Discussion





Discussion

Threading is a huge domain. This Item exists because that domain is important and needs to be explicitly acknowledged, but one Item can't do it justice and we will only summarize a few essentials; see the References for many more details and techniques. Among the most important issues are to avoid deadlocks, livelocks, and malign race conditions (including corruption due to insufficient locking).

The C++ Standard says not one word about threads. Nevertheless, C++ is routinely and widely used to write solid multithreaded code. If your application shares data across threads, do so safely:

  • Consult your target platforms' documentation for local synchronization primitives: Typical ones range from lightweight atomic integer operations to memory barriers to in-process and cross-process mutexes.

  • Prefer to wrap the platform's primitives in your own abstractions: This is a good idea especially if you need cross-platform portability. Alternatively, you can use a library (e.g., pthreads [Butenhof97]) that does it for you.

  • Ensure that the types you are using are safe to use in a multithreaded program: In particular, each type must at minimum:

    • Guarantee that unshared objects are independent: Two threads can freely use different objects without any special action on the caller's part.

    • Document what the caller needs to do to use the same object of that type in different threads: Many types will require you to serialize access to such shared objects, but some types do not; the latter typically either design away the locking requirement, or they do the locking internally themselves, in which case, you still need to be aware of the limits of what the internal locking granularity will do.

    Note that the above applies regardless of whether the type is some kind of string type, or an STL container like a vector, or any other type. (We note that some authors have given advice that implies the standard containers are somehow special. They are not; a container is just another object.) In particular, if you want to use standard library components (e.g., string, containers) in a multithreaded program, consult your standard library implementation's documentation to see whether that is supported, as described earlier.

When authoring your own type that is intended to be usable in a multithreaded program, you must do the same two things: First, you must guarantee that different threads can use different objects of that type without locking (note: a type with modifiable static data typically can't guarantee this). Second, you must document what users need to do in order to safely use the same object in different threads; the fundamental design issue is how to distribute the responsibility of correct execution (race- and deadlock-free) between the class and its client. The main options are:

  • External locking: Callers are responsible for locking. In this option, code that uses an object is responsible for knowing whether the object is shared across threads and, if so, for serializing all uses of the object. For example, string types typically use external locking (or immutability; see the third option on the next page).

  • Internal locking: Each object serializes all access to itself, typically by locking every public member function, so that callers may not need to serialize uses of the object. For example, producer/consumer queues typically use internal locking, because their whole raison d'être is to be shared across threads, and their interfaces are designed so that the appropriate level of locking is for the duration of individual member function calls (Push, Pop). More generally, note that this option is appropriate only when you know two things:

    First, you must know up front that objects of the type will nearly always be shared across threads, otherwise you'll end up doing needless locking. Note that most types don't meet this condition; the vast majority of objects even in a heavily multithreaded program are never shared across threads (and this is good; see Item 10).

    Second, you must know up front that per-member-function locking is at the right granularity and will be sufficient for most callers. In particular, the type's interface should be designed in favor of coarse-grained, self-sufficient operations. If the caller typically needs to lock several operations, rather than an operation, this is inappropriate; individually locked functions can only be assembled into a larger-scale locked unit of work by adding more (external) locking. For example, consider a container type that returns an iterator that could become invalid before you could use it, or provides a member algorithm like find that can return a correct answer that could become the wrong answer before you could use it, or has users who want to write if( c.empty() ) c.push_back(x);. (See [Sutter02] for additional examples.) In such cases, the caller needs to perform external locking anyway in order to get a lock whose lifetime spans multiple individual member function calls, and so internal locking of each member function is needlessly wasteful.

    So, internal locking is tied to the type's public interface: Internal locking becomes appropriate when the type's individual operations are complete in themselves; in other words, the type's level of abstraction is raised and expressed and encapsulated more precisely (e.g., as a producer-consumer queue rather than a plain vector). Combining primitive operations together to form coarser common operations is the approach needed to ensure meaningful but simple function calls. Where combinations of primitives can be arbitrary and you cannot capture the reasonable set of usage scenarios in one named operation, there are two alternatives: a) use a callback-based model (i.e., have the caller call a single member function, but pass in the task they want performed as a command or function object; see Items 87 to 89); or b) expose locking in the interface in some way.

  • Lock-free designs, including immutability (read-only objects): No locking needed. It is possible to design types so that no locking at all is needed (see References). One common example is immutable objects, which do not need to be locked because they never change; for example, for an immutable string type, a string object is never modified once created, and every string operation results in the creation of a new string.

Note that calling code should not need to know about your types' implementation details (see Item 11). If your type uses under-the-covers data-sharing techniques (e.g., copy-on-write), you do not need to take responsibility for all possible thread safety issues, but you must take responsibility for restoring "just enough" thread safety to guarantee that calling code will be correct if it performs its usual duty of care: The type must be as safe to use as it would be if it didn't use covert implementation-sharing. (See [Sutter04c].) As noted, all properly written types must allow manipulation of distinct visible objects in different threads without synchronization.

Particularly if you are authoring a widely-used library, consider making your objects safe to use in a multithreaded program as described above, but without added overhead in a single-threaded program. For example, if you are writing a library containing a type that uses copy-on-write, and must therefore do at least some internal locking, prefer to arrange for the locking to disappear in single-threaded builds of your library (#ifdefs and no-op implementations are common strategies).

When acquiring multiple locks, avoid deadlock situations by arranging for all code that acquires the same locks to acquire them in the same order. (Releasing the locks can be done in any order.) One solution is to acquire locks in increasing order by memory address; addresses provide a handy, unique, application-wide ordering.


     Python   SQL   Java   php   Perl 
     game development   web development   internet   *nix   graphics   hardware 
     telecommunications   C++ 
     Flash   Active Directory   Windows