libvirt is an open source API and management tool for managing platform virtualization. It is used to manage Linux KVM and Xen virtual machines through graphical interfaces such as Virtual Machine Manager and higher level tools such as oVirt. In this case the backend is KVM.
Since libvirtd distributes resources to guest machines a tight control needs to be placed on the guests to prevent the host from becoming unmanageable. Luckily through the use of CGroups this can be accomplished.
Networking is a tricky topic as there are several ways to do this in libvirtd. My personal choice is to make the host completely transparent in regards to the networking and that is bridge mode. There is also routed and NAT either way exposes the host's IP.
Routed mode also requires additional configuration in the network hardware or on the client machines and NAT requires port forwarding within the host machine. I will not cover either of these in this wiki.
Privilege separation is very important. In the event that a guest gets compromised and the attacker finds a way to break the virtualization's jail, they should not immediately gain root privileges.
Luckily this is already covered by default (at least with Fedora 12 installs) in which libvirt will drop it's privileges to the qemu user for every guest that it spawns. Each guest is also automatically given it's own SELinux label. This will further isolate each guest.
Regardless of the type of network that is chosen for guests all of their traffic goes through the forwarding chain of iptables. While I really strongly advocate against any default allow policies on anything, this is one place where I personally am willing to make an exception. To allow guests to talk to the network make the following change to the default IPTables rules:
:FORWARD DROP [0:0] --- Replace with --- :FORWARD ACCEPT [0:0]
At a high level, CGroups are a generic mechanism the kernel provides for grouping of processes and applying controls to those groups. Tunables within a cgroup are provided by what the kernel calls 'controllers', with each controller able to expose one or more tunable or control.
When mounting the cgroups filesystem it is possible to indicate what controllers are to be activated. This makes it possible to mount the filesystem several times, with each mount point having a different set of (non-overlapping) controllers.
The provided controllers are:
Under KVM all of these are not supported, however cpu, devices and memory are. Which are in my humble opinion the most important for our task.
Before creating your first guest there are a few things that I like to do. These include setting up additional storage pools and configuring networking. This section covers anything that hasn't been done in the config files and I'll save the majority of the networking information for it's own section as most of it is done outside of the virtualization console.
Since I use bridge networking and that is the only way I want my guests to talk to each other and the rest of the world, I want to get rid of the default network which performs NAT'ing so that the VMs don't suddenly get a second network interface that I don't know about.
To do this I'll want to drop into a virsh shell and issue a few commands. You can see an entire session of me removing the default network here:
[[email protected] ~]# virsh Welcome to virsh, the virtualization interactive terminal. Type: 'help' for help with commands 'quit' to quit virsh # net-destroy default Network default destroyed virsh # net-undefine default Network default has been undefined
Now that our networking problem is taken care of, on to additional storage pools. By default one storage pool (a directory) exists. This can be found at "/var/lib/libvirt/images". There is nothing wrong with just using the default for disk images. The only reason I keep a second for disk images is that I like to keep all of my VMs on an encrypted partition all to themselves.
Since I need one of my VMs at boot and won't be able to remotely provide a key until after that VM has booted it needs to be on an unencrypted partition which I still use the default for.
The rest of the VMs will live in a storage pool named "secure" mounted at "/var/lib/libvirt/images/secure". I'm using the the subdirectory of "/var/lib/libvirt/images" so that I won't have to add additional SELinux policies which is potentially very messy and could present additional security risks.
First we need to create an XML file to define the storage pool. Create the following in a file named "secure.xml":
<pool type='dir'> <name>secure</name> <target> <path>/var/lib/libvirt/images/secure</path> <permissions> <mode>0700</mode> <owner>0</owner> <group>0</group> </permissions> </target> </pool>
Next you need to tell libvirt about the pool. You can do this with the following series of commands:
[[email protected] ~]# virsh Welcome to virsh, the virtualization interactive terminal. Type: 'help' for help with commands 'quit' to quit virsh # virsh pool-define secure.xml Pool secure defined from secure.xml virsh # pool-start secure Pool secure started virsh # pool-autostart secure Pool secure marked as autostarted virsh # pool-refresh secure Pool secure refreshed
My network is a tad unusual in that I have VMs on the same box that need to be on different network segments. Each of these network segments comes back to my virtualization host through VLANs.
While Linux has long supported tagged VLANs and been able to hop on different segments as defined by the system administrator, libvirt's virtual network does not support trunking. This means that the interfaces need to be setup on the host and bridged through to the guest. I'm going to provide a working example that passes VLAN20 into a guest as it's native network card.
I'm assuming the trunked port on the host is eth1, please adjust the following configuration to match what your network has. This example is also specific to the RedHat architecture. The configuration will be different than it is here (and I have not documented it).
You will need to make sure that the package
bridge-utils is installed on the
server. I think this is a dependency of libvirt but I could be wrong so it's
best to double check.
