In forkProcess we acquire a bunch of mutexes so make sure that when we fork the child doesn't see an inconsistent state of any global data structures. However, we do not acquire the all_tasks_mutex, which means that the child might have an inconsistent view of the all_tasks list.

https://phabricator.haskell.org/D60

Sadly, I do not have a test case illustrating that this is in fact a problem. Found this while working tracking down a deadlock in forkProcess (https://ghc.haskell.org/trac/ghc/ticket/9295).