EE122: Socket Programming

Igor Ganichev Originally prepared by DK Moon

Socket API?
Q. What would you expect when learning a new Unix command (e.g., ls) ?
a) Source code b) Program options => Implementation detail => Interface

Application Programming Interface (API)

Interface to a particular service Abstracts away from implementation detail Set of functions, data structures, and constants.

Socket API
Network programming interface

Socket API
Socket API
Network programming interface Application

Transport Network

TCP IP

UDP

Socket API

BSD Socket API

From your university, UC Berkeley (1980s) Most popular network API Ported to various OSes, various languages
Windows Winsock, BSD, OS X, Linux, Solaris, Socket modules in Java, Python, Perl,

Similar to Unix file I/O API

In the form of file descriptor (sort of handle). Can share the same read()/write()/close() system calls.

Sockets
Various sockets Any similarity?

Endpoint of a connection
Identified by IP address and Port number

Primitive to implement high-level networking interfaces

e.g., Remote procedure call (RPC)

Outline
Socket API motivation, background Types of sockets (TCP vs. UDP) Elementary API functions I/O multiplexing Project Distributed Cracking Appendix (not covered in the lecture)

Types of Sockets
Stream socket (aka TCP)
Connection-oriented
Requires connection establishment & termination

Datagram socket (aka UDP)

Connection-less

Reliable delivery
In-order delivery Retransmission No duplicates

Best-effort delivery
Arbitrary order of packet delivery No retransmission Possible duplicates

High variance in latency

Cost of the reliable service

Low variance in latency

Packet-like interface
Requires packetizing

File-like interface (streaming) E.g., HTTP, SSH, FTP,

E.g., DNS, VoIP, VOD, AOD,

Types of Sockets (contd)

When sending Hi! and Hope youre well TCP treats them as a single bytes stream
Bytes stream

l l e w e poH! i H
UDP treats them as separate messages
Hope youre well Hi!
8

Types of Sockets (contd)

Thus, TCP needs application-level message boundary.
By carrying length in application-level header E.g. struct my_app_hdr {
int length
}

Bytes stream

l l e w o H 16 ! i H 3
9

Outline
Socket API motivation, background Types of sockets (TCP vs. UDP) Elementary API functions I/O multiplexing Project Distributed Cracking

Host name, IP address, Port number

Host name
Human readable name (e.g., www.eecs.berkeley.edu) Variable length Could have multiple IP addresses

IP version 4 address
Usually represented as dotted numbers for human readability
E.g., 128.32.132.214

32 bits in network byte order

E.g., 1.2.3.4 => 0x04030201

Port number
Identifies a service (or application) on a host
E.g., TCP Port 80 => web service, UDP Port 53 => name service (DNS)

16 bit unsigned number (0~65535)

Scenario #1 TCP client-server

Sequence of actions
Client
socket()

Server
bind() listen()
Termination Data transfer Connection Initialization Establishment
12

socket()

connect() send() or write() recv() or read() Time close()

accept() recv() or read() send() or write() close()

Initialization: server + client, socket()

int sock = socket(AF_INET, SOCK_STREAM, 0); if (sock < 0) { perror(socket() failed); abort(); }

socket(): returns a socket descriptor

AF_INET: IPv4 address family. (also OK with PF_INET)
C.f. IPv6 => AF_INET6

SOCK_STREAM: streaming socket type

C.f. SOCK_DGRAM

perror(): prints out an error message

Error code in Unix programming

extern int errno; // by #include <errno.h>

Many Unix system calls and library functions set errno on errors Macros for error codes (E + error name)
EINTR, EWOULDBLOCK, EINVAL, man func_name shows possible error code for the function name

Functions to convert error code into human readable msgs

void perror(const char *my_str)
Always looks for errno prints out my str: error code string

const char *strerror(int err_code)

You must provide an error code returns a string for the err_code

Initialization: server, bind()

Server needs to bind a particular port number.
struct sockaddr_in sin; memset(&sin, 0, sizeof(sin)); sin.sin_family = AF_INET; sin.sin_addr.s_addr = INADDR_ANY; sin.sin_port = htons(server_port); if (bind(sock, (struct sockaddr *) &sin, sizeof(sin)) < 0) { perror(bind failed); abort(); }
bind(): binds a socket with a particular port number.
Kernel remembers which process has bound which port(s). Only one process can bind a particular port number at a time.

struct sockaddr_in: Ipv4 socket address structure. (c.f., struct sockaddr_in6) INADDR_ANY: If server has multiple IP addresses, binds any address.

htons(): converts host byte order into network byte order.

Endianess
Q) You have a 16-bit number: 0x0A0B. How is it stored in memory?
Increasing address

0x0A
0x0B
Increasing address

0x0B
0x0A

Big Endian
Little Endian

Host byte order is not uniform

Some machines are Big endian, others are Little endian

Communicating between machines with different host byte orders is problematic

Transferred $256 (0x0100), but received $1 (0x0001)

Endianness (contd)
Network byte order: Big endian
To avoid the endian problem

We must use network byte order when sending 16bit, 32bit , 64bit numbers. Utility functions for easy conversion
uint16_t htons(uint16_t host16bitvalue); uint32_t htonl(uint32_t host32bitvalue); uint16_t ntohs(uint16_t net16bitvalue); uint32_t ntohl(uint32_t net32bitvalue);

Hint: h, n, s, and l stand for host byte order, network byte order, short(16bit), and long(32bit), respectively

Initialization: server, bind()

Server needs to bind a particular port number.
struct sockaddr_in sin; memset(&sin, 0, sizeof(sin)); sin.sin_family = AF_INET; sin.sin_addr.s_addr = INADDR_ANY; sin.sin_port = htons(server_port); if (bind(sock, (struct sockaddr *) &sin, sizeof(sin)) < 0) { perror(bind failed); abort(); }
bind(): binds a socket with a particular port number.
Kernel remembers which process has bound which port(s). Only one process can bind a particular port number at a time.

struct sockaddr_in: Ipv4 socket address structure. (c.f., struct sockaddr_in6) INADDR_ANY: If server has multiple IP addresses, binds any address.

htons(): converts host byte order into network byte order.

Reusing the same port

After TCP connection closes, waits for 2MSL, which is twice maximum segment lifetime (from 1 to 4 mins, implementation dependent). Why? Port number cannot be reused before 2MSL But server port numbers are fixed => Must be reused Solution: Put this code before bind() int optval = 1; if (setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, &optval, sizeof(optval)) < 0) { perror(reuse failed); abort(); } setsockopt(): changes socket, protocol options.
e.g., buffer size, timeout value,

Initialization: server, listen()

Socket is active, by default We need to make it passive to get connections. if (listen(sock, back_log) < 0) { perror(listen failed); abort(); } listen(): converts an active socket to passive

back_log: connection-waiting queue size. (e.g., 32)

Busy server may need a large value (e.g., 1024, )

Initialization Summary
Client
socket()

Server
socket() setsockopt(sock, SOL_SOCKET, SO_REUSEADDR) bind() listen()

Pitfalls
The order of the functions matter Do not forget to use htons() to handle port number

Scenario #1 TCP client-server

Sequence of actions
Client
socket()

Server
bind() listen()
Termination Data transfer Connection Initialization Establishment
22

socket()

connect() send() or write() recv() or read() Time close()

accept() recv() or read() send() or write() close()

Connection Establishment (client)

struct sockaddr_in sin; memset(&sin, 0 ,sizeof(sin)); sin.sin_family = AF_INET; sin.sin_addr.s_addr = inet_addr(128.32.132.214); sin.sin_port = htons(80);

if (connect(sock, (struct sockaddr *) &sin, sizeof(sin)) < 0) { perror(connection failed); abort(); }

Connect(): waits until connection establishes/fails

inet_addr(): converts an IP address string into a 32bit address number (network byte order).

Connection Establishment (server)

struct sockaddr_in client_sin; int addr_len = sizeof(client_sin); int client_sock = accept(listening_sock, (struct sockaddr *) &client_sin, &addr_len); if (client_sock < 0) { perror(accept failed); abort(); }

accept(): returns a new socket descriptor for a client connection in the connection-waiting queue.
This socket descriptor is to communicate with the client The passive socket (listening_sock) is not to communicate with a client

client_sin: contains client IP address and port number

Q) Are they in Big endian or Litten endian?

Scenario #1 TCP client-server

Sequence of actions
Client
socket()

Server
bind() listen()
Termination Data transfer Connection Initialization Establishment
25

socket()

connect() send() or write() recv() or read() Time close()

accept() recv() or read() send() or write() close()

Sending Data: server+client, send()

char *data_addr = hello, world; int data_len = 12; int sent_bytes = send(sock, data_addr, data_len, 0); if (sent_bytes < 0) { perror(send failed); }

send(): sends data, returns the number of sent bytes

Also OK with write(), writev()

data_addr: address of data to send data_len: size of the data With blocking sockets (default), send() blocks until it sends all the data. With non-blocking sockets, sent_bytes may not equal to data_len
If kernel does not have enough space, it accepts only partial data You must retry for the unsent data

Receiving Data: server+client, recv()

char buffer[4096]; int expected_data_len = sizeof(buffer); int read_bytes = recv(sock, buffer, expected_data_len, 0); if (read_bytes == 0) { // connection is closed } else if (read_bytes < 0) { // error perror(recv failed); } else { // OK. But no guarantee read_bytes == expected_data_len }

recv(): reads bytes from the socket and returns the number of read bytes.
Also OK with read() and readv()

read_bytes may not equal to expected_data_len

If no data is available, it blocks If only partial data is available, read_bytes < expected_data_len On socket close, expected_data_len equals to 0 (not error!) If you get only partial data, you should retry for the remaining portion.

Termination: server+client, close()

// after use the socket close(sock);

close(): closes the socket descriptor We cannot open files/sockets more than 1024*
We must release the resource after use

* Super user can overcome this constraint, but regular user cannot.

Scenario #2 UDP client-server

Q) What must be changed?
Client
socket()

Server
bind() listen()
Termination Data transfer Connection Initialization Establishment
29

socket()

connect() send() or write() recv() or read() Time close()

accept() recv() or read() send() or write() close()

Scenario #2 UDP client-server

A) We need a different initialization
Client
socket()

Server
bind() listen()
Termination Data transfer Connection Initialization Establishment
30

socket()

connect() send() or write() recv() or read() Time close()

accept() recv() or read() send() or write() close()

Initialization: UDP
int sock = socket(AF_INET, SOCK_DGRAM, 0); if (sock < 0) { perror(socket failed); abort(); }

UDP uses SOCK_DGRAM instead of SOCK_STREAM

Scenario #2 UDP client-server

Q) What else must be changed?
Client
socket()

Server
bind() listen()
Termination Data transfer Connection Initialization Establishment
32

socket()

connect() send() or write() recv() or read() Time close()

accept() recv() or read() send() or write() close()

Scenario #2 UDP client-server

A) UDP is connection-less. We remove all connection related steps.
Client
socket()

Server
bind() listen()
Initialization Termination Data transfer
33

socket()

send() or write() recv() or read() Time close()

recv() or read() send() or write() close()

Scenario #2 UDP client-server

A) listen() is also related to connection. Remove it.
Client
socket()

Server
bind()
Initialization Termination Data transfer
34

socket()

send() or write() recv() or read() Time close()

recv() or read() send() or write() close()

Scenario #2 UDP client-server

Q) Now its unclear where to send packets and from where I receive! Can we solve this?
Client
socket()

Server
bind()
Initialization Termination Data transfer
35

socket()

send() or write() recv() or read() Time close()

recv() or read() send() or write() close()

Scenario #2 UDP client-server

A) Give <address,port> information when sending a packet. That is, use sendto() and recvfrom() instead of send() and recv()

Client
socket()

Server
bind()
Initialization Termination Data transfer
36

socket()

sendto() recvfrom() Time close()

recvfrom() sendto() close()

Send Data Over UDP: sendto()

struct sockaddr_in sin; memset(&sin, 0, sizeof(sin)); sin.sin_family = AF_INET; sin.sin_addr.s_addr = inet_addr(128.32.132.214); sin.sin_port = htons(1234);

sent_bytes = sendto(sock, data, data_len, 0, (struct sockaddr *) &sin, sizeof(sin)); if (sent_bytes < 0) { perror(sendto failed); abort(); }

sendto(): sends a packet to a specific destination address and port

c.f., in TCP, we do this destination setting when calling connect()

As opposed to TCP, UDP packetizes data. So, sendto() sends all data or nothing.

Receive Data Over UDP: recvfrom()

struct sockaddr_in sin; int sin_len; char buffer[4096]; int read_bytes = recvfrom(sock, buffer, sizeof(buffer), 0, (struct sockaddr *) &sin, &sin_len);

if (read_bytes < 0) { perror(recvfrom failed); abort(); }

recvfrom(): reads bytes from the socket and sets the source information Reading 0 bytes does not mean connection closed unlike TCP.
Recall UDP does not have a notion of connection.

API functions Summary

TCP
Initialization
socket(AF_INET, SOCK_STREAM, 0) setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, ) bind() listen()

UDP
Initialization
socket(AF_INET, SOCK_DGRAM, 0) setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, ) bind()

Conneciton
connect() accept()

No connection Data transfer

sendto() recvfrom()

Data transfer
send() recv()

Termination
close()

Termination
close()

Outline
Socket API motivation, background Types of sockets (TCP vs. UDP) Elementary API functions I/O multiplexing Project Distributed Cracking

How to handle multiple inputs?

Data sources
Standard input (e.g., keyboard) Multiple sockets

Problem: asynchronous data arrival

Program does not know when it will arrive.

If no data available, recv() blocks. If blocked on one source, cannot handle other sources
Suppose what if a web server cannot handle multiple connections

Solutions
Polling using non-blocking socket Inefficient I/O multiplexing using select() simple Multithreading more complex. Not covered today

Polling using non-blocking socket

This approach wastes CPU cycles

int opt = fcntl(sock, F_GETFL); Gets the sockets if (opt < 0) { option perror(fcntl failed); abort(); } Updates the sockets if (fcntl(sock, F_SETFL, opt | O_NONBLOCK) < 0) { perror(fcntl failed); option with non abort(); blocking option } while (1) { int read_bytes = recv(sock, buffer, sizeof(buffer), 0); if (read_bytes < 0) { if (errno == EWOULDBLOCK) { When no data, // OK. Simply no data we see EWOULDBLOCK } else { error code. perror(recv failed); abort(); } }

I/O multiplexing using select()

fd_set read_set; struct timeval timeout FD_ZERO(&read_set); FD_SET(sock1, &read_set); FD_SET(sock2, &read_set); timeout.tv_sec = 0; timeout.tv_usec = 5000;

Initializes arguments for select()

if (select(MAX(sock1, sock2) + 1, &read_set, NULL, NULL, &time_out) < 0) { perror(select failed); abort(); Pass NULL instead of } if (FD_ISSET(sock1, &read_set)) { // sock1 has data } if (FD_ISSET(sock2, &read_set)) { // sock2 has data }

&timeout if you want to wait indefinitely

Checks I/O events.

Bibles both by W. Richard Stevens

Outline
Socket API motivation, background Types of sockets (TCP vs. UDP) Elementary API functions I/O multiplexing Project Distributed Cracking

Project Distributed Cracking

Earth is under attack by aliens. Luckily they use crypto based on prime numbers. To crack it, we need a system to check if a number is prime very quickly. In Project 1, you will build a distributed system to quickly compute prime numbers. In Project 2, you will enhance the system to store the prime numbers and to query for them.

Project 1

Check if 1234321 is prime

Hello crackers! I found that 19191 is not prime

Project 1
Part 2
Hello crackers! I found that 19191 is not prime

Check if 1234321 is prime

Part 1

Project 2
Cracker App using our prime number service

Appendix Programming Tips

Will not be covered during the lecture Please refer to these tips if youre interested

Tip #1
How to check the host byte order of my machine?
union { uint16_t number; uint8_t bytes[2]; } test; test.number = 0x0A0B; printf(%02x%02x\n, test.bytes[0], test.bytes[1]);

Tip #2
How to get IP address from host name
Use gethostbyname()
struct sockaddr_in sin; struct hostent *host; host = gethostbyname(www.berkeley.edu); sin.sin_addr.s_addr = *(unsigned *) host->h_addr_list[0];

Tip #3
By default, Unix terminates the process with SIGPIPE if you write to a TCP socket which has been closed by the other side. You can disable it by:
signal(SIGPIPE, SIG_IGN);

Tip #4 - Structure Packing

We have the following application-level packet header format (the numbers denote field size in bytes) length type source dest
addr addr 2 1 4 4

So, we define the header as struct like this:

struct my_pkt_hdr { unsigned short length; unsigned char type; unsigned int source_addr; unsigned int dest_addr; };

Q) Result of sizeof(struct my_pkt_hdr)?

Tip #4 - Structure Packing (contd)

Compiler will try to be 4-byte aligned (on 32bit machines) To avoid the previous case, we must pack struct
Windows programming style #pragma pack(push, 1) struct my_pkt_hdr { unsigned short length; unsigned char type; unsigned int source_addr; unsigned int dest_addr; }; #pragma pack(pop) GCC style struct my_pkt_hdr { unsigned short length; unsigned char type; unsigned int source_addr; unsigned int dest_addr; } __attribute__((packed));

OR

Using man pages

Best source to study system calls and library functions
Tells which header files should be included Describes how each function works Tells what the return value means and what error number can happen E.g., man connect