On 26 September 2010 15:00, Ohad Shacham <[email protected]> wrote:
> Hi,
>
>
> In addition to the behavior we experienced in function setAttribute. We also
> experienced another atomicity violation in function removeAttribute.
>
> The following code is located at the beginning of function removeAttribute
> at class ApplicationContext. This code fragment intends to check whether a
> key is already inside the “attributes” collection. If it does, then the key
> is removed from the collection and the function continues running while
> assigning the corresponding value to “value”. Otherwise, the function should
> return.
>
>
>
>
> found = attributes.containsKey(name);
>
> if (found) {
>
> value = attributes.get(name);
>
> attributes.remove(name);
>
> } else {
>
> return;
>
> }
>
>
>
> This code can behave non atomically when two different threads call
> removeAttribute function with the same attribute name. In this case, there
> exists a thread interleaving where two threads enter the “if” condition
> while the first one removes “name” from the attributes collection before the
> second one runs the get operation. This causes the function to continue
> running with value == null instead of returning.
>
>
>
> Could you please let me know whether you consider this as a bug?
>
>
>
> A possible way of modifying the code in order to execute this code
> atomically is as follows:
>
>
>
> if (attributes.containsKey(name)) {
> value = attributes.remove(name);
> if (value == null)
> return;
> }
That's not quite the same if atttributes.get() can return null, as
previously the null would be used in subsequent code.
If get() cannot return null, then I think one can just use:
value = attributes.remove(name);
if (value == null)
return;
If it is important to distinguish null entries from missing entries,
then a different approach is needed - e.g. AtomicReference as
suggested elsewhere
>
>
> Thanks,
>
> Ohad
>
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
