diff options
Diffstat (limited to 'system/doc/design_principles/statem.xml')
-rw-r--r-- | system/doc/design_principles/statem.xml | 1522 |
1 files changed, 1522 insertions, 0 deletions
diff --git a/system/doc/design_principles/statem.xml b/system/doc/design_principles/statem.xml new file mode 100644 index 0000000000..ca0fce55e2 --- /dev/null +++ b/system/doc/design_principles/statem.xml @@ -0,0 +1,1522 @@ +<?xml version="1.0" encoding="utf-8" ?> +<!DOCTYPE chapter SYSTEM "chapter.dtd"> + +<chapter> + <header> + <copyright> + <year>2016</year> + <holder>Ericsson AB. All Rights Reserved.</holder> + </copyright> + <legalnotice> + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. + + </legalnotice> + + <title>gen_statem Behaviour</title> + <prepared></prepared> + <docno></docno> + <date></date> + <rev></rev> + <file>statem.xml</file> + </header> + <marker id="gen_statem behaviour"></marker> + <p> + This section is to be read with the + <seealso marker="stdlib:gen_statem"><c>gen_statem(3)</c></seealso> + manual page in STDLIB, where all interface functions and callback + functions are described in detail. + </p> + <p> + This is a new behaviour in OTP-19.0. + It has been thoroughly reviewed, is stable enough + to be used by at least two heavy OTP applications, and is here to stay. + But depending on user feedback, we do not expect + but might find it necessary to make minor + not backwards compatible changes into OTP-20.0, + so its state can be designated as "not quite experimental"... + </p> + +<!-- =================================================================== --> + + <section> + <title>Event Driven State Machines</title> + <p> + Established Automata theory does not deal much with + how a state transition is triggered, + but in general assumes that the output is a function + of the input (and the state) and that they are + some kind of values. + </p> + <p> + For an Event Driven State Machine the input is an event + that triggers a state transition and the output + is actions executed during the state transition. + It can analogously to the mathematical model of a + Finite State Machine be described as + a set of relations of the form: + </p> + <pre> +State(S) x Event(E) -> Actions(A), State(S')</pre> + <p>These relations are interpreted as meaning:</p> + <p> + If we are in state <c>S</c> and event <c>E</c> occurs, we + are to perform actions <c>A</c> and make a transition to + state <c>S'</c>. + </p> + <p> + Note that <c>S'</c> may be equal to <c>S</c>. + </p> + <p> + Since <c>A</c> and <c>S'</c> depend only on + <c>S</c> and <c>E</c> the kind of state machine described + here is a Mealy Machine. + (See for example the corresponding Wikipedia article) + </p> + <p> + Like most <c>gen_</c> behaviours, <c>gen_statem</c> keeps + a server <c>Data</c> besides the state. This and the fact that + there is no restriction on the number of states + (assuming enough virtual machine memory) + or on the number of distinct input events actually makes + a state machine implemented with this behaviour Turing complete. + But it feels mostly like an Event Driven Mealy Machine. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <marker id="callback_modes" /> + <title>Callback Modes</title> + <p> + The <c>gen_statem</c> behaviour supports two different callback modes. + In the mode + <seealso marker="stdlib:gen_statem#type-callback_mode"> + <c>state_functions</c>, + </seealso> + the state transition rules are written as a number of Erlang + functions, which conform to the following convention: + </p> + <pre> +StateName(EventType, EventContent, Data) -> + .. code for actions here ... + {next_state, NewStateName, NewData}.</pre> + <p> + In the mode + <seealso marker="stdlib:gen_statem#type-callback_mode"> + <c>handle_event_function</c> + </seealso> + there is only one + Erlang function that implements all state transition rules: + </p> + <pre> +handle_event(EventType, EventContent, State, Data) -> + .. code for actions here ... + {next_state, State', Data'}</pre> + <p> + Both these modes allow other return tuples + that you can find in the + <seealso marker="stdlib:gen_statem#Module:StateName/3"> + reference manual. + </seealso> + These other return tuples can for example stop the machine, + execute state transition actions on the machine engine itself + and send replies. + </p> + + <section> + <title>Choosing the Callback Mode</title> + <p> + The two + <seealso marker="#callback_modes">callback modes</seealso> + gives different possibilities + and restrictions, but one goal remains: + you want to handle all possible combinations of + events and states. + </p> + <p> + You can for example do this by focusing on one state at the time + and for every state ensure that all events are handled, + or the other way around focus on one event at the time + and ensure that it is handled in every state, + or mix these strategies. + </p> + <p> + With <c>state_functions</c> you are restricted to use + atom only states, and the <c>gen_statem</c> engine dispatches + on state name for you. This encourages the callback module + to gather the implementation of all event actions particular + to one state in the same place in the code + hence to focus on one state at the time. + </p> + <p> + This mode fits well when you have a regular state diagram + like the ones in this chapter that describes all events and actions + belonging to a state visually around that state, + and each state has its unique name. + </p> + <p> + With <c>handle_event_function</c> you are free to mix strategies + as you like because all events and states + are handled in the the same callback function. + </p> + <p> + This mode works equally well when you want to focus on + one event at the time or when you want to focus on + one state at the time, but the <c>handle_event/4</c> function + quickly grows too large to handle without introducing dispatching. + </p> + <p> + The mode enables the use of non-atom states for example + complex states or even hiearchical states. + If, for example, a state diagram is largely alike + for the client and for the server side of a protocol; + then you can have a state <c>{StateName,server}</c> or + <c>{StateName,client}</c> and since you do the dispatching + yourself you make <c>StateName</c> decide where in the code + to handle most events in the state. + The second element of the tuple is then used to select + whether to handle special client side or server side events. + </p> + </section> + </section> + +<!