Re: Ingo's realtime_preempt patch causes kernel oops

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, 2006-05-24 at 11:52 -0400, Steven Rostedt wrote:
> > > As I said yesterday. You need a demultiplexer for such cases.
> > > 
> > 
> > Would IRQs stay masked until the thread has finished running?
> 
> I would say yes.  But the system is basically broken if you have the
> same interrupt line that needs both to be threaded and NODELAY.
> 
> Basically, the best I can think to have for such a case, is all
> interrupt threads that have a shared NODELAY run at MAX_PRIO (99).  So
> that they act like a NODELAY interrupt, in that they run over everything
> else, but they can still schedule.

Err. That's why you use demultiplexers. The demux handler is always
NODELAY.

shared IRQ
-> demux_handler
	disable shared irq
	identify interrupt sources
	for all sources:
		calculate the interrupt number
		irq_desc[number]->handle_irq(.....)
			disable/ack a particular source
			if NODELAY 
				call handler and reenable if appropriate
			else 
				wakeup thread
	enable shared irq

That's the way you really want to do it. Granted, that this is not
possible with the current implementation of PCI cards, but for the SoC
peripherals this is usually simple to do.

The ARM tree has tons of examples which do exactly this.

	tglx