Re: [PATCH net] cls_matchall: avoid panic when receiving a packet before filter set

[Cong Wang <xiyou.wangcong@gmail.com>]

On Wed, May 1, 2019 at 2:27 AM Matteo Croce <mcroce@redhat.com> wrote:
>
> On Tue, Apr 30, 2019 at 11:25 PM Cong Wang <xiyou.wangcong@gmail.com> wrote:
> >
> > On Mon, Apr 29, 2019 at 10:38 AM Matteo Croce <mcroce@redhat.com> wrote:
> > >
> > > When a matchall classifier is added, there is a small time interval in
> > > which tp->root is NULL. If we receive a packet in this small time slice
> > > a NULL pointer dereference will happen, leading to a kernel panic:
> >
> > Hmm, why not just check tp->root against NULL in mall_classify()?
> >
> > Also, which is the offending commit here? Please add a Fixes: tag.
> >
> > Thanks.
>
> Hi,
>
> I just want to avoid an extra check which would be made for every packet.
> Probably the benefit over a check is negligible, but it's still a
> per-packet thing.
> If you prefer a simple check, I can make a v2 that way.

Yeah, I think that is better, you can add an unlikely() for performance
concern, as NULL is a rare case.


>
> For the fixes tag, I didn't put it as I'm not really sure about the
> offending commit. I guess it's the following, what do you think?
>
> commit ed76f5edccc98fa66f2337f0b3b255d6e1a568b7
> Author: Vlad Buslov <vladbu@mellanox.com>
> Date:   Mon Feb 11 10:55:38 2019 +0200
>
>     net: sched: protect filter_chain list with filter_chain_lock mutex

I think you are right, this is the commit introduced the code
that inserts the tp before fully initializing it. Please Cc Vlad
for your v2, in case we blame a wrong commit here.


BTW, it looks like cls_cgroup needs a same fix. Please audit
other tc filters as well.

Thanks!