lock free arena allocator implementation - correct?

问题

for a simple pointer-increment allocator (do they have an official name?) I am looking for a lock-free algorithm. It seems trivial, but I'd like to get soem feedback whether my implementaiton is correct.

not threadsafe implementation:

byte * head;  // current head of remaining buffer
byte * end;   // end of remaining buffer

void * Alloc(size_t size)
{
   if (end-head < size)
     return 0; // allocation failure

   void * result = head;
   head += size;
   return head;
}

My attempt at a thread safe implementation:

void * Alloc(size_t size)
{
  byte * current;
  do 
  {
     current = head;
     if (end - current < size)
        return 0;  // allocation failure
  } while (CMPXCHG(&head, current+size, current) != current));
  return current;
}

where CMPXCHG is an interlocked compare exchange with (destination, exchangeValue, comparand) arguments, returning the original value

Looks good to me - if another thread allocates between the get-current and cmpxchg, the loop attempts again. Any comments?

回答1:

Your current code appears to work. Your code behaves the same as the below code, which is a simple pattern that you can use for implementing any lock-free algorithm that operates on a single word of data without side-effects

do
{
    original = *data; // Capture.

    result = DoOperation(original); // Attempt operation
} while (CMPXCHG(data, result, original) != original);

EDIT: My original suggestion of interlocked add won't quite work here because you support trying to allocate and failing if not enough space left. You've already modified the pointer and causing subsequent allocs to fail if you used InterlockedAdd.

来源：https://stackoverflow.com/questions/1349313/lock-free-arena-allocator-implementation-correct

标签

c++

multithreading

lock-free

allocator