mirror of
				https://git.proxmox.com/git/mirror_frr
				synced 2025-11-04 06:38:04 +00:00 
			
		
		
		
	Fold things into where they make sense. Signed-off-by: David Lamparter <equinox@diac24.net>
		
			
				
	
	
		
			817 lines
		
	
	
		
			34 KiB
		
	
	
	
		
			ReStructuredText
		
	
	
	
	
	
			
		
		
	
	
			817 lines
		
	
	
		
			34 KiB
		
	
	
	
		
			ReStructuredText
		
	
	
	
	
	
.. _command-line-interface:
 | 
						|
 | 
						|
Command Line Interface
 | 
						|
======================
 | 
						|
 | 
						|
FRR features a flexible modal command line interface. Often when adding new
 | 
						|
features or modifying existing code it is necessary to create or modify CLI
 | 
						|
commands. FRR has a powerful internal CLI system that does most of the heavy
 | 
						|
lifting for you.
 | 
						|
 | 
						|
Modes
 | 
						|
-----
 | 
						|
FRR's CLI is organized by modes. Each mode is associated with some set of
 | 
						|
functionality, e.g. EVPN, or some underlying object such as an interface. Each
 | 
						|
mode contains a set of commands that control the associated functionality or
 | 
						|
object. Users move between the modes by entering a command, which is usually
 | 
						|
different for each source and destination mode.
 | 
						|
 | 
						|
A summary of the modes is given in the following figure.
 | 
						|
 | 
						|
.. graphviz:: ../figures/nodes.dot
 | 
						|
 | 
						|
.. seealso:: :ref:`cli-data-structures`
 | 
						|
 | 
						|
Walkup
 | 
						|
^^^^^^
 | 
						|
FRR exhibits, for historical reasons, a peculiar behavior called 'walkup'.
 | 
						|
Suppose a user is in ``OSPF_NODE``, which contains only OSPF-specific commands,
 | 
						|
and enters the following command: ::
 | 
						|
 | 
						|
   ip route 192.168.100.0/24 10.0.2.2
 | 
						|
 | 
						|
This command is not defined in ``OSPF_NODE``, so the matcher will fail to match
 | 
						|
the command in that node. The matcher will then check "parent" nodes of
 | 
						|
``OSPF_NODE``. In this case the direct parent of ``OSPF_NODE`` is
 | 
						|
``CONFIG_NODE``, so the current node switches to ``CONFIG_NODE`` and the command
 | 
						|
is tried in that node. Since static route commands are defined in
 | 
						|
``CONFIG_NODE`` the command succeeds. The procedure of attempting to execute
 | 
						|
unmatched commands by sequentially "walking up" to parent nodes only happens in
 | 
						|
children (direct and indirect) below ``CONFIG_NODE`` and stops at
 | 
						|
``CONFIG_NODE``.
 | 
						|
 | 
						|
Unfortunately, the internal representation of the various modes is not actually
 | 
						|
a graph. Instead, there is an array. The parent-child relationships are not
 | 
						|
explicitly defined in any datastructure but instead are hard-coded into the
 | 
						|
specific commands that switch nodes. For walkup, there is a function that takes
 | 
						|
a node and returns the parent of the node. This interface causes all manner of
 | 
						|
insidious problems, even for experienced developers, and needs to be fixed at
 | 
						|
some point in the future.
 | 
						|
 | 
						|
Defining Commands
 | 
						|
-----------------
 | 
						|
All definitions for the CLI system are exposed in ``lib/command.h``. In this
 | 
						|
header there are a set of macros used to define commands. These macros are
 | 
						|
collectively referred to as "DEFUNs", because of their syntax:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
    DEFUN(command_name,
 | 
						|
          command_name_cmd,
 | 
						|
          "example command FOO...",
 | 
						|
          "Examples\n"
 | 
						|
          "CLI command\n"
 | 
						|
          "Argument\n")
 | 
						|
    {
 | 
						|
        // ...command handler...
 | 
						|
    }
 | 
						|
 | 
						|
DEFUNs generally take four arguments which are expanded into the appropriate
 | 
						|
constructs for hooking into the CLI. In order these are:
 | 
						|
 | 
						|
- **Function name** - the name of the handler function for the command
 | 
						|
- **Command name** - the identifier of the ``struct cmd_element`` for the
 | 
						|
  command. By convention this should be the function name with ``_cmd``
 | 
						|
  appended.
 | 
						|
- **Command definition** - an expression in FRR's CLI grammar that defines the
 | 
						|
  form of the command and its arguments, if any
 | 
						|
- **Doc string** - a newline-delimited string that documents each element in
 | 
						|
  the command definition
 | 
						|
 | 
						|
In the above example, ``command_name`` is the function name,
 | 
						|
``command_name_cmd`` is the command name, ``"example..."`` is the definition and
 | 
						|
the last argument is the doc string. The block following the macro is the body
 | 
						|
of the handler function, details on which are presented later in this section.
 | 
						|
 | 
						|
In order to make the command show up to the user it must be installed into the
 | 
						|
CLI graph. To do this, call:
 | 
						|
 | 
						|
``install_element(NODE, &command_name_cmd);``
 | 
						|
 | 
						|
This will install the command into the specified CLI node. Usually these calls
 | 
						|
are grouped together in a CLI initialization function for a set of commands, and
 | 
						|
the DEFUNs themselves are grouped into the same source file to avoid cluttering
 | 
						|
