Phase 12b: vpcd now reconciles OVN topology from NATS KV on startup
(Pass 2). After subscribing to topics, it reads all VPCs, subnets,
IGW attachments, and ENIs from KV and ensures matching OVN objects
exist. Handles reboots, OVN DB loss, and missed events. All operations
are idempotent.
Phase 16 prep: Add --gateway-ip flag to admin init for NAT mode
(specifies the OVN gateway router's external IP without DHCP).
Update reset-dev-env.sh to support EXTERNAL_MODE=nat env var for
NAT/DHCP mode testing alongside the default pool mode.
Auto-detect bridge topology: if a dedicated WAN NIC exists (not used for
SSH), add it directly to br-external instead of creating a macvlan.
Direct bridge enables distributed NAT (DNAT on VM's own chassis) and
eliminates macvlan workarounds (MAC alignment, gratuitous ARPs,
centralized NAT bottleneck).
- topology.go: bridge-mode-aware NAT and localnet port options
- vpcd.go: auto-detect bridge mode, skip MAC alignment for direct bridge
- setup-ovn.sh: --direct flag for physical NIC on br-external
- reset-dev-env.sh: SSH NIC detection via SSH_CONNECTION
Three blockers fixed for VPC public subnet E2E on single-node:
1. OVN SB stale data: reset-dev-env.sh now deletes both NB+SB DB files
instead of individual NB cleanup. Eliminates ovn-controller commit
failure loop from stale chassis/port binding entries.
2. Chassis name mismatch: reset-dev-env.sh passes --chassis-id=chassis-node1
to setup-ovn.sh, aligning OVS system-id with vpcd's chassis-{nodeName}.
3. Macvlan unicast filtering: switched to centralized NAT (no external_mac
on DNAT rules), set nat-addresses=router on localnet ports, and vpcd
aligns macvlan MAC with OVN gateway router MAC on startup. Macvlan only
delivers unicast matching its own MAC — this ensures inbound traffic
from the LAN reaches OVS/OVN.
Also fixed: DNS for VMs. admin init auto-detects host DNS servers via
resolvectl, stores them in spinifex.toml dns_servers field, and vpcd
propagates them to OVN DHCP options. Fallback: 8.8.8.8 + 1.1.1.1.
Validated E2E: inbound ping, SSH, outbound ping, DNS, curl, apt update.
The single-nic IP migration path was fragile and caused connectivity
issues during setup. macvlan is now the only bridge strategy — always
creates spx-ext-{nic} on the WAN NIC. Removed --single-nic flag,
SingleNIC config/detection fields, and updated all docs.
When setup-ovn.sh migrates the WAN IP from a physical NIC to
br-external, systemd-resolved loses its DNS association (DNS servers
are per-interface). This breaks hostname resolution for everything
on the host — git, apt, curl, SSH all fail.
Now captures DNS config from the original interface before migration
and re-associates it with br-external via resolvectl. Falls back to
8.8.8.8 if no DNS was configured on the original interface.
When setup-ovn.sh migrates the host's WAN IP to br-external, stopping
openvswitch-switch destroys the bridge and the host loses all network
connectivity — unreachable via SSH, requires console reboot.
Now checks if br-external has an IP before stopping OVS. If it does,
skip the OVS stop and only stop ovn-controller + ovn-central. OVS is
infrastructure — bridges must survive service stop/start cycles.
Discovered during 3-node E2E testing: stop-dev.sh killed all 3 nodes'
WAN connectivity simultaneously.
spx admin init now works with zero external networking flags:
- Auto-detects WAN interface, gateway, prefix from ip route
- Auto-detects single-NIC vs multi-NIC mode
- Defaults to nat mode with DHCP for gateway IP
- Prints detected topology with interface roles
setup-ovn.sh gains --dhcp flag: obtains a DHCP lease on the
macvlan/bridge interface for the OVN gateway SNAT address, then
writes gateway_ip into spinifex.toml. Tested on VM: macvlan gets
separate IP from router DHCP, host keeps its original IP.
Three tiers of external networking:
- Zero flags: auto nat+DHCP (VMs get outbound internet)
- --external-pool=start-end: pool mode (VMs get public IPs)
- --no-external: overlay only (no internet for VMs)
Add --single-nic flag to setup-ovn.sh. When used with --external-bridge,
creates a macvlan sub-interface (spx-ext-{iface}) in bridge mode off the
host's NIC instead of adding the NIC directly to br-external. The host
keeps its IP — no migration, SSH-safe.
Strategy selection:
- Dedicated NIC: --external-iface=eth1 (add directly)
- Single NIC: --external-iface=eth0 --single-nic (macvlan)
- IP migration: --external-iface=eth0 (legacy, no --single-nic)
Add single_nic bool field to VPCDConfig for per-node config. Different
nodes in a cluster can use different strategies. Unit test verifies
single_nic parsing alongside external_interface.
Add external IPAM module (CAS-based IP allocation from admin-defined
pools in NATS KV). Supports multi-pool with region/AZ scoping, pool
fallback logic, gateway IP reservation, and concurrent allocation via
CAS retries. 12 unit tests covering allocation, release, exhaustion,
CAS conflict, multi-pool, fallback, validation, and idempotent init.
Add --external-bridge flag to setup-ovn.sh: creates OVS bridge
br-external, adds WAN NIC, migrates host IP to bridge, and sets
ovn-bridge-mappings=external:br-external for OVN localnet ports.
Add commented [network] section and external_interface to spinifex.toml
config template for discoverability.
Redistribute E2E tests between single-node and multi-node suites, slim
down pseudo multi-node suite, and remove directory creation from
dev-setup.sh so spx admin init is the single source of truth for
~/spinifex/ setup.
The MIN_LOC=100 threshold checked total diff lines, not instrumentable
lines. A commit with many non-instrumentable lines (structs, comments)
but only 1-2 instrumentable lines would unfairly fail at 0% coverage.
Add a minimum instrumentable lines threshold (5) so the check skips
when the sample size is too small to be meaningful.
Extract duplicated respondWithError closures into a single package-level
function. Guard against nil ImageId in handleEC2RunInstances to prevent
panic under race conditions. Fix diff-coverage base ref for main branch.
is_multinode() counted all [nodes.*] TOML subsections (daemon, nats, awsgw,
etc.) instead of just top-level node entries. A single-node config has ~8
such sections, causing the single-node Docker E2E to use coordinated cluster
shutdown and exit with code 254.
Fix grep to match only [nodes.X] (not [nodes.X.Y]) and preserve the original
test exit code in the cleanup trap.
Ben Duncan