Virtual machine disk locking

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.

Lock manager plugins

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.

Sanlock daemon setup

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.

libvirt sanlock plugin configuration

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.

libvirt sanlock storage configuration

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
    

If your sanlock daemon happen to run under non-root privileges, you need to tell this to libvirt so it chowns created files correctly. This can be done by setting user and/or group variables in the configuration file. Accepted values range is specified in description to the same variables in /etc/libvirt/qemu.conf. For example:

      augtool -s set /files/etc/libvirt/qemu-sanlock.conf/user sanlock
      augtool -s set /files/etc/libvirt/qemu-sanlock.conf/group sanlock
    

But remember, that if this is NFS share, you need a no_root_squash-ed one for chown (and chmod possibly) to succeed.

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.

QEMU/KVM driver configuration

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>
    

Domain configuration

In case sanlock loses access to disk locks for some reason, it will kill all domains that lost their locks. This default behavior may be changed using on_lockfailure element in domain XML. When this element is present, sanlock will call sanlock_helper (provided by libvirt) with the specified action. This helper binary will connect to libvirtd and thus it may need to authenticate if libvirtd was configured to require that on the read-write UNIX socket. To provide the appropriate credentials to sanlock_helper, a client authentication file needs to contain something like the following:

[auth-libvirt-localhost]
credentials=sanlock

[credentials-sanlock]
authname=login
password=password