If you *must* use .NET System.IO.Ports.SerialPort

Microsoft .NET FrameworkAs an embedded developer who writes desktop software mostly for configuration of, and data download from, peripheral devices, I use serial data streams a lot.  Mostly USB virtual serial posts from FTDI, but also the USB Communication Device Class and real 16550-compatible UARTs on the PCI bus.  Since looking at data through an in-circuit emulator debug interface is generally a miserable experience, getting serial data communication with a custom PC application is essential to analyzing data quality and providing feedback on hardware designs.  C# and the .NET Framework provide a rapid application development that is ideal for early development that needs to track changing requirements as hardware designs evolve.  Ideal in most respects, I should say.

The System.IO.Ports.SerialPort class which ships with .NET is a glaring exception.  To put it mildly, it was designed by computer scientists operating far outside their area of core competence.  They neither understood the characteristics of serial communication, nor common use cases, and it shows.  Nor could it have been tested in any real world scenario prior to shipping, without finding flaws that litter both the documented interface and the undocumented behavior and make reliable communication using System.IO.Ports.SerialPort (henceforth IOPSP) a real nightmare.  (Plenty of evidence on StackOverflow attests to this, from devices that work in Hyperterminal but not .NET because IOPSP makes setting certain parameters mandatory, although they aren’t applicable to virtual ports, and closes the port on failure.  There’s no way to bypass or ignore failure of these settings during IOPSP initialization.)

What’s even more astonishing is that this level of failure occurred when the underlying kernel32.dll APIs are immensely better (I’ve used the WinAPI before working with .NET, and still do when I want to use a function that .NET doesn’t have a wrapper for, which notably includes device enumeration).  The .NET engineers not only failed to devise a reasonable interface, they chose to disregard the WinAPI design which was very mature, nor did they learn from two decades of kernel team experience with serial ports.

A future series of posts will present the design and implementation of a rational serial port interface built upon, and preserving the style of, the WinAPI serial port functions.  It fits seamlessly into the .NET event dispatch model, and multiple coworkers have expressed that it’s exactly how they want a serial-port class to work.  But I realize that external circumstances sometimes prohibit using a C++/CLI mixed-mode assembly.  The C++/CLI solution is incompatible with:

  • Partial trust (not really a factor, since IOPSP’s Open method also demands UnmanagedCode permission)
  • Single-executable deployment (there may be workarounds involving ILMerge or using netmodules to link the C# code into the C++/CLI assembly)
  • Development policies that prohibit third-party projects
  • .NET Compact Framework (no support for mixed-mode assemblies)

The public license (as yet undetermined) might also present a problem for some users.

Or maybe you are responsible for improving IOPSP code that is already written, and the project decision-maker isn’t ready to switch horses.  (This is not a good decision, the headaches IOPSP will cause in future maintenance far outweigh the effort of switching, and you’ll end up switching in the end to get around the unfixable bugs.)

So, if you fall into one of these categories and using the Base Class Library is mandatory, you don’t have to suffer the worst of the nightmare.  There are some parts of IOPSP that are a lot less broken that the others, but that you’ll never find in MSDN samples.  (Unsurprisingly, these correspond to where the .NET wrapper is thinnest.)  That isn’t to say that all the bugs can be worked around, but if you’re lucky enough to have hardware that doesn’t trigger them, you can get IOPSP to work reliably in limited ways that cover most usage.

I planned to start with some guidance on how to recognize broken IOPSP code that needs to be reworked, and thought of giving you a list of members that should not be used, ever.  But that list would be several pages long, so instead I’ll list just the most egregious ones and also the ones that are safe.

The worst offending System.IO.Ports.SerialPort members, ones that not only should not be used but are signs of a deep code smell and the need to rearchitect all IOPSP usage:

  • The DataReceived event (100% redundant, also completely unreliable)
  • The BytesToRead property (completely unreliable)
  • The Read, ReadExisting, ReadLine methods (handle errors completely wrong, and are synchronous)
  • The PinChanged event (delivered out of order with respect to every interesting thing you might want to know about it)

Members that are safe to use:

  • The mode properties: BaudRate, DataBits, Parity, StopBits, but only before opening the port. And only for standard baud rates.
  • Hardware handshaking control: the Handshake property
  • Port selection: constructors, PortName property, Open method, IsOpen property, GetPortNames method

And the one member that no one uses because MSDN gives no example, but is absolutely essential to your sanity:

  • The BaseStream property

The only serial port read approaches that work correctly are accessed via BaseStream.  Its implementation, the System.IO.Ports.SerialStream class (which has internal visibility; you can only use it via Stream virtual methods) is also home to the few lines of code which I wouldn’t choose to rewrite.

