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><year>2017</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 Behavior</title>
    <prepared></prepared>
    <docno></docno>
    <date></date>
    <rev></rev>
    <file>statem.xml</file>
  </header>
  <marker id="gen_statem Behaviour" />
  <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>
  <note>
    <p>
      This is a new behavior in Erlang/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.
      Depending on user feedback, we do not expect
      but can find it necessary to make minor
      not backward compatible changes into Erlang/OTP 20.0.
    </p>
  </note>

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

  <section>
    <marker id="Event-Driven State Machines" />
    <title>Event-Driven State Machines</title>
    <p>
      Established Automata Theory does not deal much with
      how a state transition is triggered,
      but 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 following form:
    </p>
    <pre>
State(S) x Event(E) -> Actions(A), State(S')</pre>
    <p>These relations are interpreted as follows:
      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>. Notice that <c>S'</c> can be equal to <c>S</c>.
    </p>
    <p>
      As <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> behaviors, <c>gen_statem</c> keeps
      a server <c>Data</c> besides the state. Because of this, and as
      there is no restriction on the number of states
      (assuming that there is enough virtual machine memory)
      or on the number of distinct input events,
      a state machine implemented with this behavior
      is in fact 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> behavior supports two callback modes:
    </p>
    <list type="bulleted">
      <item>
        <p>
          In mode
	  <seealso marker="stdlib:gen_statem#type-callback_mode"><c>state_functions</c></seealso>,
          the state transition rules are written as some Erlang
          functions, which conform to the following convention:
        </p>
        <pre>
StateName(EventType, EventContent, Data) ->
    ... code for actions here ...
    {next_state, NewStateName, NewData}.
	</pre>
	<p>
	  This form is used in most examples here for example in section
	  <seealso marker="#Example">Example</seealso>.
	</p>
      </item>
      <item>
        <p>
          In mode
	  <seealso marker="stdlib:gen_statem#type-callback_mode"><c>handle_event_function</c></seealso>,
          only one Erlang function provides all state transition rules:
        </p>
        <pre>
handle_event(EventType, EventContent, State, Data) ->
    ... code for actions here ...
    {next_state, NewState, NewData}
	</pre>
	<p>
	  See section
	  <seealso marker="#One Event Handler">One Event Handler</seealso>
	  for an example.
	</p>
      </item>
    </list>
    <p>
      Both these modes allow other return tuples; see
      <seealso marker="stdlib:gen_statem#Module:StateName/3"><c>Module:StateName/3</c></seealso>
      in the <c>gen_statem</c> manual page.
      These other return tuples can, for example, stop the machine,
      execute state transition actions on the machine engine itself,
      and send replies.
    </p>

    <section>
      <marker id="Choosing the Callback Mode" />
      <title>Choosing the Callback Mode</title>
      <p>
	The two
	<seealso marker="#Callback Modes">callback modes</seealso>
	give different possibilities
	and restrictions, but one goal remains:
	you want to handle all possible combinations of
	events and states.
      </p>
      <p>
	This can be done, for example, by focusing on one state at the time
	and for every state ensure that all events are handled.
	Alternatively, you can focus on one event at the time
	and ensure that it is handled in every state.
	You can also use a mix of these strategies.
      </p>
      <p>
	With <c>state_functions</c>, you are restricted to use
	atom-only states, and the <c>gen_statem</c> engine
	branches depending 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, which 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 all events and states are handled in the same callback function.
      </p>
      <p>
	This mode works equally well when you want to focus on
	one event at the time or on
	one state at the time, but function
	<seealso marker="stdlib:gen_statem#Module:handle_event/4"><c>Module:handle_event/4</c></seealso>
	quickly grows too large to handle without branching to
	helper functions.
      </p>
      <p>
	The mode enables the use of non-atom states, for example,
	complex states or even hierarchical states.
	If, for example, a state diagram is largely alike
	for the client side and the server side of a protocol,
	you can have a state <c>{StateName,server}</c> or
	<c>{StateName,client}</c>,
	and make <c>StateName</c> determine 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>
    <marker id="State Enter Calls" />
    <title>State Enter Calls</title>
    <p>
      The <c>gen_statem</c> behavior can regardless of callback mode
      automatically
      <seealso marker="stdlib:gen_statem#type-state_enter">
	call the state callback
      </seealso>
      with special arguments whenever the state changes
      so you can write state entry actions
      near the rest of the state transition rules.
      It typically looks like this:
    </p>
    <pre>
StateName(enter, _OldState, Data) ->
    ... code for state entry actions here ...
    {keep_state, NewData};
StateName(EventType, EventContent, Data) ->
    ... code for actions here ...
    {next_state, NewStateName, NewData}.</pre>
    <p>
      Depending on how your state machine is specified,
      this can be a very useful feature,
      but it forces you to handle the state enter calls in all states.
      See also the
      <seealso marker="#State Entry Actions">
	State Entry Actions
      </seealso>
      chapter.
    </p>
  </section>

