SerialPortStream is an independent implementation of
System.IO.Ports.SerialPort
and SerialStream
for better reliability and
maintainability, and now for portability to Mono on Linux systems.
The SerialPortStream
is a ground up implementation of a Stream
that buffers
data to and from a serial port. It uses low level Win32API (on Windows) for
managing events and asynchronous I/O, using a programming model as in the MSDN
PipeServer
example.
On Linux, it uses a support library to interface with Posix OS calls for an event loop.
These notes are for version 3.x, which is a refactoring of v2.x for better maintainability. See the end of these notes for differences.
- 1.0 Why another Serial Port implementation
- 2.0 Goals
- 2.1 Issues with MS Serial Port
- 2.2 Differences to the MS Serial Port
- 3.0 System Requirements
- 3.1 Testing
- 3.2 Compatibility
- 3.2.1 .NET Frameworks (Windows)
- 3.2.2 Mono Framework (Linux Only)
- 4.0 Installation
- 4.1 Windows
- 4.2 Linux
- 5.0 Extra Features
- 5.1 Reading and Writing - Buffering
- 6.0 Developer Notes
- 6.1 Logging
- 6.1.1 The LogSource abstraction
- 6.1.2 .NET Framework (.NET 4.0 to .NET 4.8)
- 6.1.3 .NET Core
- 6.1.3.1 Dependency Injection
- 6.1.3.2 Singleton via LogSource
- 6.1 Logging
- 7.0 Known Issues
- 7.1 General Issues
- 7.1.1 ReadTo
- 7.2 Windows
- 7.2.1 Driver Specific Issues on Windows
- 7.2.1.1 Flow Control
- 7.2.1.2 BytesToWrite
- 7.2.1 Driver Specific Issues on Windows
- 7.3 Linux
- 7.3.1 Mono on non-Windows Platforms
- 7.3.2 Driver Specific Issues on Linux
- 7.3.2.1 Parity Errors
- 7.3.2.2 Garbage Data on Open
- 7.3.2.3 Monitoring Pins and Timing Resolution
- 7.3.2.4 Close Times with Flow Control
- 7.3.2.5 Opening Ports (and some unit test case failures)
- 7.4 Guidelines on Serial Protocols
- 7.1 General Issues
Microsoft and Mono already provided a reasonable implementation for accessing the serial port. Today the main goal is to provide a buffered solution that can be used on various operating systems, the the ability to also abstract hardware. Along the way, various issues with the original implementation in .NET Framework are resolved in this library (see the next section).
This project tries to achieve the following:
- An implementation similar to the MS implementation of
SerialPort
. It's not meant to be 100% compatible, but instead provide similar functionality - Abstract the driver implementation and provide for a more reliable transport,
by making writing serial data completely buffered. With the MS implementation,
one can write data, but subsequently needs to check if all data is written or
not. If it isn't written, then it needs to be retried. The
SerialPortStream
makes this easier. - Provide for reliable and consistent behaviour. See the next section.
The SerialPortStream
tries to solve the following issues observed:
- Zach Saw describes issues regarding behaviour of the
fAbortError
flag in the SerialDCB
. TheSerialPortStream
defines this flag. - Closing a serial port, then reopening it generally causes problems. The
SerialPortStream
shouldn't have this issue. Note, there are some cases observed where the Operating System hangs, and this can't be avoided. - The
ReadTo()
implementation can subtly change the byte stream buffer, when one switches from characters to bytes. This problem occurs because the MS implementation actually converts the characters back to bytes into its buffer. So if you have UTF8, decoded some invalid characters and then have a timeout, this results in the invalid characters being converted back to bytes, resulting in "lost" data. I take some care when decoding bytes to characters to ensure a seamless and accurate transition between bytes and characters. Write()
gives the data to the serial port. If the operation is asynchronous, the call back results in the number of bytes that were actually transferred to the driver. You need to check yourself if this is valid or not. In the synchronous case, the data is simply thrown away. TheSerialPortStream
method simply copies data to a local buffer and uses asynchronous writes in a different thread. It works in the background to send out the data you provided. If the data can't be sent, then you get aTimeoutException
without any data being buffered at all. So you can implement reliable protocols and your code is simpler.- Disposing or Closing the serial port during a blocking write operation will
not abort the write operation. This implementation will abort with an
System.IO.IOException
type.
The goal is to provide a Stream
, not an API compatible replacement to the
SerialPort
.
All data is buffered internally in memory, captured using an I/O thread. The extra buffering adds delays by reading the bytes, then performing a context switch for the user code to read the buffer. This can slow down your software.
