Instance has its address and metadata, but is unreachable from outside¶
Symptom¶
The instance is ACTIVE. It got its fixed address and its metadata: the console log shows the SSH key and the
hostname set by cloud-init. From inside the tenant network, another instance reaches it. From outside the cloud,
nothing answers: no ping, no SSH, a connection timeout rather than a refusal.
$ ssh <user>@<floating-ip>
ssh: connect to host <floating-ip> port 22: Connection timed out
A client on the external network's own segment may see No route to host instead: when no router holds the
floating IP, nothing answers ARP for it. A client farther away sees a timeout.
If the instance never got an address, or got one but no metadata, see new instance gets no IP or new instance boots without metadata instead.
Impact¶
Usually one instance or one project: its users cannot reach it. When every floating IP of the cloud stops answering at once, the cause is shared (the router's gateway node, the external network) and the impact is platform-wide; check L3 agent dead, router not moved first on ML2/OVS.
Quick checks¶
With the project's credentials (or admin), from any host with the OpenStack client:
# 1. The instance's port, its fixed address, its security groups and port security
openstack port list --server <server-id> -c ID -c "Fixed IP Addresses"
openstack port show <port-id> -c fixed_ips -c security_group_ids -c port_security_enabled
# 2. Ingress rules of each security group on the port
openstack security group rule list --ingress <security-group-id>
# 3. A floating IP on this port?
openstack floating ip list --port <port-id>
# 4. The router of the tenant network: its gateway and its interfaces
openstack port list --network <tenant-network-id> --device-owner network:router_interface -f value -c ID
# HA or distributed routers use other device owners: list the network's ports without the filter if this is empty
openstack port show <router-interface-port-id> -c device_id # device_id is the router ID
openstack router show <router-id> -c external_gateway_info -c interfaces_info -c status
# 5. The external network is really external
openstack network show <external-network-id> -c router:external -c status
Each check rules out one cause. Go through them in order: most cases end at check 2 or 3.
Diagnosis¶
Security groups¶
Neutron's security groups use a default-deny policy and only contain rules that allow traffic. The default group
of each project allows all egress and denies ingress, and an instance started without a security group gets
default (Neutron's networking introduction for 2026.1). Check 2 shows the group's actual rules: any ingress rule
there that names a remote security group instead of a prefix admits only members of that group, which explains why
neighbours reach the instance and the outside does not.
In the output of check 2, look for an ingress rule that matches the traffic you test: the protocol (tcp for SSH,
icmp for ping), the port range, and a remote prefix that includes the client's address. A rule with a remote
group instead of a prefix only admits members of that group.
port_security_enabled: False disables security groups on the port altogether; it is then not the cause.
Floating IP and router¶
- No floating IP on the port (check 3 empty): the instance has only its tenant address, which is not routed from outside. Normal behaviour, not a fault.
- A floating IP exists but is associated with another port, or with another fixed address of a port with several
(
Fixed IP Addresscolumn): traffic goes elsewhere. - No router interface on the tenant network (check 4 empty), or a router without
external_gateway_info: nothing routes between the external network and the tenant subnet. Through the API, this cannot coexist with an associated floating IP: Neutron refuses the association (External network <network-id> is not reachable from subnet <subnet-id>), and refuses with HTTP 409 to remove the router interface or the gateway while a floating IP depends on them (observed on 2026.1 with ML2/OVS). This cause therefore applies when no floating IP is associated yet. An associated floating IP with no working path points at the router on its node (cause 4).
Where the router lives (only when the API records all look right)¶
ML2/OVS with the L3 agent: the router is a qrouter-<router-id> namespace on the network node the L3 agent
scheduled it on.
openstack network agent list --router <router-id> --long # which agent, alive, HA state
With L3 HA routers the namespace exists on several nodes, and only the one whose HA State is active carries the
floating IP. Check that node.
On that node, as root (without it, entering the namespace fails with Operation not permitted):
Kolla-Ansible:
sudo ip netns exec qrouter-<router-id> ip -brief addr # the qg- interface carries the floating IP
sudo ip netns exec qrouter-<router-id> iptables -t nat -S | grep -F "<floating-ip>"
Any: the same commands on the network node that runs neutron-l3-agent.
The floating IP must appear on the qg- interface and in a DNAT rule towards the fixed address. With DVR, floating
IPs live in the fip- namespace of the compute instead; DVR is not covered here.
ML2/OVN (not run on the lab): the router is a logical router neutron-<router-id> in the OVN northbound database, and the floating
IP a dnat_and_snat NAT entry on it. North-south traffic goes through the router's active gateway chassis (Neutron
OVN routing guide).
Kolla-Ansible:
docker exec ovn_northd ovn-nbctl lr-nat-list neutron-<router-id> # a dnat_and_snat line for the floating IP
Any: ovn-nbctl lr-nat-list neutron-<router-id> wherever ovn-nbctl reaches the northbound database.
