Test-and-set


In computer science, the test-and-set instruction is an instruction used to write 1 to a memory location and return its old value as a single atomic operation. If multiple processes may access the same memory location, and if a process is currently performing a test-and-set, no other process may begin another test-and-set until the first process's test-and-set is finished. A CPU may use a test-and-set instruction offered by another electronic component, such as dual-port RAM; a CPU itself may also offer a test-and-set instruction.
A lock can be built using an atomic test-and-set instruction as follows:

function Lock
The calling process obtains the lock if the old value was 0, otherwise the while-loop spins waiting to acquire the lock. This is called a spinlock. "Test and Test-and-set" is another example.
Maurice Herlihy proved that test-and-set has a finite consensus number and can solve the wait-free consensus problem for at-most two concurrent processes. In contrast, compare-and-swap offers a more general solution to this problem.

Hardware implementation of test-and-set

test-and-set instructions can work in many ways. Here are two variations, both of which describe a DPRAM which provides exactly 2 ports, allowing 2 separate electronic components access to every memory location on the DPRAM.

Variation 1

When CPU 1 issues a test-and-set instruction, the DPRAM first makes an "internal note" of this by storing the address of the memory location in a special place. If at this point, CPU 2 happens to issue a test-and-set instruction for the same memory location, the DPRAM first checks its "internal note", recognizes the situation, and issues a BUSY interrupt, which tells CPU 2 that it must wait and retry. This is an implementation of a busy waiting or spinlock using the interrupt mechanism. Since all this happens at hardware speeds, CPU 2's wait to get out of the spin-lock is very short.
Whether or not CPU 2 was trying to access the memory location, the DPRAM performs the test given by CPU 1. If the test succeeds, the DPRAM sets the memory location to the value given by CPU 1. Then the DPRAM wipes out its "internal note" that CPU 1 was writing there. At this point, CPU 2 could issue a test-and-set, which would succeed.

Variation 2

CPU 1 issues a test-and-set instruction to write to "memory location A". The DPRAM does not immediately store the value in memory location A, but instead simultaneously moves the current value to a special register, while setting the contents of memory location A to a special "flag value". If at this point, CPU 2 issues a test-and-set to memory location A, the DPRAM detects the special flag value, and as in Variation 1, issues a BUSY interrupt.
Whether or not CPU 2 was trying to access the memory location, the DPRAM now performs CPU 1's test. If the test succeeds, the DPRAM sets memory location A to the value specified by CPU 1. If the test fails, the DPRAM copies the value back from the special register to memory location A. Either operation wipes out the special flag value. If CPU 2 now issues a test-and-set, it will succeed.

Software implementation of test-and-set

Some instruction sets have an atomic test-and-set machine language instruction. Examples include x86 and IBM System/360 and its successors.
Those that do not can still implement an atomic test-and-set using a read-modify-write or compare-and-swap instruction.
The test and set instruction, when used with boolean values, uses logic like that shown in the following function, except that the function must execute atomically. That is, no other process must be able to interrupt the function mid-execution, thereby seeing a state that only exists while the function executes. That requires hardware support; it cannot be implemented as shown. Nevertheless, the code shown helps to explain the behaviour of test-and-set. NOTE: In this example, 'lock' is assumed to be passed by reference but the assignment to 'initial' creates a new value.
function TestAndSet
Not only is the code shown not atomic, in the sense of the test-and-set instruction, it also differs from the descriptions of DPRAM hardware test-and-set above. Here, the value being set and the test are fixed and invariant, and the value is updated regardless of the outcome of the test, whereas for the DPRAM test-and-set, the memory is set only when the test succeeds, and the value to set and the test condition are specified by the CPU. Here, the value to set can only be 1, but if 0 and 1 are considered the only valid values for the memory location, and "value is nonzero" is the only allowed test, then this equates to the case described for DPRAM hardware. From that viewpoint, this can, correctly, be called "test-and-set" in the full, conventional sense of that term. The essential point to note is the general intent and principle of test-and-set: a value is both tested and set in one atomic operation such that no other program thread or process can change the target memory location after it is tested but before it is set.
In the C programming language, the implementation would be like:

#define LOCKED 1
int TestAndSet

The code also shows that there are really two operations: an atomic read-modify-write and a test. Only the read-modify-write needs to be atomic.

Mutual exclusion using test-and-set

One way to implement mutual exclusion is by using a test-and-set based lock as follows:

pseudo-C implementation


volatile int lock = 0;
void Critical

The lock variable is a shared variable i.e. it can be accessed by all processors/threads. Note the volatile keyword. In absence of volatile, the compiler and/or the CPU may optimize access to lock and/or use cached values, thus rendering the above code erroneous. Conversely, and unfortunately, the presence of volatile does not guarantee that reads and writes are committed to memory. Some compilers issue memory barriers to ensure that operations are committed to memory, but since the semantics of volatile in C/C++ is quite vague, not all compilers will do that. Consult your compiler's documentation to determine if it does.
This function can be called by multiple processes, but it is guaranteed that only one process will be in the critical section at a time. The rest of the processes will keep spinning until they get the lock. It is possible that a process is never granted the lock. In such a case it will loop endlessly. This is a drawback of this implementation as it doesn't ensure fairness. These issues are further elaborated in the performance section.

Assembly implementation


enter_region: ; A "jump to" tag; function entry point.
tsl reg, flag ; Test and Set Lock; flag is the
; shared variable; it is copied
; into the register reg and flag
; then atomically set to 1.
cmp reg, #0 ; Was flag zero on entry_region?
jnz enter_region ; Jump to enter_region if
; reg is non-zero; i.e.,
; flag was non-zero on entry.
ret ; Exit; i.e., flag was zero on
; entry. If we get here, tsl
; will have set it non-zero; thus,
; we have claimed the resource
; associated with flag.
leave_region:
move flag, #0 ; store 0 in flag
ret ; return to caller

Here tsl is an atomic instruction and flag is the lock variable. The process doesn't return unless it acquires the lock.

Performance evaluation of test-and-set locks

The four major evaluation metrics for locks in general are uncontended lock-acquisition latency, bus traffic, fairness, and storage.
Test-and-set scores low on two of them, namely, high bus traffic and unfairness.
When processor P1 has obtained a lock and processor P2 is also waiting for the lock, P2 will keep incurring bus transactions in attempts to acquire the lock. When a processor has obtained a lock, all other processors which also wish to obtain the same lock keep trying to obtain the lock by initiating bus transactions repeatedly until they get hold of the lock. This increases the bus traffic requirement of test-and-set significantly. This slows down all other traffic from cache and coherence misses. It slows down the overall section, since the traffic is saturated by failed lock acquisition attempts. Test-and-test-and-set is an improvement over TSL since it does not initiate lock acquisition requests continuously.
When we consider fairness, we consider if a processor gets a fair chance of acquiring the lock when it is set free. In an extreme situation the processor might starve i.e. it might not be able to acquire the lock for an extended period of time even though it has become free during that time.
Storage overhead for TSL is next to nothing since only one lock is required. Uncontended latency is also low since only one atomic instruction and branch are needed.