Hi,
we’ve found a wrong behaviour with service_monitor record status for a
health-checked Load Balancer.
Its status can stay online forever even if virtual machine is stopped. This
leads to load balanced traffic been sent to a dead backend.
Below is the script to reproduce the issue because I doubt about the correct
place for a possible fix (my guess is it should be fixed in
controller/binding.c in function binding_lport_set_down, but I’m not sure how
this can affect VM live migration…):
# cat ./repro.sh
#!/bin/bash -x
ovn-nbctl ls-add ls1
ovn-nbctl lsp-add ls1 lsp1 -- \
lsp-set-addresses lsp1 "00:00:00:00:00:01 192.168.0.10"
ovn-nbctl lb-add lb1 192.168.0.100:80 192.168.0.10:80
ovn-nbctl set Load_balancer lb1 ip_port_mappings:192.168.0.10=lsp1:192.168.0.8
ovn-nbctl --id=@id create Load_Balancer_Health_Check vip='"192.168.0.100:80"'
-- set Load_Balancer lb1 health_check=@id
ovn-nbctl ls-lb-add ls1 lb1
ovs-vsctl add-port br-int test-lb -- set interface test-lb type=internal
external_ids:iface-id=lsp1
ip li set test-lb addr 00:00:00:00:00:01
ip a add 192.168.0.10/24 dev test-lb
ip li set test-lb up
# check service_monitor
ovn-sbctl list service_mon
# ensure state became offline
sleep 4
ovn-sbctl list service_mon
# start listen on :80 with netcat
ncat -k -l 192.168.0.10 80 &
# ensure state turned to online
sleep 4
ovn-sbctl list service_mon
# trigger binding release
ovs-vsctl remove interface test-lb external_ids iface-id
# ensure state remains online
sleep 10
ovn-sbctl list service_mon
# ensure OVS group and backend is still in bucket
ovs-ofctl dump-groups br-int | grep 192.168.0.10
————
Looking forward to hear any thoughts on this.
PS. don’t forget to kill ncat ;)
Regards,
Vladislav Odintsov
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
