Nested runBlocking deadlocks when a task depends on the nested runBlocking itself

[dovchinnikov]

What do we have now?

The following deadlocks because nested runBlocking tries to process tasks from the outer runBlocking, and one of the tasks actually depends on the completion of the nested runBlocking:

import kotlinx.coroutines.*

class Container(private val scope: CoroutineScope) {

  private val instances = HashMap<Class<*>, Deferred<Any>>() // basically a map

  suspend fun <T> instance(key: Class<out T>): T {
    return instances
      .computeIfAbsent(key, ::instanceDeferred)
      .await() as T
  }

  private fun instanceDeferred(clazz: Class<*>): Deferred<Any> = scope.async {
    val constructor = clazz.getConstructor(Container::class.java)
    constructor.isAccessible = true
    constructor.newInstance(this@Container)
  }
}

class MyService(container: Container) {
  init {
    runBlocking {
      // this can ask container for an unrelated service,
      // or in this case it's enough for it to be empty
    }
  }
}

class MyService2(container: Container) {
  val service1 = runBlocking {
    container.instance(MyService::class.java)
  }
}

fun main() {
  runBlocking {
    val container = Container(this@runBlocking)
    launch {
      println(container.instance(MyService::class.java))
    }
    launch {
      println(container.instance(MyService2::class.java).service1)
    }
  }
}

What should be instead?

runBlocking should not steal tasks from outer runBlocking.

Alternatively, please provide an API to make runBlocking not steal tasks. It may be a boolean, or a special coroutine context element. Currently we have to use internals.

Why?

Because otherwise it deadlocks.

Notes

Originally, the "stealing" was added to fix a very specific use-case #860.

I believe that if nested runBlocking "steals" tasks while being inside a limited dispatcher, then this stealing logic should apply to all limited dispatchers, i.e. runBlocking on a thread inside Dispatchers.Default should steal tasks from Dispatchers.Default (#3439). If that's not possible for any reason, then nested runBlocking should not steal tasks from the outer runBlocking for the very same reason whatever it might be.

[dkhalanskyjb]

A simplified reproducer:

fun main() {
    runBlocking {
        val job = launch {
            runBlocking {
            }
        }
        launch {
            runBlocking {
                job.join()
            }
        }
    }
}

[dkhalanskyjb]

Right now, the code I show above hangs.

If we disable work stealing and let the thread do its thing inside runBlocking, we encounter the opposite problem: the following code will hang.

fun main() {
    runBlocking {
        val job = launch {
            yield() // new thing
            runBlocking {
            }
        }
        launch {
            runBlocking {
                job.join()
            }
        }
    }
}

Right now, this code passes just fine: the runBlocking on the bottom gets entered, runs the other tasks to completion, joins the job and exits. If it couldn't steal tasks, then the runBlocking on the bottom would just wait forever until someone completed the job.

In fact, in your case, just swapping the launch blocks at the bottom is enough to "fix" the problem, but this "fix" would stop working if we disabled work stealing.

What I'm getting from this is, with runBlocking, you can't avoid topologically sorting your code by the time an inner runBlocking is entered using the "depends-on" relation: whether work stealing is enabled or not simply reverses the direction of the dependency arrows.

[dovchinnikov]

Basically, yes.

[dovchinnikov]

The important thing here is a conceptual problem: should runBlocking steal tasks from a limited dispatcher, a thread of which is blocked by the said runBlocking? I believe that if runBlocking steals from outer runBlocking, then runBlocking should steal from Dispatchers.Default if it happens on a thread which holds the CPU-token, and, in general, runBlocking should steal from any limited dispatcher.