Buffering solves the problem however, that data is read from the serial port in an arbitrary sized memory buffer, and not dependent on the driver, so a likelihood of driver underruns and overruns are reduced. This was an important aspect when writing this library.
Software has been tested and developed using:
- .NET Standard 2.1 on Windows 10 Pro x64, .NET SDK 5.x
- .NET 4.8 on Windows 10 Pro x64
- Mono 6.x from Xamarin on Ubuntu 20.04 (64-bit)
See later in these notes for known issues and changes.
The software is written originally for .NET 4.0 and should work on those platforms. It is extended for .NET 4.5 features. A version targets .NET Core with API level .NET Standard 2.1, so should work on .NET Core 2.1, 3.1 and .NET 5.0 and later.
Windows XP SP3 and later should work when using .NET 4.0. It's not possible to run the unit tests on Windows XP since the unit tests have migrated to NUnit 3.x, but was working fine prior to that with NUnit 2.x.
The SerialPortStream
should work on Linux, and it should be possible to import
the assembly into your code when running on Linux.
When using the Mono Framework, you should reference the .NET 4.0 or .NET 4.5 projects.
It has been tested to compile and unit test cases pass with the dotnet
command
on Linux.
On Windows, just reference the assembly in your project installing the NuGet version.
You first need to compile the support library libnserial.so
for your platform.
To do that, you'll need a compiler (e.g. GCC 4.8 or later) and cmake
. The
binaries for Linux are not part of the distribution, as it's operating system
specific.
After cloning the repository, execute the following:
git clone https://github.com/jcurl/serialportstream.git
cd serialportstream/dll/serialunix
./build.sh
Binaries are built and put in the bin
folder from where you ran the build
script. You can add a reference to LD_LIBRARY_PATH
to the library:
export LD_LIBRARY_PATH=`pwd`/bin/usr/local/lib:$LD_LIBRARY_PATH
and then run your Mono program from there.
Or you can build and install in your system:
cd serialportstream/dll/serialunix
mkdir mybuild
cd mybuild
cmake .. && make
sudo make install
The following features are in addition to the System.IO.Ports.SerialPort
implementation:
- You can obtain the RingIndicator pin status.
- The
Read()
andWrite()
buffers are completely independent of the low level Windows driver.- For those concerned, the buffering means that a copy must always be made on
every
Read()
andWrite()
method.
- For those concerned, the buffering means that a copy must always be made on
every
Why is it interesting to perform buffering? A driver might be configured to be 4096 or 8192 bytes (which is quite typical). Testing with the PL2303 chipset, one can't write more than about 12KB with a single write operation.
A Write buffer may be 128KB, which one writes to. The thread in the background will write the data and issue as many write calls as is necessary to get the job done. A Read buffer may be 5MB. The background thread will read from the serial port when ever data arrives and buffers into the 5MB.
So long as the I/O thread in .NET can execute every 100-200ms, it can continue to read data from the driver. Your own application doesn't need to keep to such difficult time constraints. Such issues typically arise in Automation type environments where a computer has many different peripherals. So long as the process doesn't block, your main application might sleep for 10 seconds and you've still lost no data. The MS implementation wouldn't be so simple, you have to make sure that you perform frequent read operations else the driver itself might overflow (resulting in lost data).
If you come across a problem using this library, you may be asked to provide additional debug logs. This section describes how to obtain those logs for .NET Framework and .NET Core.
Logging for SerialPortStream 3.x uses my RJCP.Diagnostics.Trace
library that
provides an implementation called LogSource
. This is a wrapper around
TraceSource
, and can provide where necessary a TraceListener
for .NET Core.
For .NET Core, it provides a factory method as a singleton as an alternative to
dependency injection. The following sections provide more details.
The library uses the TraceSource
object, so you can add tracing to your
project in the normal way. You should use the switch name
RJCPIO.Ports.SerialPortStream
. An example of an app.config
file that you
can use to enable logging:
<?xml version="1.0" encoding="utf-8" ?>
<configuration>
<system.diagnostics>
<sources>
<source name="RJCP.IO.Ports.SerialPortStream" switchValue="Verbose">
<listeners>
<clear/>
<add name="myListener"/>
</listeners>
</source>
</sources>
<sharedListeners>
<add name="myListener" type="System.Diagnostics.TextWriterTraceListener" initializeData="logfile.txt"/>
</sharedListeners>
</system.diagnostics>
</configuration>
Please note, for SerialPortStream 3.x and later, the name of the trace source has changed to include the full namespace, to be compatible with my other projects.
