GitLab

This is a copy of `doc/updating-RHL.txt' r1.2, modified on Nov 12 2003.

I'm committing it now, before I start trying to clean it up.

This change makes the configuration more portable between testbed sites.

* sslxmlrpc_server.py: A rather gross hack that needs more thought;
  pass the client IP address to the emulabserver class instantiation,
  which is passed along to the new elabinelab module ...

* emulabserver.py: A new class called elabinelab which exports some methods
  that are to be used by an inner elab. At present, the IP address of the
  client is passed along and a bunch of checks are made that restrict the
  client to the inner emulab boss node, with the credentials of the
  creator of the inner emulab. In other words, the ssl certificate of the
  elabinelab creator is placed on the inner boss, and all proxy
  operations are invoked with this certificate (as the creator) and
  only from the inner boss node.

  The elabinelab class currently exports two methods; a power method
  to power cycle an inner node; the command is handed of the power
  command, which does the permission checks. Of course, the inner boss
  does its permission checks, but ultimately, the outer boss will
  allow the power cycle only if the client is allowed to power cycle the
  node.

  The other method exported is a vlans command to setup and destroy a
  set of vlans for an inner experiment. Permissions checks are modeled as
  above, with everything passed out to new snmpit.proxy script, which
  then invokes plain snmpit.

Log `auth.info' messages (`ssh' logins) to `users.emulab.net'.

This is done because we replace ceratin user- and group-management programs
(from the `shadow_utils' package) with versions that accept uppercase letters
in user and group names.  The Linux-standard (?) versions of those utilities do
not understand the CREATE_HOME option in the `/etc/login/defs' file.  In fact,
the standard `useradd' program goes out of its way to *complain* about the
option when it is found.  The relevant source comment from `src/useradd.c':

		/*
		 * RedHat added the CREATE_HOME option in login.defs in their
		 * version of shadow-utils (which makes -m the default, with
		 * new -M option to turn it off). Unfortunately, this
		 * changes the way useradd works (it can be run by scripts
		 * expecting some standard behaviour), compared to other
		 * Unices and other Linux distributions, and also adds a lot
		 * of confusion :-(.
		 * So we now recognize CREATE_HOME and give a warning here
		 * (better than "configuration error ... notify administrator"
		 * errors in every program that reads /etc/login.defs). -MM
		 */

Rather than suffer obnoxious warnings at boot time, we just avoid the option.

(2) building the patched, Emulab-specific kernels from source, and (3) cleaning
up the `lilo.conf' file itself after the kernel building and installation
process.

testbed programs rely on libraries found in `/usr/local/lib'.

hack to deal with inner vs outer control network.

except that it's been modified to the point of making the serial console usable
in the University of Utah Emulab.

The exact contents of this file vary according to the installation environment
in any case.

although the exact contents of this file vary according to the RPMs that are
installed when the system is created.

Initial checkin.  This version is from a virgin Red Hat Linux 9 installation, except that it spawns agetty on ttyS0.

* snmpit: When ElabInELabis true, use the routines in the new
  snmpit_remote.pm library for setting up and tearing down vlans for an
  experiment. At present, only these two operations are proxied out to
  the outer emulab.

* snmpit_remote.pm: A new little library that uses the XMLPRC server on
  the outer emulab to setup and destroy vlans for an inner experiment.
  This code is used from snmpit (see above).

* snmpit_lib.pm: A couple of minor changes for the server side of the
  proxy operation.

* snmpit.proxy.in: A new perl module that is invoked from the RPC
  server.  This proxy sets up and tears down vlans for an inner elab.
  The basic model is that the container experiment will have lots of
  vlans for various individual experiments running on the inner emulab.

* swapexp: A couple of minor elabinelab hacks.

* tbswap: For elabinelab experiments, reconfig/restart dhcpd when
  tearing down the experiment, and call out to new elabinelab script
  when setting up an elabinelab experiment. There is no provision for
  swapmod at this time.

* elabinelab: A new script to create the inner emulab. Does all kinds of
  gross DB stuff then more gross stuff on the inner ops and boss.

library to call out to the outer emulab to do the actual power cycle.

the initial DB contents are loaded. Create project, the users, the
groups, etc.

   This way we can use frisbee as a fast network copy.

2. In imageunzip, don't truncate the output file until we have performed
   basic validity checks.  Otherwise, if you happened to reverse the order
   of the imagefile and the output disk, it would truncate the imagefile
   and THEN fail.

3. Strip binaries when they are installed.

Contributed by Keith Sklower at Berkeley.

port, and the location of the SSL certificate used to talk to the server

If you install GRUB as the boot loader for Linux (actually Xen),
it seems to leave behind enough LILO residue to screw up imagezip.

dying to see whats going on. These changes can go into new images; it will
be ignored in non ElabInElab experiments;

rc.inelab: A terrible hack that I now regret. The whole point of this
script to allow inner boss/ops to 1) play by the standard state machine goo
(ISUP) for the outer emulab. That way os_setup, node_reboot, etc all work
when dealing with setting up the inner elab. 2) Since the inner
control can fall on any of the interfaces (as picked by assign), we
need to still do ifc/route configuration using tmcd. At the time I did
this, I expected that swapmod/swapin/swapout would be possible and
that boss/ops would be coming in on different nodes using different
interfaces for the inner control net, and that I would want to ask
each reboot. Well, I think swapping an inner emulab is a really long
way off, maybe so far off I'll be retired by that time.

rc.mkelab: A lot of very ugly code that turns a node into an inner ops
or boss. I won't bother to describe this code other than to say the
operating model for this rev is to create a fully functional inner
emulab based on the DB state provided by the outer emulab. Currently,
the current testbed software is expected to be in the /proj tree, and
a full boss/ops install is done, followed by all kinds of
customizations (like building accounts for members of the project,
etc) to make it a fully function emulab with just a single project;
the project in which the inner emulab was created.

resource tables.

New routine to return whether experiment is an ElabInElab experiment.

frisbee MFS with a zillion possible devices.  Requires that /sbin/mknod
be installed in the MFS as well.