Re: [PATCH] net: sched: drr: dont intepret cls results when asked to drop

[Victor Nogueira]

On 15/09/2023 19:55, Jamal Hadi Salim wrote:
> On Fri, Sep 15, 2023 at 11:06 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> > On Fri, Sep 15, 2023 at 5:03 PM Pedro Tammela <pctammela@xxxxxxxxxxxx> wrote:
> > > On 15/09/2023 09:55, Eric Dumazet wrote:
> > > > On Fri, Sep 15, 2023 at 12:42 PM Ma Ke <make_ruc2021@xxxxxxx> wrote:
> > > > > If asked to drop a packet via TC_ACT_SHOT it is unsafe to
> > > > > assume res.class contains a valid pointer.
> > > > >
> > > > > Signed-off-by: Ma Ke <make_ruc2021@xxxxxxx>
> > > > > ---
> > > > >    net/sched/sch_drr.c | 2 ++
> > > > >    1 file changed, 2 insertions(+)
> > > > >
> > > > > diff --git a/net/sched/sch_drr.c b/net/sched/sch_drr.c
> > > > > index 19901e77cd3b..2b854cb6edf9 100644
> > > > > --- a/net/sched/sch_drr.c
> > > > > +++ b/net/sched/sch_drr.c
> > > > > @@ -309,6 +309,8 @@ static struct drr_class *drr_classify(struct sk_buff *skb, struct Qdisc *sch,
> > > > >           *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
> > > > >           fl = rcu_dereference_bh(q->filter_list);
> > > > >           result = tcf_classify(skb, NULL, fl, &res, false);
> > > > > +       if (result == TC_ACT_SHOT)
> > > > > +               return NULL;
> > > > >           if (result >= 0) {
> > > > >    #ifdef CONFIG_NET_CLS_ACT
> > > > >                   switch (result) {
> > > > > --
> > > > > 2.37.2
> > > >
> > > >    I do not see a bug, TC_ACT_SHOT is handled in the switch (result) just fine
> > > > at line 320 ?
> > >
> > > Following the code path (with CONFIG_NET_CLS_ACT=n in mind), it looks
> > > like there are a couple of places which return TC_ACT_SHOT before
> > > calling any classifiers, which then would cause some qdiscs to look into
> > > a uninitialized 'struct tcf_result res'.
> > > I could be misreading it... But if it's the problem the author is trying
> > > to fix, the obvious way to do it would be:
> > >          struct tcf_result res = {};
> >
> > CONFIG_NET_CLS_ACT=n, how come TC_ACT_SHOT could be used ?
> >
> > Can we get rid of CONFIG_NET_CLS_ACT, this seems obfuscation to me at
> > this point.
>
> The problem is the verdict vs return code are intermixed - not saying
> this was fixing anything useful.
> We discussed this in the past after/during commit
> caa4b35b4317d5147b3ab0fbdc9c075c7d2e9c12
> Victor worked on a patch to resolve that. Victor, maybe revive that
> patch and post as RFC?

Okk, will review, rebase on top of net-next and post.

cheers,
Victor