Notice

This document is for a development version of Ceph.

ceph-mgr module developer’s guide

Warning

This is developer documentation, describing Ceph internals that are only relevant to people writing ceph-mgr modules.

Creating a module

In pybind/mgr/, create a python module. Within your module, create a class that inherits from MgrModule. For ceph-mgr to detect your module, your directory must contain a file called module.py.

The most important methods to override are:

  • a serve member function for server-type modules. This function should block forever.

  • a notify member function if your module needs to take action when new cluster data is available.

  • a handle_command member function if your module exposes CLI commands. But this approach for exposing commands is deprecated. For more details, see Exposing commands.

Some modules interface with external orchestrators to deploy Ceph services. These also inherit from Orchestrator, which adds additional methods to the base MgrModule class. See Orchestrator modules for more on creating these modules.

Installing a module

Once your module is present in the location set by the mgr module path configuration setting, you can enable it via the ceph mgr module enable command:

ceph mgr module enable mymodule

Note that the MgrModule interface is not stable, so any modules maintained outside of the Ceph tree are liable to break when run against any newer or older versions of Ceph.

Logging

Logging in Ceph manager modules is done as in any other Python program. Just import the logging package and get a logger instance with the logging.getLogger function.

Each module has a log_level option that specifies the current Python logging level of the module. To change or query the logging level of the module use the following Ceph commands:

ceph config get mgr mgr/<module_name>/log_level
ceph config set mgr mgr/<module_name>/log_level <info|debug|critical|error|warning|>

The logging level used upon the module’s start is determined by the current logging level of the mgr daemon, unless if the log_level option was previously set with the config set ... command. The mgr daemon logging level is mapped to the module python logging level as follows:

  • <= 0 is CRITICAL

  • <= 1 is WARNING

  • <= 4 is INFO

  • <= +inf is DEBUG

We can unset the module log level and fallback to the mgr daemon logging level by running the following command:

ceph config set mgr mgr/<module_name>/log_level ''

By default, modules’ logging messages are processed by the Ceph logging layer where they will be recorded in the mgr daemon’s log file. But it’s also possible to send a module’s logging message to it’s own file.

The module’s log file will be located in the same directory as the mgr daemon’s log file with the following name pattern:

<mgr_daemon_log_file_name>.<module_name>.log

To enable the file logging on a module use the following command:

ceph config set mgr mgr/<module_name>/log_to_file true

When the module’s file logging is enabled, module’s logging messages stop being written to the mgr daemon’s log file and are only written to the module’s log file.

It’s also possible to check the status and disable the file logging with the following commands:

ceph config get mgr mgr/<module_name>/log_to_file
ceph config set mgr mgr/<module_name>/log_to_file false

Exposing commands

There are two approaches for exposing a command. The first one is to use the @CLICommand decorator to decorate the method which handles the command. like this

@CLICommand('antigravity send to blackhole',
            perm='rw')
def send_to_blackhole(self, oid: str, blackhole: Optional[str] = None, inbuf: Optional[str] = None):
    '''
    Send the specified object to black hole
    '''
    obj = self.find_object(oid)
    if obj is None:
        return HandleCommandResult(-errno.ENOENT, stderr=f"object '{oid}' not found")
    if blackhole is not None and inbuf is not None:
        try:
            location = self.decrypt(blackhole, passphrase=inbuf)
        except ValueError:
            return HandleCommandResult(-errno.EINVAL, stderr='unable to decrypt location')
    else:
        location = blackhole
    self.send_object_to(obj, location)
    return HandleCommandResult(stdout=f'the black hole swallowed '{oid}'")

The first parameter passed to CLICommand is the “name” of the command. Since there are lots of commands in Ceph, we tend to group related commands with a common prefix. In this case, “antigravity” is used for this purpose. As the author is probably designing a module which is also able to launch rockets into the deep space.

