[ipxe-devel] Xen netfront support

[Michael Brown]

I'm please to announce that iPXE now natively supports Xen netfront 
virtual NICs.  This may be of interest to anyone wanting to use iPXE 
running in a Xen PV-HVM domain.

I have one issue which I'm not sure how to handle, and would appreciate 
feedback from anyone who frequently uses Xen:

For each configured NIC, Xen exposes both an emulated PCI NIC (for OSes 
with no native Xen drivers) and a netfront virtual NIC (for OSes with 
native Xen drivers).  Xen provides a mechanism for OSes to "unplug" the 
emulated PCI NICs.  Operating systems with native Xen drivers will 
typically unplug the emulated PCI NICs to prevent confusion.

iPXE could easily unplug the emulated PCI NICs.  However, this operation 
is irreversible.  If the loaded OS does not include native Xen drivers, 
then it will not be able to see any network devices.  This is undesirable.

iPXE could easily expose both the emulated PCI NICs and the netfront 
virtual NICs (as e.g. net0 and net1).  This works perfectly (since the 
backends are entirely independent; the "hardware" really _is_ providing 
two NICs which happen to have the same MAC address and be connected to 
the same network), but it's somewhat confusing for a user to see:

   iPXE> ifstat
   net0: 56:3c:dc:73:51:7e using netfront on vif/0 (open)
     [Link:up, TX:0 TXE:0 RX:0 RXE:0]
   net1: 56:3c:dc:73:51:7e using rtl8139 on PCI00:04.0 (open)
     [Link:up, TX:0 TXE:0 RX:0 RXE:0]

iPXE cannot easily mask out just the emulated NICs, since we have no 
easy way to identify which ones they are (other than irreversibly 
unplugging them).  Masking by PCI vendor:device IDs known to be used by 
qemu doesn't work, since there's nothing preventing someone using PCI 
pass-through to expose a real physical NIC with the same vendor:device 
ID.  Also, this approach would be fragile since the list of emulated NIC 
types is known to change occasionally.

Any suggestions?

Michael