.NET Core has an implementation of TraceListener
and TraceSource
, but it
doesn't load the app.config
on start up, nor provide a singleton for
applications to use for tracing. There are two ways to enable logging for
SerialPortStream
on .NET Core.
The SerialPortStream
has a constructor where you can provide your ILogger
object for tracing.
Internally, the ILogger
is wrapped around a minimal TraceListener
implementation to keep the code common between .NET Framework and .NET Core.
When upgrading from .NET Framework to .NET Core, it can be quite difficult to refactor software from using a singleton pattern, to dependency injection, and may require large swaths of code to be refactored.
To avoid this, the LogSource
classes provide a mechanism to get an ILogger
using a factory method you provide.
You may add in your code the following:
using Microsoft.Extensions.Logging;
using RJCP.CodeQuality.NUnitExtensions.Trace;
using RJCP.Diagnostics.Trace;
internal static class GlobalLogger
{
static GlobalLogger() {
ILoggerFactory factory = LoggerFactory.Create(builder => {
builder
.AddFilter("Microsoft", LogLevel.Warning)
.AddFilter("System", LogLevel.Warning)
.AddFilter("RJCP", LogLevel.Debug)
.AddNUnitLogger();
});
LogSource.SetLoggerFactory(factory);
}
public static void Initialize() {
/* Intentially empty. By calling this method, the static constructor
will be automatically called */
}
}
The example above is from the unit test cases for SerialPortStream
, but can
be easily adapted for your own projects. The important part is that:
- In your production code, you assign the static factory with
LogSource.SetLoggerFactory()
, which is not called as part of your unit tests. It will likely be different to the example given above, as you'd want to instead, get the logging configuration (log levels) via a configuration file. - Your test code would set the
LogSource.SetLoggerFactory()
similarly as in the example provided above (the example given above is good for NUnit 3.x with the.AddNUnitLogger()
provided by myRJCP.CodeQuality
library). - Setting the factory is only needed for .NET Core. On .NET Framework, the
TraceSource
class provides the singleton functionality and will instantiate for you the correctTraceSource
from the application configuration file.
The code works by requesting to the ILoggerFactory.CreateLogger
with the name
RJCP.IO.Ports.SerialPortStream
. This is also the motivation why the trace
source was changed to include the full namespace. The code above shows that it
will create this object and log all debug level events to the NUnit logger.
The implementation of ReadTo
and other character based read events with the
SerialPortStream
is slow. It tries to calculate the size (in bytes) of each
individual character, in case the user decides to read bytes in-between. The
purpose of this was to prevent possible data loss in when a read of bytes is
mixed with a read of characters. It's recommended not to use the character based
APIs.
Some test cases are failing and needs to be investigated (although it's not clear if this is a bug in the library or in the test case).
The following issues are known:
- This is not an issue in the library, but when using the
Com0Com
for running unit tests, some specific test cases for Parity will fail. That is because Com0Com doesn't emulate data at a bit level. Using real serial hardware with a NULL modem adapter works as expected. - .NET 4.0 to .NET 4.8 and .NET Core has a minor issue in
System.Text.Decoder
that in a special circumstance it will consume too many bytes. ThePeekChar()
method is therefore a slower implementation than what it could be. Please refer to the Xamarin bug 40002. Found against Mono 4.2.3.4 and later tested to be present since .NET 4.0 on Windows XP also.
Using the FTDI chipset on Windows 10 x64 (FTDI 2.12.16.0 dated 09/Mar/2016) flow
control (RTS/CTS) doesn't work as expected. For writing small amounts of data
(1024 bytes) with CTS off, the FTDI driver will still send data. See the test
case ClosedWhenFlushBlocked
, change the buffer from 8192 bytes to 1024 and the
test case now fails. This problem is not observable with com0com 3.0. You can
see the effect in logs, there is a TX-EMPTY event that occurs, which should
never be there if no data is ever sent.
On Windows, the SerialPortStream
returns the bigger of either the internal
write buffer, or the amount of data in the output queue of the driver. Drivers
don't report the number of bytes that are in the output queue before the next
write begins, and may return sooner. This leads to the effects:
This driver indicates more bytes are in the output queue than what it will return from the current ongoing write operation. This can cause some jumps in the returned value.
CP210x Universal Windows Driver v10.1.10 1/13/2021.
