socket , a built-in Tcl command, opens a TCP network connection
This command opens a network socket and returns a channel identifier that may be used in future invocations of commands like read, puts and flush. At present only the TCP network protocol is supported; future releases may include support for additional protocols like UDP and IPX. The socket command may be used to open either the client or server side of a connection, depending on whether the -server switch is specified. See also the official documentation at http://www.purl.org/tcl/home/man/tcl8.5/TclCmd/socket.htm
Server sockets can only accept incoming connections when Tcl has an event loop running. All sockets use the system encoding by default; this is probably not what you want, but it is up to the protocol to define or negotiate something more suitable.
Something which is at first surprising is that the channel identifier returned by server -socket cannot be used to transfer any data (it can be fconfigured and closed, but that's it). Tcl inherited this peculiarity from the underlying C APIs for sockets: the datatype for open connections (to other processes, machines, etc.) over which one transfers data was reused to keep track of server ports to which clients may connect, so it was easiest to handle both of them as Tcl channels. The channel that the server side may use to transfer data is the one which is passed to the command callback. ----
Under windows, [socket] is implemented with WSAAsyncSelect up to 8.5, and through I/O completion ports starting with 8.6 [confirm]. Unix has always (well, since what I remember as 7.5) used select().
Port numbers are integers but depending upon your system configuration you may be able to use symbolic names for some services. For symbolic names, Tcl calls the getservbyname sockets API function which uses a services database file to convert between port numbers and service names. This function is quite limited on certain platforms (eg:Windows). slebetman: this is partially untrue (see comment below).
Some lists of protocols and their associated port numbers are:
On unix-type systems you very likely have an /etc/services file which is used by this function. It typically has the same syntax as the official IANA list referenced above slebetman: this is also not true, similar syntax but not the same. The typical /etc/services file has the syntax:
# comment service_name port_number/protocol alias
where comments can appear at the end of line and aliases can be used instead of the service_name when requesting for the port number (for example both "smtp" and "mail" refer to port 25). The IANA list on the other hand has the syntax:
service_name port_number/protocol description/comment # person who requested the port for the service <his@email> OR relevant documentation
In addition, for the IANA list, all text above the lines:
Port Assignments: Keyword Decimal Description References ------- ------- ----------- ----------
are to be considered as comments.
slebetman: On Windows, starting from Windows 2000 you have a C:\WINNT\SYSTEM32\DRIVERS\ETC\SERVICES or C:\WINDOWS\SYSTEM32\DRIVERS\ETC\SERVICES file serving the exact same function as the Unix counterpart and has exactly the same syntax. The currently shipping services file for Windows XP is dated 1999 so it doesn't have newer registered ports such as tcpmux and rje. Nonetheless this file is editable and can be easily updated by downloading the current 2006 IANA definitions from the net.
Stu - 2007/02/16 : The netinfo extension provides easy access to the contents of /etc/services as well as /etc/protocols and /etc/hosts.
How to use the -async option? Pasted from the Tcl'ers chat
proc verify s {
set msg [fconfigure $s -error]
if {$msg ne ""} {
# an error connecting
}
# set up to use $s
}
set s [socket -async $host $port]
fileevent $s writable [list verify $s]dgp: over the Tcl versions and platforms, I've found it helpful to test puts -nonewline $s "" and fconfigure $s -peername to be *really* sure you have a successful socket connection.
Cameron Laird wrote in comp.lang.tcl:
The most widely-used solution (for ensuring only a single instance of an app is running on one machine), I think, is to set up a do-nothing networking service. Have your application
socket -server ... 5678
when it first starts. If it's unable to do so because another service is already on 5678, take that as a previous instance of the application. Otherwise, continue. For an implementation, see How do I manage lock files in a cross platform manner in Tcl
Notice this approach is reasonably portable.
CL himself notes, somewhat later, that this is a bit "tricky", and not on the main stream of socket use; it's really socket's contribution to the theme of singleton applications.
Several challenges in effective network programming have to do with DNS pragmatics. DKF wrote, "numeric addresses are never resolved on socket creation on Unix, but they *are* on Windows (but I do not know if it is *really* doing it, as that depends on the Winsock implementation)."
As of 8.5, there's apparently no clean socket introspection. Socket handles do show up in "file channels sock*", though, so, as long as one is willing to rely on this slightly dirty match, the answer is computable.
DGP: See Tcl Feature Request 455867 .
GPS: from the Tcl Chatroom:
% proc accept {s addr port} {puts $s "Hi and bye!"; close $s}
% socket -server accept 8888Client tclsh:
% set s [socket localhost 8888]
sock3
% gets $sNow the client will block, because the server isn't running its event loop. So go to the server tclsh and type: update and you will see in the client: Hi and bye! (Wikified soon)
GPS: I would like to clarify a bit. fileevent interacts with the event loop allowing us to do other things (rather than just blocking) and have an event callback invoked when data is ready to be send/received on a socket channel.
kostix: Issues with remote disconnects and blocking sockets
There's a problem of knowing that the remote wants to disconnect since there's no possibility to register a kind of callback for getting some info about changes in the socket's state.
