Inconsistent behaviour when child coroutine attaches to the parent during "completing" -> "completed" transition

[qwwdfsad]

trafficstars

Steps to reproduce:

// Add this test to JobChildStressTest

@Test
fun testFailingChildIsAddedWhenJobFinalizesItsState() {
    // All exceptions should get aggregated here
    repeat(N_ITERATIONS) {
        runBlocking {
            val rogueJob = AtomicReference<Job?>()
            println(it)
            val deferred = CompletableDeferred<Unit>()
            launch(pool + deferred) {
                deferred.complete(Unit) // Transition deferred into "completing" state waiting for current child
                // **Asynchronously** submit task that launches a child so it races with completion
                pool.executor.execute {
                    rogueJob.set(launch(pool + deferred) {
                        println("isCancelled: " + coroutineContext.job.isCancelled)
                        throw TestException()
                    })
                }
            }

            deferred.join()
            if (rogueJob.get()?.isActive ?: false) {
                val rogue = rogueJob.get()!!
                println("Rogue job with parent " + rogue.parent + " and children list: " + rogue.parent?.children?.toList())
            }
        }
    }
}

What happens here:

• Deferred is completing, waiting for the first launch (1) ChildCompletion handler to finalize its state
• ChildCompletion invokes continueCompleting
• In parallel, the second launch (2) is attached to the deferred

1. Happy path: 2 successfully attaches to the parent, 1 detects that in continueCompleting and starts waiting for it. This situation is indistinguishable from deferred having two children

2. Unhappy path #1: 1 detects there are no children and invokes finalizeFinishingState. Then 2 attaches itself to the parent. finalizeFinishingState reaches completeStateFinalization -> notifyCompletion and cancels the child, which might have been running for some time already. This is an observable and counter-intuitive (because nothing actually failed or was cancelled explicitly) behaviour. Also, if 2 fails with an exception, it gets reported to the global exception handler.

3. Unhappy path #2: the same as above, but 2 attaches itself to the parent after it completely finalizes its state. Meaning that we have a completed deferred with no children and active non-cancelled coroutine with a parent pointing to the deferred

Note that 2) kind of emulates the behaviour "attempt to attach as a child to already completed job immediately cancels current job"