the codebase.  The names of these files follow the form ``*_vty.[ch]`` by
 | 
						|
convention. Please do not scatter individual CLI commands in the middle of
 | 
						|
source files; instead expose the necessary functions in a header and place the
 | 
						|
command definition in a ``*_vty.[ch]`` file.
 | 
						|
 | 
						|
Definition Grammar
 | 
						|
^^^^^^^^^^^^^^^^^^
 | 
						|
FRR uses its own grammar for defining CLI commands. The grammar draws from
 | 
						|
syntax commonly seen in \*nix manpages and should be fairly intuitive. The
 | 
						|
parser is implemented in Bison and the lexer in Flex. These may be found in
 | 
						|
``lib/command_lex.l`` and ``lib/command_parse.y``, respectively.
 | 
						|
 | 
						|
    **ProTip**: if you define a new command and find that the parser is
 | 
						|
    throwing syntax or other errors, the parser is the last place you want
 | 
						|
    to look. Bison is very stable and if it detects a syntax error, 99% of
 | 
						|
    the time it will be a syntax error in your definition.
 | 
						|
 | 
						|
The formal grammar in BNF is given below. This is the grammar implemented in the
 | 
						|
Bison parser. At runtime, the Bison parser reads all of the CLI strings and
 | 
						|
builds a combined directed graph that is used to match and interpret user input.
 | 
						|
 | 
						|
Human-friendly explanations of how to use this grammar are given a bit later in
 | 
						|
this section alongside information on the :ref:`cli-data-structures` constructed
 | 
						|
by the parser.
 | 
						|
 | 
						|
.. productionlist::
 | 
						|
   command: `cmd_token_seq`
 | 
						|
          : `cmd_token_seq` `placeholder_token` "..."
 | 
						|
   cmd_token_seq: *empty*
 | 
						|
                : `cmd_token_seq` `cmd_token`
 | 
						|
   cmd_token: `simple_token`
 | 
						|
            : `selector`
 | 
						|
   simple_token: `literal_token`
 | 
						|
               : `placeholder_token`
 | 
						|
   literal_token: WORD `varname_token`
 | 
						|
   varname_token: "$" WORD
 | 
						|
   placeholder_token: `placeholder_token_real` `varname_token`
 | 
						|
   placeholder_token_real: IPV4
 | 
						|
                         : IPV4_PREFIX
 | 
						|
                         : IPV6
 | 
						|
                         : IPV6_PREFIX
 | 
						|
                         : VARIABLE
 | 
						|
                         : RANGE
 | 
						|
                         : MAC
 | 
						|
                         : MAC_PREFIX
 | 
						|
   selector: "<" `selector_seq_seq` ">" `varname_token`
 | 
						|
           : "{" `selector_seq_seq` "}" `varname_token`
 | 
						|
           : "[" `selector_seq_seq` "]" `varname_token`
 | 
						|
   selector_seq_seq: `selector_seq_seq` "|" `selector_token_seq`
 | 
						|
                   : `selector_token_seq`
 | 
						|
   selector_token_seq: `selector_token_seq` `selector_token`
 | 
						|
                     : `selector_token`
 | 
						|
   selector_token: `selector`
 | 
						|
                 : `simple_token`
 | 
						|
 | 
						|
Tokens
 | 
						|
^^^^^^
 | 
						|
The various capitalized tokens in the BNF above are in fact themselves
 | 
						|
placeholders, but not defined as such in the formal grammar; the grammar
 | 
						|
provides the structure, and the tokens are actually more like a type system for
 | 
						|
the strings you write in your CLI definitions. A CLI definition string is broken
 | 
						|
apart and each piece is assigned a type by the lexer based on a set of regular
 | 
						|
expressions. The parser uses the type information to verify the string and
 | 
						|
determine the structure of the CLI graph; additional metadata (such as the raw
 | 
						|
text of each token) is encoded into the graph as it is constructed by the
 | 
						|
parser, but this is merely a dumb copy job.
 | 
						|
 | 
						|
Here is a brief summary of the various token types along with examples.
 | 
						|
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| Token type      | Syntax          | Description                                                 |
 | 
						|
+=================+=================+=============================================================+
 | 
						|
| ``WORD``        | ``show ip bgp`` | Matches itself. In the given example every token is a WORD. |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``IPV4``        | ``A.B.C.D``     | Matches an IPv4 address.                                    |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``IPV6``        | ``X:X::X:X``    | Matches an IPv6 address.                                    |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``IPV4_PREFIX`` | ``A.B.C.D/M``   | Matches an IPv4 prefix in CIDR notation.                    |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``IPV6_PREFIX`` | ``X:X::X:X/M``  | Matches an IPv6 prefix in CIDR notation.                    |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``MAC``         | ``M:A:C``       | Matches a 48-bit mac address.                               |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``MAC_PREFIX``  | ``M:A:C/M``     | Matches a 48-bit mac address with a mask.                   |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``VARIABLE``    | ``FOOBAR``      | Matches anything.                                           |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
