Re: [patch] Real-Time Preemption, -RT-2.6.10-rc1-mm3-V0.7.25-1

[Shane Shrybman <shrybman@aei.ca>]

On Fri, 2004-11-12 at 15:13, Ingo Molnar wrote:
> * Shane Shrybman <shrybman@aei.ca> wrote:
> 
> > V0.7.25-1 has been stable here with the ivtv driver for 11 hrs. No
> > sign of the ide dma time out issue either. Out of curiosity, do we
> > know what solved that problem?
> 
> could you try the attached patch - does it trigger the DMA timeouts
> again? There were 3 changes to the IOAPIC code that could have affected
> your dma-timeout problem, this patch reverts all of them.
> 

Yes it does trigger the DMA timeouts again. Just the addition of
CONFIG_SMP dep is enough to trigger it. Which isn't surprising since
the hack was to put #if 0 there wasn't it?

> Mark's suggestion sounds quite plausible too - but the question is, your
> timeout problems went away previously by tweaking io_apic.c, so it would
> be nice to see that they are still gone even with the old 'broken'
> io_apic.c logic. (none of the io_apic.c changes fixes any particular
> bug, they are only latency optimizations, so i'd be surprised if they
> really impacted your timeout problems.)
> 
> if the DMA timeouts are still gone even with this patch applied then i
> think it's safe to conclude that Mark's explanation is the correct one,
> and that it was starvation of the SCHED_OTHER IDE irq-thread that caused
> the timeouts: it _really_ was a timeout. (a workaround would be to make
> the timeout longer)
> 
> 	Ingo
> 
> --- linux/arch/i386/kernel/io_apic.c.orig2	
> +++ linux/arch/i386/kernel/io_apic.c	
> @@ -150,7 +150,7 @@ static void update_io_apic_cache(unsigne
>  	}
>  }
>  
> -#define IOAPIC_CACHE
> +// #define IOAPIC_CACHE
>  /*
>   * Some systems need a POST flush or else level-triggered interrupts
>   * generate lots of spurious interrupts due to the POST-ed write not
> @@ -188,7 +188,7 @@ static void __modify_IO_APIC_irq (unsign
>  		/*
>  		 * Force POST flush by reading:
>  	 	 */
> -		reg = *(IO_APIC_BASE(entry->apic)+4);
> +		reg = io_apic_read(entry->apic, 0x10 + pin*2);
>  #endif
>  		if (!entry->next)
>  			break;
> @@ -1940,7 +1940,7 @@ static unsigned int startup_level_ioapic
>   * unacked local APIC is dangerous on SMP as it can prevent the
>   * delivery of IPIs and can thus cause deadlocks.)
>   */
> -#if defined(CONFIG_PREEMPT_HARDIRQS) && defined(CONFIG_SMP)
> +#if defined(CONFIG_PREEMPT_HARDIRQS)
>  
>  static void mask_and_ack_level_ioapic_irq(unsigned int irq)
>  {
> 

shane