Re: net: fix routing encapsulated packets when binding a socket to a tunnel interface

[David Miller]

From: lifonghsu <lifonghsu@xxxxxxxxxxxx>
Date: Fri, 22 Mar 2019 14:29:58 +0800

> From: LiFong Hsu <lifonghsu@xxxxxxxxxxxx>
> 
> When binding a socket to a 4in6/6in4 tunnel interface, the kernel sends
> the packet to the tunnel interface without any problem, e.g.,
> ping 8.8.8.8 -I 4in6. However, after the 4in6/6in4 tunnel encapsulation,
> the encapsulated packet could be sent to the tunnel interface again when
> some fields of the skb were changed in mangle table's output chain,
> such as skb->mark and src/dest IP address. Sending to the tunnel interface
> twice is unexpected, since there are no corresponding routing rules on
> the tunnel interface for the encapsulated packet. Eventually, the encapsulated
> packet will be dropped by the tunnel interface.
> 
> This commit stops referring to sk_bound_dev_if while re-routing a packet
> with skb_iif!=0 which indicates that the packet has already been sent to
> the interface specified by sk_bound_dev_if. Instead, this commit sends
> the packet to the underlying network device.
> 
> Signed-off-by: LiFong Hsu <lifonghsu@xxxxxxxxxxxx>
> Reviewed-by: JianJhen Chen <kchen@xxxxxxxxxxxx>

skb->skb_iif is a receive side indication, why would it be changed on
transmit?

I see mac802154 doing this, but what it's doing is somewhat broken and
that doesn't come into play in your example.