The design of mdev (mini-udev for busybox).

[Rob Landley]

By the way, for those of you curious, the basic idea behind mdev is to do a 
stripped down udev for busybox.  My TODO entry about this reads:

  mdev
    Basic merge.
    Config file
      regex support
      username:group (string, not just uid:gid)
      zap cmp
      local_errno
      bb_xasprintf();
      add dir/
      add shellout
    syntaxes:
      symlink -> symlink
      dir/ (and dir| dir=)
      thing uid:gid 777
      /dev/initctl - a named pipe?
    Is the "" escaping useful here?  (Spaces?)

And translating that for everybody else:

I have some existing C code (a simple C port of the shell script posted last 
message).  The "zap cmp, local_errno, bb_xasprintf()" bits are just about 
that code being busyboxified.

There are two modes for mdev.  One scans /sys and creates devices for 
everything it finds.  (That'll be "mdev -s".)  The second is called from 
hotplug and looks at the environment variables to see what new device just 
cropped up/went away, and adds or removes it.  (That's mdev with no 
arguments.)  Both need to wash its action through the config file.

Why a config file?  Because we have to specify ownership and permissions on 
newly created devices.  The config file needs to be as sparse as possible and 
easy to parse with a small amount of code, but we do need one. 

A basic device entry looks like:
devname uid:gid permissions

So for example:
tty3 0:0 600

It dumps everything in the same directory (no devfs-style loop/0 nonsense), 
and works based on the names the kernel gives stuff.  The goal here is to be 
_simple_.  It'll need some kind of regex support to allow a single rule to 
match tty[0-9]* or hd[a-z][0-9]*.  For each device, rules are scanned top to 
bottom and we stop at the first hit, so you can specify a tty3 and then a 
default tty[0-9]*.

This is the minimum we need.  Anything beyond this should be a CONFIG_FEATURE.

If we're feeling posh, we should look up UID:GID in /etc/passwd and /etc/group 
so we can use names rather than just numbers.  That's probably a configurable 
feature.

Since mdev will generally be run on an empty tmpfs (and may in fact mount 
tmpfs itself), we may also need to create a few directories (pts, shm) and 
named pipes (initctl) and such in there.  The config file can specify these 
things (with ownership and permission) by appending the characters from "ls 
-F".  For example:

shm/ 0:0 1777  # Yes, setting the sticky bit needs to be supported.
initctl| 0:0 600

And for symlinks, we use " -> " again just like ls does, so:
cdrom -> hdc

Non-device lines like these would only be matched on the initial scan (-s), 
not when looking for a description of a new device.  And that's a second 
configurable feature: the user could also mount tmpfs from a shell script and 
do that setup before calling "mdev -s".  It is nice to have it all together 
in one config file, though.

In future we might also need a shellout syntax, and I suggest:

hdc 0:24 660 $/etc/mdev/mkhdc.sh # create symlink, call dbus, etc.
pts/ 0:0 755 $mount -t devpts /dev/pts /dev/pts

It'll automatically inherit all those nifty environment variables.  (We can 
synthesize one: MDEV_SYSFS which is the path to the sysfs entry we're 
servicing.  That way the cold boot scanning case can cope.)

Ordinarily we don't distinguish between char and block devices, but if a need 
to do this does crop up I suppose could always append @b or @c, so for 
example a catch-all rule for otherwise unspecified block devices could be:
.*@b 0:0 600

Anyway, this is what I'm thinking about.  Comments?

Rob
-- 
Steve Ballmer: Innovation!  Inigo Montoya: You keep using that word.
I do not think it means what you think it means.