On Windows, I am observing that if an async_read operation successfully completes on a serial port, and I immediately initiate another async_read opera
In short, the fundamental problem is either:
ReadFile API contract with timeouts is incorrectReadFile API contractThe easiest solution is to account for this behavior in the application code, and issue another async_read operation if the prior one completes with success and 0 bytes read. Depending on the communication driver's implementation, a 1 milliseconds sleep between reads may work.
The COMMTIMEOUTS documentation states that for the read interval timeout:
The maximum time allowed to elapse before the arrival of the next byte on the communications line, in milliseconds. If the interval between the arrival of any two bytes exceeds this amount, the
ReadFileoperation is completed and any buffered data is returned. [...]
Asio's interpretation of the documentation, namely the emphasized text, is that for a given ReadFile operation, the read interval timeout begins after the first byte has been read. The implication is that if ReadFile is requested to read more than 0 bytes, Asio does not expect the ReadFile operation to return a status indicating it has successfully read 0 bytes synchronously or asynchronously. With this interpretation, Asio's implementation configures the serial port with a read interval timeout of 1 millisecond1.
// Set up timeouts so that the serial port will behave similarly to a
// network socket. Reads wait for at least one byte, then return with
// whatever they have. Writes return once everything is out the door.
::COMMTIMEOUTS timeouts;
timeouts.ReadIntervalTimeout = 1;
timeouts.ReadTotalTimeoutMultiplier = 0;
timeouts.ReadTotalTimeoutConstant = 0;
The async_read is a composed operation implemented in zero or more calls to intermediate async_read_some operations. The async_read operation interprets an intermediate async_read_some operation completing with success and 0 bytes transferred as-if no further progress will be made for the composed operation, and thus the async_read operation completes. This interpretation becomes problematic when the underlying system call to ReadFile unexpectedly completes synchronously with success and 0 bytes read.
With this details, alternative solutions are to either:
ReadFile operation once the operation has read at least a single byteReadFile and finds that it only occurs if the ReadFile operation completes synchronously, then one may be able to patch the async_read_some() operation within win_iocp_handle_service::start_read_op. Otherwise, one could patch the various read_op specializations so that their completion predicate does not exit if 0 bytes have been read, but more than 0 had been requested.1. If the communication driver's implementation is allowing the read interval timeout that started at the last byte read for ReadFilen operation to affect ReadFilen+1 operation that is initiated within the read interval timeout gap, then sleeping between ReadFilen and ReadFilen+1 for the period of the read interval timeout would prevent the Asio read operations from completing with success and 0 bytes when the supplied buffer's size is greater than 0 bytes.