...Every time I look at this, and then think about throwing slices or
Charles' wedges into the mix too, it looks like a lot more than I want
to rely on working, in order to boot to a single-user prompt.  Or, for
that matter, to fit into install media.

My own personal take is that a lot of the problems of detaching
devices, and then reattaching them, really revolve around nonvolatile
state _in_ the device.  For a disk, that means what filesystem the
platters acutally contain -- is it your /usr, or your /home?  For a
network card, it's what subnet is reachable via the card.  (at least
for ethernet; modems are trickier.)

Regardless of one's religion about detachhing and re-attaching
devices, I'd like to have a PCMCIA-attached disk mounted at ``the
same'' mount-point if it's removed and then re-attached.  For network
devices, I'd like the network subsystem to assign the ``right'' IP
address to the card (either via a config file, or by running a
dhcp/bootp client, or whatever), and then make the appropriate
routing-table changes automagically.

Trying to force those onto bus-specific locators has alwways seemed to
me like a mistake.  To meet the goals of treating ``the same'' device
in ``the same'' way if it's detached and remounted, we need some
definition of ``the same'' that really meets the user's expectation of
"the same".  From the examples we've had here, bus locators just
aren't a good way of doing that in all situations.

I'd prefer to punt handling of removing and re-inserting ``the same''
device to a luser-land daemon, and letting the sysadmin configure what
"the same" _really_ means, in his/her local environment.

Here, the volumed (or whatever we call daemon that decides what
devices are``the same'') implements a naming hierachy layered _on top
of_ the existing autoconfig names (rather than replacing them, as
Solaris does).

The existing autoconfig device names are still around, so if the
luser-land daemon loses its brains, you can connect up the devices you
need, reboot, and manually reconfigure stuff using the
autoconfig-generated device names we know and love.  From previous
discussions about PCMCIA, that much is a non-negotiable requirement.

For network devices, the Linux PCMCIA daemon/kerneld can do pretty
much what I'd like. But networks are special in that they already have
to handle and recover from data loss and network partition.
Well-behaved apps can already cope with devices going away and coming
back, at least in the medium-term.  I dont' know of anything similar
on Unix for mounted disk devices that I'd trust quite that much.

Is that too paranoid?