Kernel

Reference Manual

Version 8.0.1

Table of Contents

gen_tcp

Module

gen_tcp

Module Summary

Interface to TCP/IP sockets.

Description

This module provides functions for communicating with sockets using the TCP/IP protocol.

The following code fragment is a simple example of a client connecting to a server at port 5678, transferring a binary, and closing the connection:

client() ->
    SomeHostInNet = "localhost", % to make it runnable on one machine
    {ok, Sock} = gen_tcp:connect(SomeHostInNet, 5678, 
                                 [binary, {packet, 0}]),
    ok = gen_tcp:send(Sock, "Some Data"),
    ok = gen_tcp:close(Sock).

At the other end, a server is listening on port 5678, accepts the connection, and receives the binary:

server() ->
    {ok, LSock} = gen_tcp:listen(5678, [binary, {packet, 0}, 
                                        {active, false}]),
    {ok, Sock} = gen_tcp:accept(LSock),
    {ok, Bin} = do_recv(Sock, []),
    ok = gen_tcp:close(Sock),
    ok = gen_tcp:close(LSock),
    Bin.

do_recv(Sock, Bs) ->
    case gen_tcp:recv(Sock, 0) of
        {ok, B} ->
            do_recv(Sock, [Bs, B]);
        {error, closed} ->
            {ok, list_to_binary(Bs)}
    end.

For more examples, see section Examples.

Note

Functions that create sockets can take an optional option; {inet_backend, Backend} that, if specified, has to be the first option. This selects the implementation backend towards the platform's socket API.

This is a temporary option that will be ignored in a future release.

The default is Backend = inet that selects the traditional inet_drv.c driver. The other choice is Backend = socket that selects the new socket module and its NIF implementation.

The system default can be changed when the node is started with the application kernel's configuration variable inet_backend.

For gen_tcp with inet_backend = socket we have tried to be as "compatible" as possible which has sometimes been impossible. Here is a list of cases when the behaviour of inet-backend inet (default) and socket are different:

  • Non-blocking send

    If a user calling gen_tcp:send/2 with inet_backend = inet, tries to send more data than there is room for in the OS buffers, the "rest data" is buffered by the inet driver (and later sent in the background). The effect for the user is that the call is non-blocking.

    This is not the effect when inet_backend = socket, since there is no buffering. Instead the user hangs either until all data has been sent or the send_timeout timeout has been reached.

  • Remote close detected by background send.

    An background send will detect a 'remote close' and (the inet driver will) mark the socket as 'closed'. No other action is taken. If the socket has active set to false (passive) at this point and no one is reading, this will not be noticed. But as soon as the socket is "activated" (active set to not false, send/2 is called or recv/2,3 is called), an error message will be sent to the caller or (socket) owner: {tcp_error, Socket, econnreset}. Any data in the OS receive buffers will be lost!

    This behaviour is not replicated by the socket implementation. A send operation will detect a remote close and immediately return this to the caller, but do nothing else. A reader will therefor be able to extract any data from the OS buffers. If the socket is set to active to not false, the data will be received as expected ({tcp, ...} and then a closed message ({tcp_closed, ...} will be received (not an error).

  • The option show_econnreset basically do not work as described when used with inet_backend = socket. The "issue" is that a remote close (as described above) do allow a reader to extract what is in the read buffers before a close is "delivered".

  • The option nodelay is a TCP specific option that is not compatible with domain = local.

    When using inet_backend = socket, trying to create a socket (via listen or connect) with domain = local (for example with option {ifaddr, {local,"/tmp/test"}}) will fail with {error, enotsup}.

    This does not actually work for inet_backend = inet either, but in that case the error is simply ignored, which is a bad idea. We have choosen to not ignore this error for inet_backend = socket.

  • Async shutdown write

    Calling gen_tcp:shutdown(Socket, write | read_write) on a socket created with inet_backend = socket will take immediate effect, unlike for a socket created with inet_backend = inet.

    See async shutdown write for more info.

