Hi Carl, Indeed, there is no way in AF_XDP to detect whether the checksum is offloaded or not. We could always add a flag but I am not sure it would worth it. Can you try host-interface (AF_PACKET) and see if it solves your issue: https://s3-docs.fd.io/vpp/22.06/cli-reference/clis/clicmd_src_vnet_devices_af_packet.html With veth, AF_XDP and AF_PACKET should have similar performance anyway.
Best ben > -----Original Message----- > From: vpp-dev@lists.fd.io <vpp-dev@lists.fd.io> On Behalf Of Carl Lei > Sent: mercredi 9 mars 2022 02:56 > To: vpp-dev@lists.fd.io > Subject: [vpp-dev] Question on checksum offload in veth+AF_XDP > > Hello list, > > I configured an address in the kernel on one side of veth and attached > the other side to VPP using AF_XDP, then I found that kernel-sent > packets into this veth pair would be forwarded by VPP without fixing > checksums. Kernel seems to believe the veth has checksum offload > features, well apparently it has not; the receiver then discards all > TCP/UDP packets, but ICMP packets are accepted. This can be fixed by > `ethtool -K interface tx off` on the kernel side, but it is quite slow > (~10Gbps); is this expected on AF_XDP with veth, or is there a toggle > to fixup checksums on an AF_XDP interface, or should I use some other > interface to connect to my running kernel? > > I have tried `create tap`, but using vpp 22.02-release package for > ubuntu focal running kernel 5.14.0-1024-oem, in rx-mode=adaptive it > misses quite some packets (~3% packet loss in ping tests). Using the > same 22.02-release in ubuntu focal userspace but running kernel > 5.16.12-arch1-1 (this is a ubuntu chroot on arch linux) it works > fine. Maybe something was fixed between the two kernels, but I want > to run ubuntu kernels without busy polling, so unfortunately I cannot > use tap for now.
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#20988): https://lists.fd.io/g/vpp-dev/message/20988 Mute This Topic: https://lists.fd.io/mt/89653493/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
