Hi, On Mon, 16 Jul 2018 16:39:55 -0700 Cong Wang <xiyou.wangc...@gmail.com> wrote:
> On Fri, Jul 13, 2018 at 2:55 AM Paolo Abeni <pab...@redhat.com> wrote: > > > > When mirred is invoked from the ingress path, and it wants to > > redirect the processed packet, it can now use the ACT_REDIRECT > > action, filling the tcf_result accordingly. > > > > This avoids a skb_clone() in the TC S/W data path giving a ~10% > > improvement in forwarding performances. Overall TC S/W performances > > are now comparable to the kernel openswitch datapath. > > Avoiding skb_clone() for redirection is cool, but why need to use > skb_do_redirect() here? > > There is a subtle difference here: > > skb_do_redirect() calls __bpf_rx_skb() which calls > dev_forward_skb(). > > while the current mirred action doesn't scrub packets when > redirecting to ingress (from egress). Although I forget if it is > intentionally. > > Also, skb->skb_iif is unset in skb_do_redirect() when > redirecting to ingress, I recall we have to set it correctly > for input routing. Probably yet another reason why we > can't scrub it, unless my memory goes wrong. :) Also dev_forward_skb() enforces MTU checks on the packet which are not done at the moment. I am aware of deployments where this would break things. Eyal.
