On 8/27/19 1:42 PM, Shmulik Ladkani wrote:

On Mon, 26 Aug 2019 19:47:40 +0200
Eric Dumazet [off-list ref] wrote:

quoted

On 8/26/19 4:07 PM, Shmulik Ladkani wrote:

quoted

  - ipv4 forwarding to dummy1, where eBPF nat4-to-6 program is attached
    at TC Egress (calls 'bpf_skb_change_proto()'), then redirect to ingress
    on same device.
    NOTE: 'bpf_skb_proto_4_to_6()' mangles 'shinfo->gso_size'  

Doing this on an skb with a frag_list is doomed, in current gso_segment() state.

A rewrite  would be needed (I believe I did so at some point, but Herbert Xu fought hard against it)

Thanks Eric,

- If a rewrite is still considered out of the question, how can one use
  eBPF's bpf_skb_change_proto() safely without disabling GRO completely?
  - e.g. is there a way to force the GROed skbs to fit into a layout that is
    tolerated by skb_segment?
  - alternatively can eBPF layer somehow enforce segmentation of the
    original GROed skb before mangling the gso_size?

- Another thing that puzzles me is that we hit the BUG_ON rather rarely
  and cannot yet reproduce synthetically. If skb_segment's handling of
  skbs with a frag_list (that have gso_size mangled) is broken, I'd expect
  to hit this more often... Any ideas?

skb_segment of a gro packet (especially with frag_list) is only supported 
if the geometry of the individual segments is not changed,
meaning that gso_size must remain the same.

- Suppose going for a rewrite, care to elaborate what's exactly missing
  in skb_segment's logic?
  I must admit I do not fully understand all the different code flows in
  this function, it seems to support many different input skbs - any
  assistance is highly appreciated.

Well, this is the point really.
The complexity of this function is so high that very few of us dare to touch it.