path: root/system/doc/design_principles/statem.xml

<?xml version="1.0" encoding="utf-8" ?>
<!DOCTYPE chapter SYSTEM "chapter.dtd">

<chapter>
  <header>
    <copyright>
      <year>2016</year>
      <holder>Ericsson AB.  All Rights Reserved.</holder>
    </copyright>
    <legalnotice>
      Licensed under the Apache License, Version 2.0 (the "License");
      you may not use this file except in compliance with the License.
      You may obtain a copy of the License at

          http://www.apache.org/licenses/LICENSE-2.0

      Unless required by applicable law or agreed to in writing, software
      distributed under the License is distributed on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
      See the License for the specific language governing permissions and
      limitations under the License.

    </legalnotice>

    <title>gen_statem Behaviour</title>
    <prepared></prepared>
    <docno></docno>
    <date></date>
    <rev></rev>
    <file>statem.xml</file>
  </header>
  <marker id="gen_statem behaviour"></marker>
  <p>
    This section is to be read with the
    <seealso marker="stdlib:gen_statem"><c>gen_statem(3)</c></seealso>
    manual page in STDLIB, where all interface functions and callback
    functions are described in detail.
  </p>
  <p>
    This is a new behaviour in OTP-19.0.
    It has been thoroughly reviewed, is stable enough
    to be used by at least two heavy OTP applications, and is here to stay.
    But depending on user feedback, we do not expect
    but might find it necessary to make minor
    not backwards compatible changes into OTP-20.0,
    so its state can be designated as "not quite experimental"...
  </p>

<!-- =================================================================== -->

  <section>
    <title>Event Driven State Machines</title>
    <p>
      Established Automata theory does not deal much with
      how a state transition is triggered,
      but in general assumes that the output is a function
      of the input (and the state) and that they are
      some kind of values.
    </p>
    <p>
      For an Event Driven State Machine the input is an event
      that triggers a state transition and the output
      is actions executed during the state transition.
      It can analogously to the mathematical model of a
      Finite State Machine be described as
      a set of relations of the form:
    </p>
    <pre>
State(S) x Event(E) -> Actions(A), State(S')</pre>
    <p>These relations are interpreted as meaning:</p>
    <p>
      If we are in state <c>S</c> and event <c>E</c> occurs, we
      are to perform actions <c>A</c> and make a transition to
      state <c>S'</c>.
    </p>
    <p>
      Note that <c>S'</c> may be equal to <c>S</c>.
    </p>
    <p>
      Since <c>A</c> and <c>S'</c> depend only on
      <c>S</c> and <c>E</c> the kind of state machine described
      here is a Mealy Machine.
      (See for example the corresponding Wikipedia article)
    </p>
    <p>
      Like most <c>gen_</c> behaviours, <c>gen_statem</c> keeps
      a server <c>Data</c> besides the state.  This and the fact that
      there is no restriction on the number of states
      (assuming enough virtual machine memory)
      or on the number of distinct input events actually makes
      a state machine implemented with this behaviour Turing complete.
      But it feels mostly like an Event Driven Mealy Machine.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <marker id="callback_modes" />
    <title>Callback Modes</title>
    <p>
      The <c>gen_statem</c> behaviour supports two different callback modes.
      In the mode
      <seealso marker="stdlib:gen_statem#type-callback_mode">
	<c>state_functions</c>,
      </seealso>
      the state transition rules are written as a number of Erlang
      functions, which conform to the following convention:
    </p>
    <pre>
StateName(EventType, EventContent, Data) ->
    .. code for actions here ...
    {next_state, NewStateName, NewData}.</pre>
    <p>
      In the mode
      <seealso marker="stdlib:gen_statem#type-callback_mode">
	<c>handle_event_function</c>
      </seealso>
      there is only one
      Erlang function that implements all state transition rules:
    </p>
    <pre>
handle_event(EventType, EventContent, State, Data) ->
    .. code for actions here ...
    {next_state, State', Data'}</pre>
    <p>
      Both these modes allow other return tuples
      that you can find in the
      <seealso marker="stdlib:gen_statem#Module:StateName/3">
	reference manual.
      </seealso>
      These other return tuples can for example stop the machine,
      execute state transition actions on the machine engine itself
      and send replies.
    </p>

