/Doc/howto/sockets.rst

  1****************************
  2  Socket Programming HOWTO
  3****************************
  4
  5:Author: Gordon McMillan
  6
  7
  8.. topic:: Abstract
  9
 10   Sockets are used nearly everywhere, but are one of the most severely
 11   misunderstood technologies around. This is a 10,000 foot overview of sockets.
 12   It's not really a tutorial - you'll still have work to do in getting things
 13   operational. It doesn't cover the fine points (and there are a lot of them), but
 14   I hope it will give you enough background to begin using them decently.
 15
 16
 17Sockets
 18=======
 19
 20Sockets are used nearly everywhere, but are one of the most severely
 21misunderstood technologies around. This is a 10,000 foot overview of sockets.
 22It's not really a tutorial - you'll still have work to do in getting things
 23working. It doesn't cover the fine points (and there are a lot of them), but I
 24hope it will give you enough background to begin using them decently.
 25
 26I'm only going to talk about INET sockets, but they account for at least 99% of
 27the sockets in use. And I'll only talk about STREAM sockets - unless you really
 28know what you're doing (in which case this HOWTO isn't for you!), you'll get
 29better behavior and performance from a STREAM socket than anything else. I will
 30try to clear up the mystery of what a socket is, as well as some hints on how to
 31work with blocking and non-blocking sockets. But I'll start by talking about
 32blocking sockets. You'll need to know how they work before dealing with
 33non-blocking sockets.
 34
 35Part of the trouble with understanding these things is that "socket" can mean a
 36number of subtly different things, depending on context. So first, let's make a
 37distinction between a "client" socket - an endpoint of a conversation, and a
 38"server" socket, which is more like a switchboard operator. The client
 39application (your browser, for example) uses "client" sockets exclusively; the
 40web server it's talking to uses both "server" sockets and "client" sockets.
 41
 42
 43History
 44-------
 45
 46Of the various forms of IPC (*Inter Process Communication*), sockets are by far
 47the most popular.  On any given platform, there are likely to be other forms of
 48IPC that are faster, but for cross-platform communication, sockets are about the
 49only game in town.
 50
 51They were invented in Berkeley as part of the BSD flavor of Unix. They spread
 52like wildfire with the Internet. With good reason --- the combination of sockets
 53with INET makes talking to arbitrary machines around the world unbelievably easy
 54(at least compared to other schemes).
 55
 56
 57Creating a Socket
 58=================
 59
 60Roughly speaking, when you clicked on the link that brought you to this page,
 61your browser did something like the following::
 62
 63   #create an INET, STREAMing socket
 64   s = socket.socket(
 65       socket.AF_INET, socket.SOCK_STREAM)
 66   #now connect to the web server on port 80
 67   # - the normal http port
 68   s.connect(("www.mcmillan-inc.com", 80))
 69
 70When the ``connect`` completes, the socket ``s`` can now be used to send in a
 71request for the text of this page. The same socket will read the reply, and then
 72be destroyed. That's right - destroyed. Client sockets are normally only used
 73for one exchange (or a small set of sequential exchanges).
 74
 75What happens in the web server is a bit more complex. First, the web server
 76creates a "server socket". ::
 77
 78   #create an INET, STREAMing socket
 79   serversocket = socket.socket(
 80       socket.AF_INET, socket.SOCK_STREAM)
 81   #bind the socket to a public host,
 82   # and a well-known port
 83   serversocket.bind((socket.gethostname(), 80))
 84   #become a server socket
 85   serversocket.listen(5)
 86
 87A couple things to notice: we used ``socket.gethostname()`` so that the socket
 88would be visible to the outside world. If we had used ``s.bind(('', 80))`` or
 89``s.bind(('localhost', 80))`` or ``s.bind(('127.0.0.1', 80))`` we would still
 90have a "server" socket, but one that was only visible within the same machine.
 91
 92A second thing to note: low number ports are usually reserved for "well known"
 93services (HTTP, SNMP etc). If you're playing around, use a nice high number (4
 94digits).
 95
 96Finally, the argument to ``listen`` tells the socket library that we want it to
 97queue up as many as 5 connect requests (the normal max) before refusing outside
 98connections. If the rest of the code is written properly, that should be plenty.
 99
100OK, now we have a "server" socket, listening on port 80. Now we enter the
101mainloop of the web server::
102
103   while 1:
104       #accept connections from outside
105       (clientsocket, address) = serversocket.accept()
106       #now do something with the clientsocket
107       #in this case, we'll pretend this is a threaded server
108       ct = client_thread(clientsocket)
109       ct.run()
110
111There's actually 3 general ways in which this loop could work - dispatching a
112thread to handle ``clientsocket``, create a new process to handle
113``clientsocket``, or restructure this app to use non-blocking sockets, and
114mulitplex between our "server" socket and any active ``clientsocket``\ s using
115``select``. More about that later. The important thing to understand now is
116this: this is *all* a "server" socket does. It doesn't send any data. It doesn't
117receive any data. It just produces "client" sockets. Each ``clientsocket`` is
118created in response to some *other* "client" socket doing a ``connect()`` to the
119host and port we're bound to. As soon as we've created that ``clientsocket``, we
120go back to listening for more connections. The two "clients" are free to chat it
121up - they are using some dynamically allocated port which will be recycled when
122the conversation ends.
123
124
125IPC
126---
127
128If you need fast IPC between two processes on one machine, you should look into
129whatever form of shared memory the platform offers. A simple protocol based
130around shared memory and locks or semaphores is by far the fastest technique.
131
132If you do decide to use sockets, bind the "server" socket to ``'localhost'``. On
133most platforms, this will take a shortcut around a couple of layers of network
134code and be quite a bit faster.
135
136
137Using a Socket
138==============
139
140The first thing to note, is that the web browser's "client" socket and the web
141server's "client" socket are identical beasts. That is, this is a "peer to peer"
142conversation. Or to put it another way, *as the designer, you will have to
143decide what the rules of etiquette are for a conversation*. Normally, the
144``connect``\ ing socket starts the conversation, by sending in a request, or
145perhaps a signon. But that's a design decision - it's not a rule of sockets.
146
147Now there are two sets of verbs to use for communication. You can use ``send``
148and ``recv``, or you can transform your client socket into a file-like beast and
149use ``read`` and ``write``. The latter is the way Java presents their sockets.
150I'm not going to talk about it here, except to warn you that you need to use
151``flush`` on sockets. These are buffered "files", and a common mistake is to
152``write`` something, and then ``read`` for a reply. Without a ``flush`` in
153there, you may wait forever for the reply, because the request may still be in
154your output buffer.
155
156Now we come the major stumbling block of sockets - ``send`` and ``recv`` operate
157on the network buffers. They do not necessarily handle all the bytes you hand
158them (or expect from them), because their major focus is handling the network
159buffers. In general, they return when the associated network buffers have been
160filled (``send``) or emptied (``recv``). They then tell you how many bytes they
161handled. It is *your* responsibility to call them again until your message has
162been completely dealt with.
163
164When a ``recv`` returns 0 bytes, it means the other side has closed (or is in
165the process of closing) the connection.  You will not receive any more data on
166this connection. Ever.  You may be able to send data successfully; I'll talk
167about that some on the next page.
168
169A protocol like HTTP uses a socket for only one transfer. The client sends a
170request, the reads a reply.  That's it. The socket is discarded. This means that
171a client can detect the end of the reply by receiving 0 bytes.
172
173But if you plan to reuse your socket for further transfers, you need to realize
174that *there is no "EOT" (End of Transfer) on a socket.* I repeat: if a socket
175``send`` or ``recv`` returns after handling 0 bytes, the connection has been
176broken.  If the connection has *not* been broken, you may wait on a ``recv``
177forever, because the socket will *not* tell you that there's nothing more to
178read (for now).  Now if you think about that a bit, you'll come to realize a
179fundamental truth of sockets: *messages must either be fixed length* (yuck), *or
180be delimited* (shrug), *or indicate how long they are* (much better), *or end by
181shutting down the connection*. The choice is entirely yours, (but some ways are
182righter than others).
183
184Assuming you don't want to end the connection, the simplest solution is a fixed
185length message::
186
187   class mysocket:
188       '''demonstration class only
189         - coded for clarity, not efficiency
190       '''
191
192       def __init__(self, sock=None):
193           if sock is None:
194               self.sock = socket.socket(
195                   socket.AF_INET, socket.SOCK_STREAM)
196           else:
197               self.sock = sock
198
199       def connect(self, host, port):
200           self.sock.connect((host, port))
201
202       def mysend(self, msg):
203           totalsent = 0
204           while totalsent < MSGLEN:
205               sent = self.sock.send(msg[totalsent:])
206               if sent == 0:
207                   raise RuntimeError, \
208                       "socket connection broken"
209               totalsent = totalsent + sent
210
211       def myreceive(self):
212           msg = ''
213           while len(msg) < MSGLEN:
214               chunk = self.sock.recv(MSGLEN-len(msg))
215               if chunk == '':
216                   raise RuntimeError, \
217                       "socket connection broken"
218               msg = msg + chunk
219           return msg
220
221The sending code here is usable for almost any messaging scheme - in Python you
222send strings, and you can use ``len()`` to determine its length (even if it has
223embedded ``\0`` characters). It's mostly the receiving code that gets more
224complex. (And in C, it's not much worse, except you can't use ``strlen`` if the
225message has embedded ``\0``\ s.)
226
227The easiest enhancement is to make the first character of the message an
228indicator of message type, and have the type determine the length. Now you have
229two ``recv``\ s - the first to get (at least) that first character so you can
230look up the length, and the second in a loop to get the rest. If you decide to
231go the delimited route, you'll be receiving in some arbitrary chunk size, (4096
232or 8192 is frequently a good match for network buffer sizes), and scanning what
233you've received for a delimiter.
234
235One complication to be aware of: if your conversational protocol allows multiple
236messages to be sent back to back (without some kind of reply), and you pass
237``recv`` an arbitrary chunk size, you may end up reading the start of a
238following message. You'll need to put that aside and hold onto it, until it's
239needed.
240
241Prefixing the message with it's length (say, as 5 numeric characters) gets more
242complex, because (believe it or not), you may not get all 5 characters in one
243``recv``. In playing around, you'll get away with it; but in high network loads,
244your code will very quickly break unless you use two ``recv`` loops - the first
245to determine the length, the second to get the data part of the message. Nasty.
246This is also when you'll discover that ``send`` does not always manage to get
247rid of everything in one pass. And despite having read this, you will eventually
248get bit by it!
249
250In the interests of space, building your character, (and preserving my
251competitive position), these enhancements are left as an exercise for the
252reader. Lets move on to cleaning up.
253
254
255Binary Data
256-----------
257
258It is perfectly possible to send binary data over a socket. The major problem is
259that not all machines use the same formats for binary data. For example, a
260Motorola chip will represent a 16 bit integer with the value 1 as the two hex
261bytes 00 01. Intel and DEC, however, are byte-reversed - that same 1 is 01 00.
262Socket libraries have calls for converting 16 and 32 bit integers - ``ntohl,
263htonl, ntohs, htons`` where "n" means *network* and "h" means *host*, "s" means
264*short* and "l" means *long*. Where network order is host order, these do
265nothing, but where the machine is byte-reversed, these swap the bytes around
266appropriately.
267
268In these days of 32 bit machines, the ascii representation of binary data is
269frequently smaller than the binary representation. That's because a surprising
270amount of the time, all those longs have the value 0, or maybe 1. The string "0"
271would be two bytes, while binary is four. Of course, this doesn't fit well with
272fixed-length messages. Decisions, decisions.
273
274
275Disconnecting
276=============
277
278Strictly speaking, you're supposed to use ``shutdown`` on a socket before you
279``close`` it.  The ``shutdown`` is an advisory to the socket at the other end.
280Depending on the argument you pass it, it can mean "I'm not going to send
281anymore, but I'll still listen", or "I'm not listening, good riddance!".  Most
282socket libraries, however, are so used to programmers neglecting to use this
283piece of etiquette that normally a ``close`` is the same as ``shutdown();
284close()``.  So in most situations, an explicit ``shutdown`` is not needed.
285
286One way to use ``shutdown`` effectively is in an HTTP-like exchange. The client
287sends a request and then does a ``shutdown(1)``. This tells the server "This
288client is done sending, but can still receive."  The server can detect "EOF" by
289a receive of 0 bytes. It can assume it has the complete request.  The server
290sends a reply. If the ``send`` completes successfully then, indeed, the client
291was still receiving.
292
293Python takes the automatic shutdown a step further, and says that when a socket
294is garbage collected, it will automatically do a ``close`` if it's needed. But
295relying on this is a very bad habit. If your socket just disappears without
296doing a ``close``, the socket at the other end may hang indefinitely, thinking
297you're just being slow. *Please* ``close`` your sockets when you're done.
298
299
300When Sockets Die
301----------------
302
303Probably the worst thing about using blocking sockets is what happens when the
304other side comes down hard (without doing a ``close``). Your socket is likely to
305hang. SOCKSTREAM is a reliable protocol, and it will wait a long, long time
306before giving up on a connection. If you're using threads, the entire thread is
307essentially dead. There's not much you can do about it. As long as you aren't
308doing something dumb, like holding a lock while doing a blocking read, the
309thread isn't really consuming much in the way of resources. Do *not* try to kill
310the thread - part of the reason that threads are more efficient than processes
311is that they avoid the overhead associated with the automatic recycling of
312resources. In other words, if you do manage to kill the thread, your whole
313process is likely to be screwed up.
314
315
316Non-blocking Sockets
317====================
318
319If you've understood the preceeding, you already know most of what you need to
320know about the mechanics of using sockets. You'll still use the same calls, in
321much the same ways. It's just that, if you do it right, your app will be almost
322inside-out.
323
324In Python, you use ``socket.setblocking(0)`` to make it non-blocking. In C, it's
325more complex, (for one thing, you'll need to choose between the BSD flavor
326``O_NONBLOCK`` and the almost indistinguishable Posix flavor ``O_NDELAY``, which
327is completely different from ``TCP_NODELAY``), but it's the exact same idea. You
328do this after creating the socket, but before using it. (Actually, if you're
329nuts, you can switch back and forth.)
330
331The major mechanical difference is that ``send``, ``recv``, ``connect`` and
332``accept`` can return without having done anything. You have (of course) a
333number of choices. You can check return code and error codes and generally drive
334yourself crazy. If you don't believe me, try it sometime. Your app will grow
335large, buggy and suck CPU. So let's skip the brain-dead solutions and do it
336right.
337
338Use ``select``.
339
340In C, coding ``select`` is fairly complex. In Python, it's a piece of cake, but
341it's close enough to the C version that if you understand ``select`` in Python,
342you'll have little trouble with it in C. ::
343
344   ready_to_read, ready_to_write, in_error = \
345                  select.select(
346                     potential_readers,
347                     potential_writers,
348                     potential_errs,
349                     timeout)
350
351You pass ``select`` three lists: the first contains all sockets that you might
352want to try reading; the second all the sockets you might want to try writing
353to, and the last (normally left empty) those that you want to check for errors.
354You should note that a socket can go into more than one list. The ``select``
355call is blocking, but you can give it a timeout. This is generally a sensible
356thing to do - give it a nice long timeout (say a minute) unless you have good
357reason to do otherwise.
358
359In return, you will get three lists. They have the sockets that are actually
360readable, writable and in error. Each of these lists is a subset (possibly
361empty) of the corresponding list you passed in. And if you put a socket in more
362than one input list, it will only be (at most) in one output list.
363
364If a socket is in the output readable list, you can be
365as-close-to-certain-as-we-ever-get-in-this-business that a ``recv`` on that
366socket will return *something*. Same idea for the writable list. You'll be able
367to send *something*. Maybe not all you want to, but *something* is better than
368nothing.  (Actually, any reasonably healthy socket will return as writable - it
369just means outbound network buffer space is available.)
370
371If you have a "server" socket, put it in the potential_readers list. If it comes
372out in the readable list, your ``accept`` will (almost certainly) work. If you
373have created a new socket to ``connect`` to someone else, put it in the
374potential_writers list. If it shows up in the writable list, you have a decent
375chance that it has connected.
376
377One very nasty problem with ``select``: if somewhere in those input lists of
378sockets is one which has died a nasty death, the ``select`` will fail. You then
379need to loop through every single damn socket in all those lists and do a
380``select([sock],[],[],0)`` until you find the bad one. That timeout of 0 means
381it won't take long, but it's ugly.
382
383Actually, ``select`` can be handy even with blocking sockets. It's one way of
384determining whether you will block - the socket returns as readable when there's
385something in the buffers.  However, this still doesn't help with the problem of
386determining whether the other end is done, or just busy with something else.
387
388**Portability alert**: On Unix, ``select`` works both with the sockets and
389files. Don't try this on Windows. On Windows, ``select`` works with sockets
390only. Also note that in C, many of the more advanced socket options are done
391differently on Windows. In fact, on Windows I usually use threads (which work
392very, very well) with my sockets. Face it, if you want any kind of performance,
393your code will look very different on Windows than on Unix.
394
395
396Performance
397-----------
398
399There's no question that the fastest sockets code uses non-blocking sockets and
400select to multiplex them. You can put together something that will saturate a
401LAN connection without putting any strain on the CPU. The trouble is that an app
402written this way can't do much of anything else - it needs to be ready to
403shuffle bytes around at all times.
404
405Assuming that your app is actually supposed to do something more than that,
406threading is the optimal solution, (and using non-blocking sockets will be
407faster than using blocking sockets). Unfortunately, threading support in Unixes
408varies both in API and quality. So the normal Unix solution is to fork a
409subprocess to deal with each connection. The overhead for this is significant
410(and don't do this on Windows - the overhead of process creation is enormous
411there). It also means that unless each subprocess is completely independent,
412you'll need to use another form of IPC, say a pipe, or shared memory and
413semaphores, to communicate between the parent and child processes.
414
415Finally, remember that even though blocking sockets are somewhat slower than
416non-blocking, in many cases they are the "right" solution. After all, if your
417app is driven by the data it receives over a socket, there's not much sense in
418complicating the logic just so your app can wait on ``select`` instead of
419``recv``.
420