At 12:58 PM 03/07/2002, Bill Pugh wrote:
>I want to start by stating something that I believe needs to be guaranteed
>by the memory model, and then see where that takes us.
>I propose that the memory model not allow "Spooky errors at a distance".
>Specifically, say you have two Runnable tasks, r1 and r2. Consider two
>ways to execute these tasks:
>1) Run each in their own thread
>2) Give both to a thread pool or Executor, where they might be executed
> sequentially by a single thread.
>Now, I will argue feature 1: assuming you aren't playing games with
>ThreadLocals or thread identity, using (2) should never be worse than
>using (1). If using (2) could be worse, Executors become unattractive and
>debugging nearly impossible.
>In particular, say that
> a) some other thread has published a reference to an object x while
> x is being constructed
> b) r1 loads a reference to x in a way that doesn't guarantee
> the object is initialized before the reference is loaded
> c) r2 loads a reference to x, but uses synchronization to
> guarantee that the reference is not loaded until x is
> completely initialized.
>Now, I argue that even if r1 and r2 are run in the same thread, r2 should see
>the correctly initialized version of x. Otherwise, "spooky action at a
>distance" can cause serious errors in your software.
>If you agree with feature 1, then I further argue for feature 2: if task
>r3 loads two difference references to x, the first incorrectly
>synchronized and the second correctly synchronized, then loads through the
>second reference should be valid, even if loads through the first
>reference are not. The reason for feature 2 is that I think it would be
>very difficult to distinguish between feature 1 and feature 2 in the semantics.
But we could require that the thread pool or Executor do something before
starting each task. Thus it is not necessary that feature 2 be
unconditional. An alternative feature 2': if task r3 loads two different
references to x, the first incorrectly synchronized and the second
correctly synchronized and does [fill in the blank] between the first and
the second, then ...
Obviously it is preferable to fill in the blank with something cheap, but
it doesn't have to be totally free.
>Before I go further in discussing the implications of these features, I
>want to see if there is general support that these two features are a
>requirement for a memory model that supports reliable software.
>JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel
JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel
This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:38 EDT