Fix RTL8125 2.5GbE ethernet that never gets a DHCP lease
Three defects, one of them fatal to checking the record at all. First, the single cited source is dead: https://github.com/basecamp/omarchy/issues/7804 returns a hard 404 with no redirect, tested with both a browser user agent and a named tool user agent, as does every other basecamp/omarchy/issues/<n> path, so the rename to omacom/omarchy is not followed for issue URLs. The live issue is https://github.com/omacom/omarchy/issues/7804, opened 2026-08-22 by donutWolf and titled "Wired ethernet stopped getting a DHCP lease after upgrading to Quattro (RTL8125 NICs)", and I read it in full through the API: it supports the symptom, the cause, the ruled-out list and the ethtool -K ... tx off plus systemd service fix almost word for word, so the record's content was faithful and only its URL was wrong. Second, the fix cannot run as written on Omarchy 4: ethtool is not installed on this workstation, /usr/bin/ethtool does not exist, pacman -Qo /usr/bin/ethtool reports no owner, and it is neither a dependency nor an optional dependency of networkmanager 1.58.1-1, so sudo ethtool -K eno1 tx off fails with command not found and the unit's ExecStart=/usr/bin/ethtool would fail 203/EXEC at every boot. It is extra/ethtool 1:7.1-1, and the corrected fix installs it through omarchy update followed by sudo pacman -S --needed ethtool, which I confirmed is not blocked: /usr/bin/omarchy-update-pacman-guard aborts only when the pacman command line carries both a sync and a sysupgrade flag. Third, the persistence mechanism ignores that NetworkManager owns the connections on Omarchy. NetworkManager 1.58.1 has ethtool.feature-tx, which I confirmed both in man nm-settings-nmcli on this machine and in the upstream nm-settings-nmcli page, and it is applied during activation ahead of the first DHCP request, so it needs no ethtool binary and does not race. The record's device-unit approach does race, because the oneshot and NetworkManager's own activation both fire when sys-subsystem-net-devices-<iface>.device appears, and I confirmed such device units exist and are active here. I kept that unit as a labelled fallback and added a third route independent of any connection manager, a .link file using TransmitChecksumOffload=false, which systemd.link(5) has carried since version 245 and this machine runs systemd 261.2-1, with the default NamePolicy and AlternativeNamesPolicy lines copied from /usr/lib/systemd/network/99-default.link because the Arch wiki warns that only the first matching .link file applies and that omitting the default content can misconfigure the interface. I set corrected_frequency to rare and left severity at high: the evidence is one self-reported, still open issue with a single comment, that comment reports the same symptom on an e1000e NIC rather than a Realtek one, modinfo -p r8169 on kernel 7.1.9 prints nothing so the driver exposes no knob for this, and the Arch wiki Ethernet page has no RTL8125 or checksum-offload section at all, so occasional overstated how often this is the answer while losing the network entirely is still high severity. The cause is rewritten to keep the mechanism but say whose diagnosis it is, and the danger now names the interface-rename trap in the .link route, which can leave a machine with no network and is a worse outcome than the offload change it was flagging. Not exercised: no RTL8125 is in this machine (the wired NIC is an Intel I219-V on e1000e), so no lease failure could be reproduced, and per instruction I ran no ethtool -K, installed nothing and changed no network configuration here.
Disabling TX checksum offload moves checksumming to the CPU. The cost is negligible at 2.5 Gbit, but it is a real behaviour change, so do not apply it to a NIC that is working. The larger risk is in the persistence step rather than in the offload. A .link file that matches your interface and omits the default NamePolicy and AlternativeNamesPolicy lines can change the interface name, and NetworkManager's saved profile then no longer matches it, which leaves the machine with no network at all. Only the first matching .link file is applied, so check the result with sudo udevadm test-builtin net_setup_link /sys/class/net/eno1 and networkctl status eno1 before you reboot, and keep a way back in that does not depend on this NIC.
Symptom
Wired ethernet shows link up and negotiates the correct speed, but never receives an IP address. nmcli/networkctl report DHCP timing out on every retry. Wi-Fi on the same machine works. The same cable and port get a lease instantly from Windows on the same hardware. sudo tcpdump -i eno1 -n udp port 67 or port 68 shows DHCPDISCOVER going out repeatedly with zero replies.
