Yes it is about the frag_size check in br_nf_ip_fragment(). As i said
without br_netfilter the packets forwarding is fine.
And I feel it is weird that br_nf_dev_queue_xmit() use out dev's mtu
to decide whether to do the fragmentation, but
then br_nf_ip_fragment() use bridge's mtu to do the actual fragmentation.
And in this case fragmented packets fit the out dev mtu but were
dropped, I think it is not right.
Nikolay Aleksandrov <nikolay@xxxxxxxxxxxxxxxxxxx> 于2019年7月30日周二 下午8:41写道：
>
> On 30/07/2019 15:25, Rundong Ge wrote:
> > Given following setup:
> > -modprobe br_netfilter
> > -echo '1' > /proc/sys/net/bridge/bridge-nf-call-iptables
> > -brctl addbr br0
> > -brctl addif br0 enp2s0
> > -brctl addif br0 enp3s0
> > -brctl addif br0 enp6s0
> > -ifconfig enp2s0 mtu 1300
> > -ifconfig enp3s0 mtu 1500
> > -ifconfig enp6s0 mtu 1500
> > -ifconfig br0 up
> >
> > multi-port
> > mtu1500 - mtu1500|bridge|1500 - mtu1500
> > A | B
> > mtu1300
> >
> > With netfilter defragmentation/conntrack enabled, fragmented
> > packets from A will be defragmented in prerouting, and refragmented
> > at postrouting.
> > But in this scenario the bridge found the frag_max_size(1500) is
> > larger than the dst mtu stored in the fake_rtable whitch is
> > always equal to the bridge's mtu 1300, then packets will be dopped.
> >
> > This modifies ip_skb_dst_mtu to use the out dev's mtu instead
> > of bridge's mtu in bridge refragment.
> >
> > Signed-off-by: Rundong Ge <rdong.ge@xxxxxxxxx>
> > ---
> > include/net/ip.h | 2 ++
> > 1 file changed, 2 insertions(+)
> >
> > diff --git a/include/net/ip.h b/include/net/ip.h
> > index 29d89de..0512de3 100644
> > --- a/include/net/ip.h
> > +++ b/include/net/ip.h
> > @@ -450,6 +450,8 @@ static inline unsigned int ip_dst_mtu_maybe_forward(const struct dst_entry *dst,
> > static inline unsigned int ip_skb_dst_mtu(struct sock *sk,
> > const struct sk_buff *skb)
> > {
> > + if ((skb_dst(skb)->flags & DST_FAKE_RTABLE) && skb->dev)
> > + return min(skb->dev->mtu, IP_MAX_MTU);
> > if (!sk || !sk_fullsock(sk) || ip_sk_use_pmtu(sk)) {
> > bool forwarding = IPCB(skb)->flags & IPSKB_FORWARDED;
> >
> >
>
> I don't think this is correct, there's a reason why the bridge chooses the smallest
> possible MTU out of its members and this is simply a hack to circumvent it.
> If you really like to do so just set the bridge MTU manually, we've added support
> so it won't change automatically to the smallest, but then how do you pass packets
> 1500 -> 1300 in this setup ?
>
> You're talking about the frag_size check in br_nf_ip_fragment(), right ?
>
