====SR-IOV====
I don't think this can be made to work.

The number of VFs per port is controlled when the module is loaded:

  options mlx4_core num_vfs=3,3,0 probe_vf=3,3,0

The triplet is the number of VFs for port 1, the number of VFs for port 2, and the number of VFs that contain both ports. **The last number only works if the card is configured for ETH/ETH operation.**

You must probe all your VFs. If you set "don't probe this" it will not be possible to change the VF settings.

The PCI devices, network devices visible via `ip a`, and devices visible via `ip link` do not match.

To change the settings on a VF, use driverctl to unbind a VF:

  [root]@[shark][00:02:55][~]# driverctl set-override 0000:0e:00.5 none
  [root]@[shark][00:03:06][~]# ip link set mlx1 vf 4 mac b6:2e:99:38:a9:66
  [root]@[shark][00:04:01][~]# driverctl set-override 0000:0e:00.5 mlx4_core

However, the **ONLY** traffic I can get the VF to pass is broadcast storm traffic if I add it to a bridge and create a broadcast loop. In all other ways it acts like it's on a network segment by itself.

====GUIDs====
The MAC address for an adapter is just the GUID with the 03:00 removed from the middle.

  * 00:02:c9:03:00:b3:7f:10 - shark ConnectX-3 FCBT with no date on the label
  * f4:52:14:03:00:2c:18:b0 - uninstalled ConnectX-3 dated 2014-01-29