On Windows, the communicate() method of subprocess.Popen is implemented with threads calling:

def _readerthread(self, fh, buffer):
    buffer.append(fh.read())
    fh.close()

where fh is one the Popen pipes: popen.stdout or popen.stderr.

For stdout=PIPE and/or stderr=PIPE, Popen._get_handles() creates a pipe with _winapi.CreatePipe(None, 0).

The fh.read() calls only completes when the write end of the pipe is closed: when all open handles of the write end of the pipe are closed.

When using Popen(stdout=PIPE), Python uses the pipe as stdout:

                startupinfo.dwFlags |= _winapi.STARTF_USESTDHANDLES
                startupinfo.hStdInput = p2cread
                startupinfo.hStdOutput = c2pwrite
                startupinfo.hStdError = errwrite

So far so good, it works well when Python spawns a single process.

--

Problems arise when the child process itself spawns new processes. In that case, the stdout pipe is inherited by other processes.

Popen.communicate() will block until all processes close the pipe.

This behavior is surprising for the following pattern:
------------------
try:
    stdout, stderr = popen.communicate(timeout=5.0)
except subprocess.TimeoutExpired:
    popen.kill()
    stdout, stderr = popen.communicate()
------------------

I would expect that the second communicate() call completes immediately since we just killed the process: Windows knows that child process is dead, Python knows that Popen object has been kill, but fh.read() continues to block (until all processes closed the pipe).

I would expect communicate() to complete immediately.

The problem is that fh.read() only returns once all data is returned at once. If fh.read() call is cancelled somehow, communicate() would just return empty data: we loose all data.

--

Eryk Sun proposed two solutions:
https://bugs.python.org/issue37531#msg350246

"For Windows, subprocess could have a _read_all(file) method that special cases a pipe. The read loop for a pipe would check whether the child has exited. Then call _winapi.PeekNamedPipe on the handle (from get_osfhandle), and do a raw read of the available bytes. If the child has exited or PeekNamedPipe fails (EPIPE), then break, join the partial reads, decode and translate newlines if it's text mode, and return."

and:
https://bugs.python.org/issue37531#msg350329

"Alternatively, instead of special casing the file type and spinning on PeekNamedPipe, the workaround could be based on a multiple-object wait that includes the child process handle. In this case, communicate() would always call _communicate in Windows, regardless of the timeout or number of pipes -- because simplistically calling either self.stdout.read() or self.stderr.read() could hang. 

The implementation would be moderately complicated. If we stop waiting on the reader threads because the process exited, we can give the threads a short time to finish reading and close the files -- maybe 250 ms is enough. But if they haven't exited at this time, we can't simply raise a TimeoutExpired exception if the call hasn't actually timed out.  To finish the _communicate call, we would have to cancel the pending reads and join() the threads.

We can force a reader thread to exit by canceling the read via WINAPI CancelIoEx. However, _readerthread has to be modified to support this. It could be implemented as a loop that calls _winapi.ReadFile to read the output in chunks that get appended to the buffer list. The read loop would break for either ERROR_BROKEN_PIPE or ERROR_OPERATION_ABORTED (canceled). 

The final step in _communicate would be to concatenate the partial reads. If it's text mode, it would also have to decode the bytes and translate newlines. The POSIX implementation of _communicate has to do this, so we already have a _translate_newlines method for this case.

Note that _winapi.WaitForMultipleObjects is interruptible in the main thread via Ctrl+C, which is a bonus improvement since Thread.join() can't be interrupted in Windows."

--

This issue has been discussed in bpo-37531 about regrtest:

https://bugs.python.org/issue37531#msg350181

It impacts the Python test suite when a test uses multiprocessing (for example) which spawns new processes.