Fix repr_traceback to work with unicode errors

[prusse-martin]

Properly handle errors with unicode messages (could cause unicode decode errors when converting to str).
Added test to check if different encodings in raised error will not cause the code to fail.

[nicoddemus]

Thanks for porting this PR from bitbucket @prusse-martin!

Could you take a look at the failing tests?

Also, please add a changelog entry and yourself to AUTHORS. 😁

[prusse-martin]

My bad. When checking travis-ci to see if some new error have been introduced got confused and ended looking at the build results of master.

[nicoddemus]

No worries!

(note, make sure to comment something here when you think the PR is ready for review again, because GitHub does not send new notification e-mails when commits are added, only for new comments on the PR).

[prusse-martin]

The tests are passing. I actually have mixed felling about the way I did it. The other possible way is to use something like:

...
try:
  msg = str(error)  # python 3 (will also work in python 2 if the error message is ascii convertible)
except UnicodeDecodeError:
  if hasattr(error, 'message'):  # if just to be safe
    msg = error.message  # no 'message' attr in python 3
    if isinstance(msg, bytes):
      msg = msg.decode(sys.getdefaultencoding(), 'replace')  # python 2.6
  else:
    msg = ''
if "maximum recursion depth exceeded" in msg:
  ...
...

Any suggestions?

[nicoddemus]

After some thought I realized that we only need to check if the exception was a "maximum recursion depth" exception, we don't need to try to come up with a generic solution to convert a solution to string first.

With that in mind how about:

# code.py
if sys.version_info[:2] >= (3, 5):  # RecursionError introduced in 3.5
    def is_recursion_error(exception):
        return isinstance(exception, RecursionError)
else:
    def is_recursion_error(exception):
        if not isinstance(exception, RuntimeError):
            return False
        try:
            return "maximum recursion depth exceeded" in str(exception)
        except UnicodeDecodeError:
            return False

class FormattedExcinfo(object):
    ...
    def repr_traceback(self, excinfo):
        traceback = excinfo.traceback
        if self.tbfilter:
            traceback = traceback.filter()
        recursionindex = None
        if is_recursion_error(excinfo.value):        
            recursionindex = traceback.recursionindex()
        ...

The line return "maximum recursion depth exceeded" in str(exception) will always successfully check for recursion error in both Python 2 and 3 because the error message for the recursion will always be the same, ascii-compatible, message.

What do you think?

[nicoddemus]

Please add a reference back to the PR, something along the lines:

* Fix maximum recursion depth detection when raised error class is not aware
  of unicode/encoded bytes. Thanks `@prusse-martin`_ for the PR (`#1506`_). 

.. _#1506: https://github.com/pytest-dev/pytest/pull/1506
.. _@prusse-martin: https://github.com/prusse-martin

[prusse-martin]

I thought something like your proposed code. I just felt a bit awkward about explicitly checking the version at that moment.

[nicoddemus]

I would say that's OK, specially because RecursionError was included in 3.5 only.

[prusse-martin]

Done. The coverage results got down a bit due the version dependent implementation (the tests for coverage have been run with python 3.4).

[nicoddemus]

Thanks a lot Martin!