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 Behavior</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>
  <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>
    <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>
      </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>
      </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>
      <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_entry_events" />
    <title>State Entry Events</title>
    <p>
      The <c>gen_statem</c> behavior can regardless of callback mode
      automatically generate an
      <seealso marker="stdlib:gen_statem#type-state_entry_mode">
	event whenever the state changes
      </seealso>
      so you can write state entry code
      near the rest of the state transition rules.
      It typically looks like this:
    </p>
    <pre>
StateName(enter, _OldState, Data) ->
    ... code for state entry 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 if you use it
      you will have to handle the state entry events in all states.
    </p>
  </section>

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

  <section>
    <title>Example</title>
    <p>
      This example starts off as equivalent to the example in section
      <seealso marker="fsm"><c>gen_fsm</c>-Behavior</seealso>.
      In later sections, additions and tweaks are made
      using features in <c>gen_statem</c> that <c>gen_fsm</c> does not have.
      The end of this chapter provides the example again
      with all the added features.
    </p>
    <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>
    <marker id="ex"></marker>
    <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},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) ->
    {ok,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, <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>
    <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>.
      <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.
      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, as
      any event cancels the event timer, so no
      time-out event occurs 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>
      10,000 is a time-out value in milliseconds.
      After this time (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 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.
    </p>
  </section>

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

  <section>
    <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},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 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>
      <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>
    <title>Actions</title>
    <p>
      In the first sections 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_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 and can do the following:
    </p>
    <list type="bulleted">
      <item>Postpone the current event</item>
      <item>Hibernate the <c>gen_statem</c></item>
      <item>Start an event time-out</item>
      <item>Reply to a caller</item>
      <item>Generate the next event to handle</item>
    </list>
    <p>
      In the example earlier was mentioned the event time-out
      and replying to a caller.
      An example of event postponing is included later in this chapter.
      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
      and generate multiple next events to handle.
    </p>
  </section>

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

  <section>
    <title>Event Types</title>
    <p>
      The previous sections 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>
      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>enter</c></tag>
      <item>
	Generated by a state transition with
	<c>OldState =/= NewState</c> when running with
	<seealso marker="#state_entry_events">state entry events</seealso>.
      </item>
      <tag><c>timeout</c></tag>
      <item>
	Generated by 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 state transition action
	<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>
    <title>State Time-Outs</title>
    <p>
      The time-out event generated by state transition action
      <c>{timeout,Time,EventContent}</c> 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>
      Often you want a timer not to be cancelled by any event
      or you want to start a timer in one state and respond
      to the time-out 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>
      For the example so far in this chapter: 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 time-out is cancelled and the button event is delivered.
      So, we choose to lock the door if this occurred.
    </p>
    <p>
      Suppose that 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 because of some other event, you can use
      <seealso marker="erts:erlang#cancel_timer/2"><c>erlang:cancel_timer(Tref)</c></seealso>.
      Notice that a time-out message cannot arrive after this,
      unless you have postponed it (see the next section) before,
      so ensure that you do not accidentally postpone such messages.
    </p>
    <p>
      Another way to cancel a timer is not to cancel it,
      but 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, 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>
      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>:
      if a change in the item value affects which events that
      are handled, then this item is to be part of the state.
    </p>
    <p>
      You want to avoid that you maybe much later decide
      to postpone an event in one state and by misfortune it is never retried,
      as 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: 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. 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>
    <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.
      This could be modelled with a separate state machine that sends
      the pre-processed events to the main state machine, or 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 and then send the pre-processed
      events as internal events to the main state machine.
    </p>
    <p>
      Another example of using self-generated events can be when you have
      a state machine specification that uses state entry actions.
      You can code that 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: everywhere you go to
      the state, you must explicitly
      insert the <c>internal</c> event
      or use a state transition function.
      This is something that can be forgotten, and if you find that
      annoying please look at the next chapter.
    </p>
    <p>
      The following is an implementation of entry actions
      using <c>internal</c> events with content <c>enter</c>
      using a helper function <c>enter/3</c> for state entry:
    </p>
    <code type="erl"><![CDATA[
...
init(Code) ->
    process_flag(trap_exit, true),
    Data = #{code => Code},
    enter(ok, locked, Data).

callback_mode() ->
    state_functions.

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>Using State Entry Events</title>
    <p>
      Here is the same example as the previous but instead using
      the built in
      <seealso marker="#state_entry_events">state entry events</seealso>.
    </p>
    <p>
      Since the state entry events are unconditionally inserted by
      the <c>gen_statem</c> engine you can not forget to insert them
      yourself and you will have to handle the state entry events
      in every state.
    </p>
    <p>
      If you want state entry code in just a few states the previous
      example may be more suitable, especially to only send internal
      events when entering just those few states.
      Note: additional discipline will be required.
    </p>
    <p>
      You can also in the previous example choose to generate
      events looking just like the events you get from using
      <seealso marker="#state_entry_events">state entry events</seealso>.
      This may be confusing, or practical,
      depending on your point of view.
    </p>
    <code type="erl"><![CDATA[
...
init(Code) ->
    process_flag(trap_exit, true),
    Data = #{code => Code},
    {ok, locked, Data}.

callback_mode() ->
    [state_functions,state_entry_events].

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) ->
    Tref = erlang:start_timer(10000, self(), lock),
    do_unlock(),
    {keep_state,Data#{timer => Tref}};
open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) ->
    {next_state, locked, Data};
...
    ]]></code>
  </section>

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

  <section>
    <title>Example Revisited</title>
    <p>
      This section includes the example after all mentioned modifications
      and some more using 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>
      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>
      <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_entry_events].

locked(enter, _OldState, #{code := Code} = Data) ->
    do_lock(),
    {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}};
        [_|_] -> % Wrong
            {keep_state,Data#{remaining := Code}}
    end;
locked(EventType, EventContent, Data) ->
    handle_event(EventType, EventContent, Data).

open(enter, _OldState, Data) ->
    Tref = erlang:start_timer(10000, self(), lock),
    do_unlock(),
    {keep_state,Data#{timer => Tref}};
open(info, {timeout,Tref,lock}, #{timer := Tref} = 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>
      <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 entry events,
        so this example first branches depending on state:
      </p>
      <code type="erl"><![CDATA[
...
-export([handle_event/4]).

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

%% State: locked
handle_event(enter, _OldState, 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
            {next_state, open, Data};
        [Digit|Rest] -> % Incomplete
            {keep_state,Data#{remaining := Rest}};
        [_|_] -> % Wrong
            {keep_state,Data#{remaining := Code}}
    end;
%%
%% State: open
handle_event(enter, _OldState, 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) ->
    {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 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
      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>
    <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 affects the event handling,
      in particular in combination with postponing events.
      We 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, timer => undefined},
    {ok, {locked,LockButton}, Data}.

callback_mode() ->
    [handle_event_function,state_entry_events].

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}};
	{{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;
%%
%% State: open
handle_event(
  EventType, EventContent,
  {open,LockButton}, #{timer := Timer} = Data) ->
    case {EventType,EventContent} of
	{enter,_OldState} ->
	    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.

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>
    <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[
...
handle_event(
  EventType, EventContent,
  {open,LockButton}, #{timer := Timer} = Data) ->
    case {EventType,EventContent} of
	{enter,_OldState} ->
	    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 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>
      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>