Data Types

option() =
    {active, true | false | once | -32768..32767} |
    {buffer, integer() >= 0} |
    {delay_send, boolean()} |
    {deliver, port | term} |
    {dontroute, boolean()} |
    {exit_on_close, boolean()} |
    {header, integer() >= 0} |
    {high_msgq_watermark, integer() >= 1} |
    {high_watermark, integer() >= 0} |
    {keepalive, boolean()} |
    {linger, {boolean(), integer() >= 0}} |
    {low_msgq_watermark, integer() >= 1} |
    {low_watermark, integer() >= 0} |
    {mode, list | binary} |
    list | binary |
    {nodelay, boolean()} |
    {packet,
     0 | 1 | 2 | 4 | raw | sunrm | asn1 | cdr | fcgi | line |
     tpkt | http | httph | http_bin | httph_bin} |
    {packet_size, integer() >= 0} |
    {priority, integer() >= 0} |
    {raw,
     Protocol :: integer() >= 0,
     OptionNum :: integer() >= 0,
     ValueBin :: binary()} |
    {recbuf, integer() >= 0} |
    {reuseaddr, boolean()} |
    {send_timeout, integer() >= 0 | infinity} |
    {send_timeout_close, boolean()} |
    {show_econnreset, boolean()} |
    {sndbuf, integer() >= 0} |
    {tos, integer() >= 0} |
    {tclass, integer() >= 0} |
    {ttl, integer() >= 0} |
    {recvtos, boolean()} |
    {recvtclass, boolean()} |
    {recvttl, boolean()} |
    {ipv6_v6only, boolean()}

If the platform implements the IPv4 option IP_PKTOPTIONS, or the IPv6 option IPV6_PKTOPTIONS or IPV6_2292PKTOPTIONS for the socket this value is returned from inet:getopts/2 when called with the option name pktoptions.

Note

This option appears to be VERY Linux specific, and its existence in future Linux kernel versions is also worrying since the option is part of RFC 2292 which is since long (2003) obsoleted by RFC 3542 that explicitly removes this possibility to get packet information from a stream socket. For comparision: it has existed in FreeBSD but is now removed, at least since FreeBSD 10.

option_name() =
    active | buffer | delay_send | deliver | dontroute |
    exit_on_close | header | high_msgq_watermark |
    high_watermark | keepalive | linger | low_msgq_watermark |
    low_watermark | mode | nodelay | packet | packet_size |
    priority |
    {raw,
     Protocol :: integer() >= 0,
     OptionNum :: integer() >= 0,
     ValueSpec ::
         (ValueSize :: integer() >= 0) | (ValueBin :: binary())} |
    recbuf | reuseaddr | send_timeout | send_timeout_close |
    show_econnreset | sndbuf | tos | tclass | ttl | recvtos |
    recvtclass | recvttl | pktoptions | ipv6_v6only

connect_option() =
    {ip, inet:socket_address()} |
    {fd, Fd :: integer() >= 0} |
    {ifaddr, inet:socket_address()} |
    inet:address_family() |
    {port, inet:port_number()} |
    {tcp_module, module()} |
    {netns, file:filename_all()} |
    {bind_to_device, binary()} |
    option()

listen_option() =
    {ip, inet:socket_address()} |
    {fd, Fd :: integer() >= 0} |
    {ifaddr, inet:socket_address()} |
    inet:address_family() |
    {port, inet:port_number()} |
    {backlog, B :: integer() >= 0} |
    {tcp_module, module()} |
    {netns, file:filename_all()} |
    {bind_to_device, binary()} |
    option()

As returned by accept/1,2 and connect/3,4.

accept(ListenSocket) -> {ok, Socket} | {error, Reason}
accept(ListenSocket, Timeout) -> {ok, Socket} | {error, Reason}

Types

ListenSocket = socket()
Returned by listen/2.
Timeout = timeout()
Socket = socket()
Reason = closed | timeout | system_limit | inet:posix()

Accepts an incoming connection request on a listening socket. Socket must be a socket returned from listen/2. Timeout specifies a time-out value in milliseconds. Defaults to infinity.

Returns:

  • {ok, Socket} if a connection is established

  • {error, closed} if ListenSocket is closed

  • {error, timeout} if no connection is established within the specified time

  • {error, system_limit} if all available ports in the Erlang emulator are in use

  • A POSIX error value if something else goes wrong, see inet(3) for possible error values

Packets can be sent to the returned socket Socket using send/2. Packets sent from the peer are delivered as messages (unless {active, false} is specified in the option list for the listening socket, in which case packets are retrieved by calling recv/2):

{tcp, Socket, Data}
Note

The accept call does not have to be issued from the socket owner process. Using version 5.5.3 and higher of the emulator, multiple simultaneous accept calls can be issued from different processes, which allows for a pool of acceptor processes handling incoming connections.

close(Socket) -> ok

Types

Socket = socket()

Closes a TCP socket.

Note that in most implementations of TCP, doing a close does not guarantee that any data sent is delivered to the recipient before the close is detected at the remote side. If you want to guarantee delivery of the data to the recipient there are two common ways to achieve this.

  1. Use gen_tcp:shutdown(Sock, write) to signal that no more data is to be sent and wait for the read side of the socket to be closed.

  2. Use the socket option {packet, N} (or something similar) to make it possible for the receiver to close the connection when it knowns it has received all the data.

connect(Address, Port, Options) -> {ok, Socket} | {error, Reason}
connect(Address, Port, Options, Timeout) ->
           {ok, Socket} | {error, Reason}

Types

Timeout = timeout()
Socket = socket()
Reason = timeout | inet:posix()

Connects to a server on TCP port Port on the host with IP address Address. Argument Address can be a hostname or an IP address.

The following options are available:

{ip, Address}

If the host has many network interfaces, this option specifies which one to use.

{ifaddr, Address}

Same as {ip, Address}. If the host has many network interfaces, this option specifies which one to use.

{fd, integer() >= 0}

If a socket has somehow been connected without using gen_tcp, use this option to pass the file descriptor for it. If {ip, Address} and/or {port, port_number()} is combined with this option, the fd is bound to the specified interface and port before connecting. If these options are not specified, it is assumed that the fd is already bound appropriately.

inet

Sets up the socket for IPv4.

inet6

Sets up the socket for IPv6.

local

Sets up a Unix Domain Socket. See inet:local_address()

{port, Port}

Specifies which local port number to use.

{tcp_module, module()}

Overrides which callback module is used. Defaults to inet_tcp for IPv4 and inet6_tcp for IPv6.

Opt

See inet:setopts/2.

Packets can be sent to the returned socket Socket using send/2. Packets sent from the peer are delivered as messages:

{tcp, Socket, Data}

If the socket is in {active, N} mode (see inet:setopts/2 for details) and its message counter drops to 0, the following message is delivered to indicate that the socket has transitioned to passive ({active, false}) mode:

{tcp_passive, Socket}

If the socket is closed, the following message is delivered:

{tcp_closed, Socket}

If an error occurs on the socket, the following message is delivered (unless {active, false} is specified in the option list for the socket, in which case packets are retrieved by calling recv/2):

{tcp_error, Socket, Reason}

The optional Timeout parameter specifies a time-out in milliseconds. Defaults to infinity.

Note

Keep in mind that if the underlying OS connect() call returns a timeout, gen_tcp:connect will also return a timeout (i.e. {error, etimedout}), even if a larger Timeout was specified.

Note

The default values for options specified to connect can be affected by the Kernel configuration parameter inet_default_connect_options. For details, see inet(3).

controlling_process(Socket, Pid) -> ok | {error, Reason}

Types

Socket = socket()
Pid = pid()
Reason = closed | not_owner | badarg | inet:posix()

Assigns a new controlling process Pid to Socket. The controlling process is the process that receives messages from the socket. If called by any other process than the current controlling process, {error, not_owner} is returned. If the process identified by Pid is not an existing local pid, {error, badarg} is returned. {error, badarg} may also be returned in some cases when Socket is closed during the execution of this function.

