Re: [PATCH 0/9] Implement wake event support on Tegra186 and later

[Marc Zyngier]

On Mon, 08 Oct 2018 08:14:53 +0100,
Stephen Boyd <swboyd@xxxxxxxxxxxx> wrote:
> 
> Quoting Lina Iyer (2018-09-25 10:16:05)
> > Thanks Linus, for bringing this to my attention.
> > 
> > Hi Thierry,
> > 
> > On Tue, Sep 25 2018 at 03:57 -0600, Thierry Reding wrote:
> > >On Tue, Sep 25, 2018 at 10:48:39AM +0200, Linus Walleij wrote:
> > >> Hi Thierry,
> > >>
> > >> thanks for working on the wakeup business!
> > >>
> > >> This patch gets me a bit confused on our different approaches
> > >> toward wakeups in the kernel, so I included Lina, Marc and Ulf
> > >> to see if we can get some common understanding.
> > >>
> > >> On Fri, Sep 21, 2018 at 12:25 PM Thierry Reding
> > >> <thierry.reding@xxxxxxxxx> wrote:
> > >>
> > >> > The following is a set of patches that allow certain interrupts to be
> > >> > used as wakeup sources on Tegra186 and later. To implement this, each
> > >> > of the GPIO controllers' IRQ domain needs to become hierarchical, and
> > >> > parented to the PMC domain. The PMC domain in turn implements a new
> > >> > IRQ domain that is a child to the GIC IRQ domain.
> > >> >
> > >> > The above ensures that the interrupt chip implementation of the PMC is
> > >> > called at the correct time. The ->irq_set_type() and ->irq_set_wake()
> > >> > implementations program the PMC wake registers in a way to enable the
> > >> > given interrupts as wakeup sources.
> > >> >
> > >> > This is based on a suggestion from Thomas Gleixner that resulted from
> > >> > the following thread:
> > >> >
> > >> >         https://lkml.org/lkml/2018/9/13/1042
> [...]
> > >
> > >Yes, there was some good discussion in that thread which helped me come
> > >up with this solution. I think it's pretty elegant because it allows all
> > >of this interaction to happen almost automatically via the existing
> > >infrastructure. I'm not sure the same could be applied to the PDC,
> > >though, because of the need to manually replay the interrupt. That's not
> > >something I think can be done with just the simple parent/child
> > >relationship that we use on Tegra.
> > >
> > I wasn't able to use the hierarchy because not all GPIOs and the summary
> > line are routed to the PDC. But I am exploring options of hierarchy as
> > well.
> > 
> 
> From reading this thread (and https://lkml.org/lkml/2018/9/17/756) it
> looks almost exactly the same. The only difference is that Nvidia Tegra
> does the replay in hardware whereas Qualcomm SDM845 decided to replay
> the irq in software. Either way, the gpio controller has two parent
> domains, one is wakeup capable (PDC or PMC) and the other is not (GIC)
> and some wakeup capable irqs only go through the PDC/PMC and then to the
> GIC (e.g. RTC) instead of through gpio first. And it sounds like not all
> gpios are wakeup capable in both designs.
> 
> The plan to have the gpio to wakeup capable irq map live in DT for the
> PMC sounds good too. That way, the wakeup domain alloc function can
> figure things out and redirect gpios by itself while the gpio controller
> doesn't need to do anything special besides ask for wakeup to be set and
> fail if the parent can't support it.
> 
> Can hierarchical irq domains entirely replace the chained irqchip code
> in gpiolib? That would be interesting.

I'm not convinced this is generally doable. Most GPIO blocks multiplex
the signalling on a single parent interrupt, meaning that although you
may be able to have a hierarchy extending to that point, it can't go
any further (at which point you're back into chained-irq land). It
doesn't mean it invalidates the above design, but it probably requires
a bit of flexibility.

I must admit having slightly lost track of the intricacies of the QC
design, but we already have a set of interrupt controllers whose sole
task is to generate wake-up events. They are well behaved though, in
the sense that they will regenerate edges that the QC HW drops on the
floor.

The main issue I can see is that the QC HW relies on some signal other
than the normal interrupt we can handle, and this completely breaks
the very notion of a hierarchy. You need some "side-band signalling"
which will re-inject the lost edges. That, on its own, is a bit of a
deal-breaker.

Thanks,

	M.

-- 
Jazz is not dead, it just smell funny.