    <section>
      <title>Choosing the Callback Mode</title>
      <p>
	The two
	<seealso marker="#callback_modes">callback modes</seealso>
	gives different possibilities
	and restrictions, but one goal remains:
	you want to handle all possible combinations of
	events and states.
      </p>
      <p>
	You can for example do this by focusing on one state at the time
	and for every state ensure that all events are handled,
	or the other way around focus on one event at the time
	and ensure that it is handled in every state,
	or mix these strategies.
      </p>
      <p>
	With <c>state_functions</c> you are restricted to use
	atom only states, and the <c>gen_statem</c> engine dispatches
	on state name for you.  This encourages the callback module
	to gather the implementation of all event actions particular
	to one state in the same place in the code
	hence to focus on one state at the time.
      </p>
      <p>
	This mode fits well when you have a regular state diagram
	like the ones in this chapter that describes all events and actions
	belonging to a state visually around that state,
	and each state has its unique name.
      </p>
      <p>
	With <c>handle_event_function</c> you are free to mix strategies
	as you like because all events and states
	are handled in the the same callback function.
      </p>
      <p>
	This mode works equally well when you want to focus on
	one event at the time or when you want to focus on
	one state at the time, but the <c>handle_event/4</c> function
	quickly grows too large to handle without introducing dispatching.
      </p>
      <p>
	The mode enables the use of non-atom states for example
	complex states or even hiearchical states.
	If, for example, a state diagram is largely alike
	for the client and for the server side of a protocol;
	then you can have a state <c>{StateName,server}</c> or
	<c>{StateName,client}</c> and since you do the dispatching
	yourself you make <c>StateName</c> decide where in the code
	to handle most events in the state.
	The second element of the tuple	is then used to select
	whether to handle special client side or server side events.
      </p>
    </section>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Example</title>
    <p>
      This is an example starting off as equivalent to the the example in the
      <seealso marker="fsm"><c>gen_fsm</c> behaviour</seealso>
      description.  In later chapters additions and tweaks are made
      using features in <c>gen_statem</c> that <c>gen_fsm</c> does not have.
      At the end of this section you can find the example again
      with all the added features.
    </p>
    <p>
      A door with a code lock can be viewed as a state machine.
      Initially, the door is locked.  Anytime someone presses a button,
      this generates an event.
      Depending on what buttons have been pressed before,
      the sequence so far can be correct, incomplete, or wrong.
    </p>
    <p>
      If it is correct, the door is unlocked for 10 seconds (10000 ms).
      If it is incomplete, we wait for another button to be pressed.  If
      it is is wrong, we start all over,
      waiting for a new button sequence.
    </p>
    <image file="../design_principles/code_lock.png">
      <icaption>Code lock state diagram</icaption>
    </image>
    <p>
      We can implement such a code lock state machine using
      <c>gen_statem</c> with the following callback module:
    </p>
    <marker id="ex"></marker>
    <code type="erl"><![CDATA[
-module(code_lock).
-behaviour(gen_statem).
-define(NAME, code_lock).
-define(CALLBACK_MODE, state_functions).

-export([start_link/1]).
-export([button/1]).
-export([init/1,terminate/3,code_change/4]).
-export([locked/3,open/3]).

start_link(Code) ->
    gen_statem:start_link({local,?NAME}, ?MODULE, Code, []).

button(Digit) ->
    gen_statem:cast(?NAME, {button,Digit}).


init(Code) ->
    do_lock(),
    Data = #{code => Code, remaining => Code},
    {?CALLBACK_MODE,locked,Data}.

locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] ->
	    do_unlock(),
            {next_state,open,Data#{remaining := Code},10000};
        [Digit|Rest] -> % Incomplete
            {next_state,locked,Data#{remaining := Rest}};
        _Wrong ->
            {next_state,locked,Data#{remaining := Code}}
    end.

open(timeout, _,  Data) ->
    do_lock(),
    {next_state,locked,Data};
open(cast, {button,_}, Data) ->
    do_lock(),
    {next_state,locked,Data}.

do_lock() ->
    io:format("Lock~n", []).
do_unlock() ->
    io:format("Unlock~n", []).

terminate(_Reason, State, _Data) ->
    State =/= locked andalso do_lock(),
    ok.
code_change(_Vsn, State, Data, _Extra) ->
    {?CALLBACK_MODE,State,Data}.
    ]]></code>
    <p>The code is explained in the next sections.</p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Starting gen_statem</title>
    <p>
      In the example in the previous section, the <c>gen_statem</c> is
      started by calling <c>code_lock:start_link(Code)</c>:
    </p>
    <code type="erl"><![CDATA[
start_link(Code) ->
    gen_statem:start_link({local,?NAME}, ?MODULE, Code, []).
    ]]></code>
    <p>
      <c>start_link</c> calls the function
      <seealso marker="stdlib:gen_statem#start_link/4">
	<c>gen_statem:start_link/4</c>
      </seealso>
      which spawns and links to a new process; a <c>gen_statem</c>.
    </p>
    <list type="bulleted">
      <item>
        <p>
	  The first argument, <c>{local,?NAME}</c>, specifies
          the name.  In this case, the <c>gen_statem</c> is locally
	  registered as <c>code_lock</c> through the macro <c>?NAME</c>.
	</p>
        <p>
	  If the name is omitted, the <c>gen_statem</c> is not registered.
          Instead its pid must be used.  The name can also be given
	  as <c>{global,Name}</c>, in which case the <c>gen_statem</c> is
	  registered using
	  <seealso marker="kernel:global#register_name/2">
	    <c>global:register_name/2</c>.
	  </seealso>
	</p>
      </item>
      <item>
        <p>
	  The second argument, <c>?MODULE</c>, is the name of
          the callback module, that is; the module where the callback
          functions are located, which is this module.
	</p>
        <p>
	  The interface functions (<c>start_link/1</c> and <c>button/1</c>)
	  are located in the same module as the callback functions
	  (<c>init/1</c>, <c>locked/3</c>, and <c>open/3</c>).
	  It is normally good programming practice to have the client
	  side and the server side code contained in one module.
	</p>
      </item>
      <item>
        <p>
	  The third argument, <c>Code</c>, is a list of digits that
	  is the correct unlock code which is passsed
	  to the callback function <c>init/1</c>.
	</p>
      </item>
      <item>
        <p>
	  The fourth argument, <c>[]</c>, is a list of options.  See the
	  <seealso marker="stdlib:gen_statem#start_link/3">
	    <c>gen_statem:start_link/3</c>
	  </seealso>
	  manual page for available options.
	</p>
      </item>
    </list>
    <p>
      If name registration succeeds, the new <c>gen_statem</c> process
      calls the callback function <c>code_lock:init(Code)</c>.
      This function is expected to return <c>{CallbackMode,State,Data}</c>,
      where
      <seealso marker="#callback_modes">
	<c>CallbackMode</c>
      </seealso>
      selects callback module state function mode, in this case
      <seealso marker="stdlib:gen_statem#type-callback_mode">
	<c>state_functions</c>
      </seealso>
      through the macro <c>?CALLBACK_MODE</c> that is; each state
      has got its own handler function.
      <c>State</c> is the initial state of the <c>gen_statem</c>,
      in this case <c>locked</c>; assuming the door is locked to begin with.
      <c>Data</c> is the internal server data of the <c>gen_statem</c>.
      Here the server data is a <seealso marker="stdlib:maps">map</seealso>
      with the key <c>code</c> that stores
      the correct button sequence and the key <c>remaining</c>
      that stores the remaining correct button sequence
      (the same as the <c>code</c> to begin with).
    </p>
    <code type="erl"><![CDATA[
init(Code) ->
    do_lock(),
    Data = #{code => Code, remaining => Code},
    {?CALLBACK_MODE,locked,Data}.
    ]]></code>
    <p>
      <seealso marker="stdlib:gen_statem#start_link/3">
	<c>gen_statem:start_link</c>
      </seealso>
      is synchronous.  It does not return until the <c>gen_statem</c>
      has been initialized and is ready to receive events.
    </p>
    <p>
      <seealso marker="stdlib:gen_statem#start_link/3">
	<c>gen_statem:start_link</c>
      </seealso>
      must be used if the <c>gen_statem</c>
      is part of a supervision tree, that is; started by a supervisor.
      There is another function;
      <seealso marker="stdlib:gen_statem#start/3">
	<c>gen_statem:start</c>
      </seealso>
      to start a standalone <c>gen_statem</c>, that is;
      a <c>gen_statem</c> that is not part of a supervision tree.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Events and Handling them</title>
    <p>The function notifying the code lock about a button event is
      implemented using
      <seealso marker="stdlib:gen_statem#cast/2">
	<c>gen_statem:cast/2</c>:
      </seealso>
    </p>
    <code type="erl"><![CDATA[
button(Digit) ->
    gen_statem:cast(?NAME, {button,Digit}).
    ]]></code>
    <p>
      The first argument is the name of the <c>gen_statem</c> and must
      agree with the name used to start it so therefore we use the
      same macro <c>?NAME</c> as when starting.
      <c>{button,Digit}</c> is the actual event content.
    </p>
    <p>
      The event is made into a message and sent to the <c>gen_statem</c>.
      When the event is received, the <c>gen_statem</c> calls
      <c>StateName(cast, Event, Data)</c>, which is expected to
      return a tuple <c>{next_state,NewStateName,NewData}</c>.
      <c>StateName</c> is the name of the current state and
      <c>NewStateName</c> is the name of the next state to go to.
      <c>NewData</c> is a new value for the server data of
      the <c>gen_statem</c>.
    </p>
    <code type="erl"><![CDATA[
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] -> % Complete
	    do_unlock(),
            {next_state,open,Data#{remaining := Code},10000};
        [Digit|Rest] -> % Incomplete
            {next_state,locked,Data#{remaining := Rest}};
        [_|_] -> % Wrong
            {next_state,locked,Data#{remaining := Code}}
    end.

