Re: [PATCH v4 3/5] irqchip: Add DT binding doc for the virtual irq demuxer chip

[Mark Rutland <mark.rutland@arm.com>]

On Wed, Feb 11, 2015 at 03:07:48PM +0000, Rafael J. Wysocki wrote:
> On Wednesday, February 11, 2015 02:14:37 PM Mark Rutland wrote:
> > On Wed, Feb 11, 2015 at 02:31:18PM +0000, Rafael J. Wysocki wrote:
> > > On Wednesday, February 11, 2015 11:15:17 AM Mark Rutland wrote:
> > > > On Wed, Feb 11, 2015 at 09:11:59AM +0000, Peter Zijlstra wrote:
> > > > > On Tue, Feb 10, 2015 at 08:48:36PM +0000, Mark Rutland wrote:
> > > > > > From f390ccbb31f06efee49b4469943c8d85d963bfb5 Mon Sep 17 00:00:00 2001
> > > > > > From: Mark Rutland <mark.rutland@arm.com>
> > > > > > Date: Tue, 10 Feb 2015 20:14:33 +0000
> > > > > > Subject: [PATCH] genirq: allow mixed IRQF_NO_SUSPEND requests
> > > > > > 
> > > > > > In some cases a physical IRQ line may be shared between devices from
> > > > > > which we expect interrupts during suspend (e.g. timers) and those we do
> > > > > > not (e.g. anything we cut the power to). Where a driver did not request
> > > > > > the interrupt with IRQF_NO_SUSPEND, it's unlikely that it can handle
> > > > > > being called during suspend, and it may bring down the system.
> > > > > > 
> > > > > > This patch adds logic to automatically mark the irqactions for these
> > > > > > potentially unsafe handlers as disabled during suspend, leaving actions
> > > > > > with IRQF_NO_SUSPEND enabled. If an interrupt is raised on a shared line
> > > > > > during suspend, only the handlers requested with IRQF_NO_SUSPEND will be
> > > > > > called. The handlers requested without IRQF_NO_SUSPEND will be skipped
> > > > > > as if they had immediately returned IRQF_NONE.
> > > > > > 
> > > > > > Cc: Boris Brezillon <boris.brezillon@free-electrons.com>
> > > > > > Cc: Jason Cooper <jason@lakedaemon.net>
> > > > > > Cc: Nicolas Ferre <nicolas.ferre@atmel.com>
> > > > > > Cc: Peter Zijlstra <peterz@infradead.org>
> > > > > > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > > > > Cc: Thomas Gleixner <tglx@linutronix.de>
> > > > > > Signed-off-by: Mark Rutland <mark.rutland@arm.com>
> > > > > 
> > > > > Aw gawd.. not that again.
> > > > 
> > > > I agree this isn't pretty, but at least it doesn't require the HW
> > > > description to know about Linux internals, and it can work for !DT
> > > > systems.
> > > > 
> > > > I'm really not happy with placing Linux implementation details into
> > > > DTBs.
> > > > 
> > > > > So Rafael and tglx went over this a few months ago I think:
> > > > > 
> > > > >   lkml.kernel.org/r/26580319.OZP7jvJnA9@vostro.rjw.lan
> > > > > 
> > > > > is the last series I could find. Maybe Rafael can summarize?
> > > > 
> > > > I can't get at any commentary from that link, unfortunately.
> > > > 
> > > > Rafael?
> > > 
> > > Well, the commentary is not there, because both I and Thomas implicitly agreed
> > > on one thing: We cannot add any suspend-related checks to the interrupt handling
> > > hot path, because that will affect everyone including people who don't use
> > > suspend at all and who *really* care about interrupt handling performance.
> > 
> > That's fair enough, and I'm happy to avoid that by other means.
> > 
> > My fundamental objection(s) to the current approach is that we create a
> > binding for a non-existent device that people will abuse without
> > considering the consequences. All we will end up with is more DTBs
> > containing the mux regardless of wether the drivers (or hardware) are
> > actually safe with a shared line.
> 
> That is a valid concern in my view.
> 
> > So with the changes moves out of the hot-path (e.g. with shuffling
> > to/from a suspended_actions list in the pm code), is there some issue
> > that I have not considered?
> 
> When we were reworking the handling of wakeup interrupts during the 3.18
> cycle, one of my proposals was to move the "suspended" irqactions to an
> "inactive" list during suspend_device_irqs() and back during
> resume_device_irqs(), but Thomas didn't like that approach.  His main
> argument was that it made the code in question overly complex which
> was fair enough to me.

I've just looked into that, and have a trivial implementation that's
contained within kernel/irq/pm.c, but it assumes that during suspend
nothing needs to modify actions.

> What about adding a new flag like I said?

That works for me. I'll respond in the other mail.

Thanks,
Mark.