1 The Erlang I/O-protocol

The I/O-protocol in Erlang specifies a way for a client to communicate with an I/O server and vice versa. The I/O server is a process that handles the requests and performs the requested task on e.g. an IO device. The client is any Erlang process wishing to read or write data from/to the IO device.

The common I/O-protocol has been present in OTP since the beginning, but has been fairly undocumented and has also somewhat evolved over the years. In an addendum to Robert Virdings rationale the original I/O-protocol is described. This document describes the current I/O-protocol.

The original I/O-protocol was simple and flexible. Demands for spacial and execution time efficiency has 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 text 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 client code need not be aware of the actual IO device the I/O server communicates with.

1.1  Protocol Basics

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 IO reply to.
  • ReplyAs can be any datum and is returned in the corresponding io_reply. The io module monitors the I/O server, and uses the monitor reference as the ReplyAs datum. A more complicated client could have several outstanding I/O requests to the same I/O server and would then use different references (or something else) to differentiate among the incoming IO replies. The ReplyAs element should be considered opaque by the I/O server. Note that the pid() of the I/O server is not explicitly present in the io_reply tuple. The reply can be sent from any process, not necessarily the actual I/O server. The ReplyAs element is the only thing that connects one I/O request with an I/O-reply.
  • Request and Reply are described below.

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

1.2  Output Requests

To output characters on an IO device, the following Requests exist:

{put_chars, Encoding, Characters}
{put_chars, Encoding, Module, Function, Args}

  • Encoding is either unicode or latin1, meaning that the characters are (in case of binaries) encoded as either UTF-8 or ISO-latin-1 (pure bytes). A well behaved I/O server should also return error if list elements contain integers > 255 when Encoding is set to latin1. Note that this does not in any way tell how characters should be put on the actual IO device or how the I/O server should handle them. Different I/O servers may handle the characters however they want, this simply 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. Note that byte-oriented data is simplest sent using the ISO-latin-1 encoding.
  • Characters are the data to be put on the IO 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 which will be called to produce the data (like io_lib:format/2). Args is a list of arguments to the function. The function should produce data in the given Encoding. The I/O server should call the function as apply(Mod, Func, Args) and will put the returned data on the IO 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 should be sent back to the client.

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

{error, Error}

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

For backward compatibility the following Requests should also be handled by an I/O server (these requests should not be present after R15B of OTP):

{put_chars, Characters}
{put_chars, Module, Function, Args}

These should behave as {put_chars, latin1, Characters} and {put_chars, latin1, Module, Function, Args} respectively.

1.3  Input Requests

To read characters from an IO 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 however 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 will always see lists of integers, never binaries, but the list may contain numbers > 255 if the 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 IO device. Prompt is often ignored by the I/O server and if set to '' it should always be ignored (and result in nothing being written to the IO device).
  • Module, Function, and ExtraArgs denote a function and arguments to determine when enough data is written. The function should take two additional arguments, the last state, and a list of characters. The function should return one of:

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

    The Result can be any Erlang term, but if it is a list(), the I/O server may 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 will be called with the data the I/O server finds on its IO device, returning {done, Result, RestChars} when enough data is read (in which case Result is sent to the client and RestChars is kept in the I/O server as a buffer for subsequent input) or {more, Continuation}, indicating that more characters are needed to complete the request. The Continuation will be sent as the state in subsequent calls to the function when more characters are available. When no more characters are available, the function shall return {done, eof, Rest}. The initial state is the empty list and the data when an end of file is reached on the IO device is the atom eof. An emulation of the get_line request could 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} ->

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

A fixed number of characters is requested using this Request:

{get_chars, Encoding, Prompt, N}

  • Encoding and Prompt as for get_until.
  • N is the number of characters to be read from the IO device.

A single line (like in the example above) is requested with this Request:

{get_line, Encoding, Prompt}

  • Encoding and Prompt as above.

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

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

{error, Error}

  • Data is the characters read, in either list or binary form (depending on the I/O server mode, see below).
  • Error describes the error to the client, which may do whatever it wants with it. The Erlang io module typically returns it as is.
  • eof is returned when input end is reached and no more data is available to the client process.

For backward compatibility the following Requests should also be handled by an I/O server (these reqeusts should not be present after R15B of OTP):

{get_until, Prompt, Module, Function, ExtraArgs}
{get_chars, Prompt, N}
{get_line, Prompt}

These should behave as {get_until, latin1, Prompt, Module, Function, ExtraArgs}, {get_chars, latin1, Prompt, N} and {get_line, latin1, Prompt} respectively.

1.4  I/O-server Modes

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 the Data element 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 will be sent in the standard Erlang Unicode format, i.e. UTF-8 (note that the function of the get_until request still gets list data regardless of the I/O server mode).

Note that i.e. 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 request is sent to an I/O server). The client has to be prepared for data received as answers to those requests to be in a variety of forms, but the I/O server should convert the results to binaries whenever possible (i.e. when the function supplied to get_until actually returns a list). The example shown later in this text does just that.