open(timeout, _, Data) ->
    do_lock(),
    {next_state,locked,Data};
open(cast, {button,_}, Data) ->
    do_lock(),
    {next_state,locked,Data}.
    ]]></code>
    <p>
      If the door is locked and a button is pressed, the pressed
      button is compared with the next correct button and,
      depending on the result, the door is either unlocked
      and the <c>gen_statem</c> goes to state <c>open</c>,
      or the door remains in state <c>locked</c>.
    </p>
    <p>
      If the pressed button is incorrect the server data
      restarts from the start of the code sequence.
    </p>
    <p>
      In state <c>open</c> any button locks the door since
      any event cancels the event timer so we will not get
      a timeout event after a button event.
    </p>
  </section>

  <section>
    <title>Event Time-Outs</title>
    <p>
      When a correct code has been given, the door is unlocked and
      the following tuple is returned from <c>locked/2</c>:
    </p>
    <code type="erl"><![CDATA[
{next_state,open,Data#{remaining := Code},10000};
    ]]></code>
    <p>
      10000 is a time-out value in milliseconds.
      After this time, that is; 10 seconds, a time-out occurs.
      Then, <c>StateName(timeout, 10000, Data)</c> is called.
      The time-out occurs when the door has been in state <c>open</c>
      for 10 seconds.  After that the door is locked again:
    </p>
    <code type="erl"><![CDATA[
open(timeout, _,  Data) ->
    do_lock(),
    {next_state,locked,Data};
    ]]></code>
  </section>

<!-- =================================================================== -->

  <section>
    <title>All State Events</title>
    <p>
      Sometimes an event can arrive in any state of the <c>gen_statem</c>.
      It is convenient to handle these in a common state handler function
      that all state functions call for events not specific to the state.
    </p>
    <p>
      Let's introduce a <c>code_length/0</c> function that returns
      the length of the correct code
      (that should not be sensitive to reveal...).
      We'll dispatch all events that are not state specific
      to the common function <c>handle_event/3</c>.
    </p>
    <code type="erl"><![CDATA[
...
-export([button/1,code_length/0]).
...

code_length() ->
    gen_statem:call(?NAME, code_length).

...
locked(...) -> ... ;
locked(EventType, EventContent, Data) ->
    handle_event(EventType, EventContent, Data).

...
open(...) -> ... ;
open(EventType, EventContent, Data) ->
    handle_event(EventType, EventContent, Data).

handle_event({call,From}, code_length, #{code := Code} = Data) ->
    {keep_state,Data,[{reply,From,length(Code)}]}.
    ]]></code>
    <p>
      This example uses
      <seealso marker="stdlib:gen_statem#call/2">
	<c>gen_statem:call/2</c>
      </seealso>
      which waits for a reply from the server.
      The reply is sent with a <c>{reply,From,Reply}</c> tuple
      in an action list in the <c>{keep_state,...}</c> tuple
      that retains the current state.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>One Event Handler</title>
    <p>
      If you use the mode <c>handle_event_function</c>
      all events are handled in <c>handle_event/4</c> and we
      may (but do not have to) use an event-centered approach
      where we dispatch on event first and then state:
    </p>
    <code type="erl"><![CDATA[
...
-define(CALLBACK_MODE, state_functions).