| ``RANGE``       | ``(X-Y)``       | Matches numbers in the range X..Y inclusive.                |
 | 
						|
+-----------------+-----------------+-------------------------------------------------------------+
 | 
						|
 | 
						|
When presented with user input, the parser will search over all defined
 | 
						|
commands in the current context to find a match. It is aware of the various
 | 
						|
types of user input and has a ranking system to help disambiguate commands. For
 | 
						|
instance, suppose the following commands are defined in the user's current
 | 
						|
context:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   example command FOO
 | 
						|
   example command (22-49)
 | 
						|
   example command A.B.C.D/X
 | 
						|
 | 
						|
The following table demonstrates the matcher's choice for a selection of
 | 
						|
possible user input.
 | 
						|
 | 
						|
+---------------------------------+---------------------------+--------------------------------------------------------------------------------------------------------------+
 | 
						|
| Input                           | Matched command           | Reason                                                                                                       |
 | 
						|
+=================================+===========================+==============================================================================================================+
 | 
						|
| ``example command eLi7eH4xx0r`` | example command FOO       | ``eLi7eH4xx0r`` is not an integer or IPv4 prefix,                                                            |
 | 
						|
|                                 |                           | but FOO is a variable and matches all input.                                                                 |
 | 
						|
+---------------------------------+---------------------------+--------------------------------------------------------------------------------------------------------------+
 | 
						|
| ``example command 42``          | example command (22-49)   | ``42`` is not an IPv4 prefix. It does match both                                                             |
 | 
						|
|                                 |                           | ``(22-49)`` and ``FOO``, but RANGE tokens are more specific and have a higher priority than VARIABLE tokens. |
 | 
						|
+---------------------------------+---------------------------+--------------------------------------------------------------------------------------------------------------+
 | 
						|
| ``example command 10.3.3.0/24`` | example command A.B.C.D/X | The user entered an IPv4 prefix, which is best matched by the last command.                                  |
 | 
						|
+---------------------------------+---------------------------+--------------------------------------------------------------------------------------------------------------+
 | 
						|
 | 
						|
Rules
 | 
						|
^^^^^
 | 
						|
There are also constructs which allow optional tokens, mutual exclusion,
 | 
						|
one-or-more selection and repetition.
 | 
						|
 | 
						|
-  ``<angle|brackets>`` -- Contain sequences of tokens separated by pipes and
 | 
						|
   provide mutual exclusion. User input matches at most one option.
 | 
						|
-  ``[square brackets]`` -- Contains sequences of tokens that can be omitted.
 | 
						|
   ``[<a|b>]`` can be shortened to ``[a|b]``.
 | 
						|
-  ``{curly|braces}`` -- similar to angle brackets, but instead of mutual
 | 
						|
   exclusion, curly braces indicate that one or more of the pipe-separated
 | 
						|
   sequences may be provided in any order.
 | 
						|
-  ``VARIADICS...`` -- Any token which accepts input (anything except WORD)
 | 
						|
   which occurs as the last token of a line may be followed by an ellipsis,
 | 
						|
   which indicates that input matching the token may be repeated an unlimited
 | 
						|
   number of times.
 | 
						|
-  ``$name`` -- Specify a variable name for the preceding token. See
 | 
						|
   "Variable Names" below.
 | 
						|
 | 
						|
Some general notes:
 | 
						|
 | 
						|
-  Options are allowed at the beginning of the command. The developer is
 | 
						|
   entreated to use these extremely sparingly. They are most useful for
 | 
						|
   implementing the 'no' form of configuration commands. Please think carefully
 | 
						|
   before using them for anything else. There is usually a better solution, even
 | 
						|
   if it is just separating out the command definition into separate ones.
 | 
						|
-  The developer should judiciously apply separation of concerns when defining
 | 
						|
   commands. CLI definitions for two unrelated or vaguely related commands or
 | 
						|
   configuration items should be defined in separate commands. Clarity is
 | 
						|
   preferred over LOC (within reason).
 | 
						|
-  The maximum number of space-separated tokens that can be entered is
 | 
						|
   presently limited to 256. Please keep this limit in mind when
 | 
						|
   implementing new CLI.
 | 
						|
 | 
						|
Variable Names
 | 
						|
^^^^^^^^^^^^^^
 | 
						|
The parser tries to fill the "varname" field on each token. This can happen
 | 
						|
either manually or automatically. Manual specifications work by appending