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

  <section>
    <marker id="Actions" />
    <title>Actions</title>
    <p>
      In the first section
      <seealso marker="#Event-Driven State Machines">
	Event-Driven State Machines
      </seealso>
      actions were mentioned as a part of
      the general state machine model. These general actions
      are implemented with the code that callback module
      <c>gen_statem</c> 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_callback_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 and can do the following:
    </p>
    <list type="bulleted">
      <item>
	<seealso marker="stdlib:gen_statem#type-postpone">
	  Postpone
	</seealso>
	the current event, see section
	<seealso marker="#Postponing Events">Postponing Events</seealso>
      </item>
      <item>
	<seealso marker="stdlib:gen_statem#type-hibernate">
	  Hibernate
	</seealso>
	the <c>gen_statem</c>, treated in
	<seealso marker="#Hibernation">Hibernation</seealso>
      </item>
      <item>
	Start a
	<seealso marker="stdlib:gen_statem#type-state_timeout">
	  state time-out</seealso>,
	  read more in section
	<seealso marker="#State Time-Outs">State Time-Outs</seealso>
      </item>
      <item>
	Start a
	<seealso marker="stdlib:gen_statem#type-generic_timeout">
	  generic time-out</seealso>,
	  read more in section
	<seealso marker="#Generic Time-Outs">Generic Time-Outs</seealso>
      </item>
      <item>
	Start an
	<seealso marker="stdlib:gen_statem#type-event_timeout">event time-out</seealso>,
	see more in section
	<seealso marker="#Event Time-Outs">Event Time-Outs</seealso>
      </item>
      <item>
	<seealso marker="stdlib:gen_statem#type-reply_action">
	  Reply
	</seealso>
	to a caller, mentioned at the end of section
	<seealso marker="#All State Events">All State Events</seealso>
      </item>
      <item>
	Generate the
	<seealso marker="stdlib:gen_statem#type-action">
	  next event
	</seealso>
	to handle, see section 
	<seealso marker="#Self-Generated Events">Self-Generated Events</seealso>
      </item>
    </list>
    <p>
      For details, see the
      <seealso marker="stdlib:gen_statem#type-action">
	<c>gen_statem(3)</c>
      </seealso>
      manual page.
      You can, for example, reply to many callers,
      generate multiple next events,
      and set time-outs to relative or absolute times.
    </p>
  </section>

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

  <section>
    <marker id="Event Types" />
    <title>Event Types</title>
    <p>
      Events are categorized in different
      <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>
      The following is a 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>state_timeout</c></tag>
      <item>
	Generated by state transition action
	<seealso marker="stdlib:gen_statem#type-state_timeout">
	  <c>{state_timeout,Time,EventContent}</c>
	</seealso>
	state timer timing out.
      </item>
      <tag><c>{timeout,Name}</c></tag>
      <item>
	Generated by state transition action
	<seealso marker="stdlib:gen_statem#type-generic_timeout">
	  <c>{{timeout,Name},Time,EventContent}</c>
	</seealso>
	generic timer timing out.
      </item>
      <tag><c>timeout</c></tag>
      <item>
	Generated by state transition action
	<seealso marker="stdlib:gen_statem#type-event_timeout">
	  <c>{timeout,Time,EventContent}</c>
	</seealso>
	(or its short form <c>Time</c>)
	event timer timing out.
      </item>
      <tag><c>internal</c></tag>
      <item>
	Generated by state transition
	<seealso marker="stdlib:gen_statem#type-action">action</seealso>
	<c>{next_event,internal,EventContent}</c>.
	All event types above can also be generated using
	<c>{next_event,EventType,EventContent}</c>.
      </item>
    </taglist>
  </section>

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

  <section>
    <marker id="Example" />
    <title>Example</title>
    <p>
      A door with a code lock can be seen as a state machine.
      Initially, the door is locked. When someone presses a button,
      an event is generated.
      Depending on what buttons have been pressed before,
      the sequence so far can be correct, incomplete, or wrong.
      If correct, the door is unlocked for 10 seconds (10,000 milliseconds).
      If incomplete, we wait for another button to be pressed. If
      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>
      This code lock state machine can be implemented using
      <c>gen_statem</c> with the following callback module:
    </p>
    <code type="erl"><![CDATA[
-module(code_lock).
-behaviour(gen_statem).
-define(NAME, code_lock).

