> I've tried both and I have a real objection to the naming scheme Bill
> suggests.  It's not good for multihomed hosts, and it's just plain
> unworkable for multi-homed hosts where the set of interfaces _or_ the
> order in which they're attached changes, because the names (and thus
> the mapping between a specific IP address in a config file and a
> device the bus, which is done via the interface nams) don't carry
> across from one change to another.
> 
> If all interfaces have the same name, that mapping can be changed by
> as little as reordering cards on a bus; or by changing the order in
> which the kernel probes for tehm; or (given LKM network drivers) the
> order in which they're loaded.

This is a mostly-bogus objection.  The same problem existed for SCSI
busses, and the same solution can be used.  This can all be handled
using the existing autoconfiguration machinery.

The reason that this is "mostly" bogus (rather than completely bogus)
is that i'm pretty sure that there would be problems trying to tie
pseudo-device network interfaces (slip, ppp, etc.) to specific device
names.  However, if there were a desire to fix that problem, it could
probably be fixed easily enough.



FWIW, I'm _not_ keen on making the proposed change, because I simply
don't like it, and because it's quite different than what's been done
in the past in BSD.  I'm not sure if the mind-shift necessary for
experienced BSD users is justified by the rather small (in my opinion)
gain that could be had.




cgd