Handler.asCoroutineDispatcher() can block runBlocking() forever

[pyricau]

Summary

CoroutineDispatcher.dispatch() kdoc says:

This method should guarantee that the given block will be eventually invoked, otherwise the system may reach a deadlock state and never leave it.

The CoroutineDispatcher instance returned by Handler.asCoroutineDispatcher() does not provide that guarantee: if the backing Handler thread has quit, Handler.post() will return false, but the dispatcher does not check for that and the coroutine does not run (source).

override fun dispatch(context: CoroutineContext, block: Runnable) {
  handler.post(block)
}

runBlocking() is an example of method that expects the coroutine dispatcher to run or cancel the coroutine, and otherwise will block forever.

Executor.asCoroutineDispatcher() implements this guarantee properly (source):

    override fun dispatch(context: CoroutineContext, block: Runnable) {
        try {
            executor.execute(wrapTask(block))
        } catch (e: RejectedExecutionException) {
            unTrackTask()
            cancelJobOnRejection(context, e)
            Dispatchers.IO.dispatch(context, block)
        }
    }

Example code

// setup
val thread = HandlerThread("stop stopping")
thread.start()
val handler = Handler(thread.looper)
val dispatcher = handler.asCoroutineDispatcher()

// Quit the thread, after this Handler.post() now returns false.
thread.quit()
// runBlocking waits for the dispatcher to run or cancel the coroutine. The current implementation does nothing.
runBlocking(dispatcher) {
  // This  never runs
}

Note: on Android, the main thread (which is a HandlerThread) never quits. However it's super common to create HandlerThread instances for single threaded work, or reuse existing instances when progressively migrating to coroutines. And it's common to quit such threads when running UI tests in CI and resetting the app in between every thread.

We ran into this edge case at Square, I posted the investigation on twitter: https://twitter.com/Piwai/status/1406301224295710720

[pyricau]

For more context, see #2003 from which the fix was implemented for Executors.asCoroutineDispatcher (#2012)