Hacker News Reader

Comment by sanjoy_das | Hacker News Reader

In (other) words, the situation is that you've just decremented the reference count of an object, because you've nulled out the only location in the heap that reached it. The reference count becomes zero after decrementing, so you know that _now_ there are no slots in the heap that point to it; but how do you know that there isn't a thread that fetched the object out of the heap before you started, and got stalled before it could increment the reference count and has been stalled since then?

(I'll try to edit the post to make this clearer ^).

I'd also like to stress that there are many ways around the problem, the only point of the post is that you'll have to solve some non-obvious problems if you try to generalize reference counting to a heap shared across threads.

View on HN

How did the other thread get the reference to the object? The only possible existing reference is the one we are using to decrement the shared counter.

Decrementing is done when the reference itself is being dropped (set to null or to point to a different object), which is a logically mutating operation (remember that only the reference count updates are atomic, the operations on the references themselves are not), thus no other acquire operation can be happening concurrently or it would be a data race.

Any concurrent operations on the reference itself must be synchronized via external means, usually a mutex. Of course concurrently mutating distinct references which refer to the same object/ref count is fine.

edit: rewording

> How did the other thread get the reference to the object? The only possible existing reference is the one we are using to decrement the shared counter.

I don't think this affects the point you're trying to make, but I suppose you could have three threads, where the logical operations are:

    ThreadA:
      obj = x->field_a;
    
    ThreadB:
      x->field_a = null;
    
    ThreadC:
      x->field_b = null;

with both field_a and field_b pointing to the same object initially. It does not affect your point, since ThreadA and ThreadB are now racing.

> Decrementing is done when the reference itself is being dropped (set to null or to point to a different object), which is a logically mutating operation (remember that only the reference count updates are atomic, the operations on the references themselves are not), thus no other acquire operation can be happening concurrently or it would be a data race.

It depends on your programming language. In Java racing on field updates (at the Java level) is well defined (but is allowed to return counter-intuitive results to some degree). That is:

    ThreadA
      int k = obj.field.hashCode()

    ThreadB
      obj.field = someOtherValue

is defined and is not allowed to have arbitrarily bad effects like crashing the VM. This is different from C++ (where these kind of accesses are UB, as you seem to imply). Generally, I think for high level languages it is better to have Java-like semantics where even racy accesses have some guarantees.

For C++, I can get the same Java-like guarantees by using `memory_order_relaxed` loads, but I suppose it is defensible for an atomic `shared_ptr` to have a complex refcounting protocol even for `memory_order_relaxed` loads and stores.

> Any concurrent operations on the reference itself must be synchronized via external means, usually a mutex.

Not sure how you're using a reference here, but if by "reference" you mean "a location in the heap" then that does not apply for Java. I personally tend to use "reference" in the same way as "pointer".

> Of course concurrently mutating distinct references which refer to the same object/ref count is fine. edit: rewording

edit: formatting

I was of course discussing the semantics of a C-like language, especially in light of the C++11 memory model.

Being memory safe at all cost, java (and any other shared-memory memory safe language) must guarantee minimum of safety to concurrent reference updates. A JVM which uses reference counting instead of a proper GC would be interesting to implement...

You can't currently put a shared_ptr in an atomic<> variable as it does not meet the concept requirements (Trivially constructible and copyable), but, IIRC, you will in C++17 and, yes, even relaxed loads will be expensive (in practice implementations are expected to lock a spinlock around every operation)

Reply on news.ycombinator.com