Re: Wrongly wakened pthread_cond_wait() on 2.4.18 : futex.c?

Am Fre, 2002-12-27 um 13.10 schrieb Liu, Bing Wei:
> Patching stock 2.4.18 with patch-kngpt-2.4.18-2.0.3-2 downloaded from
> http://www-124.ibm.com/pthreads/download/stable/kernel/, I run Dr. Uwe's
> producer/consumer case against ngpt-2.0.4 on a SMP box and found it's 
> trapped in endless loop.
> 
> According to the printed info, I noticed that either producer or consumer
> will
> be hooked to endlessly wakened pthread_cond_wait() and prevent other 
> threads from grabbing the shared mutex.  Because it is using 
> PROCESS_SHARED condition variable, I traced back to the futex lines 
> and found a possible reason.
> 
> Within futex_poll() of kernel/futex.c, line 246,247 read
> 	if (!list_empty(&q->list))
> 		ret = POLLIN | POLLRDNORM;
> which means the select() on futex_fd will return right away.
> 
> Changing to
> 	if (list_empty(&q->list))
> 		ret = POLLIN | POLLRDNORM;
> and recompile the kernel, the issue will be gone.
> 
> Please correct me if I misunderstood the functions.
> 

You are correct - and spent some time on a bug that is known
and also fixed in the patch for 2.4.19. Unluckily is isn't
applied to the one against 2.4.18 :(