Fix - e1000e NIC Detected Hardware Unit Hang (please re-open issue 10200)

[tim-van-dijkhuizen]

Sorry if I'm abusing the discussions section. I searched for ways to reach a maintainer and this seemed like the least annoying/intrusive way.

I recently encountered the problem described in this GitHub issue:
#10200

Someone suggested a pretty janky fix that works and I applied it in a much cleaner way using a new feature in Talos. I think it would be of great help if this solution gets posted on the issue. Unfortunatly it has been closed so I can't post it.

Would any of the maintainers be so kind to re-open the issue or post the solution below, thank you so much!

Disclaimer: I used ChatGPT to help me and write comments, the code was reviewed and tested by me.

apiVersion: v1alpha1
kind: EthernetConfig
name: eno1 # network interface to fix

# Disable problematic NIC offloads to prevent Intel e1000e "Detected Hardware Unit Hang" errors.
# These are safe to turn off — functionality remains the same, only CPU does more work instead of the NIC.

features:
  # TCP Segmentation Offload (TSO): lets NIC split large TCP packets.
  # Known trigger for e1000e hangs under high packet-per-second workloads.
  tx-tcp-segmentation: "off"

  # IPv6 variant of TSO: same issue as above, but for IPv6 traffic.
  tx-tcp6-segmentation: "off"

  # Generic Segmentation Offload (GSO): kernel-side batching of large packets.
  # Disabling ensures NIC never sees oversized segments that cause descriptor corruption.
  tx-generic-segmentation: "off"

  # Generic Receive Offload (GRO): kernel merges multiple small incoming packets.
  # Disabling prevents asymmetry between RX/TX rings that can lock up the NIC.
  rx-gro: "off"

  # Receive checksum offload: NIC validates checksums on inbound packets.
  # Safe to disable; kernel will do this in software. Reduces load on buggy paths in e1000e.
  rx-checksum: "off"

  # Transmit checksum offload (generic): NIC calculates checksums for outbound packets.
  # Again safe to disable, avoids faulty offload handling in driver/hardware.
  tx-checksum-ip-generic: "off"

  # VLAN hardware tagging: NIC inserts VLAN headers.
  # Disabled to avoid offload edge cases; VLAN tagging still works via software.
  tx-vlan-hw-insert: "off"

  # VLAN hardware parsing: NIC strips VLAN headers on receive.
  # Disabled for consistency; kernel will parse VLAN tags in software.
  rx-vlan-hw-parse: "off"

[asimhusain-ai]

apiVersion: v1alpha1
kind: EthernetConfig
name: eno1

features:
tx-tcp-segmentation: "off"
tx-tcp6-segmentation: "off"
tx-generic-segmentation: "off"
rx-gro: "off"
rx-checksum: "off"
tx-checksum-ip-generic: "off"
tx-vlan-hw-insert: "off"
rx-vlan-hw-parse: "off"

[rothgar]

This sounds like something that should go into the official documentation, not github issue comments or discussions. I'll label this discussion and we can figure out where it will fit.