-export([start_link/1]).
-export([button/1]).
-export([init/1,callback_mode/0,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},
    {ok, locked, Data}.

callback_mode() ->
    state_functions.

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

open(state_timeout, lock,  Data) ->
    do_lock(),
    {next_state, locked, Data};
open(cast, {button,_}, Data) ->
    {next_state, open, 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) ->
    {ok, State, Data}.
    ]]></code>
    <p>The code is explained in the next sections.</p>
  </section>

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

  <section>
    <marker id="Starting gen_statem" />
    <title>Starting gen_statem</title>
    <p>
      In the example in the previous section, <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 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 specified
          as <c>{global,Name}</c>, then the <c>gen_statem</c> is
          registered using
          <seealso marker="kernel:global#register_name/2"><c>global:register_name/2</c></seealso>
          in Kernel.
        </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
          code and the server-side code contained in one module.
        </p>
      </item>
      <item>
        <p>
          The third argument, <c>Code</c>, is a list of digits, which
          is the correct unlock code that is passed
          to callback function <c>init/1</c>.
	</p>
      </item>
      <item>
        <p>
          The fourth argument, <c>[]</c>, is a list of options.
          For the available options, see
          <seealso marker="stdlib:gen_statem#start_link/3"><c>gen_statem:start_link/3</c></seealso>.
	</p>
      </item>
    </list>
    <p>
      If name registration succeeds, the new <c>gen_statem</c> process
      calls callback function <c>code_lock:init(Code)</c>.
      This function is expected to return <c>{ok, State, Data}</c>,
      where <c>State</c> is the initial state of the <c>gen_statem</c>,
      in this case <c>locked</c>; assuming that 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 key <c>code</c> that stores
      the correct button sequence, and 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},
    {ok,locked,Data}.
    ]]></code>
    <p>Function
      <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>
      is initialized and is ready to receive events.
    </p>
    <p>
      Function
      <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.
      Another function,
      <seealso marker="stdlib:gen_statem#start/3"><c>gen_statem:start</c></seealso>
      can be used to start a standalone <c>gen_statem</c>, that is,
      a <c>gen_statem</c> that is not part of a supervision tree.
    </p>

    <code type="erl"><![CDATA[
callback_mode() ->
    state_functions.
    ]]></code>
    <p>
      Function
      <seealso marker="stdlib:gen_statem#Module:callback_mode/0"><c>Module:callback_mode/0</c></seealso>
      selects the
      <seealso marker="#Callback Modes"><c>CallbackMode</c></seealso>
      for the callback module, in this case
      <seealso marker="stdlib:gen_statem#type-callback_mode"><c>state_functions</c></seealso>.
      That is, each state has got its own handler function.
    </p>

  </section>

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

  <section>
    <marker id="Handling Events" />
    <title>Handling Events</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, we use the
      same macro <c>?NAME</c> as when starting.
      <c>{button,Digit}</c> is the 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>,
      or <c>{next_state, NewStateName, NewData, Actions}</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>, and <c>Actions</c> is a list of
      actions on the <c>gen_statem</c> engine.
    </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},
             [{state_timeout,10000,lock}]};
        [Digit|Rest] -> % Incomplete
            {next_state, locked, Data#{remaining := Rest}};
        [_|_] -> % Wrong
            {next_state, locked, Data#{remaining := Code}}
    end.

open(state_timeout, lock, Data) ->
    do_lock(),
    {next_state, locked, Data};
open(cast, {button,_}, Data) ->
    {next_state, open, Data}.
    ]]></code>
    <p>
      If the door is locked and a button is pressed, the pressed
      button is compared with the next correct button.
      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>
      If the whole code is correct, the server changes states
      to <c>open</c>.
    </p>
    <p>
      In state <c>open</c>, a button event is ignored
      by staying in the same state.  This can also be done
      by returning <c>{keep_state, Data}</c> or in this case
      since <c>Data</c> unchanged even by returning
      <c>keep_state_and_data</c>.
    </p>
  </section>

