How to Use ethtool in Linux to Diagnose Network Link Speeds, Ring Buffers, and Offload Settings

Architectural Foundation of the Linux Network Driver and ethtool Subsystem

Modern enterprise servers running 10 GbE, 25 GbE, and 100 GbE network fabrics frequently encounter subtle performance degradation—such as packet drops during traffic microbursts, duplex mismatches, and hardware offload anomalies. In Linux, ethtool serves as the primary utility for inspecting, tuning, and diagnosing physical and virtual Network Interface Cards (NICs), communicating directly with the kernel’s ethtool_ops driver callbacks via Netlink (ethnl) and legacy ioctl sockets.

When Ethernet frames reach a physical network card, hardware PHY transceivers negotiate electrical signaling parameters before transferring frames into Direct Memory Access (DMA) Descriptor Ring Buffers. If an incoming traffic burst exceeds the capacity of the card’s receive (RX) ring buffer before the CPU can service the hardware interrupt (via NAPI polling), the NIC’s internal FIFO queue overflows, resulting in silent hardware frame discards. Leveraging ethtool allows systems engineers to expand descriptor ring sizes, audit physical PHY layer errors, and tune protocol offloading engines directly at the hardware layer.

Subsystem Component	Kernel Driver Interface	Primary Diagnostic Role	Performance Impact
PHY Layer Transceiver	Auto-negotiation registers	Speed, duplex, and MDI-X signaling	Determines physical line rate (e.g., 1000 Mb/s vs 10 Gb/s)
RX/TX Ring Descriptors	DMA circular memory pointers	Queues incoming and outgoing Ethernet frames	Prevents frame discards during microburst traffic spikes
Hardware Offload (TSO/GSO)	NIC silicon compute blocks	Offloads TCP segmentation and checksumming	Frees CPU cores from intensive packet packetization
Hardware MAC Counters	Silicon error registers	Audits CRC, alignment, and FIFO errors	Isolates physical cable and transceiver degradation

Auditing Physical Link Speed and Duplex Parameters

To inspect physical link parameters and auto-negotiation states, invoke ethtool against the target interface name (e.g., eth0 or enp3s0):

# Query physical interface link state, speed, and duplex configuration
sudo ethtool eth0

Key output telemetry fields to evaluate:

Settings for eth0:
        Supported ports: [ TP ]
        Supported link modes:   1000baseT/Full, 10000baseT/Full
        Speed: 10000Mb/s
        Duplex: Full
        Auto-negotiation: on
        Port: Twisted Pair
        PHYAD: 0
        Transceiver: internal
        Link detected: yes

If Speed: reports 100Mb/s on a Gigabit-rated interface or Duplex: displays Half, the physical link has negotiated a fallback state, typically caused by damaged twisted-pair cabling or misconfigured switch port autonegotiation.

Diagnosing and Expanding NIC Ring Buffers

Packet loss occurring at high line rates while CPU utilisation remains low is classic evidence of descriptor ring buffer exhaustion. Inspect current versus maximum ring buffer allocations:

# Query current and maximum ring buffer sizes
sudo ethtool -g eth0

Example ring buffer output:

Ring parameters for eth0:
Pre-set maximums:
RX:             4096
RX Mini:        n/a
RX Jumbo:       n/a
TX:             4096
Current hardware settings:
RX:             512
TX:             512

In this scenario, the hardware supports 4,096 descriptors, but the driver is constrained to a conservative default of 512. Expand both RX and TX rings to their hardware maximums:

# Expand RX and TX ring buffers to 4096 descriptors
sudo ethtool -G eth0 rx 4096 tx 4096

# Verify the updated hardware settings
sudo ethtool -g eth0

Managing Hardware Offload Settings (TSO, GSO, and Checksumming)

Modern NICs accelerate throughput by processing TCP/IP packetization inside onboard silicon via TCP Segmentation Offload (TSO) and Generic Receive Offload (GRO).

# 1. Audit all active and fixed protocol offload settings
sudo ethtool -k eth0

# 2. Disable TCP Segmentation Offload (Useful when diagnosing bridge/NAT packet corruption)
sudo ethtool -K eth0 tso off

# 3. Enable Generic Receive Offload and hardware checksumming
sudo ethtool -K eth0 gro on rx on tx on

Auditing Hardware Error Registers with ethtool -S

Standard Linux networking commands (like ip -s link) only aggregate general packet counters. To inspect low-level silicon hardware registers, query extended statistics:

# Query vendor-specific hardware counters
sudo ethtool -S eth0 | grep -E "drop|error|crc|fifo|miss"

Critical error register indicators:

• rx_crc_errors: Signals physical layer noise, bad patch cables, or bent optical fibre pins.
• rx_missed_errors / rx_fifo_errors: Confirms that packets arrived at the physical MAC but were dropped due to exhausted host DMA ring buffers.
• tx_aborted_errors: Signals switch-side flow control backpressure (Ethernet Pause frames).

Identifying Physical Ports in Server Racks

When working in high-density data centres with dozens of identical Ethernet ports, identify the target physical port by blinking its chassis LED:

# Flash the physical port LED on eth0 for 15 seconds
sudo ethtool -p eth0 15

Troubleshooting Common ethtool and NIC Driver Faults

Observed Anomaly	Underlying Root Cause	Diagnostic Verification Method	Remediation Protocol
Link Flaps Repeatedly Between 1000 Mb/s and 100 Mb/s	Physical cable pin degradation or autonegotiation mismatch with upstream switch	Inspect ethtool eth0; speed fluctuates, rx_crc_errors increments	Replace Ethernet patch cable; verify switch port is set to auto-negotiate, not forced speed
Ring Buffer Settings Reset to 512 Upon Reboot	ethtool -G commands modify volatile runtime kernel driver state only	Check ring buffers after system reboot; settings revert to hardware defaults	Persist settings using systemd-networkd ([Link] ReceiveBufferSize=4096) or a udev rule in /etc/udev/rules.d/
Packet Checksum Errors on Virtual Bridges / VXLAN	NIC hardware checksum offload corrupted virtual network encapsulation headers	Wireshark captures report TCP Checksum Incorrect across internal bridge	Disable checksum offload on the virtual adapter: sudo ethtool -K eth0 tx-checksum-ip-generic off
“Operation not supported” Error on ethtool -G	Interface is a virtual paravirtualized device (e.g., VirtIO, veth) without physical ring hardware	Run lspci or ethtool -i eth0; driver reports virtio_net or veth	Ring buffer tuning must be executed on the physical host hypervisor NIC, not inside the guest VM