For example:
BytesToWrite = 40960 (driver 12288)
RJCP.IO.Ports.SerialPortStream Verbose: 0 : COM5: SerialThread: ProcessWriteEvent: 1024 bytes
BytesToWrite = 40412 (driver 40412)
RJCP.IO.Ports.SerialPortStream Verbose: 0 : COM5: SerialThread: DoWriteEvent: WriteFile(736, 312385272, 39936, ...) == False
BytesToWrite = 40387 (driver 40387)
BytesToWrite = 40386 (driver 40386)
The internal buffer is 40kB, the driver returned it wrote 1024 bytes, but the queue still has 40412 bytes (which is more than the 39936 bytes it should be).
It can also fluctuate without writes without calls to the OS in between.
BytesToWrite = 40418 (driver 40418)
BytesToWrite = 40393 (driver 40393)
BytesToWrite = 40392 (driver 40392)
BytesToWrite = 40391 (driver 40391)
BytesToWrite = 40390 (driver 40390)
BytesToWrite = 40389 (driver 40389)
BytesToWrite = 40428 (driver 40428)
BytesToWrite = 40427 (driver 40427)
BytesToWrite = 40426 (driver 40426)
Generally this driver reports that it has zero bytes in the output queue, but
may sometimes report the number of bytes in the last WriteFile()
call. This
is not a problem, but the number of bytes in the output queue is less than what
is still to be written, so a user may think it is complete, when it is not.
The SerialPortStream
was tested on Ubuntu 14.04 to 20.04. Feedback welcome for
other distributions!
The main functionality on Linux is provided by a support C library that
abstracts the Posix system call select()
.
Issues in the current implementation are:
- Custom baud rates are not supported. To know what baud rates are supported on
your system, look at the file
config.h
after building. - DSR and DTR handshaking is not supported. You can still set and clear the pins though.
Patches are welcome to implement these features!
Use the currently supported versions of Mono provided by the Mono project for your Linux distribution. For example, Ubuntu 14.04 ships with Mono 3.2.8 which is known to not work.
- [Mono-Dev] Mono 3.2.8 incompatibility with .NET 4.0 on Windows 7-10. The System.Text implementation for converting bytes to UTF8 don't work. If you don't use the character based methods, it may work. But the software has not been tested against this framework.
- The
DataReceived
event doesn't fire for the EOF character (0x1A). On Windows it does, as this is managed by the driver itself and not emulated by the C-Library. - Linux doesn't implement DSR.
Tests have been done using FTDO, PL2303H, PL2303RA and 16550A (some still do exist!). The following has been observed:
Some chipsets do not report properly parity errors. The 16550A chipset works as
expected. Issues observed with FTDI, PL2303H, PL2303RA. In particular, on a
parity error, more bytes are reported as having parity errors than there are in
the stream. Tested using loopback devices with comptest
.
$ ./nserialcomptest /dev/ttyUSB0 /dev/ttyUSB1`
[ RUN ] SerialParityTest.Parity7E1ReceiveError
/home/jcurl/Programming/serialportstream/dll/serialunix/libnserial/comptest/SerialParityTest.cpp:221: Failure
Value of: comparison
Actual: false
Expected: true
Unexpected byte received with Even Parity
[ FAILED ] SerialParityTest.Parity7E1ReceiveError (585 ms)
[ RUN ] SerialParityTest.Parity7O1ReceiveError
/home/jcurl/Programming/serialportstream/dll/serialunix/libnserial/comptest/SerialParityTest.cpp:373: Failure
Value of: comparison
Actual: false
Expected: true
Unexpected byte received with Even Parity
[ FAILED ] SerialParityTest.Parity7O1ReceiveError (584 ms)
[ RUN ] SerialParityTest.Parity7O1ReceiveErrorWithReplace
/home/jcurl/Programming/serialportstream/dll/serialunix/libnserial/comptest/SerialParityTest.cpp:427: Failure
Value of: comparison
Actual: false
Expected: true
Unexpected byte received with Even Parity
[ FAILED ] SerialParityTest.Parity7O1ReceiveErrorWithReplace (572 ms)
On Linux Kernel with Ubuntu 14.04 and Ubuntu 16.04, we observe that some USB-SER
drivers provide extra data depending on what a previous process was doing. It
shows itself as garbage zero's appearing at the beginning of a stream when
reading, and may be visible in your application also. There's a test case
comptest/kernelbug
that shows this behaviour on a Lenovo T61p. Affected is
PL2303H and FTDI chipsets. Chipsets that don't show this behaviour are 16550A
and PL2303RA chipsets. Invocate the test program twice and you'll see the error.