  <section>
    <marker id="State Time-Outs" />
    <title>State 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},
 [{state_timeout,10000,lock}]};
    ]]></code>
    <p>
      10,000 is a time-out value in milliseconds.
      After this time (10 seconds), a time-out occurs.
      Then, <c>StateName(state_timeout, lock, 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(state_timeout, lock,  Data) ->
    do_lock(),
    {next_state, locked, Data};
    ]]></code>
    <p>
      The timer for a state time-out is automatically cancelled
      when the state machine changes states.  You can restart
      a state time-out by setting it to a new time, which cancels
      the running timer and starts a new.  This implies that
      you can cancel a state time-out by restarting it with
      time <c>infinity</c>.
    </p>
  </section>

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

  <section>
    <marker id="All State Events" />
    <title>All State Events</title>
    <p>
      Sometimes events 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>
      Consider a <c>code_length/0</c> function that returns
      the length of the correct code
      (that should not be sensitive to reveal).
      We 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.  This return form is convenient
      when you want to stay in the current state but do not know or
      care about what it is.
    </p>
  </section>

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

  <section>
    <marker id="One Event Handler" />
    <title>One Event Handler</title>
    <p>
      If mode <c>handle_event_function</c> is used,
      all events are handled in
      <seealso marker="stdlib:gen_statem#Module:handle_event/4"><c>Module:handle_event/4</c></seealso>
      and we can (but do not have to) use an event-centered approach
      where we first branch depending on event
      and then depending on state:
    </p>
    <code type="erl"><![CDATA[
...
-export([handle_event/4]).

...
callback_mode() ->
    handle_event_function.

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},
                     [{state_timeout,10000,lock}]};
		[Digit|Rest] -> % Incomplete
		    {keep_state, Data#{remaining := Rest}};
		[_|_] -> % Wrong
		    {keep_state, Data#{remaining := Code}}
	    end;
	open ->
            keep_state_and_data
    end;
handle_event(state_timeout, lock, open, Data) ->
    do_lock(),
    {next_state, locked, Data}.

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

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

  <section>
    <marker id="Stopping" />
    <title>Stopping</title>

    <section>
      <marker id="In a Supervision Tree" />
      <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 function <c>init/1</c> set itself to trap exit signals
	by calling
	<seealso marker="erts:erlang#process_flag/2"><c>process_flag(trap_exit, true)</c></seealso>:
      </p>
      <code type="erl"><![CDATA[
init(Args) ->
    process_flag(trap_exit, true),
    do_lock(),
    ...
      ]]></code>
      <p>
	When ordered to shut down, the <c>gen_statem</c> then calls
	callback function <c>terminate(shutdown, State, Data)</c>.
      </p>
      <p>
	In this example, function <c>terminate/3</c>
	locks 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>
      <marker id="Standalone gen_statem" />
      <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 callback function
	<c>terminate/3</c> just like for a supervised server
	and waits for the process to terminate.
      </p>
    </section>
  </section>

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

  <section>
    <marker id="Event Time-Outs" />
    <title>Event Time-Outs</title>
    <p>
      A time-out feature inherited from <c>gen_statem</c>'s predecessor
      <seealso marker="stdlib:gen_fsm"><c>gen_fsm</c></seealso>,
      is an event time-out, that is,
      if an event arrives the timer is cancelled.
      You get either an event or a time-out, but not both.
    </p>
    <p>
      It is ordered by the state transition action
      <c>{timeout,Time,EventContent}</c>, or just <c>Time</c>,
      or even just <c>Time</c> instead of an action list
      (the latter is a form inherited from <c>gen_fsm</c>.
    </p>
    <p>
      This type of time-out is useful to for example act on inactivity.
      Let us restart the code sequence
      if no button is pressed for say 30 seconds:
    </p>
    <code type="erl"><![CDATA[
...

locked(
  timeout, _, 
  #{code := Code, remaining := Remaining} = Data) ->
    {next_state, locked, Data#{remaining := Code}};
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
...
        [Digit|Rest] -> % Incomplete
            {next_state, locked, Data#{remaining := Rest}, 30000};
...
     ]]></code>
    <p>
      Whenever we receive a button event we start an event time-out
      of 30 seconds, and if we get an event type <c>timeout</c>
      we reset the remaining code sequence.
    </p>
    <p>
      An event time-out is cancelled by any other event so you either
      get some other event or the time-out event.  It is therefore
      not possible nor needed to cancel or restart an event time-out.
      Whatever event you act on has already cancelled
      the event time-out...
    </p>
  </section>