-- =================================================================== --> + + <section> + <title>Example</title> + <p> + This is an example starting off as equivalent to the the example in the + <seealso marker="fsm"><c>gen_fsm</c> behaviour</seealso> + description. In later chapters additions and tweaks are made + using features in <c>gen_statem</c> that <c>gen_fsm</c> does not have. + At the end of this section you can find the example again + with all the added features. + </p> + <p> + A door with a code lock can be viewed as a state machine. + Initially, the door is locked. Anytime someone presses a button, + this generates an event. + Depending on what buttons have been pressed before, + the sequence so far can be correct, incomplete, or wrong. + </p> + <p> + If it is correct, the door is unlocked for 10 seconds (10000 ms). + If it is incomplete, we wait for another button to be pressed. If + it is is wrong, we start all over, + waiting for a new button sequence. + </p> + <image file="../design_principles/code_lock.png"> + <icaption>Code lock state diagram</icaption> + </image> + <p> + We can implement such a code lock state machine using + <c>gen_statem</c> with the following callback module: + </p> + <marker id="ex"></marker> + <code type="erl"><![CDATA[ +-module(code_lock). +-behaviour(gen_statem). +-define(NAME, code_lock). +-define(CALLBACK_MODE, state_functions). + +-export([start_link/1]). +-export([button/1]). +-export([init/1,terminate/3,code_change/4]). +-export([locked/3,open/3]). + +start_link(Code) -> + gen_statem:start_link({local,?NAME}, ?MODULE, Code, []). + +button(Digit) -> + gen_statem:cast(?NAME, {button,Digit}). + + +init(Code) -> + do_lock(), + Data = #{code => Code, remaining => Code}, + {?CALLBACK_MODE,locked,Data}. + +locked( + cast, {button,Digit}, + #{code := Code, remaining := Remaining} = Data) -> + case Remaining of + [Digit] -> + do_unlock(), + {next_state,open,Data#{remaining := Code},10000}; + [Digit|Rest] -> % Incomplete + {next_state,locked,Data#{remaining := Rest}}; + _Wrong -> + {next_state,locked,Data#{remaining := Code}} + end. + +open(timeout, _, Data) -> + do_lock(), + {next_state,locked,Data}; +open(cast, {button,_}, Data) -> + do_lock(), + {next_state,locked,Data}. + +do_lock() -> + io:format("Lock~n", []). +do_unlock() -> + io:format("Unlock~n", []). + +terminate(_Reason, State, _Data) -> + State =/= locked andalso do_lock(), + ok. +code_change(_Vsn, State, Data, _Extra) -> + {?CALLBACK_MODE,State,Data}. + ]]></code> + <p>The code is explained in the next sections.</p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Starting gen_statem</title> + <p> + In the example in the previous section, the <c>gen_statem</c> is + started by calling <c>code_lock:start_link(Code)</c>: + </p> + <code type="erl"><![CDATA[ +start_link(Code) -> + gen_statem:start_link({local,?NAME}, ?MODULE, Code, []). + ]]></code> + <p> + <c>start_link</c> calls the function + <seealso marker="stdlib:gen_statem#start_link/4"> + <c>gen_statem:start_link/4</c> + </seealso> + which spawns and links to a new process; a <c>gen_statem</c>. + </p> + <list type="bulleted"> + <item> + <p> + The first argument, <c>{local,?NAME}</c>, specifies + the name. In this case, the <c>gen_statem</c> is locally + registered as <c>code_lock</c> through the macro <c>?NAME</c>. + </p> + <p> + If the name is omitted, the <c>gen_statem</c> is not registered. + Instead its pid must be used. The name can also be given + as <c>{global,Name}</c>, in which case the <c>gen_statem</c> is + registered using + <seealso marker="kernel:global#register_name/2"> + <c>global:register_name/2</c>. + </seealso> + </p> + </item> + <item> + <p> + The second argument, <c>?MODULE</c>, is the name of + the callback module, that is; the module where the callback + functions are located, which is this module. + </p> + <p> + The interface functions (<c>start_link/1</c> and <c>button/1</c>) + are located in the same module as the callback functions + (<c>init/1</c>, <c>locked/3</c>, and <c>open/3</c>). + It is normally good programming practice to have the client + side and the server side code contained in one module. + </p> + </item> + <item> + <p> + The third argument, <c>Code</c>, is a list of digits that + is the correct unlock code which is passsed + to the callback function <c>init/1</c>. + </p> + </item> + <item> + <p> + The fourth argument, <c>[]</c>, is a list of options. See the + <seealso marker="stdlib:gen_statem#start_link/3"> + <c>gen_statem:start_link/3</c> + </seealso> + manual page for available options. + </p> + </item> + </list> + <p> + If name registration succeeds, the new <c>gen_statem</c> process + calls the callback function <c>code_lock:init(Code)</c>. + This function is expected to return <c>{CallbackMode,State,Data}</c>, + where + <seealso marker="#callback_modes"> + <c>CallbackMode</c> + </seealso> + selects callback module state function mode, in this case + <seealso marker="stdlib:gen_statem#type-callback_mode"> + <c>state_functions</c> + </seealso> + through the macro <c>?CALLBACK_MODE</c> that is; each state + has got its own handler function. + <c>State</c> is the initial state of the <c>gen_statem</c>, + in this case <c>locked</c>; assuming the door is locked to begin with. + <c>Data</c> is the internal server data of the <c>gen_statem</c>. + Here the server data is a <seealso marker="stdlib:maps">map</seealso> + with the key <c>code</c> that stores + the correct button sequence and the key <c>remaining</c> + that stores the remaining correct button sequence + (the same as the <c>code</c> to begin with). + </p> + <code type="erl"><![CDATA[ +init(Code) -> + do_lock(), + Data = #{code => Code, remaining => Code}, + {?CALLBACK_MODE,locked,Data}. + ]]></code> + <p> + <seealso marker="stdlib:gen_statem#start_link/3"> + <c>gen_statem:start_link</c> + </seealso> + is synchronous. It does not return until the <c>gen_statem</c> + has been initialized and is ready to receive events. + </p> + <p> + <seealso marker="stdlib:gen_statem#start_link/3"> + <c>gen_statem:start_link</c> + </seealso> + must be used if the <c>gen_statem</c> + is part of a supervision tree, that is; started by a supervisor. + There is another function; + <seealso marker="stdlib:gen_statem#start/3"> + <c>gen_statem:start</c> + </seealso> + to start a standalone <c>gen_statem</c>, that is; + a <c>gen_statem</c> that is not part of a supervision tree. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Events and Handling them</title> + <p>The function notifying the code lock about a button event is + implemented using + <seealso marker="stdlib:gen_statem#cast/2"> + <c>gen_statem:cast/2</c>: + </seealso> + </p> + <code type="erl"><![CDATA[ +button(Digit) -> + gen_statem:cast(?NAME, {button,Digit}). + ]]></code> + <p> + The first argument is the name of the <c>gen_statem</c> and must + agree with the name used to start it so therefore we use the + same macro <c>?NAME</c> as when starting. + <c>{button,Digit}</c> is the actual event content. + </p> + <p> + The event is made into a message and sent to the <c>gen_statem</c>. + When the event is received, the <c>gen_statem</c> calls + <c>StateName(cast, Event, Data)</c>, which is expected to + return a tuple <c>{next_state,NewStateName,NewData}</c>. + <c>StateName</c> is the name of the current state and + <c>NewStateName</c> is the name of the next state to go to. + <c>NewData</c> is a new value for the server data of + the <c>gen_statem</c>. + </p> + <code type="erl"><![CDATA[ +locked( + cast, {button,Digit}, + #{code := Code, remaining := Remaining} = Data) -> + case Remaining of + [Digit] -> % Complete + do_unlock(), + {next_state,open,Data#{remaining := Code},10000}; + [Digit|Rest] -> % Incomplete + {next_state,locked,Data#{remaining := Rest}}; + [_|_] -> % Wrong + {next_state,locked,Data#{remaining := Code}} + end. + +open(timeout, _, Data) -> + do_lock(), + {next_state,locked,Data}; +open(cast, {button,_}, Data) -> + do_lock(), + {next_state,locked,Data}. + ]]></code> + <p> + If the door is locked and a button is pressed, the pressed + button is compared with the next correct button and, + depending on the result, the door is either unlocked + and the <c>gen_statem</c> goes to state <c>open</c>, + or the door remains in state <c>locked</c>. + </p> + <p> + If the pressed button is incorrect the server data + restarts from the start of the code sequence. + </p> + <p> + In state <c>open</c> any button locks the door since + any event cancels the event timer so we will not get + a timeout event after a button event. + </p> + </section> + + <section> + <title>Event Time-Outs</title> + <p> + When a correct code has been given, the door is unlocked and + the following tuple is returned from <c>locked/2</c>: + </p> + <code type="erl"><![CDATA[ +{next_state,open,Data#{remaining := Code},10000}; + ]]></code> + <p> + 10000 is a time-out value in milliseconds. + After this time, that is; 10 seconds, a time-out occurs. + Then, <c>StateName(timeout, 10000, Data)</c> is called. + The time-out occurs when the door has been in state <c>open</c> + for 10 seconds. After that the door is locked again: + </p> + <code type="erl"><![CDATA[ +open(timeout, _, Data) -> + do_lock(), + {next_state,locked,Data}; + ]]></code> + </section> + +<!-- =================================================================== --> + + <section> + <title>All State Events</title> + <p> + Sometimes an event can arrive in any state of the <c>gen_statem</c>. + It is convenient to handle these in a common state handler function + that all state functions call for events not specific to the state. + </p> + <p> + Let's introduce a <c>code_length/0</c> function that returns + the length of the correct code + (that should not be sensitive to reveal...). + We'll dispatch all events that are not state specific + to the common function <c>handle_event/3</c>. + </p> + <code type="erl"><![CDATA[ +... +-export([button/1,code_length/0]). +... + +code_length() -> + gen_statem:call(?NAME, code_length). + +... +locked(...) -> ... ; +locked(EventType, EventContent, Data) -> + handle_event(EventType, EventContent, Data). + +... +open(...) -> ... ; +open(EventType, EventContent, Data) -> + handle_event(EventType, EventContent, Data). + +handle_event({call,From}, code_length, #{code := Code} = Data) -> + {keep_state,Data,[{reply,From,length(Code)}]}. + ]]></code> + <p> + This example uses + <seealso marker="stdlib:gen_statem#call/2"> + <c>gen_statem:call/2</c> + </seealso> + which waits for a reply from the server. + The reply is sent with a <c>{reply,From,Reply}</c> tuple + in an action list in the <c>{keep_state,...}</c> tuple + that retains the current state. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>One Event Handler</title> + <p> + If you use the mode <c>handle_event_function</c> + all events are handled in <c>handle_event/4</c> and we + may (but do not have to) use an event-centered approach + where we dispatch on event first and then state: + </p> + <code type="erl"><![CDATA[ +... +-define(CALLBACK_MODE, state_functions). + +... +-export([handle_event/4]). + +... + +handle_event(cast, {button,Digit}, State, #{code := Code} = Data) -> + case State of + locked -> + case maps:get(remaining, Data) of + [Digit] -> % Complete + do_unlock(), + {next_state,open,Data#{remaining := Code},10000}; + [Digit|Rest] -> % Incomplete + {keep_state,Data#{remaining := Rest}}; + [_|_] -> % Wrong + {keep_state,Data#{remaining := Code}} + end; + open -> + do_lock(), + {next_state,locked,Data} + end; +handle_event(timeout, _, open, Data) -> + do_lock(), + {next_state,locked,Data}. + +... + ]]></code> + </section> + +<!-- =================================================================== --> + + <section> + <title>Stopping</title> + + <section> + <title>In a Supervision Tree</title> + <p> + If the <c>gen_statem</c> is part of a supervision tree, + no stop function is needed. + The <c>gen_statem</c> is automatically terminated by its supervisor. + Exactly how this is done is defined by a + <seealso marker="sup_princ#shutdown">shutdown strategy</seealso> + set in the supervisor. + </p> + <p> + If it is necessary to clean up before termination, the shutdown + strategy must be a time-out value and the <c>gen_statem</c> must + in the <c>init/1</c> function set itself to trap exit signals + by calling + <seealso marker="erts:erlang#process_flag/2"> + <c>process_flag(trap_exit, true)</c>. + </seealso> + When ordered to shutdown, the <c>gen_statem</c> then calls + the callback function + <c>terminate(shutdown, State, Data)</c>: + </p> + <code type="erl"><![CDATA[ +init(Args) -> + process_flag(trap_exit, true), + do_lock(), + ... + ]]></code> + <p> + In this example we let the <c>terminate/3</c> function + lock the door if it is open so we do not accidentally leave the door + open when the supervision tree terminates. + </p> + <code type="erl"><![CDATA[ +terminate(_Reason, State, _Data) -> + State =/= locked andalso do_lock(), + ok. + ]]></code> + </section> + + <section> + <title>Standalone gen_statem</title> + <p> + If the <c>gen_statem</c> is not part of a supervision tree, + it can be stopped using + <seealso marker="stdlib:gen_statem#stop/1"> + <c>gen_statem:stop</c>, + </seealso> + preferably through an API function: + </p> + <code type="erl"><![CDATA[ +... +-export([start_link/1,stop/0]). + +... +stop() -> + gen_statem:stop(?NAME). + ]]></code> + <p> + This makes the <c>gen_statem</c> call the <c>terminate/3</c> + callback function just like for a supervised server + and waits for the process to terminate. + </p> + </section> + </section> + +<!-- =================================================================== --> + + <section> + <title>Actions</title> + <p> + In the first chapters we mentioned actions as a part of + the general state machine model, and these actions + are implemented with the code the <c>gen_statem</c> + callback module executes in an event handling + callback function before returning + to the <c>gen_statem</c> engine. + </p> + <p> + There are more specific state transition actions + that a callback function can order the <c>gen_statem</c> + engine to do after the callback function return. + These are ordered by returning a list of + <seealso marker="stdlib:gen_statem#type-action"> + actions + </seealso> + in the + <seealso marker="stdlib:gen_statem#type-state_function_result"> + return tuple + </seealso> + from the + <seealso marker="stdlib:gen_statem#Module:StateName/3"> + callback function. + </seealso> + These state transition actions affect the <c>gen_statem</c> + engine itself. They can: + </p> + <list type="bulleted"> + <item>Postpone the current event.</item> + <item>Hibernate the <c>gen_statem</c>.</item> + <item>Start an event timeout.</item> + <item>Reply to a caller.</item> + <item>Generate the next event to handle.</item> + </list> + <p> + We have mentioned the event timeout + and replying to a caller in the example above. + An example of event postponing comes in later in this chapter. + See the + <seealso marker="stdlib:gen_statem#type-action"> + reference manual + </seealso> + for details. You can for example actually reply to several callers + and generate multiple next events to handle. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Event Types</title> + <p> + So far we have mentioned a few + <seealso marker="stdlib:gen_statem#type-event_type"> + event types. + </seealso> + Events of all types are handled in the same callback function, + for a given state, and the function gets + <c>EventType</c> and <c>EventContent</c> as arguments. + </p> + <p> + Here is the complete list of event types and where + they come from: + </p> + <taglist> + <tag><c>cast</c></tag> + <item> + Generated by + <seealso marker="stdlib:gen_statem#cast/2"> + <c>gen_statem:cast</c>. + </seealso> + </item> + <tag><c>{call,From}</c></tag> + <item> + Generated by + <seealso marker="stdlib:gen_statem#call/2"> + <c>gen_statem:call</c> + </seealso> + where <c>From</c> is the reply address to use + when replying either through the state transition action + <c>{reply,From,Msg}</c> or by calling + <seealso marker="stdlib:gen_statem#reply/1"> + <c>gen_statem:reply</c>. + </seealso> + </item> + <tag><c>info</c></tag> + <item> + Generated by any regular process message sent to + the <c>gen_statem</c> process. + </item> + <tag><c>timeout</c></tag> + <item> + Generated by the state transition action + <c>{timeout,Time,EventContent}</c> (or its short form <c>Time</c>) + timer timing out. + </item> + <tag><c>internal</c></tag> + <item> + Generated by the state transition action + <c>{next_event,internal,EventContent}</c>. + In fact all event types above can be generated using + <c>{next_event,EventType,EventContent}</c>. + </item> + </taglist> + </section> + +<!