On Sat, 2020-01-18 at 09:08 -0500, Ryan Goodfellow wrote: > On Sat, Jan 18, 2020 at 11:14:05AM +0100, Jesper Dangaard > Brouerwrote: > > > > I'm wondering why did you choose/need AF_XDP technology for doing > > forwarding? > > This is just a sample program used to demonstrate moving packets > between > different interfaces efficiently using AF_XDP. > > Our actual use case is performing network emulation in userspace. For > example, > representing impaired links or entire networks with link-by-link > shaping > specifications. We are using AF_XDP to get packets to/from our > network emulation > software as quickly as possible without having to go through the > entire network > stack, as the emulation host's network configuration does not > influence the > networks it's emulating. > > Traditionally we've used DPDK for this, but are porting to AF_XDP for > the > relative simplicity and flexibility it provides. Some specific > benefits for us > are: > > - Can attach to VTEPs which allows us to hook into some EVPN/VXLAN > based > networks we have easily. Alternatively with the BPF program > flexibility, we > also have the option to split out BGP control plane traffic from > overlay > traffic when attaching to the physical interface and pass it > through to the > kernel. Both of these approaches let the kernel manage the FDB for > VTEPs as > well as taking care of encap/decap (potentially offloaded to the > NIC itself) > and let our software focus on emulation. > > - Using XDP in virtual machines in our testing environment is > straightforward, > while this is possible with DPDK and virtio, the setup was rather > convoluted. I'm in a very similar situation. The per-packet work we do is complicated. Doing it all in BPF would be much more painful if even possible. For us, AF_XDP is a nice, simple DPDK. We have no need for BPF in the AF_XDP path.
|