Finally, some code.

Here’s the (wrong) way the examples show to receive data:

port.DataReceived += port_DataReceived;

// (later, in DataReceived event)
try {
    byte[] buffer = new byte[port.BytesToRead];
    port.Read(buffer, 0, buffer.Length);
    raiseAppSerialDataEvent(buffer);
}
catch (IOException exc) {
    handleAppSerialError(exc);
}

Here’s the right approach, which matches the way the underlying Win32 API is intended to be used:

byte[] buffer = new byte[blockLimit];
Action kickoffRead = delegate {
    port.BaseStream.BeginRead(buffer, 0, buffer.Length, delegate (IAsyncResult ar) {
        try {
            int actualLength = port.BaseStream.EndRead(ar);
            byte[] received = new byte[actualLength];
            Buffer.BlockCopy(buffer, 0, received, 0, actualLength);
            raiseAppSerialDataEvent(received);
        }
        catch (IOException exc) {
            handleAppSerialError(exc);
        }
        kickoffRead();
    }, null);
}
kickoffRead();

It looks like a little bit more, and more complex code, but it results in far fewer p/invoke calls, and doesn’t suffer from the unreliability of the BytesToRead property.  (Yes, the BytesToRead version can be adjusted to handle partial reads and bytes that arrive between inspecting BytesToRead and calling Read, but those are only the most obvious problems.)

Starting in .NET 4.5, you can instead call ReadAsync on the BaseStream object, which calls BeginRead and EndRead internally.

Calling the Win32 API directly we would be able to streamline this even more, for example by reusing a kernel event handle instead of creating a new one for each block.  We’ll look at that issue and many more in future posts exploring the C++/CLI replacement.

22 Comments

22 Responses to If you *must* use .NET System.IO.Ports.SerialPort

  1. Bob Hourigan says:

    Couldn’t agree with you more about the System.IO.Ports functionality. I tried using it with some success then discovered little quirky problems that you can’t put your finger on, i.e. extreme overrun when using Xon/Xoff handshake. I tried using the built in handshake functions and my own to no avail. Also noted that the reading of the hardware I/O input pins was also unreliable. I am now using the Marshall software libraries that uses the API. Their interface with the Visual Studio environment isn’t great but I have working reliable Comms and will fix the interface when time permits.

    • Ben Voigt says:

      Bob, are you using a publicly available software library for wrapping the Win32 serial port API? Or did you mean P/invoke and Microsoft’s System.Runtime.InteropServices.Marshal class?

  2. Kris Janssen says:

    Dear Ben,

    I eventually stumbled on this post by reading many of your answers on StackOverflow. Clearly you know COM :)

    Until now I have often used a particular implementation of IOPSP that is available in the “Termie” project on CodeProject (http://www.codeproject.com/Articles/23656/Termie-A-Simple-RS-Terminal).

    But as it also relies on Read() I have indeed found it to be quite unreliable at times.

    Rolling my own WinAPI wrapper would probably take me too long given my limited experience so I am quite interested to learn more on your take on this approach!

  3. Kris Janssen says:

    Bob might be referring to a commercial library indeed (http://www.marshallsoft.com)

  4. Jon Yutkowitz says:

    Ben, I am currently in the process of porting a VB6 serial port app to .NET and have found the performance to be inadequate. The app sends a series of messages with responses, before posting the next message. The intermessage gap is several milliseconds larger for the .NET version.

  5. Jon Yutkowitz says:

    My app is using the DataReceived event, as well as Read(), Write(), and BytesToRead. I am curious if I will see better performance using BaseStream ReadAsync() but not exactly clear on how to handle new data arriving as efficiently as possible.

    • Ben Voigt says:

      Reading from the BaseStream will get the data to your application with low overhead. But that’s generally a small part of the processing cost; ultimately how efficiently your serial processing is depends on your code that buffers, packetizes, and parses it.

  6. WB says:

    Could you expand upon line 8 in the recommended approach? Should it work like the IOSPS DataReceived event or something else entirely?

