related: msg217596 (bufsize=1 is broken if subprocess module uses Python io)

TextIOWrapper.write behavior:

_pyio.py [1]:

        if self._line_buffering and (haslf or "\r" in s):
            self.flush()

textio.c [2]:

    if (self->write_through)
        needflush = 1;
    else if (self->line_buffering &&
        (haslf ||
         PyUnicode_FindChar(text, '\r', 0, PyUnicode_GET_LENGTH(text), 1) != -1))
        needflush = 1;

C implementation calls flush() if write_through=True, Python implementation doesn't.

[1]: http://hg.python.org/cpython/file/0b2e199ad088/Lib/_pyio.py#l1636
[2]: http://hg.python.org/cpython/file/0b2e199ad088/Modules/_io/textio.c#l1333

Thanks for the report, Akira. Feel free to submit a patch! (with a test)

I've uploaded the patch that makes C implementation behave according
to the docs like Python implementation with the corresponding tests.

Issue #21332 is a dependency for this issue: subprocess' test_universal_newlines needs to be updated to work with Python version.


For Reference
-------------

issue #12591 introduces `write_through` to support subprocess' stdin
pipe in text mode with bufsize=0 i.e., TextIOWrapper.buffer is
unbuffered (raw) and Python and C implementation behave the same in
this particular case.

C implementation (pseudo-code)::

  if self.write_through:
      flush_text_buffer()
      buffer.flush() # <-- undocumented

Python implementation::

   if self.write_through:
       pass # _pyio.TextIOWrapper.write() calls buffer.write() directly

behaves according to the current documentation [1]:

   If *write_through* is ``True``, calls to :meth:`write` are guaranteed
   not to be buffered: any data written on the :class:`TextIOWrapper`
   object is immediately handled to its underlying binary *buffer*
      
[1]: hg.python.org/cpython/file/2a56d3896d50/Doc/library/io.rst

For reference, here's how subprocess.py uses write_through [2]::

  self.stdin = io.open(pipe, 'wb', bufsize)
  if universal_newlines=True:
      self.stdin = io.TextIOWrapper(self.stdin,
          write_through=True,
          line_buffering=(bufsize == 1)) # <-- issue #21332

http://hg.python.org/cpython/rev/9ce8fa0a0899/ - introduce io.TextIOWrapper in subprocess
http://hg.python.org/cpython/rev/5cc536fbd7c1  - introduce write_through

[2]: http://hg.python.org/cpython/file/2a56d3896d50/Lib/subprocess.py#l849

The patch looks basically fine. I will make a few tweaks to the comments in the test case.

Actually, with the patch, the universal_newlines tests in test_subprocess hang (quite logically, since they lack a flush()). I will fix them as well.

New changeset 39f2a78f4357 by Antoine Pitrou in branch '3.4':
Issue #21396: Fix TextIOWrapper(..., write_through=True) to not force a flush() on the underlying binary stream.
http://hg.python.org/cpython/rev/39f2a78f4357

New changeset 37d0c41ed8ad by Antoine Pitrou in branch 'default':
Issue #21396: Fix TextIOWrapper(..., write_through=True) to not force a flush() on the underlying binary stream.
http://hg.python.org/cpython/rev/37d0c41ed8ad

Thank you for the patch!