On 22/06/2017 16:11, Dave Barach (dbarach) wrote:
Folks,
I’m having a hard time trying to convince an Intel Fortville (i40e)
PMD / NIC to compute and insert TCP correct TCP checksums.
In addition to the typical struct rte_mbuf setup, I add the following:
if (b->flags & VLIB_BUFFER_TCP_CHECKSUM_OFFLOAD)
{
mb->packet_type =
RTE_PTYPE_L2_ETHER | RTE_PTYPE_INNER_L3_IPV4 |
RTE_PTYPE_INNER_L4_TCP;
mb->ol_flags = PKT_TX_TCP_CKSUM | PKT_TX_IPV4;
mb->l2_len = 14;
mb->l3_len = 20;
mb->outer_l2_len = 0;
mb->outer_l3_len = 0;
mb->l4_len = 0;
}
The inbound tcp->checksum is known to be zero.
The packets involved are the simplest case: 14-byte ethernet headers
w/ type 0x800, 20-byte ipv4 headers, TCP (SYN,ACK) packet with some
options.
Independent of port numbers / across multiple trials wireshark reports
a constant checksum error of 0xe31.
abs(received checksum - computed checksum) = 0xe31.
Does this ring any bells?
Hi Dave,
Assuming that you are not using vector functions but the default xmit
function with offload enabled, are you computing the pseudo-header checksum?
Also no need to setup PTYPE for TX.
Regards,
Sergio
Thanks… Dave
_______________________________________________
vpp-dev mailing list
vpp-dev@lists.fd.io
https://lists.fd.io/mailman/listinfo/vpp-dev
_______________________________________________
vpp-dev mailing list
vpp-dev@lists.fd.io
https://lists.fd.io/mailman/listinfo/vpp-dev