 | 
						|
``$name`` after the input specifier:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   foo bar$cmd WORD$name A.B.C.D$ip
 | 
						|
 | 
						|
Note that you can also assign variable names to fixed input tokens, this can be
 | 
						|
useful if multiple commands share code. You can also use "$name" after a
 | 
						|
multiple-choice option:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   foo bar <A.B.C.D|X:X::X:X>$addr [optionA|optionB]$mode
 | 
						|
 | 
						|
The variable name is in this case assigned to the last token in each of the
 | 
						|
branches.
 | 
						|
 | 
						|
Automatic assignment of variable names works by applying the following rules:
 | 
						|
 | 
						|
-  manual names always have priority
 | 
						|
-  a ``[no]`` at the beginning receives ``no`` as varname on the ``no`` token
 | 
						|
-  ``VARIABLE`` tokens whose text is not ``WORD`` or ``NAME`` receive a cleaned
 | 
						|
   lowercase version of the token text as varname, e.g. ``ROUTE-MAP`` becomes
 | 
						|
   ``route_map``.
 | 
						|
-  other variable tokens (i.e. everything except "fixed") receive the text of
 | 
						|
   the preceding fixed token as varname, if one can be found.  E.g.
 | 
						|
   ``ip route A.B.C.D/M INTERFACE`` assigns "route" to the ``A.B.C.D/M`` token.
 | 
						|
 | 
						|
These rules should make it possible to avoid manual varname assignment in 90% of
 | 
						|
the cases.
 | 
						|
 | 
						|
Doc Strings
 | 
						|
^^^^^^^^^^^
 | 
						|
Each token in a command definition should be documented with a brief doc string
 | 
						|
that informs a user of the meaning and/or purpose of the subsequent command
 | 
						|
tree. These strings are provided as the last parameter to DEFUN macros,
 | 
						|
concatenated together and separated by an escaped newline (``\n``). These are
 | 
						|
best explained by example.
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   DEFUN (config_terminal,
 | 
						|
          config_terminal_cmd,
 | 
						|
          "configure terminal",
 | 
						|
          "Configuration from vty interface\n"
 | 
						|
          "Configuration terminal\n")
 | 
						|
 | 
						|
The last parameter is split into two lines for readability. Two newline
 | 
						|
delimited doc strings are present, one for each token in the command. The second
 | 
						|
string documents the functionality of the ``terminal`` command in the
 | 
						|
``configure`` subtree.
 | 
						|
 | 
						|
Note that the first string, for ``configure`` does not contain documentation for
 | 
						|
'terminal'. This is because the CLI is best envisioned as a tree, with tokens
 | 
						|
defining branches. An imaginary ``start`` token is the root of every command in
 | 
						|
a CLI node. Each subsequent written token descends into a subtree, so the
 | 
						|
documentation for that token ideally summarizes all the functionality contained
 | 
						|
in the subtree.
 | 
						|
 | 
						|
A consequence of this structure is that the developer must be careful to use the
 | 
						|
same doc strings when defining multiple commands that are part of the same tree.
 | 
						|
Commands which share prefixes must share the same doc strings for those
 | 
						|
prefixes. On startup the parser will generate warnings if it notices
 | 
						|
inconsistent doc strings. Behavior is undefined; the same token may show up
 | 
						|
twice in completions, with different doc strings, or it may show up once with a
 | 
						|
random doc string. Parser warnings should be heeded and fixed to avoid confusing
 | 
						|
users.
 | 
						|
 | 
						|
The number of doc strings provided must be equal to the amount of tokens present
 | 
						|
in the command definition, read left to right, ignoring any special constructs.
 | 
						|
 | 
						|
In the examples below, each arrowed token needs a doc string.
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   "show ip bgp"
 | 
						|
    ^    ^  ^
 | 
						|
 | 
						|
   "command <foo|bar> [example]"
 | 
						|
    ^        ^   ^     ^
 | 
						|
 | 
						|
DEFPY
 | 
						|
^^^^^
 | 
						|
``DEFPY(...)`` is an enhanced version of ``DEFUN()`` which is preprocessed by
 | 
						|
:file:`python/clidef.py`. The python script parses the command definition
 | 
						|
string, extracts variable names and types, and generates a C wrapper function
 | 
						|
that parses the variables and passes them on. This means that in the CLI
 | 
						|
function body, you will receive additional parameters with appropriate types.
 | 
						|
 | 
						|
This is best explained by an example. Invoking ``DEFPY`` like this:
 | 
						|
 | 
						|
.. code-block:: c
 | 
						|
 | 
						|
   DEFPY(func, func_cmd, "[no] foo bar A.B.C.D (0-99)$num", "...help...")
 | 
						|
 | 
						|
defines the handler function like this:
 | 
						|
 | 
						|
.. code-block:: c
 | 
						|
 | 
						|
   func(self, vty, argc, argv,  /* standard CLI arguments */
 | 
						|
        const char *no,         /* unparsed "no" */
 | 
						|
        struct in_addr bar,     /* parsed IP address */
 | 
						|
        const char *bar_str,    /* unparsed IP address */
 | 
						|
        long num,               /* parsed num */
 | 
						|
        const char *num_str)    /* unparsed num */
 | 
						|
 | 
						|