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

  <section>
    <marker id="Generic Time-Outs" />
    <title>Generic Time-Outs</title>
    <p>
      The previous example of state time-outs only work if
      the state machine stays in the same state during the
      time-out time.  And event time-outs only work if no
      disturbing unrelated events occur.
    </p>
    <p>
      You may want to start a timer in one state and respond
      to the time-out in another, maybe cancel the time-out
      without changing states, or perhaps run multiple
      time-outs in parallel. All this can be accomplished with
      <seealso marker="stdlib:gen_statem#type-generic_timeout">generic time-outs</seealso>.
      They may look a little bit like
      <seealso marker="stdlib:gen_statem#type-event_timeout">event time-outs</seealso>
      but contain a name to allow for any number of them simultaneously
      and they are not automatically cancelled.
    </p>
    <p>
      Here is how to accomplish the state time-out
      in the previous example by instead using a generic time-out
      named <c>open_tm</c>:
    </p>
    <code type="erl"><![CDATA[
...
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] ->
	    do_unlock(),
            {next_state, open, Data#{remaining := Code},
	     [{{timeout,open_tm},10000,lock}]};
...

open({timeout,open_tm}, lock, Data) ->
    do_lock(),
    {next_state,locked,Data};
open(cast, {button,_}, Data) ->
    {keep_state,Data};
...
    ]]></code>
    <p>
      Just as
      <seealso marker="#State Time-Outs">state time-outs</seealso>
      you can restart or cancel a specific generic time-out
      by setting it to a new time or <c>infinity</c>.
    </p>
    <p>
      Another way to handle a late time-out can be to not cancel it,
      but to ignore it if it arrives in a state
      where it is known to be late.
    </p>
  </section>

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

  <section>
    <marker id="Erlang Timers" />
    <title>Erlang Timers</title>
    <p>
      The most versatile way to handle time-outs is to use
      Erlang Timers; see
      <seealso marker="erts:erlang#start_timer/4"><c>erlang:start_timer3,4</c></seealso>.
      Most time-out tasks can be performed with the
      time-out features in <c>gen_statem</c>,
      but an example of one that can not is if you should need
      the return value from
      <seealso marker="erts:erlang#cancel_timer/2"><c>erlang:cancel_timer(Tref)</c></seealso>, that is; the remaining time of the timer.
    </p>
    <p>
      Here is how to accomplish the state time-out
      in the previous example by instead using an Erlang Timer:
    </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,maps:remove(timer, Data)};
open(cast, {button,_}, Data) ->
    {keep_state,Data};
...
    ]]></code>
    <p>
      Removing the <c>timer</c> key from the map when we
      change to state <c>locked</c> is not strictly
      necessary since we can only get into state <c>open</c>
      with an updated <c>timer</c> map value.  But it can be nice
      to not have outdated values in the state <c>Data</c>!
    </p>
    <p>
      If you need to cancel a timer because of some other event, you can use
      <seealso marker="erts:erlang#cancel_timer/2"><c>erlang:cancel_timer(Tref)</c></seealso>.
      Note that a time-out message cannot arrive after this,
      unless you have postponed it before (see the next section),
      so ensure that you do not accidentally postpone such messages.
      Also note that a time-out message may have arrived
      just before you cancelling it, so you may have to read out
      such a message from the process mailbox depending on
      the return value from 
      <seealso marker="erts:erlang#cancel_timer/2"><c>erlang:cancel_timer(Tref)</c></seealso>.
    </p>
    <p>
      Another way to handle a late time-out can be to not cancel it,
      but to ignore it if it arrives in a state
      where it is known to be late.
    </p>
  </section>

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

  <section>
    <marker id="Postponing Events" />
    <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, that is, <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 are 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>
      Since a postponed event is only retried after a state change,
      you have to think about where to keep a state data item.
      You can keep it in the server <c>Data</c>
      or in the <c>State</c> itself,
      for example by having two more or less identical states
      to keep a boolean value, or by using a complex state with
      <seealso marker="#Callback Modes">callback mode</seealso>
      <seealso marker="stdlib:gen_statem#type-callback_mode"><c>handle_event_function</c></seealso>.
      If a change in the value changes the set of events that is handled,
      then the value should be kept in the State.
      Otherwise no postponed events will be retried
      since only the server Data changes.
    </p>
    <p>
      This is not important if you do not postpone events.
      But if you later decide to start postponing some events,
      then the design flaw of not having separate states
      when they should be, might become a hard to find bug.
    </p>

