GitLab

We also make sure that all dirs in /var/emulab are really dirs and
not symlinks. Having /var/emulab/vms be a symlink (left over from
a vnode setup) caused vast confusion when trying to use the new
images for vnodes.

Note: you really, really should not image the vhost in an experiment
with vnodes. Prepare in such an environment is destructive and you
won't get a clean image anyway (due to LVM state). Swap the Xen image
in on a node directly and image that.

You do this by setting the nodeos of "xenhost", ala:

    set elabinelab_nodeos("xenhost") XEN44-64-BETA2

    set elabinelab_nodeos("boss") FBSD100-64-STD
    set elabinelab_nodeos("ops") FBSD100-64-STD

Needed for pygrub.

Stick the pid file into /run/sendsigs.omit.d. If xenstored is gone, any
attempt to use "xl" will hang un-killably, at least on Xen 4.4.

Note that this does not help shutdown to single user, which runs a different
"killprocs" script as it enters single-user. That kills all processes with
no omissions (/etc/rc1.d/S30killprocs)! We may want to consider running the
Xen start script following that to at least re-start xenstored (and which
will free up any hung processes).

I was having to hack grub.cfg to make this work anyway since I don't know
how to properly configure /etc/default/grub to use ttyS0 for regular Linux
and hvc0 for Linux under Xen.

This change would mess up the console on Xen 4.1 and Xen 4.3 image too.

we rely on that version of the tools in the code. Need to be
fixed at some point.

experiment.

I've seen about a dozen cases in the last two days of bad rspecs,
but not catching the fact that we cannot get peer ip from the
manifest on the other aggregate.

only visible to admin users. The control flow is not how we
want it, and I don't want to spend too much time on it until
Jon gets his stuff committed, which I think will include some
better ajax code.

Magical udev and upstart overrides.

Also, more tweaks for Xen 4.4.

only one of the tags, not all of them.

then 10 connections from the same source in the last 100 seconds. Note
this is just to the physical host itself, it does not affect traffic
to the containers.

pidfile is gone, or the process is gone; both imply
the container is gone, so don't throw an error.

actually being done.

1. Change hackwaitandexit on the client, to return zero if the guest
   has not finished setting up. We used to treat 30 seconds as too
   long must have failed, but this is really not the case, especially
   on busy machines.

2. Fix up vnode_setup exit code handling, we were losing non-zero
   status cause of not shifting it down, and so failures were never
   being reported.

   New: If the vnode setup does return failure, set its event state to
   TBFAILED to cut short the wait in os_setup and the IG monitor
   process. On the surface this seems like an obviously good idea, but
   I'm sure it will come and bite me when I least expect it.

3. Change GeniAggregate Start/Restart to ignore vnode_setup failures,
   and let the monitor watch for TBFAILED or timeout. There are just
   too many ways for it to fail, and we want to allow vnodes that did
   not fail to set up normally, and give the user the choice to
   restart the ones that failed.

4. Don't let frisbee run forever, protect with timeout. I need to use
   Mike's new -T option, but not till I actually get new frisbee
   pushed out.

If we don't get any messages from our server after <sec> seconds,
we exit. Odd that this didn't already exist, huh?

I re-purposed the -T option which used to be used to specify a prefix for
a trace file. Since I am the only one that ever used that option, I don't
feel bad about changing it. Traceprefix is now specified with -U.