Re: Lack of suspend/resume/shutdown ordering between GPIO providers and consumers
From: Rafael J. Wysocki
Date: Wed Apr 25 2018 - 17:32:38 EST
On Wednesday, April 25, 2018 8:14:35 PM CEST Dmitry Torokhov wrote:
> On Wed, Apr 25, 2018 at 10:00:31AM -0500, Rob Herring wrote:
> > On Tue, Apr 24, 2018 at 5:58 PM, Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
> > > Hi Linus, Rafael, all
> > >
> > > Our GPIO controller driver: gpio-brcmstb.c has a shutdown callback which
> > > gets invoked when the system is brought into poweroff aka S5. So far so
> > > good, except that we also wish to use gpio_keys.c as a possible wake-up
> > > source, so we may have a number of GPIO pins declared as gpio-keys that
> > > allow the system to wake-up from deep slumber.
> > >
> > > Recently we noticed that we could easily get into a state where
> > > gpio-brcmstb.c::brcmstb_gpio_shutdown() gets called first, and then
> > > gpio_keys.c::gpio_keys_suspend() gets called later, which is too late to
> > > have the enable_irq_wake() call do anything sensible since we have
> > > suspend its parent interrupt controller before. This is completely
> > > expected unfortunately because these two drivers are both platform
> > > device instances with no connection to one another except via Device
> > > Tree and the use of the GPIOLIB APIs.
> > >
> > > First solution is to make sure that gpio-keys nodes are declared in
> > > Device Tree *before* the GPIO controller. This works because Device Tree
> > > nodes are probed in the order in which they are declared in Device Tree
> > > and that directly influences the order in which platform devices are
> > > created. Problem with that is that this is easy to miss and it may not
> > > work with overlays, kexec reconstructing DT etc. etc.
> > I'm going to make of_platform_populate randomize the order it creates devices...
> > > Another possible solution would be have the GPIO controller nodes have
> > > the GPIO consumers nodes such as gpio-keys, gpio-leds etc., and that
> > > would allow the Linux device driver model to create an appropriate
> > > child/parent relationship. This would unfortunately require Device Tree
> > > changes everywhere to make that consistent, and it would be a special
> > > case, because not all GPIO consumers are eligible as child nodes of
> > > their parent GPIO controller, there are plenty of other consumers that
> > > are not suitable for being moved under a parent GPIO controller node.
> > > This would also mean that we need to "probe" GPIO controller nodes to
> > > populate their child nodes (e.g: of_platform_bus_populate).
> > >
> > > I am thinking a more generic solution might involve some more complex
> > > tracking of the provider <-> consumer, but there is room for breakage.
> > That's what device connections are for. It probably just needs the
> > GPIO core to create the links. (but I've not looked into it at all).
> Not all APIs accept device as parameter to easily create links. But I
> wonder, for cases like this, if we could not simply move the device to
> the end of the dpm list after successful binding it to a driver. The
> assumption that when GOPIs or other resources are not ready they'll
> return -EPROBE_DEFER and probing would fail.
Not just to the end of dpm_list if shutdown is involved.
Also if you need runtime PM to follow the dependencies, this isn't