From: Tony Williams 

Frank Haber wrote:
> I remember zombie processes in UNIX. I'm astonished that that page you
> referred us to seems to think that a kill can be defined as "the message
> to go die was sent and probably received."  Sounds like typical
> programmer optimism to me, filtered through a rosy haze of object
> orientation.  You may now throw things.

The message almost certainly will have been received by the kernel, but if
the process never gets to a point where the kernel can safely terminate it
that doesn't help much.

> Modern hung stuff seems more driver- or hardware-based, usually (gut
> feeling).

Exactly the cases where it can be unsafe to interrupt the process
(depending on the specifics of processor hardware and the device driver's
execution environment).

> Microsoft is now talking about "new cancel standards" in Vista, with
> file-based stuff to back up the thread-based cancels.  Apparently the
> snoop or monitor processes can't be trusted to know the score.
> Concurrency is truly a bitch, and I'm gradually learning why the
> definition of "race condition" has expanded so much.  It
apparently can
> even now include a cancel that happens just as the target has
> successfully completed.

I find that last one hard to believe. Haven't they heard of spinlocks and semaphores?

--
Tony

> What a nightmare.  A lock is not a lock, and a
> block is not a block. OMG.

--- BBBS/NT v4.01 Flag-5