Sun, 16 Mar 2014
Just for my own memory (attaching 400G to vm named git0):
pool-define-as hdd --type dir --target /mnt/libvirt pool-start hdd pool-autostart hdd vol-create-as hdd gitdisk 400G attach-disk git0 /mnt/libvirt/gitdisk vdb --persistent --driver qemu --subdriver qcow2
The driver and subdriver options are essential. Otherwise the disk is attached as raw disk and the guest will only see several kilobytes (kb) of disk not the full 400G.
Sun, 02 Feb 2014
I own multiple root servers distributed in different data centers across the world. Each server runs libvirt and serves multiple VMs in private subnets. The goal of this exercise is to have the virtual machines be able to securely talk to each other. The default configuration for libvirt enables NAT from a private network (192.168.122.0/24) to the Internet.
This post describes step-by-step how to establish secure tunnels between three hosts (called b0, c0 and d0 respectively) but could very well be extended further. It enables hosting root servers in multiple heterogeneous data centers and not worry about traffic between VMs. They will be able to communicate with each other securely, just like being on one single trusted router.
To establish the secure connection openswan is used and the following instructions assume and are tested with Debian/wheezy. Openswan as of writing this is at version 2.6.37-3 and libvirt is at 0.9.12-11. No further packages should be necessary (apt-get install openswan libvirt-bin).
First of all different subnets need to be assigned to each host. For this use virsh net-edit default and change any occurrence of 192.168.122 by a unique number. The following is my own configuration used for the further examples:
Host Public-IP Subnet ---- --------- ---------------- b0 10.2.2.2 192.168.122.0/24 c0 10.3.3.3 192.168.123.0/24 d0 10.4.4.4 192.168.124.0/24
Next restart libvirt and verify using ifconfig/ip that the configuration has been applied. If any VM had a static IP address assigned it might need to be changed.
In the next step ipsec needs to be configured on each host. Open /etc/ipsec.conf in your favorite editor and add the following lines for each pair of hosts needing to directly communicate with each other:
conn b0-to-c0 authby=secret auto=start type=tunnel left=10.2.2.2 leftid=10.2.2.2 leftsubnet=192.168.122.0/24 right=10.3.3.3 rightid=10.3.3.3 rightsubnet=192.168.123.0/24 conn b0-to-d0 ... left=10.2.2.2 leftsubnet=192.168.122.0/24 ... right=10.4.4.4 rightsubnet=192.168.124.0/24 ... conn c0-to-d0 ... left=10.3.3.3 leftsubnet=192.168.123.0/24 ... right=10.4.4.4 rightsubnet=192.168.124.0/24 ...
The same file can be used for every host because non-existing interfaces will just be ignored. The left/right designation is arbitrary.
If multiple gateways exist on the host, the following might be useful (in addition to a listen 10.2.2.2 in the config setup section):
Next up configure authentication tokens between the hosts. It is recommended to use host keys for the connection, but for the sake of simplicity this uses the same preshared key between all hosts (DO NOT DO THIS IN PRODUCTION). Edit /etc/ipsec.secrets and add:
10.2.2.2 %any: "s3cr3t" 10.3.3.3 %any: "s3cr3t" 10.4.4.4 %any: "s3cr3t"
Restart ipsec (/etc/init.d/ipsec restart) on both ends. Execute ipsec auto --up b0-to-c0 on each side and use ipsec look and /var/log/daemon.log to verify that the tunnel comes up.
In theory it should be possible to ping both ends of the tunnel. But libvirt installs a pesky NAT rule by default to circumvent this execute the following (can go into /etc/rc.local):
iptables -I POSTROUTING 1 -d 192.168.0.0/16 -j RETURN -t nat
This effectively disables the NAT rules for the whole private subnet starting with 192.168. Now try:
ping -I 192.168.123.1 192.168.122.1
If it does not work see above how to verify that the tunnel came up. If this is inconclusive tcpdump 'udp port 500' and on port 4500 should show in- and outgoing packets. Verify that all IP addresses are correct and no other firewall rules are blocking communication on these ports.
It is necessary to specify -I 192.168.123.1 for ping because ipsec doesn't install a default route for the network connection (different protocol layer). Thus execute a final:
route add -net 192.168.0.0/16 virbr0
on the host to route any traffic to and from the virtual machines correctly. Now ping should work without the -I parameter. Make it permanent by adding it to /etc/rc.local.
Communicating from within the VMs should now just work. Verify that pinging around different network produces the desired results. If not verify that IP forwarding is actually enabled. /etc/sysctl.conf should contain the following line:
As always with security, cross check this against other documentation and verify using tcpdump etc. that the traffic is actually encrypted.