Thread Safety: Requirements and Design

Last modified 12 Mar 2021 10:22 +01:00

The purpose of this document is to guide the development of midPoint code regarding the aspect of thread safety.

For the time being, this is just a skeleton.

Requirements

What we want to achieve in the area of thread safety? Which objects will be shared and which need not be?

Data structures

We need midPoint to be massively scalable, so it is understandable that majority of its actions will be executed in parallel threads. Fortunately, many of the data structures can be confined to a single thread, but there are others, which have to be safely shared.

Objects Shared Description

Mutable prism structures

No

Mutable prism structures (including constituent generated beans) are not to be shared. We assume that the cost of making them thread-safe outweighs benefits of doing so. These structures are often short-lived, created e.g. for processing a single user, and destroyed afterwards. On the other side, their use in GUI is limited to a single user session, so again, they do not need to be thread-safe.[1]

Immutable prism structures

Yes

On the other hand, there is a lot of data that is long-lived, and potentially shared among threads. For example, all objects that are cached (e.g. repository objects, "live" resource and connector objects) are used by many threads simultaneously.

Prism Definitions

Yes, for frozen

Prism context definitions must be thread-safe

Prism Deltas

Yes, for frozen

Prism Queries

Yes

Prism queries should be freezable/immutable.

RunningTask instances

Yes

If a task is running, it is accessed both by task manager and the respective task handler. (Probably also by handlers of lightweight subtasks.) This applies to all data structures that constitute the task and are reachable through it.

Other task instances

No

A "plain" task should be accessed only by the thread that as associated to it.

Operation result

Yes, single writer, multiple readers

Any given operation result should be accessed only by a single thread at a time. (As these results form trees, it is legal for multiple threads access different parts of such a tree.)

This can present some complications related to tasks, because operation result is a part of any running task. This is something to be resolved in the future.

TODO …​

…​

…​

MidPoint components

The majority of components represented by Spring beans (like RepositoryService, ProvisioningService, ModelService, and so on) are accessed in parallel, so they must be thread-safe.

TODO …​

Design guidelines

What specific rules should we obey when constructing thread-safe code? (We can skip listing obvious things here. But what is obvious in this are, anyway?)

Majority of these guidelines stem from Java Concurrency in Practice (JCIP) book.
  1. Each class (except for evident cases) should have a synchronization policy, typically:

    • Thread-confined. Objects of this class are accessed only from a single thread at any given time. (The "ownership" may change, but only in a well-defined and safe way.)

    • Shared read-only. These objects are accessed by multiple threads, but never modified.

    • Shared thread-safe. These objects perform any needed synchronization internally, so they are safely used from multiple threads.

    • Guarded. These objects can be accessed by multiple threads, provided that a specified lock is held upon access.

  2. If possible, a synchronization policy should be declared using appropriate class-level annotation, like @ThreadSafe, @NotThreadSafe, @Immutable.[2]

  3. Generally, we prefer immutability. All classes should be immutable unless strictly needed. (In other words, immutability should be considered as the first option.) Even if the class as a whole cannot be mutable, one should make final as many of its fields as possible.

  4. When using locking, it should be clear which locks guard which state variables, for example by using @GuardedBy annotation. Quoting JCIP:

    The design process for a thread-safe class should include these three basic elements:
    * Identify the variables that form the object’s state;
    * Identify the invariants that constrain the state variables;
    * Establish a policy for managing concurrent access to the object’s state.
  5. Besides avoiding concurrent access to shared data structures, memory visibility must be addressed as well. Each publication of a shared object must be safe. Again, from JCIP:

    To publish an object safely, both the reference to the object and the object’s
    state must be made visible to other threads at the same time. A properly constructed object
    can be safely published by:
    * Initializing an object reference from a static initializer;
    * Storing a reference to it into a volatile field or AtomicReference;
    * Storing a reference to it into a final field of a properly constructed object; or
    * Storing a reference to it into a field that is properly guarded by a lock.
  6. …​ TODO …​

To be considered as well:

  • For shared thread-safe classes, the mechanism(s) for ensuring thread safety must be (at least very briefly) described in the class-level comment.

  • Should we (at least in minimal way) specify class invariants for thread-safe classes with more complex mutable state?


1. TODO: Is it true that during a single user session there is no concurrent access to prism objects? What about wicket/servlet serialization? Could it somehow collide with the "main" processing of session objects?
2. See also Appendix A of JCIP.
Was this page helpful?
YES NO
Thanks for your feedback