ConnectX-5 link stays down after repeated DAC plug/unplug cycles until mlxconfig reset
A dual-port ConnectX-5 (MT27800, firmware 16.35.1012, MLNX_OFED 5.8 on BC-Linux 8.2) connected to a switch with a DAC would not recover after the cable had been connected and disconnected about fifteen times: the port stayed in State: Polling with Physical state ETH_AN_FSM_ABILITY_DETECT, and mlxlink recommended Negotiation failure (or Other issues with auto-negotiation off). Only a NIC reset restored the link.
- Hardware: Mellanox ConnectX-5 MT27800
- Transceivers: SFP28/QSFP28 DAC
- Form factor: DAC
Symptoms / messages:
State: Polling
ETH_AN_FSM_ABILITY_DETECT
Negotiation failure
Comments 1
NVIDIA support: verify the cable against the supported list, run mlxconfig -d /dev/mst/mt4119_pciconf0 reset and reboot; the reset fixed it and the problem did not return after restoring the original settings, pointing to an internal state rather than the configuration values.