Re: [PATCH v8 3/5] amba: Don't unprepare the clocks if device driver wants IRQ safe runtime PM

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Tuesday, November 04, 2014 10:11:35 AM Ulf Hansson wrote:
> On 4 November 2014 02:57, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> > On Monday, November 03, 2014 10:41:02 AM Alan Stern wrote:
> >> On Mon, 3 Nov 2014, Russell King - ARM Linux wrote:
> >>
> >> > That makes it pretty horrid from the point of view of having bus
> >> > management code, because we now have the management of the bus clock
> >> > split between the bus layer and the device driver.
> >> >
> >> > This is /really/ a problem for runtime PM.  Runtime PM permits there
> >> > to be a bus layer involved - and runtime PM can also be coupled up
> >> > to PM domains as well.  For all this stuff, the context which the
> >> > callbacks are called in depends on whether the driver itself has
> >> > marked the device as having IRQ-safe callbacks.
> >> >
> >> > That's fine, but the bus and PM domain level code then /really/ needs
> >> > to know what context they're being called in, so they know whether
> >> > they can sleep or not, or they must to be written to always use
> >> > non-sleeping functions so they work in both contexts.  If we assume
> >> > the former, then that implies that the irq-safe flag must never change
> >> > state between a suspend and a resume.
> >>
> >> If a bus subsystem or PM domain is going to allow its drivers to choose
> >> between IRQ-safe and non-IRQ-safe runtime PM, then it is up to the
> >> subsystem to come up with a way for drivers to indicate their choice.
> >>
> >> I tend to agree with Rafael that testing dev->power.irq_safe should be
> >> good enough, with no real need for a wrapper.  But the subsystem can
> >> use a different mechanism if it wants.
> >>
> >> Bear in mind, however, that once the irq_safe flag has been set, the
> >> runtime PM core offers no way to turn it off again.
> >
> > There is a problem with it, though.  Say, a driver handles a device that
> > may or may not be in a power domain.  Or in other words, the power domain
> > the device is in may or may not be always on.  If the domain is always on,
> > the runtime PM callbacks are IRQ-safe (they depend on the driver only).
> > If it isn't, they may not be IRQ-safe.  How's the driver going to decide
> > whether or not to set power.irq_safe?
> 
> From my point of view; the decision whether the driver will set the
> IRQ safe flag is in principle a software design choice.
> 
> Currently genpd isn't able to power off, if one of its devices are IRQ
> safe configured. That's a limitation in genpd which we need to fix and
> it's on my TODO list.
> 
> My point is thus, I don't think the driver should care about PM
> domains at all regarding using the IRQ safe option. Does that make
> sense?

Yes, it does, and that's the heart of the problem above.  The driver should
not care about wherther or not the device is in a power domain, but it needs
to know that when deciding whether or not to set power.irq_safe.  Catch 22.

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.