...
-export([handle_event/4]).

...

handle_event(cast, {button,Digit}, State, #{code := Code} = Data) ->
    case State of
	locked ->
	    case maps:get(remaining, Data) of
		[Digit] -> % Complete
		    do_unlock(),
		    {next_state,open,Data#{remaining := Code},10000};
		[Digit|Rest] -> % Incomplete
		    {keep_state,Data#{remaining := Rest}};
		[_|_] -> % Wrong
		    {keep_state,Data#{remaining := Code}}
	    end;
	open ->
	    do_lock(),
	    {next_state,locked,Data}
    end;
handle_event(timeout, _, open, Data) ->
    do_lock(),
    {next_state,locked,Data}.

...
    ]]></code>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Stopping</title>

    <section>
      <title>In a Supervision Tree</title>
      <p>
	If the <c>gen_statem</c> is part of a supervision tree,
	no stop function is needed.
	The <c>gen_statem</c> is automatically terminated by its supervisor.
	Exactly how this is done is defined by a
	<seealso marker="sup_princ#shutdown">shutdown strategy</seealso>
        set in the supervisor.
      </p>
      <p>
	If it is necessary to clean up before termination, the shutdown
        strategy must be a time-out value and the <c>gen_statem</c> must
	in the <c>init/1</c> function set itself to trap exit signals
	by calling
	<seealso marker="erts:erlang#process_flag/2">
	  <c>process_flag(trap_exit, true)</c>.
	</seealso>
	When ordered to shutdown, the <c>gen_statem</c> then calls
	the callback function
	<c>terminate(shutdown, State, Data)</c>:
      </p>
      <code type="erl"><![CDATA[
init(Args) ->
    process_flag(trap_exit, true),
    do_lock(),
    ...
      ]]></code>
      <p>
	In this example we let the <c>terminate/3</c> function
	lock the door if it is open so we do not accidentally leave the door
	open when the supervision tree terminates.
      </p>
      <code type="erl"><![CDATA[
terminate(_Reason, State, _Data) ->
    State =/= locked andalso do_lock(),
    ok.
      ]]></code>
    </section>

    <section>
      <title>Standalone gen_statem</title>
      <p>
	If the <c>gen_statem</c> is not part of a supervision tree,
	it can be stopped using
	<seealso marker="stdlib:gen_statem#stop/1">
	  <c>gen_statem:stop</c>,
	</seealso>
	preferably through an API function:
      </p>
      <code type="erl"><![CDATA[
...
-export([start_link/1,stop/0]).

...
stop() ->
    gen_statem:stop(?NAME).
      ]]></code>
      <p>
	This makes the <c>gen_statem</c> call the <c>terminate/3</c>
	callback function just like for a supervised server
	and waits for the process to terminate.
      </p>
    </section>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Actions</title>
    <p>
      In the first chapters we mentioned actions as a part of
      the general state machine model, and these actions
      are implemented with the code the <c>gen_statem</c>
      callback module executes in an event handling
      callback function before returning
      to the <c>gen_statem</c> engine.
    </p>
    <p>
      There are more specific state transition actions
      that a callback function can order the <c>gen_statem</c>
      engine to do after the callback function return.
      These are ordered by returning a list of
      <seealso marker="stdlib:gen_statem#type-action">
	actions
      </seealso>
      in the
      <seealso marker="stdlib:gen_statem#type-state_function_result">
	return tuple
      </seealso>
      from the
      <seealso marker="stdlib:gen_statem#Module:StateName/3">
	callback function.
      </seealso>
      These state transition actions affect the <c>gen_statem</c>
      engine itself.  They can:
    </p>
    <list type="bulleted">
      <item>Postpone the current event.</item>
      <item>Hibernate the <c>gen_statem</c>.</item>
      <item>Start an event timeout.</item>
      <item>Reply to a caller.</item>
      <item>Generate the next event to handle.</item>
    </list>
    <p>
      We have mentioned the event timeout
      and replying to a caller in the example above.
      An example of event postponing comes in later in this chapter.
      See the
      <seealso marker="stdlib:gen_statem#type-action">
	reference manual
      </seealso>
      for details.  You can for example actually reply to several callers
      and generate multiple next events to handle.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Event Types</title>
    <p>
      So far we have mentioned a few
      <seealso marker="stdlib:gen_statem#type-event_type">
	event types.
      </seealso>
      Events of all types are handled in the same callback function,
      for a given state, and the function gets
      <c>EventType</c> and <c>EventContent</c> as arguments.
    </p>
    <p>
      Here is the complete list of event types and where
      they come from:
    </p>
    <taglist>
      <tag><c>cast</c></tag>
      <item>
	Generated by
	<seealso marker="stdlib:gen_statem#cast/2">
	  <c>gen_statem:cast</c>.
	</seealso>
      </item>
      <tag><c>{call,From}</c></tag>
      <item>
	Generated by
	<seealso marker="stdlib:gen_statem#call/2">
	  <c>gen_statem:call</c>
	</seealso>
	where <c>From</c> is the reply address to use
	when replying either through the state transition action
	<c>{reply,From,Msg}</c> or by calling
	<seealso marker="stdlib:gen_statem#reply/1">
	  <c>gen_statem:reply</c>.
	</seealso>
      </item>
      <tag><c>info</c></tag>
      <item>
	Generated by any regular process message sent to
	the <c>gen_statem</c> process.
      </item>
      <tag><c>timeout</c></tag>
      <item>
	Generated by the state transition action
	<c>{timeout,Time,EventContent}</c> (or its short form <c>Time</c>)
	timer timing out.
      </item>
      <tag><c>internal</c></tag>
      <item>
	Generated by the state transition action
	<c>{next_event,internal,EventContent}</c>.
	In fact all event types above can be generated using
	<c>{next_event,EventType,EventContent}</c>.
      </item>
    </taglist>
  </section>

