Test and test-and-set

In computer science, the test-and-set CPU instruction is used to implement mutual exclusion in multiprocessor environments. Although a correct lock can be implemented with test-and-set, it can lead to resource contention in busy lock (caused by bus locking and cache invalidation when test-and-set operation needs to access memory atomically).

To lower the overhead a more elaborate locking protocol test and test-and-set is used. The main idea is not to spin in test-and-set but increase the likelihood of successful test-and-set by using the following entry protocol to the lock:

boolean locked := false // shared lock variable
procedure EnterCritical() {
  do {
    while (locked == true) skip // spin until lock seems free
  } while TestAndSet(locked) // actual atomic locking
}

Exit protocol is:

procedure ExitCritical() {
  locked := false
}

The entry protocol uses normal memory reads to spin, waiting for the lock to become free. Test-and-set is only used to try to get the lock when normal memory read says it's free. Thus the expensive atomic memory operations happens less often than in simple spin around test-and-set.

If the programming language used supports short-circuit evaluation, the entry protocol could be implemented as:

 procedure EnterCritical() {
   while ( locked == true or TestAndSet(locked) == true )
     skip // spin until locked
 }

Caveat

Although this optimization is useful in system programming it should be avoided in high level concurrent programming unless all constraints are clear and understood. One example of bad usage is a similar idiom called double-checked locking, which, under certain conditions, can be an anti-pattern.

See also

• Parallel processor
• Parallel programming
• Mutual exclusion
• Test-and-set
• Fetch-and-add

References

• Gregory R. Andrews, Foundations of Multithreaded, Parallel, and Distributed Programming, pp. 100–101. Addison-Wesley, 2000. ISBN 0-201-35752-6.