Re: [PATCH v3 1/2] PM: domains: Add devres variant for dev_pm_domain_attach()

[Rafael J. Wysocki]

On Mon, Jun 16, 2025 at 1:37 PM Claudiu Beznea <claudiu.beznea@xxxxxxxxx> wrote:
>
>
>
> On 16.06.2025 14:18, Rafael J. Wysocki wrote:
> > On Mon, Jun 16, 2025 at 11:37 AM Claudiu Beznea
> > <claudiu.beznea@xxxxxxxxx> wrote:
> >>
> >> Hi, Rafael,
> >>
> >> On 13.06.2025 13:02, Rafael J. Wysocki wrote:
> >>> On Fri, Jun 13, 2025 at 9:39 AM Claudiu Beznea <claudiu.beznea@xxxxxxxxx> wrote:
> >>>>
> >>>> Hi, Rafael,
> >>>>
> >>>> On 09.06.2025 22:59, Rafael J. Wysocki wrote:
> >>>>> On Sat, Jun 7, 2025 at 3:06 PM Jonathan Cameron <jic23@xxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> On Fri, 6 Jun 2025 22:01:52 +0200
> >>>>>> "Rafael J. Wysocki" <rafael@xxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> Hi Rafael,
> >>>>>>
> >>>>>>> On Fri, Jun 6, 2025 at 8:55 PM Dmitry Torokhov
> >>>>>>> <dmitry.torokhov@xxxxxxxxx> wrote:
> >>>>>>>>
> >>>>>>>> On Fri, Jun 06, 2025 at 06:00:34PM +0200, Rafael J. Wysocki wrote:
> >>>>>>>>> On Fri, Jun 6, 2025 at 1:18 PM Claudiu <claudiu.beznea@xxxxxxxxx> wrote:
> >>>>>>>>>>
> >>>>>>>>>> From: Claudiu Beznea <claudiu.beznea.uj@xxxxxxxxxxxxxx>
> >>>>>>>>>>
> >>>>>>>>>> The dev_pm_domain_attach() function is typically used in bus code alongside
> >>>>>>>>>> dev_pm_domain_detach(), often following patterns like:
> >>>>>>>>>>
> >>>>>>>>>> static int bus_probe(struct device *_dev)
> >>>>>>>>>> {
> >>>>>>>>>>     struct bus_driver *drv = to_bus_driver(dev->driver);
> >>>>>>>>>>     struct bus_device *dev = to_bus_device(_dev);
> >>>>>>>>>>     int ret;
> >>>>>>>>>>
> >>>>>>>>>>     // ...
> >>>>>>>>>>
> >>>>>>>>>>     ret = dev_pm_domain_attach(_dev, true);
> >>>>>>>>>>     if (ret)
> >>>>>>>>>>         return ret;
> >>>>>>>>>>
> >>>>>>>>>>     if (drv->probe)
> >>>>>>>>>>         ret = drv->probe(dev);
> >>>>>>>>>>
> >>>>>>>>>>     // ...
> >>>>>>>>>> }
> >>>>>>>>>>
> >>>>>>>>>> static void bus_remove(struct device *_dev)
> >>>>>>>>>> {
> >>>>>>>>>>     struct bus_driver *drv = to_bus_driver(dev->driver);
> >>>>>>>>>>     struct bus_device *dev = to_bus_device(_dev);
> >>>>>>>>>>
> >>>>>>>>>>     if (drv->remove)
> >>>>>>>>>>         drv->remove(dev);
> >>>>>>>>>>     dev_pm_domain_detach(_dev);
> >>>>>>>>>> }
> >>>>>>>>>>
> >>>>>>>>>> When the driver's probe function uses devres-managed resources that depend
> >>>>>>>>>> on the power domain state, those resources are released later during
> >>>>>>>>>> device_unbind_cleanup().
> >>>>>>>>>>
> >>>>>>>>>> Releasing devres-managed resources that depend on the power domain state
> >>>>>>>>>> after detaching the device from its PM domain can cause failures.
> >>>>>>>>>>
> >>>>>>>>>> For example, if the driver uses devm_pm_runtime_enable() in its probe
> >>>>>>>>>> function, and the device's clocks are managed by the PM domain, then
> >>>>>>>>>> during removal the runtime PM is disabled in device_unbind_cleanup() after
> >>>>>>>>>> the clocks have been removed from the PM domain. It may happen that the
> >>>>>>>>>> devm_pm_runtime_enable() action causes the device to be runtime-resumed.
> >>>>>>>>>
> >>>>>>>>> Don't use devm_pm_runtime_enable() then.
> >>>>>>>>
> >>>>>>>> What about other devm_ APIs? Are you suggesting that platform drivers
> >>>>>>>> should not be using devm_clk*(), devm_regulator_*(),
> >>>>>>>> devm_request_*_irq() and devm_add_action_or_reset()? Because again,
> >>>>>>>> dev_pm_domain_detach() that is called by platform bus_remove() may shut
> >>>>>>>> off the device too early, before cleanup code has a chance to execute
> >>>>>>>> proper cleanup.
> >>>>>>>>
> >>>>>>>> The issue is not limited to runtime PM.
> >>>>>>>>
> >>>>>>>>>
> >>>>>>>>>> If the driver specific runtime PM APIs access registers directly, this
> >>>>>>>>>> will lead to accessing device registers without clocks being enabled.
> >>>>>>>>>> Similar issues may occur with other devres actions that access device
> >>>>>>>>>> registers.
> >>>>>>>>>>
> >>>>>>>>>> Add devm_pm_domain_attach(). When replacing the dev_pm_domain_attach() and
> >>>>>>>>>> dev_pm_domain_detach() in bus probe and bus remove, it ensures that the
> >>>>>>>>>> device is detached from its PM domain in device_unbind_cleanup(), only
