+Stanislav, bpf From: Zhengchao Shao <shaozhengchao@xxxxxxxxxx> Date: Fri, 21 Oct 2022 16:40:58 +0800 > As [0] see, it will cause null-ptr-deref issue. > The following is the process of triggering the problem: > fq_codel_enqueue() > ... > idx = fq_codel_classify() --->if idx != 0 > flow = &q->flows[idx]; > flow_queue_add(flow, skb); --->add skb to flow[idex] > q->backlogs[idx] += qdisc_pkt_len(skb); --->backlogs = 0 > ... > fq_codel_drop() --->set sch->limit = 0, always > drop packets > ... > idx = i --->because backlogs in every > flows is 0, so idx = 0 > ... > flow = &q->flows[idx]; --->get idx=0 flow > ... > dequeue_head() > skb = flow->head; --->flow->head = NULL > flow->head = skb->next; --->cause null-ptr-deref > > So, only need to discard the packets whose len is 0 on dropping path of > enqueue. Then, the correct flow id can be obtained by fq_codel_drop() on > next enqueuing. > > [0]: https://syzkaller.appspot.com/bug?id=0b84da80c2917757915afa89f7738a9d16ec96c5 This can be caused by BPF, but there seems to be no consensus yet. https://lore.kernel.org/netdev/CAKH8qBsOMxVaemF0Oy=vE1V0vKO8ORUcVGB5YANS3HdKOhVjjw@xxxxxxxxxxxxxx/ """ I think the consensus here is that the stack, in general, doesn't expect the packets like this. So there are probably more broken things besides fq_codel. Thus, it's better if we remove the ability to generate them from the bpf side instead of fixing the individual users like fq_codel. """ > > Fixes: 4b549a2ef4be ("fq_codel: Fair Queue Codel AQM") > Signed-off-by: Zhengchao Shao <shaozhengchao@xxxxxxxxxx> > --- > net/sched/sch_fq_codel.c | 8 ++++++++ > 1 file changed, 8 insertions(+) > > diff --git a/net/sched/sch_fq_codel.c b/net/sched/sch_fq_codel.c > index 99d318b60568..3bbe7f69dfb5 100644 > --- a/net/sched/sch_fq_codel.c > +++ b/net/sched/sch_fq_codel.c > @@ -187,6 +187,7 @@ static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch, > struct fq_codel_sched_data *q = qdisc_priv(sch); > unsigned int idx, prev_backlog, prev_qlen; > struct fq_codel_flow *flow; > + struct sk_buff *drop_skb; We can move this into the if-block below or remove. > int ret; > unsigned int pkt_len; > bool memory_limited; > @@ -222,6 +223,13 @@ static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch, > > /* save this packet length as it might be dropped by fq_codel_drop() */ > pkt_len = qdisc_pkt_len(skb); > + > + /* drop skb if len = 0, so fq_codel_drop could get the right flow idx*/ > + if (unlikely(!pkt_len)) { > + drop_skb = dequeue_head(flow); > + __qdisc_drop(drop_skb, to_free); just __qdisc_drop(dequeue_head(flow), to_free); > + return NET_XMIT_SUCCESS; > + } > /* fq_codel_drop() is quite expensive, as it performs a linear search > * in q->backlogs[] to find a fat flow. > * So instead of dropping a single packet, drop half of its backlog > -- > 2.17.1
