Mighty Sockets

 

Sockets are indeed the window to networking and networking drives the whole world some way or other, that's why the name "Mighty Sockets".

In simple terms, a socket is an endpoint of communication. And a connection is represented by a socket pair. Remember, not all communication is connection oriented. So, you can have a single socket also doing most of the job for you (e.g. UDP sockets). A process deals with the socket in the same way as it deals with files. We'll see more of this similarity in coming sections.

1. HOW DO WE CREATE THEM?

As always we have a system call to do the job for us :)

int sockfd = socket(AF_INET, SOCK_STREAM, 0);

//socket(2) prototype: socket(int domain, int type, int protocol)

What it says is - give me a socket of domain INET (AF_INET is for Address Famiy - INET) and of type SOCK_STREAM (stream socket) and default protocol that goes with streaming sockets (it's TCP).

Socket system call does a lot of things behind the scene. It creates socket structure in kernel memory. This structure is then filled by various protocol independent data and other data depending upon the address family, type and protocol.

2. NAMING THE SOCKET

Naming or addressing of a socket is very important. This is how data is supposed to reach the right process.

How sockets are addressed depends upon the domain of the socket. For example, an INET domain socket is identified by an IP address and a port combined. Of course, an IP address is required if you want to reach a machine -- IP address is used by the inernet protocol to route data from one machine to another machine. Port is required by the kernel to route packets to the right process.

So as you can see, naming of a socket is important. But, you don't have to do this yourself always. If a specific port number is not required by an application process (as is the case for most of the client applications), the kernel may assign a port to the socket internally. However, if you are looking for a specific port, for example in case of a server, you need to register it through the 'bind(2)' system call:

int bind(int sockfd, struct sockaddr *my_addr,
         socklen_t addrlen);
"sockaddr" structure has fields like socket family (AF_INET), IP address (for IP) and Port. struct sockaddr is a generic structure. Generally we fill information in a more specific structure, for example sockaddr_in for domain INET.

This way we bind the socket to a specific address, or in other words, we give it a name. As mentioned earlier, this is required generally only in case of server, when we want a specific port for ourselves.

3. CONNECTION ESTABLISHMENT:

Connection establishment is required for all connection oriented protocols (like TCP). In TCP terminology, a connection is established using a three-way handshake. A SYN packet is sent from one side (generally client), and when that packet arrives at the receiving end (generally a server), that end sends a [SYN,ACK] to the client, and when server receives a final ACK from the client, connection is considered to be complete.

                active          passive
                |      SYN      |
                |-------------->|
                |   [SYN, ACK]  |
                |<------------- |
                |      ACK      |
                |-------------->|
                |               |

There are 2 possible ways to establish a connection: Either you initiate the connection or wait for a connection. If you choose to wait for a connection (e.g. you are a server), you do a 'passive open' and if you want to initiate the connection (e.g. you are client and want something done from server), you do an 'active open'

3.1 PASSIVE OPEN:

If you are doing passive open, you are waiting for the connection. Waiting for a connection is also called 'listening'. Your server starts listening when it executes:

listen(backlog);

"The 'listen' function converts an unconnected socket into a passive socket, indicating that the kernel should accept incoming connection requests directed to this socket."

For a given listening socket kernel maintains 2 queues:

  1. An 'incomplete conneciton queue', which contains an entry for all unacknowledged connections. Basically, I have received the SYN from the other machine, sent the [SYN, ACK] and waiting for an ACK from other side. These sockets are in the SYN_RCVD state.
  2. A 'completed connection queue', which contains an entry for each machine with whom the TCP three-way handshake has been completed. These sockets are in the ESTABLISHED state.
Sum of the both queues cannot exceed backlog.

		 _______________________
		|     	 _______	|
		|	|Server	|	|
		|	|_______|	|
listening <-----|------	(*.1200,*.*)	|
socket		|_______________________|
			fig. 1

Above figure, represents a server listening at port 1200. No connection has been established yet.

Fetching Connections:

Connections, if any, are still within the TCP stack. To return a connection to the application

int new_sock = accept(int orig_sock, struct sockaddr* addr, 
		      socklen_t *addrlen);

'accept' is called by an application to return the next completed connection from the front of the completed connection queue. If the completed connection queue is empty the process is put to sleep (assuming the default of a blocking socket).

If 'accept' is successful, its return value is a brand new descriptor that was automatically created by kernel. This new descriptor refers to the TCP connection with the client. Listening socket remains untouched. Client socket address and length information is filled by the "accept" call. Server may or may not use them.

So you have 2 sockets now, one listening and other one already connected. You may close orig_sock, if you have no use for it now (eg. don't want to continue listening).

	 _______________________	 _______________________			
	|     	 _______	|	|	 _______	|
	|	|Server	|	|	|----->	|Client	|	|
	|	|_______|	|      /|	|_______|	|
listening----->(*.1200,*.*)	|     /	|(3.209.127.100:1500,	|
socket	|	 _______	|    /	| 3.209.129.47:1200)	|
	|	|Server	|	|   /	|_______________________|	
	|	|_______|	|  /		3.209.127.100
	| (3.209.129.47:1200,<--|-/
	| 3.209.127.100:1500)	|
	|_______________________|
		3.209.129.47
				Figure 2	

3.2 ACTIVE OPEN:

When you initiate a connection, you are doing an 'active open'. You just need to call 'connect(2)' to initiate a connection.

int connect(int sockfd, const struct sockaddr *serv_addr,
	    socklen_t addrlen);

This initiates the process of connection (3-way handshake in TCP teminology) and returns only when connection is complete (or something goes wrong with your luck like signals and all). If not able to establish connection, it returns -1 and sets errno.

When you call connect, the kernel implicitly allocates a port to the socket (naming of socket). Fig. 2 has a client connected to the server. Port 1500 has been allocated to the client implicitly by the kernel.

4. DATA TRANSFER:

Now next obvious step will be to get some work done with this socket.

Mechanism for sending and receiving data is actually very simple. You have to just read, write the socket descriptor like file descriptor. But, what happens behind the scenes is more important (like always ;)). A lot can be said about sockets, but I'll just talk about the packets demultiplexing next, which is handled by the TCP-IP stack.

5. DEMULTIPLEXING:

Ok, so now you have a connection, and client is beginning to sends some data to the server, say a request for a web page in case of a web client. How does this data reach the server? Packets carry the information like source address (ip and port) and destination address. IP layer does all the work to make the packet reach the destination machine, using the "dst" IP address.

But now what, we have 2 sockets on the server with the same name ie. with same ports (remember one socket listening and other connected to the client). There can be many more sockets with the same name depending upon the number of clients connected at a time. Here comes the TCP layer to rescue. TCP demultiplexes incoming segments for us.

TCP cannot demulitplex incoming segments by looking at just the destination port number. TCP must look at all 4 elements in the socket pair to determine which endpoint receives the arriving segment.

						
	 _______________________	 _______________________			
	|     	 _______	|	|	 _______	|
	|	|Server	|	|     --|----->	|Client	|	|
	|	|_______|	|    /	|	|_______|	|
	|	(*.1200,*.*)	|   /	|(3.209.127.100:1500,	|
	|	 _______	|  /	| 3.209.129.47:1200)	|
	|	|Server	|  <----|-/	|	 _______	|	
	|	|__c_1__|	|     --|----->	|Client	|	|
	| (3.209.129.47:1200,   |    /	|	|_______|	|
	| 3.209.127.100:1500)	|   /	|(3.209.127.100:1501,	|
	|	 _______	|  /	| 3.209.129.47:1200)	|
	|	|Server	|  <----|-/	|_______________________|	
	|	|__c_2__|	|	      3.209.127.100
	| (3.209.129.47:1200,   |
	| 3.209.127.100:1501)	|
	|		 	|
	|_______________________|
	      3.209.129.47
				Figure 3

In figure 3, we have 2 connections open between server machine and client machine. It's a case of concurrent servers (most typical way to handle clients). In this case, when a new connection arrives, server forks out and lets its child handle the connection and while it itself keeps listening for new connections.

Here we have 3 sockets with the same local port (1200). If a segment arrives from 3.209.127.100 port 1500 destined for port 1200, it is delivered to first child. If a segment arrives from 3.209.127.100 port 1501 destined for 3.209.129.47 port 1200, it is delivered to the second child. All other segments destined for port 1200 are delivered to the original server with the listening socket.

There is a lot more to talk about sockets. But, there is a lot more documentation too. This is just an attempt to explain what most other documents miss to explain.


Copyright 2006 .