> [On the topic of scsi address mapping]
> 
> To add another thing that I have see to the fray... Some systems have
> a /dev/root which is good for the GENERIC case...

kernfs provides /kern/rootdev and /kern/rrootdev, for easy access to
this.


One of the biggest problems with pinning all devices nodes to specific
targets is: how do you do it?  in the generic kernels, do you assume
that each bus should have its own set of (pinned) targets?  (that's
hell on device nodes, and in config files.)  do you assume that you
just have one bus?  (that doesn't work on some machines.)  etc.

the point of having the generic kernels work the way they do now is
that they work, easily, in just about any situation that you can throw
at them.  no matter how your drives are scattered, they _will_ be
usable.  when you config your own kernel -- and i say "when," not if,
because generic kernels really aren't meant for everyday use, if only
because they waste memory -- you can and should pin devices to the
unit numbers you prefer.

i think it boils down to: when you're trying to get the system going
with a generic kernel, be it while installing, or while recovering
from a disaster, having fixed mappings wins you nothing, and can make
life impossibly hard.


chris