If the socket is set in active mode, this function will transfer any messages in the mailbox of the caller to the new controlling process. If any other process is interacting with the socket while the transfer is happening, the transfer may not work correctly and messages may remain in the caller's mailbox. For instance changing the sockets active mode before the transfer is complete may cause this.

listen(Port, Options) -> {ok, ListenSocket} | {error, Reason}

Types

ListenSocket = socket()
Reason = system_limit | inet:posix()

Sets up a socket to listen on port Port on the local host.

If Port == 0, the underlying OS assigns an available port number, use inet:port/1 to retrieve it.

The following options are available:

list

Received Packet is delivered as a list.

binary

Received Packet is delivered as a binary.

{backlog, B}

B is an integer >= 0. The backlog value defines the maximum length that the queue of pending connections can grow to. Defaults to 5.

{ip, Address}

If the host has many network interfaces, this option specifies which one to listen on.

{port, Port}

Specifies which local port number to use.

{fd, Fd}

If a socket has somehow been connected without using gen_tcp, use this option to pass the file descriptor for it.

{ifaddr, Address}

Same as {ip, Address}. If the host has many network interfaces, this option specifies which one to use.

inet6

Sets up the socket for IPv6.

inet

Sets up the socket for IPv4.

{tcp_module, module()}

Overrides which callback module is used. Defaults to inet_tcp for IPv4 and inet6_tcp for IPv6.

Opt

See inet:setopts/2.

The returned socket ListenSocket should be used in calls to accept/1,2 to accept incoming connection requests.

Note

The default values for options specified to listen can be affected by the Kernel configuration parameter inet_default_listen_options. For details, see inet(3).

recv(Socket, Length) -> {ok, Packet} | {error, Reason}
recv(Socket, Length, Timeout) -> {ok, Packet} | {error, Reason}

Types

Socket = socket()
Length = integer() >= 0
Timeout = timeout()
Packet = string() | binary() | HttpPacket
Reason = closed | timeout | inet:posix()
HttpPacket = term()
See the description of HttpPacket in erlang:decode_packet/3 in ERTS.

Receives a packet from a socket in passive mode. A closed socket is indicated by return value {error, closed}.

Argument Length is only meaningful when the socket is in raw mode and denotes the number of bytes to read. If Length is 0, all available bytes are returned. If Length > 0, exactly Length bytes are returned, or an error; possibly discarding less than Length bytes of data when the socket is closed from the other side.

The optional Timeout parameter specifies a time-out in milliseconds. Defaults to infinity.

send(Socket, Packet) -> ok | {error, Reason}

Types

Socket = socket()
Packet = iodata()
Reason = closed | {timeout, RestData} | inet:posix()
RestData = binary()

Sends a packet on a socket.

There is no send call with a time-out option, use socket option send_timeout if time-outs are desired. See section Examples.

The return value {error, {timeout, RestData}} can only be returned when inet_backend = socket.

Note

Non-blocking send.

If the user tries to send more data than there is room for in the OS send buffers, the 'rest data' is put into (inet driver) internal buffers and later sent in the background. The function immediately returns ok (not informing the caller that not all of the data was actually sent). Any issue while sending the 'rest data' is maybe returned later.

When using inet_backend = socket, the behaviour is different. There is no buffering done (like the inet-driver does), instead the caller will "hang" until all of the data has been sent or send timeout (as specified by the send_timeout option) expires (the function can hang even when using 'inet' backend if the internal buffers are full).

If this happens when using packet =/= raw, we have a partial package written. A new package therefor must not be written at this point, as there is no way for the peer to distinguish this from the data portion of the current package. Instead, set package to raw, send the rest data (as raw data) and then set package to the wanted package type again.

shutdown(Socket, How) -> ok | {error, Reason}

Types

Socket = socket()
How = read | write | read_write
Reason = inet:posix()

Closes a socket in one or two directions.

How == write means closing the socket for writing, reading from it is still possible.

If How == read or there is no outgoing data buffered in the Socket port, the socket is shut down immediately and any error encountered is returned in Reason.

If there is data buffered in the socket port, the attempt to shutdown the socket is postponed until that data is written to the kernel socket send buffer. If any errors are encountered, the socket is closed and {error, closed} is returned on the next recv/2 or send/2.

