Smart pointers: who owns the object?

Question

C++ is all about memory ownership - aka ownership semantics.

It is the responsibility of the owner of a chunk of dynamically allocated memory to release that memory. So the question really becomes who owns the memory.

In C++ ownership is documented by the type a raw pointer is wrapped inside thus in a good (IMO) C++ program it is very rare (rare, not never) to see raw pointers passed around (as raw pointers have no inferred ownership thus we can not tell who owns the memory and thus without careful reading of the documentation you can't tell who is responsible for ownership).

Conversely, it is rare to see raw pointers stored in a class each raw pointer is stored within its own smart pointer wrapper. (N.B.: If you don't own an object you should not be storing it because you can not know when it will go out of scope and be destroyed.)

So the question:

• What type of ownership semantic have people come across?
• What standard classes are used to implement those semantics?
• In what situations do you find them useful?

Lets keep 1 type of semantic ownership per answer so they can be voted up and down individually.

Summary:

Conceptually, smart pointers are simple and a naive implementation is easy. I have seen many attempted implementations, but invariably they are broken in some way that is not obvious to casual use and examples. Thus I recommend always using well tested smart pointers from a library rather than rolling your own. std::auto_ptr or one of the Boost smart pointers seem to cover all my needs.

std::auto_ptr<T>:

Single person owns the object. Transfer of ownership is allowed.

Usage: This allows you to define interfaces that show the explicit transfer of ownership.

boost::scoped_ptr<T>

Single person owns the object. Transfer of ownership is NOT allowed.

Usage: Used to show explicit ownership. Object will be destroyed by destructor or when explicitly reset.

boost::shared_ptr<T> (std::tr1::shared_ptr<T>)

Multiple ownership. This is a simple reference counted pointer. When the reference count reaches zero, the object is destroyed.

Usage: When an object can have multiple owers with a lifetime that can not be determined at compile time.

boost::weak_ptr<T>:

Used with shared_ptr<T> in situations where a cycle of pointers may happen.

Usage: Used to stop cycles from retaining objects when only the cycle is maintaining a shared refcount.

Solution 1

For me, these 3 kinds cover most of my needs:

shared_ptr - reference-counted, deallocation when the counter reaches zero

weak_ptr - same as above, but it's a 'slave' for a shared_ptr, can't deallocate

auto_ptr - when the creation and deallocation happen inside the same function, or when the object has to be considered one-owner-only ever. When you assign one pointer to another, the second 'steals' the object from the first.

I have my own implementation for these, but they are also available in Boost.

I still pass objects by reference (const whenever possible), in this case the called method must assume the object is alive only during the time of call.

There's another kind of pointer that I use that I call hub_ptr. It's when you have an object that must be accessible from objects nested in it (usually as a virtual base class). This could be solved by passing a weak_ptr to them, but it doesn't have a shared_ptr to itself. As it knows these objects wouldn't live longer than him, it passes a hub_ptr to them (it's just a template wrapper to a regular pointer).

Solution 2

Simple C++ Model

In most modules I saw, by default, it was assumed that receiving pointers was not receiving ownership. In fact, functions/methods abandoning ownership of a pointer were both very rare and explicitly expressed that fact in their documentation.

This model assumes that the user is owner only of what he/she explicitly allocates. Everything else is automatically disposed of (at scope exit, or through RAII). This is a C-like model, extended by the fact most pointers are owned by objects that will deallocate them automatically or when needed (at said objects destruction, mostly), and that the life duration of objects are predictable (RAII is your friend, again).

In this model, raw pointers are freely circulating and mostly not dangerous (but if the developer is smart enough, he/she will use references instead whenever possible).

• raw pointers
• std::auto_ptr
• boost::scoped_ptr

Smart Pointed C++ Model

In a code full of smart pointers, the user can hope to ignore the lifetime of objects. The owner is never the user code: It is the smart pointer itself (RAII, again). The problem is that circular references mixed with reference counted smart pointers can be deadly, so you have to deal both with both shared pointers and weak pointers. So you have still ownership to consider (the weak pointer could well point to nothing, even if its advantage over raw pointer is that it can tell you so).

• boost::shared_ptr
• boost::weak_ptr

Conclusion

No matter the models I describe, unless exception, receiving a pointer is not receiving its ownership and it is still very important to know who owns who. Even for C++ code heavily using references and/or smart pointers.

Solution 3

Don't have shared ownership. If you do, make sure it's only with code you don't control.

That solves 100% of the problems, since it forces you to understand how everything interacts.

Solution 4

From boost, there's also the pointer container library. These are a bit more efficient and easier to use than a standard container of smart pointers, if you'll only be using the objects in the context of their container.

On Windows, there are the COM pointers (IUnknown, IDispatch, and friends), and various smart pointers for handling them (e.g. the ATL's CComPtr and the smart pointers auto-generated by the "import" statement in Visual Studio based on the _com_ptr class).

Solution 5

std::tr1::shared_ptr<Blah> is quite often your best bet.

View more solutions

Comments

[Martin York]

shared_ptr is the most common. But there are many more. Each has its own usage pattern and good and bad places to sue. A bit more description would be nice.

[Branan]

If you're stuck with an older compiler, boost::shared_ptr<blah> is what std::tr1::shared_ptr<blah> is based on. It's a simple enough class that you can probably rip it from Boost and use it even if your compiler isn't supported by the latest version of Boost.

[Michel]

Instead of creating your own pointer class (hub_ptr), why don't you just pass *this to these objects and let them store it as a reference? Since you even acknowledge that the objects will be destroyed at the same time as the owning class, I don't understand the point of jumping through so many hoops.

[Daniel Earwicker]

That's exactly what I said in the last paragraph.

[Fabio Ceconello]

It's basically a design contract to make the things clear. When the child object receives the hub_ptr, it knows that the pointed object won't be destroyed during the child's lifetime, and has no ownership to it. Both the contained and container objects agree to a clear set of rules. If you use a naked pointer, the rules can be documented, but won't be enforced by the compiler and the code.

[Fabio Ceconello]

Also note that you can have #ifdefs to make hub_ptr be typedef'd to a naked pointer in release builds, so the overhead will exist only in the debug build.

[Drake]

Note that the Boost documentation contradicts your description of scoped_ptr. It states that it is noncopyable and that ownership can not be transferred.

[Fabio Ceconello]

@Alec Thomas, You're right. I was thinking about auto_ptr and wrote scoped_ptr. Corrected.