Capabilities in CephFS
When a client wants to operate on an inode, it will query the MDS in various ways, which will then grant the client a set of capabilities. These grant the client permissions to operate on the inode in various ways. One of the major differences from other network file systems (e.g NFS or SMB) is that the capabilities granted are quite granular, and it’s possible that multiple clients can hold different capabilities on the same inodes.
Types of Capabilities
There are several “generic” capability bits. These denote what sort of ability the capability grants.
/* generic cap bits */
#define CEPH_CAP_GSHARED 1 /* (metadata) client can reads (s) */
#define CEPH_CAP_GEXCL 2 /* (metadata) client can read and update (x) */
#define CEPH_CAP_GCACHE 4 /* (file) client can cache reads (c) */
#define CEPH_CAP_GRD 8 /* (file) client can read (r) */
#define CEPH_CAP_GWR 16 /* (file) client can write (w) */
#define CEPH_CAP_GBUFFER 32 /* (file) client can buffer writes (b) */
#define CEPH_CAP_GWREXTEND 64 /* (file) client can extend EOF (a) */
#define CEPH_CAP_GLAZYIO 128 /* (file) client can perform lazy io (l) */
These are then shifted by a particular number of bits. These denote a part of the inode’s data or metadata on which the capability is being granted:
/* per-lock shift */
#define CEPH_CAP_SAUTH 2 /* A */
#define CEPH_CAP_SLINK 4 /* L */
#define CEPH_CAP_SXATTR 6 /* X */
#define CEPH_CAP_SFILE 8 /* F */
Only certain generic cap types are ever granted for some of those “shifts”, however. In particular, only the FILE shift ever has more than the first two bits.
| AUTH | LINK | XATTR | FILE
2 4 6 8
From the above, we get a number of constants, that are generated by taking each bit value and shifting to the correct bit in the word:
#define CEPH_CAP_AUTH_SHARED (CEPH_CAP_GSHARED << CEPH_CAP_SAUTH)
These bits can then be or’ed together to make a bitmask denoting a set of capabilities.
There is one exception:
#define CEPH_CAP_PIN 1 /* no specific capabilities beyond the pin */
The “pin” just pins the inode into memory, without granting any other caps.
Graphically:
+---+---+---+---+---+---+---+---+
| p | _ |As x |Ls x |Xs x |
+---+---+---+---+---+---+---+---+
|Fs x c r w b a l |
+---+---+---+---+---+---+---+---+
The second bit is currently unused.
Abilities granted by each cap
While that is how capabilities are granted (and communicated), the important bit is what they actually allow the client to do:
PIN: this just pins the inode into memory. This is sufficient to allow the client to get to the inode number, as well as other immutable things like major or minor numbers in a device inode, or symlink contents.
AUTH: this grants the ability to get to the authentication-related metadata. In particular, the owner, group and mode. Note that doing a full permission check may require getting at ACLs as well, which are stored in xattrs.
LINK: the link count of the inode
XATTR: ability to access or manipulate xattrs. Note that since ACLs are stored in xattrs, it’s also sometimes necessary to access them when checking permissions.
FILE: this is the big one. These allow the client to access and manipulate file data. It also covers certain metadata relating to file data – the size, mtime, atime and ctime, in particular.
Shorthand
Note that the client logging can also present a compact representation of the capabilities. For example:
pAsLsXsFs
The ‘p’ represents the pin. Each capital letter corresponds to the shift values, and the lowercase letters after each shift are for the actual capabilities granted in each shift.
The relation between the lock states and the capabilities
In MDS there are four different locks for each inode, they are simplelock, scatterlock, filelock and locallock. Each lock has several different lock states, and the MDS will issue capabilities to clients based on the lock state.
In each state the MDS Locker will always try to issue all the capabilities to the clients allowed, even some capabilities are not needed or wanted by the clients, as pre-issuing capabilities could reduce latency in some cases.
If there is only one client, usually it will be the loner client for all the inodes. While in multiple clients case, the MDS will try to calculate a loner client out for each inode depending on the capabilities the clients (needed | wanted), but usually it will fail. The loner client will always get all the capabilities.
The filelock will control files’ partial metadatas’ and the file contents’ access permissions. The metadatas include mtime, atime, size, etc.
Fs: Once a client has it, all other clients are denied Fw.
Fx: Only the loner client is allowed this capability. Once the lock state transitions to LOCK_EXCL, the loner client is granted this along with all other file capabilities except the Fl.
Fr: Once a client has it, the Fb capability will be already revoked from all the other clients.
If clients only request to read the file, the lock state will be transferred to LOCK_SYNC stable state directly. All the clients can be granted Fscrl capabilities from the auth MDS and Fscr capabilities from the replica MDSes.
If multiple clients read from and write to the same file, then the lock state will be transferred to LOCK_MIX stable state finally and all the clients could have the Frwl capabilities from the auth MDS, and the Fr from the replica MDSes. The Fcb capabilities won’t be granted to all the clients and the clients will do sync read/write.
Fw: If there is no loner client and once a client have this capability, the Fsxcb capabilities won’t be granted to other clients.
If multiple clients read from and write to the same file, then the lock state will be transferred to LOCK_MIX stable state finally and all the clients could have the Frwl capabilities from the auth MDS, and the Fr from the replica MDSes. The Fcb capabilities won’t be granted to all the clients and the clients will do sync read/write.
Fc: This capability means the clients could cache file read and should be issued together with Fr capability and only in this use case will it make sense.
While actually in some stable or interim transitional states they tend to keep the Fc allowed even the Fr capability isn’t granted as this can avoid forcing clients to drop full caches, for example on a simple file size extension or truncating use case.
Fb: This capability means the clients could buffer file write and should be issued together with Fw capability and only in this use case will it make sense.
While actually in some stable or interim transitional states they tend to keep the Fc allowed even the Fw capability isn’t granted as this can avoid forcing clients to drop dirty buffers, for example on a simple file size extension or truncating use case.
Fl: This capability means the clients could perform lazy io. LazyIO relaxes POSIX semantics. Buffered reads/writes are allowed even when a file is opened by multiple applications on multiple clients. Applications are responsible for managing cache coherency themselves.