GitLab

aggregate in the DDC.

Remove obsolete stitching rspecs.

I appear to have broken it while doing moonshot support...

from directory to filename when IMAGEDIRECTORIES is turned off, or
not admin|feature enabled.

Perl realpath returns undef or '' (take your pick) if the path does not
exist, rendering it pretty much useless for our check. So we run realpath
on the directory part of the path (which should exist) and then do some
other paranoid checks on the filename part (no funky chars, cannot be a
symlink).

Also, clean up some nits.

types in the images/default_typelist sitevar.

In the case the upstream server was not responding, this would slowly eat
away FDs and ephemeral ports.

The hard-coded per-model offsets aren't always correct.  Need to deal
with this in a different way, but this is a quick hack to move things along
with the Utah stitcher aggregate.

Soon, we will have images with both full images and deltas, for the same
image version. To make this possible, the image path will now be a
directory instead of a file, and all of the versions (ndz,sig,sha1,delta)
files will reside in the directory.

A new config variable IMAGEDIRECTORIES turns this on, there is also a check
for the ImageDiretories feature. This is applied only when a brand new
image is created; a clone version of the image inherits the path it started
with. Yes, you can have a mix of directory based and file based image
descriptors.

When it is time to convert all images over, there is a script called
imagetodir that will go through all image descriptors, create the
directory, move/rename all the files, and update the descriptors.
Ultimately, we will not support file based image paths.

I also added versioning to the image metadata descriptors so that going
forward, old clients can handle a descriptor from a new server.

Similar to Ubuntu 14, but new and different UIDs, and...

systemd!  systemd has made it into Ubuntu.  This affects our control net
discovery.  Our /etc/network/interfaces hack marks eth[0-9] as "auto".
New Ubuntu versions have systemd (or upstart, if your Ubuntu 15 uses
that -- our images will not) job that waits for any interfaces marked
auto to come up, and doesn't timeout for 2 mins.  This doesn't work for
us, because our use of auto is abuse :).

So, we enter a new level of abuse.  We change the systemd unit
dependencies, making the network-online.target service want (depend on)
our special ifup-wait-emulab-cnet.service (which looks for /run/cnet to
get written), instead of on ifup-wait-all-auto.service.  This means
network-online.target happens as soon as the control net is up, which is
the primary thing we're concerned about.

Now, if experimenters would ever add their own special "auto" interfaces
to /etc/network/interfaces, the rest of the startup scripts that depend
on network-online.target could possibly race (although you would think
that our control net discovery will be slower than any static
configuration...).

But for now, this solution appealed more that "faking" that eth[0-9] are
all up by writing the right files --- or by finding a better mechanism
instead of abusing "auto".

The previous commit to disable it per-interface doesn't work when forwarding
packets. Apparently the kernel looks at the "all" value when doing this.
Now we just append the settings to /etc/sysctl.conf which is present in
Ubuntu and CentOS.