This page describes how to ensure a single disk cannot be used by more than one running VM at a time, across any host in a network. This is critical to avoid data corruption of guest files systems that are not cluster aware.
libvirt includes a pluggable framework for lock managers, which hypervisor drivers can use to ensure safety for guest domain disks, and potentially other resources. At this time there are only two plugin implementations, a "no op" implementation which does absolutely nothing, and a sanlock implementation which uses the Disk Paxos algorithm to ensure safety.
On many operating systems, the sanlock plugin
is distributed in a sub-package which needs to be installed
separately from the main libvirt RPM. On a Fedora/RHEL host
this can be done with the yum command
$ su - root
# yum install libvirt-lock-sanlock
The next step is to start the sanlock daemon. For maximum safety sanlock prefers to have a connection to a watchdog daemon. This will cause the entire host to be rebooted in the event that sanlock crashes / terminates abnormally. To start the watchdog daemon on a Fedora/RHEL host the following commands can be run:
$ su - root
# chkconfig wdmd on
# service wdmd start
Once the watchdog is running, sanlock can be started as follows
# chkconfig sanlock on
# service sanlock start
Note: if you wish to avoid the use of the
watchdog, add the following line to /etc/sysconfig/sanlock
before starting it
SANLOCKOPTS="-w 0"
The sanlock daemon must be started on every single host that will be running virtual machines. So repeat these steps as necessary.
Once the sanlock daemon is running, the next step is to
configure the libvirt sanlock plugin. There is a separate
configuration file for each libvirt driver that is using
sanlock. For QEMU, we will edit /etc/libvirt/qemu-sanlock.conf
There is one mandatory parameter that needs to be set,
the host_id. This is a integer between
1 and 2000, which must be set to a unique
value on each host running virtual machines.
$ su - root
# augtool -s set /files/etc/libvirt/qemu-sanlock.conf/host_id 1
Repeat this on every host, changing 1 to a unique value for the host.
The sanlock plugin needs to create leases in a directory
that is on a filesystem shared between all hosts running
virtual machines. Obvious choices for this include NFS
or GFS2. The libvirt sanlock plugin expects its lease
directory be at /var/lib/libvirt/sanlock
so update the host's /etc/fstab to mount
a suitable shared/cluster filesystem at that location
$ su - root
# echo "some.nfs.server:/export/sanlock /var/lib/libvirt/sanlock nfs hard,nointr 0 0" >> /etc/fstab
# mount /var/lib/libvirt/sanlock
In terms of storage requirements, if the filesystem
uses 512 byte sectors, you need to allow for 1MB
of storage for each guest disk. So if you have a network
with 20 virtualization hosts, each running 50 virtual
machines and an average of 2 disks per guest, you will
need 20*50*2 == 2000 MB of storage for
sanlock.
On one of the hosts on the network is it wise to setup
a cron job which runs the virt-sanlock-cleanup
script periodically. This scripts deletes any lease
files which are not currently in use by running virtual
machines, freeing up disk space on the shared filesystem.
Unless VM disks are very frequently created + deleted
it should be sufficient to run the cleanup once a week.
The QEMU/KVM driver is fully integrated with the lock
manager framework as of release 0.9.3.
The out of the box configuration, however, currently
uses the nop lock manager plugin.
To get protection for disks, it is thus necessary
to reconfigure QEMU to activate the sanlock
driver. This is achieved by editing the QEMU driver
configuration file (/etc/libvirt/qemu.conf)
and changing the lock_manager configuration
tunable.
$ su - root
# augtool -s set /files/etc/libvirt/qemu.conf/lock_manager sanlock
# service libvirtd restart
If all went well, libvirtd will have talked to sanlock and created the basic lockspace. This can be checked by looking for existence of the following file
# ls /var/lib/libvirt/sanlock/
__LIBVIRT__DISKS__
Every time you start a guest, additional lease files will appear
in this directory, one for each virtual disk. The lease
files are named based on the MD5 checksum of the fully qualified
path of the virtual disk backing file. So if the guest is given
a disk backed by /var/lib/libvirt/images/demo.img
expect to see a lease /var/lib/libvirt/sanlock/bfa0240911bc17753e0b473688822159
It should be obvious that for locking to work correctly, every
host running virtual machines should have storage configured
in the same way. The easiest way to do this is to use the libvirt
storage pool capability to configure any NFS volumes, iSCSI targets,
or SCSI HBAs used for guest storage. Simply replicate the same
storage pool XML across every host. It is important that any
storage pools exposing block devices are configured to create
volume paths under /dev/disks/by-path to ensure
stable paths across hosts. An example iSCSI configuration
which ensures this is:
<pool type='iscsi'>
<name>myiscsipool</name>
<source>
<host name='192.168.254.8'/>
<device path='your-iscsi-target-iqn'/>
</source>
<target>
<path>/dev/disk/by-path</path>
</target>
</pool>