An I/O-server in binary mode will affect the data sent to the client, so that it has to be able to handle binary data. For convenience, it is possible to set and retrieve the modes of an I/O server using the following I/O requests:

{setopts, Opts}

  • Opts is a list of options in the format recognized by proplists (and of course by the I/O server itself).

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)

- of which the binary and encoding options are common for all I/O servers in OTP, while echo and expand are valid only for this I/O server. It is worth noting that the unicode option notifies how characters are actually put on the physical IO device, i.e. if the terminal per se 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 should 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, this request is used:


The getopts 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 is always an atom.

1.5  Multiple I/O Requests

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

{requests, Requests}

  • Requests is a list of valid io_request tuples for the protocol, they shall be executed in the order in which they appear in the list and the execution should continue until one of the requests result 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 valid results in the reply:

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

- depending on the actual requests in the list.

1.6  Optional I/O Requests

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

{get_geometry, Geometry}

  • Geometry is either the atom rows or the atom columns.

The I/O server should send the Reply as:

{ok, N}
{error, Error}

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

1.7  Unimplemented Request Types

If an I/O server encounters a request it does not recognize (i.e. the io_request tuple is in the expected format, but the actual Request is unknown), the I/O server should 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 backwards compatible.

1.8  An Annotated and Working Example I/O Server

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 enough, a process handling incoming requests, usually both I/O-requests and other IO device-specific requests (for i.e. positioning, closing etc.).

Our example I/O server stores characters in an ETS table, making up a fairly crude ram-file (it is probably not useful, but working).

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 might be using the io module to send requests). For each request the function request/2 is called and a reply is eventually sent using the reply/3 function.

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 IO 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 might be confusing to the reader.

Let us look at the reply function first...

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

Simple enough, it sends the io_reply tuple back to the client, providing the ReplyAs element received in the request along with the result of the request, as described above.

Now look at the different requests we need to handle. 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 tells us 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 the unicode:characters_to_list/2 function. The conversion function conveniently accepts the encoding types unicode or latin1, so we can use Encoding directly.

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

Let us handle the requests for retrieving data too:

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 might be too inefficient, but that of course depends on the frequency of the different requests. Before we start actually implementing the functions put_chars/2 and get_until/5, let us look into 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});

The get_geometry request has no meaning for this I/O server, so the reply will be {error, enotsup}. The only option we handle is the binary/list option, 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.

What is left is to handle backward compatibility and the file module (which uses the old requests until backward compatibility with pre-R13 nodes is no longer needed). Note that the I/O server will not work with a simple file:write/2 if these are not added:

request({put_chars,Chars}, State) ->
    request({put_chars,latin1,Chars}, State);
request({put_chars,M,F,As}, State) ->
    request({put_chars,latin1,M,F,As}, State);
request({get_chars,Prompt,N}, State) ->
    request({get_chars,latin1,Prompt,N}, State);
request({get_line,Prompt}, State) ->
    request({get_line,latin1,Prompt}, State);
request({get_until, Prompt,M,F,As}, State) ->
    request({get_until,latin1,Prompt,M,F,As}, State);

OK, what is left now is to return {error, request} if the request is not recognized:

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

Let us move further and actually 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 the function io_reply).

The getopts and setopts requests are also simple to handle, we just change or read our 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.

The getopts request should return a list of {Option, Value} tuples, which 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 our I/O server has been fairly generic (except for the rewind request handled in the main loop and the creation of an ETS table). Most I/O servers contain code similar to the one above.

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

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 just split the list in runs of a predefined size and put each run in the table at the current position (and forward). The functions split_data/3 and apply_update/2 are implemented below.

Now we want to read data from the table. The get_until/5 function reads data and applies the function until it says 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 the setopts request. By default, all OTP I/O servers send data back to the client as lists, but switching mode to binary might increase efficiency if the I/O server handles it in an appropriate way. The implementation of get_until is hard 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. This example does not optimize anything however. It is important though that the returned data is of the right type depending on the options set, so we convert the lists to binaries in the correct encoding if possible before returning. The function supplied in the get_until request tuple may, as its final result return anything, so only functions actually returning lists can get them converted to binaries. If the request contained the encoding tag unicode, the lists can contain all Unicode codepoints and the binaries should be in UTF-8, if the encoding tag was latin1, the client should only get characters in the range 0..255. The function check/2 takes care of not returning arbitrary Unicode codepoints in lists if the encoding was given as latin1. If the function did not return a list, the check cannot be performed and the result will be that of the supplied function untouched.

Now we are more or less done. We implement the utility functions below to actually manipulate the table:

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

The function check takes care of providing an error tuple if Unicode codepoints above 255 is to be returned if the client requested latin1.

The two functions until_newline/3 and until_enough/3 are helpers used together with the get_until/5 function 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 should be provided in get_until requests.

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

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 obviously 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 i.e. the io module or even the file module. It is as simple as that to implement a fully fledged I/O server in Erlang.