[dpdk-users] vhost, vxlan, and nic TSO trouble.

Matthias Gatto matthias.gatto at outscale.com
Mon Apr 25 16:24:50 CEST 2016

I'm working on a project using DPDK:

We are receiving packets from a vhost interface, then we encapsulate
them in vxlan packets, and finally we burst them into an ixgbe nic.

We set txq_flags of ixgbe nic to 0 in order to use offloading capability.

During encapsulation, we add vxlan header length to l2_len field inside mbuf.
It seems that l3_len, l4_len and tso_segsz are already set by vhost.

When we use iperf3 to test our architecture, it seems that:
- Packets with PKT_TX_TCP_CKSUM and without PKT_TX_TCP_SEG  have a
correct checksum.
- Packets with PKT_TX_TCP_CKSUM and with PKT_TX_TCP_SEG are segmented
but have a bad checksum.

Do you think I've missed something ?

Thanks for your help

-- Matthias Gatto

More information about the users mailing list