A security group rule has no effect. Traffic it allows is still blocked, or traffic it should block still passes¶
Symptom¶
A rule was added to a security group, and the traffic it allows still does not reach the instance (a port stays closed, ping fails). Or a rule was removed, or never existed, and the traffic still passes. The instance itself is up, and other traffic works.
Impact¶
The instances whose ports use that security group. Traffic that should pass is cut, or traffic that should be filtered reaches the instance: a security exposure when filtering is expected.
Quick checks¶
Read-only. With credentials for the instance's project (or admin):
# 1. The instance's ports, their security groups, and whether port security is on
openstack port list --server <server-id> -c ID -c "Fixed IP Addresses"
openstack port show <port-id> -c security_group_ids -c port_security_enabled -c allowed_address_pairs
# 2. The rules of each of those groups
openstack security group rule list <security-group-id>
# 3. The agents on the instance's compute (ML2/OVS: the Open vSwitch agent; ML2/OVN: the OVN controller)
openstack network agent list --host <compute>
Read each rule against the traffic: direction (ingress for traffic to the instance), ethertype (IPv4 or
IPv6), protocol and port range, and the remote side (remote_ip_prefix, or remote_group_id: then only traffic from
ports in that group matches).
Diagnosis¶
How Neutron applies security groups (Neutron 2026.1 networking guide)¶
- A security group is a set of allow rules on a port, default deny: traffic matches if any rule of any of the port's groups allows it. Groups add up.
- Rules are stateful: allowing ingress TCP 22 also allows the return traffic.
- The
defaultgroup of a project allows all egress, and allows ingress only from ports indefaultitself (IPv4 and IPv6, every protocol and port): Neutron creates it from its default security group rule templates, two of which haveremote_group_idset to the group itself (neutron/db/securitygroups_db.py). So instances that sharedefaultreach each other on every port. An admin can change these templates (openstack default security group rule list). A port created without a group getsdefault. - Neutron also adds basic and anti-spoofing rules: a port can only send from its own MAC and IP addresses, or those in
its
allowed_address_pairs. - With
port_security_enabledset toFalseon a port, none of this applies: no security group, no anti-spoofing, everything passes.
Where the rules are enforced¶
- ML2/OVS with Kolla-Ansible 2026.1:
openvswitch_agent.inisetsfirewall_driver = neutron.agent.linux.iptables_firewall.OVSHybridIptablesFirewallDriver. Each port gets a Linux bridge (qbr…) between the instance andbr-int, and iptables chains on the compute, named after the port ID (its first 10 characters):neutron-openvswi-i<id>for ingress,…-o<id>for egress,…-s<id>for anti-spoofing; unmatched traffic goes tosg-fallback, which drops it (neutron/agent/linux/iptables_firewall.py). The Open vSwitch agent writes them; while it is dead, rule changes are not applied on that compute. - ML2/OVS with the native driver (
firewall_driver = openvswitch): OpenFlow rules inbr-int. The networking guide lists two differences with iptables: traffic that conntrack marks invalid is blocked, and multicast to224.0.0.Xis always allowed. - ML2/OVN: each security group is an OVN port group named
pg_<security group ID, with _ instead of ->with ACLs; ports with port security are also inneutron_pg_drop, which drops everything not allowed (neutron/common/ovn/utils.py).ovn-northdandovn-controlleron the compute turn them into flows; if either is down, changes are not applied.
To see what is enforced:
Kolla-Ansible, ML2/OVS: on the compute, inside the agent's container, as root (the container's default user gets
Could not fetch rule set generation id: Permission denied (you must be root); a host using another iptables
backend may not show what the container wrote),
sudo docker exec -u root neutron_openvswitch_agent iptables-save | grep -i <first-10-characters-of-port-id>
sudo docker exec -u root neutron_openvswitch_agent ip6tables-save | grep -i <first-10-characters-of-port-id>
On an Ubuntu 24.04 host, where the host and the container both use iptables v1.8.10 (nf_tables), sudo
iptables-save on the host showed the same chains.
Kolla-Ansible, ML2/OVN: on a host running the OVN northbound database,
sudo docker exec ovn_nb_db ovn-nbctl --no-leader-only acl-list pg_<security_group_id_with_underscores>
Any: the same commands without the container prefix.
Possible causes¶
Cause 1: the rule does not match the traffic, or is on another group¶
The rule is in a group the port does not use; or its direction, ethertype, protocol, port range or remote side does
not match (a common one: remote_group_id set, while the traffic comes from outside that group; or an IPv4 rule for
IPv6 traffic).
Cause 2: port security is off on the port¶
port_security_enabled is False: no rule applies, and everything passes. Removing a rule changes nothing.
Cause 3: the change did not reach the compute¶
ML2/OVS: the Open vSwitch agent on the compute is dead or lagging (quick check 3), so the iptables chains still hold
the old rules. ML2/OVN: ovn-northd or the compute's ovn-controller is down, so the ACLs are not turned into flows.
Cause 4: the traffic is filtered elsewhere¶
The guest's own firewall, a router or its gateway, a floating IP that is not associated, or a physical firewall: the security group allows it, something else drops it (unreachable from outside).
Cause 5: an allowed address pair or a second port lets it through¶
Traffic from an address listed in allowed_address_pairs, or arriving on another port of the instance with other
groups, is allowed by design.
Cause 6: both instances are in the default group¶
Traffic that should be blocked passes between two instances of the project because both ports use default, which
allows all ingress from its own members (see How Neutron applies security groups). Quick checks 1 and 2 show it: an
ingress rule with default itself as remote group (its own ID in the Remote Security Group column).
