On Mon, 2010-10-25 at 21:15 +0200, Jan Kiszka wrote:
> Am 25.10.2010 21:08, Philippe Gerum wrote:
> > On Mon, 2010-10-25 at 20:10 +0200, Jan Kiszka wrote:
> >> Am 25.10.2010 18:48, Philippe Gerum wrote:
> >>> On Wed, 2010-10-13 at 16:52 +0200, Philippe Gerum wrote: 
> >>>>>
> >>>>> Should we test IPIPE_STALL_FLAG on all but current CPUs?
> >>>>
> >>>> That would solve this particular issue, but we should drain the pipeline
> >>>> out of any Xenomai critical section. The way it is done now may induce a
> >>>> deadlock (e.g. CPU0 waiting for CPU1 to acknowledge critical entry in
> >>>> ipipe_enter_critical when getting some IPI, and CPU1 waiting hw IRQs off
> >>>> for CPU0 to release the Xenomai lock that annoys us right now).
> >>>>
> >>>> I'll come up with something hopefully better and tested in the next
> >>>> days.
> >>>>
> >>>
> >>> Sorry for the lag. In case that helps, here is another approach, based
> >>> on telling the pipeline to ignore the irq about to be detached, so that
> >>> it passes all further occurrences down to the next domain, without
> >>
> >> Err, won't this irritate that next domain, ie. won't Linux dump warnings
> >> about a spurious/unhandled IRQ? I think either the old handler shall
> >> receive the last event or no one.
> > 
> > Flipping the IRQ modes within a ipipe_critical_enter/exit section gives
> > you that guarantee. You are supposed to have disabled the irq line
> > before detaching, and critical IPIs cannot be acknowledged until all
> > CPUs have re-enabled interrupts at some point. Therefore, there are only
> > two scenarii:
> > 
> > - irq was disabled before delivery, and a pending interrupt is masked by
> > the PIC and never delivered to the CPU.
> > 
> > - an interrupt sneaked in before disabling, it is currently processed by
> > the pipeline in the low handler on some CPU, in which case interrupts
> > are off, so a critical IPI could be acked yet, and the irq mode bits
> > still allow dispatching to the target domain on that CPU. The assumption
> > which is happily made is that only head domains are interested in
> > un-virtualizing irqs, so the dispatch will happen immediately, while the
> > handler is still valid (actually, we are not allowed to un-virtualize
> > root irqs, and intermediate Adeos domains are already considered as
> > endangered species, so this is fine).
> > 
> >>
> >> Why this complex solution, why not simply draining (via critical_enter
> >> or whatever) - but _after_ xnintr_irq_detach, ie. while the related
> >> resources are still valid?
> >>
> > 
> > Because it's already too late. You have cleared the handler pointer when
> > un-virtualizing via xnarch_release_irq, and the wired irq dispatcher or
> > the log syncer on another CPU could then branch to eip $0.
> 
> Just make ipipe_virtualize_irq install a nop handler instead of NULL.

This does not solve the issue of the last interrupt which should be
processed. You don't want to miss it.

> 
> Jan
> 
> > 
> > And the solution is - reasonably - complex because xnintr_detach has
> > quite a few inter-deps. Typically, you may not drop the lock Xenomai
> > holds on the irq descriptor before calling xnarch_release_irq, to avoid
> > a race with xnintr_irq_handler in SMP (you could get a NULL cookie
> > there).
> > 
> > I would have preferred to have ipipe_virtualize_irq drain the
> > interrupts, but you just can't rely on a critical IPI while holding a
> > lock other CPUs might spin on irqs off. And you do need this code to
> > happen in a critical enter section, to act as a barrier wrt IRQ
> > dispatching. So the operation is unfold, the irq barrier first with irqs
> > on, then un-virtualizing the irq (for the relevant domain) with irqs
> > off.
> > 
> >> Jan
> >>
> > 
> 

-- 
Philippe.



_______________________________________________
Xenomai-help mailing list
[email protected]
https://mail.gna.org/listinfo/xenomai-help

Reply via email to