These are sections that do not fit into the rest of the document.
This is a brief note on a potential future feature.
Clixon is mainly a stand-alone app tightly coupled to the application/device with "shared fate", that is, if clixon fails, so does the application.
- That said, the primary state is the backend holding the configuration datastore that can be shared in several ways. This is not implemented in Clixon, but potential implementation strategies include:
- Active/standby: With a standard failure/liveness detection of a master backend, a standby could be started when the master fails using "-s running" (just picking up the state from the failed master). The cache write-through mechanism supports this. Would suffer from outage during standby boot. One way already possible today is using existing detection/restart mechanisms in SystemD for example.
- Active/active: Would need an active cache coherence algorithm. Would also need a failure/liveness detection.
In both cases the config-db would be a single-point-of-failure but could be mitigated by a replicated file system, for example.
- Regarding clients:
- the CLI and NETCONF clients are stateless and spun up on demand.
- the RESTCONF daemon is stateless and can run as multiple instances (with a load-balancer)
Clixon has a simple internal process handling currently used for :ref:`internal restconf <clixon_restconf>` but can also be used for user applications. The process data structure has a unique name with pids created for "active" processes. There are three states:
- STOPPED
- pid=0, No process running
- RUNNING
- pid is set, Process started and assumed running
- EXITING
- pid set, Process is killed by parent but not waited for (eg not
There are three operations that a client can perform on processes:
- start
- Start a process
- restart
- Restart a process
- stop
- Stop a process
The state machine for a process is as follows:
--> STOPPED --(re)start--> RUNNING(pid) ^ <--1.wait(kill)--- | ^ | stop/| | | restart| | restart | v | wait(stop) ------- EXITING(dying pid)
The Process struct is created by calling clixon_process_register() with static info such as name, description, namespace, start arguments, etc. Starts in STOPPED state:
--> STOPPED
On operation "start" or "restart" it gets a pid and goes into RUNNING state:
STOPPED -- (re)start --> RUNNING(pid)
When running, several things may happen:
It is killed externally: the process gets a SIGCHLD triggers a wait and it goes to STOPPED:
RUNNING --sigchld/wait--> STOPPEDIt is stopped due to a rpc or configuration remove: The parent kills the process and enters EXITING waiting for a SIGCHLD that triggers a wait, therafter it goes to STOPPED:
RUNNING --stop--> EXITING --sigchld/wait--> STOPPEDIt is restarted due to rpc or config change (eg a server is added, a key modified, etc). The parent kills the process and enters EXITING waiting for a SIGCHLD that triggers a wait, therafter a new process is started and it goes to RUNNING with a new pid:
RUNNING --restart--> EXITING --sigchld/wait + restart --> RUNNING(pid)
Clixon implements RFC 5277 NETCONF Event Notifications
The main example illustrates an EXAMPLE stream notification that triggers every 5s. First, declare a notification in YANG:
notification event { description "Example notification event."; leaf event-class { type string; description "Event class identifier."; } ...
To start a notification stream via netconf:
<rpc><create-subscription xmlns="urn:ietf:params:xml:ns:netmod:notification"><stream>EXAMPLE</stream></create-subscription></rpc>]]>]]> <rpc-reply><ok/></rpc-reply>]]>]]> <notification xmlns="urn:ietf:params:xml:ns:netconf:notification:1.0"><eventTime>2019-01-02T10:20:05.929272</eventTime><event><event-class>fault</event-class><reportingEntity><card>Ethernet0</card></reportingEntity><severity>major</severity></event></notification>]]>]]>
This can also be triggered via the CLI:
clixon_cli -f /usr/local/etc/clixon/example.xml cli> notify cli> event-class fault; reportingEntity { card Ethernet0; } severity major; cli> no notify cli>
Restconf notifications are also supported.
While only XML and JSON are currently supported as datastore formats, Clixon also supports CLI and TEXT formats for printing, and saving and loading files.
The main example contains example code showing how to load and save a config using other formats.
Example of showing a config as XML, JSON, TEXT and CLI:
cli> show configuration xml <table xmlns="urn:example:clixon"> <parameter> <name>a</name> <value>17</value> </parameter> <parameter> <name>b</name> <value>99</value> </parameter> </table> cli> show configuration json { "clixon-example:table": { "parameter": [ { "name": "a", "value": "17" }, { "name": "b", "value": "99" } ] } } cli> show configuration text clixon-example:table { parameter a { value 17; } parameter b { value 99; } } cli> show configuration cli set table parameter a set table parameter a value 17 set table parameter b set table parameter b value 99
Save and load a file using TEXT:
cli> save foo.txt text cli> load foo.txt replace text
CLI show and save commands uses an internal API for print, save and load of the formats. Such CLI functions include: cli_show_config, cli_pagination, load_config_file, save_config_file.
The following internal C API is available for output:
- XML:
clixon_xml2file()
andclixon_xml2cbuf()
to file and memory respectively. - JSON:
clixon_json2file()
andclixon_json2cbuf()
- CLI:
clixon_cli2file()
- TEXT:
clixon_txt2file()
The arguments of these functions are similar with some local variance. For example:
int clixon_xml2file(FILE *f, cxobj *xn, int level, int pretty, clicon_output_cb *fn, int skiptop, int autocliext)
where:
- f is the output stream (such as stdout)
- xn is the top-level XML node
- level is indentation level to start with, normally 0
- pretty makes the output indented and use newlines
- fn is the output function to use. NULL means fprintf, cligen_output is used for scrolling in CLI
- skiproot only prints the children by skipping the top-level XML node xn
- autocliext Set if you want to activate autocli extensions (eg hide extensions)