The core concept of structured concurrency is that when control splits into concurrent tasks that they join up again. If a “main task” splits into several concurrent sub-tasks scheduled to be executed in fibers then those fibers must terminate before the main task can complete.
The main benefit of structured concurrency is abstraction. A caller of a method that is invoked to do a task should not care if the method decomposes the task and schedules a million fibers. When the method completes, any fibers scheduled by the method should have terminated.
In Project Loom, a prototype API has been developed called FiberScope that is a scope in which fibers are scheduled. Here is a basic example:
A thread or fiber enters a scope by calling the FiberScope.cancellable method (we will explain cancellation later). It exits the scope when the code in the block completes and any fibers scheduled in the scope have terminated. The example schedules two fibers. The thread/fiber executing the above code may have to wait (in the FiberScope’s close method) until the two fibers have terminated.
FiberScopes can be nested, consider the following:
In this example, a thread or fiber enters scope1, schedules a fiber, then enters scope2 where it schedules a fiber in that scope. The execution cannot exit scope2 until the fiber scheduled in that scope terminates. When it exits, it is back in scope1, where it scheduled another fiber. It cannot exit scope1 until the two fibers scheduled in that scope have terminated.
More generally, the structured approach leads to trees of tasks. Consider the following method x that schedules two fibers to execute foo and bar. The code in foo and bar each schedule two fibers.
fiber1 is scheduled in scope1 to execute foo. It enters scope2 and schedules two fibers. It exits scope2 (and returns to scope1) when the two fibers terminate. fiber2 is scheduled in scope1 to execute bar. It enters scope3 and schedules two fibers. It exits scope3 (and returns to scope1) when the two fibers terminate. The thread or fiber executing will not exit scope1 until fiber1 and fiber2 have terminated.
As an escape hatch, the FiberScope API defines the detached() method to return a scope which can be used to schedule fibers that are intended to outline the context where they are initially scheduled.
A fiber executing in cancellable scope may be cancelled by invoking its cancel method to set the fiber’s cancel status and unpark the fiber. Cancellation works cooperatively: a Fiber needs to check for cancellation (say, when doing blocking operations), and throw an appropriate exception for the context that it is running in. Examples of operation operations that check for cancellation are Thread.sleep and blocking socket operations such as connect, accept, read, and write.
There are rare, but important, cases such as shutdown or recovery steps where a fiber may need to be shielded from cancellation. To support this, FiberScope defines the nonCancellable() method to enter a non-cancellable scope. A Fiber that checks for cancellation in a non-cancellation scope will get back “false”. If its cancel status is set then it will observe this when it pops back to a cancellable scope.
Decomposing deadline or timeouts is very difficult to get right.
FiberScope supports entering a scope with a deadline, expressed as a java.time.Instant. If the deadline expires before the thread/fiber exits the scope then all fibers scheduled in the scope are cancelled and the close method throws an exception (or it gets added as a suppressed exception when exiting with an exception).
Deadlines work with nested scopes. Consider the following:
scope1 is entered with a deadline that is now + 10s. It schedules a fiber in scope2 and cannot exit to scope1 until the fiber terminates. If the deadline expires in the meantime then the fiber will be cancelled and the thread/fiber will throw CancelledException(“Deadline expired”) when existing scope1. If the inner scope had a deadline that was further into the future that the deadline then the deadline for the outer scope will expire first.