condition notify_all problem

From: joshuaboelter (yahoo_at_[hidden])
Date: 2002-11-15 04:05:43

I've just started using the boost libraries, threads so far -- great 
libraries.  I have a test app that uses thread groups of producers 
and consumers working with a queue.  It seems simple enough, but I'm 
encountering a deadlock problem using notify_all() vs looping with 
notify_one() when ending the consumer threads.  I'm not sure my 
implementation is correct though (link to source below).

I can reliably cause deadlocks on all Windows XP systems I've tried 
(single, dual and quad machines) within about a minute of 
execution.  However, I only had one deadlock on Windows 2000 (single 
or quad) in over 6 hours of execution and still going.

Attaching or running the executable in the VS7 debugger shows all 
the threads and the main thread stopped in ntdll.dll via 
kernel32.dll.  The stack shows the last user code executed by wmain 
was boost::thread_group::join_all from ctg.join_all().  The threads 
last ran boost::condition::do_wait before entering the kernel.

Swapping lines 217-218 for 222-227 appears to fix the deadlock by 
looping with notify_one().  I would guess that something fundamental 
is wrong with my implementation, and my fix just changes the 
timing.  Any help is appreciated.

The source is posted at the link below.  I can provide further 
information if necessary.

Thanks,
Joshua Boelter

www.boelter.org/data/t_threadgroup.cpp

Using Visual Studio 7.0.9466, stlport 4.5.3 (dll), boost libraries 
1.29.0 on various WinXP Pro and Win2k systems.