dynamic_lib is not thread safe

[alexcrichton]

I was looking back at all the recent failures on the BSD bots, and all of them were in the dynamic_lib libstd tests. If you take a look at these logs:

http://buildbot.rust-lang.org/builders/auto-bsd-64-opt/builds/1029/steps/test/logs/stdio
http://buildbot.rust-lang.org/builders/auto-bsd-64-opt/builds/1030/steps/test/logs/stdio
http://buildbot.rust-lang.org/builders/auto-bsd-64-opt/builds/1040/steps/test/logs/stdio
http://buildbot.rust-lang.org/builders/auto-bsd-64-opt/builds/1027/steps/test/logs/stdio

The problem is always that in the test_errors_do_not_crash test, the call to DynamicLibrary::open actually returns Ok when it should be returning Err. The runtime failure then happens when the destructor re-invokes fail!().

This only fails sometimes, and I'm led to believe that this is a race in the check_for_errors_in function. This function is not thread-safe in that it is possible for one thread to get the errors of another. Basically there should not be any interleaving inside of this function, but it's a very real possibility that there is.

For now I'm going to ignore the tests, but this needs a real fix at some point.

cc @sstewartgallus

[mstewartgallus]

Right. I recognized this possibility when I wrote the library but was mistaken about how to fix this. I already wrote a related issue #8156 but this issue is more specific. I plan to fix this after I finish up with the issue where the unit tests use the cos function instead of an internal function.

[alexcrichton]

Ah, closing in favor of #8156 in that case.