Re: devfs - why not ?

[Alexander Viro]

On Thu, 13 Apr 2000, Peter Samuelson wrote:
> [Alexander Viro]
> > Funny, that... Where did I hear about something similar? Ah, yes.
> > automounter. Somebody, tell Richard that Sun had stolen his idea many
> > moons ago...
> 
> I can't speak for Richard of course, but I honestly had never thought
> of (ab)using autofs for device nodes.  If it's been mentioned here
> before, it must have been when I wasn't reading l-k.  Certainly the
> autofs protocol could provide at least some of the functionality of the
> devfsd protocol.

Umhm. Notice that use of automounter allows you to use per-driver
filesystems (with ramfs or equivalent acting as library for all metadata
stuff, so that driver would only have to define its nodes and tell the
layout of tree). That way procfs-located part of tree could also go there,
killing all granularity problems and devfs vs. procfs wars.

> > BTW, _what_ was that wrong with using autofs for that instead of
> > reinventing the wheel?
> 
> Perhaps because adding a devfsd mechanism to devfs is more lightweight
> on the kernel side.  fs/autofs/ is 68 blocks on my box, fs/devfs/ is
> just under twice that.  I know obj side != source size, but I'd guess a
> lot less than half of devfs is devoted to the devfsd protocol....

autofs also contains a thing or two that can be cleaned up. And it's 42k
on this box (counting comments).