    <section>
      <marker id="Fuzzy State Diagrams" />
      <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: 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>
      <marker id="Selective Receive" />
      <title>Selective Receive</title>
      <p>
        Erlang's selective receive statement is often used to
        describe simple state machine examples in straightforward
        Erlang code. The following 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 implicitly <c>open</c>
      to postpone any events to the <c>locked</c> state.
    </p>
    <p>
      A selective receive cannot be used from a <c>gen_statem</c>
      behavior as for any <c>gen_*</c> behavior,
      as the receive statement is within the <c>gen_*</c> engine itself.
      It must be there because all
      <seealso marker="stdlib:sys"><c>sys</c></seealso>
      compatible behaviors 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 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>
      Both mechanisms have the same theoretical
      time and memory complexity, while the selective receive
      language construct has smaller constant factors.
    </p>
    </section>
  </section>

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

  <section>
    <marker id="State Entry Actions" />
    <title>State Entry Actions</title>
    <p>
      Say you have a state machine specification
      that uses state entry actions.
      Allthough you can code this using self-generated events
      (described in the next section), especially if just
      one or a few states has got state entry actions,
      this is a perfect use case for the built in
      <seealso marker="#State Enter Calls">state enter calls</seealso>.
    </p>
    <p>
      You return a list containing <c>state_enter</c> from your
      <seealso marker="stdlib:gen_statem#Module:callback_mode/0"><c>callback_mode/0</c></seealso>
      function and the <c>gen_statem</c> engine will call your
      state callback once with the arguments
      <c>(enter, OldState, ...)</c> whenever the state changes.
      Then you just need to handle these event-like calls in all states.
    </p>
    <code type="erl"><![CDATA[
...
init(Code) ->
    process_flag(trap_exit, true),
    Data = #{code => Code},
    {ok, locked, Data}.

callback_mode() ->
    [state_functions,state_enter].

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

open(enter, _OldState, _Data) ->
    do_unlock(),
    {keep_state_and_data, [{state_timeout,10000,lock}]};
open(state_timeout, lock, Data) ->
    {next_state, locked, Data};
...
    ]]></code>
    <p>
      You can repeat the state entry code by returning one of
      <c>{repeat_state, ...}</c>, <c>{repeat_state_and_data,_}</c>
      or <c>repeat_state_and_data</c> that otherwise behaves
      exactly like their <c>keep_state</c> siblings.
      See the type
      <seealso marker="stdlib:gen_statem#type-state_callback_result">
	<c>state_callback_result()</c>
      </seealso>
      in the reference manual.
    </p>
  </section>

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

  <section>
    <marker id="Self-Generated Events" />
    <title>Self-Generated Events</title>
    <p>
      It can sometimes be beneficial 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 action
      <c>next_event</c>. Hence, it cannot 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 for this is to pre-process incoming data, for example
      decrypting chunks or collecting characters up to a line break.
      Purists may argue that this should be modelled with a separate
      state machine that sends pre-processed events
      to the main state machine.
      But to decrease overhead the small pre-processing state machine
      can be implemented in the common state event handling
      of the main state machine using a few state data variables
      that then sends the pre-processed events as internal events
      to the main state machine.
    </p>
    <p>
      The following example uses an input model where you give the lock
      characters with <c>put_chars(Chars)</c> and then call
      <c>enter()</c> to finish the input.
    </p>
    <code type="erl"><![CDATA[
...
-export(put_chars/1, enter/0).
...
put_chars(Chars) when is_binary(Chars) ->
    gen_statem:call(?NAME, {chars,Chars}).

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

...

locked(enter, _OldState, Data) ->
    do_lock(),
    {keep_state,Data#{remaining => Code, buf => []}};
...

handle_event({call,From}, {chars,Chars}, #{buf := Buf} = Data) ->
    {keep_state, Data#{buf := [Chars|Buf],
     [{reply,From,ok}]};
handle_event({call,From}, enter, #{buf := Buf} = Data) ->
    Chars = unicode:characters_to_binary(lists:reverse(Buf)),
    try binary_to_integer(Chars) of
        Digit ->
            {keep_state, Data#{buf := []},
             [{reply,From,ok},
              {next_event,internal,{button,Chars}}]}
    catch
        error:badarg ->
            {keep_state, Data#{buf := []},
             [{reply,From,{error,not_an_integer}}]}
    end;
...
    ]]></code>
    <p>
      If you start this program with <c>code_lock:start([17])</c>
      you can unlock with <c>code_lock:put_chars(&lt;&lt;"001">>),
      code_lock:put_chars(&lt;&lt;"7">>), code_lock:enter()</c>.
    </p>
  </section>