Option {exit_on_close, false} is useful if the peer has done a shutdown on the write side.

Note

Async shutdown write (write or read_write).

If the shutdown attempt is made while the inet-driver is sending buffered data in the background, the shutdown is postponed until all buffered data has been sent. The function immediately returns ok and the caller is not informed (that the shutdown has not yet been performed).

When using inet_backend = socket, the behaviour is different. A shutdown with How == write | read_write, the operation will take immediate effect (unlike the inet-driver, which basically saves the operation for later).

Examples

The following example illustrates use of option {active,once} and multiple accepts by implementing a server as a number of worker processes doing accept on a single listening socket. Function start/2 takes the number of worker processes and the port number on which to listen for incoming connections. If LPort is specified as 0, an ephemeral port number is used, which is why the start function returns the actual port number allocated:

start(Num,LPort) ->
    case gen_tcp:listen(LPort,[{active, false},{packet,2}]) of
        {ok, ListenSock} ->
            start_servers(Num,ListenSock),
            {ok, Port} = inet:port(ListenSock),
            Port;
        {error,Reason} ->
            {error,Reason}
    end.

start_servers(0,_) ->
    ok;
start_servers(Num,LS) ->
    spawn(?MODULE,server,[LS]),
    start_servers(Num-1,LS).

server(LS) ->
    case gen_tcp:accept(LS) of
        {ok,S} ->
            loop(S),
            server(LS);
        Other ->
            io:format("accept returned ~w - goodbye!~n",[Other]),
            ok
    end.

loop(S) ->
    inet:setopts(S,[{active,once}]),
    receive
        {tcp,S,Data} ->
            Answer = process(Data), % Not implemented in this example
            gen_tcp:send(S,Answer),
            loop(S);
        {tcp_closed,S} ->
            io:format("Socket ~w closed [~w]~n",[S,self()]),
            ok
    end.

Example of a simple client:

client(PortNo,Message) ->
    {ok,Sock} = gen_tcp:connect("localhost",PortNo,[{active,false},
                                                    {packet,2}]),
    gen_tcp:send(Sock,Message),
    A = gen_tcp:recv(Sock,0),
    gen_tcp:close(Sock),
    A.

The send call does not accept a time-out option because time-outs on send is handled through socket option send_timeout. The behavior of a send operation with no receiver is mainly defined by the underlying TCP stack and the network infrastructure. To write code that handles a hanging receiver that can eventually cause the sender to hang on a send do like the following.

Consider a process that receives data from a client process to be forwarded to a server on the network. The process is connected to the server through TCP/IP and does not get any acknowledge for each message it sends, but has to rely on the send time-out option to detect that the other end is unresponsive. Option send_timeout can be used when connecting:

...
{ok,Sock} = gen_tcp:connect(HostAddress, Port,
                            [{active,false},
                             {send_timeout, 5000},
                             {packet,2}]),
                loop(Sock), % See below
...

In the loop where requests are handled, send time-outs can now be detected:

loop(Sock) ->
    receive
        {Client, send_data, Binary} ->
            case gen_tcp:send(Sock,[Binary]) of
                {error, timeout} ->
                    io:format("Send timeout, closing!~n",
                              []),
                    handle_send_timeout(), % Not implemented here
                    Client ! {self(),{error_sending, timeout}},
                    %% Usually, it's a good idea to give up in case of a 
                    %% send timeout, as you never know how much actually 
                    %% reached the server, maybe only a packet header?!
                    gen_tcp:close(Sock);
                {error, OtherSendError} ->
                    io:format("Some other error on socket (~p), closing",
                              [OtherSendError]),
                    Client ! {self(),{error_sending, OtherSendError}},
                    gen_tcp:close(Sock);
                ok ->
                    Client ! {self(), data_sent},
                    loop(Sock)
            end
    end.

Usually it suffices to detect time-outs on receive, as most protocols include some sort of acknowledgment from the server, but if the protocol is strictly one way, option send_timeout comes in handy.