    • Ben Voigt says:

      At a high level it is a similar concept to the DataReceived event because it allows the application to respond to incoming data in an event-driven fashion. Important differences are that the event actually carries the received data with it, so the application is all set for processing serial data, and the implementation does this with a single API call, which is far lower overhead.

  7. WB says:

    Could you add more context? I get a compile error on line 13: use of unassigned local variable. I’m using VS 2010 with .Net 4.0.

    BTW, in first comment, IOSPS should be IOPSP.

    • Ben Voigt says:

      That looks like a limitation in the compiler’s dataflow analysis; the variable is definitely assigned before use. Initialization by itself isn’t adequate, if the initializer expression passes the variable by reference to another function. But this code doesn’t, it only constructs a delegate using the variable; it doesn’t actually execute it.

      In any case you can work around it by splitting up initialization and delegate creation.

      Action kickoffRead = null;
      kickoffRead = delegate {

  8. Marya says:

    Thanks Ben for all the explanation. Do you have any sample code on how to use “ReadAsync” ? I do not seem to find a good document on how to use it.

    Thanks

  9. dominikjeske says:

    Can You give full example on sending a message and receiving a response using BaseStream. I have tried this but it is not working for me? Sample would be very helpful

    • Marya says:

      This might be useful:

      for writing, you can use
      await sp.BaseStream.WriteAsync(buffer ,0, buffer.Length);

      for reading,
      byte[] buffer = new byte [7];
      int read = 0;
      while (read < 7)
      read += await sp.BaseStream.ReadAsync(buffer , read, 7 – read);

      these commands should be used within a function that uses "async" as a modifier. For example:
      private async void DoSomething()
      {
      // Read or write like explained earlier
      }

    • dominikjeske says:

      Unfortunately I have to use .NET 4.0. I’m looking for some full example from send to receive an answer. I have sent some data but I have timeout when I’m waiting for response.

    • Ben Voigt says:

      If you ask a question on StackOverflow you are welcome to bring it to my attention here. But the blog platform is not designed for Q&A. Besides, at SO you’ll get input from multiple experts. To address your particular case, appropriate timeouts are specific to the device you’re communicating with; I could give sample code that works perfectly in my environment and you could still get timeouts.

  10. Marya says:

    Is there any way the we read from the buffer without knowing how long is the buffer? Thanks

    • Ben Voigt says:

      Sure, if you set a read timeout of zero and issue a read request, you’ll get as much data transferred to your buffer as is already received (up to the limit of the size you made your buffer). This is a much better approach than querying the buffer size. In general, just getting the data is more efficient and involves fewer race conditions than asking questions about the data and then retrieving it.

  11. Marya says:

    As far as I know, readAsync or WriteAsync does not support Readtimeout or Writetimeout. Even creating a manual timeout and setting a condition would not work since await operator is sitting there to get more data, although reading task is complete.

    • Ben Voigt says:

      “As far as I know” leads to false assumptions. You can avoid bad assumptions by either testing, or inspecting the implementation logic (JetBrains dotPeek and Red Gate .NET Reflector are good options here). It seems like you did neither.

      There are no serial port functions in the Win32 API that don’t respect timeout. The only way the .NET wrapper would not finish on timeout is if it detected a partial read and looped to read the rest. But I used dotPeek to verify that there is no such logic.

      Stream’s ReadAsync is just a wrapper around BeginRead/EndRead, and that returns once a Win32 ReadFile operation completes. There’s no loop to fill the entire buffer. A timeout will cause the ReadFile operation to complete, which means SerialStream.EndRead is called, and Stream.ReadAsync’s task completes also.

      Now, it is true that SerialStream.ReadTimeout is not as flexible as Win32 SetCommTimeouts. But it does support the desired combination “A (ReadIntervalTimeout) value of MAXDWORD, combined with zero values for both the ReadTotalTimeoutConstant and ReadTotalTimeoutMultiplier members, specifies that the read operation is to return immediately with the bytes that have already been received, even if no bytes have been received.” You get that from setting SerialStream.ReadTimeout=0.

  12. Alex says:

    hi Ben, do you know how to properly use SerialPort using .NET compact framework? MSDN said that “The .NET Compact Framework does not support the asynchronous model with base streams.”

Add your thoughts