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

  <section>
    <marker id="Example Revisited" />
    <title>Example Revisited</title>
    <p>
      This section includes the example after most of the mentioned
      modifications and some more using state enter calls,
      which deserves a new state diagram:
    </p>
    <image file="../design_principles/code_lock_2.png">
      <icaption>Code Lock State Diagram Revisited</icaption>
    </image>
    <p>
      Notice that this state diagram does not specify how to handle
      a button event in the state <c>open</c>. So, you need to
      read somewhere else that unspecified events
      must be ignored as in not consumed but handled in some other state.
      Also, the state diagram does not show that the <c>code_length/0</c>
      call must be handled in every state.
    </p>

    <section>
      <marker id="Callback Mode: state_functions" />
      <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).

-export([start_link/1,stop/0]).
-export([button/1,code_length/0]).
-export([init/1,callback_mode/0,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},
    {ok, locked, Data}.

callback_mode() ->
    [state_functions,state_enter].

locked(enter, _OldState, #{code := Code} = Data) ->
    do_lock(),
    {keep_state, Data#{remaining => Code}};
locked(
  timeout, _, 
  #{code := Code, remaining := Remaining} = Data) ->
    {keep_state, Data#{remaining := Code}};
locked(
  cast, {button,Digit},
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] -> % Complete
            {next_state, open, Data};
        [Digit|Rest] -> % Incomplete
            {keep_state, Data#{remaining := Rest}, 30000};
        [_|_] -> % Wrong
            {keep_state, Data#{remaining := Code}}
    end;
locked(EventType, EventContent, Data) ->
    handle_event(EventType, EventContent, Data).

open(enter, _OldState, _Data) ->
    do_unlock(),
    {keep_state_and_data, [{state_timeout,10000,lock}]};
open(state_timeout, lock, Data) ->
    {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)}]}.

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) ->
    {ok,State,Data}.
      ]]></code>
    </section>

    <section>
      <marker id="Callback Mode: handle_event_function" />
      <title>Callback Mode: handle_event_function</title>
      <p>
        This section describes what to change in the example
        to use one <c>handle_event/4</c> function.
        The previously used approach to first branch depending on event
	does not work that well here because of the state enter calls,
        so this example first branches depending on state:
      </p>
      <code type="erl"><![CDATA[
...
-export([handle_event/4]).

...
callback_mode() ->
    [handle_event_function,state_enter].

%% State: locked
handle_event(
  enter, _OldState, locked,
  #{code := Code} = Data) ->
    do_lock(),
    {keep_state, Data#{remaining => Code}};
handle_event(
  timeout, _, locked,
  #{code := Code, remaining := Remaining} = Data) ->
    {keep_state, Data#{remaining := Code}};
handle_event(
  cast, {button,Digit}, locked,
  #{code := Code, remaining := Remaining} = Data) ->
    case Remaining of
        [Digit] -> % Complete
            {next_state, open, Data};
        [Digit|Rest] -> % Incomplete
            {keep_state, Data#{remaining := Rest}, 30000};
        [_|_] -> % Wrong
            {keep_state, Data#{remaining := Code}}
    end;
%%
%% State: open
handle_event(enter, _OldState, open, _Data) ->
    do_unlock(),
    {keep_state_and_data, [{state_timeout,10000,lock}]};
handle_event(state_timeout, lock, open, Data) ->
    {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>
      Notice that postponing buttons from the <c>locked</c> state
      to the <c>open</c> state feels like a strange thing to do
      for a code lock, but it at least illustrates event postponing.
    </p>
  </section>

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

  <section>
    <marker id="Filter the State" />
    <title>Filter the State</title>
    <p>
      The example servers so far in this chapter
      print the full internal state in the error log, for example,
      when killed by an exit signal or because of an internal error.
      This state contains both the code lock code
      and which digits that remain to unlock.
    </p>
    <p>
      This state data can be regarded as sensitive,
      and maybe not what you want in the error log
      because of some unpredictable event.
    </p>
    <p>
      Another reason to filter the state can be
      that the state is too large to print, as 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 function
      <seealso marker="stdlib:gen_statem#Module:format_status/2"><c>Module:format_status/2</c></seealso>,
      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>,
      in this example containing sensitive information.
    </p>
  </section>

