sockfs version alpha2 - permissions filesystem for privileged sockets

               Copyright (c) 1997-1998, Malcolm Beattie

    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation; either version 1, or (at your option)
    any later version.

sockfs is a pseudo-filesystem which allows the setting of
owner, group and permissions for each reserved internet-domain
port (i.e. family AF_INET, ports 1-1024). When mounted,
directory entries for each reserved port appear (named 1, 2, 3,
etc., rather like the procfs entries for each process named by PID).

The filesystem replaces the kernel privilege check for binding
reserved ports. Instead of the default check which allows only
root to bind to reserved ports, it checks whether the
appropriate entry in the sockfs filesystem is writable by the
process attempting the bind(). Owner, group and "other" bits
are checked just as for ordinary file permission checks. The
superuser (in fact, the fsuser just as for ordinary filesystems)
is always granted permission. The filesystem allows the owner,
group and permission bits to be changed by whoever has write
access to the root of the mounted filesystem (usually root alone).

The filesystem can be configured either to have the privilege
check restored to the default one at unmount time or to preserve
the permissions across mount/unmount (in which case the permissions
check only reverts when the filesystem module is unloaded).

INSTALLATION

(1) Extract the files from the sockfs distribution anywhere away
    from the kernel source tree. Apply the enclosed patch
    sockfs.patch to your kernel. The patch was made against
    2.0.35. Slight changes made in the networking code between
    2.0.29 and 2.0.32 will probably stop this patch applying to
    earlier kernels, although it should be pretty easy to fix up.
        cd /usr/src/linux
        patch -p1 < /foo/bar/sockfs.patch

(2) Rebuild your kernel with the new patch applied (and maybe
    rebuild your modules too in case module versioning kicks in).

(3) cd back to the sockfs directory and type
        make

    The make will probably bomb out when it tries to symlink the
    module into the kernel source tree somewhere (a side effect
    of the hack used to persuade the kernel's makefiles to allow
    a build away from the kernel source tree). However, it should
    have built the module sockfs.o by that time which is all you
    need. If the Makefile doesn't work then take a look at the
    file "build" which contains the three commands necessary for
    one of my configurations to build it.

(4) Put the sockfs.o module in an appropriate place (for example,
    /lib/modules/2.0.35-2/net/sockfs.o if you're using kernel
    2.0.35 release 2 on a Red Hat 5.1 system) and reboot on the
    new kernel.

USAGE

You can decide at module-load time whether you want the new-style
permission checks to persist across umount/mount (until rmmod time)
or whether you want the permission checks to revert to the default
each time the filesystem is unmounted. To do the former, load the
module with
    # insmod sockfs.o permanent=1
To do the latter, omit the permanent=1 argument.

Now mount the filesystem somewhere: /sockfs is a good bet.
    # mount -t sockfs sockfs /sockfs
If you now do
    # ls /sockfs
you'll see an entry for each reserved port. You can use chown,
chgrp and chmod to make any given reserved port available for
binding by a given user/group. The permissions check for binding
looks only for writability of the entry: the other bits (readable,
executable, setuid, setgid and sticky) are not used by the
permissions check but are stored as-is by the fileystem.

One final technicality: the filesystem needs only 6K to store the
owner/group/mode information (1024 structures each holding a
uid_t, gid_t and mode_t). Usually, the module delays allocating
the memory until the filesystem is mounted. If your memory is
severely fragmented, the allocation could fail. In fact, this is
very unlikely but future versions of the filesystem may cope with
other socket families or such like. You can append the argument
    alloc_now=1
to the insmod command to force the module to allocate the
necessary memory at module-load time instead of delaying until
the filesystem is mounted.

CHANGE HISTORY

21 Oct 1998  Version alpha2
  alpha1 turned out to be a bit broken. This one may be better,
  although I haven't even compiled it so, then again, it may not be.

16 Sep 1997 Version alpha1

Malcolm Beattie
mbeattie@sable.ox.ac.uk
21 October 1998