Try to make test_wsgiref less fragile against environment changes by other tests

[python.git] / Doc / howto / sockets.rst

****************************
  Socket Programming HOWTO
****************************

:Author: Gordon McMillan


.. topic:: Abstract

   Sockets are used nearly everywhere, but are one of the most severely
   misunderstood technologies around. This is a 10,000 foot overview of sockets.
   It's not really a tutorial - you'll still have work to do in getting things
   operational. It doesn't cover the fine points (and there are a lot of them), but
   I hope it will give you enough background to begin using them decently.


Sockets
=======

Sockets are used nearly everywhere, but are one of the most severely
misunderstood technologies around. This is a 10,000 foot overview of sockets.
It's not really a tutorial - you'll still have work to do in getting things
working. It doesn't cover the fine points (and there are a lot of them), but I
hope it will give you enough background to begin using them decently.

I'm only going to talk about INET sockets, but they account for at least 99% of
the sockets in use. And I'll only talk about STREAM sockets - unless you really
know what you're doing (in which case this HOWTO isn't for you!), you'll get
better behavior and performance from a STREAM socket than anything else. I will
try to clear up the mystery of what a socket is, as well as some hints on how to
work with blocking and non-blocking sockets. But I'll start by talking about
blocking sockets. You'll need to know how they work before dealing with
non-blocking sockets.

Part of the trouble with understanding these things is that "socket" can mean a
number of subtly different things, depending on context. So first, let's make a
distinction between a "client" socket - an endpoint of a conversation, and a
"server" socket, which is more like a switchboard operator. The client
application (your browser, for example) uses "client" sockets exclusively; the
web server it's talking to uses both "server" sockets and "client" sockets.


History
-------

Of the various forms of IPC (*Inter Process Communication*), sockets are by far
the most popular.  On any given platform, there are likely to be other forms of
IPC that are faster, but for cross-platform communication, sockets are about the
only game in town.

They were invented in Berkeley as part of the BSD flavor of Unix. They spread
like wildfire with the Internet. With good reason --- the combination of sockets
with INET makes talking to arbitrary machines around the world unbelievably easy
(at least compared to other schemes).


Creating a Socket
=================

Roughly speaking, when you clicked on the link that brought you to this page,
your browser did something like the following::

   #create an INET, STREAMing socket
   s = socket.socket(
       socket.AF_INET, socket.SOCK_STREAM)
   #now connect to the web server on port 80
   # - the normal http port
   s.connect(("www.mcmillan-inc.com", 80))

When the ``connect`` completes, the socket ``s`` can now be used to send in a
request for the text of this page. The same socket will read the reply, and then
be destroyed. That's right - destroyed. Client sockets are normally only used
for one exchange (or a small set of sequential exchanges).

What happens in the web server is a bit more complex. First, the web server
creates a "server socket". ::

   #create an INET, STREAMing socket
   serversocket = socket.socket(
       socket.AF_INET, socket.SOCK_STREAM)
   #bind the socket to a public host,
   # and a well-known port
   serversocket.bind((socket.gethostname(), 80))
   #become a server socket
   serversocket.listen(5)

A couple things to notice: we used ``socket.gethostname()`` so that the socket
would be visible to the outside world. If we had used ``s.bind(('', 80))`` or
``s.bind(('localhost', 80))`` or ``s.bind(('127.0.0.1', 80))`` we would still
have a "server" socket, but one that was only visible within the same machine.

A second thing to note: low number ports are usually reserved for "well known"
services (HTTP, SNMP etc). If you're playing around, use a nice high number (4
digits).

Finally, the argument to ``listen`` tells the socket library that we want it to
queue up as many as 5 connect requests (the normal max) before refusing outside
connections. If the rest of the code is written properly, that should be plenty.

OK, now we have a "server" socket, listening on port 80. Now we enter the
mainloop of the web server::

   while 1:
       #accept connections from outside
       (clientsocket, address) = serversocket.accept()
       #now do something with the clientsocket
       #in this case, we'll pretend this is a threaded server
       ct = client_thread(clientsocket)
       ct.run()

There's actually 3 general ways in which this loop could work - dispatching a
thread to handle ``clientsocket``, create a new process to handle
``clientsocket``, or restructure this app to use non-blocking sockets, and
mulitplex between our "server" socket and any active ``clientsocket``\ s using
``select``. More about that later. The important thing to understand now is
this: this is *all* a "server" socket does. It doesn't send any data. It doesn't
receive any data. It just produces "client" sockets. Each ``clientsocket`` is
created in response to some *other* "client" socket doing a ``connect()`` to the
host and port we're bound to. As soon as we've created that ``clientsocket``, we
go back to listening for more connections. The two "clients" are free to chat it
up - they are using some dynamically allocated port which will be recycled when
the conversation ends.


IPC
---

If you need fast IPC between two processes on one machine, you should look into
whatever form of shared memory the platform offers. A simple protocol based
around shared memory and locks or semaphores is by far the fastest technique.

If you do decide to use sockets, bind the "server" socket to ``'localhost'``. On
most platforms, this will take a shortcut around a couple of layers of network
code and be quite a bit faster.


Using a Socket
==============

The first thing to note, is that the web browser's "client" socket and the web
server's "client" socket are identical beasts. That is, this is a "peer to peer"
conversation. Or to put it another way, *as the designer, you will have to
decide what the rules of etiquette are for a conversation*. Normally, the
``connect``\ ing socket starts the conversation, by sending in a request, or
perhaps a signon. But that's a design decision - it's not a rule of sockets.

Now there are two sets of verbs to use for communication. You can use ``send``
and ``recv``, or you can transform your client socket into a file-like beast and
use ``read`` and ``write``. The latter is the way Java presents their sockets.
I'm not going to talk about it here, except to warn you that you need to use
``flush`` on sockets. These are buffered "files", and a common mistake is to
``write`` something, and then ``read`` for a reply. Without a ``flush`` in
there, you may wait forever for the reply, because the request may still be in
your output buffer.

Now we come the major stumbling block of sockets - ``send`` and ``recv`` operate
on the network buffers. They do not necessarily handle all the bytes you hand
them (or expect from them), because their major focus is handling the network
buffers. In general, they return when the associated network buffers have been
filled (``send``) or emptied (``recv``). They then tell you how many bytes they
handled. It is *your* responsibility to call them again until your message has
been completely dealt with.

When a ``recv`` returns 0 bytes, it means the other side has closed (or is in
the process of closing) the connection.  You will not receive any more data on
this connection. Ever.  You may be able to send data successfully; I'll talk
about that some on the next page.

A protocol like HTTP uses a socket for only one transfer. The client sends a
request, the reads a reply.  That's it. The socket is discarded. This means that
a client can detect the end of the reply by receiving 0 bytes.

But if you plan to reuse your socket for further transfers, you need to realize
that *there is no "EOT" (End of Transfer) on a socket.* I repeat: if a socket
``send`` or ``recv`` returns after handling 0 bytes, the connection has been
broken.  If the connection has *not* been broken, you may wait on a ``recv``
forever, because the socket will *not* tell you that there's nothing more to
read (for now).  Now if you think about that a bit, you'll come to realize a
fundamental truth of sockets: *messages must either be fixed length* (yuck), *or
be delimited* (shrug), *or indicate how long they are* (much better), *or end by
shutting down the connection*. The choice is entirely yours, (but some ways are
righter than others).

Assuming you don't want to end the connection, the simplest solution is a fixed
length message::

   class mysocket:
       '''demonstration class only
         - coded for clarity, not efficiency
       '''

       def __init__(self, sock=None):
           if sock is None:
               self.sock = socket.socket(
                   socket.AF_INET, socket.SOCK_STREAM)
           else:
               self.sock = sock

       def connect(self, host, port):
           self.sock.connect((host, port))

       def mysend(self, msg):
           totalsent = 0
           while totalsent < MSGLEN:
               sent = self.sock.send(msg[totalsent:])
               if sent == 0:
                   raise RuntimeError("socket connection broken")
               totalsent = totalsent + sent

       def myreceive(self):
           msg = ''
           while len(msg) < MSGLEN:
               chunk = self.sock.recv(MSGLEN-len(msg))
               if chunk == '':
                   raise RuntimeError("socket connection broken")
               msg = msg + chunk
           return msg

The sending code here is usable for almost any messaging scheme - in Python you
send strings, and you can use ``len()`` to determine its length (even if it has
embedded ``\0`` characters). It's mostly the receiving code that gets more
complex. (And in C, it's not much worse, except you can't use ``strlen`` if the
message has embedded ``\0``\ s.)

The easiest enhancement is to make the first character of the message an
indicator of message type, and have the type determine the length. Now you have
two ``recv``\ s - the first to get (at least) that first character so you can
look up the length, and the second in a loop to get the rest. If you decide to
go the delimited route, you'll be receiving in some arbitrary chunk size, (4096
or 8192 is frequently a good match for network buffer sizes), and scanning what
you've received for a delimiter.

One complication to be aware of: if your conversational protocol allows multiple
messages to be sent back to back (without some kind of reply), and you pass
``recv`` an arbitrary chunk size, you may end up reading the start of a
following message. You'll need to put that aside and hold onto it, until it's
needed.

Prefixing the message with it's length (say, as 5 numeric characters) gets more
complex, because (believe it or not), you may not get all 5 characters in one
``recv``. In playing around, you'll get away with it; but in high network loads,
your code will very quickly break unless you use two ``recv`` loops - the first
to determine the length, the second to get the data part of the message. Nasty.
This is also when you'll discover that ``send`` does not always manage to get
rid of everything in one pass. And despite having read this, you will eventually
get bit by it!

In the interests of space, building your character, (and preserving my
competitive position), these enhancements are left as an exercise for the
reader. Lets move on to cleaning up.


Binary Data
-----------

It is perfectly possible to send binary data over a socket. The major problem is
that not all machines use the same formats for binary data. For example, a
Motorola chip will represent a 16 bit integer with the value 1 as the two hex
bytes 00 01. Intel and DEC, however, are byte-reversed - that same 1 is 01 00.
Socket libraries have calls for converting 16 and 32 bit integers - ``ntohl,
htonl, ntohs, htons`` where "n" means *network* and "h" means *host*, "s" means
*short* and "l" means *long*. Where network order is host order, these do
nothing, but where the machine is byte-reversed, these swap the bytes around
appropriately.

In these days of 32 bit machines, the ascii representation of binary data is
frequently smaller than the binary representation. That's because a surprising
amount of the time, all those longs have the value 0, or maybe 1. The string "0"
would be two bytes, while binary is four. Of course, this doesn't fit well with
fixed-length messages. Decisions, decisions.


Disconnecting
=============

Strictly speaking, you're supposed to use ``shutdown`` on a socket before you
``close`` it.  The ``shutdown`` is an advisory to the socket at the other end.
Depending on the argument you pass it, it can mean "I'm not going to send
anymore, but I'll still listen", or "I'm not listening, good riddance!".  Most
socket libraries, however, are so used to programmers neglecting to use this
piece of etiquette that normally a ``close`` is the same as ``shutdown();
close()``.  So in most situations, an explicit ``shutdown`` is not needed.

One way to use ``shutdown`` effectively is in an HTTP-like exchange. The client
sends a request and then does a ``shutdown(1)``. This tells the server "This
client is done sending, but can still receive."  The server can detect "EOF" by
a receive of 0 bytes. It can assume it has the complete request.  The server
sends a reply. If the ``send`` completes successfully then, indeed, the client
was still receiving.

Python takes the automatic shutdown a step further, and says that when a socket
is garbage collected, it will automatically do a ``close`` if it's needed. But
relying on this is a very bad habit. If your socket just disappears without
doing a ``close``, the socket at the other end may hang indefinitely, thinking
you're just being slow. *Please* ``close`` your sockets when you're done.


When Sockets Die
----------------

Probably the worst thing about using blocking sockets is what happens when the
other side comes down hard (without doing a ``close``). Your socket is likely to
hang. SOCKSTREAM is a reliable protocol, and it will wait a long, long time
before giving up on a connection. If you're using threads, the entire thread is
essentially dead. There's not much you can do about it. As long as you aren't
doing something dumb, like holding a lock while doing a blocking read, the
thread isn't really consuming much in the way of resources. Do *not* try to kill
the thread - part of the reason that threads are more efficient than processes
is that they avoid the overhead associated with the automatic recycling of
resources. In other words, if you do manage to kill the thread, your whole
process is likely to be screwed up.


Non-blocking Sockets
====================

If you've understood the preceeding, you already know most of what you need to
know about the mechanics of using sockets. You'll still use the same calls, in
much the same ways. It's just that, if you do it right, your app will be almost
inside-out.

In Python, you use ``socket.setblocking(0)`` to make it non-blocking. In C, it's
more complex, (for one thing, you'll need to choose between the BSD flavor
``O_NONBLOCK`` and the almost indistinguishable Posix flavor ``O_NDELAY``, which
is completely different from ``TCP_NODELAY``), but it's the exact same idea. You
do this after creating the socket, but before using it. (Actually, if you're
nuts, you can switch back and forth.)

The major mechanical difference is that ``send``, ``recv``, ``connect`` and
``accept`` can return without having done anything. You have (of course) a
number of choices. You can check return code and error codes and generally drive
yourself crazy. If you don't believe me, try it sometime. Your app will grow
large, buggy and suck CPU. So let's skip the brain-dead solutions and do it
right.

Use ``select``.

In C, coding ``select`` is fairly complex. In Python, it's a piece of cake, but
it's close enough to the C version that if you understand ``select`` in Python,
you'll have little trouble with it in C. ::

   ready_to_read, ready_to_write, in_error = \
                  select.select(
                     potential_readers,
                     potential_writers,
                     potential_errs,
                     timeout)

You pass ``select`` three lists: the first contains all sockets that you might
want to try reading; the second all the sockets you might want to try writing
to, and the last (normally left empty) those that you want to check for errors.
You should note that a socket can go into more than one list. The ``select``
call is blocking, but you can give it a timeout. This is generally a sensible
thing to do - give it a nice long timeout (say a minute) unless you have good
reason to do otherwise.

In return, you will get three lists. They have the sockets that are actually
readable, writable and in error. Each of these lists is a subset (possibly
empty) of the corresponding list you passed in. And if you put a socket in more
than one input list, it will only be (at most) in one output list.

If a socket is in the output readable list, you can be
as-close-to-certain-as-we-ever-get-in-this-business that a ``recv`` on that
socket will return *something*. Same idea for the writable list. You'll be able
to send *something*. Maybe not all you want to, but *something* is better than
nothing.  (Actually, any reasonably healthy socket will return as writable - it
just means outbound network buffer space is available.)

If you have a "server" socket, put it in the potential_readers list. If it comes
out in the readable list, your ``accept`` will (almost certainly) work. If you
have created a new socket to ``connect`` to someone else, put it in the
potential_writers list. If it shows up in the writable list, you have a decent
chance that it has connected.

One very nasty problem with ``select``: if somewhere in those input lists of
sockets is one which has died a nasty death, the ``select`` will fail. You then
need to loop through every single damn socket in all those lists and do a
``select([sock],[],[],0)`` until you find the bad one. That timeout of 0 means
it won't take long, but it's ugly.

Actually, ``select`` can be handy even with blocking sockets. It's one way of
determining whether you will block - the socket returns as readable when there's
something in the buffers.  However, this still doesn't help with the problem of
determining whether the other end is done, or just busy with something else.

**Portability alert**: On Unix, ``select`` works both with the sockets and
files. Don't try this on Windows. On Windows, ``select`` works with sockets
only. Also note that in C, many of the more advanced socket options are done
differently on Windows. In fact, on Windows I usually use threads (which work
very, very well) with my sockets. Face it, if you want any kind of performance,
your code will look very different on Windows than on Unix.


Performance
-----------

There's no question that the fastest sockets code uses non-blocking sockets and
select to multiplex them. You can put together something that will saturate a
LAN connection without putting any strain on the CPU. The trouble is that an app
written this way can't do much of anything else - it needs to be ready to
shuffle bytes around at all times.

Assuming that your app is actually supposed to do something more than that,
threading is the optimal solution, (and using non-blocking sockets will be
faster than using blocking sockets). Unfortunately, threading support in Unixes
varies both in API and quality. So the normal Unix solution is to fork a
subprocess to deal with each connection. The overhead for this is significant
(and don't do this on Windows - the overhead of process creation is enormous
there). It also means that unless each subprocess is completely independent,
you'll need to use another form of IPC, say a pipe, or shared memory and
semaphores, to communicate between the parent and child processes.

Finally, remember that even though blocking sockets are somewhat slower than
non-blocking, in many cases they are the "right" solution. After all, if your
app is driven by the data it receives over a socket, there's not much sense in
complicating the logic just so your app can wait on ``select`` instead of
``recv``.