On Mon, 2007-10-08 at 23:04 -0400, Steven Rostedt wrote:
> On Mon, Oct 08, 2007 at 11:45:23AM -0700, Mike Kravetz wrote:
> > On Fri, Oct 05, 2007 at 07:15:48PM -0700, Mike Kravetz wrote:
> > > After applying the fix to try_to_wake_up() I was still seeing some large
> > > latencies for realtime tasks.
> >
> > I've been looking for places in the code where reschedule IPIs should
> > be sent in the case of 'overload' to redistribute RealTime tasks based
> > on priority. However, an even more basic question to ask might be: Are
> > the use of reschedule IPIs reliable enough for this purpose. In the
> > code, there is the following comment:
> >
> > /*
> > * this function sends a 'reschedule' IPI to another CPU.
> > * it goes straight through and wastes no time serializing
> > * anything. Worst case is that we lose a reschedule ...
> > */
> >
> > After a quick read of the code, it does appear that reschedule's can
> > be lost if the the IPI is sent at just the right time in schedule
> > processing. Can someone confirm this is actually the case?
> >
> > The issue I see is that the 'rt_overload' mechanism depends on reschedule
> > IPIs for RealTime scheduling semantics. If this is not a reliable
> > mechanism then this can lead to breakdowns in RealTime scheduling semantics.
> >
> > Are these accurate statements? I'll start working on a reliable delivery
> > mechanism for RealTime scheduling. But, I just want to make sure that
> > is really necessary.
>
> For i386 I don't think so. Seems that the interrupt handler will set the
> current task to "need_resched" and on exit of the interrupt handler, the
> schedule should take place. I don't see the race (that doesn't mean
> there is one).
>
> For x86_64 though, I don't think that we schedule. All the reschedule
> vector does is return with a comment:
>
> /*
> * Reschedule call back. Nothing to do,
> * all the work is done automatically when
> * we return from the interrupt.
> */
> asmlinkage void smp_reschedule_interrupt(void)
> {
> ack_APIC_irq();
> }
>
> I'm thinking that this was the case for i386 a while back, and we fixed
> it for RT.
>
> /me does a quick search...
>
> http://lkml.org/lkml/2005/5/13/174
>
> Yep! This is a bug in x86_64. I'll fix this up tomorrow and send out a
> patch.
Hmm, my understanding is that the IPI caller needs to set
TIF_NEED_RESCHED before issuing the IPI.
So I'm inclined to not like this 'fix'.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
