Difference between revisions of "Atomic Operation"
Chris Tyler (talk | contribs) (Created page with 'Category:Computer Architecture An ''atomic operation'' is one that cannot be interrupted. These operations are typically used for acquiring locks. Consider the following cod…') |
Chris Tyler (talk | contribs) |
||
| (2 intermediate revisions by the same user not shown) | |||
| Line 5: | Line 5: | ||
if (lock == 0) { | if (lock == 0) { | ||
| − | lock++ | + | lock++; |
/* Perform exclusive action */ | /* Perform exclusive action */ | ||
| − | lock-- | + | lock--; |
} else { | } else { | ||
| − | abort() | + | abort(); |
} | } | ||
| Line 19: | Line 19: | ||
GCC provides [[Compiler Intrinsics|intrinsics]] for some atomic operations. | GCC provides [[Compiler Intrinsics|intrinsics]] for some atomic operations. | ||
| + | |||
| + | (Note that in the above example, even the <code>lock++</code> statement may not be atomic - it will typically compile into the three-operation sequence "load-increment-store", and its possible that two processes or threads could each perform this operation in an interleaved fashion and increment <code>lock</code> by just one instead of two). | ||
Latest revision as of 21:23, 9 October 2019
An atomic operation is one that cannot be interrupted. These operations are typically used for acquiring locks.
Consider the following code, intended to set a "lock" in shared memory so that only one thread or process performs a particular action at a time:
if (lock == 0) {
lock++;
/* Perform exclusive action */
lock--;
} else {
abort();
}
If another process or thread writes to the lock variable between the lock comparison (lock == 0) and setting the lock (lock++), then the lock could fail to ensure exclusive action.
There are two ways to accomplish an atomic operation:
- Use a single machine instruction.
- Tell the operating system in advance not to interrupt the operation.
GCC provides intrinsics for some atomic operations.
(Note that in the above example, even the lock++ statement may not be atomic - it will typically compile into the three-operation sequence "load-increment-store", and its possible that two processes or threads could each perform this operation in an interleaved fashion and increment lock by just one instead of two).
