https://kotlinlang.org logo
#coroutines
Title
# coroutines
g

groostav

05/09/2018, 5:01 PM
unrelated: so, I was talking with somebody (apologies, history goes fast in the slack channel and I'm bad with names, suffice to say it was somebody smarter than me) in this channel a while ago about a
SingleThreadedDispatcher
that would allow you to take a thread pool and put a facade on it such that any job submitted would be serialized (read: only run after the previous job had been completed). such a dispatcher would get you a similar effect as a
kotlinx.coroutines...sync.Mutex
as it relates to the behaviour of
sequentialDispatcherFacade.submit
and
mutex.withLock
. This issue of using
mutex
without the other more traditional garbage stuff in
java.concurrent
sounds like it can induce some insidious problems with threads caching values, unless somebody here tells me I'm mistaken which I'm desperately hoping they do. consider a simple little class:
Copy code
class SomeStatefulClass {
  val mutex: Mutex = Mutex()

  var state: Int = 0

  fun doStatefulTransform(){
    mutex.withLock {
      val newState = state + 1
      state = newState //this is actuall just state++, but it illustrates how we might get a stale-read. 
    }
  }
}

@Test fun doStuff(){
  val statefulThing = SomeStatefulClass()
  val job1 = launch(JavaFX) { for(i in 0 until 10_000) statefulThing.doStatefulTransform() }
  val job2 = launch(CommonPool) { for(i in 0 until 10_000) statefulThing.doStatefulTransform() } 
  //a better test might be val jobs = 0 .. 10k .map { launch { ...

  job1.await(); job2.await()

  assertThat(statefulThing.state).isEqualTo( ??????? )
}
Such a test would likely pass on x86 with a good deal of regularity because my understanding is that x86 doesn't make CPU-local caches as aggressively as some platforms. But arm on the other hand might release the mutex acquired by
doStatefulTransform()
before flushing its local change in its L2 cache (or similar) through to main memory, thus letting the other thread acquire the mutex but read a stale
state
value. Is there something in "happens-before" semantics that I'm missing that would protect us from this? Or do you need to mark
state
with
@Volatile
(or use an
AtomicInt
or
Unsafe
etc etc)? Similarly, could we leverage jcstress or similar with a CI service on kotlinx.coroutines to discover this? Is it worth investing the time? If I could steal some time on jetbrain's quasi-public teamcity infrastructure, I'd be happy to try and set this up, if the kotlinx.coroutines commander-in-chief thinks its a good idea.
v

Vsevolod Tolstopyatov [JB]

05/10/2018, 1:29 PM
Hi,
Is there something in “happens-before” semantics that I’m missing that would protect us from this?
Yes, releasing the
Mutex
happens-before every subsequent acquisition of that mutex, you don’t have to do any kind of synchronization on data protected by a mutex. I will improve our documentation.
Similarly, could we leverage jcstress or similar with a CI service on kotlinx.coroutines
Currently
jcstress
is not properly supported by Gradle+Koltin due to plugins inference. Usually we write hand-rolled stress tests (which are, of course, less strict than
jcstress
, but much easier to write and to integrate into CI pipeline). For “hardcore” concurrency tests (e.g. to check linearizability) we’re using
lin-check
(https://github.com/Devexperts/lin-check) which is based on jcstress, but is simpler to use: you don’t have to mark events as “interesting”,
lin-check
knows how to produce sequential execution history and compares it with concurrent one.
Is it worth investing the time?
If you see something is not properly covered by stress/lin-check tests, feel free to submit a PR, they are always welcome. You can find a lot of stress tests in repo as a reference (everything which ends with
StressTest
or
LinearizabilityTest
)
6 Views