Here's what occur:
The amount of data that can be sent depends on whether the remote wants to continue to receive the data from our endpoint, and which method of buffering is set for the socket.
Most probably, the remote will send a packet with the RST flag set if it gets any not-yet-acknowledged packet from our endpoint w/o the FIN flag set. This will forcibly tear down the connection and any subsequent attempts to send the data to the remote will fail with EPIPE or ECONNRESET.
The data to the remote may pour, however, until the local TCP/IP stack finally gets that RST packet. At least 2 data packets can be sent to the remote before we get this RST packet: one with the ACK on the initial FIN, and one that the remote will consider as bad (and will send its RST).
So, from the standpoint of tracing the remote disconnects, the different buffering schemes should be used as described:
It would be cool to have some real means for getting the socket status, at least passive (through fconfigure or the like) but it seems current Tcl lacks this.
In my Tcl 8.4.9 on Linux, fconfigure $sock -error doesn't indicate that half-closed connection state. May be it should do so.
Update: at least on 8.5.2 under Linux fconfigure $sock -error seem to return "broken pipe" earlier than puts errors out. The error appears to be set after the first attempt to put data into unbuffered socket which was closed remotely. For some reason only the second attempt to send data to such socket will actually throw the "broken pipe" error.
JMN 2006-10-17 related, and worse IMO.. there's apparently no way to trigger a half close on a Tcl Socket. It seems to me that Tcl is simply incapable of talking to certain TCP daemons such as spammassassin's spamc/spamd 1.0 protocol. This daemon relies on the client closing its write half of the channel to trigger the daemon to send a response. Tcl only lets you close the whole socket - so you never get the response. Just sending an EOF doesn't work. Quite a serious limitation of Tcl sockets.. no?
AMG: Agreed. My suggestion is to create a command to unbundle a read/write channel into a pair of channels, one read-only and the other write-only. Then the halves can be separately closed. For symmetry, there could also be a command to bundle a read-only channel (e.g. stdin) and a write-only channel (e.g. stdout) into a read/write channel. This would make it possible to pass stdin/stdout to a procedure that accepts as an argument a single read/write channel (such as a socket).
Lars H 2008-05-06: For splitting channels, there is TIP 301 . There doesn't seem to be any implementation of that yet, however.
kostix 2008-09-25: TIP 301 was rejected and superseded by TIP 332 .
ferrieux 2008-09-26: not formally rejected, withdrawn ;-) To answer JMN's concern about missing half-close for socket: good news, I think I will finish #332 pretty soon because there already is machinery at the C level for a generic notion of half-closes. They are even already implemented for pipes. So we just need to implement the half-close proc for sockets (with shutdown()) and expose the whole at Tcl level. For the impatient though, there is a workaround: on unix do exec shut w >@ $sok where 'shut' is standalone binary of a few lines of C doing the shutdown() on its fd 1. On Windows use TWAPI or ffidl.
kostix 2009-09-06: TIP #332 "Half-Close for Bidirectional Channels" [L3 ] is implemented as part of Tcl 8.6.
Ro 2007-12-21: Why does the following line block for a few seconds before returning to the prompt when run from the tclkit console on windows? Is it resolving its own dns name? If so, is there a way around this? socket -server accept -myaddr 127.0.0.1 10001
UZ 2007-12-21: I had similar issues with socket clients and servers on the local host, and found that it is due to the antivirus sw (I am using NOD32). After indicating to the AV sw to skip the executable, connection was immediate. Hope it helps.
Ro 2007-12-21: Thank you, Umberto! I have not played with server sockets on windows in a while and this was unexpected. Your solution worked perfectly, I told the IMON component of NOD32 to skip this tclkit executable I was using and now it works fast and smooth. Many thanks!
KL 2008-05-02: I just wondered why under Windows XP (with SP2), multiple socket connection were not working like they do under Linux/WIN2000. After searching for some time, I found the solution (just for other people who have the same problem). There is a restriction of maximal half open sockets introduced in Windows XP SP2 to increase the security against the spreading of worms and virus. I think that's not the proper way to achieve security since that's not the root cause. For more information and a patch to increase the number of connection, please look here [L4 ] or google for "Event ID 4226".
An example script for IP-Scan process can be found at ipscan.
MHo 2015-03-02: At least on windows, if a connect could not be made, one have to wait several seconds (with a synchronous connect) before a timeout occurs. The solution is to use -async sockets, but that requires a running eventloop. So, a simple and fast socket test (waiting for 1 second or so) isn't possible for programs without an eventloop. The solution I posted years ago at the page Testing socketports is dangerous in some constructs because it can lead to nested eventloops... I don't know the underlying API, but if possible, a timeout option should be exposed to the scripting layer for synchronous sockets.
HaO 2015-04-01: Eventually see socket -async use cases - async without event loop.