cql_builder::build fail race

[mhfrantz]

This PR contains two commits.

The first commit is a unit test that reproduces a race. One could argue whether this belongs in the unit test suite, since it actually attempts network communication. It uses a port that is expected to contain no server. Perhaps it should be off by default, or moved to integration tests, although it does not require the CCM bridge stuff.

With the unit test, I was able to generate a few different fatal errors. Running the test under valgrind revealed that there were accesses to deleted memory.

The second commit fixes the bug by using Boost enable_shared_from_this. The callbacks registered with the io_service are thus protected from the race because they contain a shared_ptr rather than a raw pointer.

The change goes further by using the same technique with all of the io_service callbacks that I could find. In fact, I think that some of the complexity of _is_disposed and the cql_trashcan could potentially be eliminated. As far as I can tell, those are mechanisms to address races between the io_service thread and sessions that are closed on the main thread.

[jstasiewicz]

Thanks that you took the time to fix this. I'll review and merge your changes soon.

[mhfrantz]

@jstasiewicz , it looks like you incorporated these ideas into your "dev" branch over the past week. Do you want me to pull out the unit test into a separate PR, or are you still deciding where it should go?

[jstasiewicz]

Your patch was fine; I thought it over and recognized that this is the right time to exterminate all raw pointers (well, most of them ATM). It was a risky challenge, so it began starting from dev branch.

Regarding the test - I'm not a Git virtuoso, so I can either put it by hand into the tests suite or merge some separate PR of current dev (which would include the test). This second option is the way to go if you have some spare time and want to see your name on the list of commiters ;).
Integration tests seem like a good place for it (indeed, we do test some really basic features there - vide exponential_policy_construction_test).

[mpenick]

Hi! Thanks for your pull request. We just released a new beta version of the driver: https://github.com/datastax/cpp-driver/tree/1.0. Work continues on that branch. I will be moving this branch to "deprecated".