GitLab

This reverts commit e709749894d85768044ce0631588385ad06ac06b.
Don't need it after figuring out imagezip

This completes linkdelay support for linux, dropping IMQ in favor
of IFB and netem (so that we do not need to build a special kernel).

Since the pcpg-i2 nodes are so very flaky, lets try something that
does not require them to rebooted or imaged!

The key change is that on these remote nodes, we do not bridge the
physical control interface to the VM control interfaces. There is no
point since there are no routable IPs we can use, nor is there a
192.168 network that would be useful.

However, we still want to give the VMs their 192.168 address and we
still want multiple VMs on the same host to talk to each other, and we
we still want the VMs to be able to access the outside world with NAT.
So we still create the xenbr0 bridge, and give it the router address
(192.168.0.1). Any traffic heading out will be NAT's as normal, and
you can ssh into the VM using the physical host IP and the per VM
sshd port number.

Prior to this commit, XEN guests disks were single partition, no MBR,
the bits dumped into the lvm. This makes a snapshot of a XEN node,
look completely different then a physical disk image, especially if
users want more disk space (mkextrafs) inside the guest, and then want
to take a snapshot of that, and then run it on a physical node (which
was not possible).

With these changes, guests now use the same MBR layout as our version
two MBR, which makes them interchangeable with physical disk images.
In fact, the goal is to be able to switch back and forth as needed,
based on physical resource availability.

and let it throw an error later if all space is consumed. After
all, each VM can use an arbitrary amount of disk, and eventually
we will have to track it with assign.

causing the ciscos to drop the packets with an undersize error.

Make use of (localized) pygrub to find the boot kernel and ramdisk
inside the guest. For BSD guests, we mount the FS and look for a
xen kernel or then the standard kernel. We faill back to the old
method if we cannot find a kernel.

Localize guests just like we do with slicefix.

For ubuntu kernels, we have to fix the ramdisk to local the xen
block driver (which is in the ramdisk, but not used). 

Use IFBS for BW capping in dom0. Guests can still do their own link
shaping, but we cap bandwith outside the guest.

A bunch of non functional gre tunnel stuff based on the openvz code.
This is all going to come out and be replaced with openvswitch, but
want it in the repo.

script, which I am using from the XEN code to localize a guest.

whacky slice images.

a single slice image, since we can now pull the kernel (ramdisk) out
from the guest filesystem (using pygrub for linux, or just mounting
BSD filesystems). This is a lot faster and easier to deal with. I
added an option to the newimage page so that people can set this, but
in general we need a better way to guess that we need it. Always set
for EC2 images.

Take care of an iSCSI race condition in Fedora15.

kernel command line, according to new loadinfo variable returned from
tmcd. This is so we can tailor the amount of memory per node type.

Mike's instruction. Faster, better, slicker then snot.

The Storage Subsystem Whack-a-Mole game continues! I smack down the
bugs in the LVM version, and up pops ZFS problems. Wham! Down goes
ZFS, up pops Vinum...

Use ZFS rather than gvinum as the LVM on 64-bit FreeBSD. One funkiness:
if the blockstore has a mountpoint (implying a filesystem), we create a
ZFS filesystem. If they do not, we create a ZFS zvol (a block device).
The filesystem has much better performance than the zvol, hence using it
when we can.

In gvinum, if all disks are the same size, stripe them rather than
concanonate them. For us, this would only happen on the d820s if you
said NONSYSVOL (and you aren't using ZFS).

Support the unspecified (zero) length blockstore which means "as much
space as is available".

Handle mounting of SAN-based storage. We cannot just throw the mount
into /etc/fstab since the SAN device would not exist when fsck or
mount is done at boot time. You could probably configure it to work
on Linux, but I didn't see how on Ubuntu. The iSCSI initiator and
associated infrastructure on FreeBSD is so primitive that there is
no way you could do it automatically.

Add a "fullreset" command to rc.storage for debugging help. It attempts
to completely teardown the storage no matter what state it is in. The
regular "reset" command is conservative and quits at the first sign of
trouble.

using the same routing trick we us on the openvz image. But
lets save it in case we figure it out.

This replaces the previous revision that did this with an extra bridge
and IMQ device. We like IFBs better now! Note that we enforce just the
maximum BW; all other shaping is done in the current way, by mapping
an IMQ into the container and doing linkdelays as we do now.