The type annotations for the method parameters are mandatory here, so the usage of the command can be properly reported to the ceph CLI, and the manager daemon can convert the serialized command parameters sent by the clients to the expected type before passing them to the handler method. With properly implemented types, one can also perform some sanity checks against the parameters!

The names of the parameters are part of the command interface, so please try to take the backward compatibility into consideration when changing them. But you cannot change name of inbuf parameter, it is used to pass the content of the file specified by ceph --in-file option.

The docstring of the method is used for the description of the command.

The manager daemon cooks the usage of the command from these ingredients, like:

antigravity send to blackhole <oid> [<blackhole>]  Send the specified object to black hole

as part of the output of ceph --help.

In addition to @CLICommand, you could also use @CLIReadCommand or @CLIWriteCommand if your command only requires read permissions or write permissions respectively.

The second one is to set the COMMANDS class attribute of your module to a list of dicts like this:

COMMANDS = [
    {
        "cmd": "foobar name=myarg,type=CephString",
        "desc": "Do something awesome",
        "perm": "rw",
        # optional:
        "poll": "true"
    }
]

The cmd part of each entry is parsed in the same way as internal Ceph mon and admin socket commands (see mon/MonCommands.h in the Ceph source for examples). Note that the “poll” field is optional, and is set to False by default; this indicates to the ceph CLI that it should call this command repeatedly and output results (see ceph -h and its --period option).

Each command is expected to return a tuple (retval, stdout, stderr). retval is an integer representing a libc error code (e.g. EINVAL, EPERM, or 0 for no error), stdout is a string containing any non-error output, and stderr is a string containing any progress or error explanation output. Either or both of the two strings may be empty.

Implement the handle_command function to respond to the commands when they are sent:

MgrModule.handle_command(inbuf, cmd)

Called by ceph-mgr to request the plugin to handle one of the commands that it declared in self.COMMANDS

Return a status code, an output buffer, and an output string. The output buffer is for data results, the output string is for informative text.

Parameters
  • inbuf (str) – content of any “-i <file>” supplied to ceph cli

  • cmd (dict) – from Ceph’s cmdmap_t

Return type

Union[HandleCommandResult, Tuple[int, str, str]]

Returns

HandleCommandResult or a 3-tuple of (int, str, str)

Configuration options

Modules can load and store configuration options using the set_module_option and get_module_option methods.

Note

Use set_module_option and get_module_option to manage user-visible configuration options that are not blobs (like certificates). If you want to persist module-internal data or binary configuration data consider using the KV store.

You must declare your available configuration options in the MODULE_OPTIONS class attribute, like this:

MODULE_OPTIONS = [
    Option(name="my_option")
]

If you try to use set_module_option or get_module_option on options not declared in MODULE_OPTIONS, an exception will be raised.

You may choose to provide setter commands in your module to perform high level validation. Users can also modify configuration using the normal ceph config set command, where the configuration options for a mgr module are named like mgr/<module name>/<option>.

If a configuration option is different depending on which node the mgr is running on, then use localized configuration ( get_localized_module_option, set_localized_module_option). This may be necessary for options such as what address to listen on. Localized options may also be set externally with ceph config set, where they key name is like mgr/<module name>/<mgr id>/<option>

If you need to load and store data (e.g. something larger, binary, or multiline), use the KV store instead of configuration options (see next section).

Hints for using config options:

  • Reads are fast: ceph-mgr keeps a local in-memory copy, so in many cases you can just do a get_module_option every time you use a option, rather than copying it out into a variable.

  • Writes block until the value is persisted (i.e. round trip to the monitor), but reads from another thread will see the new value immediately.

  • If a user has used config set from the command line, then the new value will become visible to get_module_option immediately, although the mon->mgr update is asynchronous, so config set will return a fraction of a second before the new value is visible on the mgr.

  • To delete a config value (i.e. revert to default), just pass None to set_module_option.

