fix(hostagent): check PF interface before writing sriov_numvfs

[tsorya]

Summary

• Guard SetNumOfVFs with a PF interface check before writing sriov_numvfs
• Prevents a catastrophic VF probe cascade (~46 VFs × 120s timeout each) when mlx5_core is in error state after firmware fatal error
• The 30s reconcile loop naturally retries once the DPU firmware recovers

Problem

When the DPU firmware hits a fatal error during provisioning, mlx5_core on the host enters error state and tears down the netdev — but remains bound to the PCI device. The network manager's CreateP0VF step writes sriov_numvfs regardless, triggering the kernel to enumerate dozens of VFs that each block 120s in wait_fw_init, resulting in 90+ minutes of stuck kernel threads and failed provisioning.

Fix

Check that the PF has a net interface (proving the driver is healthy and firmware is responsive) before calling SetNumOfVFs. If the interface is missing, return an error early — the network manager retries every 30s and will succeed once the DPU finishes booting.

Test plan

• Verify unit tests pass (go test ./internal/provisioning/hostagent/util/)
• Deploy on a cluster and provision a DPU — confirm VFs are created normally when PF is healthy
• Simulate firmware error scenario (or trigger BFB install) — confirm no VF cascade, provisioning recovers after DPU boot

Made with Cursor