Cause
One reporter traced this to TX checksum offload in the in-kernel r8169 driver on an RTL8125D: the NIC computes bad checksums on outgoing packets, the router drops every DHCP request, and no reply ever comes back. A tcpdump capture cannot rule that out, because the capture is taken before the hardware inserts the checksum, so the request looks sound in the capture whatever the card later puts on the wire. The same reporter ruled out firewall rules, EEE, PCIe ASPM, the DHCP client identifier and the choice of NetworkManager versus systemd-networkd, which are the usual suspects for these symptoms. Treat the attribution as one person's diagnosis rather than a confirmed driver bug: the upstream issue is still open with no maintainer response, r8169 on kernel 7.1 exposes no module parameters to tune this, and the only other person on the thread reports the same symptom on an Intel e1000e NIC, which points at something wider than one Realtek revision. The test is cheap and reversible, so it is worth trying before spending hours on the router.
Fix
Confirm the chipset and the driver first:
lspci -k | grep -A3 -i ethernet # look for RTL8125 and "Kernel driver in use: r8169"
ip -br link show # the interface name, eno1 in the examples below
Omarchy 4 does not ship ethtool, so install it before testing anything. Do not reach for pacman -Sy, which is a partial upgrade, and note that pacman -Syu is aborted by Omarchy's ALPM guard:
omarchy update # the supported full sync and upgrade
sudo pacman -S --needed ethtool
Test the theory without persisting anything:
sudo ethtool -K eno1 tx off
sudo nmcli connection up "Wired connection 1"
ip addr show eno1 # expect an IPv4 address within a few seconds
If that gets a lease, make it persistent. **On Omarchy 4 NetworkManager owns the connection**, so put the setting on the connection profile. NetworkManager applies it during activation, before it sends the first DHCP request, and it needs neither the ethtool binary nor an extra unit:
nmcli -g NAME,DEVICE connection show --active
sudo nmcli connection modify "Wired connection 1" ethtool.feature-tx off
sudo nmcli connection up "Wired connection 1"
nmcli -f ethtool connection show "Wired connection 1"
To undo it, clear the property rather than setting it back to on, so the kernel default returns:
sudo nmcli connection modify "Wired connection 1" ethtool.feature-tx ""
**If the machine does not use NetworkManager**, or the NIC has to come up with the offload already off before any connection manager touches it, use a systemd link file, which udev applies as the device appears. Copy the default NamePolicy and AlternativeNamesPolicy lines from /usr/lib/systemd/network/99-default.link, because only the first matching .link file is applied and dropping them can rename the interface:
sudo tee /etc/systemd/network/50-rtl8125-txoff.link >/dev/null <<'EOF'
[Match]
Driver=r8169
[Link]
NamePolicy=keep kernel database onboard slot path
AlternativeNamesPolicy=database onboard slot path mac
MACAddressPolicy=persistent
TransmitChecksumOffload=false
EOF
sudo udevadm control --reload
sudo udevadm trigger --subsystem-match=net --action=add
Driver=r8169 matches every r8169 NIC in the machine. To pin it to one card, match on the address instead with MACAddress=aa:bb:cc:dd:ee:ff in the [Match] section.
The systemd service the upstream issue suggests also works once ethtool is installed, but it can race the first DHCP attempt, because the service and NetworkManager both start when the device appears:
sudo tee /etc/systemd/system/rtl8125-txoff@.service >/dev/null <<'EOF'
[Unit]
Description=Disable TX checksum offload on %i
BindsTo=sys-subsystem-net-devices-%i.device
After=sys-subsystem-net-devices-%i.device
[Service]
Type=oneshot
ExecStart=/usr/bin/ethtool -K %i tx off
RemainAfterExit=yes
[Install]
WantedBy=sys-subsystem-net-devices-%i.device
EOF
sudo systemctl daemon-reload
sudo systemctl enable --now rtl8125-txoff@eno1.service
Verify
Check the setting took:
ethtool -k eno1 | grep tx-checksumming # expect: off
nmcli -f ethtool connection show "Wired connection 1" | grep feature-tx
Then reboot cold and confirm the lease arrives on its own, every time:
ip addr show eno1 # an IPv4 address within a couple of seconds
journalctl -b -u NetworkManager | grep -i dhcp # a lease, not repeated timeouts