MgrModule.get_module_option(key, default=None)

Retrieve the value of a persistent configuration setting

Return type

Union[bool, int, float, str, None]

MgrModule.set_module_option(key, val)

Set the value of a persistent configuration setting

Parameters

key (str) –

Raises

ValueError – if val cannot be parsed or it is out of the specified range

Return type

None

MgrModule.get_localized_module_option(key, default=None)

Retrieve localized configuration for this ceph-mgr instance

Return type

Union[bool, int, float, str, None]

MgrModule.set_localized_module_option(key, val)

Set localized configuration for this ceph-mgr instance :param str key: :param str val: :rtype: None :return: str

KV store

Modules have access to a private (per-module) key value store, which is implemented using the monitor’s “config-key” commands. Use the set_store and get_store methods to access the KV store from your module.

The KV store commands work in a similar way to the configuration commands. Reads are fast, operating from a local cache. Writes block on persistence and do a round trip to the monitor.

This data can be access from outside of ceph-mgr using the ceph config-key [get|set] commands. Key names follow the same conventions as configuration options. Note that any values updated from outside of ceph-mgr will not be seen by running modules until the next restart. Users should be discouraged from accessing module KV data externally – if it is necessary for users to populate data, modules should provide special commands to set the data via the module.

Use the get_store_prefix function to enumerate keys within a particular prefix (i.e. all keys starting with a particular substring).

MgrModule.get_store(key, default=None)

Get a value from this module’s persistent key value store

Return type

Optional[str]

MgrModule.set_store(key, val)

Set a value in this module’s persistent key value store. If val is None, remove key from store

Return type

None

MgrModule.get_localized_store(key, default=None)
Return type

Optional[str]

MgrModule.set_localized_store(key, val)
Return type

None

MgrModule.get_store_prefix(key_prefix)

Retrieve a dict of KV store keys to values, where the keys have the given prefix

Parameters

key_prefix (str) –

Return type

Dict[str, str]

Returns

str

Accessing cluster data

Modules have access to the in-memory copies of the Ceph cluster’s state that the mgr maintains. Accessor functions as exposed as members of MgrModule.

Calls that access the cluster or daemon state are generally going from Python into native C++ routines. There is some overhead to this, but much less than for example calling into a REST API or calling into an SQL database.

There are no consistency rules about access to cluster structures or daemon metadata. For example, an OSD might exist in OSDMap but have no metadata, or vice versa. On a healthy cluster these will be very rare transient states, but modules should be written to cope with the possibility.

Note that these accessors must not be called in the modules __init__ function. This will result in a circular locking exception.

MgrModule.get(data_name)

Called by the plugin to fetch named cluster-wide objects from ceph-mgr.

Parameters

data_name (str) – Valid things to fetch are osdmap_crush_map_text, osd_map, osd_map_tree, osd_map_crush, config, mon_map, fs_map, osd_metadata, pg_summary, io_rate, pg_dump, df, osd_stats, health, mon_status, devices, device <devid>, pg_stats, pool_stats, pg_ready, osd_ping_times, mgr_map, mgr_ips, modified_config_options, service_map, mds_metadata, have_local_config_map, osd_pool_stats, pg_status.

Note:

All these structures have their own JSON representations: experiment or look at the C++ dump() methods to learn about them.

MgrModule.get_server(hostname)

Called by the plugin to fetch metadata about a particular hostname from ceph-mgr.

This is information that ceph-mgr has gleaned from the daemon metadata reported by daemons running on a particular server.

Parameters

hostname – a hostname

Return type

Dict[str, Union[str, List[Dict[str, str]]]]

MgrModule.list_servers()

Like get_server, but gives information about all servers (i.e. all unique hostnames that have been mentioned in daemon metadata)

Returns

a list of information about all servers

Return type

list

MgrModule.get_metadata(svc_type, svc_id, default=None)

Fetch the daemon metadata for a particular service.

ceph-mgr fetches metadata asynchronously, so are windows of time during addition/removal of services where the metadata is not available to modules. None is returned if no metadata is available.

