Someone else also made a related point about device permissions.
Personally, I favour a kernel-generated /dev, for two major reasons.
Firstly, it makes standardisation of device names easier, and enforces
that standardisation more widely. Secondly, it avoids the terrible
clutter that tends to accumulate in /dev by having only supported
devices represented, and also avoids the problems that occur when a
supported device has no node in the filesystem.
The kernel would necessarily have to allow chmod(2) and symlink(2) to
work in the device filesystem, and possibly also mkdir(2). The
required permissions and symlinks can then be easily reconstructed by a
shell script at each boot. A kernel-generated filesystem would also
have the advantage of allowing for convenience devices that can be used
to chmod groups of devices -- /dev/unallocated_ptys, for example.
And look at it this way -- if a kernel-generated /dev doesn't get into
the standard kernel, someone -- possibly me -- will write a devfs
module for userfs. It would be feasible to have a stripped-down
`physical' /dev, with the kernel-generated one mounted over the top of
it at boot time.
-zefram