LAN Routing Fix: Tailscale 10GbE Throughput Issue

Problem

Hosts with host-level Tailscale on the 192.168.0.0/24 LAN have their local traffic intercepted by Tailscale's policy routing table 52. Instead of going directly over the physical 10GbE link, traffic gets routed through the WireGuard tunnel via Calypso's advertised 192.168.0.0/24 subnet route.

Root Cause

Calypso (Headscale node ID:12) advertises 192.168.0.0/24 as a subnet route so remote nodes (Moon, Seattle, NUC) can reach LAN devices over Tailscale. However, machines that are already on that LAN also accept this route into Tailscale's routing table 52 (ip rule priority 5270), causing local traffic to hairpin through the tunnel.

Diagnosis:

# Shows traffic going through tailscale0 instead of the physical NIC
ip route get 192.168.0.200
# → 192.168.0.200 dev tailscale0 table 52 src 100.75.252.64

# Table 52 has the LAN subnet routed through Tailscale
ip route show table 52 | grep 192.168.0
# → 192.168.0.0/24 dev tailscale0

Affected Hosts

Host	LAN IP	Physical NIC	Affected?
Guava (TrueNAS)	192.168.0.100	enp1s0f0np0 (10GbE)	YES — fixed
homelab-vm	192.168.0.210	ens18	YES — fixed
Atlantis	192.168.0.200	eth2/ovs_eth2 (10GbE)	No (Synology OVS)
Calypso	192.168.0.250	ovs_eth2	No (Synology OVS)
Pi-5	192.168.0.66	eth0	No (not accepting route)
NUC	192.168.68.100	eno1	No (different subnet)

Measured Impact (Guava → Atlantis)

Route	Throughput	Retransmits
Before fix (via Tailscale)	1.39 Gbps	6,891
After fix (direct LAN)	7.61 Gbps	5,066

5.5x improvement — from WireGuard-encapsulated tunnel to direct 10GbE.

Fix Applied

Add an ip policy rule at priority 5200 (before Tailscale's table 52 at 5270) that forces LAN traffic to use the main routing table, which routes via the physical NIC:

sudo ip rule add to 192.168.0.0/24 lookup main priority 5200

This means: for any traffic destined to 192.168.0.0/24, check the main table first. The main table has 192.168.0.0/24 dev <physical-nic>, so traffic goes direct. All Tailscale traffic to 100.x.x.x nodes is unaffected.

Verification

# Should show physical NIC, not tailscale0
ip route get 192.168.0.200

# Should get sub-1ms ping
ping -c 3 192.168.0.200

# Confirm rule is in place
ip rule show | grep 5200

Revert

sudo ip rule del to 192.168.0.0/24 lookup main priority 5200

Persistence

Guava (TrueNAS)

Init script added via TrueNAS API (ID: 2):

Type: COMMAND
When: POSTINIT
Command: ip rule add to 192.168.0.0/24 lookup main priority 5200
Comment: Bypass Tailscale routing for LAN traffic (direct 10GbE)

Manage via TrueNAS UI: System → Advanced → Init/Shutdown Scripts

homelab-vm (Ubuntu 24.04)

Systemd service at /etc/systemd/system/lan-route-fix.service:

[Unit]
Description=Ensure LAN traffic bypasses Tailscale routing table
After=network-online.target tailscaled.service
Wants=network-online.target

[Service]
Type=oneshot
ExecStart=/sbin/ip rule add to 192.168.0.0/24 lookup main priority 5200
ExecStop=/sbin/ip rule del to 192.168.0.0/24 lookup main priority 5200
RemainAfterExit=yes

[Install]
WantedBy=multi-user.target

Enabled with sudo systemctl enable lan-route-fix.service.

Notes

Remote nodes (Moon, Seattle, NUC) that are not on 192.168.0.0/24 are unaffected — they correctly use Calypso's subnet route to reach LAN devices via Tailscale.
If a new host is added to the LAN with host-level Tailscale, the same fix will need to be applied.
The Synology boxes (Atlantis, Calypso) use Open vSwitch bridging and don't exhibit this issue.

3.7 KiB Raw Blame History