collectors: add sriov metrics collector

[aharivel]

Add a new collector that exports SR-IOV metrics for physical functions and virtual functions. Supports VFs bound to both network drivers and vfio-pci by reading stats from the parent PF when direct VF stats are unavailable.

Exported metrics include traffic counters (rx/tx bytes, unicast, multicast, broadcast), error counters (dropped, allocation failures), and TX performance metrics. Each metric includes NUMA node information for topology-aware monitoring.

Parses per-VF statistics from Intel PF drivers (ixgbe, i40e, ice) which use different naming conventions for VF stats.

[rjarry]

I wonder if it would make sense and/or be simpler to query the statistics using https://github.com/vishvananda/netlink.

The stats are parsed here: https://github.com/vishvananda/netlink/blob/c6faf428e8f84dcb73774e7c77a1e4fe38bbdb4d/link_linux.go#L3991

That way, no need to deal with hardware specific ethtool extended stats.

[aharivel]

I wonder if it would make sense and/or be simpler to query the statistics using https://github.com/vishvananda/netlink.

The stats are parsed here: https://github.com/vishvananda/netlink/blob/c6faf428e8f84dcb73774e7c77a1e4fe38bbdb4d/link_linux.go#L3991

That way, no need to deal with hardware specific ethtool extended stats.

The problem is when the VF is bond to vfio_pci and netlink communicate with the kernel only . That's why I used "ethtool -s" to get the metrics directly from the PF.

From what I see, it works only for mlx_5 driver. because each VF has a representator on the eswitch (e.g., eth0_0, eth0_1, or enp3s0f0_0) and when the VF is bond to vfio_pci, the representor still exists and has stats.

So it's going to be tricky to filter which VF is intel and which is mlx in order to retrieves the metrics the right way.

Any other idea ?