Resolution¶
For cause 1: add the rule the traffic needs, to the group the port uses¶
⚠️ State-changing. Changes what reaches every port that uses this group, in every instance of the project that has it. Undo:
openstack security group rule delete <rule-id>.
openstack security group rule create --ingress --ethertype IPv4 --protocol tcp --dst-port <port> \
--remote-ip <cidr> <security-group-id>
Or add a group to the port (openstack port set --security-group <group> <port-id>, same warning).
For cause 2: turn port security back on, if it should be on¶
⚠️ State-changing. Filtering starts on that port: traffic not allowed by its groups, or sent from addresses that are not the port's own, is dropped at once. Check the groups first, and the allowed address pairs if the instance uses other addresses (a VRRP address, a router guest). Undo:
openstack port set --no-security-group --disable-port-security <port-id>.
openstack port set --enable-port-security --security-group <group> <port-id>
For cause 3: bring the agent back¶
⚠️ State-changing. Restarting an agent touches every port on that host; read the warnings in the page linked below first.
Start the Open vSwitch agent or ovn-controller on the compute, or ovn-northd
(port binding failed, cause 1, for the restarts and their warnings). On Kolla-Ansible
2026.1, ML2/OVS: sudo systemctl start kolla-neutron_openvswitch_agent-container.service. Once it is back, it
applies the current rules (Agent has just been revived. Doing a full sync. in neutron-openvswitch-agent.log; on
the lab the pending rule applied 18 s after the start); check with the commands under Where the rules are
enforced.
For causes 4 and 5¶
Fix the filter that drops the traffic, or the pair or port that lets it through.
For cause 6: give the instances their own group¶
⚠️ State-changing. The instance loses the access
defaultgave it from its peers, and any rule ofdefaultit relied on. Check the new group's rules first. Undo:openstack port set --no-security-group --security-group default <port-id>.
openstack security group create <group>
openstack security group rule create --ingress --ethertype IPv4 --protocol tcp --dst-port <port> \
--remote-ip <cidr> <group>
openstack port set --no-security-group --security-group <group> <port-id>
Or, as admin, change the templates for groups created from now on (openstack default security group rule list,
create, delete); existing default groups keep their rules.
Verification¶
From a host allowed by the rule, the traffic passes; from a host not allowed, it does not (for example nc -zv
<instance-address> <port> from both). On the compute, the port's chains (ML2/OVS) or the port group's ACLs (ML2/OVN)
show the rule.
Lab observations¶
Kolla-Ansible 22.2.0 (OpenStack 2026.1) all-in-one, ML2/OVS with OVSHybridIptablesFirewallDriver, 2026-10-27. A
test instance in a group allowing TCP 22 only, a listener on port 8080 in the guest, tests with nc -zv from the
network's DHCP namespace:
- Symptom:
nc: connect to <instance-address> port 8080 (tcp) timed out: Operation now in progress, port 22 succeeded. The port's chainsneutron-openvswi-i<id>,-o<id>,-s<id>ended withSend unmatched traffic to the fallback chain.(-j neutron-openvswi-sg-fallback). - Cause 1: after
security group rule create ... --dst-port 8080 --remote-ip 0.0.0.0/0, the agent loggedRefresh firewall rules,--dport 8080 -j RETURNappeared in the port's ingress chain, and port 8080 passed 1 s after the command returned. Deleting the rule blocked it again within 3 s. - Cause 2:
port set --no-security-group --disable-port-securitylet port 8080 through within 5 s with no rule;port set --enable-port-security --security-group <group>blocked it again within 3 s. - Cause 3: with the agent stopped (
systemctl stop kolla-neutron_openvswitch_agent-container.service, which took 61 s and ended with exit 137), the agent showedAliveXXXand a new 8080 rule had no effect and no iptables line. Aftersystemctl start, the rule applied 18 s later. - Cause 6, read-only:
openstack default security group rule listshowed the two ingress templates with remotePARENT; the project'sdefaultgroup's ingress rules carried its own ID as remote group. - Not run: causes 4 and 5, the resolution of cause 6, ML2/OVN and the native OVS driver, the conntrack behaviour on rule removal.
Validity and limits¶
- Written from Neutron 2026.1 (networking guide on security groups, Open vSwitch native firewall guide,
neutron/agent/linux/iptables_firewall.py,neutron/common/ovn/utils.py), Kolla-Ansible 2026.1 (openvswitch_agent.ini.j2), python-openstackclient 2026.1, and the default security group rule templates (neutron/db/securitygroups_db.py, migrationc33da356b165). Causes 1, 2 and 3 reproduced and resolved on a Kolla-Ansible 2026.1 all-in-one with ML2/OVS and the iptables hybrid driver (Lab observations). Not reproduced on a lab: ML2/OVN and the native OVS driver (theovn-nbctl acl-listcheck,ovn-northd/ovn-controllerfor cause 3), causes 4 and 5, the resolution of cause 6. - Connections already established: when a rule is removed, the iptables driver deletes the matching conntrack
entries (
delete_conntrack_state_by_ruleiniptables_firewall.py); with other drivers, notably OVN's stateful ACLs, a flow conntrack already tracks may keep passing. Not verified on a lab. - Not covered: stateless security groups, security group logging, FWaaS.
Related¶
- Instance has its address and metadata, but is unreachable from outside
- New instance boots without metadata: removing the default egress rules blocks the metadata service.
- Port binding failed
- Neutron 2026.1: security groups, native OVS firewall
Provenance: written by agent:ozarland-author, from the Neutron, Kolla-Ansible and python-openstackclient 2026.1 documentation and source. Reviewed in its pull request (see the page history). Checked by a human operator: not yet.
Worked for you, or not? Tell us in one line: your release, your deployment method, and what happened.