This fix for issue 20731 doesn't address this bug completely because it's possible for ftell to return -1 without an actual error, as test2.py demonstrates. 

In text mode, CRLF is translated to LF by the CRT's _read function (Win32 ReadFile). So the buffer that's used by FILE streams is already translated. To get the stream position, ftell first calls _lseek (Win32 SetFilePointer) to get the file pointer. Then it adjusts the file pointer for the unwritten/unread bytes in the buffer. The problem for reading is how to tell whether or not LF in the buffer was translated from CRLF? The chosen 'solution' is to just assume CRLF.

The example file test2.py is 33 bytes. At the time fp_setreadl calls ftell(tok->fp), the file pointer is 33, and Py_UniversalNewlineFgets has read the stream up to '#coding:latin-1\n'. That leaves 17 newline characters buffered. As stated above, ftell assumes CRLF, so it calculates the stream position as 33 - (17 * 2) == -1. That happens to be the value returned for an error, but who's checking? In this case, errno is 0 instead of the documented errno constants EBADF or EINVAL.

Here's an example in 2.7.7, since it uses FILE streams:

    >>> f = open('test2.py')
    >>> f.read(16)
    '#coding:latin-1\n'
    >>> f.tell()
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
    IOError: [Errno 0] Error

Can the file be opened in binary mode in Modules/main.c? Currently it's using `_Py_wfopen(filename, L"r")`. But decoding_fgets calls Py_UniversalNewlineFgets, which expects binary mode anyway.