What's the pattern to dispose of a resource when a suspend function needs to allocated a disposable resource. How do I handle the case where the cancellation happens after the resource is allocated but right before it's returned?

Cancellation is cooperative in coroutines, it cannot happen in every possible instant. The places where it can happen are predictable. But in any case, cancellation can be handled like any other exception. If you have a resource that needs cleanup, use a

try/finally

block to clean it up in every scenario (normal completion and exceptions)

It cannot happen when returning from one suspend function to another?

It cannot happen when returning from one suspend function to another?

No. For example, if you run this `foo`:

suspend fun foo() {
    val job = coroutineContext.job
    println("In foo, calling bar (isActive=${job.isActive})")
    bar()
    println("In foo, after bar returned (isActive=${job.isActive})")

}

suspend fun bar() {
    println("In bar, cancelling job…")
    coroutineContext.job.cancel()
    println("In bar, job cancelled, returning to foo…")
}

You’ll get this sequence of messages:

In foo, calling bar (isActive=true)
In bar, cancelling job…
In bar, job cancelled, returning to foo…
In foo, after bar returned (isActive=false)

The issue can happen in innocent looking code when unsing `withContext`:

suspend fun allocateResource() = withContext(<http://Dispatchers.io|Dispatchers.io>) {
  // allocate resource. E.g. open a database
  return resource
}

suspend fun leakResource() {
  val ressource = allocateResource()
  try {
    doSomething(resource)
  } finally {
    ressource.close()
  }
}

withContext

would catch the cancelation before returning and leak the resource even before it is returned from

allocateRessource

. Nothing much you can do about it.

relatedly, 1. in a suspend fun with a long chain of blocking calls, you may need

ensureActive()

or

runInterruptible {}

to allow for cancellation in the middle 2. you should not make suspend calls or launch new coroutines in a

finally

(or

catch (Exception)

block if it could contain

CancellationException

) because the containing job may be cancelled which would cause such calls to immediately propagate cancellation instead

@ephemient the issue in above exmaple is after the withContext block terminates. Not sure if withContext(NonCancellable) would skip cancelation checks on exit as well. If anyone is up to test it let us know the results, please.

you just need something like

var resource: Resource? = null
try {
    withContext(NonCancellable) {
        resource = allocateResource()
    }
    doSomething(resource)
} finally {
    resource?.close()
}

Correct. The pitfall is though, that the caller needs to take care of the NonCancellable context. And has to understand the issue enough to put the assignment of resource into the try block

So I guess the implicit continuations from one suspend to another do not go through a dispatcher

the caller gets resumed (with exception) even in case of cancellation. the only case I can see where this wouldn't happen is if the scope got garbage collected away

But if I understand you correctly the caller cannot get resumed with cancellation on return

what do you mean by "cancellation in return"

Well a return from a suspend function could really be a continuation under the hood

as Joffrey said, the model is that cancellation is cooperative and only happens at suspend points

@fitermay calling a suspend function does not suspend (until the called function gets to a suspension pint itself) so returning from a suspend function does not resume.

When one suspend function invokes another the return is not done via a stack

yes

yes, but the caller will always* be resumed

That's only if you return without suspending first

Suspend functions do calls and returns through the stack. I just double checked through the kotlin bytecode decompiler. No resuming on return, no cancelation on return.

How can it return through the stack after suspending?

suspend only actually happen when functions like

suspend*Coroutine

are called. I like to call these the ‘leaf’ suspend functions. All other suspend functions are only suspend when the functions they call suspend. I.e. you build up a ‘real’ call stack until you run into a `suspend*Coroutine`call or alike.

I understand that it only suspends there but the return statement might be after it suspends

can you provide a concrete example? I think you may have a misunderstanding of the coroutine model

suspend fun foo() : Int {
 println("Before suspend")
 delay(1)
 println("After suspend")
 return 1
}

All the magic happens around

delay

. entering foo and exiting does not suspend and does not resume

(I meant to ask that to Yuli, if it wasn't clear)

sorry @ephemient

I see now

Exactly

it's not really different from the equivalent in blocking APIs

fun allocate(): Resource {
    val future = executor.submit { ... }
    return future.get()
}

which can also leak if the current thread is interrupted

But I think the implementation detail of how return is handled is not obvious. They could've also implemented it by inverting the stack through a continuation

to truly suspend, there's no choice but to unwind the blocking call stack somehow. the way continuations work could be different but the "return to trampoline" part is inevitable

Right, looking at the disassembly again the return goes through the trampoline which invokes the return continuations