<!-- =================================================================== -->

  <section>
    <title>State Timeouts</title>
    <p>
      The timeout event generated by the state transition action
      <c>{timeout,Time,EventContent}</c> is an event timeout,
      that is; if an event arrives the timer is cancelled.
      You get either an event or a timeout but not both.
    </p>
    <p>
      Often you want a timer to not be cancelled by any event
      or you want to start a timer in one state and respond
      to the timeout in another.  This can be accomplished
      with a regular erlang timer:
      <seealso marker="erts:erlang#start_timer/4">
	<c>erlang:start_timer</c>.
      </seealso>
    </p>
    <p>
      Looking at the example in this chapter so far; using the
      <c>gen_statem</c> event timer has the consequence that
      if a button event is generated while in the <c>open</c> state,
      the timeout is cancelled and the button event is delivered.
      Therefore we chose to lock the door if this happended.
    </p>
    <p>
      Suppose we do not want a button to lock the door,
      instead we want to ignore button events in the <c>open</c> state.
      Then we start a timer when entering the <c>open</c> state
      and wait for it to expire while ignoring button events:
    </p>
    <code type="erl"><![CDATA[
...
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] ->
	    do_unlock(),
	    Tref = erlang:start_timer(10000, self(), lock),
            {next_state,open,Data#{remaining := Code, timer := Tref}};
...

open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) ->
    do_lock(),
    {next_state,locked,Data};
open(cast, {button,_}, Data) ->
    {keep_state,Data};
...
    ]]></code>
    <p>
      If you need to cancel a timer due to some other event you can use
      <seealso marker="erts:erlang#cancel_timer/2">
	<c>erlang:cancel_timer(Tref)</c>.
      </seealso>
      Note that a timeout message can not arrive after this,
      unless you have postponed it (see the next section) before,
      so make sure you do not accidentally postpone such messages.
    </p>
    <p>
      Another way to cancel a timer is to not cancel it,
      but instead to ignore it if it arrives in a state
      where it is known to be late.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Postponing Events</title>
    <p>
      If you want to ignore a particular event in the current state
      and handle it in a future state, you can postpone the event.
      A postponed event is retried after the state has
      changed i.e <c>OldState =/= NewState</c>.
    </p>
    <p>
      Postponing is ordered by the state transition
      <seealso marker="stdlib:gen_statem#type-action">
	action
      </seealso>
      <c>postpone</c>.
    </p>
    <p>
      In this example, instead of ignoring button events
      while in the <c>open</c> state we can postpone them
      and they will be queued and later handled in the <c>locked</c> state:
    </p>
    <code type="erl"><![CDATA[
...
open(cast, {button,_}, Data) ->
    {keep_state,Data,[postpone]};
...
    ]]></code>
    <p>
      The fact that a postponed event is only retried after a state change
      translates into a requirement on the event and state space:
      if you have a choice between storing a state data item
      in the <c>State</c> or in the <c>Data</c>;
      should a change in the item value affect which events that
      are handled, then this item ought to be part of the state.
    </p>
    <p>
      What you want to avoid is that you maybe much later decide
      to postpone an event in one state and by misfortune it is never retried
      because the code only changes the <c>Data</c> but not the <c>State</c>.
    </p>

    <section>
      <title>Fuzzy State Diagrams</title>
      <p>
	It is not uncommon that a state diagram does not specify
	how to handle events that are not illustrated
	in a particular state in the diagram.
	Hopefully this is described in an associated text
	or from the context.
      </p>
      <p>
	Possible actions may be; ignore as in drop the event
	(maybe log it) or deal with the event in some other state
	as in postpone it.
      </p>
    </section>

    <section>
      <title>Selective Receive</title>
      <p>
	Erlang's selective receive statement is often used to
	describe simple state machine examples in straightforward
	Erlang code.  Here is a possible implementation of
	the first example:
      </p>
    <code type="erl"><![CDATA[
-module(code_lock).
-define(NAME, code_lock_1).
-export([start_link/1,button/1]).

start_link(Code) ->
    spawn(
      fun () ->
	      true = register(?NAME, self()),
	      do_lock(),
	      locked(Code, Code)
      end).

button(Digit) ->
    ?NAME ! {button,Digit}.

locked(Code, [Digit|Remaining]) ->
    receive
	{button,Digit} when Remaining =:= [] ->
	    do_unlock(),
	    open(Code);
	{button,Digit} ->
	    locked(Code, Remaining);
	{button,_} ->
	    locked(Code, Code)
    end.

open(Code) ->
    receive
    after 10000 ->
	    do_lock(),
	    locked(Code, Code)
    end.

