Alex Rivera | Logout

Fully thread-safe shared_ptr implementation

Asked 2009-01-14T03:28:41.833
13

Does anybody know of a fully thread-safe shared_ptr implementation? E.g. boost implementation of shared_ptr is thread-safe for the targets (refcounting) and also safe for simultaneous shared_ptr instance reads, but not writes or for read/write.

(see Boost docs, examples 3, 4 and 5).

Is there a shared_ptr implementation that is fully thread-safe for shared_ptr instances?

Strange that boost docs say that:

shared_ptr objects offer the same level of thread safety as built-in types.

But if you compare an ordinary pointer (built-in type) to smart_ptr, then simultaneous write of an ordinary pointer is thread-safe, but simultaneous write to a smart_ptr isn't.

EDIT: I mean a lock-free implementation on x86 architecture.

EDIT2: An example use case for such a smart pointer would be where there are a number of worker threads which update a global shared_ptr with a their current work item and a monitor thread that takes random samples of the work items. The shared-ptr would own the work item until another work item pointer is assigned to it (thereby destroying the previous work item). The monitor would get ownership of the work item (thereby preventing the work item to be destroyed) by assigning it to its own shared-ptr. It can be done with XCHG and manual deletion, but would be nice if a shared-ptr could do it.

Another example is where the global shared-ptr holds a "processor", and is assigned by some thread, and used by some other thread. When the "user" thread sees that the processor shard-ptr is NULL, it uses some alternative logic to do the processing. If it's not NULL, it prevents the processor from being destroyed by assigning it to its own shared-ptr.

Edit
Report

2 Answers

2

Simultaneous write to a built-in pointer is certainly not thread safe. Consider the implications of writing to the same value with respect to memory barriers if you really want to drive yourself crazy (for instance, you could have two threads thinking the same pointer had different values).

RE: Comment - the reason built-ins aren't double deleting is because they aren't deleting at all (and the implementation of boost::shared_ptr I use wouldn't double delete, since it uses a special atomic increment and decrement, so it would only single delete, but then the result would could have the pointer from one and the ref count of the other. Or pretty much any combination of the two. It would be bad.). The statement in the boost docs is correct as it is, you get the same guarantees as you do with a built-in.

RE: EDIT2 - The first situation you are describing are very different between using built-ins and shared_ptrs. In one (XCHG and manual delete) there's no reference count; you are assuming you are the one and only owner when you do this. If using shared pointers, you are saying other threads might have ownership, which makes things far more complex. I believe it is possible with a compare-and-swap, but this would be very non-portable.

C++0x is coming out with an atomics library, which should make it much easier to write generic multi-threaded code. You'll probably have to wait till that comes out to see good cross-platform reference implementations of thread-safe smart pointers.

answered 2009-01-14T03:58:13.147
0

You can use this implementation Atomic Reference Counting Pointers to at least implement the Reference Counting mechanism.

answered 2010-05-04T19:41:06.227

Your Answer