openvswitch / ovs-issues

Issue tracker repo for Open vSwitch
10 stars 3 forks source link

Dropped packets on Window Server 2k12 R2 in Openstack #166

Open neeseius opened 5 years ago

neeseius commented 5 years ago

We've had this problem for about a year and haven't been able to figure out the cause and are desperate.

Our 2 Active Directory Domain Controllers running 2k12 R2 have replication issues. They can't manage each other and replication is slow, they get out of sync easily. After many many tcpdumps we discovered that when the server tries to send a tcp-ack, it seems to get dropped/lost between the linux bridge and the vxlan_sys_4789 interface. So I am assuming the packets are being dropped in ovs bridge(s) (interface stats say they are). Setting net.bridge.bridge-nf-call-iptables=0 did not make a difference so I don't believe this is related to security groups, and the security group permits all traffic. Windows firewall is disabled.

We are running a very small Openstack Pike cluster on Centos7 with iptables_hybrid security groups and OpenVswitch utilizing DVR.

Text files with evidence and configs are attached. Any help is much appreciated! packetcapture.txt interface-stats.txt ovs-vsctl-show.txt neutron-cfgs.txt packages-and-lsmod.txt

neeseius commented 5 years ago

This was caused by the MTU being set to 1500 instead of 1450 in Windows. netns interface ipv4 show interface netns interface ipv4 set subinterface Ethernet mtu=1450 store=persistent It seems like the cloudbase-init script sets this every boot but we removed cloudbase-init because it was changing the server name when we didn't want it to change but that then lead to this new issue.

Anway, resolved.

alexpilotti commented 5 years ago

The MTU is set by cloudbase-init according to what your Neutron configuration sets (passed as DHCP option), so that's the place where the right MTU value should be set. Alternatively you can just disable the MTU plugin in cloudbase-init, not need to remove it entirely :-)