Re: [boost] win32 mutex and condition variable algorithms

From: Anthony Williams (anthony_w.geo_at_[hidden])
Date: 2006-09-20 08:58:36

Alexander Terekhov <terekhov_at_[hidden]> writes:

> Anthony Williams wrote:
> [...]
>>     // Any APC will break the sleep, so keep sleeping until we've been
>>     // notified, or we've timed out
>>     while(!atomic_read(new_node.notified)
>>         && SleepEx(milliseconds_until(timeout), true)==WAIT_IO_COMPLETION);
>> 
>>     lock(gate)
>
> (generally, access *this after waiting without any sync with dtor) 
>
> The gate being just a mutex (in 8a the gate is a semaphore 
> unlocked/posted by last exiting waiter in the case of no-NOP 
> signaling). You've got a problem here regarding POSIX safety with 
> respect to CV destruction. POSIX says that it "shall be safe to 
> destroy an initialized condition variable upon which no threads are 
> currently blocked" and even gives an example illustrating it using 
> broadcast() followed by destroy(). 8a handles it by synchronizing 
> with exiting waiters on a gate semaphore in CV dtor.

I currently haven't thought about destruction much. This is an important
issue, so thanks for raising it. I'll post afresh when I've got a solution.

Anthony
-- 
Anthony Williams
Software Developer
Just Software Solutions Ltd
http://www.justsoftwaresolutions.co.uk