do_lock() ->
    io:format("Locked~n", []).
do_unlock() ->
    io:format("Open~n", []).
    ]]></code>
    <p>
      The selective receive in this case causes <c>open</c>
      to implicitly postpone any events to the <c>locked</c> state.
    </p>
    <p>
      A selective receive can not be used from a <c>gen_statem</c>
      behaviour just as for any <c>gen_*</c> behavior
      since the receive statement is within the <c>gen_*</c> engine itself.
      It has to be there because all
      <seealso marker="stdlib:sys"><c>sys</c></seealso>
      compatible behaviours must respond to system messages and therefore
      do that in their engine receive loop,
      passing non-system messages to the callback module.
    </p>
    <p>
      The state transition
      <seealso marker="stdlib:gen_statem#type-action">
	action
      </seealso>
      <c>postpone</c> is designed to be able to model
      selective receives.  A selective receive implicitly postpones
      any not received events, but the <c>postpone</c>
      state transition action explicitly postpones one received event.
    </p>
    <p>
      Other than that both mechanisms have got the same theoretical
      time and memory complexity, while the selective receive
      language construct has got smaller constant factors.
    </p>
    </section>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Self Generated Events</title>
    <p>
      It may be beneficial in some cases to be able to generate events
      to your own state machine.
      This can be done with the state transition
      <seealso marker="stdlib:gen_statem#type-action">
	action
      </seealso>
      <c>{next_event,EventType,EventContent}</c>.
    </p>
    <p>
      You can generate events of any existing
      <seealso marker="stdlib:gen_statem#type-action">
	type,
      </seealso>
      but the <c>internal</c> type can only be generated through the
      <c>next_event</c> action and hence can not come from an external source,
      so you can be certain that an <c>internal</c> event is an event
      from your state machine to itself.
    </p>
    <p>
      One example of using self generated events may be when you have
      a state machine specification that uses state entry actions.
      That you could code using a dedicated function
      to do the state transition.  But if you want that code to be
      visible besides the other state logic you can insert
      an <c>internal</c> event that does the entry actions.
      This has the same unfortunate consequence as using
      state transition functions that everywhere you go to
      the state in question you will have to explicitly
      insert the <c>internal</c> event
      or use state transition function.
    </p>
    <p>
      Here is an implementation of entry actions
      using <c>internal</c> events with content <c>enter</c>
      utilizing a helper function <c>enter/3</c> for state entry:
    </p>
    <code type="erl"><![CDATA[
...
-define(CALLBACK_MODE, state_functions).

...

init(Code) ->
    process_flag(trap_exit, true),
    Data = #{code => Code},
    enter(?CALLBACK_MODE, locked, Data).

...

locked(internal, enter, _Data) ->
    do_lock(),
    {keep_state,Data#{remaining => Code}};
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] ->
            enter(next_state, open, Data);
...

open(internal, enter, _Data) ->
    Tref = erlang:start_timer(10000, self(), lock),
    do_unlock(),
    {keep_state,Data#{timer => Tref}};
open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) ->
    enter(next_state, locked, Data);
...

enter(Tag, State, Data) ->
    {Tag,State,Data,[{next_event,internal,enter}]}.
    ]]></code>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Example Revisited</title>
    <p>
      Here is the example after all mentioned modifications
      and some more utilizing the entry actions,
      which deserves a new state diagram:
    </p>
    <image file="../design_principles/code_lock_2.png">
      <icaption>Code lock state diagram revisited</icaption>
    </image>
    <p>
      Note that this state diagram does not specify how to handle
      a button event in the state <c>open</c>, so you will have to
      read some other place that is here that unspecified events
      shall be ignored as in not consumed but handled in some other state.
      Nor does it show that the <c>code_length/0</c> call shall be
      handled in every state.
    </p>

    <section>
      <title>Callback Mode: state_functions</title>
      <p>
	Using state functions:
      </p>
      <code type="erl"><![CDATA[
-module(code_lock).
-behaviour(gen_statem).
-define(NAME, code_lock_2).
-define(CALLBACK_MODE, state_functions).

-export([start_link/1,stop/0]).
-export([button/1,code_length/0]).
-export([init/1,terminate/3,code_change/4]).
-export([locked/3,open/3]).

start_link(Code) ->
    gen_statem:start_link({local,?NAME}, ?MODULE, Code, []).
stop() ->
    gen_statem:stop(?NAME).

button(Digit) ->
    gen_statem:cast(?NAME, {button,Digit}).
code_length() ->
    gen_statem:call(?NAME, code_length).

init(Code) ->
    process_flag(trap_exit, true),
    Data = #{code => Code},
    enter(?CALLBACK_MODE, locked, Data).

locked(internal, enter, #{code := Code} = Data) ->
    do_lock(),
    {keep_state,Data#{remaining => Code}};
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] -> % Complete
            enter(next_state, open, Data);
        [Digit|Rest] -> % Incomplete
            {keep_state,Data#{remaining := Rest}};
        [_|_] -> % Wrong
            {keep_state,Data#{remaining := Code}}
    end;
locked(EventType, EventContent, Data) ->
    handle_event(EventType, EventContent, Data).

open(internal, enter, Data) ->
    Tref = erlang:start_timer(10000, self(), lock),
    do_unlock(),
    {keep_state,Data#{timer => Tref}};
open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) ->
    enter(next_state, locked, Data);
open(cast, {button,_}, _) ->
    {keep_state_and_data,[postpone]};
open(EventType, EventContent, Data) ->
    handle_event(EventType, EventContent, Data).

handle_event({call,From}, code_length, #{code := Code}) ->
    {keep_state_and_data,[{reply,From,length(Code)}]}.
enter(Tag, State, Data) ->
    {Tag,State,Data,[{next_event,internal,enter}]}.

do_lock() ->
    io:format("Locked~n", []).
do_unlock() ->
    io:format("Open~n", []).

terminate(_Reason, State, _Data) ->
    State =/= locked andalso do_lock(),
    ok.
code_change(_Vsn, State, Data, _Extra) ->
    {?CALLBACK_MODE,State,Data}.
      ]]></code>
    </section>

    <section>
      <title>Callback Mode: handle_event_function</title>
      <p>
	What to change to use one <c>handle_event/4</c> function.
	Here a clean first-dispatch-on-event approach
	does not work that well due to the generated
	entry actions:
      </p>
      <code type="erl"><![CDATA[
...
-define(CALLBACK_MODE, handle_event_function).