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

  <section>
    <marker id="Complex State" />
    <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 section
      <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 when changed should cancel the
      <seealso marker="#State Time-Outs">state time-out</seealso>,
      or one that affects the event handling
      in combination with postponing events.
      We will complicate the previous example
      by introducing a configurable lock button
      (this is the state item in question),
      which 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 until now was not the lock button.
      The sensible thing can be to say that
      the button was pressed too early so it is
      not to be recognized as the lock button.
      However, then it can 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 we make the <c>button/1</c> function synchronous
      by using
      <seealso marker="stdlib:gen_statem#call/2"><c>gen_statem:call</c></seealso>
      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 does not return until the state transits to <c>locked</c>,
      as it is there the event is handled and the reply is sent.
    </p>
    <p>
      If a process now calls <c>set_lock_button/1</c>
      to change the lock button while another process
      hangs in <c>button/1</c> with the new lock button,
      it can 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 that when we change the lock button, the state changes
      and all postponed events are 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).

-export([start_link/2,stop/0]).
-export([button/1,code_length/0,set_lock_button/1]).
-export([init/1,callback_mode/0,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},
    {ok, {locked,LockButton}, Data}.

callback_mode() ->
    [handle_event_function,state_enter].

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)}]};
%%
%% State: locked
handle_event(
  EventType, EventContent,
  {locked,LockButton}, #{code := Code, remaining := Remaining} = Data) ->
    case {EventType, EventContent} of
	{enter, _OldState} ->
	    do_lock(),
	    {keep_state, Data#{remaining := Code}};
        {timeout, _} ->
            {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}, 30000]};
		[_|_] -> % Wrong
		    {keep_state, Data#{remaining := Code},
		     [{reply,From,ok}]}
	    end
    end;
%%
%% State: open
handle_event(
  EventType, EventContent,
  {open,LockButton}, Data) ->
    case {EventType, EventContent} of
	{enter, _OldState} ->
	    do_unlock(),
	    {keep_state_and_data, [{state_timeout,10000,lock}]};
	{state_timeout, lock} ->
	    {next_state, {locked,LockButton}, Data};
	{{call,From}, {button,Digit}} ->
	    if
		Digit =:= LockButton ->
		    {next_state, {locked,LockButton}, Data,
		     [{reply,From,locked}]};
		true ->
		    {keep_state_and_data,
		     [postpone]}
	    end
    end.

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) ->
    {ok,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 can be an ill-fitting model for a physical code lock
      that the <c>button/1</c> call can hang until the lock
      is locked. But for an API in general it is not that strange.
    </p>
  </section>

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

  <section>
    <marker id="Hibernation" />
    <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 the memory footprint of a server
      can be mimimized by hibernating it through
      <seealso marker="stdlib:proc_lib#hibernate/3"><c>proc_lib:hibernate/3</c></seealso>.
    </p>
    <note>
      <p>
        It is rather costly to hibernate a process; see
        <seealso marker="erts:erlang#hibernate/3"><c>erlang:hibernate/3</c></seealso>.
        It is not something you want to do after every event.
      </p>
    </note>
    <p>
      We can in this example hibernate in the <c>{open,_}</c> state,
      because what normally occurs in that state is that
      the state time-out after a while
      triggers a transition to <c>{locked,_}</c>:
    </p>
    <code type="erl"><![CDATA[
...
%% State: open
handle_event(
  EventType, EventContent,
  {open,LockButton}, Data) ->
    case {EventType, EventContent} of
        {enter, _OldState} ->
            do_unlock(),
            {keep_state_and_data,
             [{state_timeout,10000,lock},hibernate]};
...
    ]]></code>
    <p>
      The atom
      <seealso marker="stdlib:gen_statem#type-hibernate"><c>hibernate</c></seealso>
      in the 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 rehibernate, so the server stays
      awake after any event.
    </p>
    <p>
      To change that we would need to insert
      action <c>hibernate</c> 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>
      Another not uncommon scenario is to use the event time-out
      to triger hibernation after a certain time of inactivity.
    </p>
    <p>
      This server probably does not use
      heap memory worth hibernating for.
      To gain anything from hibernation, your server would
      have to produce some garbage during callback execution,
      for which this example server can serve as a bad example.
    </p>
  </section>

</chapter>