Re: [v4,1/9] ACPI / PM: Let acpi_dev_pm_detach() return an error code

[Dmitry Torokhov <dmitry.torokhov@gmail.com>]

On Thu, Sep 18, 2014 at 01:20:49AM +0200, Ulf Hansson wrote:
> On 17 September 2014 22:10, Dmitry Torokhov <dmitry.torokhov@gmail.com> wrote:
> > On Wed, Sep 17, 2014 at 08:25:44PM +0200, Ulf Hansson wrote:
> >> On 16 September 2014 01:36, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> >> > On Monday, September 15, 2014 09:53:59 AM Dmitry Torokhov wrote:
> >> >> On Sun, Sep 14, 2014 at 06:38:58PM +0200, Rafael J. Wysocki wrote:
> >> >> > On Friday, September 12, 2014 02:05:53 PM Dmitry Torokhov wrote:
> >> >> > > Hi Ulf,
> >> >> > >
> >> >> > > On Tue, Sep 09, 2014 at 01:36:02PM +0200, Ulf Hansson wrote:
> >> >> > > > To give callers the option of acting on a errors while removing the
> >> >> > > > pm_domain ops for the device in the ACPI PM domain, let
> >> >> > > > acpi_dev_pm_detach() return an int to provide the error code.
> >> >> > >
> >> >> > > So how would callers handle the errors? As far as I can see
> >> >> > > acpi_dev_pm_detach() is called from ->remove() and ->shutdown() methods, where
> >> >> > > there is no meaningful strategy to handle errors as you are past the point of
> >> >> > > no return and you keep on tearing down the device.
> >>
> >> The benefit is only relevant when ACPI and genpd PM domains would
> >> co-exist. In that case we might be able to skip genpd_dev_pm_detach()
> >> if acpi_dev_pm_detach() succeeds. So, currently there are  no benefit,
> >> but still it doesn't hurt.
> >
> > It doe snot have any negative material effect, the drawback is purely
> > from API perspective.
> >
> >>
> >> >> >
> >> >> > This is specifically for what patch [3/9] is doing AFAICS.
> >> >> >
> >> >> > The existing callers don't need to worry about this.
> >> >>
> >> >> OK, so I have the very same comment about patch 3 then: we have
> >> >> dev_pm_domain_detach() returning error. How would the callers handle errors?
> >> >
> >> > Ulf?
> >>
> >> I see your point. How about making dev_pm_domain_detach() to be a void
> >> function instead?
> >
> > Yes, please.
> 
> OK!
> 
> >
> >>
> >> >
> >> >> WRT this patch: I'd rater we did not just return generic "error code" just
> >> >> because we do not know who manages PD for the device. Can we add API to check
> >> >> if we are using ACPI to manage power domains? Then patch #3 could check if it
> >> >> needs to use ACPI or generic power domain API.
> >>
> >> The problem is scalability. If we have other PM domains implementation
> >> in future, each of them need to be checked prior invoking the attach
> >> functions.
> >> Also, how would we distinguish between genpd and a new PM domain XYZ?
> >
> > I do not think that trying all available methods to detach a pm domain,
> > i.e.
> >
> >         err = acpi_dev_pm_detach();
> >         if (err)
> >                 err = blah_dev_pm_detach();
> >         if (err)
> >                 err = flab_dev_pm_detach();
> >         if (err)
> >                 err = gen_dev_pm_detach();
> >
> > is any better from scalability point of view. If you need to do that you
> > will probably have to store something like "struct pd_ops *pd_ops" in
> > your device and call appropriate implementation via it.
> 
> No, that's not needed. Go ahead and have look at both ACPI and genpd,
> the interesting part is the validation of struct dev_pm_domain pointer
> in the struct device. That's all there is to it, no additional data
> are required.

OK, so can you simply put the needed method into struct dev_pm_domain and then
dev_pm_domain_detach() would become:

void dev_pm_domain_detach(struct device *dev, bool power_off)
{
	if (dev->pm_domain)
		dev->pm_domain->detach(dev, power_off);
}

Thanks.

-- 
Dmitry