Re: [boost] Boost Threads Condition may block forever on clockchange

From: Johan Nilsson (r.johan.nilsson_at_[hidden])
Date: 2007-11-29 06:13:56

Anthony Williams wrote:
> Johan Nilsson <r.johan.nilsson <at> gmail.com> writes:
>

[snip]

>> For Windows: As you said yourself earlier, WM_TIMECHANGE is
>> available. Create a hidden window and run the message loop in a
>> worker thread to catch those msgs. This could presumably be
>> implemented using some kind of lazy init to work transparently even
>> for the DLL version of Boost.Thread.
>
> Thanks for the suggestion, but I don't like the idea of forcing the
> creation of a background thread with a hidden Window just to handle
> the rare event of the system clock being changed during a wait. Also,
> you've still got to wake the waiting threads. You could do this with
> a manual-reset event, but how do you know to reset it?

[Caveat: just thinking out loud here]

What about using user-mode APCs? Queue the user APCs from the background 
thread when the time changes. Whenever a thread is created, they could 
register a APC do-nothing callback (and remove on destruction). For all 
timed waits inside the thread library, use the alertable Wait functions 
(WaitXxxEx) and handle the APC delivery accordingly.

>
> Another option is just to have the wait divide the wait up so it only
> waits for a maximum of some small period (e.g. 100ms or 10s or
> something) before waking up and checking the system time, and
> resuming waiting if the time has not expired. Again, this adds
> overhead.

Well, that's also susceptible to a race. You really can't know for sure how 
much time that elapses between your sleeps and wake-ups (and adjust the 
total according to that).

/ Johan