Naive reference counting is “easy” to implement on a system that does not share objects between threads, but thinking about reference counting in systems that do share objects between threads, two problems (other than the standard “increments and decrements need to be atomic operations”) come to mind. So far the contents of this post (which are not novel) have lived in one-off tweets and emails, but I think it is time to write them down in an organized way.

Problem 0: Stores to the Heap need to be XCHGes

(Edit: I initially had a few mistakes here – I’d claimed that the stores need to be CAS’es when an XCHG would be sufficient. The order between the increment and the decrement was also incorrect. Thanks @barrkel for pointing these out!)

Executing a store to the heap requires incrementing the refcount of the object stored, and decrementing the refcount of the object overwritten. On a single threaded application this can be implemented in the obvious way as

old_val = *heap_addr
*heap_addr = new_val

Naively extending this to a multi-threaded system requires doing a XCHG

old_val = Atomic_XCHG(heap_addr, new_val)

which has a fairly high overhead, especially since the programmer did not ask for an atomic operation, and the extra synchronization is purely “dead” overhead.

Solutions I’m aware of

“An on-the-fly reference counting garbage collector for Java.”1 enumerates a solution that involves synchronizing the collector and the mutator, à la checkpoints2 or ragged safepoints3. A solution like that is feasible in a JVM, but would be difficult to implement for an uncooperative environment, e.g. for a thread-safe version of std::shared_ptr<T>.

Problem 1: Racing Increments and Decrements

Consider two threads racing to update a slot in the heap:

  val = object->field;
  val->refcount++;  // either because you actually track refs
                    // on the stack or because you're about to
                    // publish val to some slot on the heap


  ;; Semantically, this is "object.field = null"
  old_val = Atomic_XCHG(&(object->field), null)
  if (--old_val->refcount == 0)
    delete old_val;

There is a race between Thread_A and Thread_B if the refcount of the initial value of object->field is 1 (i.e. the initial value of object->field is reachable only from object):

Thread_A: val = object->field;
Thread_B: old_val = Atomic_XCHG(&(object->field), null)
Thread_B: --old_val->refcount == 0 // == true
Thread_B: delete old_val;
Thread_A: val->refcount++; // == CRASH!

In (other) words Thread_B just decremented the reference count of an object O because it overwrote a slot in the heap that reached it. The reference count becomes zero after decrementing, so it knows that now there are no slots in the heap that point to O (and the slot it overwrote was the only slot that contained a pointer to O). But it still needs to know that there isn’t a Thread_A that fetched O out of the heap before Thread_B overwrote the slot, and got stalled before it could increment the reference count. When Thread_A started O was reachable normally and had a reference count of 1; but that is not relevant here.

Solutions I’m aware of

There are three solutions to this that I’m aware of:

  1. Hazard pointers4val->refcount++ in Thread_A counts (no pun intended!) as a hazardous access, and could be protected by a published hazard pointer. However, I think there are some subtleties here, discussed below.

  2. Pass the buck5 – I don’t understand this solution quite yet, but it looks like a generalization of hazard pointers.

  3. ThreadScan6 – this was pointed out to me by @Matt. The subtlety with hazard pointers mentioned below also applies to ThreadScan, as far as I can tell.

These should not be fundamentally difficult to implement in an uncooperative environment (e.g. for a thread-safe std::shared_ptr<T>), but they’re still very tricky to get right.

A Subtlety with Hazard Pointers

I think there is an issue with using hazard pointers for reference counting – a “node” in our “data structure” (the heap) can go from “unreachable from the heap” to “reachable from the heap”. This means if we do something like this for obj.field = null (in thread A, say):

// Trying to set obj->field to null
do {
  old_val = obj->field
} while(CAS(&(object->field), old_val, null) != Success);
if ((--old_val->refcount) <= 0)

then we have a race between another thread (B, say) loading obj->field and linking it back to the heap: that operation could have started before thread A unlinked old_val from the heap, and finished before A called hazard_ptr_free. Since B no longer has a hazard pointer to old_val, A would end up freeing something reachable from the heap.

Note: in this example I’ve had to use CAS instead of XCHG, since I need to guarantee that obj->field is old_val after old_val has been published as a hazard pointer. There may be a way around this – I haven’t spent too much time thinking about it.

The issue seems solvable though, perhaps we need to be careful to not increment refcounts of objects with a zero refcount? That would mean the increment operation needs to be something like an xadd instead of an add.

Other Solutions

I’m interested in hearing about other solutions to these problems. If you’re aware of any, please comment here, drop me an email, or tweet at me – I’ll update this section with appropriate credits.

  1. Levanoni, Yossi, and Erez Petrank. “An on-the-fly reference counting garbage collector for Java.” ACM SIGPLAN Notices 36.11 (2001): 367-380. 

  2. Click, Cliff, Gil Tene, and Michael Wolf. “The pauseless GC algorithm.” Proceedings of the 1st ACM/USENIX international conference on Virtual execution environments. ACM, 2005. 

  3. Pizlo, Filip, et al. “Schism: fragmentation-tolerant real-time garbage collection.” ACM Sigplan Notices. Vol. 45. No. 6. ACM, 2010. 

  4. Michael, Maged M. “Hazard pointers: Safe memory reclamation for lock-free objects.” Parallel and Distributed Systems, IEEE Transactions on 15.6 (2004): 491-504. 

  5. Herlihy, Maurice, Victor Luchangco, and Mark Moir. “The repeat offender problem: a mechanism for supporting dynamic-sized lock-free data structures.” (2002). 

  6. Alistarh, Dan, et al. “ThreadScan: Automatic and Scalable Memory Reclamation.” (2015).