> >>>>>>>>>> after all driver's devres-managed resources have been release.
> >>>>>>>>>>
> >>>>>>>>>> For flexibility, the implemented devm_pm_domain_attach() has 2 state
> >>>>>>>>>> arguments, one for the domain state on attach, one for the domain state on
> >>>>>>>>>> detach.
> >>>>>>>>>
> >>>>>>>>> dev_pm_domain_attach() is not part driver API and I'm not convinced at
> >>>>>>>>
> >>>>>>>> Is the concern that devm_pm_domain_attach() will be [ab]used by drivers?
> >>>>>>>
> >>>>>>> Yes, among other things.
> >>>>>>
> >>>>>> Maybe naming could make abuse at least obvious to spot? e.g.
> >>>>>> pm_domain_attach_with_devm_release()
> >>>>>
> >>>>> If I'm not mistaken, it is not even necessary to use devres for this.
> >>>>>
> >>>>> You might as well add a dev_pm_domain_detach() call to
> >>>>> device_unbind_cleanup() after devres_release_all().  There is a slight
> >>>>> complication related to the second argument of it, but I suppose that
> >>>>> this can be determined at the attach time and stored in a new device
> >>>>> PM flag, or similar.
> >>>>>
> >>>>
> >>>> I looked into this solution. I've tested it for all my failure cases and
> >>>> went good.
> >>>
> >>> OK
> >>>
> >>>>> Note that dev->pm_domain is expected to be cleared by ->detach(), so
> >>>>> this should not cause the domain to be detached twice in a row from
> >>>>> the same device, but that needs to be double-checked.
> >>>>
> >>>> The genpd_dev_pm_detach() calls genpd_remove_device() ->
> >>>> dev_pm_domain_set(dev, NULL) which sets the dev->pm_domain = NULL. I can't
> >>>> find any other detach function in the current code base.
> >>>
> >>> There is also acpi_dev_pm_detach() which can be somewhat hard to find,
> >>> but it calls dev_pm_domain_set(dev, NULL) either.
> >>>
> >>>> The code I've tested for this solution is this one:
> >>>>
> >>>> diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> >>>> index b526e0e0f52d..5e9750d007b4 100644
> >>>> --- a/drivers/base/dd.c
> >>>> +++ b/drivers/base/dd.c
> >>>> @@ -25,6 +25,7 @@
> >>>>  #include <linux/kthread.h>
> >>>>  #include <linux/wait.h>
> >>>>  #include <linux/async.h>
> >>>> +#include <linux/pm_domain.h>
> >>>>  #include <linux/pm_runtime.h>
> >>>>  #include <linux/pinctrl/devinfo.h>
> >>>>  #include <linux/slab.h>
> >>>> @@ -552,8 +553,11 @@ static void device_unbind_cleanup(struct device *dev)
> >>>>         dev->dma_range_map = NULL;
> >>>>         device_set_driver(dev, NULL);
> >>>>         dev_set_drvdata(dev, NULL);
> >>>> -       if (dev->pm_domain && dev->pm_domain->dismiss)
> >>>> -               dev->pm_domain->dismiss(dev);
> >>>> +       if (dev->pm_domain) {
> >>>> +               if (dev->pm_domain->dismiss)
> >>>> +                       dev->pm_domain->dismiss(dev);
> >>>> +               dev_pm_domain_detach(dev, dev->pm_domain->detach_power_off);
> >>>
> >>> I would do the "detach" before the "dismiss" to retain the current ordering.
> >>
> >> I applied on my local development branch all your suggestions except this
> >> one because genpd_dev_pm_detach() as well as acpi_dev_pm_detach() set
> >> dev->pm_domain = NULL.
> >>
> >> Due to this I would call first ->dismiss() then ->detach(), as initially
> >> proposed. Please let me know if you consider it otherwise.
> >
> > This is a matter of adding one more dev->pm_domain check AFAICS, but OK.
>
> I don't know all the subtleties around this, my concern with adding one
> more check on dev->pm_domain was that the
> dev->pm_domain->dismiss() will never be called if the ->detach() function
> will be called before ->dismiss() and it will set dev->pm_domain = NULL (as
> it does today (though genpd_dev_pm_detach() and acpi_dev_pm_detach())).
>
> For platform drivers that used to call dev_pm_domain_detach() in platform
> bus remove function, if I'm not wrong, the dev->pm_domain->dismiss() was
> never called previously. If that is a valid scenario, the code proposed in
> this thread will change the behavior for devices that have ->dismiss()
> implemented.

->dismiss() and ->detach() are supposed to be mutually exclusive, so
this should not be a problem either way and in practice so far the
only user of ->dismiss() has been acpi_lpss_pm_domain which doesn't do
->detach() at all.

So this is your call.