Parameters
  • svc_type (str) – service type (e.g., ‘mds’, ‘osd’, ‘mon’)

  • svc_id (str) – service id. convert OSD integer IDs to strings when calling this

Return type

dict, or None if no metadata found

MgrModule.get_daemon_status(svc_type, svc_id)

Fetch the latest status for a particular service daemon.

This method may return None if no status information is available, for example because the daemon hasn’t fully started yet.

Parameters
  • svc_type (str) – string (e.g., ‘rgw’)

  • svc_id (str) – string

Return type

Dict[str, str]

Returns

dict, or None if the service is not found

MgrModule.get_perf_schema(svc_type, svc_name)

Called by the plugin to fetch perf counter schema info. svc_name can be nullptr, as can svc_type, in which case they are wildcards

Parameters
  • svc_type (str) –

  • svc_name (str) –

Return type

Dict[str, Dict[str, Dict[str, Union[str, int]]]]

Returns

list of dicts describing the counters requested

MgrModule.get_counter(svc_type, svc_name, path)

Called by the plugin to fetch the latest performance counter data for a particular counter on a particular service.

Parameters
  • svc_type (str) –

  • svc_name (str) –

  • path (str) – a period-separated concatenation of the subsystem and the counter name, for example “mds.inodes”.

Return type

Dict[str, List[Tuple[float, int]]]

Returns

A dict of counter names to their values. each value is a list of of two-tuples of (timestamp, value). This may be empty if no data is available.

MgrModule.get_mgr_id()

Retrieve the name of the manager daemon where this plugin is currently being executed (i.e. the active manager).

Return type

str

Returns

str

Exposing health checks

Modules can raise first class Ceph health checks, which will be reported in the output of ceph status and in other places that report on the cluster’s health.

If you use set_health_checks to report a problem, be sure to call it again with an empty dict to clear your health check when the problem goes away.

MgrModule.set_health_checks(checks)

Set the module’s current map of health checks. Argument is a dict of check names to info, in this form:

{
  'CHECK_FOO': {
    'severity': 'warning',           # or 'error'
    'summary': 'summary string',
    'count': 4,                      # quantify badness
    'detail': [ 'list', 'of', 'detail', 'strings' ],
   },
  'CHECK_BAR': {
    'severity': 'error',
    'summary': 'bars are bad',
    'detail': [ 'too hard' ],
  },
}
Parameters

list – dict of health check dicts

Return type

None

What if the mons are down?

The manager daemon gets much of its state (such as the cluster maps) from the monitor. If the monitor cluster is inaccessible, whichever manager was active will continue to run, with the latest state it saw still in memory.

However, if you are creating a module that shows the cluster state to the user then you may well not want to mislead them by showing them that out of date state.

To check if the manager daemon currently has a connection to the monitor cluster, use this function:

MgrModule.have_mon_connection()

Check whether this ceph-mgr daemon has an open connection to a monitor. If it doesn’t, then it’s likely that the information we have about the cluster is out of date, and/or the monitor cluster is down.

Return type

bool

Reporting if your module cannot run

If your module cannot be run for any reason (such as a missing dependency), then you can report that by implementing the can_run function.

static MgrModule.can_run()

Implement this function to report whether the module’s dependencies are met. For example, if the module needs to import a particular dependency to work, then use a try/except around the import at file scope, and then report here if the import failed.

This will be called in a blocking way from the C++ code, so do not do any I/O that could block in this function.

:return a 2-tuple consisting of a boolean and explanatory string

Return type

Tuple[bool, str]

Note that this will only work properly if your module can always be imported: if you are importing a dependency that may be absent, then do it in a try/except block so that your module can be loaded far enough to use can_run even if the dependency is absent.

Sending commands

A non-blocking facility is provided for sending monitor commands to the cluster.