Note that as documented in the previous section, ``bar`` is automatically
 | 
						|
applied as variable name for ``A.B.C.D``. The Python script then detects this as
 | 
						|
an IP address argument and generates code to parse it into a ``struct in_addr``,
 | 
						|
passing it in ``bar``. The raw value is passed in ``bar_str``. The range/number
 | 
						|
argument works in the same way with the explicitly given variable name.
 | 
						|
 | 
						|
Type rules
 | 
						|
""""""""""
 | 
						|
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| Token(s)                   | Type                           | Value if omitted by user |
 | 
						|
+============================+================================+==========================+
 | 
						|
| ``A.B.C.D``                | ``struct in_addr``             | ``0.0.0.0``              |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``X:X::X:X``               | ``struct in6_addr``            | ``::``                   |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``A.B.C.D + X:X::X:X``     | ``const union sockunion *``    | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``A.B.C.D/M``              | ``const struct prefix_ipv4 *`` | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``X:X::X:X/M``             | ``const struct prefix_ipv6 *`` | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``A.B.C.D/M + X:X::X:X/M`` | ``const struct prefix *``      | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``(0-9)``                  | ``long``                       | ``0``                    |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``VARIABLE``               | ``const char *``               | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| ``word``                   | ``const char *``               | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
| *all other*                | ``const char *``               | ``NULL``                 |
 | 
						|
+----------------------------+--------------------------------+--------------------------+
 | 
						|
 | 
						|
Note the following details:
 | 
						|
 | 
						|
-  Not all parameters are pointers, some are passed as values.
 | 
						|
-  When the type is not ``const char *``, there will be an extra ``_str``
 | 
						|
   argument with type ``const char *``.
 | 
						|
-  You can give a variable name not only to ``VARIABLE`` tokens but also to
 | 
						|
   ``word`` tokens (e.g. constant words). This is useful if some parts of a
 | 
						|
   command are optional. The type will be ``const char *``.
 | 
						|
-  ``[no]`` will be passed as ``const char *no``.
 | 
						|
-  Pointers will be ``NULL`` when the argument is optional and the user did not
 | 
						|
   use it.
 | 
						|
-  If a parameter is not a pointer, but is optional and the user didn't use it,
 | 
						|
   the default value will be passed. Check the ``_str`` argument if you need to
 | 
						|
   determine whether the parameter was omitted.
 | 
						|
-  If the definition contains multiple parameters with the same variable name,
 | 
						|
   they will be collapsed into a single function parameter. The python code will
 | 
						|
   detect if the types are compatible (i.e. IPv4 + IPv6 variants) and choose a
 | 
						|
   corresponding C type.
 | 
						|
-  The standard DEFUN parameters (``self, vty, argc, argv``) are still present
 | 
						|
   and can be used. A DEFUN can simply be **edited into a DEFPY without further
 | 
						|
   changes and it will still work**; this allows easy forward migration.
 | 
						|
-  A file may contain both ``DEFUN`` and ``DEFPY`` statements.
 | 
						|
 | 
						|
Getting a parameter dump
 | 
						|
""""""""""""""""""""""""
 | 
						|
The clidef.py script can be called to get a list of DEFUNs/DEFPYs with the
 | 
						|
parameter name/type list:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   lib/clippy python/clidef.py --all-defun --show lib/plist.c > /dev/null
 | 
						|
 | 
						|
The generated code is printed to stdout, the info dump to stderr. The
 | 
						|
``--all-defun`` argument will make it process DEFUN blocks as well as DEFPYs,
 | 
						|
which is useful prior to converting some DEFUNs. **The dump does not list the
 | 
						|
``_str`` arguments** to keep the output shorter.
 | 
						|
 | 
						|
Note that the ``clidef.py`` script cannot be run with python directly, it needs
 | 
						|
to be run with *clippy* since the latter makes the CLI parser available.
 | 
						|
 | 
						|
Include & Makefile requirements
 | 
						|
"""""""""""""""""""""""""""""""
 | 
						|
A source file that uses DEFPY needs to include the ``*_clippy.c`` file **before
 | 
						|
all DEFPY statements**:
 | 
						|
 | 
						|
.. code-block:: c
 | 
						|
 | 
						|
   /* GPL header */
 | 
						|
   #include ...
 | 
						|
   ...
 | 
						|
   #ifndef VTYSH_EXTRACT_PL
 | 
						|
   #include "daemon/filename_clippy.c"
 | 
						|
   #endif
 | 
						|
 | 
						|
   DEFPY(...)
 | 
						|
   DEFPY(...)
 | 
						|
 | 
						|
   install_element(...)
 | 
						|
 | 
						|
This dependency needs to be marked in ``Makefile.am`` or ``subdir.am``: (there
 | 
						|
is no ordering requirement)
 | 
						|
 | 
						|
.. code-block:: make
 | 
						|
 | 
						|
   # ...
 | 
						|
 | 
						|
   # if linked into a LTLIBRARY (.la/.so):
 | 
						|
   filename.lo: filename_clippy.c