...
-export([handle_event/4]).

...

%% State: locked
handle_event(internal, enter, locked, #{code := Code} = Data) ->
    do_lock(),
    {keep_state,Data#{remaining => Code}};
handle_event(
  cast, {button,Digit}, locked,
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] -> % Complete
            enter(next_state, open, Data);
        [Digit|Rest] -> % Incomplete
            {keep_state,Data#{remaining := Rest}};
        [_|_] -> % Wrong
            {keep_state,Data#{remaining := Code}}
    end;
%%
%% State: open
handle_event(internal, enter, open, Data) ->
    Tref = erlang:start_timer(10000, self(), lock),
    do_unlock(),
    {keep_state,Data#{timer => Tref}};
handle_event(info, {timeout,Tref,lock}, open, #{timer := Tref} = Data) ->
    enter(next_state, locked, Data);
handle_event(cast, {button,_}, open, _) ->
    {keep_state_and_data,[postpone]};
%%
%% Any state
handle_event({call,From}, code_length, _State, #{code := Code}) ->
    {keep_state_and_data,[{reply,From,length(Code)}]}.

...
      ]]></code>
    </section>
    <p>
      Note that postponing buttons from the <c>locked</c> state
      to the <c>open</c> state feels like the wrong thing to do
      for a code lock, but it at least illustrates event postponing.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Filter the State</title>
    <p>
      The example servers so far in this chapter will for example
      when killed by an exit signal or due to an internal error
      print out the full internal state in the error log.
      This state contains both the code lock code
      and which digits that remains to unlock.
    </p>
    <p>
      This state data can be regarded as sensitive,
      and maybe not what you want in the error log
      because of something unpredictable happening.
    </p>
    <p>
      Another reason to filter the state can be
      that the state is too big to print out since it fills
      the error log with uninteresting details.
    </p>
    <p>
      To avoid this you can format the internal state
      that gets in the error log and gets returned from
      <seealso marker="stdlib:sys#get_status/1">
	<c>sys:get_status/1,2</c>
      </seealso>
      by implementing the
      <seealso marker="stdlib:gen_statem#Module:format_status/2">
	<c>Module:format_status/2</c>
      </seealso>
      function, for example like this:
    </p>
    <code type="erl"><![CDATA[
...
-export([init/1,terminate/3,code_change/4,format_status/2]).
...

format_status(Opt, [_PDict,State,Data]) ->
    StateData =
	{State,
	 maps:filter(
	   fun (code, _) -> false;
	       (remaining, _) -> false;
	       (_, _) -> true
	   end,
	   Data)},
    case Opt of
	terminate ->
	    StateData;
	normal ->
	    [{data,[{"State",StateData}]}]
    end.
    ]]></code>
    <p>
      It is not mandatory to implement a
      <seealso marker="stdlib:gen_statem#Module:format_status/2">
	<c>Module:format_status/2</c>
      </seealso>
      function.  If you do not a default implementation is used that
      does the same as this example function without filtering
      the <c>Data</c> term that is: <c>StateData = {State,Data}</c>.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Complex State</title>
    <p>
      The callback mode
      <seealso marker="stdlib:gen_statem#type-callback_mode">
	<c>handle_event_function</c>
      </seealso>
      enables using a non-atom state as described in
      <seealso marker="#callback_modes">
	Callback Modes,
      </seealso>
      for example a complex state term like a tuple.
    </p>
    <p>
      One reason to use this is when you have
      a state item that affects the event handling
      in particular when combining that with postponing events.
      Let us complicate the previous example
      by introducing a configurable lock button
      (this is the state item in question)
      that in the <c>open</c> state immediately locks the door,
      and an API function <c>set_lock_button/1</c> to set the lock button.
    </p>
    <p>
      Suppose now that we call <c>set_lock_button</c>
      while the door is open,
      and have already postponed a button event
      that up until now was not the lock button;
      the sensible thing might be to say that
      the button was pressed too early so it should
      not be recognized as the lock button,
      but then it might be surprising that a button event
      that now is the lock button event arrives (as retried postponed)
      immediately after the state transits to <c>locked</c>.
    </p>
    <p>
      So let us make the <c>button/1</c> function synchronous
      by using <c>gen_statem:call</c>,
      and still postpone its events in the <c>open</c> state.
      Then a call to <c>button/1</c> during the <c>open</c>
      state will not return until the state transits to <c>locked</c>
      since it is there the event is handled and the reply is sent.
    </p>
    <p>
      If now one process calls <c>set_lock_button/1</c>
      to change the lock button while some other process
      hangs in <c>button/1</c> with the new lock button
      it could be expected that the hanging lock button call
      immediately takes effect and locks the lock.
      Therefore we make the current lock button a part of the state
      so when we change the lock button the state will change
      and all postponed events will be retried.
    </p>
    <p>
      We define the state as <c>{StateName,LockButton}</c>
      where <c>StateName</c> is as before
      and <c>LockButton</c> is the current lock button:
    </p>
    <code type="erl"><![CDATA[
-module(code_lock).
-behaviour(gen_statem).
-define(NAME, code_lock_3).
-define(CALLBACK_MODE, handle_event_function).