MgrModule.send_command(result, svc_type, svc_id, command, tag, inbuf=None)

Called by the plugin to send a command to the mon cluster.

Parameters
  • result (CommandResult) – an instance of the CommandResult class, defined in the same module as MgrModule. This acts as a completion and stores the output of the command. Use CommandResult.wait() if you want to block on completion.

  • svc_type (str) –

  • svc_id (str) –

  • command (str) – a JSON-serialized command. This uses the same format as the ceph command line, which is a dictionary of command arguments, with the extra prefix key containing the command name itself. Consult MonCommands.h for available commands and their expected arguments.

  • tag (str) – used for nonblocking operation: when a command completes, the notify() callback on the MgrModule instance is triggered, with notify_type set to “command”, and notify_id set to the tag of the command.

  • inbuf (str) – input buffer for sending additional data.

Return type

None

Receiving notifications

The manager daemon calls the notify function on all active modules when certain important pieces of cluster state are updated, such as the cluster maps.

The actual data is not passed into this function, rather it is a cue for the module to go and read the relevant structure if it is interested. Most modules ignore most types of notification: to ignore a notification simply return from this function without doing anything.

MgrModule.notify(notify_type, notify_id)

Called by the ceph-mgr service to notify the Python plugin that new state is available.

Parameters
  • notify_type (str) – string indicating what kind of notification, such as osd_map, mon_map, fs_map, mon_status, health, pg_summary, command, service_map

  • notify_id (str) – string (may be empty) that optionally specifies which entity is being notified about. With “command” notifications this is set to the tag from send_command.

Return type

None

Accessing RADOS or CephFS

If you want to use the librados python API to access data stored in the Ceph cluster, you can access the rados attribute of your MgrModule instance. This is an instance of rados.Rados which has been constructed for you using the existing Ceph context (an internal detail of the C++ Ceph code) of the mgr daemon.

Always use this specially constructed librados instance instead of constructing one by hand.

Similarly, if you are using libcephfs to access the file system, then use the libcephfs create_with_rados to construct it from the MgrModule.rados librados instance, and thereby inherit the correct context.

Remember that your module may be running while other parts of the cluster are down: do not assume that librados or libcephfs calls will return promptly – consider whether to use timeouts or to block if the rest of the cluster is not fully available.

Implementing standby mode

For some modules, it is useful to run on standby manager daemons as well as on the active daemon. For example, an HTTP server can usefully serve HTTP redirect responses from the standby managers so that the user can point his browser at any of the manager daemons without having to worry about which one is active.

Standby manager daemons look for a subclass of StandbyModule in each module. If the class is not found then the module is not used at all on standby daemons. If the class is found, then its serve method is called. Implementations of StandbyModule must inherit from mgr_module.MgrStandbyModule.

The interface of MgrStandbyModule is much restricted compared to MgrModule – none of the Ceph cluster state is available to the module. serve and shutdown methods are used in the same way as a normal module class. The get_active_uri method enables the standby module to discover the address of its active peer in order to make redirects. See the MgrStandbyModule definition in the Ceph source code for the full list of methods.

For an example of how to use this interface, look at the source code of the dashboard module.

Communicating between modules

Modules can invoke member functions of other modules.

MgrModule.remote(module_name, method_name, *args, **kwargs)

Invoke a method on another module. All arguments, and the return value from the other module must be serializable.

Limitation: Do not import any modules within the called method. Otherwise you will get an error in Python 2:

RuntimeError('cannot unmarshal code objects in restricted execution mode',)
Parameters
  • module_name (str) – Name of other module. If module isn’t loaded, an ImportError exception is raised.

  • method_name (str) – Method name. If it does not exist, a NameError exception is raised.

  • args (Any) – Argument tuple

  • kwargs (Any) – Keyword argument dict

Raises
  • RuntimeErrorAny error raised within the method is converted to a RuntimeError

  • ImportError – No such module

Return type

Any

Be sure to handle ImportError to deal with the case that the desired module is not enabled.

