Lecture 15: Network System Calls, Library Functions

• Hostname Resolution: IPv4
• Linux C includes directives to convert host names (e.g. "www.facebook.com") to IPv4 address (e.g. "31.13.75.17") and vice versa. Functions called gethostbyname and gethostbyaddr, while technically deprecated, are still so prevalent that you should know how to use them.
• In fact, your B&O textbook only mentions these deprecated functions:

• Each function populates a statically allocated struct hostent with network information about some machine on the Internet.
  • gethostbyname assumes its argument is a host name (e.g. "www.google.com").
  • gethostbyaddr assumes the first argument is a binary representation of an IP address (e.g. not the string "171.64.64.137", but the base address of a character array with ASCII values of 171, 64, 64, and 137 laid down side by side in network byte order, which is the same as big endian order.
    For IPv4, the second argument is usually 4 (or rather, sizeof(struct in_addr)) and the third is typically the AF_INET constant.

Lecture 15: Network System Calls, Library Functions

• Hostname Resolution: IPv4
  • The struct hostent packages all of the information about a particular host:

• The struct in_addr is a one-field record modeling an IPv4 address.
  • The s_addr field packs each figure of a dotted quad (e.g. 171.64.64.136) into one of its four bytes. Each of these four numbers numbers can range from 0 up through 255.
• The struct hostent is used for all IP addresses, not just IPv4 addresses. For non-IPv4 addresses, h_addrtype, h_length, and h_addr_list carry different types of data than they do for IPv4.

Lecture 15: Network System Calls, Library Functions

• Users prefer the host naming scheme behind "www.facebook.com", but network communication ultimately works with IP addresses like "31.13.75.17".
  • Not surprisingly, gethostbyname and gethostbyaddr are used to manage translations between the two.
  • Here's the core of a larger program that continuously polls the user for hostnames and responds by publishing the set of the one or more IP addresses each hostname is bound to:

Lecture 15: Network System Calls, Library Functions

• h_addr_list is typed to be a char * array, implying it's an array of C strings, perhaps dotted quad IP addresses. However, that's not correct. For IPv4 records, h_addr_list is an array of struct in_addr *s.
• The inet_ntop function places a traditional C string presentation of an IP address into the provided character buffer, and returns the the base address of that buffer.
• The while loop crawls over the h_addr_list array until it lands on a NULL.

Lecture 15: Network System Calls, Library Functions

Hostname Resolution: IPv4

• A sample run of our hostname resolver is presented on the right.
• In general, you see that most of the hostnames we recognize are in fact the officially recorded hostnames.
• www.yale.edu is the exception. It looks like Yale relies on a content delivery network called Cloudflare, and www.yale.edu is catalogued as an alias.
• Google's IP address is different by geographical location, which is why it exposes only one IP address.
• Apparently billions of people use okcupid every millisecond, though, which is why it necessarily exposes five.

Lecture 15: Network System Calls, Library Functions

Hostname Resolution: IPv6

• Because IPv4 addresses are 32 bits, there are 2^32, or roughly 4 billion different IP addresses. That may sound like a lot, but it was recognized decades ago that we'd soon run out of IPv4 addresses.
• In contrast, there are 340,282,366,920,938,463,463,374,607,431,768,211,456 IPv6 addresses. That's because IPv6 addresses are 128 bits.
• Here are a few IPv6 addresses:
  • Google's 2607:f8b0:4005:80a::2004
  • MIT's 2600:1406:1a:396::255e and 2600:1406:1a:38d::255e
  • Berkeley's 2600:1f14:436:7801:15f8:d879:9a03:eec0 and 2600:1f14:436:7800:4598:b474:29c4:6bc0
  • The White House's 2600:1406:1a:39e::fc4 and 2600:1406:1a:39b::fc4

A more generic version of gethostbyname—inventively named gethostbyname2—can be used to extract IPv6 address information about a hostname.

Lecture 15: Network System Calls, Library Functions

• There are only two valid address types that can be passed as the second argument to gethostbyname2: AF_INET and AF_INET6.
• A call to gethostbyname2(host, AF_INET) is equivalent to a call to gethostbyname(host)
• A call to gethostbyname2(host, AF_INET6) still returns a struct hostent *, but the struct hostent is populated with different values and types:
• the h_addrtype field is set to AF_INET6,
• the h_length field houses a 16 (or rather, sizeof(struct in6_addr)), and
• the h_addr_list field is really an array of struct in6_addr pointers, where each struct in6_addr looks like this:

Lecture 15: Network System Calls, Library Functions

• Here is the IPv6 version of the publishIPAddressInfo we wrote earlier (we call it publishIPv6AddressInfo).

• Notice the call to gethostbyname2, and notice the explicit use of AF_INET6, struct in6_addr, and INET6_ADDRSTRLEN.
• Full program is right here.

Lecture 15: Network System Calls, Library Functions

• A sample run of our IPv6 hostname resolver is presented below.
• Note that many hosts aren't IPv6-compliant yet, so they don't admit IPv6 addresses.

Lecture 15: Network System Calls, Library Functions

• The three data structures presented below are in place to model the IP address/port pairs:

The sockaddr_in is used to model IPv4 address/port pairs.

• The sin_family field should always be initialized to be AF_INET, which is a constant used to be clear that IPv4 addresses are being used. If it feels redundant that a record dedicated to IPv4 needs to store a constant saying everything is IPv4, then stay tuned.
• The sin_port field stores a port number in network byte (i.e. big endian) order.
• The sin_addr field stores an IPv4 address as a packed, big endian int, as you saw with gethostbyname and the struct hostent.
• The sin_zero field is generally ignored (though it's often set to store all zeroes). It exists to pad the record up to 16 bytes.

Lecture 15: Network System Calls, Library Functions

• The three data structures presented below are in place to model the IP address/port pairs:

• The sin6_family field should always be set to AF_INET6. As with the sin_family field, sin6_family field occupies the first two bytes of surrounding record.
• The sin6_port field holds a two-byte, network-byte-ordered port number, just like sin_port does.
• A struct in6_addr is also wedged in there to manage a 128-bit IPv6 address.
• sin6_flowinfo and sin6_scope_id are beyond the scope of what we need, so we'll ignore them.

Lecture 15: Network System Calls, Library Functions

• The three data structures presented below are in place to model the IP address/port pairs:

• You rarely if ever declare variables of type struct sockaddr, but many system calls will accept parameters of type struct sockaddr *.
• Rather than define a set of network system calls for IPv4 addresses and a second set of system calls for IPv6 addresses, Linux defines one set for both.
• If a system call accepts a parameter of type struct sockaddr *, it really accepts the address of either a struct sockaddr_in or a struct sockaddr_in6. The system call relies on the value within the first two bytes—the sa_family field—to determine what the true record type is.

Lecture 15: Network System Calls, Library Functions

At this point, we know most of the directives needed to implement and understand how to implement createClientSocket and createServerSocket.

• createClientSocket is the easier of the two, so we'll implement that one first. (For simplicity, we'll confine ourselves to an IPv4 world.)
• Fundamentally, createClientSocket needs to:
• • Confirm the host of interest is really on the net by checking to see if it has an IP address. gethostbyname does this for us.
  • Allocate a new descriptor and configure it to be a socket descriptor. We'll rely on the socket system call to do this.
  • Construct an instance of a struct sockaddr_in that packages the host and port number we're interested in connecting to.
  • Associate the freshly allocated socket descriptor with the host/port pair. We'll rely on an aptly named system call called connect to do this.
  • Return the fully configured client socket.
• The full implementation of createClientSocket is on the next slide (and right here).

Lecture 15: Network System Calls, Library Functions

Lecture 15: Network System Calls, Library Functions

Here are a few details about my implementation of createClientSocket worth calling out:

• We call gethostbyname first before we call socket, because we want to confirm the host has a registered IP address—which means it's reachable—before we allocate any system resources.
• Recall that gethostbyname is intrinsically IPv4. If we wanted to involve IPv6 addresses instead, we would need to use gethostbyname2.
• The call to socket finds, claims, and returns an unused descriptor. AF_INET configures it to be compatible with an IPv4 address, and SOCK_STREAM configures it to provide reliable data transport, which basically means the socket will reorder data packets and requests missing or garbled data packets to be resent so as to give the impression that data that is received in the order it's sent.
  • The first argument could have been AF_INET6 had we decided to use IPv6 addresses instead. (Other arguments are possible, but they're less common.)
  • The second argument could have been SOCK_DGRAM had we preferred to collect data packets in the order they just happen to arrive and manage missing and garbled data packets ourselves. (Other arguments are possible, though they're less common.)

Lecture 15: Network System Calls, Library Functions

Here are a few more details:

• address is declared to be of type struct sockaddr_in, since that's the data type specifically set up to model IPv4 addresses. Had we been dealing with IPv6 addresses, we'd have declared a struct sockaddr_in6 instead.
  • It's important to embed AF_INET within sin_family, since those two bytes are examined by system calls to determine the type of socket address structure.
  • The sin_port field is, not surprisingly, designed to hold the port of interest. htons—that's an abbreviation for host-to-network-short—is there to ensure the port is stored in network byte order (which is big endian order). On big endian machines, htons is implemented to return the provided short without modification. On little endian machines (like the myths), htons returns a figure constructed by exchanging the two bytes of the incoming short. In addition to htons, Linux also provided htonl for four-byte longs, and it also provides ntohs and ntohl to restore host byte order from network byte ordered figures.
• The call to connect associates the descriptor s with the host/IP address pair modeled by the supplied struct sockaddr_in *. The second argument is downcast to a struct sockaddr *, since connect must accept a pointer to any type within the entire struct sockaddr family, not just struct sockaddr_ins.  connect will return -1 (with errno set to ECONNREFUSED) if the server of interest isn't running.

Lecture 15: Network System Calls, Library Functions

Here is the full implementation of createServerSocket (and online right here):

Lecture 15: Network System Calls, Library Functions

• The call to socket is precisely the same here as it was in createClientSocket. It allocates a descriptor and configures it to be a socket descriptor within the AF_INET namespace.
• The address of type struct sockaddr_in here is configured in much the same way it was in createClientSocket, except that the sin_addr.s_addr field should be set to a local IP address, not a remote one. The constant INADDR_ANY is used to state that address should represent all local addresses.
• The bind call simply assigns the set of local IP addresses represented by address to the provided socket s. Because we embedded INADDR_ANY within address, bind associates the supplied socket with all local IP addresses. That means once createServerSocket has done its job, clients can connect to any of the machine's IP addresses via the specified port.
• The listen call is what converts the socket to be one that's willing to accept connections via accept. The second argument is a queue size limit, which states how many pending connection requests can accumulate and wait their turn to be accepted. If the number of outstanding requests is at the limit, additional requests are simply refused.