datapath: Treat GSO skbs as if they were fragments

In dp_output_control() UDP GSO skbs are split into fragments which are
passed to userspace.  So the resulting flow set-up by the controller
(I am using ovs-vswitchd) is created based on a fragment.  This means
that the UDP source and destination port of the flow is zero.

In order for the datapath to match the resulting flow flow_extract() needs
to treat UDP GSO skbs as if they are fragments.  That is, set the UDP
source and destination port to 0.

A flow established for a UDP GSO skb with this change won't match any
subsequent non-GSO skbs, they will need to be passed to the controller and
a new flow established. But without this change no UDP GSO skbs will ever
match any flow.

I noticed this while using KVM using virtio with VhostNet and netperf's
UDP_STREAM test. The result was that the test sent ~5Gbit/s but only a
small fraction of that was received by the other side. Much less than the
1Gbit/s available on the physical link between the host (and guest) and the
machine running netserver. 100% of one of the host's CPUs was consumed, 50%
for the host and 50% for the guest.  The host consumption was contributed
to largely by ovs-vswitchd.

With this change I get a much nicer result of a fraction under 1Gbit/s sent
and almost all packets ending up at the other end.

Signed-off-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Jesse Gross <jesse@nicira.com>

diff --git a/datapath/flow.c b/datapath/flow.c
index b33b315..fb339e7 100644 (file)
--- a/datapath/flow.c
+++ b/datapath/flow.c
@@ -333,7 +333,8 @@ int flow_extract(struct sk_buff *skb, u16 in_port, struct odp_flow_key *key,
                key->nw_proto = nh->protocol;
 
                /* Transport layer. */
-               if (!(nh->frag_off & htons(IP_MF | IP_OFFSET))) {
+               if (!(nh->frag_off & htons(IP_MF | IP_OFFSET)) &&
+                   !(skb_shinfo(skb)->gso_type & SKB_GSO_UDP)) {
                        if (key->nw_proto == IPPROTO_TCP) {
                                if (tcphdr_ok(skb)) {
                                        struct tcphdr *tcp = tcp_hdr(skb);