Avoiding Common Networking Mistakes

When writing networking-based software, developers often make a few common design and usage mistakes that can cause serious performance problems, crashes, and other misbehavior. This chapter highlights a few of those mistakes and describes how to avoid or fix them.

Clean Up Your Connections

TCP connections remain open until either the connection is explicitly closed or a timeout occurs. Unless TCP keepalive is enabled for the connection, a timeout occurs only if there is data waiting to be transmitted that cannot be delivered. This means that if you do not close your idle TCP connections, they will remain open until your program quits.

The recommended way to enable TCP keepalive is by setting the SO_KEEPALIVE flag on the socket with setsockopt.

Avoid POSIX Sockets and CFSocket on iOS Where Possible

Using POSIX sockets directly has both advantages and disadvantages relative to using them through a higher-level API. The main advantages are:

The main disadvantages are:

The most appropriate times to use sockets directly are when you are developing a cross-platform tool or high-performance server software. In other circumstances, you typically should use a higher-level API.

Avoid Synchronous Networking Calls on the Main Thread

If you are performing network operations on your main thread, you must use only asynchronous calls.

Network communication is inherently prone to delays. For example, a DNS request that times out can take upwards of half a minute, and a connect can take even longer. If you perform a synchronous networking call—a call that waits for a response and then returns the data—the thread that made the call becomes blocked in the kernel until the operation either completes or fails. If that thread happens to be your program’s main thread, your program becomes unresponsive.

In an OS X GUI app, this causes the spinning wait cursor to appear. Menus become unresponsive, clicks and keystrokes are delayed or lost, and your users may become frustrated.

In iOS, your app is killed by the watchdog timer if it doesn’t respond to user interface events for a predetermined amount of time. The timeouts for most networking operations are longer than that of the iOS watchdog timer. Thus, your app is guaranteed to be killed if your network connection fails during a synchronous networking call.

If your iOS app generates a crash report with the exception code 0x8badfood, this means that the iOS watchdog timer killed your app because it was unresponsive; such a crash may have been caused by a synchronous networking call.

Cocoa (Foundation) and CFNetwork (Core Foundation) Code

The easiest and most common way to perform networking asynchronously is to schedule your networking object in the current thread’s run loop.

All Foundation or CFNetwork networking objects—including NSURLConnection, NSStream / CFStream, NSNetService / CFNetService, and CFSocket—have built-in run-loop integration. Each of these objects has a set of delegate methods or callback functions that are called throughout your object’s network communication.

If you want to perform a computationally intensive task (such as processing a large downloaded file) over the course of your network communication, you should do so on a separate thread. The NSOperation class lets you encapsulate such a task in an operation object, which you can easily run on a separate thread by adding it to an NSOperationQueue object.

For more information, read Run Loops in Threading Programming Guide. For sample code, see LinkedImageFetcher and ListAdder.


POSIX networking presents some unique challenges, and thus has some unique tips:

Create sockets correctly. The proper call for creating a TCP socket is:


You can create a UDP socket by replacing IPPROTO_TCP with IPPROTO_UDP in the snippet above and replacing SOCK_STREAM with SOCK_DGRAM.

If you use the select system call, keep track of which sockets are actually in use. A number of poorly written socket-based programs incorrectly assume that they own every socket or file descriptor from 3 (or 0) up through the highest-numbered socket that they currently have open, and use a simple for loop to fill in the file descriptor set. This not only can result in poor performance, but also can cause incorrect behavior if the daemon opens files that may end up with file descriptor numbers in that range.

Instead, you should keep two pieces of state in your networking code:

  • An integer that keeps track of the highest-numbered socket that you currently have open. You must update this (if needed) whenever you open or close a new socket.

  • A complete FD_SET in which the relevant bit is set for every open socket. Instead of passing this set to the select system call (which destroys the contents of any descriptor sets passed as parameters), you should copy the descriptor set with FD_COPY and pass the copy to select.

    Alternatively, if you have an array of file descriptors (or some other way to keep track of them), you can construct a new FD_SET from that array.

Use run loops and nonblocking I/O to manage asynchronous reads. In all but the most trivial command-line tools, POSIX networking should always be performed in a run-loop-based fashion using either GCD, the kqueue API, or the select system call.

Avoid synchronous networking on the main thread. If you are using synchronous calls, POSIX networking should never be performed on the main program thread of any GUI application or other interactive tool. This includes:

  • Reading from and writing to sockets that are not set to non-blocking.

  • Connecting a socket.

