this place.
> 
> With your suggestion, I think it doesn't work correctly if
> j1939_can_rx_register() fails and we go to out_priv_put. The priv is kept
> but the parallel thread which may have already grabbed it thinks that
> j1939_can_rx_register() has succeeded when actually it hasn't succeed.
> Moreover, j1939_priv_set() makes it NULL on error path so that priv cannot
> be accessed from ndev.
> 
> I also considered the alternatives where we don't have to serialize access
> to j1939_can_rx_register() and subsequently introduce mutex. But with
> current j1939_netdev_start() implementation I can't see how to fix the
> racy bug without it.
 
Ok, it make sense.

I'll try to do some testing next week. If i'll forget it, please feel
free to ping me.

Regards,
Oleksij
-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |