2  The Erlang I/O Protocol

2 The Erlang I/O Protocol

The I/O protocol in Erlang enables bi-directional communication between clients and servers.

  • The I/O server is a process that handles the requests and performs the requested task on, for example, an I/O device.

  • The client is any Erlang process wishing to read or write data from/to the I/O device.

The common I/O protocol has been present in OTP since the beginning, but has been undocumented and has also evolved over the years. In an addendum to Robert Virding's rationale, the original I/O protocol is described. This section describes the current I/O protocol.

The original I/O protocol was simple and flexible. Demands for memory efficiency and execution time efficiency have triggered extensions to the protocol over the years, making the protocol larger and somewhat less easy to implement than the original. It can certainly be argued that the current protocol is too complex, but this section describes how it looks today, not how it should have looked.

The basic ideas from the original protocol still hold. The I/O server and client communicate with one single, rather simplistic protocol and no server state is ever present in the client. Any I/O server can be used together with any client code, and the client code does not need to be aware of the I/O device that the I/O server communicates with.

As described in Robert's paper, I/O servers and clients communicate using io_request/io_reply tuples as follows:

{io_request, From, ReplyAs, Request}
{io_reply, ReplyAs, Reply}

The client sends an io_request tuple to the I/O server and the server eventually sends a corresponding io_reply tuple.

  • From is the pid() of the client, the process which the I/O server sends the I/O reply to.

  • ReplyAs can be any datum and is returned in the corresponding io_reply. The io module monitors the the I/O server and uses the monitor reference as the ReplyAs datum. A more complicated client can have many outstanding I/O requests to the same I/O server and can use different references (or something else) to differentiate among the incoming I/O replies. Element ReplyAs is to be considered opaque by the I/O server.

    Notice that the pid() of the I/O server is not explicitly present in tuple io_reply. The reply can be sent from any process, not necessarily the actual I/O server.

  • Request and Reply are described below.

When an I/O server receives an io_request tuple, it acts upon the Request part and eventually sends an io_reply tuple with the corresponding Reply part.

To output characters on an I/O device, the following Requests exist:

{put_chars, Encoding, Characters}
{put_chars, Encoding, Module, Function, Args}
  • Encoding is unicode or latin1, meaning that the characters are (in case of binaries) encoded as UTF-8 or ISO Latin-1 (pure bytes). A well-behaved I/O server is also to return an error indication if list elements contain integers > 255 when Encoding is set to latin1.

    Notice that this does not in any way tell how characters are to be put on the I/O device or handled by the I/O server. Different I/O servers can handle the characters however they want, this only tells the I/O server which format the data is expected to have. In the Module/Function/Args case, Encoding tells which format the designated function produces.

    Notice also that byte-oriented data is simplest sent using the ISO Latin-1 encoding.

  • Characters are the data to be put on the I/O device. If Encoding is latin1, this is an iolist(). If Encoding is unicode, this is an Erlang standard mixed Unicode list (one integer in a list per character, characters in binaries represented as UTF-8).

  • Module, Function, and Args denote a function that is called to produce the data (like io_lib:format/2).

    Args is a list of arguments to the function. The function is to produce data in the specified Encoding. The I/O server is to call the function as apply(Mod, Func, Args) and put the returned data on the I/O device as if it was sent in a {put_chars, Encoding, Characters} request. If the function returns anything else than a binary or list, or throws an exception, an error is to be sent back to the client.

The I/O server replies to the client with an io_reply tuple, where element Reply is one of:

{error, Error}
  • Error describes the error to the client, which can do whatever it wants with it. The io module typically returns it "as is".

To read characters from an I/O device, the following Requests exist:

{get_until, Encoding, Prompt, Module, Function, ExtraArgs}
  • Encoding denotes how data is to be sent back to the client and what data is sent to the function denoted by Module/Function/ExtraArgs. If the function supplied returns data as a list, the data is converted to this encoding. If the function supplied returns data in some other format, no conversion can be done, and it is up to the client-supplied function to return data in a proper way.

    If Encoding is latin1, lists of integers 0..255 or binaries containing plain bytes are sent back to the client when possible. If Encoding is unicode, lists with integers in the whole Unicode range or binaries encoded in UTF-8 are sent to the client. The user-supplied function always sees lists of integers, never binaries, but the list can contain numbers > 255 if Encoding is unicode.

  • Prompt is a list of characters (not mixed, no binaries) or an atom to be output as a prompt for input on the I/O device. Prompt is often ignored by the I/O server; if set to '', it is always to be ignored (and results in nothing being written to the I/O device).

  • Module, Function, and ExtraArgs denote a function and arguments to determine when enough data is written. The function is to take two more arguments, the last state, and a list of characters. The function is to return one of:

    {done, Result, RestChars}
    {more, Continuation}

    Result can be any Erlang term, but if it is a list(), the I/O server can convert it to a binary() of appropriate format before returning it to the client, if the I/O server is set in binary mode (see below).

    The function is called with the data the I/O server finds on its I/O device, returning one of:

    • {done, Result, RestChars} when enough data is read. In this case Result is sent to the client and RestChars is kept in the I/O server as a buffer for later input.

    • {more, Continuation}, which indicates that more characters are needed to complete the request.

    Continuation is sent as the state in later calls to the function when more characters are available. When no more characters are available, the function must return {done, eof, Rest}. The initial state is the empty list. The data when an end of file is reached on the IO device is the atom eof.

    An emulation of the get_line request can be (inefficiently) implemented using the following functions:

    -export([until_newline/3, get_line/1]).
    until_newline(_ThisFar,eof,_MyStopCharacter) ->
    until_newline(ThisFar,CharList,MyStopCharacter) ->
            lists:splitwith(fun(X) -> X =/= MyStopCharacter end,  CharList)
    	{L,[]} ->
    	{L2,[MyStopCharacter|Rest]} ->
    get_line(IoServer) ->
        IoServer ! {io_request,
                    {get_until, unicode, '', ?MODULE, until_newline, [$\n]}},
            {io_reply, IoServer, Data} ->

    Notice that the last element in the Request tuple ([$\n]) is appended to the argument list when the function is called. The function is to be called like apply(Module, Function, [ State, Data | ExtraArgs ]) by the I/O server.

A fixed number of characters is requested using the following Request:

{get_chars, Encoding, Prompt, N}
  • Encoding and Prompt as for get_until.

  • N is the number of characters to be read from the I/O device.

A single line (as in former example) is requested with the following Request:

{get_line, Encoding, Prompt}
  • Encoding and Prompt as for get_until.

Clearly, get_chars and get_line could be implemented with the get_until request (and indeed they were originally), but demands for efficiency have made these additions necessary.

The I/O server replies to the client with an io_reply tuple, where element Reply is one of:

{error, Error}
  • Data is the characters read, in list or binary form (depending on the I/O server mode, see the next section).

  • eof is returned when input end is reached and no more data is available to the client process.

  • Error describes the error to the client, which can do whatever it wants with it. The io module typically returns it as is.

Demands for efficiency when reading data from an I/O server has not only lead to the addition of the get_line and get_chars requests, but has also added the concept of I/O server options. No options are mandatory to implement, but all I/O servers in the Erlang standard libraries honor the binary option, which allows element Data of the io_reply tuple to be a binary instead of a list when possible. If the data is sent as a binary, Unicode data is sent in the standard Erlang Unicode format, that is, UTF-8 (notice that the function of the get_until request still gets list data regardless of the I/O server mode).

Notice that the get_until request allows for a function with the data specified as always being a list. Also, the return value data from such a function can be of any type (as is indeed the case when an io:fread/2,3 request is sent to an I/O server). The client must be prepared for data received as answers to those requests to be in various forms. However, the I/O server is to convert the results to binaries whenever possible (that is, when the function supplied to get_until returns a list). This is done in the example in section An Annotated and Working Example I/O Server.

An I/O server in binary mode affects the data sent to the client, so that it must be able to handle binary data. For convenience, the modes of an I/O server can be set and retrieved using the following I/O requests:

{setopts, Opts}
  • Opts is a list of options in the format recognized by the proplists module (and by the I/O server).

As an example, the I/O server for the interactive shell (in group.erl) understands the following options:

{binary, boolean()} (or binary/list)
{echo, boolean()}
{expand_fun, fun()}
{encoding, unicode/latin1} (or unicode/latin1)

Options binary and encoding are common for all I/O servers in OTP, while echo and expand are valid only for this I/O server. Option unicode notifies how characters are put on the physical I/O device, that is, if the terminal itself is Unicode-aware. It does not affect how characters are sent in the I/O protocol, where each request contains encoding information for the provided or returned data.

The I/O server is to send one of the following as Reply:

{error, Error}

An error (preferably enotsup) is to be expected if the option is not supported by the I/O server (like if an echo option is sent in a setopts request to a plain file).

To retrieve options, the following request is used:


This request asks for a complete list of all options supported by the I/O server as well as their current values.

The I/O server replies:

{error, Error}
  • OptList is a list of tuples {Option, Value}, where Option always is an atom.

The Request element can in itself contain many Requests by using the following format:

{requests, Requests}
  • Requests is a list of valid io_request tuples for the protocol. They must be executed in the order that they appear in the list. The execution is to continue until one of the requests results in an error or the list is consumed. The result of the last request is sent back to the client.

The I/O server can, for a list of requests, send any of the following valid results in the reply, depending on the requests in the list:

{ok, Data}
{ok, Options}
{error, Error}

The following I/O request is optional to implement and a client is to be prepared for an error return:

{get_geometry, Geometry}
  • Geometry is the atom rows or the atom columns.

The I/O server is to send the Reply as:

{ok, N}
{error, Error}
  • N is the number of character rows or columns that the I/O device has, if applicable to the I/O device handled by the I/O server, otherwise {error, enotsup} is a good answer.

If an I/O server encounters a request that it does not recognize (that is, the io_request tuple has the expected format, but the Request is unknown), the I/O server is to send a valid reply with the error tuple:

{error, request}

This makes it possible to extend the protocol with optional requests and for the clients to be somewhat backward compatible.

An I/O server is any process capable of handling the I/O protocol. There is no generic I/O server behavior, but could well be. The framework is simple, a process handling incoming requests, usually both I/O-requests and other I/O device-specific requests (positioning, closing, and so on).

The example I/O server stores characters in an ETS table, making up a fairly crude RAM file.

The module begins with the usual directives, a function to start the I/O server and a main loop handling the requests:


-export([start_link/0, init/0, loop/1, until_newline/3, until_enough/3]).

-define(CHARS_PER_REC, 10).

-record(state, {
	  position, % absolute
	  mode % binary | list

start_link() ->

init() ->
    Table = ets:new(noname,[ordered_set]),
    ?MODULE:loop(#state{table = Table, position = 0, mode=list}).

loop(State) ->
	{io_request, From, ReplyAs, Request} ->
	    case request(Request,State) of
		{Tag, Reply, NewState} when Tag =:= ok; Tag =:= error ->
		    reply(From, ReplyAs, Reply),
		{stop, Reply, _NewState} ->
		    reply(From, ReplyAs, Reply),
	%% Private message
	{From, rewind} ->
	    From ! {self(), ok},
	    ?MODULE:loop(State#state{position = 0});
	_Unknown ->

The main loop receives messages from the client (which can use the the io module to send requests). For each request, the function request/2 is called and a reply is eventually sent using function reply/3.

The "private" message {From, rewind} results in the current position in the pseudo-file to be reset to 0 (the beginning of the "file"). This is a typical example of I/O device-specific messages not being part of the I/O protocol. It is usually a bad idea to embed such private messages in io_request tuples, as that can confuse the reader.

First, we examine the reply function:

reply(From, ReplyAs, Reply) ->
    From ! {io_reply, ReplyAs, Reply}.

It sends the io_reply tuple back to the client, providing element ReplyAs received in the request along with the result of the request, as described earlier.

We need to handle some requests. First the requests for writing characters:

request({put_chars, Encoding, Chars}, State) ->
request({put_chars, Encoding, Module, Function, Args}, State) ->
	request({put_chars, Encoding, apply(Module, Function, Args)}, State)
	_:_ ->
	    {error, {error,Function}, State}

The Encoding says how the characters in the request are represented. We want to store the characters as lists in the ETS table, so we convert them to lists using function unicode:characters_to_list/2. The conversion function conveniently accepts the encoding types unicode and latin1, so we can use Encoding directly.

When Module, Function, and Arguments are provided, we apply it and do the same with the result as if the data was provided directly.

We handle the requests for retrieving data:

request({get_until, Encoding, _Prompt, M, F, As}, State) ->
    get_until(Encoding, M, F, As, State);
request({get_chars, Encoding, _Prompt, N}, State) ->
    %% To simplify the code, get_chars is implemented using get_until
    get_until(Encoding, ?MODULE, until_enough, [N], State);
request({get_line, Encoding, _Prompt}, State) ->
    %% To simplify the code, get_line is implemented using get_until
    get_until(Encoding, ?MODULE, until_newline, [$\n], State);

Here we have cheated a little by more or less only implementing get_until and using internal helpers to implement get_chars and get_line. In production code, this can be inefficient, but that depends on the frequency of the different requests. Before we start implementing functions put_chars/2 and get_until/5, we examine the few remaining requests:

request({get_geometry,_}, State) ->
    {error, {error,enotsup}, State};
request({setopts, Opts}, State) ->
    setopts(Opts, State);
request(getopts, State) ->
request({requests, Reqs}, State) ->
     multi_request(Reqs, {ok, ok, State});

Request get_geometry has no meaning for this I/O server, so the reply is {error, enotsup}. The only option we handle is binary/list, which is done in separate functions.

The multi-request tag (requests) is handled in a separate loop function applying the requests in the list one after another, returning the last result.

{error, request} must be returned if the request is not recognized:

request(_Other, State) ->
    {error, {error, request}, State}.

Next we handle the different requests, first the fairly generic multi-request type:

multi_request([R|Rs], {ok, _Res, State}) ->
    multi_request(Rs, request(R, State));
multi_request([_|_], Error) ->
multi_request([], Result) ->

We loop through the requests one at the time, stopping when we either encounter an error or the list is exhausted. The last return value is sent back to the client (it is first returned to the main loop and then sent back by function io_reply).

Requests getopts and setopts are also simple to handle. We only change or read the state record:

setopts(Opts0,State) ->
    Opts = proplists:unfold(
    case check_valid_opts(Opts) of
	true ->
	        case proplists:get_value(binary, Opts) of
		    true ->
		    false ->
		    _ ->
	false ->
check_valid_opts([]) ->
check_valid_opts([{binary,Bool}|T]) when is_boolean(Bool) ->
check_valid_opts(_) ->

getopts(#state{mode=M} = S) ->
    {ok,[{binary, case M of
		      binary ->
		      _ ->

As a convention, all I/O servers handle both {setopts, [binary]}, {setopts, [list]}, and {setopts,[{binary, boolean()}]}, hence the trick with proplists:substitute_negations/2 and proplists:unfold/1. If invalid options are sent to us, we send {error, enotsup} back to the client.

Request getopts is to return a list of {Option, Value} tuples. This has the twofold function of providing both the current values and the available options of this I/O server. We have only one option, and hence return that.

So far this I/O server is fairly generic (except for request rewind handled in the main loop and the creation of an ETS table). Most I/O servers contain code similar to this one.

To make the example runnable, we start implementing the reading and writing of the data to/from the ETS table. First function put_chars/3:

put_chars(Chars, #state{table = T, position = P} = State) ->
    R = P div ?CHARS_PER_REC,
    C = P rem ?CHARS_PER_REC,
    [ apply_update(T,U) || U <- split_data(Chars, R, C) ],
    {ok, ok, State#state{position = (P + length(Chars))}}.

We already have the data as (Unicode) lists and therefore only split the list in runs of a predefined size and put each run in the table at the current position (and forward). Functions split_data/3 and apply_update/2 are implemented below.

Now we want to read data from the table. Function get_until/5 reads data and applies the function until it says that it is done. The result is sent back to the client:

get_until(Encoding, Mod, Func, As, 
	  #state{position = P, mode = M, table = T} = State) ->
    case get_loop(Mod,Func,As,T,P,[]) of
	{done,Data,_,NewP} when is_binary(Data); is_list(Data) ->
		M =:= binary -> 
		     unicode:characters_to_binary(Data, unicode, Encoding),
		     State#state{position = NewP}};
		true ->
		    case check(Encoding, 
		               unicode:characters_to_list(Data, unicode))
			{error, _} = E ->
			    {error, E, State};
			List ->
			    {ok, List,
			     State#state{position = NewP}}
	{done,Data,_,NewP} ->
	    {ok, Data, State#state{position = NewP}};
	Error ->
	    {error, Error, State}

get_loop(M,F,A,T,P,C) ->
    {NewP,L} = get(P,T),
    case catch apply(M,F,[C,L|A]) of
	{done, List, Rest} ->
	    {done, List, [], NewP - length(Rest)};
	{more, NewC} ->
	_ ->

Here we also handle the mode (binary or list) that can be set by request setopts. By default, all OTP I/O servers send data back to the client as lists, but switching mode to binary can increase efficiency if the I/O server handles it in an appropriate way. The implementation of get_until is difficult to get efficient, as the supplied function is defined to take lists as arguments, but get_chars and get_line can be optimized for binary mode. However, this example does not optimize anything.

It is important though that the returned data is of the correct type depending on the options set. We therefore convert the lists to binaries in the correct encoding if possible before returning. The function supplied in the get_until request tuple can, as its final result return anything, so only functions returning lists can get them converted to binaries. If the request contains encoding tag unicode, the lists can contain all Unicode code points and the binaries are to be in UTF-8. If the encoding tag is latin1, the client is only to get characters in the range 0..255. Function check/2 takes care of not returning arbitrary Unicode code points in lists if the encoding was specified as latin1. If the function does not return a list, the check cannot be performed and the result is that of the supplied function untouched.

To manipulate the table we implement the following utility functions:

check(unicode, List) ->
check(latin1, List) ->
	[ throw(not_unicode) || X <- List,
				X > 255 ],
	throw:_ ->
	    {error,{cannot_convert, unicode, latin1}}

The function check provides an error tuple if Unicode code points > 255 are to be returned if the client requested latin1.

The two functions until_newline/3 and until_enough/3 are helpers used together with function get_until/5 to implement get_chars and get_line (inefficiently):

until_newline([],eof,_MyStopCharacter) ->
until_newline(ThisFar,eof,_MyStopCharacter) ->
until_newline(ThisFar,CharList,MyStopCharacter) ->
        lists:splitwith(fun(X) -> X =/= MyStopCharacter end,  CharList)
	{L,[]} ->
	{L2,[MyStopCharacter|Rest]} ->

until_enough([],eof,_N) ->
until_enough(ThisFar,eof,_N) ->
  when length(ThisFar) + length(CharList) >= N ->
    {Res,Rest} = my_split(N,ThisFar ++ CharList, []),
until_enough(ThisFar,CharList,_N) ->

As can be seen, the functions above are just the type of functions that are to be provided in get_until requests.

To complete the I/O server, we only need to read and write the table in an appropriate way:

get(P,Tab) ->
    R = P div ?CHARS_PER_REC,
    C = P rem ?CHARS_PER_REC,
    case ets:lookup(Tab,R) of
	[] ->
	[{R,List}] ->
	    case my_split(C,List,[]) of
		{_,[]} ->
		{_,Data} ->

my_split(0,Left,Acc) ->
my_split(_,[],Acc) ->
my_split(N,[H|T],Acc) ->

split_data([],_,_) ->
split_data(Chars, Row, Col) ->
    {This,Left} = my_split(?CHARS_PER_REC - Col, Chars, []),
    [ {Row, Col, This} | split_data(Left, Row + 1, 0) ].

apply_update(Table, {Row, Col, List}) ->     
    case ets:lookup(Table,Row) of
	[] ->
	    ets:insert(Table,{Row, lists:duplicate(Col,0) ++ List});
	[{Row, OldData}] ->
	    {Part1,_} = my_split(Col,OldData,[]),
	    {_,Part2} = my_split(Col+length(List),OldData,[]),
	    ets:insert(Table,{Row, Part1 ++ List ++ Part2})

The table is read or written in chunks of ?CHARS_PER_REC, overwriting when necessary. The implementation is clearly not efficient, it is just working.

This concludes the example. It is fully runnable and you can read or write to the I/O server by using, for example, the io module or even the file module. It is as simple as that to implement a fully fledged I/O server in Erlang.