Kernel Korner - Network Programming in the Kernel

All Linux distributions provide a wide range of network applications—from dæmons that provide a variety of services such as WWW, mail and SSH to client programs that access one or more of these services. These programs are written in user mode and use the system calls provided by the kernel to perform various operations like network read and write. Although this is the traditional method of writing programs, there is another interesting way to develop these applications by implementing them in the kernel. The TUX Web server is a good example of an application that runs inside the kernel and serves static content. In this article, we explain the basics of writing network applications within the kernel and their advantages and disadvantages. As an example, we explain the implementation of an in-kernel FTP client.

Why would one want to implement applications within the kernel? Here are a few advantages:

On the other hand, in-kernel implementations have certain disadvantages:

Network programming is usually done with sockets. A socket serves as a communication end point between two processes. In this article, we describe network programming with TCP/IP sockets.

Server programs create sockets, bind to well-known ports, listen and accept connections from clients. Servers are usually designed to accept multiple connections from clients—they either fork a new process to serve each client request (concurrent servers) or completely serve one request before accepting more connections (iterative servers). Client programs, on the other hand, create sockets to connect to servers and exchange information.

Let's take a quick look at how an FTP client and server are implemented in user mode. We discuss only active FTP in this article. The differences between active and passive FTP are not relevant to our discussion of network programming here.

Here is a brief explanation of the design of an FTP client and server. The server program creates a socket using the socket() system call. It then binds on a well-known port using bind() and waits for connections from clients using the listen() system call. The server then accepts incoming requests from clients using accept() and forks a new process (or thread) to serve each incoming client request.

The client program creates a control socket using socket() and next calls connect() to establish a connection with the server. It then creates a separate socket for data transfer using socket() and binds to an unprivileged port (>1024) using bind(). The client now listen()s on this port for data transfer from the server. The server now has enough knowledge to honor a data transfer request from the client. Finally, the client uses accept() to accept connections from the server to send and receive data. For sending and receiving data, the client and server use the write() and read() or sendmsg() and recvmsg() system calls. The client issues close() on all open sockets to tear down its connection to the server. Figure 1 sums it up.

Figure 1. The FTP protocol uses two sockets: one for control messages and one for data.