Manuel,
A minor correction on your last statement:
Currently the last SFF is NOT popping the NSH header. Instead, we
send the packet to Netvirt with the NSH header. Netvirt processes
the packet and strips the NSH header. Currently, if we dont do it
this way, and the classifier is on the same bridge as the SFF, the
packet will get classified again, and we'll have a nasty loop.
Back to the original discussion:
When the packet enters the SFF from the SF on the vxgpe port, OVS stores
the vxlan info in what I believe is called the packet metadata (not to
be confused with the openflow metadata). I suspect that when the packet
is then egressed out the vxlan port, it is using the some of the same
packet metadata info as when it came in the vxgep port. If this is the
case, could it be that its using the same tp_dst port of 6633? We may
have to make a change to netvirt to explicitly set the tp_dst port to
4789 to overwrite the previous value.
Regards,
Brady
On 23/11/16 10:48, Manuel Buil wrote:
Hi Georgios,
Don’t be sorry, it is great to see more people participating in the
discussion J.
The last SF should not egress the packet to another compute, it should
egress the packet to the SFF. Then, if it is the last SFF, it pops the
NSH header and hands over the packet to the Netvirt pipeline. That
pipeline is not aware of a vxlan-gpe port.
Regards,
Manuel
*From:* Paraskevopoulos Georgios [mailto:[email protected]]
*Sent:* Wednesday, November 23, 2016 10:16 AM
*To:* Manuel Buil <[email protected]>; Yang, Yi Y
<[email protected]>
*Cc:* [email protected]
*Subject:* RE: Potential bug in NSH patch for OVS
Hi,
Sorry to step in. Just a remark.
It seems to me that the problem is that when the last SF egresses the
packet to another compute, it incorrectly does it through vxlan_gpe,
rather than vxlan.
So it might be that the problem is not that some value is hardcoded,
rather that OVS does an extra NSH pop on the output of the last SF.
Cheers,
George
*From:* Manuel Buil [mailto:[email protected]]
*Sent:* Wednesday, November 23, 2016 10:34 AM
*To:* Yang, Yi Y
*Cc:* [email protected]
<mailto:[email protected]>; Georgios Paraskevopoulos
([email protected] <mailto:[email protected]>)
*Subject:* RE: Potential bug in NSH patch for OVS
Hi Yi Yang,
I think we have different views here. The packet was already processed
by the SFC pipeline and now it is again in the Netvirt pipeline. Why
should then the packet be sent using the vxlan-gpe interface from the
compute? The packet should use the vxlan interface of Netvirt which is
what actually happens. The problem is that OVS confuses the port and
sends it with port 6633 instead of 4789.
/Manuel
*From:* Yang, Yi Y [mailto:[email protected]]
*Sent:* Wednesday, November 23, 2016 1:21 AM
*To:* Manuel Buil <[email protected]
<mailto:[email protected]>>
*Cc:* [email protected]
<mailto:[email protected]>; Georgios Paraskevopoulos
([email protected] <mailto:[email protected]>)
<[email protected] <mailto:[email protected]>>
*Subject:* RE: Potential bug in NSH patch for OVS
I think that is NetVirt classifier issue, in sfc103 demo, we have on
SF use case. In your case, you must have two classifiers, one is in
compute1, another is in compute2, for compute1, I think you have
enabled NetVirt to send back the packet to the client, but for
compute2, NetVirt can’t support this, right? Classifier2 must have
VxLAN-gpe port for this in compute2.
*From:* Manuel Buil [mailto:[email protected]]
*Sent:* Wednesday, November 23, 2016 1:12 AM
*To:* Yang, Yi Y <[email protected] <mailto:[email protected]>>
*Cc:* [email protected]
<mailto:[email protected]>; Georgios Paraskevopoulos
([email protected] <mailto:[email protected]>)
<[email protected] <mailto:[email protected]>>
*Subject:* Potential bug in NSH patch for OVS
Hi Yi Yang,
I am observing that sometimes our SFC OPNFV test case fails. I am
trying to investigate why and I think I found the problem.
### SET-UP ###
The environment is the following: the http traffic from client gets
classified and assigned to a chain with one SF. After the traffic
visits that SF, the SFC encapsulation is removed and the packet is
sent to the netvirt pipeline. The scenario that fails is the
following: client and SF are in compute1, the server is in compute2.