 | 
						|
 | 
						|
   # if linked into an executable or static library (.a):
 | 
						|
   filename.o: filename_clippy.c
 | 
						|
 | 
						|
Handlers
 | 
						|
^^^^^^^^
 | 
						|
The block that follows a CLI definition is executed when a user enters input
 | 
						|
that matches the definition. Its function signature looks like this:
 | 
						|
 | 
						|
.. code-block:: c
 | 
						|
 | 
						|
   int (*func) (const struct cmd_element *, struct vty *, int, struct cmd_token *[]);
 | 
						|
 | 
						|
The first argument is the command definition struct. The last argument is an
 | 
						|
ordered array of tokens that correspond to the path taken through the graph, and
 | 
						|
the argument just prior to that is the length of the array.
 | 
						|
 | 
						|
The arrangement of the token array has changed from Quagga's CLI implementation.
 | 
						|
In the old system, missing arguments were padded with ``NULL`` so that the same
 | 
						|
parts of a command would show up at the same indices regardless of what was
 | 
						|
entered. The new system does not perform such padding and therefore it is
 | 
						|
generally *incorrect* to assume consistent indices in this array. As a simple
 | 
						|
example:
 | 
						|
 | 
						|
Command definition:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   command [foo] <bar|baz>
 | 
						|
 | 
						|
User enters:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   command foo bar
 | 
						|
 | 
						|
Array:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   [0] -> command
 | 
						|
   [1] -> foo
 | 
						|
   [2] -> bar
 | 
						|
 | 
						|
User enters:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   command baz
 | 
						|
 | 
						|
Array:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   [0] -> command
 | 
						|
   [1] -> baz
 | 
						|
 | 
						|
 | 
						|
.. _cli-data-structures:
 | 
						|
 | 
						|
Data Structures
 | 
						|
---------------
 | 
						|
On startup, the CLI parser sequentially parses each command string definition
 | 
						|
and constructs a directed graph with each token forming a node. This graph is
 | 
						|
the basis of the entire CLI system. It is used to match user input in order to
 | 
						|
generate command completions and match commands to functions.
 | 
						|
 | 
						|
There is one graph per CLI node (not the same as a graph node in the CLI graph).
 | 
						|
The CLI node struct keeps a reference to its graph (see :file:`lib/command.h`).
 | 
						|
 | 
						|
While most of the graph maintains the form of a tree, special constructs
 | 
						|
outlined in the Rules section introduce some quirks. ``<>``, ``[]`` and ``{}``
 | 
						|
form self-contained 'subgraphs'. Each subgraph is a tree except that all of the
 | 
						|
'leaves' actually share a child node. This helps with minimizing graph size and
 | 
						|
debugging.
 | 
						|
 | 
						|
As a working example, here is the graph of the following command: ::
 | 
						|
 | 
						|
   show [ip] bgp neighbors [<A.B.C.D|X:X::X:X|WORD>] [json]
 | 
						|
 | 
						|
.. figure:: ../figures/cligraph.png
 | 
						|
   :align: center
 | 
						|
 | 
						|
   Graph of example CLI command
 | 
						|
 | 
						|
 | 
						|
``FORK`` and ``JOIN`` nodes are plumbing nodes that don't correspond to user
 | 
						|
input. They're necessary in order to deduplicate these constructs where
 | 
						|
applicable.
 | 
						|
 | 
						|
Options follow the same form, except that there is an edge from the ``FORK``
 | 
						|
node to the ``JOIN`` node. Since all of the subgraphs in the example command are
 | 
						|
optional, all of them have this edge.
 | 
						|
 | 
						|
Keywords follow the same form, except that there is an edge from ``JOIN`` to
 | 
						|
``FORK``. Because of this the CLI graph cannot be called acyclic. There is
 | 
						|
special logic in the input matching code that keeps a stack of paths already
 | 
						|
taken through the node in order to disallow following the same path more than
 | 
						|
once.
 | 
						|
 | 
						|
Variadics are a bit special; they have an edge back to themselves, which allows
 | 
						|
repeating the same input indefinitely.
 | 
						|
 | 
						|
The leaves of the graph are nodes that have no out edges. These nodes are
 | 
						|
special; their data section does not contain a token, as most nodes do, or
 | 
						|
``NULL``, as in ``FORK``/``JOIN`` nodes, but instead has a pointer to a
 | 
						|
``cmd_element``.  All paths through the graph that terminate on a leaf are
 | 
						|
guaranteed to be defined by that command. When a user enters a complete command,
 | 
						|
the command matcher tokenizes the input and executes a DFS on the CLI graph. If
 | 
						|
it is simultaneously able to exhaust all input (one input token per graph node),
 | 
						|
and then find exactly one leaf connected to the last node it reaches, then the
 | 
						|
input has matched the corresponding command and the command is executed. If it
 | 
						|
finds more than one node, then the command is ambiguous (more on this in
 | 
						|
deduplication). If it cannot exhaust all input, the command is unknown. If it
 | 
						|
exhausts all input but does not find an edge node, the command is incomplete.
 | 
						|
 | 
						|