-export([start_link/2,stop/0]).
-export([button/1,code_length/0,set_lock_button/1]).
-export([init/1,terminate/3,code_change/4,format_status/2]).
-export([handle_event/4]).

start_link(Code, LockButton) ->
    gen_statem:start_link(
        {local,?NAME}, ?MODULE, {Code,LockButton}, []).
stop() ->
    gen_statem:stop(?NAME).

button(Digit) ->
    gen_statem:call(?NAME, {button,Digit}).
code_length() ->
    gen_statem:call(?NAME, code_length).
set_lock_button(LockButton) ->
    gen_statem:call(?NAME, {set_lock_button,LockButton}).

init({Code,LockButton}) ->
    process_flag(trap_exit, true),
    Data = #{code => Code, remaining => undefined, timer => undefined},
    enter(?CALLBACK_MODE, {locked,LockButton}, Data, []).

handle_event(
  {call,From}, {set_lock_button,NewLockButton},
  {StateName,OldLockButton}, Data) ->
    {next_state,{StateName,NewLockButton},Data,
     [{reply,From,OldLockButton}]};
handle_event(
  {call,From}, code_length,
  {_StateName,_LockButton}, #{code := Code}) ->
    {keep_state_and_data,
     [{reply,From,length(Code)}]};
handle_event(
  EventType, EventContent,
  {locked,LockButton}, #{code := Code, remaining := Remaining} = Data) ->
    case {EventType,EventContent} of
	{internal,enter} ->
	    do_lock(),
	    {keep_state,Data#{remaining := Code}};
	{{call,From},{button,Digit}} ->
	    case Remaining of
		[Digit] -> % Complete
		    next_state(
		      {open,LockButton}, Data,
		      [{reply,From,ok}]);
		[Digit|Rest] -> % Incomplete
		    {keep_state,Data#{remaining := Rest},
		     [{reply,From,ok}]};
		[_|_] -> % Wrong
		    {keep_state,Data#{remaining := Code},
		     [{reply,From,ok}]}
	    end
    end;
handle_event(
  EventType, EventContent,
  {open,LockButton}, #{timer := Timer} = Data) ->
    case {EventType,EventContent} of
	{internal,enter} ->
	    Tref = erlang:start_timer(10000, self(), lock),
	    do_unlock(),
	    {keep_state,Data#{timer := Tref}};
	{info,{timeout,Timer,lock}} ->
	    next_state({locked,LockButton}, Data, []);
	{{call,From},{button,Digit}} ->
	    if
		Digit =:= LockButton ->
		    erlang:cancel_timer(Timer),
		    next_state(
		      {locked,LockButton}, Data,
		      [{reply,From,locked}]);
		true ->
		    {keep_state_and_data,
		     [postpone]}
	    end
    end.

next_state(State, Data, Actions) ->
    enter(next_state, State, Data, Actions).
enter(Tag, State, Data, Actions) ->
    {Tag,State,Data,[{next_event,internal,enter}|Actions]}.

do_lock() ->
    io:format("Locked~n", []).
do_unlock() ->
    io:format("Open~n", []).

terminate(_Reason, State, _Data) ->
    State =/= locked andalso do_lock(),
    ok.
code_change(_Vsn, State, Data, _Extra) ->
    {?CALLBACK_MODE,State,Data}.
format_status(Opt, [_PDict,State,Data]) ->
    StateData =
	{State,
	 maps:filter(
	   fun (code, _) -> false;
	       (remaining, _) -> false;
	       (_, _) -> true
	   end,
	   Data)},
    case Opt of
	terminate ->
	    StateData;
	normal ->
	    [{data,[{"State",StateData}]}]
    end.
    ]]></code>
    <p>
      It may be an ill-fitting model for a physical code lock
      that the <c>button/1</c> call might hang until the lock
      is locked.  But for an API in general it is really not
      that strange.
    </p>
  </section>

<!-- =================================================================== -->

  <section>
    <title>Hibernation</title>
    <p>
      If you have many servers in one node
      and they have some state(s) in their lifetime in which
      the servers can be expected to idle for a while,
      and the amount of heap memory all these servers need
      is a problem; then it is possible to minimize
      the memory footprint of a server by hibernating it through
      <seealso marker="stdlib:proc_lib#hibernate/3">
	<c>proc_lib:hibernate/3</c>.
      </seealso>
    </p>
    <note>
      <p>
	To hibernate a process is rather costly. See
	<seealso marker="erts:erlang#hibernate/3">
	  <c>erlang:hibernate/3</c>.
	</seealso>
	It is in general not something you want to do
	after every event.
      </p>
    </note>
    <p>
      We can in this example hibernate in the <c>{open,_}</c> state
      since what normally happens in that state is that
      the state timeout after a while
      triggers a transition to <c>{locked,_}</c>:
    </p>
    <code type="erl"><![CDATA[
...
handle_event(
  EventType, EventContent,
  {open,LockButton}, #{timer := Timer} = Data) ->
    case {EventType,EventContent} of
	{internal,enter} ->
	    Tref = erlang:start_timer(10000, self(), lock),
	    do_unlock(),
	    {keep_state,Data#{timer := Tref},[hibernate]};
...
    ]]></code>
    <p>
      The
      <seealso marker="stdlib:gen_statem#type-hibernate">
	<c>[hibernate]</c>
      </seealso>
      action list on the last line
      when entering the <c>{open,_}</c> state is the only change.
      If any event arrives in the <c>{open,_},</c> state we
      do not bother to re-hibernate, so the server stays
      awake after any event.
    </p>
    <p>
      To change that we would need to insert
      the <c>hibernate</c> action in more places,
      for example for the state independent <c>set_lock_button</c>
      and <c>code_length</c> operations that then would have to
      be aware of using <c>hibernate</c> while in the
      <c>{open,_}</c> state which would clutter the code.
    </p>
    <p>
      This server probably does not use an amount of
      heap memory worth hibernating for.
      To gain anything from hibernation your server would
      have to actually produce some garbage during callback execution,
      for which this example server may serve as a bad example.
    </p>
  </section>

</chapter>