Alternate Implementations Possible¶

This volume of IEEE Std 1003.1-2001 supports several alternative implementations of mutexes. An implementation may store the lock directly in the object of type pthread_mutex_t. Alternatively, an implementation may store the lock in the heap and merely store a pointer, handle, or unique ID in the mutex object. Either implementation has advantages or may be required on certain hardware configurations. So that portable code can be written that is invariant to this choice, this volume of IEEE Std 1003.1-2001 does not define assignment or equality for this type, and it uses the term "initialize" to reinforce the (more restrictive) notion that the lock may actually reside in the mutex object itself.

Note that this precludes an over-specification of the type of the mutex or condition variable and motivates the opaqueness of the type.

An implementation is permitted, but not required, to have pthread_mutex_destroy() store an illegal value into the mutex. This may help detect erroneous programs that try to lock (or otherwise reference) a mutex that has already been destroyed.

Tradeoff Between Error Checks and Performance Supported¶

Many of the error checks were made optional in order to let implementations trade off performance versus degree of error checking according to the needs of their specific applications and execution environment. As a general rule, errors or conditions caused by the system (such as insufficient memory) always need to be reported, but errors due to an erroneously coded application (such as failing to provide adequate synchronization to prevent a mutex from being deleted while in use) are made optional.

A wide range of implementations is thus made possible. For example, an implementation intended for application debugging may implement all of the error checks, but an implementation running a single, provably correct application under very tight performance constraints in an embedded computer might implement minimal checks. An implementation might even be provided in two versions, similar to the options that compilers provide: a full-checking, but slower version; and a limited-checking, but faster version. To forbid this optionality would be a disservice to users.

By carefully limiting the use of "undefined behavior" only to things that an erroneous (badly coded) application might do, and by defining that resource-not-available errors are mandatory, this volume of IEEE Std 1003.1-2001 ensures that a fully-conforming application is portable across the full range of implementations, while not forcing all implementations to add overhead to check for numerous things that a correct program never does.

Why No Limits are Defined¶

Defining symbols for the maximum number of mutexes and condition variables was considered but rejected because the number of these objects may change dynamically. Furthermore, many implementations place these objects into application memory; thus, there is no explicit maximum.

Static Initializers for Mutexes and Condition Variables¶

Providing for static initialization of statically allocated synchronization objects allows modules with private static synchronization variables to avoid runtime initialization tests and overhead. Furthermore, it simplifies the coding of self-initializing modules. Such modules are common in C libraries, where for various reasons the design calls for self-initialization instead of requiring an explicit module initialization function to be called. An example use of static initialization follows.

Without static initialization, a self-initializing routine foo() might look as follows:

static pthread_once_t foo_once = PTHREAD_ONCE_INIT;
static pthread_mutex_t foo_mutex;

void foo_init()
{


    pthread_mutex_init(&foo_mutex, NULL);
}

void foo()
{


    pthread_once(&foo_once, foo_init);


    pthread_mutex_lock(&foo_mutex);


   /* Do work. */


    pthread_mutex_unlock(&foo_mutex);
}

With static initialization, the same routine could be coded as follows:

static pthread_mutex_t foo_mutex = PTHREAD_MUTEX_INITIALIZER;

void foo()
{


    pthread_mutex_lock(&foo_mutex);


   /* Do work. */


    pthread_mutex_unlock(&foo_mutex);
}

Note that the static initialization both eliminates the need for the initialization test inside pthread_once() and the fetch of &foo_mutex to learn the address to be passed to pthread_mutex_lock() or pthread_mutex_unlock().

Thus, the C code written to initialize static objects is simpler on all systems and is also faster on a large class of systems; those where the (entire) synchronization object can be stored in application memory.

Yet the locking performance question is likely to be raised for machines that require mutexes to be allocated out of special memory. Such machines actually have to have mutexes and possibly condition variables contain pointers to the actual hardware locks. For static initialization to work on such machines, pthread_mutex_lock() also has to test whether or not the pointer to the actual lock has been allocated. If it has not, pthread_mutex_lock() has to initialize it before use. The reservation of such resources can be made when the program is loaded, and hence return codes have not been added to mutex locking and condition variable waiting to indicate failure to complete initialization.

This runtime test in pthread_mutex_lock() would at first seem to be extra work; an extra test is required to see whether the pointer has been initialized. On most machines this would actually be implemented as a fetch of the pointer, testing the pointer against zero, and then using the pointer if it has already been initialized. While the test might seem to add extra work, the extra effort of testing a register is usually negligible since no extra memory references are actually done. As more and more machines provide caches, the real expenses are memory references, not instructions executed.

Alternatively, depending on the machine architecture, there are often ways to eliminate all overhead in the most important case: on the lock operations that occur after the lock has been initialized. This can be done by shifting more overhead to the less frequent operation: initialization. Since out-of-line mutex allocation also means that an address has to be dereferenced to find the actual lock, one technique that is widely applicable is to have static initialization store a bogus value for that address; in particular, an address that causes a machine fault to occur. When such a fault occurs upon the first attempt to lock such a mutex, validity checks can be done, and then the correct address for the actual lock can be filled in. Subsequent lock operations incur no extra overhead since they do not "fault". This is merely one technique that can be used to support static initialization, while not adversely affecting the performance of lock acquisition. No doubt there are other techniques that are highly machine-dependent.

The locking overhead for machines doing out-of-line mutex allocation is thus similar for modules being implicitly initialized, where it is improved for those doing mutex allocation entirely inline. The inline case is thus made much faster, and the out-of-line case is not significantly worse.

Besides the issue of locking performance for such machines, a concern is raised that it is possible that threads would serialize contending for initialization locks when attempting to finish initializing statically allocated mutexes. (Such finishing would typically involve taking an internal lock, allocating a structure, storing a pointer to the structure in the mutex, and releasing the internal lock.) First, many implementations would reduce such serialization by hashing on the mutex address. Second, such serialization can only occur a bounded number of times. In particular, it can happen at most as many times as there are statically allocated synchronization objects. Dynamically allocated objects would still be initialized via pthread_mutex_init() or pthread_cond_init().

Finally, if none of the above optimization techniques for out-of-line allocation yields sufficient performance for an application on some implementation, the application can avoid static initialization altogether by explicitly initializing all synchronization objects with the corresponding pthread_*_init() functions, which are supported by all implementations. An implementation can also document the tradeoffs and advise which initialization technique is more efficient for that particular implementation.

Destroying Mutexes¶

A mutex can be destroyed immediately after it is unlocked. For example, consider the following code:

struct obj {
pthread_mutex_t om;


    int refcnt;


    ...
};

obj_done(struct obj *op)
{


    pthread_mutex_lock(&op->om);


    if (--op->refcnt == 0) {


        pthread_mutex_unlock(&op->om);
(A)     pthread_mutex_destroy(&op->om);
(B)     free(op);


    } else
(C)     pthread_mutex_unlock(&op->om);
}

In this case obj is reference counted and obj_done() is called whenever a reference to the object is dropped. Implementations are required to allow an object to be destroyed and freed and potentially unmapped (for example, lines A and B) immediately after the object is unlocked (line C).