cloudstack-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dave Scott <>
Subject Re: Support pure Xen as a hypervisor follow-up
Date Fri, 16 May 2014 14:22:59 GMT

On 14 May 2014, at 09:53, sebgoa <> wrote:

> On Apr 9, 2014, at 2:37 PM, Dave Scott <> wrote:
>> Hi,
>> Following up from Tim's "Support pure Xen as a hypervisor" proposal last month[1]
I'd like to start working on this and maybe even make a little bit of progress while I'm at
CCC in Denver.
>> Helpfully James Bulpin managed to get CS + libvirt + xen to start an instance in
a simple configuration. Although the patches[2] are not intended to be production-ready :)
they help highlight some of the areas we need to change.
> Dave, just to let you know that Tim has done some important "refactoring" to split up
XenServer hypervisor in CS between Xen and XenServer. That way we could keep using xapi for
XS but start moving to libvirt for Xen.
> Tim worked in the xen2server branch (don't ask about the name, I messed it up…:) ).
> Would be nice to see some of the libvirt stuff in that branch to handle a new driver
for Xen.
> Since the two hypervisors will be split up, we could still drop in some early libvirt
patches to handle Xen and put this in 4.5 as a wip.

Thanks for the links.

I’m slowly building up a set of patches here:

I think once I’ve gotten to a stable-ish point I’ll rebase on top of Tim’s branch.

So far I’ve
* changed the hypervisor detection to use ‘virsh capabilities’ in one place and ‘cat
/sys/hypervisor/type’ in another. Thinking about it again, I think it’s probably best
to standardise on /sys/hypervisor/type since that will succeed irrespective of whether the
libvirtd service is chkconfig’d on or not.

* in the python cloudutils system setup stuff isKvmEnabled() has become isHypervisorEnabled()

* added a XenLibvirtDiscoverer similar to the LXC one

* fixed what I believe is a race in sshExecuteCmdOneShotWithExitCode (which seems to hit me
every time, I don’t know why other people seem to be immune from it): see CLOUDSTACK-6621
and review board request 21261

* added the new hypervisor to hypervisor.list and system.vm.default.hypervisor, so it appears
in the UI properly

* registered a system VM template in the database, using the same qcow2 image as KVM

For my test host I’m using a XenServer nightly snapshot which comes with a nice modern xen
and kernel, and is easy to install bleeding-edge libvirt on top. I had to tweak the kernel
configuration and the network configuration but I’m hoping to make it work out of the box
in future.

When I deploy my ‘datacenter’ the discovery phase works, the agent connects and looks
healthy in the logs and the UI. The next step is to figure out why the system VM template
isn’t being copied to primary storage — for some reason the copy isn’t being attempted
but I can’t see any obvious reason why.


> -Sebastien
>> Some of the areas are:
>> 1. hypervisor detection
>> Where we currently look for KVM specifically ("lsmod | grep kvm") we could switch
to either detecting any Linux hypervisor (by reading /sys/hypervisor/type) and assuming if
a hypervisor is present then we can use libvirt on it (is this a fair assumption?) Or we could
white-list “kvm” or “xen”. Or we could query libvirt directly (perhaps via 'virsh
>> 2. fiddling with the domain.xml
>> When starting a domain via libvirt the XML configuration has hypervisor-specific
stuff in it. Some of this is easy to change like:
>> <domain type='kvm'>
>> obviously becomes
>> <domain type='xen'>
>> and
>> <emulator>/usr/libexec/qemu-kvm</emulator>
>> should probably be
>> <emulator>/some/other/path/qemu-dm</emulator>
>> Some is a bit more invasive (to the VM) such as the virtual hardware type should
be switched from "virtio" to "xen" (and the block device in Linux will change from /dev/vd*
to /dev/xvd*) and we'll have to either implement or work around the lack of
>> <channel type='unix'> ...
>> -- I presume this is a control channel into the system VM. Perhaps we could implement
this in libvirt/libxl using vchan?
>> 3. system VMs?
>> It would be very convenient if the system VM images could work on both xen and KVM.
This is probably doable as long as we don't bake in virtual hardware specific information
(such as /dev/vda) in the image. We could use the qcow2 format in both cases. What do you
>> … and I’m sure there’s more.
>> Anyway, feedback would be welcome. If anyone else in Denver wants to chat, then come
grab me later!
>> Cheers,
>> Dave Scott
>> [1]
>> [2]

View raw message