From mboxrd@z Thu Jan  1 00:00:00 1970
From: will.deacon@arm.com (Will Deacon)
Date: Fri, 4 Mar 2011 11:47:50 -0000
Subject: [PATCH 6/6] ARM: nmk: update GPIO chained IRQ handler to use EOI
	in parent chip
In-Reply-To: <alpine.LFD.2.00.1103021806560.2701@localhost6.localdomain6>
References: <1298900022-21516-1-git-send-email-will.deacon@arm.com>
	<1298900022-21516-7-git-send-email-will.deacon@arm.com>
	<20110228140327.GA1937@n2100.arm.linux.org.uk>
	<001401cbd772$a1f93ae0$e5ebb0a0$@deacon@arm.com>
	<alpine.LFD.2.00.1102281949050.2701@localhost6.localdomain6>
	<20110228214445.GD1937@n2100.arm.linux.org.uk>
	<alpine.LFD.2.00.1103011140560.2701@localhost6.localdomain6>
	<20110301201904.GA27107@n2100.arm.linux.org.uk>
	<alpine.LFD.2.00.1103012133530.2701@localhost6.localdomain6>
	<001301cbd8ef$29403c10$7bc0b430$@deacon@arm.com>
	<alpine.LFD.2.00.1103021806560.2701@localhost6.localdomain6>
Message-ID: <000401cbda61$fd8759e0$f8960da0$@deacon@arm.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

Hi Thomas,

> > > irqchip1.c
> > >
> > > struct irq_chip1;
> > >
> > > handle_primary_irq(int irq, struct irq_desc *desc)
> > > {
> > > 	chip->irq_ack();
> > > 	desc->demux();
> > > }
> > >
> > > init()
> > > {
> > > 	irq_set_chip(PRIMARY_IRQ, &irq_chip1);
> > > 	irq_set_primary_handler(PRIMARY_IRQ, handle_primary_irq);
> > > }
> >
> > I think with this approach you get the exact opposite problem; that
> > is the primary irq_chip doesn't know which IRQs are going to be
> > demuxed so it cannot know at init time which IRQs need their primary
> > handler set. Is the idea that you set_primary_handler for all IRQs,
> > stash that in the descriptor somewhere and then replace handle_irq
> > with the primary handler when a demux handler is registered?
> >
> > I guess I'm missing something here,
> 
> No, I missed that. Darn, yes this would need storing the
> primary_handler in some separate pointer or having a callback into the
> chip implementation to make this fully distangled.

You could remove the need for pointer switching by having a bitmap
indicating the flow control supported by the primary chip and then
having a single primary handler implementation which checks the map.
Not especially nice but it does allow for demux handlers to exist
where they are sufficient.

Will