RBD Exclusive Locks

Exclusive locks are a mechanism designed to prevent multiple processes from accessing the same Rados Block Device (RBD) in an uncoordinated fashion. Exclusive locks are heavily used in virtualization (where they prevent VMs from clobbering each others’ writes), and also in RBD mirroring (where they are a prerequisite for journaling).

Exclusive locks are enabled on newly created images by default, unless overridden via the rbd_default_features configuration option or the --image-feature flag for rbd create.

In order to ensure proper exclusive locking operations, any client using an RBD image whose exclusive-lock feature is enabled should be using a CephX identity whose capabilities include profile rbd.

Exclusive locking is mostly transparent to the user.

  1. Whenever any librbd client process or kernel RBD client starts using an RBD image on which exclusive locking has been enabled, it obtains an exclusive lock on the image before the first write.

  2. Whenever any such client process gracefully terminates, it automatically relinquishes the lock.

  3. This subsequently enables another process to acquire the lock, and write to the image.

Note that it is perfectly possible for two or more concurrently running processes to merely open the image, and also to read from it. The client acquires the exclusive lock only when attempting to write to the image. To disable transparent lock transitions between multiple clients, it needs to acquire the lock specifically with RBD_LOCK_MODE_EXCLUSIVE.

Blacklisting

Sometimes, a client process (or, in case of a krbd client, a client node’s kernel thread) that previously held an exclusive lock on an image does not terminate gracefully, but dies abruptly. This may be due to having received a KILL or ABRT signal, for example, or a hard reboot or power failure of the client node. In that case, the exclusive lock is never gracefully released. Thus, when a new process starts and attempts to use the device, it needs a way to break the previously held exclusive lock.

However, a process (or kernel thread) may also hang, or merely lose network connectivity to the Ceph cluster for some amount of time. In that case, simply breaking the lock would be potentially catastrophic: the hung process or connectivity issue may resolve itself, and the old process may then compete with one that has started in the interim, accessing RBD data in an uncoordinated and destructive manner.

Thus, in the event that a lock cannot be acquired in the standard graceful manner, the overtaking process not only breaks the lock, but also blocklists the previous lock holder. This is negotiated between the new client process and the Ceph Mon: upon receiving the blocklist request,

  • the Mon instructs the relevant OSDs to no longer serve requests from the old client process;

  • once the associated OSD map update is complete, the Mon grants the lock to the new client;

  • once the new client has acquired the lock, it can commence writing to the image.

Blocklisting is thus a form of storage-level resource fencing.

In order for blocklisting to work, the client must have the osd blocklist capability. This capability is included in the profile rbd capability profile, which should generally be set on all Ceph client identities using RBD.