This document is for a development version of Ceph.
CephFS Distributed Metadata Cache
While the data for inodes in a Ceph file system is stored in RADOS and accessed by the clients directly, inode metadata and directory information is managed by the Ceph metadata server (MDS). The MDS’s act as mediator for all metadata related activity, storing the resulting information in a separate RADOS pool from the file data.
CephFS clients can request that the MDS fetch or change inode metadata on its behalf, but an MDS can also grant the client capabilities (aka caps) for each inode (see Capabilities in CephFS).
A capability grants the client the ability to cache and possibly manipulate some portion of the data or metadata associated with the inode. When another client needs access to the same information, the MDS will revoke the capability and the client will eventually return it, along with an updated version of the inode’s metadata (in the event that it made changes to it while it held the capability).
Clients can request capabilities and will generally get them, but when there is competing access or memory pressure on the MDS, they may be revoked. When a capability is revoked, the client is responsible for returning it as soon as it is able. Clients that fail to do so in a timely fashion may end up blocklisted and unable to communicate with the cluster.
Since the cache is distributed, the MDS must take great care to ensure that no client holds capabilities that may conflict with other clients’ capabilities, or operations that it does itself. This allows cephfs clients to rely on much greater cache coherence than a filesystem like NFS, where the client may cache data and metadata beyond the point where it has changed on the server.
Client Metadata Requests
When a client needs to query/change inode metadata or perform an operation on a directory, it has two options. It can make a request to the MDS directly, or serve the information out of its cache. With CephFS, the latter is only possible if the client has the necessary caps.
Clients can send simple requests to the MDS to query or request changes to certain metadata. The replies to these requests may also grant the client a certain set of caps for the inode, allowing it to perform subsequent requests without consulting the MDS.
Clients can also request caps directly from the MDS, which is necessary in order to read or write file data.
Distributed Locks in an MDS Cluster
When an MDS wants to read or change information about an inode, it must gather the appropriate locks for it. The MDS cluster may have a series of different types of locks on the given inode and each MDS may have disjoint sets of locks.
If there are outstanding caps that would conflict with these locks, then they must be revoked before the lock can be acquired. Once the competing caps are returned to the MDS, then it can get the locks and do the operation.
On a filesystem served by multiple MDS’, the metadata cache is also distributed among the MDS’ in the cluster. For every inode, at any given time, only one MDS in the cluster is considered authoritative. Any requests to change that inode must be done by the authoritative MDS, though non-authoritative MDS can forward requests to the authoritative one.
Non-auth MDS’ can also obtain read locks that prevent the auth MDS from changing the data until the lock is dropped, so that they can serve inode info to the clients.
The auth MDS for an inode can change over time as well. The MDS’ will actively balance responsibility for the inode cache amongst themselves, but this can be overridden by pinning certain subtrees to a single MDS.