### PROBLEM ###
All the SFC process happens in one OVS because both classifier and SF
are in that OVS. Then the packet leaves SFC pipeline and gets into
Netvirt pipeline which egress the packet using the vxlan tunnel it has
to connect to the compute2. The problem is that instead of using udp
port 4789 as it should, it uses udp port 6633. Compute2 has no SFC
service so there is no vxlan port listening on udp port 6633 and thus
the packet is dropped.
I captured the packet using ovs-dpctl and everything is correct except
the tp_dst which is 6633 even though it should be 4789:
recirc_id(0),nsh(nsi=254,nsp=3c,nsh_mdtype=1,nsh_np=3,nshc1=c0a80006,nshc2=12),tunnel(tun_id=0x3c,src=11.0.0.6,dst=10.20.0.11,vxlan(gpe(np=0x3c)),flags(-df+csum+key)),in_port(8),skb_mark(0),eth(src=fa:16:3e:6b:08:c7,dst=fa:16:3e:4a:be:0e),eth_type(0x0800),ipv4(dst=11.0.0.4/255.255.255.254,tos=0/0x3,frag=no),
packets:0, bytes:0, used:never,
actions:set(tunnel(tun_id=0x12,src=192.168.2.4,dst=192.168.2.3,ttl=64,tp_src=49676,tp_dst=6633,vxlan(gpe(np=0x4,flags=0)),flags(df|csum|key))),pop_nsh,4
These are the ports (you can see that it egresses on 4 è vxlan_sys_4789):
port 0: ovs-system (internal)
port 1: br-ex (internal)
port 2: p_aec4d661-0 (internal)
port 3: br-int (internal)
port 4: vxlan_sys_4789 (vxlan)
port 5: tapbbd70ea7-7b
port 6: tapd5cbb645-01
port 7: tap3f9a7479-51
port 8: vxlan_sys_6633 (vxlan)
If I sniff the packets, I can see that all the traffic coming from
compute1 and going to compute2 is sent through vxlan and udp port
4789, except for the packets that traversed SFC, which are sent with
port 6633:
16:30:45.508076 IP 192.168.2.4.60213 > *192.168.2.3.4789*: VXLAN,
flags [I] (0x08), vni 0
LLDP, length 99: openflow:119801980855940
16:30:45.508288 IP 192.168.2.3.36551 > *192.168.2.4.4789*: VXLAN,
flags [I] (0x08), vni 0
LLDP, length 97: openflow:93452943833469
16:30:46.356412 IP 192.168.2.4.52024 > *192.168.2.3.6633*: VXLAN,
flags [I] (0x08), vni 18
IP 11.0.0.4.38652 > 11.0.0.5.http: Flags [S], seq 340357625, win
28200, options [mss 1410,sackOK,TS val 561484 ecr 0,nop,wscale 7],
length 0
16:30:46.356679 IP 192.168.2.3 > 192.168.2.4: ICMP 192.168.2.3 udp
port 6633 unreachable, length 118
16:30:47.354970 IP 192.168.2.4.52024 > *192.168.2.3.6633*: VXLAN,
flags [I] (0x08), vni 18
IP 11.0.0.4.38652 > 11.0.0.5.http: Flags [S], seq 340357625, win
28200, options [mss 1410,sackOK,TS val 561734 ecr 0,nop,wscale 7],
length 0
16:31:20.508294 IP 192.168.2.3.36551 > *192.168.2.4.4789*: VXLAN,
flags [I] (0x08), vni 0
LLDP, length 97: openflow:93452943833469
16:31:20.509431 IP 192.168.2.4.60213 > *192.168.2.3.4789*: VXLAN,
flags [I] (0x08), vni 0
LLDP, length 99: openflow:119801980855940
16:31:25.507641 IP 192.168.2.3.36551 > *192.168.2.4.4789*: VXLAN,
flags [I] (0x08), vni 0
LLDP, length 97: openflow:93452943833469
16:31:25.507728 IP 192.168.2.4.60213 > *192.168.2.3.4789:* VXLAN,
flags [I] (0x08), vni 0
LLDP, length 99: openflow:119801980855940
Can you investigate if this problem might be related to your NSH patch
for OVS? Perhaps packets that traverse the SFC pipeline have L4 port
destination hardcoded as 6633?
Thanks,
Manuel
_______________________________________________
sfc-dev mailing list
[email protected]
https://lists.opendaylight.org/mailman/listinfo/sfc-dev
_______________________________________________
sfc-dev mailing list
[email protected]
https://lists.opendaylight.org/mailman/listinfo/sfc-dev
