Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2026-06-23
20:40:21 sean-k-mooney ya its a partial mesh using ovs
20:40:44 sean-k-mooney so all multicast becoms bradcast which si copied intrenlly t o all spokes in the mesh
20:40:51 fungi i think vxlan relies on multicast groups underneath to carry broadcasts though? and our cloud providers don't let the multicast traffic through
20:40:57 sean-k-mooney so its a littele inefficnet but it shoudl work
20:41:02 fungi ah okay
20:41:06 sean-k-mooney fungi: not with ovs
20:41:11 sean-k-mooney but with linux bridge yes
20:41:17 sean-k-mooney ovs only supprot unicast tunnels
20:41:28 fungi you are far more familiar with all of this than i am
20:41:46 sean-k-mooney but that actully makes it more portable since you do not rely on the cloud provdier supproting multicast grouping at the fabic level
20:42:21 fungi all i can really say is afaik we didn't intentionally omit support for sending ipv6 packets over the bridge, so if it's not working i'd consider that a bug worth fixing
20:42:25 sean-k-mooney it comes form addign ovs-dpdk suppreot neutron/nova and maintianing it for years:)
20:42:43 melwitt sean-k-mooney: oh yeah, that makes sense. in pre, I was doing it for hosts: all but didn't put under both
20:43:39 fungi `hosts: all` should run the same task on every test node (but not the zuul executor, which needs `hosts: localhost` instead, not really relevant to this playbook anyway)
20:45:57 sean-k-mooney fungi: i have also impletend my own version fo the role 2 before although the most recent version i got ai to do :)
20:46:28 sean-k-mooney fungi: https://github.com/SeanMooney/ard/tree/master/ansible/roles/ard_multinode_bridge does basically the same thign but it uses linux bridge and gre tunnels to avoid installign ovs
20:46:57 sean-k-mooney i used to just use the zuul jobs one
20:48:19 sean-k-mooney im not entrily sure im happy with how the new version works but i was tryign to make it compatiabe with our operator based isntaller and i dint want to fight with its installation and management of ovs
20:49:54 sean-k-mooney that fact that ovs survies reboot makes the ovs based approch we use upstream less complex in many ways
20:56:19 sean-k-mooney my the way the mtu caluation does have to be adjuscted for ipv6 but the role aslready subtratc 70 bytes which si enough to have ipv6 geneve trafic go over the bridge
20:56:23 sean-k-mooney melwitt: https://opendev.org/zuul/zuul-jobs/src/branch/master/roles/multi-node-bridge/tasks/common.yaml#L103-L139
20:58:00 melwitt hm ok
20:59:17 sean-k-mooney techinaly the 70 bytes of overad is the bare minium for geneve to work over the bridge
20:59:25 sean-k-mooney and it shoudl work for neutron in general
20:59:57 clarkb its not geneve its vxlan iirc
21:00:02 clarkb and that is why there isn't ipv6
21:00:04 sean-k-mooney i use a 100 byte safty buffer in my verions
21:00:07 fungi sean-k-mooney: yeah, we originally used gre i believe, but it isn't allowed in all out providers
21:00:22 clarkb there is/was work to convert it to geneve and drop the use of the ansible bridge module
21:00:24 sean-k-mooney clarkb: yes the overly is
21:00:32 fungi at least in some openstack clouds, the kernel conntrack plugin for gre isn't running
21:00:43 sean-k-mooney clarkb: actully that not the part the matters
21:00:59 sean-k-mooney we have other code in devstack playbooks to calulate teh neutron mtu config
21:01:13 sean-k-mooney that the one that need to accomidate geneve whwen ovn is used
21:01:15 clarkb https://review.opendev.org/c/zuul/zuul-jobs/+/959393
21:01:36 sean-k-mooney fungi: ya gre is not an ip protocol so its often blockded
21:01:51 clarkb yes the original implementation used gre and linux bridges
21:02:10 sean-k-mooney i depebated about gre or vxlan and for 8 bytpes of mtu vxlan is definly more portabl since its just udp form the network layer
21:02:15 clarkb that didn't work because we couldn't get gre to pass even with wide open security groups so we switched to vxlan as it is udp
21:02:39 sean-k-mooney clarkb: :) yep i remember
21:02:53 clarkb and now ansible has dropped supprot for much of the machinery that we relied on to setup vxlan and ovs so we're switching back to linux bridge and taking the opportunity to use geneve in that change
21:03:11 clarkb that work has clearly stalled out and probably should be picked up again. I think it is very close if not ready
21:04:33 clarkb and the reason we didn't support ipv6 is taht netiher vxlan nor gre could support it iirc
21:04:41 clarkb or maybe it was just vxlan that couldn't
21:04:47 sean-k-mooney both can
21:04:58 clarkb I think vxlan can now
21:05:00 sean-k-mooney at both the underlay and over the mesh
21:05:01 clarkb but it was a v4 only protocol
21:05:13 sean-k-mooney it dependied on the kernel version
21:05:19 sean-k-mooney and ovs version
21:05:24 clarkb in any case if you want to make changes like this getting 959393 in first is probably a prereq
21:05:36 sean-k-mooney vxlan suppoted it but we could supprot it on all distos we supproted if i recll
21:05:58 sean-k-mooney well it does not stricly ned the refacto to linux bridge
21:06:06 sean-k-mooney but that cerntly viable
21:06:09 clarkb it does not but I will nto review a chnge that creatse a merge conflict
21:06:17 clarkb maybe other people are willing to review it and land it
21:06:25 opendevreview melanie witt proposed openstack/nova master: Add non-voting IPv6 ceph live migration job https://review.opendev.org/c/openstack/nova/+/984623
21:06:28 sean-k-mooney that is one of my version of that https://github.com/SeanMooney/ard/tree/master/ansible/roles/ard_multinode_bridge
21:06:39 sean-k-mooney i have a much older version too somehwere
21:06:42 clarkb the current code only works because I vendored the entirety of the ansible module that is no longer supported into that role
21:07:05 clarkb we really need to stop using that stuff and leaving it as is is fine. But if you want to make changes then we should do the right thing to fix the underlying ansible stuff
21:07:05 sean-k-mooney ya i replaced all of tha twith a systemd unit file that calls a bashcript
21:07:13 sean-k-mooney https://github.com/SeanMooney/ard/blob/master/ansible/roles/ard_multinode_bridge/templates/ard-bridge-setup.sh.j2
21:07:24 clarkb and that work is basically done so there isn't any reason not to
21:08:27 sean-k-mooney clarkb: ya https://review.opendev.org/c/zuul/zuul-jobs/+/959393 cerntely cleans up a lot
21:08:59 clarkb rereading things it stalled beacuse people using ovs as a hard dependency weren't properly installing/settting up that dep and relying on that role instead
21:09:08 clarkb I think we can basically say sorry you need to fix that and move forward
21:09:23 clarkb which is also what I basically said back when I was trying to move this forward
21:09:32 sean-k-mooney ya. is ther a reason it swapt to geneve?