The parser uses an incremental strategy to build the CLI graph for a node. Each
 | 
						|
command is parsed into its own graph, and then this graph is merged into the
 | 
						|
overall graph. During this merge step, the parser makes a best-effort attempt to
 | 
						|
remove duplicate nodes. If it finds a node in the overall graph that is equal to
 | 
						|
a node in the corresponding position in the command graph, it will intelligently
 | 
						|
merge the properties from the node in the command graph into the
 | 
						|
already-existing node. Subgraphs are also checked for isomorphism and merged
 | 
						|
where possible. The definition of whether two nodes are 'equal' is based on the
 | 
						|
equality of some set of token properties; read the parser source for the most
 | 
						|
up-to-date definition of equality.
 | 
						|
 | 
						|
When the parser is unable to deduplicate some complicated constructs, this can
 | 
						|
result in two identical paths through separate parts of the graph. If this
 | 
						|
occurs and the user enters input that matches these paths, they will receive an
 | 
						|
'ambiguous command' error and will be unable to execute the command. Most of the
 | 
						|
time the parser can detect and warn about duplicate commands, but it will not
 | 
						|
always be able to do this.  Hence care should be taken before defining a new
 | 
						|
command to ensure it is not defined elsewhere.
 | 
						|
 | 
						|
struct cmd\_token
 | 
						|
^^^^^^^^^^^^^^^^^
 | 
						|
 | 
						|
.. code-block:: c
 | 
						|
 | 
						|
   /* Command token struct. */
 | 
						|
   struct cmd_token
 | 
						|
   {
 | 
						|
           enum cmd_token_type type; // token type
 | 
						|
           uint8_t attr;             // token attributes
 | 
						|
           bool allowrepeat;         // matcher can match token repetitively?
 | 
						|
 | 
						|
           char *text;               // token text
 | 
						|
           char *desc;               // token description
 | 
						|
           long long min, max;       // for ranges
 | 
						|
           char *arg;                // user input that matches this token
 | 
						|
           char *varname;            // variable name
 | 
						|
   };
 | 
						|
 | 
						|
This struct is used in the CLI graph to match input against. It is also used to
 | 
						|
pass user input to command handler functions, as it is frequently useful for
 | 
						|
handlers to have access to that information. When a command is matched, the
 | 
						|
sequence of ``cmd_tokens`` that form the matching path are duplicated and placed
 | 
						|
in order into ``*argv[]``. Before this happens the ``->arg`` field is set to
 | 
						|
point at the snippet of user input that matched it.
 | 
						|
 | 
						|
For most nontrivial commands the handler function will need to determine which
 | 
						|
of the possible matching inputs was entered. Previously this was done by looking
 | 
						|
at the first few characters of input. This is now considered an anti-pattern and
 | 
						|
should be avoided. Instead, the ``->type`` or ``->text`` fields for this logic.
 | 
						|
The ``->type`` field can be used when the possible inputs differ in type. When
 | 
						|
the possible types are the same, use the ``->text`` field. This field has the
 | 
						|
full text of the corresponding token in the definition string and using it makes
 | 
						|
for much more readable code. An example is helpful.
 | 
						|
 | 
						|
Command definition:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   command <(1-10)|foo|BAR>
 | 
						|
 | 
						|
In this example, the user may enter any one of:
 | 
						|
- an integer between 1 and 10
 | 
						|
- "foo"
 | 
						|
- anything at all
 | 
						|
 | 
						|
If the user enters "command f", then:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   argv[1]->type == WORD_TKN
 | 
						|
   argv[1]->arg  == "f"
 | 
						|
   argv[1]->text == "foo"
 | 
						|
 | 
						|
Range tokens have some special treatment; a token with ``->type == RANGE_TKN``
 | 
						|
will have the ``->min`` and ``->max`` fields set to the bounding values of the
 | 
						|
range.
 | 
						|
 | 
						|
struct cmd\_element
 | 
						|
^^^^^^^^^^^^^^^^^^^
 | 
						|
 | 
						|
.. code-block:: c
 | 
						|
 | 
						|
   struct cmd_node {
 | 
						|
           /* Node index. */
 | 
						|
           enum node_type node;
 | 
						|
 | 
						|
           /* Prompt character at vty interface. */
 | 
						|
           const char *prompt;
 | 
						|
 | 
						|
           /* Is this node's configuration goes to vtysh ? */
 | 
						|
           int vtysh;
 | 
						|
 | 
						|
           /* Node's configuration write function */
 | 
						|
           int (*func)(struct vty *);
 | 
						|
 | 
						|
           /* Node's command graph */
 | 
						|
           struct graph *cmdgraph;
 | 
						|
 | 
						|
           /* Vector of this node's command list. */
 | 
						|
           vector cmd_vector;
 | 
						|
 | 
						|
           /* Hashed index of command node list, for de-dupping primarily */
 | 
						|
           struct hash *cmd_hash;
 | 
						|
   };
 | 
						|
 | 
						|
This struct corresponds to a CLI mode. The last three fields are most relevant
 | 
						|
here.
 | 
						|
 | 
						|
cmdgraph
 | 
						|
