Re: [RFC patch 2/3] genirq: Add buslock support for irq chips on slow busses

[Ingo Molnar <mingo@elte.hu>]

* Thomas Gleixner <tglx@linutronix.de> wrote:

> On Sat, 15 Aug 2009, Ingo Molnar wrote:
> > * Mark Brown <broonie@opensource.wolfsonmicro.com> wrote:
> > 
> > > On Fri, Aug 14, 2009 at 12:20:49PM +0200, Thomas Gleixner wrote:
> > > > On Fri, 14 Aug 2009, Pavel Machek wrote:
> > > 
> > > > > AFAICT this means that driver would need to know what kind of IRQ it
> > > > > is hooked to, right? That will lead to some ugly code in drivers that
> > > > > can handle both normal and slowbus irqs, right?
> > > 
> > > > Are there such drivers in reality ?
> > > 
> > > Yes.  The GPIO based stuff is the prime example but there's other 
> > > examples - one is the WM831x touchscreen (no driver in mainline 
> > > yet) which can use interrupts via the main interrupt controller on 
> > > the CPU but also has the option of bringing the interrupt signals 
> > > out to dedicated pins on the chip for direct connection to the CPU 
> > > precisely to avoid the overheads of these slow interrupt 
> > > controllers.
> > 
> > This would call for Thomas's first version of the patch, that is 
> > transparent to drivers - the IRQ subsystem will know how to lock 
> > access to the line.
> > 
> > How about implementing that first patch in a cleaner way - can we 
> > somehow express the slow-bus property purely via the irqchip? Or is 
> > that too lowlevel?
> 
> The problem here is that the management functions serialize via 
> irqdesc->lock and call the chip level function with the lock held 
> (preemption and interrupts disabled). So we can not access the 
> slow bus chips from these low level functions.
> 
> If the low level functions just store the information and schedule 
> it for bus access then there is no serialization anymore. Lets 
> look at disable_irq():
> 
> 	spin_lock_irqsave(&desc->lock);
> 	....
> 	desc->chip->mask();
> 		schedule bus access;
> 	...
> 	spin_lock_irqsave(&desc->lock);
> 
> So now we return, but the mask has not reached the chip.
> 
> The idea of bus_lock/bus_sync_unlock() was to provide well defined
> synchronization points for the chip level implementation to do the bus
> update and have this serialized against other management functions.
> 
> 	desc->chip->bus_lock();
> 		Take the chip->bus mutex
> 
> 	spin_lock_irqsave(&desc->lock);
> 	....
> 	desc->chip->mask();
> 		store mask information
> 	...
> 	spin_lock_irqsave(&desc->lock);
> 
> 	desc->chip->bus_sync_unlock();
> 		Update the mask via the slow bus
> 		Release chip->bus mutex
> 
> That way we have made sure that the change to the chip actually 
> hits the hardware before we allow further management operations 
> and it simplifies the code for the chip implementation as the bus 
> access can be done in the context of the caller w/o the need of an 
> extra thread/workqueue ...

Given the alternatives i'd prefer this one now ...

	Ingo