Fixes for multithreaded execution

[jacquelinekay]

Known issues:

to reproduce:

./build/test_rclcpp/gtest_multithreaded__rmw_opensplice_cpp --gtest_filter=*publisher_intra_process --gtest_repeat=-1

Issues only seem to occur in Opensplice in the intra process test cases, and much more frequently in the multi_access_publisher_intra_process case.

• AnyExecutable occasionally gets garbage in a few of its shared_ptr fields.

timer: {_M_ptr = 0x106cbf8, _M_refcount = {_M_pi = 0x106cbe0}}, <No data fields>}
subscription: {_M_ptr = 0x0, _M_refcount = {_M_pi = 0x2}}, <No data fields>}
subscription_intra_process: {_M_ptr = 0x100003b31, _M_refcount = {_M_pi = 0x0}}

It looks to me like:
timer is a valid shared_ptr. The pointer value changes between runs but it always looks like a valid shared_ptr.
subscription is a null pointer with a refcount of 2 (not good). These values stay the same in every occurrence of the bug.
subscription_intra_process is an invalid pointer (value evaluates to 2^32+15153 in decimal) with no ref count. The ref count is always 0, the pointer value changes between runs (but it's always over 2^32 aka 0x100000000).

Timer is not always set, sometimes the error occurs when the any_executable already has null ptr fields and it's trying to set subscription. However, it segfaults because it sees that subscription has a non-zero refcount and tries to release already null memory.

• Random deadlock.
• Random segfaults due to ownership issues of an rmw_guard_condition (probably have to extend locking of the interrupt_guard_condition in Executor).

[esteve]

in progress. don't look here!

Too late! What should I do now?

[jacquelinekay]

Now that you've looked at this issue, you have to help me fix it! Bwahahaha!