On 2012-09-14, at 03:18, Chris Muller <asqueaker at gmail.com> wrote:

>> Interesting, though completely wrong ;)
> 
> It couldn't be "completely" wrong since locking the image is a lot
> more wrong than not.

And a broken clock is right twice a day ;)

It changed the semantics of the whole system instead of only fixing the problem. We would have traded a clear-cut lockup for hard-to-debug problems caused by code running in a different process.

But I agree with Hannes that this was a great collaboration. John identified the problem and came up with a potential fix. That made it easy for me to write a correct version.

>> I just pushed a fix. It simply takes evaluating the alarm out of the critical section. Now the error shows just fine.
>> 
>> Don't know how to reproduce your double-firing problem.
> 
> The double-firing doesn't seem to happen with your fix.  Thanks!  I'm
> really glad to be able to put this lurking spectre bed.

Curious. My fix should not have any detectable effect, except for the error case. Hmm, maybe if an alarm callback schedules another alarm? That might be different.

- Bert -