21:09:56 sean-k-mooney i know it says "This is a more modern and flexible approach than the
21:09:58 sean-k-mooney previous vxlan implementation."
21:09:58 clarkb I think because geneve is the future and more flexible
21:10:08 clarkb right vxlan had a bunch of gotchas like the ipv6 support or lack of it
21:10:14 clarkb geneve exists to solve all those problems with a new protocol
21:10:21 clarkb so we should use it since those problems hit us previously
21:10:29 sean-k-mooney the main one that geneve solve is actully fragmentation
21:10:37 sean-k-mooney vxlan explity does not allow fragmentation
21:10:51 sean-k-mooney of the vxlan packet
21:10:58 sean-k-mooney geneve i belvie does
21:11:21 sean-k-mooney which makes it more flexible acorss l3 subnets
21:12:15 sean-k-mooney clarkb: but i guess i agree building on that patch woudl be good for the reason above
21:12:21 sean-k-mooney anyway dinner is ready o/
21:13:02 clarkb ya I think the next step is to followup with ykarel to do what we can to properly capture ovs dependencies then proceed with the switch in zuul-jobs then add ipv6 support
21:13:09 clarkb probably don't even need to make ipv6 support optional just add it
21:14:05 clarkb ykarel: can you followup on https://review.opendev.org/c/zuul/zuul-jobs/+/959393 and indicate if properly managing jobs' dependency on ovs si a possibility? That really shouldn't be something we rely on a side effect from a base job role to handle
22:00:05 opendevreview melanie witt proposed openstack/nova master: Add non-voting IPv6 ceph live migration job https://review.opendev.org/c/openstack/nova/+/984623
22:51:31 opendevreview melanie witt proposed openstack/nova master: Add non-voting IPv6 ceph live migration job https://review.opendev.org/c/openstack/nova/+/984623
23:52:14 opendevreview melanie witt proposed openstack/nova master: Add non-voting IPv6 ceph live migration job https://review.opendev.org/c/openstack/nova/+/984623
#openstack-nova - 2026-06-24
01:38:38 opendevreview sean mooney proposed openstack/nova master: doc: Add guide for NVIDIA vGPU with SR-IOV variant drivers https://review.opendev.org/c/openstack/nova/+/994160
01:38:39 opendevreview sean mooney proposed openstack/nova master: libvirt: Support Cyborg PCI ARQ managed mode https://review.opendev.org/c/openstack/nova/+/994579
02:23:15 opendevreview melanie witt proposed openstack/nova master: Add non-voting IPv6 ceph live migration job https://review.opendev.org/c/openstack/nova/+/984623
03:07:32 opendevreview Takashi Kajinami proposed openstack/nova master: compute: Validate SEV/SEV-ES support at start up https://review.opendev.org/c/openstack/nova/+/994342
03:07:59 opendevreview Takashi Kajinami proposed openstack/nova master: libvirt: Ignore SEV-ES when SEV-SNP is detected https://review.opendev.org/c/openstack/nova/+/994343
03:11:31 opendevreview Ghanshyam Maan proposed openstack/placement master: Update policy scope enforcement testing https://review.opendev.org/c/openstack/placement/+/987568
03:16:47 opendevreview Ghanshyam Maan proposed openstack/nova master: Update policy scope enforcement testing https://review.opendev.org/c/openstack/nova/+/986946
03:17:13 opendevreview Ghanshyam Maan proposed openstack/placement master: Update policy scope enforcement testing https://review.opendev.org/c/openstack/placement/+/987568
03:17:18 opendevreview Takashi Kajinami proposed openstack/nova master: libvirt: Ignore SEV-ES when SEV-SNP is detected https://review.opendev.org/c/openstack/nova/+/994343
04:25:02 ykarel clarkb, thx for reminding that, i had completely forgot about that :)
04:25:53 ykarel Will check on this in next few days and see how it can be unblocked
04:26:15 ykarel will start with the impacted job and see

Earlier   Later