On Fri, Jun 10, 2022 at 12:07 AM Di Zhu <[email protected]> wrote:
>
> Syzbot found an issue [1]: fq_codel_drop() try to drop a flow whitout any
> skbs, that is, the flow->head is null.
> The root cause is that: when the first queued skb with pkt_len 0, backlogs
> of the flow that this skb enqueued is still 0 and if sch->limit is set to
> 0 then fq_codel_drop() will be called. At this point, the backlogs of all
> flows are all 0, so flow with idx 0 is selected to drop, but this flow have
> not any skbs.
> skb with pkt_len 0 can break existing processing logic, so just discard
> these invalid skbs.
>
> LINK: [1] https://syzkaller.appspot.com/bug?id=0b84da80c2917757915afa89f7738a9d16ec96c5
>
> Reported-by: [email protected]
> Signed-off-by: Di Zhu <[email protected]>
> ---
> net/sched/sch_fq_codel.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/net/sched/sch_fq_codel.c b/net/sched/sch_fq_codel.c
> index 839e1235db05..c0f82b7358e1 100644
> --- a/net/sched/sch_fq_codel.c
> +++ b/net/sched/sch_fq_codel.c
> @@ -191,6 +191,9 @@ static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
> unsigned int pkt_len;
> bool memory_limited;
>
> + if (unlikely(!qdisc_pkt_len(skb)))
> + return qdisc_drop(skb, sch, to_free);
> +
This has been discussed in the past.
Feeding ndo_start_xmit() in hundreds of drivers with zero-length
packets will crash anyway.
We are not going to add such silly tests in all qdiscs, and then all
ndo_start_xmit(), since qdiscs are not mandatory.
Please instead fix BPF layer, instead of hundreds of drivers/qdiscs.
On Fri, Jun 10, 2022 at 12:32 AM Eric Dumazet <[email protected]> wrote:
>
> On Fri, Jun 10, 2022 at 12:07 AM Di Zhu <[email protected]> wrote:
> >
> > Syzbot found an issue [1]: fq_codel_drop() try to drop a flow whitout any
> > skbs, that is, the flow->head is null.
> > The root cause is that: when the first queued skb with pkt_len 0, backlogs
> > of the flow that this skb enqueued is still 0 and if sch->limit is set to
> > 0 then fq_codel_drop() will be called. At this point, the backlogs of all
> > flows are all 0, so flow with idx 0 is selected to drop, but this flow have
> > not any skbs.
> > skb with pkt_len 0 can break existing processing logic, so just discard
> > these invalid skbs.
> >
> > LINK: [1] https://syzkaller.appspot.com/bug?id=0b84da80c2917757915afa89f7738a9d16ec96c5
> >
> > Reported-by: [email protected]
> > Signed-off-by: Di Zhu <[email protected]>
> > ---
> > net/sched/sch_fq_codel.c | 3 +++
> > 1 file changed, 3 insertions(+)
> >
> > diff --git a/net/sched/sch_fq_codel.c b/net/sched/sch_fq_codel.c
> > index 839e1235db05..c0f82b7358e1 100644
> > --- a/net/sched/sch_fq_codel.c
> > +++ b/net/sched/sch_fq_codel.c
> > @@ -191,6 +191,9 @@ static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
> > unsigned int pkt_len;
> > bool memory_limited;
> >
> > + if (unlikely(!qdisc_pkt_len(skb)))
> > + return qdisc_drop(skb, sch, to_free);
> > +
>
>
> This has been discussed in the past.
>
https://www.spinics.net/lists/netdev/msg777503.html
> Feeding ndo_start_xmit() in hundreds of drivers with zero-length
> packets will crash anyway.
>
> We are not going to add such silly tests in all qdiscs, and then all
> ndo_start_xmit(), since qdiscs are not mandatory.
>
> Please instead fix BPF layer, instead of hundreds of drivers/qdiscs.