[racket-dev] Kill-safe, single-write, blocking box (was Re: scheme_sema_post_all)

From: Matthew Flatt (mflatt at cs.utah.edu)
Date: Sat Oct 22 11:42:06 EDT 2011

I think you could get this behavior by creating a manager thread when
you create the new kind of box. If threads are too heavyweight, though,
you can get the effect of a primitive by using `ffi/unsafe/atomic'.

At Sat, 22 Oct 2011 10:24:27 -0400, Tony Garnock-Jones wrote:
> On 2011-10-22 9:43 AM, Tony Garnock-Jones wrote:
> > Nothing like the 20 seconds or so after a post to make one question
> > oneself. Could it be that semaphore-peek-evt could be used to get what I
> > need? I'll experiment.
> The answer is "almost", i.e. "no". But scheme_sema_post_all doesn't do
> what I want either. And I don't think having a thread issue an infinite
> sequence of (channel-put)s can be used either. I think I need something
> else. Something primitive, maybe.
>  - If I use semaphore-peek-evt or scheme_sema_post_all, I still have a
>    problem with kill safety, because I have to do something like:
>      (when (semaphore-try-wait? (blocking-box-used b))
>        (set-blocking-box-cell! b the-value)
>        (semaphore-post (blocking-box-ready b)))
>    ...which might be killed between the try-wait and the post.
>  - If I use a thread issuing an infinite sequence of channel-puts,
>     (thread (lambda ()
>               (when (semaphore-try-wait? (blocking-box-used b))
>                 (let loop ()
>                   (channel-put c v)
>                   (loop)))))
>    ...the custodian could be shut down at some point. Trying the same
>    trick as the buffered async channels doesn't work here, because I'd
>    need to know which thread to thread-resume when I checked the box's
>    value, and to do that I'd need a kill-safe box that can be written
>    into only once, which is an infinite regress.
> It looks like I need something like a cross between CAS and a semaphore.
> Perhaps I'm having imagination failure here. Is there something I'm
> overlooking that would get me an event to wait on until a value arrives,
> and that enforces that second and subsequent value-setting attempts do
> not succeed?
> (This is closely related to E's Promises and less closely related to
> Scheme's delay/force.)
> Regards,
>   Tony

Posted on the dev mailing list.