marshal reports EOFError for I/O errors when reading from a FILE*

[serhiy-storchaka]

Bug report

The functions which read marshalled data from a FILE* never check ferror(), so a genuine I/O error is reported as if the file had simply ended.

>>> import _testcapi, os
>>> os.mkdir('d')
>>> _testcapi.pymarshal_read_object_from_file('d')
Traceback (most recent call last):
  ...
EOFError: EOF read where object expected

The real error is EISDIR. PyMarshal_ReadLongFromFile(), PyMarshal_ReadShortFromFile() and PyMarshal_ReadLastObjectFromFile() behave the same way.

Signals are not checked either, so a read interrupted by Ctrl-C is reported as EOFError instead of raising KeyboardInterrupt.

Doc/c-api/marshal.rst also documents only EOFError, ValueError and TypeError for these functions, which is inaccurate even today -- MemoryError is raised when the read buffer cannot be allocated.

Related: gh-101006 stopped EOFError from overriding an already set exception, but nothing sets OSError in the first place.

Linked PRs

• gh-155907: Raise OSError and KeyboardInterrupt when reading marshal data #155909

[serhiy-storchaka]

The writing side has the same defect, so gh-155909 fixes it too.

w_flush() and w_string() discard the result of fwrite(), so a failed or interrupted write to a FILE* is silently ignored:

>>> import _testcapi, marshal
>>> _testcapi.pymarshal_write_object_to_file(b'x' * 100000, '/dev/full', marshal.version)
>>>

PyMarshal_WriteObjectToFile() never converts the recorded WFERR_* code into an exception either, so an unmarshallable value is ignored as well, even though the function is documented to set the error indicator on failure. Only a lone TYPE_UNKNOWN byte reaches the file:

>>> _testcapi.pymarshal_write_object_to_file(object(), 'f', marshal.version)
>>> open('f', 'rb').read()
b'\xbf'