Previous debugging shows the socket accepting all the data we send it, which is just under 2MB. Maybe travis containers have very large default socket buffers? Up the data size to about 4MB.

Pre-allocate the (biggec) data before we start taking timing measurements, and be sure to close the socket when done.

Add Tox environments for running linters and include one in the default env list. Add a Tox environment for running leak tests and one for doing both, but do not run them by default. Unify the Travis and Tox configurations, and make Travis run both leak and lint tests. Switch the leak test from using gettotalrefcount (only available in debug builds) to using len(gc.get_objects()); while this is not exactly the same thing, it make be close enough, and it's avaialble to every Python developer unlike gettotalrefcount.

gevent.greenlet.joinall ignored `count` when `raise_error` was False

Should pass count param to `wait` function when `raise_error` is False.

I have spent several days to debug a wired behavior of my program caused by this bug. Hope this could save other people some time in similar situation.

Experimenting with Travis container-based builds. We lose the gettotalrefcount method running in the debug build of Python.

- ./util/cythonpp.py:19:1: E731 do not assign a lambda expression, use a def
- ./gevent/pywsgi.py:401:13: E266 too many leading '#' for block comment
- everywhere: E402 module level import not at top of file

Fix #349: threaded resolver hangs due to import lock

Patch by @ajdavis. https://github.com/gevent/gevent/issues/349#issuecomment-31665013

Close #350.

Add Python 3.4 to test set. Fixes #408. Fixes #487. The build is green for PyPy, Py26 and Py34. Py27 intermittently displays a timing issue; Py33 intermittently displays a refcount "leak".

Mark test__socket_dns as FLAKY: It works locally, but under Travis (due to different DNS configuration?) the IPV6 values differ by one for several of the hosts, and Python 3 returns those values here.