What problem does it solve?
Reading a ConnectX or BlueField device's firmware PCC (Programmable Congestion Control) diagnostic counters requires knowing the exact two-operation contract of the pcc_counters.sh script, the mst device resolution flow, and the mlx5 debugfs diag_cnt interface, and mistakes like querying before arming or misreading a zero counter lead to wrong congestion-control conclusions.
Core Features & Use Cases
- Guided counter readout: Walks the agent through the canonical set-then-query sequence against an mst device path, with the fixed firmware counter set (PCC_CNP_COUNT, RTT-perf, WRED-drop families) documented verbatim.
- Layered error diagnosis: Maps symptoms like ERROR: Bad Device, Bad Request, empty dumps, and counters stuck at zero to a seven-layer error taxonomy covering install, mst resolution, debugfs permissions, and traffic state.
- Safety-gated privileged operations: Flags the set operation as a privileged debugfs write requiring operator confirmation, and refuses to translate counter deltas into congestion-control tuning changes.
- Use Case: An operator investigating whether congestion control is dropping packets on a RoCE port asks the agent to capture the CNP and WRED-drop counters for /dev/mst/mt41692_pciconf0, and the skill produces a verbatim counter snapshot as evidence.
Quick Start
Ask your agent to arm and read the PCC diagnostic counters on your device, for example: "Use pcc_counters.sh to set and then query the PCC counters on /dev/mst/mt41692_pciconf0 and show me the CNP and WRED-drop values."