Not yet checked on a lab: on a multinode or clustered northbound database, ovn-nbctl may need
--db=<nb-connection> to find it from that container.
Inside the guest¶
If everything above is right, the guest itself may drop the traffic: a host firewall (ufw, firewalld,
nftables) or a service not listening. Look at it from the console (openstack console url show <server-id>) or
from a neighbour instance in the tenant network: if a neighbour cannot connect either, while the security groups
allow it, the guest is the cause.
The external network¶
If several projects lose their floating IPs at once, check the external network itself: the provider VLAN or flat network on the network nodes, and the upstream route towards the floating IP range. These are outside OpenStack and outside this page.
Possible causes¶
| Cause | How to tell |
|---|---|
| 1. No security group rule admits the traffic | Check 2: no ingress rule for the protocol, port and source |
| 2. No floating IP, or one associated elsewhere | Check 3 empty, or another port or fixed address |
| 3. No router path, before any floating IP is associated: no interface on the tenant network, or no gateway | Check 4; the association fails with External network … is not reachable from subnet … |
| 4. The API shows a router path and an associated floating IP, but the router is not on its node (agent down, NAT missing) | Namespace or OVN NAT checks |
| 5. The guest drops the traffic | All checks pass; a neighbour cannot connect either |
| 6. The external network or upstream routing | Every floating IP affected; nothing wrong inside OpenStack |
Resolution¶
For cause 1: add the missing ingress rule¶
⚠️ State-changing. Opens the instance (and every other port using this security group) to the traffic the rule describes. Use the narrowest source prefix that works, never
0.0.0.0/0for administration ports unless the project accepts it. Undo:openstack security group rule delete <rule-id>.
# SSH from one admin network
openstack security group rule create --ingress --protocol tcp --dst-port 22 --remote-ip <client-prefix> <security-group-id>
# ICMP (ping) from the same network
openstack security group rule create --ingress --protocol icmp --remote-ip <client-prefix> <security-group-id>
Rules are stateful: the reply traffic is allowed without an egress rule.
For cause 2: associate a floating IP with the right port¶
⚠️ State-changing. Moves the floating IP to this port: if it was associated elsewhere, that other port loses it and anything using it there is cut off. Undo: associate it back with
openstack floating ip set --port.
openstack floating ip create <external-network-id> # if none is free
openstack floating ip set --port <port-id> <floating-ip-id>
# for a port with several fixed addresses, choose which one:
openstack floating ip set --port <port-id> --fixed-ip-address <fixed-ip> <floating-ip-id>
For cause 3: give the tenant network a path to the external network, then associate the floating IP¶
⚠️ State-changing. Setting a gateway gives every instance of the router's networks outbound access (SNAT) to the external network, and consumes an address on it. Adding an interface connects the tenant subnet to the router and its other subnets. Undo:
openstack router unset --external-gateway <router-id>,openstack router remove subnet <router-id> <subnet-id>.
openstack router set --external-gateway <external-network-id> <router-id>
openstack router add subnet <router-id> <tenant-subnet-id>
Then associate the floating IP as for cause 2.
For cause 4: the router is missing on its node¶
(Not run on the lab.)
On ML2/OVS, follow L3 agent dead, router not moved. On ML2/OVN, a missing NAT entry while the API shows the floating IP associated points at the synchronisation between Neutron and the OVN northbound database; this page does not cover repairing it.
For cause 5: the guest¶
(Not run on the lab.)
The instance's owner fixes the guest firewall or the service. Nothing to change in OpenStack.
Verification¶
openstack floating ip list --port <port-id> # one floating IP, on the right fixed address
openstack security group rule list --ingress <security-group-id>
From outside the cloud, from an address inside the prefix you allowed:
ping -c 3 <floating-ip>
ssh <user>@<floating-ip>
Validity and limits¶
Written from the Neutron 2026.1 documentation (networking introduction for the security group defaults, the
self-service deployment guide, the OVN routing guide), the Kolla-Ansible 2026.1 Neutron reference (ovn-nbctl from
the ovn_northd container) and the python-openstackclient 2026.1 source for every option used. Not observed on a
platform when first written.
A lab review on 2026.1 (Kolla-Ansible 22.2.0, all-in-one, ML2/OVS with the L3 agent, legacy router, CirrOS) ran the symptom, the quick checks, the ML2/OVS namespace checks, causes 1 and 2 with their resolutions, and Neutron's refusals described under cause 3. The OVN steps and causes 4 to 6 were not run; a lab run on ML2/OVN is wanted.
Not covered: DVR, BGP-announced floating IPs, IPv6 without NAT, provider networks attached directly to instances (no router, no floating IP), and repairing the Neutron to OVN synchronisation.
Related¶
- New instance gets no IP address
- New instance boots without metadata
- L3 agent dead, instances still reachable, router not moved
- Neutron 2026.1: networking introduction, OVN routing
Provenance: written by agent:ozarland-author, from upstream documentation. 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.