On Wed, Aug 24, 2016 at 5:33 AM, Yuval Mintz [off-list ref] wrote:

quoted

This patch assumes that the bnx2x hardware will ignore existing IPv4/v6 header
fields for length and checksum as well as the length and checksum fields for
outer UDP and GRE headers.

I have no means of testing this as I do not have any bnx2x hardware but thought
I would submit it as an RFC to see if anyone out there wants to test this and see
if this does in fact enable this functionality allowing us to to segment tunneled
frames that have an outer checksum.

Signed-off-by: Alexander Duyck <redacted>

So it took me some [well, a lot] time to reach this, but I've finally  gave it a try.
I saw a performance boost with the partial support -
Throughput for vxlan tunnels with and without udpcsum were almost identical
after this, whereas without this patch the udpcsum prevented GSO and
a TCP/IPv4 connection on top of it got roughly half the throughput.

However, I did encounter one oddity I couldn't explain -
After I've disabled tx-udp_tnl-segmentation via ethtool on the base interface,
got left with:
   tx-gso-partial: on
   tx-udp_tnl-segmentation: off
   tx-udp_tnl-csum-segmentation: on

When I ran traffic over both vxlan tunnels the one with the udpcsum was still
Passing gso aggregations to base device to transmit [and the throughput was
same as before], where's the tunnel without the udpcsum showed only
MTU-sized packets reaching the base interface for transmission [which is what
I've expected]

Any idea why that happened?

So the way they are implemented tx-udp_tnl-segmentation and
tx-udp_tnl-csum-segmentation are treated as two separate features.
The kernel currently gives them the same treatment as NETIF_F_TSO and
NETIF_F_TSO6.  You can disable one and the other still functions.

Now if you disable tx-gso-partial you should expect to see
tx-udp_tnl-csum-segmentation be disabled because it is dependent on
the partial GSO offload.

- Alex