   This is a pointer to the command graph that was described in the first part
 | 
						|
   of this section. It is the datastructure used for matching user input to
 | 
						|
   commands.
 | 
						|
 | 
						|
cmd_vector
 | 
						|
   This is a list of all the ``struct cmd_element`` defined in the mode.
 | 
						|
 | 
						|
cmd_hash
 | 
						|
   This is a hash table of all the ``struct cmd_element`` defined in the mode.
 | 
						|
   When ``install_element`` is called, it checks that the element it is given is
 | 
						|
   not already present in the hash table as a safeguard against duplicate calls
 | 
						|
   resulting in a command being defined twice, which renders the command
 | 
						|
   ambiguous.
 | 
						|
 | 
						|
All ``struct cmd_node`` are themselves held in a static vector defined in
 | 
						|
:file:`lib/command.c` that defines the global CLI space.
 | 
						|
 | 
						|
Command Abbreviation & Matching Priority
 | 
						|
----------------------------------------
 | 
						|
It is possible for users to elide parts of tokens when the CLI matcher does not
 | 
						|
need them to make an unambiguous match. This is best explained by example.
 | 
						|
 | 
						|
Command definitions:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   command dog cow
 | 
						|
   command dog crow
 | 
						|
 | 
						|
User input:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   c d c         -> ambiguous command
 | 
						|
   c d co        -> match "command dog cow"
 | 
						|
 | 
						|
 | 
						|
The parser will look ahead and attempt to disambiguate the input based on tokens
 | 
						|
later on in the input string.
 | 
						|
 | 
						|
Command definitions:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   show ip bgp A.B.C.D
 | 
						|
   show ipv6 bgp X:X::X:X
 | 
						|
 | 
						|
User enters:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   s i b 4.3.2.1         -> match "show ip bgp A.B.C.D"
 | 
						|
   s i b ::e0            -> match "show ipv6 bgp X:X::X:X"
 | 
						|
 | 
						|
Reading left to right, both of these commands would be ambiguous since 'i' does
 | 
						|
not explicitly select either 'ip' or 'ipv6'. However, since the user later
 | 
						|
provides a token that matches only one of the commands (an IPv4 or IPv6 address)
 | 
						|
the parser is able to look ahead and select the appropriate command. This has
 | 
						|
some implications for parsing the ``*argv[]`` that is passed to the command
 | 
						|
handler.
 | 
						|
 | 
						|
Now consider a command definition such as:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   command <foo|VAR>
 | 
						|
 | 
						|
'foo' only matches the string 'foo', but 'VAR' matches any input, including
 | 
						|
'foo'. Who wins? In situations like this the matcher will always choose the
 | 
						|
'better' match, so 'foo' will win.
 | 
						|
 | 
						|
Consider also:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   show <ip|ipv6> foo
 | 
						|
 | 
						|
User input:
 | 
						|
 | 
						|
::
 | 
						|
 | 
						|
   show ip foo
 | 
						|
 | 
						|
``ip`` partially matches ``ipv6`` but exactly matches ``ip``, so ``ip`` will
 | 
						|
win.
 | 
						|
 | 
						|
Inspection & Debugging
 | 
						|
----------------------
 | 
						|
 | 
						|
Permutations
 | 
						|
^^^^^^^^^^^^
 | 
						|
It is sometimes useful to check all the possible combinations of input that
 | 
						|
would match an arbitrary definition string. There is a tool in
 | 
						|
:file:`tools/permutations` that reads CLI definition strings on ``stdin`` and
 | 
						|
prints out all matching input permutations. It also dumps a text representation
 | 
						|
of the graph, which is more useful for debugging than anything else. It looks
 | 
						|
like this:
 | 
						|
 | 
						|
.. code-block:: shell
 | 
						|
 | 
						|
   $ ./permutations "show [ip] bgp [<view|vrf> WORD]"
 | 
						|
 | 
						|
   show ip bgp view WORD
 | 
						|
   show ip bgp vrf WORD
 | 
						|
   show ip bgp
 | 
						|
   show bgp view WORD
 | 
						|
   show bgp vrf WORD
 | 
						|
   show bgp
 | 
						|
 | 
						|
This functionality is also built into VTY/VTYSH; :clicmd:`list permutations`
 | 
						|
will list all possible matching input permutations in the current CLI node.
 | 
						|
 | 
						|
Graph Inspection
 | 
						|
^^^^^^^^^^^^^^^^
 | 
						|
When in the Telnet or VTYSH console, :clicmd:`show cli graph` will dump the
 | 
						|
entire command space of the current mode in the DOT graph language. This can be
 | 
						|
fed into one of the various GraphViz layout engines, such as ``dot``,
 | 
						|
``neato``, etc.
 | 
						|
 | 
						|
For example, to generate an image of the entire command space for the top-level
 | 
						|
mode (``ENABLE_NODE``):
 | 
						|
 | 
						|
.. code-block:: shell
 | 
						|
 | 
						|
   sudo vtysh -c 'show cli graph' | dot -Tjpg -Grankdir=LR > graph.jpg
 | 
						|
 | 
						|
To do the same for the BGP mode:
 | 
						|
 | 
						|
.. code-block:: shell
 | 
						|
 | 
						|
   sudo vtysh -c 'conf t' -c 'router bgp' -c 'show cli graph' | dot -Tjpg -Grankdir=LR > bgpgraph.jpg
 | 
						|
 | 
						|
This information is very helpful when debugging command resolution, tracking
 | 
						|
down duplicate / ambiguous commands, and debugging patches to the CLI graph
 | 
						|
builder.
 |