Re: [PATCH net] net: sched: act_mirred: drop skb's dst_entry in ingress redirection

[Eyal Birger <eyal.birger@gmail.com>]

Hi,

On Mon, 21 Oct 2019 13:50:13 -0700
Cong Wang <xiyou.wangcong@gmail.com> wrote:

> On Fri, Oct 18, 2019 at 2:25 PM Eyal Birger <eyal.birger@gmail.com>
> wrote:
> >
> > Hi,
> >
> > On Fri, 18 Oct 2019 00:33:53 +0800
> > Zhiyuan Hou <zhiyuan2048@linux.alibaba.com> wrote:
> >
> > > On 2019/10/16 8:13 下午, Eyal Birger wrote:
> > > > Hi,
> > > >
> > > > On Wed, 16 Oct 2019 01:22:01 +0800
> > > > Zhiyuan Hou <zhiyuan2048@linux.alibaba.com> wrote:
> > > >
> > > >> On 2019/10/15 1:57 上午, Cong Wang wrote:
> > > >>> On Sat, Oct 12, 2019 at 12:16 AM Zhiyuan Hou
> > > >>> <zhiyuan2048@linux.alibaba.com> wrote:
> > > >>>> diff --git a/net/sched/act_mirred.c b/net/sched/act_mirred.c
> > > >>>> index 9ce073a05414..6108a64c0cd5 100644
> > > >>>> --- a/net/sched/act_mirred.c
> > > >>>> +++ b/net/sched/act_mirred.c
> > > >>>> @@ -18,6 +18,7 @@
> > > >>>>    #include <linux/gfp.h>
> > > >>>>    #include <linux/if_arp.h>
> > > >>>>    #include <net/net_namespace.h>
> > > >>>> +#include <net/dst.h>
> > > >>>>    #include <net/netlink.h>
> > > >>>>    #include <net/pkt_sched.h>
> > > >>>>    #include <net/pkt_cls.h>
> > > >>>> @@ -298,8 +299,10 @@ static int tcf_mirred_act(struct sk_buff
> > > >>>> *skb, const struct tc_action *a,
> > > >>>>
> > > >>>>           if (!want_ingress)
> > > >>>>                   err = dev_queue_xmit(skb2);
> > > >>>> -       else
> > > >>>> +       else {
> > > >>>> +               skb_dst_drop(skb2);
> > > >>>>                   err = netif_receive_skb(skb2);
> > > >>>> +       }
> > > >>> Good catch!
> > > > Indeed! Thanks for fixing this!
> > > >
> > > >>> I don't want to be picky, but it seems this is only needed
> > > >>> when redirecting from egress to ingress, right? That is,
> > > >>> ingress to ingress, or ingress to egress is okay? If not,
> > > >>> please fix all the cases while you are on it?
> > > >> Sure. But I think this patch is also needed when redirecting
> > > >> from ingress to ingress. Because we cannot assure that a skb
> > > >> has null dst in ingress redirection path. For example, if
> > > >> redirecting a skb from loopback's ingress to other device's
> > > >> ingress, the skb will take a dst.
> > > >>
> > > >> As commit logs point out, skb with valid dst cannot be made
> > > >> routing decision in following process. original dst may cause
> > > >> skb loss or other unexpected behavior.
> > > > On the other hand, removing the dst on ingress-to-ingress
> > > > redirection may remove LWT information on incoming packets,
> > > > which may be undesired.
> > > Sorry, I do not understand why lwt information is needed on
> > > ingress-to-ingress redirection. lwt is used on output path, isn't
> > > it? Can you please give more information?
> >
> > On rx path tunnelled packets parameters received on a collect_md
> > tunnel device are kept in a metadata dst. See ip_tunnel_rcv()
> > 'tun_dst' parameter.
> >
> > The rx metadata dst can be matched by a number of mechanisms like
> > routing rules, eBPF, OVS, and netfilter.
> 
> Should this meta information be kept when redirecting? The dest device
> may be a non-tunnel device, so I don't know if it is still useful when
> for non-tunnel devices.

I think that on ingress-to-ingress redirect it would make sense to keep the
metadata.

The dest device does not have to be a tunnel device AFAICT in order to use
tunnel info as skb_tunnel_info() does not observe skb->dev.

I don't see why going through mirred redirect should prevent the admin from
matching the packet based on LWT metadata - a packet may arrive on a collect_md
tunnel device, be ingress-redirected to different devices based on different
criteria, then routed based also on the tunnel parameters.

Eyal.