If the remote method raises a python exception, this will be converted to a RuntimeError on the calling side, where the message string describes the exception that was originally thrown. If your logic intends to handle certain errors cleanly, it is better to modify the remote method to return an error value instead of raising an exception.

At time of writing, inter-module calls are implemented without copies or serialization, so when you return a python object, you’re returning a reference to that object to the calling module. It is recommend not to rely on this reference passing, as in future the implementation may change to serialize arguments and return values.

Shutting down cleanly

If a module implements the serve() method, it should also implement the shutdown() method to shutdown cleanly: misbehaving modules may otherwise prevent clean shutdown of ceph-mgr.

Limitations

It is not possible to call back into C++ code from a module’s __init__() method. For example calling self.get_module_option() at this point will result in an assertion failure in ceph-mgr. For modules that implement the serve() method, it usually makes sense to do most initialization inside that method instead.

Debugging

Apparently, we can always use the Logging facility for debugging a ceph-mgr module. But some of us might miss PDB and the interactive Python interpreter. Yes, we can have them as well when developing ceph-mgr modules! ceph_mgr_repl.py can drop you into an interactive shell talking to selftest module. With this tool, one can peek and poke the ceph-mgr module, and use all the exposed facilities in quite the same way how we use the Python command line interpreter. For using ceph_mgr_repl.py, we need to

  1. ready a Ceph cluster

  2. enable the selftest module

  3. setup the necessary environment variables

  4. launch the tool

Following is a sample session, in which the Ceph version is queried by inputting print(mgr.version) at the prompt. And later timeit module is imported to measure the execution time of mgr.get_mgr_id().

$ cd build
$ MDS=0 MGR=1 OSD=3 MON=1 ../src/vstart.sh -n -x
$ bin/ceph mgr module enable selftest
$ ../src/pybind/ceph_mgr_repl.py --show-env
   $ export PYTHONPATH=/home/me/ceph/src/pybind:/home/me/ceph/build/lib/cython_modules/lib.3:/home/me/ceph/src/python-common:$PYTHONPATH
   $ export LD_LIBRARY_PATH=/home/me/ceph/build/lib:$LD_LIBRARY_PATH
$ export PYTHONPATH=/home/me/ceph/src/pybind:/home/me/ceph/build/lib/cython_modules/lib.3:/home/me/ceph/src/python-common:$PYTHONPATH
$ export LD_LIBRARY_PATH=/home/me/ceph/build/lib:$LD_LIBRARY_PATH
$ ../src/pybind/ceph_mgr_repl.py
$ ../src/pybind/ceph_mgr_repl.py
Python 3.9.2 (default, Feb 28 2021, 17:03:44)
[GCC 10.2.1 20210110] on linux
Type "help", "copyright", "credits" or "license" for more information.
(MgrModuleInteractiveConsole)
[mgr self-test eval] >>> print(mgr.version)
ceph version Development (no_version) quincy (dev)
[mgr self-test eval] >>> from timeit import timeit
[mgr self-test eval] >>> timeit(mgr.get_mgr_id)
0.16303414600042743
[mgr self-test eval] >>>

If you want to “talk” to a ceph-mgr module other than selftest using this tool, you can either add a command to the module you want to debug exactly like how mgr self-test eval command was added to selftest. Or we can make this simpler by promoting the eval() method to a dedicated Mixin class and inherit your MgrModule subclass from it. And define a command with it. Assuming the prefix of the command is mgr my-module eval, one can just put

../src/pybind/ceph_mgr_repl.py --prefix "mgr my-module eval"

Is something missing?

The ceph-mgr python interface is not set in stone. If you have a need that is not satisfied by the current interface, please bring it up on the ceph-devel mailing list. While it is desired to avoid bloating the interface, it is not generally very hard to expose existing data to the Python code when there is a good reason.