From: Daniel Borkmann <[email protected]>
Date: Thu, 13 Apr 2017 17:57:06 +0200
> On 04/12/2017 08:54 PM, David Miller wrote:
> [...]
>> +static u32 netif_receive_generic_xdp(struct sk_buff *skb,
>> + struct bpf_prog *xdp_prog)
>> +{
>> + struct xdp_buff xdp;
>> + u32 act = XDP_DROP;
>> + void *orig_data;
>> + int hlen, off;
>> +
>> + if (skb_linearize(skb))
>
> Btw, given the skb can come from all kind of points in the stack,
> it could also be a clone at this point. One example is act_mirred
> which in fact does skb_clone() and can push the skb back to
> ingress path through netif_receive_skb() and thus could then go
> into generic xdp processing, where skb can be mangled.
>
> Instead of skb_linearize() we would therefore need to use something
> like skb_ensure_writable(skb, skb->len) as equivalent, which also
> makes sure that we unclone whenever needed.
We could use skb_cow() for this purpose, which deals with cloning as
well as enforcing headroom.
However, thinking further about this, the goal is to make generic XDP
match precisely how in-driver-XDP behaves. Therefore, such redirects
from act_mirred would never flow through the XDP path.
No other possibility can cause us to see a cloned packet here, we
are before network taps are processed, etc. So in my opinion the
thing to do is to elide generic XDP if the SKB is cloned.