-- =================================================================== --> + + <section> + <title>State Timeouts</title> + <p> + The timeout event generated by the state transition action + <c>{timeout,Time,EventContent}</c> is an event timeout, + that is; if an event arrives the timer is cancelled. + You get either an event or a timeout but not both. + </p> + <p> + Often you want a timer to not be cancelled by any event + or you want to start a timer in one state and respond + to the timeout in another. This can be accomplished + with a regular erlang timer: + <seealso marker="erts:erlang#start_timer/4"> + <c>erlang:start_timer</c>. + </seealso> + </p> + <p> + Looking at the example in this chapter so far; using the + <c>gen_statem</c> event timer has the consequence that + if a button event is generated while in the <c>open</c> state, + the timeout is cancelled and the button event is delivered. + Therefore we chose to lock the door if this happended. + </p> + <p> + Suppose we do not want a button to lock the door, + instead we want to ignore button events in the <c>open</c> state. + Then we start a timer when entering the <c>open</c> state + and wait for it to expire while ignoring button events: + </p> + <code type="erl"><![CDATA[ +... +locked( + cast, {button,Digit}, + #{code := Code, remaining := Remaining} = Data) -> + case Remaining of + [Digit] -> + do_unlock(), + Tref = erlang:start_timer(10000, self(), lock), + {next_state,open,Data#{remaining := Code, timer := Tref}}; +... + +open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) -> + do_lock(), + {next_state,locked,Data}; +open(cast, {button,_}, Data) -> + {keep_state,Data}; +... + ]]></code> + <p> + If you need to cancel a timer due to some other event you can use + <seealso marker="erts:erlang#cancel_timer/2"> + <c>erlang:cancel_timer(Tref)</c>. + </seealso> + Note that a timeout message can not arrive after this, + unless you have postponed it (see the next section) before, + so make sure you do not accidentally postpone such messages. + </p> + <p> + Another way to cancel a timer is to not cancel it, + but instead to ignore it if it arrives in a state + where it is known to be late. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Postponing Events</title> + <p> + If you want to ignore a particular event in the current state + and handle it in a future state, you can postpone the event. + A postponed event is retried after the state has + changed i.e <c>OldState =/= NewState</c>. + </p> + <p> + Postponing is ordered by the state transition + <seealso marker="stdlib:gen_statem#type-action"> + action + </seealso> + <c>postpone</c>. + </p> + <p> + In this example, instead of ignoring button events + while in the <c>open</c> state we can postpone them + and they will be queued and later handled in the <c>locked</c> state: + </p> + <code type="erl"><![CDATA[ +... +open(cast, {button,_}, Data) -> + {keep_state,Data,[postpone]}; +... + ]]></code> + <p> + The fact that a postponed event is only retried after a state change + translates into a requirement on the event and state space: + if you have a choice between storing a state data item + in the <c>State</c> or in the <c>Data</c>; + should a change in the item value affect which events that + are handled, then this item ought to be part of the state. + </p> + <p> + What you want to avoid is that you maybe much later decide + to postpone an event in one state and by misfortune it is never retried + because the code only changes the <c>Data</c> but not the <c>State</c>. + </p> + + <section> + <title>Fuzzy State Diagrams</title> + <p> + It is not uncommon that a state diagram does not specify + how to handle events that are not illustrated + in a particular state in the diagram. + Hopefully this is described in an associated text + or from the context. + </p> + <p> + Possible actions may be; ignore as in drop the event + (maybe log it) or deal with the event in some other state + as in postpone it. + </p> + </section> + + <section> + <title>Selective Receive</title> + <p> + Erlang's selective receive statement is often used to + describe simple state machine examples in straightforward + Erlang code. Here is a possible implementation of + the first example: + </p> + <code type="erl"><![CDATA[ +-module(code_lock). +-define(NAME, code_lock_1). +-export([start_link/1,button/1]). + +start_link(Code) -> + spawn( + fun () -> + true = register(?NAME, self()), + do_lock(), + locked(Code, Code) + end). + +button(Digit) -> + ?NAME ! {button,Digit}. + +locked(Code, [Digit|Remaining]) -> + receive + {button,Digit} when Remaining =:= [] -> + do_unlock(), + open(Code); + {button,Digit} -> + locked(Code, Remaining); + {button,_} -> + locked(Code, Code) + end. + +open(Code) -> + receive + after 10000 -> + do_lock(), + locked(Code, Code) + end. + +do_lock() -> + io:format("Locked~n", []). +do_unlock() -> + io:format("Open~n", []). + ]]></code> + <p> + The selective receive in this case causes <c>open</c> + to implicitly postpone any events to the <c>locked</c> state. + </p> + <p> + A selective receive can not be used from a <c>gen_statem</c> + behaviour just as for any <c>gen_*</c> behavior + since the receive statement is within the <c>gen_*</c> engine itself. + It has to be there because all + <seealso marker="stdlib:sys"><c>sys</c></seealso> + compatible behaviours must respond to system messages and therefore + do that in their engine receive loop, + passing non-system messages to the callback module. + </p> + <p> + The state transition + <seealso marker="stdlib:gen_statem#type-action"> + action + </seealso> + <c>postpone</c> is designed to be able to model + selective receives. A selective receive implicitly postpones + any not received events, but the <c>postpone</c> + state transition action explicitly postpones one received event. + </p> + <p> + Other than that both mechanisms have got the same theoretical + time and memory complexity, while the selective receive + language construct has got smaller constant factors. + </p> + </section> + </section> + +<!-- =================================================================== --> + + <section> + <title>Self Generated Events</title> + <p> + It may be beneficial in some cases to be able to generate events + to your own state machine. + This can be done with the state transition + <seealso marker="stdlib:gen_statem#type-action"> + action + </seealso> + <c>{next_event,EventType,EventContent}</c>. + </p> + <p> + You can generate events of any existing + <seealso marker="stdlib:gen_statem#type-action"> + type, + </seealso> + but the <c>internal</c> type can only be generated through the + <c>next_event</c> action and hence can not come from an external source, + so you can be certain that an <c>internal</c> event is an event + from your state machine to itself. + </p> + <p> + One example of using self generated events may be when you have + a state machine specification that uses state entry actions. + That you could code using a dedicated function + to do the state transition. But if you want that code to be + visible besides the other state logic you can insert + an <c>internal</c> event that does the entry actions. + This has the same unfortunate consequence as using + state transition functions that everywhere you go to + the state in question you will have to explicitly + insert the <c>internal</c> event + or use state transition function. + </p> + <p> + Here is an implementation of entry actions + using <c>internal</c> events with content <c>enter</c> + utilizing a helper function <c>enter/3</c> for state entry: + </p> + <code type="erl"><![CDATA[ +... +-define(CALLBACK_MODE, state_functions). + +... + +init(Code) -> + process_flag(trap_exit, true), + Data = #{code => Code}, + enter(?CALLBACK_MODE, locked, Data). + +... + +locked(internal, enter, _Data) -> + do_lock(), + {keep_state,Data#{remaining => Code}}; +locked( + cast, {button,Digit}, + #{code := Code, remaining := Remaining} = Data) -> + case Remaining of + [Digit] -> + enter(next_state, open, Data); +... + +open(internal, enter, _Data) -> + Tref = erlang:start_timer(10000, self(), lock), + do_unlock(), + {keep_state,Data#{timer => Tref}}; +open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) -> + enter(next_state, locked, Data); +... + +enter(Tag, State, Data) -> + {Tag,State,Data,[{next_event,internal,enter}]}. + ]]></code> + </section> + +<!-- =================================================================== --> + + <section> + <title>Example Revisited</title> + <p> + Here is the example after all mentioned modifications + and some more utilizing the entry actions, + which deserves a new state diagram: + </p> + <image file="../design_principles/code_lock_2.png"> + <icaption>Code lock state diagram revisited</icaption> + </image> + <p> + Note that this state diagram does not specify how to handle + a button event in the state <c>open</c>, so you will have to + read some other place that is here that unspecified events + shall be ignored as in not consumed but handled in some other state. + Nor does it show that the <c>code_length/0</c> call shall be + handled in every state. + </p> + + <section> + <title>Callback Mode: state_functions</title> + <p> + Using state functions: + </p> + <code type="erl"><![CDATA[ +-module(code_lock). +-behaviour(gen_statem). +-define(NAME, code_lock_2). +-define(CALLBACK_MODE, state_functions). + +-export([start_link/1,stop/0]). +-export([button/1,code_length/0]). +-export([init/1,terminate/3,code_change/4]). +-export([locked/3,open/3]). + +start_link(Code) -> + gen_statem:start_link({local,?NAME}, ?MODULE, Code, []). +stop() -> + gen_statem:stop(?NAME). + +button(Digit) -> + gen_statem:cast(?NAME, {button,Digit}). +code_length() -> + gen_statem:call(?NAME, code_length). + +init(Code) -> + process_flag(trap_exit, true), + Data = #{code => Code}, + enter(?CALLBACK_MODE, locked, Data). + +locked(internal, enter, #{code := Code} = Data) -> + do_lock(), + {keep_state,Data#{remaining => Code}}; +locked( + cast, {button,Digit}, + #{code := Code, remaining := Remaining} = Data) -> + case Remaining of + [Digit] -> % Complete + enter(next_state, open, Data); + [Digit|Rest] -> % Incomplete + {keep_state,Data#{remaining := Rest}}; + [_|_] -> % Wrong + {keep_state,Data#{remaining := Code}} + end; +locked(EventType, EventContent, Data) -> + handle_event(EventType, EventContent, Data). + +open(internal, enter, Data) -> + Tref = erlang:start_timer(10000, self(), lock), + do_unlock(), + {keep_state,Data#{timer => Tref}}; +open(info, {timeout,Tref,lock}, #{timer := Tref} = Data) -> + enter(next_state, locked, Data); +open(cast, {button,_}, _) -> + {keep_state_and_data,[postpone]}; +open(EventType, EventContent, Data) -> + handle_event(EventType, EventContent, Data). + +handle_event({call,From}, code_length, #{code := Code}) -> + {keep_state_and_data,[{reply,From,length(Code)}]}. +enter(Tag, State, Data) -> + {Tag,State,Data,[{next_event,internal,enter}]}. + +do_lock() -> + io:format("Locked~n", []). +do_unlock() -> + io:format("Open~n", []). + +terminate(_Reason, State, _Data) -> + State =/= locked andalso do_lock(), + ok. +code_change(_Vsn, State, Data, _Extra) -> + {?CALLBACK_MODE,State,Data}. + ]]></code> + </section> + + <section> + <title>Callback Mode: handle_event_function</title> + <p> + What to change to use one <c>handle_event/4</c> function. + Here a clean first-dispatch-on-event approach + does not work that well due to the generated + entry actions: + </p> + <code type="erl"><![CDATA[ +... +-define(CALLBACK_MODE, handle_event_function). + +... +-export([handle_event/4]). + +... + +%% State: locked +handle_event(internal, enter, locked, #{code := Code} = Data) -> + do_lock(), + {keep_state,Data#{remaining => Code}}; +handle_event( + cast, {button,Digit}, locked, + #{code := Code, remaining := Remaining} = Data) -> + case Remaining of + [Digit] -> % Complete + enter(next_state, open, Data); + [Digit|Rest] -> % Incomplete + {keep_state,Data#{remaining := Rest}}; + [_|_] -> % Wrong + {keep_state,Data#{remaining := Code}} + end; +%% +%% State: open +handle_event(internal, enter, open, Data) -> + Tref = erlang:start_timer(10000, self(), lock), + do_unlock(), + {keep_state,Data#{timer => Tref}}; +handle_event(info, {timeout,Tref,lock}, open, #{timer := Tref} = Data) -> + enter(next_state, locked, Data); +handle_event(cast, {button,_}, open, _) -> + {keep_state_and_data,[postpone]}; +%% +%% Any state +handle_event({call,From}, code_length, _State, #{code := Code}) -> + {keep_state_and_data,[{reply,From,length(Code)}]}. + +... + ]]></code> + </section> + <p> + Note that postponing buttons from the <c>locked</c> state + to the <c>open</c> state feels like the wrong thing to do + for a code lock, but it at least illustrates event postponing. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Filter the State</title> + <p> + The example servers so far in this chapter will for example + when killed by an exit signal or due to an internal error + print out the full internal state in the error log. + This state contains both the code lock code + and which digits that remains to unlock. + </p> + <p> + This state data can be regarded as sensitive, + and maybe not what you want in the error log + because of something unpredictable happening. + </p> + <p> + Another reason to filter the state can be + that the state is too big to print out since it fills + the error log with uninteresting details. + </p> + <p> + To avoid this you can format the internal state + that gets in the error log and gets returned from + <seealso marker="stdlib:sys#get_status/1"> + <c>sys:get_status/1,2</c> + </seealso> + by implementing the + <seealso marker="stdlib:gen_statem#Module:format_status/2"> + <c>Module:format_status/2</c> + </seealso> + function, for example like this: + </p> + <code type="erl"><![CDATA[ +... +-export([init/1,terminate/3,code_change/4,format_status/2]). +... + +format_status(Opt, [_PDict,State,Data]) -> + StateData = + {State, + maps:filter( + fun (code, _) -> false; + (remaining, _) -> false; + (_, _) -> true + end, + Data)}, + case Opt of + terminate -> + StateData; + normal -> + [{data,[{"State",StateData}]}] + end. + ]]></code> + <p> + It is not mandatory to implement a + <seealso marker="stdlib:gen_statem#Module:format_status/2"> + <c>Module:format_status/2</c> + </seealso> + function. If you do not a default implementation is used that + does the same as this example function without filtering + the <c>Data</c> term that is: <c>StateData = {State,Data}</c>. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Complex State</title> + <p> + The callback mode + <seealso marker="stdlib:gen_statem#type-callback_mode"> + <c>handle_event_function</c> + </seealso> + enables using a non-atom state as described in + <seealso marker="#callback_modes"> + Callback Modes, + </seealso> + for example a complex state term like a tuple. + </p> + <p> + One reason to use this is when you have + a state item that affects the event handling + in particular when combining that with postponing events. + Let us complicate the previous example + by introducing a configurable lock button + (this is the state item in question) + that in the <c>open</c> state immediately locks the door, + and an API function <c>set_lock_button/1</c> to set the lock button. + </p> + <p> + Suppose now that we call <c>set_lock_button</c> + while the door is open, + and have already postponed a button event + that up until now was not the lock button; + the sensible thing might be to say that + the button was pressed too early so it should + not be recognized as the lock button, + but then it might be surprising that a button event + that now is the lock button event arrives (as retried postponed) + immediately after the state transits to <c>locked</c>. + </p> + <p> + So let us make the <c>button/1</c> function synchronous + by using <c>gen_statem:call</c>, + and still postpone its events in the <c>open</c> state. + Then a call to <c>button/1</c> during the <c>open</c> + state will not return until the state transits to <c>locked</c> + since it is there the event is handled and the reply is sent. + </p> + <p> + If now one process calls <c>set_lock_button/1</c> + to change the lock button while some other process + hangs in <c>button/1</c> with the new lock button + it could be expected that the hanging lock button call + immediately takes effect and locks the lock. + Therefore we make the current lock button a part of the state + so when we change the lock button the state will change + and all postponed events will be retried. + </p> + <p> + We define the state as <c>{StateName,LockButton}</c> + where <c>StateName</c> is as before + and <c>LockButton</c> is the current lock button: + </p> + <code type="erl"><![CDATA[ +-module(code_lock). +-behaviour(gen_statem). +-define(NAME, code_lock_3). +-define(CALLBACK_MODE, handle_event_function). + +-export([start_link/2,stop/0]). +-export([button/1,code_length/0,set_lock_button/1]). +-export([init/1,terminate/3,code_change/4,format_status/2]). +-export([handle_event/4]). + +start_link(Code, LockButton) -> + gen_statem:start_link( + {local,?NAME}, ?MODULE, {Code,LockButton}, []). +stop() -> + gen_statem:stop(?NAME). + +button(Digit) -> + gen_statem:call(?NAME, {button,Digit}). +code_length() -> + gen_statem:call(?NAME, code_length). +set_lock_button(LockButton) -> + gen_statem:call(?NAME, {set_lock_button,LockButton}). + +init({Code,LockButton}) -> + process_flag(trap_exit, true), + Data = #{code => Code, remaining => undefined, timer => undefined}, + enter(?CALLBACK_MODE, {locked,LockButton}, Data, []). + +handle_event( + {call,From}, {set_lock_button,NewLockButton}, + {StateName,OldLockButton}, Data) -> + {next_state,{StateName,NewLockButton},Data, + [{reply,From,OldLockButton}]}; +handle_event( + {call,From}, code_length, + {_StateName,_LockButton}, #{code := Code}) -> + {keep_state_and_data, + [{reply,From,length(Code)}]}; +handle_event( + EventType, EventContent, + {locked,LockButton}, #{code := Code, remaining := Remaining} = Data) -> + case {EventType,EventContent} of + {internal,enter} -> + do_lock(), + {keep_state,Data#{remaining := Code}}; + {{call,From},{button,Digit}} -> + case Remaining of + [Digit] -> % Complete + next_state( + {open,LockButton}, Data, + [{reply,From,ok}]); + [Digit|Rest] -> % Incomplete + {keep_state,Data#{remaining := Rest}, + [{reply,From,ok}]}; + [_|_] -> % Wrong + {keep_state,Data#{remaining := Code}, + [{reply,From,ok}]} + end + end; +handle_event( + EventType, EventContent, + {open,LockButton}, #{timer := Timer} = Data) -> + case {EventType,EventContent} of + {internal,enter} -> + Tref = erlang:start_timer(10000, self(), lock), + do_unlock(), + {keep_state,Data#{timer := Tref}}; + {info,{timeout,Timer,lock}} -> + next_state({locked,LockButton}, Data, []); + {{call,From},{button,Digit}} -> + if + Digit =:= LockButton -> + erlang:cancel_timer(Timer), + next_state( + {locked,LockButton}, Data, + [{reply,From,locked}]); + true -> + {keep_state_and_data, + [postpone]} + end + end. + +next_state(State, Data, Actions) -> + enter(next_state, State, Data, Actions). +enter(Tag, State, Data, Actions) -> + {Tag,State,Data,[{next_event,internal,enter}|Actions]}. + +do_lock() -> + io:format("Locked~n", []). +do_unlock() -> + io:format("Open~n", []). + +terminate(_Reason, State, _Data) -> + State =/= locked andalso do_lock(), + ok. +code_change(_Vsn, State, Data, _Extra) -> + {?CALLBACK_MODE,State,Data}. +format_status(Opt, [_PDict,State,Data]) -> + StateData = + {State, + maps:filter( + fun (code, _) -> false; + (remaining, _) -> false; + (_, _) -> true + end, + Data)}, + case Opt of + terminate -> + StateData; + normal -> + [{data,[{"State",StateData}]}] + end. + ]]></code> + <p> + It may be an ill-fitting model for a physical code lock + that the <c>button/1</c> call might hang until the lock + is locked. But for an API in general it is really not + that strange. + </p> + </section> + +<!-- =================================================================== --> + + <section> + <title>Hibernation</title> + <p> + If you have many servers in one node + and they have some state(s) in their lifetime in which + the servers can be expected to idle for a while, + and the amount of heap memory all these servers need + is a problem; then it is possible to minimize + the memory footprint of a server by hibernating it through + <seealso marker="stdlib:proc_lib#hibernate/3"> + <c>proc_lib:hibernate/3</c>. + </seealso> + </p> + <note> + <p> + To hibernate a process is rather costly. See + <seealso marker="erts:erlang#hibernate/3"> + <c>erlang:hibernate/3</c>. + </seealso> + It is in general not something you want to do + after every event. + </p> + </note> + <p> + We can in this example hibernate in the <c>{open,_}</c> state + since what normally happens in that state is that + the state timeout after a while + triggers a transition to <c>{locked,_}</c>: + </p> + <code type="erl"><![CDATA[ +... +handle_event( + EventType, EventContent, + {open,LockButton}, #{timer := Timer} = Data) -> + case {EventType,EventContent} of + {internal,enter} -> + Tref = erlang:start_timer(10000, self(), lock), + do_unlock(), + {keep_state,Data#{timer := Tref},[hibernate]}; +... + ]]></code> + <p> + The + <seealso marker="stdlib:gen_statem#type-hibernate"> + <c>[hibernate]</c> + </seealso> + action list on the last line + when entering the <c>{open,_}</c> state is the only change. + If any event arrives in the <c>{open,_},</c> state we + do not bother to re-hibernate, so the server stays + awake after any event. + </p> + <p> + To change that we would need to insert + the <c>hibernate</c> action in more places, + for example for the state independent <c>set_lock_button</c> + and <c>code_length</c> operations that then would have to + be aware of using <c>hibernate</c> while in the + <c>{open,_}</c> state which would clutter the code. + </p> + <p> + This server probably does not use an amount of + heap memory worth hibernating for. + To gain anything from hibernation your server would + have to actually produce some garbage during callback execution, + for which this example server may serve as a bad example. + </p> + </section> + +</chapter> |