  • Performing DNS lookups (particularly with getaddrinfo, getnameinfo, getaddrinfo, and gethostbyaddr).

Instead, either perform the synchronous calls on a separate thread or use an API that performs these operations asynchronously, such as GCD, the kqueue API, or higher-level APIs that integrate with Cocoa run loops (CFSocket or CFStream, for example).

For more information about the options available to you, read Assessing Your Networking Needs and Using Sockets and Socket Streams.

Set appropriate timeouts if you are using select. The select system call allows you to specify a period of time to wait before returning control to your code. The value you choose for this timeout is very important:

  • If your code needs to perform other operations in the background, this timeout value determines how often those other operations run. The shorter the value, the more frequently your code can do other things in the background, but the more CPU cycles it uses while waiting.

    In general, if your app is waking up more than a few times per second (or even once per second on iOS), you should probably be doing that work on a separate thread. Moreover, this is usually a sign that you’re doing something wrong, such as polling to see whether a file has been deleted or modified instead of using a more appropriate notification-based technology such as the kqueue API.

    If you are using this wakeup to check to see if another thread within your application has done something, you should consider using a pair of connected sockets instead. To do this, first a socket pair with socketpair. Then add one of the connected sockets to your descriptor set. When the other thread needs to tell your networking thread about an event, it can wake your networking thread by writing data into the other socket. Be sure to keep track of which sockets were created in this way so that your code can handle the data differently depending on whether it came from an outside connection or from within your app.

  • If your code does not need to perform any operations in the background, you should pass NULL. Passing NULL ensures that your code takes as little CPU as possible while waiting for incoming connections or data.

In addition, if you are using timeouts, be aware that the select system call returns EINTR when the timer fires. Your code should be prepared for this return value and should not treat it as an error.

Use POSIX sockets efficiently (if at all). If you are using POSIX sockets directly:

  • Handle or disable SIGPIPE.

    When a connection closes, by default, your process receives a SIGPIPE signal. If your program does not handle or ignore this signal, your program will quit immediately. You can handle this in one of two ways:

    • Ignore the signal globally with the following line of code:

      signal(SIGPIPE, SIG_IGN);
    • Tell the socket not to send the signal in the first place with the following lines of code (substituting the variable containing your socket in place of sock):

      int value = 1;
      setsockopt(sock, SOL_SOCKET, SO_NOSIGPIPE, &value, sizeof(value));

      For maximum compatibility, you should set this flag on each incoming socket immediately after calling accept in addition to setting the flag on the listening socket itself.

  • Use nonblocking sockets where possible.

  • Keep the socket’s send buffer full to the extent possible.

  • Handle incoming data early and often to keep the socket’s receive buffer empty.

  • Support both IPv4 and IPv6.

  • Check return values from socket reads and writes.

    Be prepared to handle EAGAIN and EWOULDBLOCK errors, which indicate that no data is available when reading, or that no output buffer space is available when writing. These errors are normal, non-fatal errors; your program should not close the connection when it gets them.

For an example of POSIX code that follows these rules, see Writing a TCP-Based Server in Networking Programming Topics.

Avoid Resolving DNS Names Before Connecting to a Host

The preferred way to connect to a host is with an API that accepts a DNS name, such as CFHost or CFNetService.

Although your program can resolve a DNS name and then connect to the resulting IP address, you should generally avoid doing so. DNS lookups usually return multiple IP addresses, and (at the application layer) it is not always obvious which IP address is best for connecting to the remote host. For example:

If you cannot avoid resolving a DNS name yourself, first check to see whether the CFHost API fulfills your requirements; it provides a list of addresses for a given host instead of just a single address. If the CFHost API does not meet your needs, use the DNS Service Discovery API.

For more information, read Networking Concepts, Writing a TCP-Based Server in Networking Programming Topics, CFHost Reference, and DNS Service Discovery C Reference. For sample code, see SRVResolver.

Do Not Use NSSocketPort (OS X) or NSFileHandle for General Socket Communication

Despite its name, the NSSocketPort class (available in OS X only) is not intended for general network communication. The NSSocketPort class is part of Cocoa’s distributed objects system, which is intended for controlled communication between Cocoa applications on a single machine or on a local network. For more information on the distributed objects system, see Distributed Objects Programming Topics.

Similarly, the NSFileHandle class is not designed for general networking. The NSFileHandle class circumvents the standard networking stack, which carries the following drawbacks:

Instead, use NSStream for remote connections and CFSocket for listening. For details, see Writing a TCP-Based Server in Networking Programming Topics.