runBlocking with its default semantics may cause dispatcher starvation if it is called on a worker thread.
For example, if runBlocking happens to block all Dispatchers.Default workers, it may lead to a deadlock in the application:
there may be tasks in the CPU queue that runBlockings await, but there are none CPU workers available to run them.
This patch changes the behavior of runBlocking so that it always releases associated computation permits before it parks,
and reacquires them after unpark. It works for every CoroutineDispatcher that is built using library primitives:
plain Dispatcher.* objects or .limitedParallelism limited dispatchers that are on top of them.
This change in behavior comes with a cost. Permit reacquiring mechanism may need an additional thread park/unpark. Worker threads that release their computational permits always let go of the local task queue, which means less benefit from locality, higher contention and transactional costs at the very least.
This patch doesn't change the fact that runBlocking should still be used carefully.