First eth1 needs to be configured (or unconfigured in this case). This can be
done by editing
/etc/sysconfig/network-scripts/ifcfg-eth1. The entirety of
the files contents should be replaced with the following:
DEVICE=eth1 BOOTPROTO=none HWADDR=XX:XX:XX:XX:XX:XX ONBOOT=yes
Next we need to tell the server about the VLAN. Create the file
/etc/sysconfig/network-scripts/ifcfg-vlan20 if it doesn't exist already, and
populate it with the following:
VLAN=yes VLAN_NAME_TYPE=VLAN_PLUS_VID_NO_PAD DEVICE=vlan20 PHYSDEV=eth1 BOOTPROTO=none ONBOOT=yes TYPE=Ethernet BRIDGE=br20
Note above that I've specified which bridge the vlan device is going to be connected to. The 20 in br20 can be changed to any other number you'd like, however for simplicity sake I like to have the bridge reflect what VLAN it is connected to.
The last step is to configure the bridge interface. This is needed so that a
guest may make a connection to it. Create the file
/etc/sysconfig/network-scripts/ifcfg-br20 if it doesn't already exist and
replace the contents with the following:
# br20 - vlan20 - Some non-existant subnet DEVICE=br20 NM_CONTROLLED=no ONBOOT=yes TYPE=Bridge BOOTPROTO=static IPADDR="10.13.37.30" NETMASK="255.255.255.0" GATEWAY="10.13.37.1" DEFROUTE=yes DNS1="10.13.37.1" DNS2="10.13.37.2" IPV4_FAILURE_FATAL=yes IPV6INIT=no NAME="Some Bridged Network"
Careful with this one, the value for TYPE is case sensitive. Putting 'bridge' or 'BRIDGE' will cause the scripts to ignore that this is a bridge and the interface will not come up.
In my experience adding a VLAN (and it's the same for each one after) adds a couple seconds to the boot process each time the server boots up. If uptime/downtime is important and you need servers coming back up as fast as they can a large number of VLANs can significantly delay the process.
When creating a guest you'll want to add the following line to the command to give the guest an interface on the appropriate VLAN:
You can also define a network in virsh so that it'll show up to utilities like
virt-manager by putting the following in an xml file and running
net-define bridge20.xml assuming you name the file
net-autostart LANBridge which assumes you use the same name that I've chosen
for this bridge below:
<network> <name>LANBridge</name> <uuid>5ba71de3-c507-4512-93c3-65fd3a38e112</uuid> <forward mode="bridge" /> <bridge name="br20" /> </network>
Things you'll need to determine before creating a new native KVM guest:
I've found that headless linux guests don't need a whole lot of RAM to function well. Depending on the function of the server I usually give my guests between 128 and 384Mb of RAM.
Disk space will have to be determined on a guest by guest basis, however I do like to the qcow2 image for my disk images for a variety of reasons. Snapshots, compression, encryption, and copy-on-write images are all handy little features of the qcow2 image format.
There is a small performance hit, I've measured that a qcow2 image is about 8% slower on reads, and about 12% slower on writes. The added bonuses however more than make up for that for me. You'll have to decide on your own how you'd like to handle it.
I have never used more than one core for a guest as they've never needed it, I've also heard some things that I've never checked into that some guests become unstable with more than one core defined. You'll have to figure that out on your own.
Networks need to be created before the guests. If using bridging (like I always do) then you need the bridge created and figure out which ones you want available on your guest.
The OS is less important and can even be omitted, however libvirt has some internal optimizations for certain OS's so if it's available its best to list what type of OS is going to be run in the guest during the initial creation.
I don't normally need a graphical environment from my guests, preferring to administer from the command line as it is more powerful and less resource hungry. In some cases it is unavoidable. You'll notice in the example I give that the creation commands has the following in it:
--graphics none --serial pty --extra-args="console=ttyS0 serial"
If your going to be using graphics you will want to omit those two options and replace them with:
This will allow you to VNC into the guest (there is no authentication or encryption although these can be configured. These are not documented here.)
Finally the name of the guest, this can not contain spaces and should not contain any special characters as they may interfere with issuing commands. I like the names to be short and descriptive referring to either the name of the server or the primary purpose.
The following command creates a Fedora 16 guest with the name 'example', with 1 CPU, 512Mb of RAM, no graphics, a network card on the LANBridge network, and a 20Gb raw sparse hard disk image.
virt-install --connect qemu:///system \ --name example \ --description="This is an example VM intended for the wiki" \ --ram 512 \ --arch=x86_64 \ --vcpus=1,maxvcpus=2 \ --check-cpu \ --os-type=linux \ --os-variant=fedora16 \ --hvm \ --graphics none \ --serial pty \ --location="http://mirror.chpc.utah.edu/pub/fedora/linux/releases/17/Fedora/x86_64/os" \ --extra-args="console=ttyS0 serial ks=http://example.org/kickstarts/ks.cfg" \ --disk path=/var/lib/libvirt/images/example.img,format=raw,size=20,sparse=true \ --network=network:LANBridge