This is reported to
Ubuntu
$ kernelbug /dev/ttyUSB0 /dev/ttyUSB1
Offset: 4
Flushing...
Writing Complete...
Reading complete...
Comparison MATCH <---- PASS
Flushing...
Reading complete...
Complete...
$ kernelbug /dev/ttyUSB0 /dev/ttyUSB1
Offset: 108
Flushing...
Flush 2 bytes
Writing Complete...
Reading complete...
ERROR: Comparison mismatch <---- ERROR
Flushing...
Flush 510 bytes
Reading complete...
Complete...
Monitoring of pins CTS, DSR, RI and DCD is not 100% reliable for some chipsets
and workarounds are in place. In particular, the chips PL2303H, PL2303RA do not
support the ioctl(TIOCGICOUNT)
, so on a pin toggle, we cannot reliably detect
if they have changed if the pulse is too short. For 16550A and FTDI chips, this
ioctl()
does work and so we can always detect a change. To check if your
driver supports the ioctl(TIOCGICOUNT)
call, run the small test program
comptest/icount
.
$ ./icount /dev/ttyS0
Your driver supports TIOCGICOUNT
ocounter.cts=0
ocounter.dsr=0
ocounter.rng=3
ocounter.dcd=0
or in the case it's not supported:
$ ./icount /dev/ttyUSB0
Your driver doesn't support TIOCGICOUNT
Error: 25 (Inappropriate ioctl for device)
Some times closing the serial port may take a long time (observed from 5s to
21s) if it is write blocked due to hardware flow control. In particular, the
C-library function serial_close()
appears to take an excessive time when
calling close(handle->fd)
on Ubuntu 16.04. This issue appears related to the
Linux driver and not the MONO framework.
The .NET Test Cases that show this behaviour are (blocking on write):
ClosedWhenBlocked
CloseWhenFlushBlocked
DisposeWhenBlocked
DisposeWhenFlushBlocked
This issue is not reproducible with the 16550A UART when it is write blocked. In this case, the times for closing are usually not more than 20ms.
When testing continuously to open a port, send data, and then receive on another port using a NULL-modem cable, minor issues can occur that result in test case failures. These issues would also be visible in real-world programs and are driver dependent.
The test scenario is on Linux (Ubuntu 20.04) with various USB serial port
devices. The test case ReadToWithMbcs
from SerialPortStreamNativeTest is
modified to run 2000 times with [Repeat(2000)]
. The command to run the test
after building is then:
dotnet test RJCP.SerialPortStreamNativeTest.dll --filter Name=ReadToWithMbcs
(please note, not only this test case is affected, but it is easy to reproduce.)
- FTDI: After opening the serial port, in about 1% of the cases, data is not
sent (observed using .NET Mono 4.0 and 4.5). It is confirmed that the system
call
write()
was called, and all data was given to the kernel via the librarylibnserial
. However, on the other serial port, data is never received. Waiting 15ms after opening would resolve the problem - this workaround will not be part of theSerialPortStream
as it appears to be very specific to this driver and similar behaviour is not observed on other drivers. - PL2303H: Sometimes on connecting the serial port, a spurios 0xFF is sent. This causes the test case to fail, as data that was not sent is received and affects the output.
- PL2303RA: The test cases appear to run about 10x slower than any other driver, but no errors were observed.
- CP2101: Seems to work flawlessly.
Given the issues listed in this section, one can come up with the following recommendations for protocol design over the serial port:
- Assume that at Layer 2 (the serial port bus), data can be inserted, modified, or deleted.
- Define data as frames. There should be a marker byte indicating the start of a frame, a length to know how much data should be received, and a checksum (at least a CRC16) that can be used to check the integrity of the frame.
- Define the protocol to be able to resend data in case of lost data if needed, or can continue if data is lost.
I recommend to not use hardware or software flow control, but define in the serial protocol frames, like a link control protocol (LCP) that can manage this. Do not use parity, and instead opt to use checksum bytes within a frame.
- Hardware flow control can lead to deadlocks in software. No flow control just means data can be lost, and can be replaced using a LCP. Software flow control can also cause complications in the protocol, which can be more generically handled using a LCP.
- Parity can insert arbitrary bytes and corrupted data, especially with USB serial devices. Use frame checksums (FCS) instead.
Allow bundling of frames one after the other, and decode separately. Lots of
small frames that need to be acknowledged can lead to delays between frames, and
longer transmission times for an already "slow" bus speed. The
SerialPortStream
is buffered, so performance is impacted by lots of context
switches between sending data, and waiting for a response, as there is a buffer
thread used in-between.